Agricultural Production Functions.

AGRICULTURAL
PRODUCTION FUNCTIONS
Earl O. Heady
Professor of Economics
and C. F. Curtiss Distinguished Professor
of Agriculture,
Iowa State University
John L. Dillon
Research Officer,
Commonwealth Scientific and
Industrial Research Organization,
Melbourne, Australia
Iowa State University Press, Âmes, Iowa

© 1961 by The Iowa State University Press.
All rights reserved.
Library of Congress Catalog Card Number: 60-11128.
LITHOPRINTED IN THE UNITED STATES OF AMERICA BY

CUSHING - MALLOY, INC., ANN ARBOR, MICHIGAN, 1961
Preface
The application of formal production function concepts in agricul-

tural research is a relatively recent development* The area of analysis
was initiated by W. J. Spillman and other pioneer economists and physi-
cal scientists in agriculture. However, economists and physical scien-
tists fairly well “went their own way* and several decades went by with
littleco-ordinated research in the prediction of physical production and
in economic application of the results. However, with increased com-
mercialization of agriculture and greater economic literacy of farm
operators there is need for design of physical and biological research
so that the results can have greater economic interpretation and appli-
cation.
This text, considered as one in agricultural science rather than
purely in economics, summarizes certain concepts and methods relat-
ing to the prediction and use of agricultural production functions. It
also reports, in summary form, the results from selected production
function studies. Emphasis is on concepts, principles, and methodolog-
ical results. Practical applicaj£.cjn is considered to be a second step in
communicating principles and prediction for farmer use. However, a
few illustrative examples are included to indicate how complex data and
concepts can be interpreted and presented for practical use.
The studies reported are largely those resulting from co-operative
research efforts at Iowa State University over the past decade. These
studies include an important group of products and are drawn together
in this monograph for use of other scientists in economic and biological
phases of agriculture. Numerous studies represent pioneering efforts
in predicting production surfaces and in adapting them for economic in-
terpretations. The designs explained are not suggested as optimum for
the purposes, but are simply those which appeared to be appropriate at
the time a particular experiment was initiated or which were consistent
with the time and research funds available. Predictions for some of
these experiments, if designs such as those summarized in Chapter 5
might have been used, could have been improved. Subsequent research
has employed these designs or modifications of them. Functional forms
are not presented as those most appropriate under all environmental
conditions. Rather, they are used to illustrate the types of relation-
ships and recommendations which result when particular forms of
y
vi PREFACE
functions are employed, or serve most efficiently under a given set of

conditions.
The studies of empirical production functions relating to crops,
livestock, and poultry were possible only because of the scientific in-
terest, knowledge, and professional abilities of Damon V. Catron, John
T. Pesek, Stanley Balloun, and Norman L. Jacobson. Along with Gordon
C. Ashton, Roger Woodworth, Solomon Bloom, Vaughn C. Speer, John
A. Schnittker, William G. Brown, Robert McAlexander, Dean E. McKee,
Joseph Stritzel, John P. Doll, Gerald W. Dean, Harold O. Carter, C. C.
Culbertson, Owen W. McCarthy and R. P. Nicholson, these persons
should be considered as co-authors of this monograph. We have only
drawn together the numerous studies conducted by these persons, add-
ing a few chapters of interest to those conducting similar studies and
listing a collection of reading pertinent to agricultural production func-
tions. Accordingly, authorship of the persons named above is recog-
nized in the appropriate chapters and we should more properly be con-
sidered as editors of this monograph. Too, relative to the array of
farm -firm production function estimates presented in the final chapter,
we are grateful to Takashi Takayama, K. S. Suryanarayana, W.
Darcovich, T. Godsell, G. D. Agrawal, Y. Wang, George Mason, and
especially, Lennart Hjelm, Eje Sandqvist, and Yair Mundlak, for sup-
plementary information willingly given.
The authors hope that this text will serve to stimulate more and
improved research in agricultural production functions. Too, we hope
that it serves as foundation for further co-operative effort among per-
sonnel of biological, physical, and economic sciences. The concepts,
principles, and quantities presented are those relevant both for greater
scientific knowledge and more efficient and practical use of certain ag-
ricultural research.
In drawing together the set of studies conducted at Iowa State Uni-
versity, we do not depreciate or overlook those which have been com-
pleted or are underway at other research organizations and institu-
tions. Time, space, and publication costs posed restraints which could
not be eliminated at the time of preparing this monograph.
Earl O. Heady
John L. Dillon
May, 1960
Contents
1. Development of Production Function Studies 1
2. Economic Applications 31
3. Forms of Production Functions 73
4. Data Analysis for Production Function Estimation • . 108
5. Data Collection for Production Function Estimation 142
6. Economic Specification of the Production Function 195
7. Miscellaneous Empirical Problems Relating to

Estimation of Production Functions 218
8. Pork Production Functions for Hogs Fed in Drylot 266
9. Pork Production Functions and Substitution Coefficients

forHogs on Pasture 302
10. Production Functions, Least-cost Rations, and Optimum

and Optimum Marketing Weights for Broilers 330
11. Least-cost Rations and Optimum Marketing Weights

for Turkeys 374
12. Milk Production Functions and Marginal Rates

of Substitution Between Forage and Grain 404
13. Production Functions and Substitution Coefficients

for Beef 452
14. Crop Response Surfaces and Economic Optima

Use
in Fertilizer 475
vii
viii CONTENTS
15. Surfaces, Isoquants, and Isoclines From Fertilization 526
16. Functions for Fixed Plants and Other Farm Situations 554
17. Comparison of Production Function Estimates

From Farm Samples Over the World 585
Bibliography 645
Index 665
CHAPTER 1
Development of Production
Function Studies
HIS BOOK deals with agricultural production functions. Its pur-
T pose is to summarize certain concepts, empirical methods, and

quantitative researches which relate to or have been derived for
farm production functions. It covers both physical production functions
based on experiments with crops and livestock and firm production
functions based on cross-sectional or time series samples. In particu-
lar, it emphasizes research completed in these areas at Iowa State
University.
USE OF PRODUCTION FUNCTION CONCEPTS
The production function is a concept in physical and biological sci-

ence.However, it was largely developed and, until recently, used
mainly by economists. Historically, refinements in concepts relating
to production functions grew out of economics probably because of the
following reasons. (1) The nature of production functions is important
in economic development and in determining the extent to which national
products can be increased from given resource stocks. (2) The magni-
tude of production coefficients serve as the base for determining opti-
mum patterns of international or interregional trade. (3) The concept
is basic to certain theories in the functional distribution of income.
The conditions under which a total output can be imputed to the factors
from which it is produced with the product just exhausted depends on
the nature of the production function. (4) The production function pro-
vides half or one of two general categories of the data needed in deter-
mining or ^specifying the use of resources and the pattern of outputs
which maximize firm profits. (5) The algebraic nature of supply func-
tions rests, in large part, upon the nature of the production function.
Research workers in the physical and biological sciences of land-
grant colleges and the United States Department of Agriculture have
long conducted research providing information on the nature of agricul-
tural production functions. However, historically, this research was
designed and conducted somewhat apart from the formal concept of
production functions represented by regression equations. More typi-
cally, research was designed on the basis of discrete phenomena
1
2 DEVELOPMENT OF STUDIES
wherein two or a few treatments were used to provide point estimates
of crop or livestock output resulting from input of factors (materials
representing treatments) In some instances, although not designed for
.
these purposes, the data were sufficient for deriving simple regression
equations or input -output curves. More frequently, the experimental
designs and statistical procedures used have only allowed indication of
whether mathematically significant differences exist between the yield
or output level of two or three discrete treatments or input levels.
From these differences could be computed the relative profitability of
the few treatments or inputs. However, it was generally impossible to
apply refined economic principles in determining the most profitable
level of output and input, or the most profitable combination of inputs
for a specified output.
These designs and approaches have proved useful in the past and
may continue to do so under certain conditions. In many cases, re-
search workers in biological fields have been concerned only with esti-
mating the output from a specific quantity of new material which serves
as an innovation. Here the goal of the research often has been to
answer the question: Does the material or resource, used at any level
whatsoever, give a response? Much early research on fertilizer fell
in this framework. In some cases, the practice or treatment under
consideration represents a resource or material of discrete and limi-
tational nature. Artificial insemination is an example. Here there is
no important question of “dosage” and a formal production function ap-
proach is inappropriate. For other materials used in production, the
phenomena under consideration could have been estimated as a con-
tinuous function, but there was little need to so represent it. For ex-
ample, fertilizer rates recommended to farmers, based on a few point
estimates, have not always been as high as those which would maximize
profits in the classical sense of a farm operating with unlimited capital
under static conditions. However, because farmers operate in a
decision-making framework of uncertainty and limited funds, they often
have failed to use fertilizer inputs even as large as those recommended
on the basis of trials which include only a few discrete treatment levels.
A final reason might be given to justify experiments designed to give
point estimates of yield or output from a few discrete input levels: the
results provide points similar to the “straight line” segments assumed
in linear programming. The optima selected within this framework
of
assumed physical relationships always fall “on the corners,” repre-
sented by the point estimates. However, refinements allowing optima
to fall between these “corners” may not be important where price
un-
certainty is so great that ex poste accuracy in decisions can never
be
attained.
There are many biological and physical areas, however, where con-
tinuous relationships are involved and the data lend themselves
to for-
mal production function analysis. Also, in many of these areas, recom-
mendations to farmers could be made with greater economic meaning
if
the experimental design and statistical analysis were of a form
to allow
DEVELOPMENT OF STUDIES 3
prediction of the production functions involved. In the past, these pro-

cedures were not often used for several reasons. (1) The scientists
conducting the research often have used criteria other than economics
in interpreting their findings and in making recommendations to
farmers. In the past, for example, the criteria used often have been
rations which gave (a) the largest gain per pound of feed, (b) the great-
est output per cow or hen, or (c) the fastest daily gain. The most prof-
itable output or resource input is seldom identical with these maxima
or minima. However, even where the objective is prediction of a phys-
ical maximum or minimum, the exact quantity can be estimated more
accurately where the data, if of appropriate nature, are used to esti-
mate the regression equations representing the production functions in-
volved. Derivatives then can be computed and equated to zero, with the
appropriate magnitude of input then derived. (2) The statistical
methods serving as a guide for research were based on early biological
procedures which supposed the data to be discrete phenomena most ap-
propriate for point estimates. Early texts on experimental design and
statistical analysis emphasized these methods as appropriate for ex-
periments in biological sciences. (Early emphasis on continuous rela-
tionships and regression analysis, as characteristic of physical data
representative of production functions, was made especially by econo-
metricians.) (3) Many physical scientists have not been acquainted
with production function concepts and the economic principles which
define profit maximization or cost minimization. Since they were par-
ticularly concerned with conditions of profit maximization and compet-
ing economic alternatives within the farm or firm, economists have
been concerned with marginal products, marginal rates of substitution,
isoquants, and isoclines; quantities which are derived from continuous
production functions.
In the past decade, however, an increasing number of physical and
biological scientists in agriculture have become acquainted with pro-
duction function concepts. They have 'become interested in interpreting
basic relationships of nature accordingly and in using them to make
recommendations to farmers. This increased interest and activity
partly grows out of scientific advance and is parallel to that in the field
of agricultural economics. Generally, new fields of research start with
a less formal and precise set of concepts and models. Over the dec-
ades, continuous scientific investigation and thought provide refine-
ment to the theories, concepts, and models which serve as the basis of
designs in experiments and of principles in decision making. Pioneer
agricultural economists, like other agricultural scientists, also were
concerned with physical relationships which relate to production func-
tions. Without a well-defined set of concepts, they derived principles
termed “factors affecting farm profits.” The “factors affecting farm
profits” principles implicitly supposed certain conditions in respect to
physical production functions. Generally, these early findings in agri-
cultural economics served sufficiently in leading farmers to improve
their farming operations. But these less formal principles also have,
with the advent of time and improved knowledge by both farmers and
research workers, been replaced by more exact principles of profit
maximization; principles which directly require knowledge of the rele-
vant production and price quantities. Given recognition of these princi-
ples, both physical scientists and economists will have greater future
interest in quantities derived from production functions. While the re-
lationships represented by production functions are physical phenom-
ena, economic principle is involved when recommendations to farmers,
or decisions by them, relate to the quantity or mix of resources and
products to be used and produced respectively. Accordingly, more co-
operative research and educational activities can be expected among
economists, agronomists, engineers, and animal scientists.
NEED FOR PRODUCTION FUNCTION RESEARCH

This text emphasizes production function estimates of the two fol-
lowing types: (a) biological functions derived from experiments where
the plant is fixed at the magnitude of an animal or acre and (b) farm pro-
duction functions derived from samples where plant size, as measured
by “fixity” of a particular resource, is either fixed or variable. Per-
haps the most appropriate use of biological functions is that of guiding
farmers in their individual decisions. However, these same data can be
of extreme use for certain purposes of policy and economic develop-
ment. For example, a nation such as India may have limited facilities
for producing or purchasing fertilizer materials. With a given amount
of fertilizer for annual allocation to agriculture, how much should be
distributed to various soil, geographic, and climatic areas? If the goal
is to maximize the food product available from limited fertilizer re-
sources, production functions derived for major regions and crops, with
an indication of marginal coefficients, can provide a basis or guide for
attaining the goal. Similarly, the same type of information can be used
during wartime or during any other emergency period when material
shortages exist. On the other hand, farm production functions of the
type discussed in this text probably best serve for the diagnostic pur-
poses pointed out later. In this vein, they can provide general guidance
for farmers’ decisions, credit policy formulation, readjustment of agri-
cultural regions, etc. However, aside from studies based on highly re-
fined samples, such as some mentioned in later chapters, they seldom
can be used to indicate “exact equilibrium use” of resources by an in-
dividual farmer.
With the passage of time, a greater knowledge level of farmers, and
an increased commercialization of agriculture, there is increased need
for experimental designs and research in biological fields which lend
themselves to estimation of production functions. Data can then be
better adapted for economic interpretations and recommendations. An
increasing number of farmers has knowledge of principles of profit
maximization and wish to use fertilizer, feed, and similar resources in
a manner to approach formal economic optima. Not only does a greater

proportion of individual farmers have this knowledge, but also highly
specialized operations found in vegetable, cotton, grain, and large scale
and integrated livestock and poultry farming generally have manage-
ment personnel who seek out and use the appropriate economic princi-
ples and physical data. Under these highly commercial operations,
management is not satisfied to stop with use of a quantity of fertilizer
or a feed ration which is profitable. They are concerned with the most
profitable quantity and mix of ingredients. Under large-scale opera-
tions, large profits may be sacrificed unless fertilizer is used at levels
which give most profitable yields, or unless livestock rations are those
which minimize costs for given gains.
The need for data in a form which allows the correct or appropriate
economic recommendation is itself a sufficient reason why more bio-
logical research might well be based on production function concepts or
models. But there also are other reasons why this approach is desira-
ble. If data provided in this form came at great cost to society and
farmers, the appropriateness of the experimental designs and statisti-
cal procedures implied might well be questioned. The important de-
cision then hinges around this: Is the added gain from improved recom-
mendations and decisions greater than the added costs of the research
refinements? In many, perhaps the numerical majority of experiments,
adaptations of designs to allow estimates of production functions or
continuous input-output relationships have a very low marginal cost.
Frequently the marginal cost may be zero. Hence, as long as the ex-
periment is being conducted anyway, it might as well be designed to
allow the appropriate representation of the physical relationships in-
volved, and to allow application of appropriate economic principles.
Adaptations to allow production function estimates ordinarily require
few additional treatments, if the general range of information is to be
the same in both cases. The goal of some research is to determine the
magnitude of intra-treatment variance. But where this is not neces-
sary, the added cost of adapting an experiment to allow regression es-
timate of the functional input -output relationship can be at zero cost.
For example, a fertilizer experiment with three treatments replicated
three times might be converted to one with nine different treatments
with each representing a different fertilization level. The latter ap-
proach would not allow computation of error terms represented by
variance within individual treatments, but it would allow estimate of
the standard error of the regression coefficient and other probability
criteria which are appropriate for choice and decision making.
Where design of experiments to allow production function analysis
may result in slightly higher total costs in conducting experiments, the
cost per unit of scientific information obtained will generally be lower.
Formal estimation of production functions allows derivation of marginal
products, physical maxima or minima, isoquants, isoclines, marginal
rates of substitution, and other quantities which cannot be computed
from experiments which imply discrete data and which are designed
only to provide point estimates. From this standpoint, production func-
tion analysis has scientific as well as practical importance. Wherever
biological or physical relationships of nature are of continuous func-
tional form it would appear that the interest of even pure science would
be that of estimating them in this manner. With a large annual input of
funds already devoted to biological or physical research in land-grant
colleges, the United States Department of Agriculture, and other re-
search organizations, it would seem desirable that more of this be di-
verted to alternative but more complete methods of estimating the
basic relationships involved.
The amount of resources devoted to estimating farm and industry
production functions is much smaller than that devoted to technical
functions. The method is only one of many for providing knowledge of
agricultural structure and the foundation on which choice and decision
can be based. It is not a substitute for budgeting, programming, or
planning procedures which provide more broader estimates of optim um
resource use within a farm. It has important limitations as an empiri-
cal' procedure for accomplishing these ends. Accordingly, it will never
claim the absolute amount or proportion of research resources devoted
to estimating purely physical quantities and relationships. Yet it does
have some uses in assessing resource productivity in agriculture and
for decision making by individual farmers and policy administrators.
INTERDISCIPLINARY CO-OPERATION
IN PRODUCTION FUNCTION ESTIMATES
As mentioned
previously, production functions derived for fertiliza-
tion, irrigation, livestock feeding,
and similar agricultural processes
essentially are problems in technical sciences. However, most of the
formal production function research completed to date has resulted
from co-operative effort among scientists in physical and economic
fields. This has held true perhaps largely because the technological
models involved have had greatest conceptual use and refinement in
economic analysis and because there has been some need for technical
data better adapted for economic interpretation and use. In the
future it
is entirely likely that
physical scientists will independently conduct
most of these research projects. However, since much of the same
data will have greater applicability in economic recommendations
and
choice, continued co-operative effort between scientists in
physical and
economic fields appears desirable. Accordingly, we include some
notes relating to success of co-operative endeavors.
Biological Functions
Co-operative work on production functions among physical scien-

tists and economists at Iowa State University has been more or less
spontaneous. Also, it has mostly been on an informal basis between in-

dividuals, rather than on the basis of detailed and formal interdepart-
mental agreements. In most cases, the work has been done entirely
outside the realm of any type of joint project; the activity, research in-
terest, and mutual trust of the individuals serving alone as the "joint
ingredient.” An important element of the success has hinged on the
ability and interests of the physical scientists involved.
The first requirement for successful co-operative research is an
understanding by the physical scientist of the analytical models which
thus far have been emphasized in economics. The physical scientist
will be more enthusiastic over use of these models as a basis for his
experiments if he actually understands them. Although graduate study
is a help, it is not necessary that the physical scientist have graduate
courses in economics. The relevant number of concepts is small and
can be acquired in a few informal seminars. Informal sessions with
two or four people meeting four to six times often are more productive
than formal seminars with a room full of people where much time is
spent informing those on the "fringe of interest” who will not be en-
gaged in the research. Insufficient time may be devoted to the inter-
ests of the scientist who will actually initiate an experiment.
Second, the persons involved need to be genuinely interested in the
investigation. Successful co-operative research seldom comes about
by administrative edict alone wherein staff members are simply in-
formed that a joint project is to be put into effect. Informal arrange-
ment of projects at Iowa State University has been possible because all
parties have had a genuine interest in their work. Generally, persons
in relevant departments who promise to have an interest in the ap-
proach get together and go over the concepts and possible designs.
Then, if all see sufficient mutual interest in the approach, they get in
contact with the proper administrative officials.
A statement by administrators of the need for the research may fa-
cilitate co-operative effort, but will not assure highly productive co-
operative research. The real stimulus for scientific integration must
come from the individual research workers. They must see the need
for the prospective product; they must be willing to spend several
hours in informal seminars, prior to the actual initiation of the empiri-
cal work, in learning and understanding each other’s concepts, models,
and laws. The economist has as much to learn about the biological
phenomena and the experimental difficulties as the physical scientist
has to learn about the economic concepts. Frequently the agronomist
or animal scientist already has a grasp of the basic relationships in-
volved, but has not formalized them into the theoretical skeleton used
by the economist.
Third, the people involved must not be of the "lone -wolf” type. Co-
operative research requires that some people "run in packs.” We do,
of course, need lone -wolf research. Most important fundamental dis-
coveries come from the minds and imaginations of individuals. Seldom
can two individuals be put together and give birth to a joint idea.
Concepts are usually formed in each individual mind. Most discoveries
in fundamental science will continue to be born in the minds of individ-
uals. However, in the case of applied research relating to eventual
farm recommendation, the manager generally makes his choice of
practices or resource use patterns from more than one field of science.
Problems or managerial decisions of any magnitude always involve
phenomena from two or more fields. Applied research needs to be
planned accordingly.
A fourth requirement is scientific objectivity. One scientist must
not have a special interest in one kind of finding. * Results, whether neg-
ative or positive to the cause of one scientist, should be accepted ob-
jectively by both. If each wants to prove his predetermined notions, he
may not be willing to accept outcomes in terms of probabilities.
A fifth requirement is to establish a satisfactory division of labor.
In some cases the largest input of time and experimental funds is in
conducting the experiment. In other cases, the analysis of the data is
more time-consuming. Where the largest burden is in conducting the
experiment, funds and personnel should be provided accordingly. The
economist may need to help provide funds or personnel for use by the
project leader from the other department. Generally, a division of
labor between conducting the experiment and processing of the data,
with joint effort on interpretation and presentation of results, may be
satisfactory,
A final requirement for successful co-operative research is that
the objectives of the two research workers be sufficiently complemen-
tary to allow for an integrated approach. Otherwise a basis does not
exist for a co-operative project. Interests, however, do not need to be
identical. An animal scientist might have prior interest in digestibility
of feed materials fed in different proportions, while the economist has
prior interest in estimating marginal productivities of feeds. In a case
such as this, interests may be sufficiently complementary that the same
design and experimental layout can accomplish both ends. In other
cases, objectives of research may be partly complementary and partly
competitive and the research workers will need to determine which of
the competitive phases will be given priority.
Farm Functions
The discussion above has centered on physical production functions.

Similar considerations of co-operative effort are not required for farm
production functions since they generally are conducted by economists
alone. However, there are important cases where the economist
might
best seek out the aid and co-operation of the physical scientist even
for
these types of studies. In a physical sense, what input categories can
be aggregated? Over what range of soil types should the sample
extend
if a single production function, rather than a
hybrid of several different
production functions, is to be estimated? These physical questions
are,
in fact, some of the more important ones which need

to be answered for
efficient estimates offarm production function. In the future, perhaps,
as physical scientists assume more largely the role of estimating
physical production functions, the need for economists to request the
aid of agronomists, engineers, and animal scientists will be greater
than the opposite.
HISTORIC SUMMARY
The amount of resources devoted to predicting production functions
has increased greatly in the last decade. The concept is no longer
simply one for the classroom, with the quantities themselves being un-
known. Great progress has been made in estimating productivity coef-
ficients for both physical and farm units. And while progress in putting
these into practical use may have lagged somewhat behind the rate of
derivation, they are coming into increased use for farmer decisions
and economic policy projections. Still it should not be implied that
concepts of production function forms and derivation of coefficients are
only of recent origin. Some of the early articles and empirical studies
on production functions had lasting effects on subsequent thought and
research.
A summary review of historic developments in production function
estimates follows. No attempt has been made to review the details of
all studies and statements which, over time, have had direct or indirect
bearing on subsequent empirical analysis. Those reviewed are mainly
the ones which fall in a classical category, either in respect to the al-
gebraic forms which they propose or the empirical procedures implied.
Studies reviewed are mainly those in which the functional or algebraic
nature of production relationships has been estimated or the magnitude
of coefficients has been interpreted. This procedure, used because of
space limitations, excludes mention of a large number of studies in
which certain related quantities, ordinarily using different empirical
procedures, have been provided in agricultural experiments, farm man-
agement and industry studies. The period reviewed ends approximately
at the period of World War II.
The review is not limited to functions derived for agricultural pro-
duction. Review is made of the early industry studies by Douglas and
associates because these gave impetus to estimation of farm functions.
They also gave rise to questions and discussions of methodology which
are pertinent in estimation of agricultural production functions. Hence,
the more important methodological discussions also are summarized.
Physical Production Functions
We first review studies relating to physical production functions of

technical units in agriculture. This procedure is followed because
some of these studies predate those made for firms and industries.
Justus Von Liebig’s “law of the minimum” was the first attempt to
define the fundamental relationship between fertilizer or nutrient inputs
and crop yields. 1 Von Liebig stated that a soil containing all nutrients
necessary for plant growth except one is barren for all crops for which
the lacking nutrient is indispensable. He believed that crop yields were
proportional to the amount of nutrients supplied to or provided from the
soil, and that when all soil nutrients are present in sufficient supply,
addition of one or more would not increase yield. While Von Liebig did
not suggest algebraic equations representing his concept of the crop
production function, the algebraic form implied is somewhat obvious.
Each nutrient would serve as a limitational factor to the others. The
production surface would reduce to a simple “knife’s edge” with a con-
stant slope to the maximum per acre yield. In general, this type of
crop production function has been disproved. However, as data pre-
sented later in this test show, most crop production functions probably
have a “Von Liebig point” where all nutrients are limitational and addi-
tion of any one will not increase yield. This “Von Liebig point” is at
the maximum or peak of the production surface where the isoclines
converge and the isoquant defining this output reduces to a single point.
Baule suggests that Von Liebig’s law of the minimum supposed that
plants took up nutrients only in a given ratio and that yield would vary
directly with the quantity of nutrient available in smallest supply.2
Bondorff and Plessing interpreted his law as represented by the alge-
3
braic form
(1.1) y = ax
where y is yield response, x is the quantity of the nutrient, and a is

the constant or coefficient defining the transformation ratio. Boresch
proposed Von Liebig’s law to be represented as: 4
(1.2) y = c + ax
where y is total yield, c is yield level in absence of applications of x,

the nutrient added. Extension of these concepts to more than one nutri-
ent would provide, in the case of equation 1.1, the function
y = x3 ax^
for three nutrients and yield would be proportional to input of any one
fertilizer resource. Yield would be zero in absence of any one. How-
ever, this algebraic form would not directly define the yield maximum
discussed by Von Liebig.
1
Liebig, Justus von. Die grundsatze der agricultur - chemie mit rucksicht
auf die in
England angestellten untersuchungen. Friedrich Viewig und Sohn. Braunschweig.
2
1855.
Baule, B. Zu Mitscherlichs gesetz der physiologischen beziechungen. Landw.
Jahrb ’
51: 363-85.
Bondorff, K. A. Det kvantitative porhold me lie m
planternes ernaering og stofproduk-
tion. Den Kongelige Veterinaer - og Landboke j skale Aarskrift.
1924. Pp. 293-336; and
Plessing, H. C. Udbyttekurver med saerlight henblik pad matematisk formulering
of landru-
gets udbyttelov. Nordisk Jordbrugsfarskning. Vol. 25: 399-424.
4
Boresch, K. Uber ertragsgesetze bei pflanzen. Ergebnisse Bioiogie,
4: 130-204.
The first attempt to define the algebraic nature of the fertilizer -

5
crop production function was that of Mitscherlich in 1909. He, per-
haps, was the first agriculturist to suggest a nonlinear production func-
tion relating nutrient input and crop output. With the aid of Baule, a
mathematician, he proposed the equation
(1.3) log A - log (A - y) = cx
to explain fertilizer response allowing diminishing marginal produc-

tivity. A is total yield when the nutrient, x, is not de-
In this equation,
ficient (i.e., A is the maximum yield attainable from addition of x) and
c is a proportionality constant, defining the rate at which marginal
yields decline. Mitscherlich proposed the coefficient c was a constant
for all crops, unaffected by type of crop, climate, or other environ-
mental factors. A debate on these theories was conducted by Briggs,
6
Rippel, and Mitscherlich. One objection to the Mitscherlich equation
was that it did not allow negative marginal products, or diminishing
total yields. Accordingly, Mitscherlich made an adaptation to allow
7
this condition. This equation can be stated as
cx _kx2 c
(1.4) y = (1 - 10“ )(10 )(10 )
where k is a “damage factor” due to excessive magnitudes of x.

Working independently and without knowledge of this development,
Spillman proposed an exponential yield equation similar to that of
8
Mitscherlich. In main form, Spillman’s total yield equation is
(1.5) Y = M - AR X
where M is the maximum total yield attainable by increasing the nutri-
ent input x, A
a constant defining the maximum response (the sum of
is
marginal yields) attainable from use of x, and R is the coefficient de-
fining the ratio by which marginal productivity of x declines. He de-
veloped this equation after examining results from fertilizer experi-
ments on tobacco in North Carolina. In contrast to Mitscherlich, he
believed that the constants in the production function varied with en-
vironmental conditions. In the above equation, total attainable output
can be defined as M
= y 0 + y where y 0 is total yield when x = 0 and y is
the addition to yield corresponding to various nonzero values of x.
5
Mitscherlich, JR. Das gesetz des mmimums und das gesetz des abnehmenden bodener-
trages. Landw. Jahrb., 38: 537-52.
Briggs, G. E. Plant yield and the intensity of external factors — Mitscherlich^
8
“Wirkungs gesetz . Ann. Bot. 39: 475-502; and Rippel, A. Zwei experimentelle wider legungen
des Mitscherlich — Bauleschen wirkungsgesetzes der wachstumsfaktoren. Zeits Planzen
Bung, und Boden, Series A, 8: 65-76.
7
Mitscherlich, E. A. Second approximation of the law of action. Zeits. Pflanzen Bung,
und Boden, Series A, 12: 373-82.
8
Spillman, W. J. Application of the law of diminishing returns to some fertilizer and
feed data. Jour. Farm Econ., 5: 36-52.
Hence, we can define the output forthcoming in the absence of the vari-
able nutrient as y 0 = M
- A while the addition to yield or response is
(1.8) y = A(1 - Rx )
with the latter defining the magnitude of incremental yields associated

with various values of x. Using these components, total yield is Y= y 0
+ y = A+M- x A(l-R )=M
- AR X , the quantity indicated in equation
x
1.5. The term 1 - R can be defined as a “percentage sufficiency”
quantity. If x is zero, response also will be zero. As x increases in
magnitude, the fraction 1 - R x , or the percentage represented by it,
declines and the response approaches the mathematical limit, A.
Hence, the input-output curve is asymptotic to A when measurement is
of response to a variable factor and to when measurement is of re- M
sponse to variable plus output forthcoming from fixed factors.
Baule extended the Mitseherllch equation to include n variables. 9
His equation was based on the “percentage sufficiency” concept for in-
dividual nutrients. Based on Baule’s work, Spillman developed the
equation
(1.7) y = A(1 - R x )(l - Rz )
where y is output from variable inputs, A is the maximum increase

from use of variable resource, R is the ratio defining marginal prod-
ucts, with the unit of variables defined to make R = .8, and x and z are
quantities of two variable inputs. For this equation, the variable inputs
serve as limitational factors at zero magnitude of either, since 1 - R°
= 0. Again, the function allows for diminishing marginal productivity
of
any resource but does not allow for negative marginal products; the re-
sulting surface approaching a height asymptotic to A.
Pfeiffer and Frolich are reported to have used a single -variable
polynomial of quadratic form in 1912, in relating total crop yield to
supply of nitrogen. Panse used an equation of this type to fit response
data for cotton while Sukhatme used it for rice. 10 Briggs suggested the
use of the hyperbola function of the general form 11
1 . 8) - fr + b E >
( v
y
x + b+ h
where E is a maximum yield, b is the quantity of x initially in the soil,

and h optimal supply of the input. Somewhat similar hyperbolic
is the
type equations were suggested by Boresch, Balmukand, Baule and
B ' Zu Mitscherlich s gesetz der physiologischen bezichungen. Landw.

’
Jahrb.,
51; 363-85.
10
Panse, V. G., et al. Co-ordinated manurial trials on rainfed cotton in peninsular India.
Indian Jour. Agr. Sci., 21; 113-35, and Sukhatme, P. V. Economics of manuring. Indian
Jour. Agr. Sci., 11; 325-37,
11
Briggs, G. E. Plant yield and intensity of external factors - Mitscherlich’s “Wirkunes
gesetz.” Ann. Bot., 39; 475-502.
~
12
Bondorff. Thelau is quoted as suggesting an elliptical function to
13
characterize response to fertilizer. Other literature also could be
cited, but the above represent the main early concepts of the algebraic
nature of crop response to fertilizer.
E. M. Crowther and F. Yates analyzed all 1 year fertilizer experi-
ments conducted in Great Britain between 1900 and 1941 and similar
experiments conducted in Europe as a basis for formulation of wartime
14
fertilizer policy. They emphasized that final conclusions on fertilizer
response must be based on a series of experiments conducted in differ-
ent years, on different crops, and under varying farm and soil situa-
tions., The function used for certain of the estimates was y = y 0 + d(l
- 10
KX
), a modification of the Mitscherlich formula, where y 0 is yield
without fertilizer, d is the limit in response, x is amount of nutrient
applied, and k is a constant value for each nutrient. The average k
values determined for the British experiments were 1.1 for nitrogen,
.8 for phosphorus, .8 for potash, and .04 for manure. The authors seem
to suggest that these k values are constants for different locations and
conditions, as confirmed by analysis of additional experiments from
Denmark and Sweden.
Soil scientists more often than animal scientists have been con-
cerned with the functional nature of technical production functions.
Discussion of production functions by soil scientists predated interests
of agricultural economists. However, efforts of animal scientists in
estimating production functions have been largely co-operative efforts
with economists. A classical livestock production function study based
15
on experimental feeding was published by Jensen et al. in 1942. This
study dealt with input -output relationships in milk production. The al-
gebraic form used was that developed by Spillman for a single input
category. The data came from a large-scale co-operative experiment
between several land-grant colleges and the United States Department
of Agriculture. The results of the study were used to indicate how
feeding rates to maximize profits per cow should be varied from
Haecker standards as prices of feed and milk vary. Marginal feed pro-
ductivity was estimated for cows of both high and low producing capac-
ity at stations with herds of different output levels. This study, made
jointly by dairy scientists and agricultural economists, was important
in helping to change technical views on the nature of the feed-milk pro-
duction function. The 367 cows included in the experiments at nine sta-
tions received a variety of feeds in the form of grains, silage, hay, and
pasture. These feeds were all converted to a total digestible nutrient
12
Boresch, K. Liber ertagsgesetz bei pflanzen. Erge. Biol., 4: 130-204; Balkumand, B. H.
Studies in crop variation. Indian Jour. Agr. Sci., 18: 602-27; Boule, F. Liber die Weiter nick-
lung. Zeit. Ahen Pflanzen, 96: 173-86; and Bondorff, F. (as reported in Plessing, Udbytte-
kurver med saerlight henblik. Nordisk Jord., 25: 399-424).
13
As quoted by Plessing, ibid.
14
Crowther, E. M. and Yates, F. Fertilizer policy in wartime. The fertilizer require-
ments of arable crops. Empire Journal of Experimental Agriculture, 9: 77-98.
15
Jensen, E., et. al. Input-output relationships in milk production. Tech. Bui. 815.
USDA, Washington, D. C. 1941.
basis, a procedure completely satisfactory only if all feeds substitute
at a constant marginal rate at the transformation ratio used, with the
single feed input category measured in pounds of TDN. The authors did
not provide the numerical magnitudes of the coefficients in the derived
production functions but some of these can be estimated from the tabu-
lar and graphic materials presented. Since only grain was varied and
cows generally were allowed to consume all the forage they desired,
the TDN -milk transformation relationship is probably not a “true” pro-
duction function in the theoretical sense. Instead, it more nearly ex- .
presses the relationship between TDN input and milk output along a
stomach limit line such as explained by Heady. 16
Another production function study in feeding levels also was origi-
nally inspired by considerations of farm profit maximization. Atkinson
and Klein made a study of optimum marketing weight of hogs relative to
feed transformation ratios and price relationships. The input -output or
production function relationship derived relates weight per hog to feed
intake measured as pounds of concentrates. The basic data are from
swine feeding trials although the original experiments were not de-
signed specifically for formal production function analysis. Numerous
experiments from five land-grant colleges were aggregated for deriva-
tion of an average input -output curve involving a single feed variable.
The several grains and protein supplements in the experiments were
converted into a concentrate variable on the basis of weight. The
single -variable algebraic function used also was of the type developed
by Spillman. While the coefficients for the equation expressing hog
gain relative to pounds of concentrate intake is not given, an algebraic
function relating weight to age of hogs is published. It is
A
(1.9) W = 588(.0097)
,9923
where W
is hog weight in pounds and A is age in days from birth. The
estimates derived in this study also were useful for wartime guides and
policies relative to optimum marketing weights for hogs, and, hence,
for structuring of administered price differentials for hogs of vary ing;
weights.
Another livestock study inspired by wartime food shortages and
need for basic data in determining livestock feeding and price policy
was conducted t?y Nelson. 17 This study between animal gain and feed
intake for calves, yearlings, and 2 year old beef animals also was
based on feeding experiments. Again, the experiments were conducted
in a number of years prior to the analysis and were not designed
spe-
cifically for production function analysis. A single -variable
production
function was derived for the several classes of cattle with the various
grains, protein supplements, and forages converted to a single input
18
Heady, Earl O. Economics of agricultural production and resource use.
Prentice-Hall’
Inc. New York. 1952. Pp. 156-57.
17
Nelson, A. Relationship of feed consumed to food products produced by
fattening cattle.
Tech. Bui. 900. USDA, Washington, D. C. 1945.
category measured as total digestible nutrients. Estimates were de-

rived for both live and dressed weight of cattle. The live weight re-
gressions indicate diminishing marginal and average productivity of
feed from the outset of the feeding period. The dressed weight regres-
sions indicate, because dressing percentage of animals increases with
weight, ranges of both increasing and decreasing marginal productivity
of feed. Like the previous studies cited for hogs and dairy cattle, pro-
ductivity coefficients cannot be specified for different feed categories,
nor can marginal rates of substitution be specified for different feeds.
The single -variable functions derived, adaptations of the Spillman
function, were
OO2Bf
(1.10) Calves: w = 1.446 - l,049e“-°
OO28f
(1.11) Yearlings: w = 1,446 - 805e"'°
0037f
(1.12) 2 year: w = 1,446 - 610e”'°
where w is live weight in pounds after start of the feeding period, e is

the base of the natural system of logarithms, and f is feed intake per
animal measured as pounds of total digestible nutrients. These live
weight functions were transformed, on the basis of published data on
dressing percentages, into quantities representing relationships be-
tween dressed weight and feed intake per animal.
Industry Functions
Numerous early general economists, such as Smith, Ricardo, and

Malthus, suggested hypotheses about the general nature of the produc-
tion function for the agricultural industry. Knut Wicksell was perhaps
the first to suggest an algebraic form or physical nature for the agri-
18
cultural production function. He refined the notions of classical
economists by stating that increasing returns to labor and capital are
possible when fertilizer inputs also are applied to nutrient -short soils.
He stated that the agricultural research stations should provide the
necessary data for establishing the functions, but that they had not done
so at that time. Wicksell indicated that agricultural output was clearly
a mathematical function of the quantities of labor, land, and capital
used and if these are expressed respectively as a, b, and c, with pro-
duction per annum indicated as p, the production function can be ex-
pressed as
(1.13) P = f(a, b, c) .
18
Wicksell, Knut. Den “kritiska punkten” i lagen fOr jordbrukets aftagande produktivitet.
Ekonomisk Tidskrift. 1916. Pp. 285-292.
However, he stated that this function is homogeneous of first degree,
denoting constant returns to scale. Hence, if input of all factors is
doubled, it becomes f (2a, 2b, 2c) = 2p, and output also is doubled. He
expresses the belief that Rohtlieb, 19 in a previous paper, was also dis-
cussing the possibility of constant scale returns for factors increased
in fixed proportions. However, Wicksell states, in argument with
Rohtlieb, that if a simultaneous increase in the three factors leads to a
proportion increase in output, then increase in one such as fertilizer
cannot lead to the same kind of increase and, therefore, that classical
economists were correct in their notions of diminishing agricultural
productivity because they considered land to be fixed in quantity. In his
terms, technical change simply gives rise to a new production function
</>( a, b, c), which has constant scale returns for all factors increased
by fixed proportion, but diminishing productivity for one factor in-

creased relative to fixed quantities of the others. Here he is discussing
the production function of a farm, but states that the production func-
tion, F(A, B, C) = P, for a nation need not be homogeneous of first de-
gree. Possibilities of labor specialization, for example, in a colony as
its population grows and more land is taken up, may allow increasing
returns. Wicksell also implies that Rohtlieb’s empirical analysis in-
volves comparisons of different production functions, rather than of ef-
fects of different input levels for a given production function.
Actually, the equation which has come to be known as the Cobb-
Douglas function traces to Wicksell. He stated, in a footnote, the func-
tion as
a
(1*14) p = a b^c r
and, as in the early attempts of Douglas, said that a and y should

, 0 ,
sum to 1.0.
While the first general hypothesis about the algebraic nature of an
industry production function was proposed for agriculture, the first
empirical attempt was for nonfarm industries. Cobb and Douglas ap-
plied a function similar to Wicksell’ s to data for American manufac-
turing industries over the period 1899 -1922. 20 Evidently, this was the
first formal empirical production function fitted to time series
data.
The function fitted was of the form
k 1-* k
(1.15) P' = bL C
where P' was the predicted index of manufacturing output over the pe-
riod, L was the index of employment in manufacturing industries, and C
was the index of fixed capital in industry. The function derived from
the time series data was
19
Rohtlieb, C. E. Om grans en ftJr jordbrukets intensifiering. Ekonomisk Tidskrift.
1916:189.
Cobb, Charles W. and Douglas, Paul H. A theory of production.
American Economic
Review. Vol. 18, (March, supplement): 139-65. 1928.
( 1 . 16 ) P' = i.oil-ts c- 25 .
Cobb and Douglas computed the marginal productivities of capital

and labor which are respectively .25PC -1 and .75PL"1 They used .
these quantities to impute the shares of the actual total product to labor
and capital during the period studied. On the basis of these quantities,
they imputed the total product to capital and labor in the proportions
.25P and .75P, respectively, where P is the actual index of production
in any one year (as compared to P' which is the index of production es-
timated from the derived function). The annual quantities so imputed
from the empirical production function then were compared with esti-
mated shares of the total manufacturing product actually distributed to
labor and capital over the period. Cobb and Douglas also explored the
possibilities that k = .67 and 1 - k = .33 and decided that the derived
values k = .75 and 1 - k = .25 described the actual process of produc-
tion in manufacturing in a fairly accurate manner. The value of R, the
multiple correlation coefficient, for the function with these elasticities
was .97.
They selected this function with its restraint that the sums of elas -
ticitiesor regression coefficients should total 1.0 because they were
interested in imputing the total product back to the two factor catego-
ries. With a sum of elasticities either greater or smaller than 1.0, the
total product would have been respectively less or greater than the
total amount imputed to the resources. Cobb and Douglas did not claim
to have solved the laws of production or to have employed the most ap-
propriate data. They stated that the form of function used was only one
alternative and that formulas should be devised which would not require
“constant relative contributions of each factor to total product but
would allow for variations from year to year.” Douglas indicated that
one of his co-workers suggested a modification of the function as early
21
as 1926. He gave credit to Wilcox for suggesting the function
“h
(1.17) P' = bR 1_k L k Ch where R = (L
2
+ C 2 )-5
and the sum of k and h is not forced to unity. In applying this to the
data used for equation 1.16, Douglas obtained the values b = 1.06, 1 - k
- h = -.146, k = .788, and h = .358. However, Douglas and his associ-
ates did not use this specific function for later time series analyses.
Douglas and his associates used time series data to estimate sev-
eral industry production functions of the form of equation 1.15 where
the sum of the exponents are forced to unity. In all these cases, they
were interested in the condition of equation 1.18 where the marginal
products of labor and capital,
5P
and
5P —
respectively,
^
L C
(1.18) ? =
f •
f *
21
Douglas, Paul H. The theory of wages. Macmillan Co., New York. 1934, P. 152.
multiplied by the quantities of labor and capital, L and C, would allow
a siim equal to the total product, P. The following values were obtained
22
for other sets of time series data.
Country and Period b k 1-k
Massachusetts (1890-1926) 1.007 .74 .26

Victoria (1902-1929) - .71 .29
New South Wales (1901-1927) 1.018 .65 .35
New Zealand (1915-35) - .52 .48
In an article published in 1936, Edelberg concerned himself with a

somewhat similar type of function in relation to time series data for
23
industry. The function he discussed was
r
(1.19) P = cs“x^L
where L and X were respectively the rates at which land and labor are
applied in production, s was the time or period of production and P was
the rate of production. Capital letters were used to denote quantities
for the economy, although he indicated that firm conditions might be
represented by the same algebraic form and the same variables de-
noted by p, s, and 1. He obtained the following empirical functions
over the period 1850-1910 for Great Britain and the United States
where, like Cobb and Douglas, he assumed the function to have constant
returns to scale in respect
(1.20) U.K. P = cs*

33 wX
L'
32
(1.21) U.S. P = cs' L'^X'
toL and X but not in respect to s. He indicated that the original Cobb-
Douglas function might be criticized for not including the time element
and "circulating” capital. Edelberg was interested mainly in deriving
labor demand functions.
Douglas and associates then relaxed the restraint that the sum of
the elasticities in the production function should sum to unity and em-
ployed, at the suggestion of Durand, 24 the function
(1.22) P = bL k CJ
22
Gunn, Grace T. and Douglas, Paul H. The production function for American manu-
facturing in 1919. American Economic Review, 31 (March): 67-80. 1941; Gunn, Grace T.
and Douglas, Paul H. The production function for Australian manufacturing. Quarterly
Journal of Economics, 56: 108-29. 1941; and Douglas, Paul H. The theory of wages. Mac-
millan Co., New York. 1934. Ch. 6-7.
23
Edelberg, Victor. An econometric model of production and distribution. Econometrica,
4: 210-25. 1936.
24
Durand, David. Some thoughts on marginal productivity with special reference to
Professor Douglas’ analysis. Journal of Political Economy, 45: 740-58. 1937.
where k and j could take any value. The resulting power function,
which is linear in logs, has commonly come to be known as the Cobb -
Douglas function. It has been applied in numerous production function
studies for technical units and firms of agriculture. Douglas and asso-
ciates applied it mainly to cross-sectional data where observations on
indices of product, labor input, and capital investment were obtained
from individual manufacturing industries. Statistics for some of these
cross-sectional studies are as follows: 25
Country, region, Number of industries

and year
b k J
for observations
United States (1904) 336 2.08 .65 .31

United States (1909) 90 .90 .74 .32
United States (1914) 340 1.97 .61 .36
United States (1919) 556 2.39 .76 .25
Victoria (1910-11) 34 — « .74 .25
Victoria (1923-24) 38 _ _ .62 .30
Victoria (1927-28) 35 .59 .27
New South Wales (1933-34) 125 - - .65 .34
Australia (1934-35) 138 — .64 .36
Bronfenbrenner and Douglas also computed separate cross-sectional

production functions for several individual industries based on 1909
26
data. These included:
Industry Number of observations b k j
Clothing -textiles 16 2.043 .978 -.070

Foods and beverages 12 .365 .715 .350
Metals and machinery 22 1.007 .714 .256
Verhulst estimated production functions from cross-sectional data

27
for twenty -five French gas firms in the fourth quarter of 1945. His
purpose was to determine the nature of the production function of a gas
firm in “the neighborhood of the point of equilbrium.” His input varia-
bles included the net prime costs, x, of each firm and the index of capi-
tal charges of the same firms. His output, p, was represented by the
quantity of gas sold. In estimating the production function in this
25
See Gunn, Grace T. and Douglas, Paul H. The production function for American manu-
facturing in 19X9. American Economic Review, 31 (March): 67-80. 1941; Gunn, Grace T.
and Douglas, Paul H. The production function for Australian manufacturing. Quarterly
Journal of Economics, 56: 108-29. 1941; Douglas, Paul H. The theory of wages. Macmillan
Co., New York. 1934. Ch. 5-9; Daly, Patricia, Olson, Ernest, and Douglas, Paul H. The
production function for manufacturing in the United States, 1904. Journal of Political
Economy, 51; 61-65. 1943.
26
Bronfenbrenner, M. and Douglas, Paul H. Cross-section studies in the Cobb-Douglas
function. Journal of Political Economy, 47: 761-85. 1939.
27
Verhulst, M. J. J. The pure theory of production applied to the French gas industry.
Econometrica, 16: 295-308. 1948.
“neighborhood,” the method of simultaneous equations outlined by
Marschak and Andrews where the production function was the modified
Cobb -Douglas type: 28
(1.23) p = ax ^"
0
His interest centered on methodological problems and no particular at-

tempt was made to interpret the quantitative elasticities derived from
the sample.
Criticisms of Industry Studies
Numerous criticisms of the functions applied by Douglas and his

associates appeared in the literature after parameters were estimated
from industry data on both a time series and cross-sectional basis.
Some of these referred to inaccuracies in the available data but others
were in terms of comparison of these results with those theorized for a
firm. It has been pointed out that the resulting empirical functions,
based on time series data wherein observations are drawn from indi-
vidual years as aggregates for all industry or on cross-sectional data
with observations representing aggregates of different industries, apply
only to groups of industries and not to individual firms.
Durand initially criticized the form of the Cobb-Douglas function
which required a sum of individual elasticities equal to unity. 29 He ex-
pressed doubt that the enterprise production function is homogeneous of
first degree and indicated that the validity of this condition was yet to be
established. If this condition actually held true, equilibrium output and
prices could not be attained under pure competition. As indicated ear-
lier, Douglas and associates then changed the function so that k +
j / 1.0,
to allow increasing, constant, or decreasing returns to scale. Durand
also pointed out that the function (1) is not of the type ordinarily dis-
cussed in economic theory since it applies to groups of industries
rather than to individual firms, and (2) is expressed in terms of undif-
ferentiated labor and capital instead of specific resources actually em-
ployed. Durand also pointed out that methods of measuring deprecia-
tion and the fact that management inputs were not quantified would
likely result in k and k - 1 values which did not bear the correct pro-
portion to each other.
Bronfenbrenner and Douglas summarized other criticisms of the
Cobb-Douglas functions as follows: Labor and capital are not on the
same footing because the first is measured by quantity actually used in
production while the second is measured by quantity available for pro-
duction. The function cannot be statistically fitted in three dimensions,
P, L, and C, because (1) the data are not sufficiently accurate,
(2) the
28
Marschak, J. and Andrews, W. H. Random simultaneous equations
and the theory of
production. Econometrica, 12: 143-205. 1944.
29
Durand, op. cit.
functions were fitted to index numbers rather than to actual quantities

of input and output, (3) the function as actually derived was only in the
two directions, P/C and L/C, and (4), the minimization of deviations
might, given the type of industry data, be in the direction of the L and C
axes as appropriately as in the direction of the P axis. The relation-
ships actually existing between P, L, and C in the time series data are
simply functions of time as it relates to technological change, popula-
tion growth, and economic development. Thus the original Cobb-
Douglas function might not be a production function at all, but a spuri-
30
ous correlation of these three series with time.
Bronfenbrenner and Douglas examined these criticisms, elaborating
on some and attempting to disprove others. They suggested that the
term “production function” need not apply to a particular aggregate of
activity such as a single plant, but might be applied to an entire in-
dustry or other strata and aggregates of production. The criticism of
capital measured in available quantity was dismissed as unimportant in
years of prosperity and full employment. The limitation of observa-
tions measured in index numbers was eliminated in the cross-sectional
studies since actual values of P, L, and C were used. The limitation of
a function homogeneous of first degree was eliminated through modifi-
cation of the original equation. The problem of multicollinearity or
spurious correlations revolving around time were eliminated in use of
cross-sectional data.
Menderhausen discussed Douglas’ production function results
31
largely from a statistical standpoint. He found a high degree of multi-
collinearity between the variables in the time series data. He pre-
sented a three-dimensional figure to show that the product quantities,
P, lie practically in a straight line over the input plane rather than
forming a surface dispersed over the resource plane. Menderhausen
illustrated that Douglas’ three variates (log P, log L, and log C) can be
represented as linear functions of time (t) as
(1.24) log P = at (a = .0156)
(1.25) log L = /3t 03 = .0112)
(1.26) log C = yt (y = .0281)
and the differences in the trend-slope lines for P, L, and C are
(1.27) logP - log C = a - y
(1.28) log L - log C = 0 - y
30
Bronfenbrenner, M. and Douglas, Paul H. Cross-section studies in the Cobb-Douglas
function. Journal of Political Economy, 47: 761-85. 1939.
31
Menderhausen, Horst. On the significance of Professor Douglas’ production function.
Econometrica, 6: 143-53. 1938.
with the ratio being equal to k of the Douglas time series func-
tion where k is simply a ratio of the two trend -slope differences. He

showed that this ratio did, in fact, turn out to be .741, a quantity simi-
lar to the k value of .75 found by Douglas. Hence, he concludes that
the k value in the Cobb -Douglas original function is simply an expres-
sion for the trend in technical development. He also mentioned that the
measure of capital, C, over the time period analyzed represented the
fixed capital present rather than that actually used, particularly in
years of underemployment in industry. Tinbergen pointed out that
Douglas* results would be perfectly acceptable if it could be assumed
on a priori economic grounds that k = 1 - j where k and j are respec-
32
tively the derived elasticities for labor and capital. He states that the
problem is then reduced to one of simple correlation. Both Mender -
hausen and Tinbergen emphasized the time series estimates of Douglas
gave rise to measurement difficulties. Menderhausen also emphasized
that it was as relevant to minimize the deviations from the direction of
the L and C axes, as well as from the P direction, and that the results
were so different that no real credence could be given to the latter.
Several other economists have emphasized the difference between
inter -firm, intra-firm, and inter -industry aspects of production func-
tions such as those derived by Douglas. These studies were based on
aggregate observations for groups of industries, rather than upon ob-
servations for individual firms as has been the case in more recent
studies made for agriculture. Reder indicated that these industry func-
tions differ conceptually in three ways from the production functions of
economic theory: (1) In theory, the physical production function shows
the functional relationship between the input quantities and the output of
a firm. The functions derived by Douglas are the loci of input -output
quantities of all firms used in the particular study. (2) The theoretical
function is in terms of physical quantities while the functions fitted to
industry data represent the value added in manufacturing to the physical
inputs of labor and the value of plant and equipment owned. (3) The
marginal value productivity of a factor is the partial derivative of
total product in respect to the particular factor multiplied times the
marginal revenue for this output. 33 The marginal value productivities
derived from the Douglas functions can be derived directly, without
multiplying by a corresponding marginal revenue times the marginal
physical productivity, because of the method of measurement. Accord-
ing to Reder, the latter should be called inter -firm marginal produc-
tivities whereas the theoretical concepts refer to intra-firm productiv-
ities. Bronfenbrenner, in discussing the inter -firm aspects of these
functions, attempts to show how results from inter -industry observa-
tions should be the same as those derived from inter-firm and
32
Tinbergen, J. Econometrics. Blakeston Co., New York. 1951. P.123.
33
Reder, M. W. An alternative interpretation of the Cobb-Douglas production function.
Econometrica, 11: 259-64. 1943.
34
intra-firm observations in competitive industries. He illustrated his
propositions for competitive industries by drawing input -output curves
representing individual firm production functions and associating these
with the type of relationship derived in the classical Cobb -Douglas
cross-sectional production function. He points out that, under compe-
tition, the slope of the production function -under equilibrium should be
the same between industries and firms. Smith pointed out some of the
conditions necessary for fitting meaningful statistical production func-
35
tions. Some of his observations are applicable to functions derived
from firm samples, as well as to industry observations or measure-
ments. The optimum set of data would involve experiments with indi-
vidual firms where labor and capital are combined in numerous propor-
tions. With inability to conduct experiments of type, he indicated that
the conditions of perfect competition and static equilibrium would serve
as a substitute. However, Smith doubted that this assumption is rele-
vant for industry observations of the type employed by Douglas and as-
sociates. If perfect competition did prevail, the production function of
the firm would be of the same degree whether input-output relationships
are expressed in physical or value terms and would differ only by a
constant product representing price relationships. Under perfect com-
petition and perfect factor markets, Smith stated, a long-run statistical
function from cross-sectional data should allow estimates of the con-
ventional intra-firm marginal value productivities. Under these con-
ditions, the average and marginal value productivities should equal the
factor prices paid by each firm. But under the conditions of imperfect
markets expected in industry studies, Smith concluded that the degree
of the value production function would be less than the degree of the
physical production function. Smith points out several other ways in
which the function derived from industry observations may differ from
the theoretical physical production function. (1) Particular problems
arise in measurement of capital input. Theoretically, the relevant input
is annual capital use but capital investment ordinarily is used. If an-
nual capital input always bore a constant ratio to investment in capital
goods, the elasticities should be the same. (2) Firms in a cross-
section study may employ different techniques, particularly due to fixed
plants inherited from the past, and the long-run production functions so
derived may represent “mongrels” or “hybrids.” Some firm adjust-
ments to price and other conditions are of long-run nature while some
are of short-run nature and the differences among observations of this
type cannot be separated conveniently. If short-run or long-run pro-
duction functions were desired, the degree of adaptation in labor and
capital would need to be limited in terms of concepts of the relevant pe-
riod of time and change in plant structure. (3) The observations result
from management decisions rather than experimentation or physical
manipulation of resource mixes, to measure their effect on output.
34
Bronfenbrenner, M. Production functions: Cobb-Douglas mterfirm, intrafirm.
Econometrica, 12: 35-40. 1944.
35
Smith, Victor E. The statistical production function. Econometrica, 59: 543-62. 1945.
-
Accordingly, observations are available for only a small segment of the
production surface and errors may arise due to changes in the function
itself, to delayed or incomplete adjustments to price changes, and to
errors on the part of management.
May stated conditions necessary for estimating a firm production
function of technical or physical nature. The function must relate in-
puts and outputs in an individual productive unit. The variables of input
and output must be scalars of definite units associated with observable
and measurable quantities. There must be a scalar representing the
input quantity of each factor, aggregates being excluded except in the
trivial case where components of the aggregate are single valued func-
tions of the aggregate. For fixed capital inputs there should be two
variables, with one representing the stock of capital and the other its
extent of use. The variables should be in physical rather than value
units. Factors other than inputs should be excluded as variables and
held fixed and include such considerations as time distribution of inputs
over the calendar period, morale factors, intensity of effort, techniques,
geographical environment, and others. He suggests that production
functions might well be estimated for subprocesses of the over -all pro-
36
duction process.
Marschak and Andrews in 1944 discussed further the estimational
procedures involved in deriving production functions such as those esti-
mated by Douglas. 37 In this connection they proposed the possibility of
using systems of equations in attempting to predict production func-
tions, revenue functions, and factor demand functions. They outlined
possible procedures in estimating an entire set of relevant equations,
including the production function, for individual firms when data are
drawn from an inter -firm sample where differences exist in production
functions.
Contribution of Cobb -Douglas Functions
The discussion immediately above refers mainly to the types of in-

dustry functions fitted to time series or cross-sectional data by Douglas
and his associates. In general, the data used in these types of studies
have not been similar to those used for agricultural production func-
tions. The latter have been based almost entirely on physical experi-
ments and samples of firms. The literature surrounding the industry
studies lists certain limitations which do not apply to agricultural pro-
duction functions estimated from experimental and farm sample obser-
vations. However, some of the more technical points relating to inter
firm and intra-firm difficulties may have application in farm production
functions. The literature relating to the Cobb-Douglas industry
36
May, Kenneth, Structure of production function of the firm. Econometrica, 17: 186-87.
1949.
57
Marschak, Jacob and Andrews, William H. Random simultaneous equations and the
theory of production. Econometrica, 12: 143-205. 1944.
functions has been reviewed both to indicate (1) developments leading

to estimates of value functionsfrom economic observations, and (2) a
few limitations which may also have application to farm value functions.
One of the main contributions of Cobb and Douglas, and also of
Durand, to production function studies may have been development of
the particular algebraic form of equation. It has been used widely be-
cause of its convenience in interpreting elasticities of production, be-
cause estimation of parameters involve fewer degrees of freedom than
other algebraic forms which allow increasing or decreasing returns to
scale, and because its use involves simple computations. It has not
been used in agriculture in the original form requiring a sum of indi-
vidual elasticities equal to 1.0 and it has seldom had use in estimating
the functional distribution of the product among factors. The term
Cobb -Douglas function might best be reserved for reference to the
particular algebraic form. Unfortunately, however, this term is some-
times used synonymously with any production function estimated for
firms or industries. Hence, an impression sometimes prevails that all
of the limitations outlined above apply to value functions estimated from
nonexperimental data. Obviously, all of the problems relating to multi-
collinearity in time series data, aggregation across industries, imper-
fect competition in the market, and others do not apply to functions de-
rived from a sample of competitive firms such as those found in
agriculture. Similarly, it is not necessary that a function fitted to a
firm sample be of the algebraic form used by Douglas and his associ-
ates. Hence, in the chapters which follow, the term Cobb-Douglas
function will refer to a particular algebraic form of equation, regard-
less of whether its parameters are estimated from firm industry or
experimental observations. It will not be used, as is sometimes the
convention, to indicate all production functions derived from nonexperi-
mental observations.
Early Functions for Agricultural Enterprises and Firms
We now turn to a summary of early studies dealing with formal esti-

mates of production functions from farm data. In contrast to studies
reviewed previously, they are based on neither experimental data nor
industry samples but on cross-sectional observations of enterprises or
farms.
One of the first major attempts by agricultural scientists to fit pro-
duction functions to farm data, in contrast to physical data, was pub-
38
lished in 1924 by Tolley, Black, and Ezekiel. The functions analyzed
were largely single -variable physical relationships for enterprises.
But unlike other physical production functions based on experimental
38
m
Tolley, H. R., Black, J. D., and Ezekiel, M. J. B. Input as related to output farm
organization and cost of production studies. Tech. Bui. 1277. USDA, Washington, D. C.
1924.
Table 1.1. Daily Gains in Pounds per Head as a Function of Corn and
Hay Input From Tolley, Black, and Ezekiel Study
Hay, Pounds per Day

Corn, Pounds per Day
8 12 16 20
10 1.61 1.81 1.98 2.13
15 1.96 2.16 2.33 2.48
20 2.27 2.47 2.64 2.79
25 2.41 2.61 2.78
data, these were based on farm data, and although regression equations
were not always computed, inputs analyzed included such categories as
labor, fertilizer, and feed. Inputs were measured both in physical and
monetary terms. A few linear regressions were derived relating total
gain to total feed intake per hog. Data also were used to derive a sim-
ple hog production function which showed diminishing marginal produc-
tivity. They presented a “tabular production surface,” showing daily
gain for steers as a function of corn and hay input. The results are
shown in Table 1.1. These estimates of intra -animal or intra-
enterprise surfaces were derived from inter -farm or cross-sectional
data. While they did not actually derive an algebraic production sur-
face, they did present several two -dimension cost surfaces.
While such concepts as isoclines, isoquants, marginal rates of sub-
stitution, and production surface had not yet come into the literature,
these early agricultural economists did come close to producing them
in their analysis. They, along with Spillman, also used regression
techniques. Hence, an interesting question arises: What caused the
time relapse in estimation of farm and other agricultural production
functions?
Firm functions
Perhaps the earliest production function study representative of es-

timates for farm firms was published by K. Kamiya of Tokyo University
39
in 1941. He analyzed data from a 1939 survey of paddy farms in the
Tokoku and Seinan districts of Japan. The respective Cobb-Douglas
functions fitted to these data are
73 07
(1.29) P = 924T* L"*
- *53
(1.30) P = 1,1011* •* L
where P is receipts from rice production measured in yen (after deduc-

tion of expenses for fertilizers, seed and other materials), T is land
area measured in cho, and L is labor in days. Data of both districts
39
Kamiya, K. Productivity of labor. Journal of Rural Economics, 21(3); 22-44.
Table 1.2. Elasticities (Regression Coefficients) for 1939 Random Sample

of Iowa Farms
|
Miscellaneous
Equip- Livestock Operating Sum of
Farm Group Land Labor ment and Feed Expense Elasticities
Northeast dairy .2081* -.0449 .0871* .6136* .0223* .8862
Cash grain .3425* .1131 .0757* .4115* .0174* .9502
Western meat .1037* .0107 .0888 .5511* .0310* .7853
Southern pasture .1624* -.0508 .1325 .4944* .0438* .7823
Eastern meat .1918* .2364 .0952 .4263* .0245* .9742
Crop .2412* -.0748 .0761* .5294 .0166* .9381
Hog .0691* .0209 .0965 .7449* .0343* .9657
Dual purposes 4 *
f
and dairy .1029 .0074 .0636 .6294' .0178 .8211
General .1737* .1192 .1586* .4555* .0302* .9372
Special .2970* -.0070 .1097* .3512* .0379* .7888
“Large” .2814* .0001 .1126* .5378* .0286* .9605
“Small” .2131* .0482 .0843* .4271* .0324* .8051
All farms .2316* .0282 .0844* .4767* .0317* .8526
* Significant at the .05 level of probability,

tsignificant at the .01 level of probability.
resulted in coefficients for labor which denote negative marginal pro-

ductivities. Kamiya explained that the negative marginal productivities
for labor might result from the fixity of family labor on Japanese farms
and the size of farms. Under the family labor and land system, he pro-
posed that the farmers actually purchase a large proportion of their
own product through their labor. However, he also estimated that mar-
ginal productivity of labor would exist for farms with more than four cho
of land in Tokoku district and more than three cho in Seinan district.
Evidently the first empirical estimates of production functions for
agricultural firms in the United States were included in Iowa studies.
Heady derived production functions for a 1939 random sample of 738
40
Iowa farms. Functions were derived both for types of farms and
areas of the state. In all cases the inputs were land, labor, power and
equipment, livestock and feed, and operating expenses, all measured in
dollars. Output was measured in dollar value of product. A Cobb-
Douglas function without the restraint of constant returns to scale was
employed. The computed elasticities are shown in Table 1.2.
Some of the elasticities were negative. While it is hardly conceiva-
ble that total production would decrease were more of any individual
40
Heady, Earl O. Production functions from a random sample of farms. Journal of Farm
Economics, 28: 989-1004. 1946.
factor employed, none of the negative elasticities shown in Table 1.2
were statistically significant at the 5 per cent level of probability.
Hence, these elasticities could have arisen with a probability of one in
twenty even if the true population elasticities were zero. In every case
the sum of the elasticities was less than one, denoting diminishing re-
turns to scale.
Management was not included as an input since there was no objec-
tive measure available in the data. The results might well have dif-
fered had it been possible to measure the input of this factor. It is the
common judgment that better managers are found on’ larger farms. Ac-
cordingly, the true elasticity outcome would differ depending on whether
the actual input of management for the farms studied increased at an
increasing or at a decreasing rate, as the input of other resources or
output of product increased by a given percentage.
Marginal productivities, computed at the geometric mean, and their
fiducial limits are presented in Table 1.3 for geographic groups of
farms.
Limitations discussed in the study included aggregation of inputs,
lack of homogeneity in farms sampled for estimating intra-firm func-
tions, measurement of labor available rather than labor used, measure-
ment of capital inputs, specification of management input, and form of
function.
Tintner and Brownlee derived a Cobb -Douglas function for 468 Iowa
farms which kept records in 1939. 41 They derived the following mean
marginal productivities where inputs and product were measured in
dollars: land, .10; labor, .09; buildings, .05; liquid assets, .18; working
assets, .16; and cash operating expenses, .28. Tintner derived a simi-
lar production function for 609 Iowa farms which kept records in 1942. 42
The magnitude of elasticities is indicated in the following empirical re-
sult where product is measured in value terms and inputs are, respec-
tively, land, labor, farm improvements, liquid assets, working assets,
and cash operating expense. All inputs but land acres and labor months
were measured in dollars.
29 16 05 21 01 16
(1.31) P = aA* B* C* D* E~' F*
In all three of these exploratory studies, crop and livestock prod-

ucts were aggregated into a single output. Inputs were aggregated
largely on the basis of accounting procedures of the time. They were
computed from existing data available and the authors recognized that
the samples were heterogeneous and likely gave rise to mongrel func-
tions. Also, it was recognized that inputs were not always measured in
a logical fashion and thus may have given rise to law or negative pro-
ductivities. However, it also was recognized that these same
41
Tintner, G. and Brownlee, O. H. Production functions derived from farm records.
Journal of Farm Economics, 26: 566-74. 1944.
42
Tintner, G. A note on the derivation of production functions from farm record data.
Econometrica, 16: 295-304. 1944.
Table 1.3. Marginal Productivities and Fiducial Limits at the 5 Per Cent Level
of Probability (per Dollar of Input)
Miscellaneous
Livestock Operating
Land Labor Equipment and Feed Expense
A B C D E
All farms
mean .0465 .0791 .2013 .8390 .3931
upper limit .0578 .2329 .2855 1.0772 .4320
lower limit .0351 -.0742 .1171 .6008 .3543
Northeast dairy area
mean .0331 .0973 .1484 .6588 .3783
upper limit .0518 .3359 .2850 1.0723 .4493
lower limit .0144 -.1413 .0018 .2452 .3074
Cash gram area
mean .0615 .1066 .1809 .4177 .3693
upper limit .0862 .1460 .3570 .6798 .4662
lower limit .0368 .0672 .0048 .1556 .2725
Western meat area
mean .0382 .0302 .2410 .7130 .4037
upper limit .0791 .3207 .3930 1.2075 .4738
lower limit -.0027 .2603 .0890 .2185 .3337
Southern pasture area
mean .0187 -.1091 .3133 2.6419 .4025
upper limit .0485 .1965 .5634 3.8121 .5160
lower limit -.0111 -.4146 .0631 1.4718 .2989
Eastern meat area
mean .0398 .0685 .2147 .5012 .3407
upper limit .0657 .0891 .4473 .9738 .4159
lower limit .0140 .0479 -.0178 .0286 .2655
limitations applied to other types of farm management studies using

other empirical techniques.
Nicholls used weekly time series data to derive production functions
of a midwestern meat packing firm over the period 1938-39 and 1939-
3
40. He used data from 14 departments within the firm and had 101
successive weekly or time series observations for each. Production
functions were derived from groups of individual departments, with the
variables being hogs processed as output and total man-hours per week
(H) and man-hours per man per week (N) the main inputs. As an exam-
ple, the 1938-39 production function for departments 1-6 was
(1.32) X = -15.30 + 4.73N - .79N

2
+ 5.02H - .44H
2
.
This is a short-run production function where not only plant size but
also department capital is fixed in quantity. Nicholls also computed
43
Nicholls, W. H. Labor productivity functions m meat packing. University of Chicago
Press, Chicago. 1948.
some “ultra short-run” functions where only man-hours per man was
the variable. He compared results of linear, quadratic, and Cobb-
Douglas forms. From equations with three variables, he computed iso-
quants and estimated combinations of “number of workers” and “hours
per week” which minimized costs at specified wage and overtime rates.
Tintner conducted an industry study for American agriculture for
44
the period 1920-41. In contrast to the farm functions based on time
series data cited above, his was on industry studies. It was based on
time series data and paralleled the nonfarm industry studies of Douglas.
The derived function was
(1*33) Xx = -.187 + 1.698X 2 + .813X S + .007X4
where the variables are measured in logarithms with X x being the vol-
ume of agricultural productions, X 2 the employment of agriculture, X3
the operating capital of agriculture, and X 4 being time with origin in
1930-31. Tintner obtained an elasticity for labor over twice as great as
that for capital. His exponential time trend represents a yearly in-
crease of 1.6 per cent due to innovation and related factors.
Lomax made a similar industry study for English agriculture cover-
45
ing the period 1924-47. His also was a study based on time series
data and paralleled that of Douglas and associates for nonfarm sectors.
His function was
P = aE a82 C*
374
e
*°io3 R
(1.34)
where P the predicted index of output, E is the index of labor input,

is
C and R is a residual, being a time trend
is the index of capital input,
expressed in years. After examining another variable F, the input of
bought feeds and fertilizer, he concluded that its effect on production
was comparatively unimportant and could be neglected. He obtained an
elasticity of capital about double that of labor. His annual increase in
production due to trend was 1.03 per cent. Lomax discussed several
limitations of the data and estimating procedures.
Recent Production Function Studies
The above review includes only early production function studies for
agriculture or those which have contributed methodology for this pur-
pose. The review ends with a period approximating the Second World
War. Many more studies have been completed since that time. Be-
cause of space limitations, these are not reviewed but are listed in the
selected bibliography of recent analyses.
44
Tintner, Gerhard. Econometrics. John Wiley and Sons, New York. 1952. Pp. 303-304.
45
Lomax, K. S. An agricultural production function for the United Kingdom, 1924 to 1947.
The Manchester School of Economic and Social Studies, 17; 146-60. 1949.
CHAPTER 2
Economic Applications
O NE REASON for estimating agricultural production functions is to

provide basic scientific knowledge. With the large investment
made by society for estimating technical relationships in biologi-
cal and physical sciences and economic structure in farm management
science, it is reasonable to expect some of this investment should be
devoted to the basic functional relationships as they exist in nature or
phenomena. However, most scientific effort is expected to have even-
tual economic application and productivity. Hence, this chapter is de-
voted to review and summary of various economic principles and rela-
tionships as they relate to choices resting on the coefficients implicit in
production function studies. Over time, society is likely to invest but
small amounts in production function estimates unless economic cri-
teria can be employed in putting them to use in decisions which benefit
farmers and/or consumers.
Agriculture is surrounded by conditions of uncertainty, especially in
respect to price and weather. It is virtually impossible to equate price
and productivity ratios in a manner to exactly attain the profit maxi-
mizing conditions to be outlined later under these conditions of imper-
fect knowledge. Since this is true, the question may be posed: Is the
scientific interest paramount in highly refined production function esti-
mates, since farmers have limited capital and can only approximate
certain principles of profit maximization? The answer may be partly
positive, for example, in crop fertilization where price and yield from
fertilizer inputs cannot be predicted accurately over the crop season
ahead. A sufficient number of point estimates on the production function
may suffice under these conditions. Selection of one of these discrete
points in the lack of complete knowledge of the relevant functional rela-
tionship, while the optimum actually lies between it and another point,
may result in less profit sacrifice than errors in predicting price and
yield levels — and, hence, in the selection of the wrong point. Still, as
is pointed out later, more complete knowledge of the production function
may come at little cost. Consequently, there may be utility in estimat-
ing the continuous function, even though greatest errors in profit maxi-
mization may result from price and weather uncertainty. But for prob-
lems in livestock rations, such as the proportion of corn and protein
supplement to minimize feed costs for 150-200 pound hogs, the answer
31
.
32 ECONOMIC APPLICATIONS
is more nearly no. Here decisions are made in respect to inputs and
outputs of the near future where price and production can be known with
a fair degree of certainty. But even in the case of crop fertilization,
the biologically correct relationship is one of a continuous function and
the experimental and computational cost, over the long run, is probably
no greater in estimating a continuous production function than in esti-
mating a few points on it. Too, knowledge of the appropriate relation-
ships and economic principles can lead to more practical recommenda-
tions and inferences. For example, classical recommendations on
fertilization were for a single rate of nutrients for a particular soil and
climatic situation. No regard was given to prices or the farmer’s capi-
tal or tenure situation. The same was true for livestock wherein a
specific feed mix was recommended regardless of the relative prices
of the feeds. The application of principles which recognize price, capi-
tal and other market quantities, does not require that information be
retailed to farmers in complicated mathematical form.
The practical presentation of findings and recommendations should
not be confounded with the scientific estimation and presentation of
functions. These are two different activities. But the first is facili-
tated by accomplishment of the second. Hence, in this chapter we re-
view the various economic quantities and principles which are relevant
in application of production function knowledge. Emphasis is on basic
principles and relationships. Considerations in practical application
are presented later. The method used Is to illustrate economic princi-
ples and relationships with simple types of algebraic equations. These
simple algebraic equations, in contrast to use of greater mathematical
abstraction where particular algebraic forms are not specified, is to
illustrate the steps involved in deriving from production functions the
quantities of importance in economic decisions. The same steps and
procedures would be used for other forms of production functions. The
functions used are those which adapt themselves to simple manipulation
and, therefore, require minimum space in the text and minimum check-
ing by the reader. But before the principles can be illustrated, deriva-
tion of relevant quantities from production functions is illustrated.
These quantities are illustrated in the immediate following section, for
application in the economic principles illustrated in later sections.
QUANTITIES FROM PRODUCTION FUNCTIONS

Production functions, when used for economic analysis and recom-
mendation, provide one of the two sets of information needed for choice
and decision-making. The other set of information needed is price data
or other quantities which serve as economic criteria. Economic prin-
ciples and relationships using both types of information are outlined
later in this chapter. Since the physical quantities derived from pro-
duction functions are a necessary part of the data needed to apply these
principles, this section explains the quantities which can be derived
from production functions
ECONOMIC APPLICATIONS 33
Production Surfaces
As outlined in a later chapter, physical production ordinarily is a

function of many resources. To simplify the discussion which follows,
and because additional details are spelled out in later chapters, we
suppose a production function where only two resources, X
and Z, are
variable and a single product, Y, results. The setting might be nitro-
gen and phosphate used on an acre of land to produce cotton, corn and
forage fed to a cow to produce milk or labor and simple capital forms
used on a farm of fixed size. With output a function of the two input
categories, numerous types of production surfaces may result. Some
of these are explained later. For the moment suppose that the surface
takes the three-dimensional form of Figure 2.1 where resource magni-
tudes are measured along the horizontal axes or input plane while out-
put is measured vertically or in product space. For this surface or any
other surface, and also in the case of production functions representing
more than two variable resources, several sets of interrelated quanti-
ties can be derived.
Input -Output Curves
One relationship is that represented by z 2 l 1 in Figure 2.1. It repre-

sents a vertical slice through the surface, parallel to the X
axis. This
QUANTITY OF Z
Figure 2.1. Production surface representative of a function with
two variable input categories and a single product.
QUANTITY OF X
Figure 2.2. Input - output relationship between Y and X
with Z constant.
line or curve expresses output in relation to variable input of X when

the Z resource is held fixed in the quantity oz 2 . Many such vertical
slices are possible. For example, similar slices could be made where
input of Z is held fixed at oz x , oz 3 , oz 4 , etc. In fact, line z 4 l 2 is an
input -output curve, relating output to the single variable resource X,
when Z is fixed at the level oz 4 . These input-output curves can be rep-
resented in two dimensions as in Figure 2.2. A different input -output
curve, showing output for each input of X, exists for each constant level
of Z. With vertical slices parallel to the X axis in Figure 2.1, a set of
input-output curves similar to Figure 2.2 can be derived, showing the
relation between output and quantity of variable resource Z when X is
fixed in various magnitudes.
The slopes of the individual input -output curves, such as those in
Figure 2.2, indicate the marginal products of the variable input. For
the surface in Figure 2.1, marginal products can be derived simultane-
ously for both resources. The marginal products indicate the amount
added to total product by each successive unit of the variable resource.
These marginal products (i.e., the slopes over the production surface
or along the input -output curves) will be used later in applying eco-
nomic principles in specifying optimum resource use.
Product Isoquants
A second relationship of importance can be represented by hori-

zontal slices through the production surface in Figure 2.1. The height
at which we “horizontally slice” the surface represents a specific
Figure 2.3. Product contours or isoquants and isoclines.
output level. Slices at y x , y2 , and y3 in Figure 2.1 would represent

three different levels of output. Associated with each vertical slice is
a contour around the surface at constant height or product level. Hence,
it is called an isoquant. Line ab in Figure 2.1 is such a product con-
tour or isoquant. It represents an output level of y 3 and the path it

traces over the input or resource plane indicates all of the possible
quantities of X and Z which will produce this given output. Many such
vertical slices and isoquants can be derived for a particular production
surface. They also can be represented in two dimensions. Figure 2.3
provides such a map of product contours or isoquants. The curves yx ,
y 2 and y 3 are such isoquants plotted against quantities of the two re-
,
sources. The production surface is now represented by a contour map

with output measured by the specified quantities of yx , rather than in
vertical distance as in Figure 2.1. Graphed over the resource or input
plane and with each representing a particular output level, the individ-
ual contours indicate all possible combinations of the two inputs or fac-
tors which will produce the specified output. The resource X might
represent corn while Z represents oil meal. With y 2 representing 100
pounds of pork, this contour would show all possible combinations of
corn and oil meal which will produce 100 pounds of gain on a pig. Or, if
X is nitrogen, Z is phosphate and y 2 is 30 bushels, the contour would
show all possible combinations of the two materials which will produce
a 30-bushel product. The resources X and Z might represent labor and
capital respectively, while the contours specify level of cotton output on
a farm. The other contours, specified as y x and y2 , indicate the input
combinations and quantities which will produce these other pork, crop,
or farm outputs.
The slopes of each isoquant indicate the rate at which one resource
substitutes for or replaces the other, if output is to be maintained at
the specified level. When the isoquants are curved, the rate at which
one resource substitutes for the other declines, as the given output is
produced with more of the former and less of the latter. Indeed, these
slopes are crucial in specifying the optimum combinations of labor and
capital in farming, the optimum proportions of feeds in livestock pro-
duction, the optimum proportions of nutrients in crop fertilization, etc.
Similarly, the slopes of the input -output curves, in Figure 2.2, the
slopes over the surface in Figure 2.1, are important in specifying the
optimum quantity of resources to use per farm, the optimum fertiliza-
tion level and the best marketing weight or milk output of an animal. If
the contours are linear, or straight lines with a constant slope, gener-
ally only one resource or material should be used in producing the
specified output. If the input-output curves are linear, there is no limit
to the level of input and output which is profitable, if it is profitable at
all. Evidently agriculturists have sometimes supposed linear relation-
ships when, in fact, nonlinear relationships exist for particular phe-
nomena.
The lines in Figure 2.3 indicated by k x are isoclines or expansion
paths. They are isoclines in the sense that they connect points of equal
slope on successively higher isoquants. Hence, they also connect
points on the isoquants or product contours which denote equal mar-
ginal rates of substitution between factors. For this reason, they also
are termed expansion paths: they show the path which the mix of inputs
should follow if output is to be expanded. If the isoclines are linear and
intersect the origin, the combination of inputs should remain the same
as output is expanded to conform to higher product prices or lower re-
source prices. If the isoclines are curved or if they are linear and do
not pass through the origin, the mix of resources should change as out-
put is expanded or contracted to conform with changes in product or
factor prices. Two isoclines may serve as ridge lines, denoting zero
and infinite rates of substitution between resources. Isoclines k and k
x 2
in Figure 2.3 are ridge lines: k 2 intersects at slopes of infinite magni-
tude on the contours while k2 intersects at slopes of zero magnitude.
Hence, k x or k2 indicates that addition of more of X or Z, respectively,
will not replace any of the other in producing the output level specified.
If the production surface rises to a distinct peak, the
isoclines will
converge to a single point as in Figure 2.3. Rates of substitution and
marginal productivities also are zero at this point. Production sur-
faces which slope to a plateau do not have a similar convergence point,
a limitational input, for isoclines.
Algebraic Derivation
Production functions derived from experimental, sample or other

data allow algebraic estimation and arithmetic expression of all the
quantities specified above. Hence, we will specify a production
function
for two variable factors and illustrate derivation of these quantities and
relationships. While it is not taken to represent agricultural situations,
suppose the production function is known to be that in equation 2.1 be-
low. The constants a, b, and c would be derived
2
(2.1) Y = aXZ - bX2 - cZ
as regression coefficients in experiments and cross-sectional surveys

or from time series data. They would have numerical values. With
knowledge of them, we could set X and Z at various levels and estimate
a series of Y or output values.
If we plot these values of Y against quantities of X and Z, the pro-
duction surface can be drawn. Hence, equation 2.1 is the algebraic

counterpart of the geometric representation of the production function
such as in Figure 2.1.
With knowledge of the production function with two variable re-
sources, such as in equation 2.1, we can specify the algebraic input
output relationship for one resource by holding the other constant at
some specified level. Suppose that we fix Z at the level z 2 The prod- .
2
ucts aZ = az 2 and cZ = cz 2 then will become constants and can be re-
defined as mând m 2 The resulting single variable production function
.
or input -output relationship is then
(2.2) Y = m.X - bX 2 - m 2 .
If we set X at
various levels and compute the corresponding quantities
of Y, the latter can be plotted to provide a single input -output curve
such as the lower curve in Figure 2.2. By setting Z at other levels, ad-
ditional input -output curves can be derived.
The marginal physical product of a factor is measured as the first
derivative of output with respect to input. Hence, for the function in
equation 2.2, the marginal product equation for the variable resource X
is
(2.3)
H = m x - 2bX .
This equation measures the slope on the input -output curve for various
magnitudes of X. However, this marginal product equation exists for a
specified constant level of Z. We can return to the original production
function in equation 2.1, and derive the marginal product equations for
both resources. These are the first partial derivatives
(2.4) = aZ - 2bX and = aX - 2cZ .

oX oZ
From these equations, representing slopes over the production surface,

we can specify the marginal productivity of one factor when the other is
fixed at any specific level.
The maximum physical product is attained when the marginal prod-

ucts, or first derivatives, are zero. Hence, in the case of equation 2.3,
we could set this equation to equal zero, and solve for the quantity of X
which is consistent with a zero marginal product or first derivative. In
this case, a value ofX equal to drives the first derivative to
zero and defines a maximum physical product. The second derivative
can be used to define the inflexion point where both increasing and de-
creasing marginal productivity of the factor hold true.
The isoquant equation can be derived from the production function
equation by expressing input of one factor as a function of output level
and quantity of the other resource. Using X as the “dependent* re-
source we can derive an isoquant equation from equation 2.1. Collect-
ing terms and completing the square, the isoquant equation based on
equation 2.1 is
/ 2 2
(a - 4cb)Z - 4b
2b
This equation is the algebraic counterpart of contours such as those in

Figure 2.3. By setting Y at different levels, a number of product con-
tours or isoquants can be derived accordingly. Setting Y to be a spe-
cific value, various magnitudes of Z can be plugged into equation 2.5.
The corresponding values of X are then computed. If these values of X
are plotted against the “plugged in* values of Z, a single contour or
isoquant results. Setting Y at different levels and following the same
procedure, other isoquants can be derived similarly.
The marginal rates of substitution along a given contour are repre-
sented by the derivative of the isoquant equation. Also, they can be
measured as the “inverse* ratio of marginal products with a minus
sign attached. Hence, the equation of marginal substitution rates, cor-
responding to isoquant equation 2.5 is
/o an 5X ~
aX - 2cZ
{Z ' b) ~ *
5Z aZ - 2bX
Having computed isoquant sets from equation 2.5 for plotting of the
product contours, these values of X and Z can then be plugged into par-
tial derivative equation 2.6. The quantities so computed represent the
slopes of the isoquant and the marginal rates of substitution for the
corresponding points on the isoquant.
The isocline equation can be computed as follows: Set the equation
of marginal rates of substitution, the derivative of one resource varia-
ble with respect to the other, to equal the substitution rate or constant
level desired. Following the algebraic example being discussed, we
set equation 2.6 to equal -k below. Both sides can be multiplied by -1,
aX - 2cZ
(2.7) "
aZ _ 2bX
to remove the negative coefficient. Transposing from equation 2.7 and

solving for X, we obtain
2 . 8) _ 2c + ak
(
a + 2bk
Now
selecting a value for k (say that the marginal rate of substitution
is of Z for two units of X), we can plug different values of Z
one unit
into isocline equation 2.8. The computed values of X can be plotted
against the “plugged in” values of Z, with the result being an isocline
for a given k or substitution rate. Setting k at another level and fol-
lowing the same procedure, another isocline can be defined. In the
case of equation 2.8, the isoclines are linear but do not pass through
the origin.
Production Elasticities
Another quantity of some importance in production function studies,

particularly in analyses relating to product imputation, is the elasticity
coefficient. The elasticity of production indicates the change in output
relative to the change in input. If output increases by a greater per-
centage than input, the ratio is greater than 1.0; if output increases by
the same rate as input, it is 1.0; for a percentage increase in output
less than the percentage increase in input, the ratio is less than unity.
These three magnitudes of the elasticity coefficient, if all input cate-
gories are increased in the same proportion, indicate respectively in-
creasing, constant and decreasing returns to scale. The elasticity
coefficient, E , is defined as
p
( 2 9) *
^ = Si
for a single variable resource X. In other words, it is the product of
the marginal product and the reciprocal of the average product. Apply-
ing this definition for X in the function of equation 2.1, we multiply the
derivative of Y with respect to X by the ratio XY" 1 . Using the deriva-
tive in equation 2.3, where Z is fixed at z 2 level, the production elas-
ticity of X is as follows, expressed in terms of X:
m X 1 - 2bX
2
2 . 10 ) EP =
(
%X
1 - bX 2
- m 2
‘
Substituting the production function in equation 2.1, with Z fixed at z

2
level, for Y
in equation 2.10 and plugging in various values for X, we
can compute the elasticity for each value of X. For this function, a dif-
ferent elasticity of X will exist for each magnitude of this factor, and
for each level at which Z is fixed. Similar computations would lead to

the production elasticity for Z.
Elasticity of production magnitudes divides the production function
into three stages for particular phenomena. However, all three of
these stages are not present in all statistical production functions.
Stage I of production corresponds to the range of inputs and outputs
over which the elasticity of production is greater than 1.0 and, if all
input categories have been included, increasing returns to scale pre-
vail. (Generally, stages of production exist only where some factors
are fixed.) In this range, the marginal product is greater than the
average product. This is an uneconomic range of production where
fixed factors are involved since some of the fixed factor can be dis-
carded. Use of more variable resources on a smaller mix of fixed
factors then will increase total product and average product. More
output can be obtained from a given amount of variable resource.
Stage II corresponds to the output and input range where the elasticity
coefficient is less than 1.0 but greater than zero. It is within this
range in which economical combinations or use of resource is found;
within this range, the principles outlined in this chapter can be applied
to maximize profits. Stage III corresponds to output ranges where the
elasticity coefficient is less than zero. Use of more variable resource
depresses total output and average product. Hence, range HI never de-
notes economical production since discarding some variable resource
1
will increase return to both fixed and variable resources.
PROFIT MAXIMIZING QUANTITIES OF INPUTS
Recommendations in agriculture ordinarily are in terms of inputs.

The agriculturist suggests quantities of fertilizer, seed or insecticide
to applyon a fixed producing unit such as an acre. Or, he recommends
daily rations for a dairy cow. Hence, we review principles which em-
phasize specification of the magnitude of optimum inputs rather than
direct specification of optimum outputs. (Obviously, specification of
the optimum input implies designation of the optimum output and vice
versa.) First, consider the case of a single variable resource where
production is related to a fixed plant such as an acre or an animal.
Here we suppose that the production function and product and factor
prices are known. Profit is maximized for the fixed unit if the mar-
ginal value product of the factor is equal to the marginal cost of the
factor. We will use a form of production function involving only simple
calculations, although the steps and principles illustrated would apply
to any function. Assuming the production function in equation 2.11,
with Y representing magnitude of output, X magnitude of input, and b
(2.11) Y = a + bX - cX2
1
For additional details on stages of production, see Heady, Earl O. Economics of agri-
cultural production. Prentice -Hall, Inc., New York. 1952. Ch. 3.
and c positive constants, the total value product, V, is given by equa-

tion 2.12 where output in equation 2.11 is multiplied by Py ,
the product
price.
s
(2.12) V = Py Y = aP bPy X - cPy X
y +
The marginal value product is the derivative of equation 2.12 while the
marginal cost of the resource is P x , the price of the resource in a
competitive market. Equating the marginal value product of the re-
source with its marginal cost, we obtain equation 2.13.
(2.13)
^ = bPy - 2cP y X = Px
Transposing terms and dividing by the coefficient of X, we specify the

magnitude of input which will maximize profit as that in equation 2.14.
bPy - PX
(2.14)
2cPy
As is indicated in later chapters, the coefficients b and c will or-

dinarily be derived from regression equations estimated from experi-
mental or sample data. With these estimates available and knowing the
product and factor prices as P and Px , respectively, these quantities
y
can be substituted into equation 2.14, with the magnitude of X specified
accordingly. This value of X indicates the magnitude of input or treat-
ment which will maximize profit from the fixed factor. It might be, for
example, 40 pounds of nitrogen on an acre of corn. Substituting this
value, X = 40, back into the production function in equation 2.11, we
could also compute the optimum output or yield level.
The optimum quantity of input is a function of factor and product
price, with b and c being constants in the production function. As Px
increases, the optimum magnitude of X declines. As P increases, the
y
optimum level of X increases. A decline in factor and product prices
will have the opposite effect on the optimum magnitude of X.
If fixed costs are involved in use of the variable factor, profit may
be examined better from the standpoint of a profit function. Profit from

the use of a factor is the difference between total return from the fac-
tor and its total cost. The total cost of using the factor is equal to the
sum of fixed costs, K, involved in its use plus the value of the quantity
of the variable factor used, or K + Py X. Subtracting this total cost
function from the total revenue function, we can define total profit, II ,
as in equation 2.15.
(2.15) n= aPy + bPy X - cPy X2

- (K + P x X)
Profit is maximized when the marginal profit from use of the resource
is zero. Hence, defining marginal profit as the derivative of the profit
function with respect to the variable resource, we obtain equation 2.16.
(2.16) bpy - px - 2cp

y
x
1£r=
Equating equation 2.16 to zero and solving for X, the optimum input (the
profit of factor), is again that in equation 2.14.

maximizing quantity
The same quantity of input is specified because fixed costs do not affect
the magnitude of marginal costs or marginal profits. Since this is true,
a more apparent relationship between factor and product prices and
production coefficients can be indicated. Profit from use of the factor
also is maximized when its marginal product is equated to the price
ratio, price of factor divided by price of product or
The marginal product of the factor is the first derivative of the produc-
tion function in equation 2.11. Setting the derivative of equation 2.11 to
equal the factor to product price ratio, we obtain equation 2.18.
x P
b - 2cX = -~~-
(2.18)
Fy
Transposing and dividing by 2c, we obtain the quantity of X which maxi-

2
mizes profit. It again is that determined for equation 2.14. But from
equation 2.18 it is obvious that to maximize profit the marginal product
will need to decrease, through addition of X, as factor price decreases
relative to product price. An increase in the magnitude of Px relative
to P will call for an increase in magnitude of the marginal product, b
y
- 2cX, by a reduction in magnitude of X.
Classically, economists have defined total revenue and total costs
as a function of output and have defined optimum production in terms of
output level which maximizes profits. While, as indicated earlier, this
quantity can be derived from the steps, described by equations 2.11
through 2.18, we illustrate it with the determination now being first in
terms of specifying the optimum level of output. Hence, multiplying
output, Y, by P the price per unit of output in a competitive market,
y ,
the total revenue function is
(2.19) R = Py Y .
2
That solution by the approach of equation 2.18 is the same as that by equation 2.15 is
easily illustrated. From equation 2.15 we obtain equation 2.16 and set it equal to zero, or
bP
y - Px - 2cP
y
X = 0 .
Adding Px to both sides, we have

bP
y - 2cP
y
X = Px .
Now dividing through by Py we obtain the quantity already expressed in equation 2.18; indi-
cating that if we follow through from the profit equation in 2.15, we actually equate the mar-
ginal product with the factor to product price ratio.
The total cost function is the cost per unit of Y multiplied by the mag-
nitude of Y. The total cost function was represented in equation 2.15 as
(2.20) T = K + Px X ,
a function of X. However, since profit maximization is now to be de-

fined in terms of output, the total cost equation must be expressed as a
function of Y. Thus from the production function in equation 2.11, we
first express input as a function of output. For the particular algebraic
form, this is accomplished by transposing terms with X and 2 to the X
left and Y to the right of the equality and completing the square. Solv-
ing the quadratic equation for value of X, we obtain equation 2.21.
ml 2
(2.21) X = (2c ) [b t V 4c(a - Y) + b ]
With input defined as this function of output, we substitute the value of X

in equation 2.21 into the cost equation in equation 2.20. This quantity is
subtracted from total revenue to provide the profit equation in equation
2 . 22 .
Px [b t V 4c (a - Y) + b2 J
( 2 22 )
. n = Py Y - K
-
+
2c
Profit again is at a maximum when marginal profit is zero. Marginal

profit now is the first derivative of the total profit equation in equation
2.12, or the first derivative of II with respect to Y as represented by
equation 2.23.
dn px
(2.23)
y (4ac - 4cY + b 2 )- 5
Equating equation 2.23 to zero, and solving for Y, we obtain the equi-
librium output as that in equation 2.24.
2
b Px
(2.24) 2
4c 4cP
Now if we substitute the value of Y in equation 2.11 for that on the left
in equation 2.24 and solve for X, we obtain the equilibrium input already
indicated in equation 2.14.
Other Criteria
The principles outlined above suppose that capital is unlimited and

profit from the fixed unit can be maximized. However, typical farmers
do not have unlimited capital. They must employ other, but related,
economic criteria in specifying the optimum quantity of variable factor
to be used per unit of fixed factor. The optimum input defined above is
the maximum quantity of the resource which should be used. Most

farmers will need to use less, if they maximize profit in their limited
capital situation. Similarly, a minimum quantity of input can be defined,
with farmers needing to use more if they are to make any profit. In case
the production function involves only decreasing marginal returns and
fixed costs are not involved, the minimum quantity of input to be used
is zero. More than zero quantity should be used only if the price of the
product is sufficiently large relative to the price of the factor. How-
ever, if fixed costs are involved, the minimum quantity of input which
should be used differs from zero even if the production function in-
cludes only diminishing returns. The minimum input is defined by
equation of total cost of and total revenue (total return) from use of the
factor. In a diagram of total costs and total revenue related to use of a
variable factor such as Figure 2.4, this minimum quantity of input is
ox^ With fixed costs of oc 1? none of the factor should be used if price
of the product is sufficiently low (i.e., if the total revenue curve is
lower throughout than the total cost curve for the resource). But with
a product price which is sufficiently high relative to the factor price,
the minimum quantity of factor which should be used is ox 1 For inputs .
larger than this, revenue exceeds costs and net profit is generated.
Profit from the fixed producing unit can be increased until input
reaches the magnitude of ox 2 the level which maximizes profit for the
,
fixed producing unit. At ox 2 the difference between total revenue and

,
total cost, rs, is at a maximum. At this input, marginal cost and mar-
ginal revenue from using the resource are equal, as denoted by equality
of the slopes of the two curves R and T. The optimum input quantity
for a farmer with limited capital and other profitable investment oppor-
tunities will fall somewhere between ox x and ox 2 Given knowledge of
.
the production function, the minimum input which is profitable can be

determined algebraically. With the production function in equation 2.11,
the total revenue from resource use is equation 2.12. With the total
cost function in equation 2.20, where K is fixed cost from use of the re-
source and Px is the price per unit of resource, total revenue and total
costs are equated in equation 2.25 below.
2
(2.25) aPy + bPyX - cPyX = K + Px X
Transposing terms, we obtain equation 2.26. Completing the square and

solving for the roots of X, we obtain equation 2.27 which provides the X
quantity defining the minimum quantity of resource which can be used
profitably.
(2.26) CPyX 2 + (Px - bPy)X = aPy - K
(2.27) X =
Ll!
(/)
cr
O X, X2
QUANTITY OF RESOURCE
Figure 2.4. Total revenue from and total costs of using a
variable resource in relation to minimum and
maximum quantities of input.
If the production function has ranges of both increasing and decreas-
ing marginal productivity (i.e., ranges over which the elasticity of pro-
duction is both greater and smaller than 1.0), the minimum input which
should be used is defined at the point where the marginal product is
equal to the average product (i.e., the input where the elasticity of pro-
duction is equal to 1.0). Details indicating why this quantity of resource
is minimum for a production function with a range of increasing returns
3
are given elsewhere. Again, the quantity can be derived algebraically
for a given production function. Assuming the algebraic form of pro-
duction function equation in 2.28, the marginal product equation or de-
rivative is equation 2.29. The average product is the total product di-
vided by total input, or the right-hand side of equation 2.28 divided by X.
2 3
(2.28) Y = bX + cX - dX
3
Heady, Earl O. Economics of agricultural production and resource use. Prentice-
Hall, Inc., New York. 1952. Pp. 91-92.
2
(2.29) ~| = b + 2cX - 3dX
Defining the average product thus and equating it to the marginal prod-
uct in equation 2.29, we obtain equation 2.30. Completing the division
2 *?X + cX 2 - dX3
(2.30) b + 2cX - 3dX =
on the right-hand side of equation 2.30, collecting terms and solving

for X, we obtain equation 2.31, indicating the minimum quantity of fac-
tor which should be used per unit of fixed resource, if input is to be
greater than zero.
(2.31) X = .5cd“x
Return for Investment in Variable Factor
For many farmers with limited capital, the quantity of variable in-
put which maximizes the rate of return on investment in the variable
input is perhaps as relevant as the input which maximizes profit in the
sense of equating marginal costs and revenues. For a function with
only diminishing marginal returns, the quantity of input which maxi-
mizes return per dollar invested in it is zero. However, when there
are fixed costs, K, of using the factor and some nonzero quantity of
factor is profitable, the input which maximizes return on investment in
the variable factor is not zero. The input of variable factor which max-
imizes rate of return on investment is fixed and variable costs for it
can be defined as follows: total monetary return is determined by mul-
tiplying output by product price as in equation 2.12 for the production
function in equation 2.11. The total revenue function can be divided by
the total cost function, to provide an equation defining return per dollar
invested in the factor, a rate of return equation. The marginal rate of
return per dollar invested is the derivative of the rate of return equa-
tion. Setting this derivative, the rate of return in respect to the varia-
ble factor, to equal zero and solving for the value of input, we can indi-
cate the quantity of variable factor which maximizes rate of return on
investment in the variable resource. The production function in equa-
tion 2.11 can be used to illustrate the procedure. The total revenue
function is equation 2.12 and we divide it by the total cost function in
equation 2.20 to define the rate of return on investment, r, as different
inputs are used, as in equation 2.32.
aPy + bPyX - CPyX 2

(2.32)
KT-Plj
The marginal rate of return, as a function ofinput, is the derivative of
r with respect to X as shown in equation 2.33. Setting equation 2.33 to
2
dr_ (bK - aPx )PV - 2cKPv X - cP v Px X
=
dX K2 + 2Px KX + P2 X 2
equal zero, collecting terms, completing the square, and solving for the
roots of the equation, we obtain the value of X in equation 2.34. This is
-2cKPy t [4c
2
K2 Py +
2
4cP P2 (bK - aPx )]'
(2.34) X = _____
the quantity of input which will maximize the return per dollar invested
in the variable factor, considering the magnitude of fixed costs. It in-
creases as product price increases or factor price decreases. It de-
creases for price movements in the opposite direction.
As a "rule of thumb* criterion, to be used apart from analysis of
the whole farm business and in making input recommendations to
farmers who are short on capital, the value of input which maximizes
return on investment in the variable factor is perhaps more relevant
than the input magnitude which defines the minimum input described
earlier or the maximum input defined by a marginal product which is
equal to the factor to product price ratio. The farmer with limited cap-
ital will, of course, maximize profits if he invests in each input cate-
gory such that the marginal return per dollar invested is equal among
all input categories. (We have been considering only a single variable
input here. Hence, the problem posed falls more correctly in sections
to be outlined later.) Thus specialists working separately (such as
specialists in fertilization, hog feeding, or dairy nutrition) on a partic-
ular input category, and without attempting to program profit maximiz-
ing programs for the entire farm, might best provide information of
this type for farmer use under the following procedure: Each special-
ist should provide the function indicating the marginal "rate of return*
per dollar invested in his specialized resource (or resource collection)
as different quantities of it are used. The farmer could take these
several sets of information and, considering his capital and the num-
bers of different types of producing units such as acres and animals,
decide on how much of each type of variable resource to use with each
fixed acre and animal, if profit from the farm is to be maximized. He
would allocate investment to each category of variable resource so that
the marginal "rate of return* is equal among them. A means of doing
this is provided by equations such as equation 2.34. By inserting dif-
ferent values of X in such an equation, marginal returns per dollar in-
vested in the variable factor can be determined as level of X increases.
Given this information for numerous input categories (provided by sep-
arate agricultural scientists) such as feed, fertilizer, and insecticide,
the farmer could estimate rather quickly the amount of each which
should be used if he were to allocate his limited funds optimally among
them. Bu]t if this degree of refinement were to be used, programming
procedures might be more appropriate.
Allocation of Limited Capital
An allocation of limited capital among product investment opportu-

nities when continuous production functions are known can be illustrated
for a simple situation. Here we might suppose that a single variable
resource such as fertilizer is involved, although the basic principle
would be the same if other types of resources and producing units also
were involved. Suppose that we have three crops, Yx Y2 and Y3 , ,
.
Each has a production function of the form in equation 2.11, Hence, a

subscript is attached to the constants in each equation to indicate the
product to which it refers. Similar subscripts indicate the quantity of
the resource (e.g., nitrogen fertilizer) used for each product. Funds
are available for only a fixed quantityôf the resource indicated as X.
What proportion of the total quantity X should be allocated to each
crop? The marginal physical products of the variable resource are in-
dicated in the first column of equations in system 2.35. By multiplying
these equations by Px P 2 , and P3 the prices of the three crops, we
, ,
obtain the three equations for the value of the marginal products
(VHP’s) which are identical with the marginal value productivities for
a competitive industry such as agriculture, in the second column of
equations in system 2.35.
bA - c1 X 1 VMPX = b1 V 1 - cx P x X1
dX
(2.35)
—dY 2
= b2 - c 2 X2 VMP2 = b2 P 2 - c2 P 2 Xg
dY 3
dXo
h* - C3 X 3 VMP 3
= - C3P3X3
With a limited quantity, X, of the variable resource, we wish to al-

locate a quantity to each crop so that the marginal value productivity of
the resource will be equal for the three crops. We do not know the
magnitude of the marginal value product which will exist when re-
sources are allocated to each crop so that this quantity is equal for all.
Hence, we will denote it as m, and set all three marginal value produc-
tivities to equal this value as in the first three equations of system 2.36.
kj.Pi - c iP iî = m
(2.36)
fr 2 P2 -X2 C2 P 2 = m
b3 P ~ C 3 P 3 X3
3
= m
X, + X2 + X3 = X
But we also have a restriction on resource allocation to be met. It is
indicated in the last equation of system 2.36, stating that the sum of the
quantities of the resources allocated to each crop must equal the total
available quantity, X, of the resource. The unknown quantities in this
set of four equations are X X2 X3x, , ,
and m. All other quantities are
known as parameters of price, resource supply or of the production
functions. Hence, system 2.36 is a system of simultaneous equations
which can be solved for the unknown values. This may be done as fol-
lows: Transpose all terms of known quantities to the right-hand side
of the equations and all terms with unknown quantities to the left-hand
side, as in system 2.37; we have also introduced each unknown quantity
-c.P.X, + GX 2 + 0X 3 - 1m -bi*x
OX, - C
2
P 2 X2 + 0 X3 - 1m -b 2 P 2
(2.37)
0X x + 0X 2 c3P3 X3 - lm "^ 3^3
1X X + IX 2 + 1X 3 + Om = X
in all equations, by giving a zero coefficient for the equation in which
it
it did not previously occur.Such a formulation more clearly suggests

the algebraic manipulations which follow. The coefficients of the X*
and m
form a coefficient matrix for equation 2.38 while the Xi form a
vector of unknowns to be solved from the vector of known parameters
4
on the right of the equals sign .
f-CxPx 0 0-1 xf *1*1

0 1 hj to 0 1 X2 =
-b 2 P 2
1110
0 0 -C3P3 -1 X
m
3
- 1*3
X
P3
Denoting by A the matrix of coefficients, by X the vector of unknowns,

and by S the vector on the right, we have the matrix product:
(2.39) AX = S
Now inverting A, the matrix of known coefficients, to obtain A"*1 ,

its in-
verse, we have
*" 1
(2.40) X = A S
from which we form a system of scalar equations expressing the value

of X,, X 2 X 3 and m. Letting cy represent the elements of this 4x4
, ,
order inverse matrix, we can compute the values of X 1? X 2 and X 3 , the ,
4
Readers unacquainted with solving systems of equations are referred to the relatively
simple explanation given m
Chapter 13 of Heady, Earl O. and Candler, Wilfred V. Linear
programming methods. Iowa State University Press, Ames. 1958.
quantities of resource to be allocated to each crop and m, the magni-
tude of the marginal value product, as follows:
X 1
= -c
u bi P i - c
12
b P
2 2
- c
13
b P + c
3 3 14
“C ~ C 22^2^2 - C 23^3^3 + C
(2.41) 21 î 24
Xs = -c sl b x Pi “ C 32^2^2 " C 33^3^3 + C 34
m = -c b P
4i x x
~ C42^2^2 ~ C 43^3^3 + C44
Hence, we have established the quantity of the variable resource to al-

locate to each cropJ:o provide equal marginal value productivities. With
the total quantity, X, allocated to provide an equal marginal value prod-
uct, m, for each crop, profit from use of the fixed quantity of the re-
source is at a maximum. This procedure in resource allocation is ap-
propriate for farmers who have limited funds and cannot apply variable
resources to each fixed producing unit until marginal physical products
are equal to factor to product price ratios. While the principle has
been illustrated with a particular algebraic form of function, the same
procedures generally apply to other types of equations, although certain
other forms of equations are more difficult to manipulate. This prob-
lem could also be approached by the empirical technique outlined later,
using Lagrangian multipliers to equate the returns from several re-
sources when expenditures on them are limited by a fixed capital re-
straint.
Perhaps more often the farmer has a limited fund with which he can
purchase different types of physical resources, such as different nutri-
ents or different mixes of nutrients, which he may apply on different
enterprises. Then, in contrast to the case of equations 2.35 through
2.41 where a fixed quantity of a given type of physical resource is allo-
cated to different crops, the problem is to allocate the different re-
sources so that the marginal value return is the same per dollar in-
vested in each type of resource for each enterprise. In this case, an
equation such as 2.32 should be constructed for each crop. Then setting
the partial derivatives of these with respect to the X*s equal to v, a
quantity expressing a given magnitude of marginal value for each type
of resource used for each crop, a system of equations similar to sys-
tem 2.36 can be formulated. Using a restraint such that the value of
the resources used, the sum of the products of prices times the quanti-
ties of resources, must equal the fixed expenditure, an equilibrium so-
lution can be derived by the same procedure. It will indicate the quan-
tity of each resource which should be purchased for each crop if profits
are to be maximized. In many cases, practical solutions to problems of
where many investment opportunities exist,
this type, particularly may
be more easily obtained by linear programming procedures. 5
5
See Heady, Earl O. and Candler, Wilfred V. Linear programming methods. Iowa State
University Press, Ames.
Production Possibilities
An alternative approach also allows specifying optimum allocation

of a given quantity of resource such as fertilizer, among competing
crops. This is by means of a production possibility curve. For exam-
ple, suppose that we have the two production functions
bz
(2.42) Y 1 = ajX^ 1 and Y2 = a2X
where Yx and Y 2 are the outputs of two crops, X x and X 2 are the amounts
of a resource such as fertilizer allocated to each and X is the total
quantity of resource available. From the production functions in equa-
tion 2.42, we can derive the requirements equations in equation 2.43 in-
dicating the total amount of resource required to produce a given output
jl
bl 2
(2.43) X1= a
1
X2
of each product. Now the restraint
(2.44) Xx + X2 = X
exists^jlndicating that the sum of resources used on each product must
equal X, the total amount of resource available. Now substituting into
equation 2.45 the values of Xx and X2 in equation 2.43, we obtain
(2.45)
r
}1 + a.
2
Y2 = X
which also is a resource restraint or requirements equation expressed
in terms of the variable outputs Yx and Y 2 By subtracting the product .
containing Y 2 from both sides of equation 2.45 and dividing by the coef-
ficient of Y 19 we derive the production possibility equation 2.46, with
1 x
\ b.
(2.46) r, (•* X - a. ?)
output of Y 1 expressed as_ a function of Y2 when quantity of variable re-
source is fixed at level X. This equation shows the amount of Yx , which
can be produced when Y 2 is at different levels. From this equation, we
can derive an equation of marginal rate of substitution, indicating the
amount of Yx which must be sacrificed for each unit gain in Y2 . The
latter equation is the derivative in equation 2.47. Now setting it to
b i -1
dY, _i 2 \ f -1
1
(2.47)
dY2 (- b 2 )
equal -P2 ,
the ratio formed by dividing the price of Y2 by the price
/Px
of x
Y ,
we
specify the combination of Y x and Y 2 which will maximize
profits. Setting 2.47 to equal the price ratio and solving for Y2 , we
have the value below.
b2
bi-b 2
(2.48) Y2 =
Substituting this value into equation 2.46, the value of Yx can be derived
similarly. From these quantities substituted into equation 2.43, the
quantities of resource to be used for either enterprise can be deter-
mined as a function of the calculated output of either crop and of the
prices of the two crops. But, again, alternative empirical approaches
such as those mentioned above and linear programming may be compu-
tationally desirable.
TWO OR MORE RESOURCES

Profit maximizing and related quantities of inputs have been speci-
fied above from production functions involving a single variable re-
source. Input quantities which represent an economic optimum can also
be derived when two or more resources are variable. But in addition,
another problem exists; the proportion of inputs which minimize the
cost of any output must be determined. We explain these two steps im-
mediately below. While the illustrations suppose a production function
with two variable resources, the steps outlined can be generalized to n
variables.
Profit Maximizing Quantities of Inputs
Given a production function where output is a function of two re-

sources, as in equation 2.49, profit is maximized from the fixed pro-
ducing unit when the marginal products of the two resources, or n re-
sources if more than two are varied, are simultaneously equal to their
(2.49) Y = a + - b 2Xi + b 3 X2 - b4 X^ + b^X-s

respective prices divided by the prices of the product. This condition
is illustrated in equation 2.50 where the marginal products on the left
6Y
= t>x - 2b2 X, + b.X, =
SK X f-
(2.50)
,6Y,
= b3 - 2b4 X, + b 5 X, =
ex 2 J-
(with the first equation for the marginal product of X ± and the second
equation for Xg ), are equated to the price ratios. Px is the price per
unit of the resource represented by X x , P 2 is the price for resource X 2 ,
and P y is the price for the product. Solving these two equations simul-
taneously, because two unknowns are included in the two equations, we
have the equilibrium quantities of X x and X 2 below. These are the profit
maximizing quantities of the two resources which should be used, for
1 x
X, = [(2b4 P x + b5 P2 )P^ - (2b, b4 + b3 b 5 )] (b l- 4b2 b4 r
(2.51)
X2 = (2b2 P 2 + bjPJP" 1
- (2b 2 b3 + b,b 5 )] (bf - 4b 2 b4
)~1
[
the particular form of production function under analysis, when capital

is not limiting. The procedure has been illustrated with a particular
6
form of production function and two input variables. But the same pro-
cedure applies to n input variables. The marginal products, partial
derivatives of output with respect to inputs, are obtained and equated to
the price ratio. If interaction between inputs is not present in the pro-
duction function equation, the amount of each factor can be solved di-
rectly from its partial derivative equated to the price ratio. If inter-
action is present, the n partial derivative equations must be solved
simultaneously. With more than two variables for the function used we
would need to invert the matrix of coefficients to solve simultaneously
for the unknowns.
Inputs Under Limited Capital
funds are limited, the criteria of minimum inputs and the input
If
levels which maximize rate of return on investment, as outlined for a
single resource, can be computed for two resources and may be as rel-
evant as the maximum quantities described in equation 2.51. But with
two resources an additional problem exists: namely, using a quantity of
each such that the marginal returns on investment are equal for the
two. These quantities can be computed by procedures similar to those
outlined for equations 2.37. The procedure for determining profit
6
For n variables, we can write a profit equation where II is profit, is the price and
(a) n = Pyv Y - 2 P,
A
X. - K
1=1
Xi is the amount of the i-th factor while K is fixed costs, if any. Profit is maximized when
the n partial derivatives of the nature of (b) are set to equal zero and the values of X are
1
5JL . 5Y
(b) p
6X { " y 5X i ~
determined; from single equations if no interaction and simultaneously if interaction is

present. Obviously, setting equation (b) to equal zero, we can derive MPP = factor to prod-
5Y p.
uct price ratio or for all n variable resources, the same form of equation as in
0A£ Fy
equation 2.7 of the text.
maximizing input quantities in equation 2.51, with capital considered to

be unlimited, results in the same return on the last dollar invested in
both X x and X 2 Where capital is unlimited, attainment of this condition
.
for fertilization of a crop acre with two nutrients or for marketing

weight of an animal fed n kinds of feeds could be attained for every
producing unit and every enterprise on the farm under static conditions.
But since uncertainty gives rise to capital restrictions, profit maximiz-
ing inputs cannot be applied to each animal, acre or enterprise (i.e., to
equate profit derivatives to zero) Profit from the farm as a whole is
.
maximized when inputs are applied to all enterprises so that the ratios
P v 5Y
-pf are equal for all inputs to a particular animal or acre and for
all individual inputs, to different producing units and enterprises.
,
Under this condition the last dollar invested in each input for each pro-
ducing unit and enterprise is the same. Differential equations can be
used to express this equilibrium condition but such detailed data, even
if uncertainty did not exist, are not available for farms. The most
practical empirical procedure for approximating this broad equilibrium

condition, considering types of data usually available and the lack of
precision necessitated by uncertainty, is linear programming where ac-
tivities considered are fewer than those supposed by a set of n contin-
uous functions.
More typically, a farmer with limited capital may wish to answer
such questions as the following: “Given two or more input categories
how much of each should I use when I have a fixed amount of capital to
invest in them?” In other words, how much of two nutrients should he
apply to an acre when he has an amount of capital equal to C. He can-
not attain the per acre profit maximizing quantities explained above.
But he should use an amount of each nutrient such that return on the
last dollar spent on each is equal. Without capital limitations, the
profit equation is equation 2.52 where Pi is the price and Xi is the
(2.52) n = Py Y - (Px Xi + P2 X2 + ... Pn Xn) = PyY - 2 P Xi

t
i=i
quantity of the i-th resource. Of course, in solving this quantity, the

production function equation is substituted for Y with the partial deriv-
ative for each factor equated to zero to give
5Y
(2.53) p
Fy _ Pi -
6Xi
But when capital is limited to some level C and we wish to equalize the
marginal return per dollar invested in each factor, the profit equation
must be modified as in equation 2.54 where X is a Lagrange multiplier.
n n
(2.54) n = PyY . 2 PiXj + X( 2 Pi Xi - C)
i=i
C is the fixed capital level and P* and are as defined above. Again
in actual calculations, we would substitute the computed production
n
function equation for Y. The term 2 Px X i - C indicates the restraint
i=i
n
that the total amount spent, 2 Pi Xi, on the factors can be no greater
i-i
n
than the amount of capital C. When 2 Pi Xi and C are equal, then the
i=i
term is zero. To determine the profit maximizing amount of each input

(i.e.,with return on the last dollar spent on each equal), we take the
partial derivatives of profit with respect to each input category and X
and set them to equal zero. Hence for equation 2.54 we would substi-
tute in the production function equations for Y, compute the partial de-
rivatives and set them equal to zero as:
an
" P y ®x_ ' Px + XP x = 0
aX x 6x x
an
ax
= P y ST
5X
_ p2 + = 0
2 2
(2.55)
6x;=
py
W n
- p» + xp n= 0
an n
hr m & *xi-c.o .
These equations can be solved for the Xi and X to determine the magni-
tudes of inputs which maximize profits under the restraint set out
above. With a number of resources and interaction terms in each of
the differential equations, we would invert the matrix of known coeffi-
cients as outlined for system 2.37 and solve the Xi and X values from
the resulting system of scalar equations as in system 2.41. The value
of X can be obtained by solving one of the equations for it, with this
value of X substituted into the other equations. Having solved for these
values of the they can be substituted back into the original produc-
tion function equations and its marginal product equation to indicate re-
spectively the expected yield and the marginal productivities, both
physical and value, for the various nutrients or resources.
The inputs specified are, of course, on an isocline defining the min-
imum cost mix of factors for the particular output. Marginal value
productivities of resources, relative to prices of factors, are always
equal when the least -cost mix has been specified along an isocline. The
method outlined above simply indicates where, along the isocline, input
expansion should stop when capital available for purchase of inputs is

limited. This is easily shown by setting the marginal rate of substitu-
tion, the partial derivative of one factor with respect to the other, for
two factors, X and Z, equal to their inverse price ratio:
6X Pz_
(2.56) = “
az px “
Since the marginal rate of substitution is the (minus and inverse) ratio
of the marginal products, the same condition is expressed in equation
2.57.
5Y i 5Y _
Pz
(2.57) '
“
az / ax "
px *
From equation 2.57, we can derive equation 2.58.
6Y i
_
ay i
(2.58)
az p 2 ax p x ’
The latter states that the ratio of output is equal relative to the value of
input for each factor.
m Products and n Factors
The procedure outlined above, or that explained for the problem of

equation 2.35, also can be used for m
products and n factors. In this
case the profit function is equation 2.59
m n n
(2.59) n PyjYj - .2 PijXij + X(.2 P X £j C)
=
£ y -
where Pyj price of the j -th product, Yj Py* is the price of the
is the ,
i-th resource, Xy, used

for the j -th product, X is a Lagrange multi-
plier, and C is the fixed capital amount available for expenditure on all
factors used for all products. Profit is at a maximum when the partial
derivative for each factor used for each product is equated to zero as
in equation 2.60.
( 2 . 6 °) 0
Hence, computing the (m x n) +1 differential equations for manageable

types of functions, if there is one for each of the n factors used on
each of the m
products plus one for the Lagrange multiplier, we can
solve for the X values as outlined above; inverting the coefficient ma-
y
trix and computing the scalar equations. While not all factors need to
be used on all products, the same procedure still applies. But as men-
tioned previously, linear programming or related approaches may pro-
vide sufficient accuracy for decisions on inputs allocated to agricultural
products where price and weather uncertainty is large. Sufficient pre-
cision, within this imperfect knowledge situation, may be attained by
defining a number of discrete points on isoquants and input -output
curves as separate activities.
TENANT INPUTS
Optimum input quantities under situations of limited capital have
been defined above. Another situation calling for slight modification of
principles defining optimum input quantities is that of rented farms.
Numerous arrangements exist on the proportion of (a) the price of in-
puts paid, and (b) the share of the product received by tenants or land-
lords. We could define optimum allocation under all of the conditions
outlined above for an owner. However, we will restrict the example to
the case of input quantities to maximize profit from a fixed producing
unit. Suppose that the tenant (landlord) receives a share of output equal
to s and pays a proportion of input price equal to r. Thus, his share
of total output is the production function multiplied by s, or
(2.61) Y = sf(X x , X 2 ... X n )
and the share of marginal product to him is the full marginal physical
product of each factor multiplied by s as in equation 2.62
6Y
(2.62) s • MPP i
= s
6Xi
for the i-th factor. Hence, his task is to equate this marginal product
to the factor to product price ratio multiplied by r for all factors as in
equation 2.63.
5Y
(2.63)
6X^
His optimum magnitude of marginal product is thus
(2.64) MPP, =
f ^ J-fL
for the i-th factor. Setting the marginal product to equal this quantity,
the magnitude of Xican be solved by the procedures outlined previously.
If the tenant pays the entire price or cost of factors and receives only a
half share of the product, so that r is one and s is 0.5, the tenant’s
optimum level of input for the i-th factor is defined by a marginal
product equal to twice that of an owner. Or we can generalize as fol-

lows: If r > s, the tenant marginal product should be greater than for
owner. If r = s, the tenant marginal product should equal that of owner
If r < s, the tenant marginal product should be smaller than for owner.
Minimizing Cost of Output and the Optimum Resource Mix
Under conditions of capital limitations, the problem of the input mix

to be used for a given output is especially relevant. The minimum cost
combination of resources is determined by equating the marginal rate
of substitution ofresources to their inverse price ratio. Hence, for the
production function, equation 2.49, we can derive isoquant equation 2.65.
2
b x + b 5X 2 ± V4b 2 (a - Y) + 4b 2b3X 2 - 4b 2b4Xa + Oh + b*X2 )
(2 - 65) xi= 2b;
Itexpresses resource x
X
as a function of resource X 2 for a specified
level of output, Y. It shows the amount of Xx which must be used for a
given level of 2
X
if output is to be maintained at some fixed level Y.
The marginal rate of substitution of X 2 for X x is the derivative of equa-

tion 2.65 with respect to X 2 . With the marginal rate of substitution
equation equated to the price ratio, we obtain equation 2.66 where P is
1
6X x b3 - 2b4 X2 + b 5 X x P2
(2 . 66 )
5XT S 2b 2 X X2 =
' -
x
+ b5
“X
the price per unit of X and P 2 is the price per unit of X, 7 Solving for
x
X x in terms of X 2 from equation 2.66, we obtain the isocline equation in
equation 2.67. It indicates the amount of X x which should be used for
(2.67)
(b 1 P~
1
P2 - b 3 ) + (2b4 + b^ 1
p2 )X 2
“1
(1 + 2b 2 P1 P2 )
the factor prices Pj and P2 if the output represented by Y in equation

2.65 is to be produced at minimum cost and X is to be at the specified
2
value. This same point was illustrated earlier by equating the deriva-
tive of the isoquant to a constant substitution rate.
Equation 2.67 defines the amount of X which is optim um , in
x
7
The equality in equation 2.66 obviously is that obtained if we divide the marginal prod-
uct equation of equation 2.50 for 2
X
by the marginal product equation for X, in equation 2.50.
Performing this division, we have
/«Y ji /R h. lx.
(a)
6X*/ eXi' Py/ Py" Py P,
and since the Py cancel out on the right of (a), we have, without the minus sign, equation
’s
2.66. Obviously, then, if factor inputs are employed to maximize profits as explained for
equation 2.50, the corresponding output also is produced at minimum cost as explained for
equation 2.66.
minimizing cost of a given output, when X 2 also is defined for a given

output. Hence, we can substitute this value of X x from equation 2.67
into the isoquant equation 2.65. The resulting equation then includes
only X2 as an unknown and we can ascertain its value, for the stated Y
value, from the known coefficients. Substituting this value of 2 into X
the isocline equation 2.67, we solve for 1
. X
The result indicates the
magnitudes of X x and X 2 which minimize costs for a given output. If we
have already computed a large number of X± and X% input pairs corre-
sponding to a given Y or output level, and if the corresponding marginal
rates of substitution have been computed, we also can pick the substitu-
tion rate which is nearest the factor price ratio in question. This
method would suffice where only approximate values are required.
COST CURVES
While not so frequently used for these purposes, production func-

tions also can be used to compute cost curves and to provide knowledge
of some quantities underlying supply functions. First, we illustrate the
relationship between production functions and cost functions when a
single resource is variable.
For convenience in computations, we use the production function in
equation 2.68
(2.68) Y = aX b Z d
but suppose that Z is at some fixed level and that aZ d is therefore a

constant equal to c. The single variable production function in equation
2.69 results.
(2.69) Y = cX b
From equation 2.69, we now derive the requirements equation in 2.70,
i x_
(2.70) X = c b Yb
indicating the amount of X required to
produce a specified level of out-
put, Y. Now total cost, T, is the sum K plus the product
of fixed costs
formed by multiplying the per unit price, Px , of the factor by its quan-
tity as in equation 2.71.
(2.71) T = K + Px X
Substituting the value of X

from equation 2.70 for that in equation 2.71,
we obtain the short-run total cost function in equation 2.72
(2.72) T ~ IC + c""
11
Yn P
.
where cost is expressed as a function of output and, for convenience, we
define n = j- . Since average cost is equal to total cost divided by total
output, we divide equation 2.72 by Y to obtain equation 2.73.
1 n n"1
(2.73) A = KY' + c' Y Px
Here A, average cost per unit of output, is expressed as a function of

output level Y. It represents^ a short-run cost function. A will decline
n 21 X
until the magnitude of c Y PX , the variable cost per unit, is greater
1 8
than KY"* , the fixed cost per unit.
From the total cost function we can derive the short-run marginal
cost function, M, marginal cost being the derivative of total cost with
respect to output. Hence, we obtain
(2.74) M = nc'^Y^Px ,
the equation of marginal costs, as the derivative of equation 2.72. 9 Both

equation 2.73, the average cost, and in equation 2.74, the marginal cost
8
An alternative in deriving the short-run cost function is this: From the production
function in equation 2.68, express X as a function of Z and Y as in (a).
_ 1 _c 1
(a) X = a"* Z~S Yb
Now total cost is the sum times their quantities or T= Px X + P z Z and

of prices of factors
average cost is A = (P x X + P z Z) Y"1
or total cost divided by output. Now substituting the
,
value of X from (a) into the total and average cost functions we obtain (b) and (c)
i ci
(b) T = (a b Z b Yb)px + P z Z
- 1
-C 1
-1
E
(c) A = (a z"
B
Y B ")PX + Pz ZY'
K Z is fixed at a particular level while X is variable and we let a"* b z~b p = an(j p 2 m
Z
- K, the average cost equation is (d), with cost per unit expressed solely as a function o f
mYb “1 1
(d) A = + KY’
output Y. This is the same as the average cost equation 2.73 if we suppose that c" n P = m,
9
x
Profit from the fixed factor or plant is maximized when the marginal cost equation 2.74
is equated with marginal revenue. Under the competitive conditions of agriculture, marginal
revenue is identical with average revenue or price. Hence, when the equality in (a) holds
(a) nc" n Y n_:L Px = Py

true, profit is maximized, just as it is for the condition of equation 2.14 where the marginal
product is equated to the factor to product price ratio. As an illustration, we set the mar-
ginal product, the derivative, for equation 2.69 equal to the price ratio (b). Equations (a) m
(b) bcX b_1 = Px /Py
and (b) then state the same thing, except the former is in terms of level of Y while the latter
is in terms of level of X. Substituting the value of Y from equation 2.69 into (a), we
obtain (c)
when X is variable, will depend on the level at which Z is fixed, since

d
the value of c, as indicated for 2.69, is c = aZ . The cost function in
equation 2.72 is a short-run cost function for the particular production
function in equation 2.69 because Z is at the fixed input level specified
earlier. The costs attached to Z at the fixed level represent part of the
fixed costs in K for equation 2.71. K also may include other fixed
costs, such as those attaching to use of any nonzero quantity of Z. A
different short-run cost function, such as that of average cost in equa-
tion 2.73, when X
is variable will exist for each level at which Z is
fixed. Also, as is evidenced by the presence of P x in the three cost
functions above, a different cost curve will exist for each level of price
for the variable factor.
Static Supply Basis
The marginal cost function in equation 2.74 would provide the basis
for the short-run supply function if capital were not limited, uncertainty
of production and price were absent, and the producer otherwise made
decisions under perfect knowledge. Profit is at a maximum, for a
competitive firm as in agriculture, when the marginal cost is equal to
product price. Hence, we set marginal cost from equation 2.74 to equal
product price, P , in equation 2.75.
y
n
(2.75) nc" Yn_1 Px = Py
But since a supply function indicates the amount which will be produced
ateach product price level, the short -rim supply function is more aptly
expressed as in equation 2.76.
~
(n^c ?” )P y ] n 1
11 1
(2.76) Y = [
For a given fixed plant, equation 2.76 shows what output, Y, should be
produced for each level of product or factor price. Derived from a
production function, it would give some notion of the possible algebraic
relation between production conditions in agriculture and short-run
supply functions. But in practice, production coefficients and prices
are variable and uncertain. Too, other restraints and nonprofit goals
are in operation. Hence, the actual response function will seldom cor-
respond exactly, although it is related, to the computed supply function
-n . b n-i
(c) nc (cX ) Px = Vy
b
Px bcX
(d)
Py ' X
which reduces to (d) and is equivalent to (b) . We have derived the equality in (b) from that
in (a).
corresponding to a given production function. In general, because of

uncertainty and discounting, actual response is expected to be less than
that postulated from a curve such as in equation 2.76. A different
short-run "supply basis” will exist ior each level at which Z in equa-
tions 2.68 and 2.69 is set because the magnitude of c in equation 2.76
is affected accordingly. The " supply basis” curve will "flatten” and
"move to the right” as Z increases.
Static Factor Demand
Similarly, a factor demand function, showing the amount of resource

which should be used for each level of factor price, can be derived from
the production function. Derived thus it has the same modifications in
uncertainty restraints and goals mentioned above. Under perfect
knowledge and profit maximization, it would specify the quantity of fac-
tor which should be used at each level of factor price. It can be de-
rived by these steps: From the production function we equate the mar-
ginal product, or derivative of total product with respect to factor input,
with the factor to product price ratio. For the production function in
equation 2.69 this is accomplished by equating its first derivative to the
price ratio as in equation 2.77.
b_l
(2.77) bcX = Px /Py
Dividing both sides of equation 2.77 by be, taking the (b-1) root, and
i
1 1 1 D’1
letting (b~ c’" Py )
= k, we have the factor demand function equation
2.78 where X, input level, is a function of factor price, Px .
(Stated
i
b_1
(2.78) X = kP
otherwise, factor demand is also a function of product price.) Since,

under normal short-run situations diminishing returns hold true, and b
or the elasticity coefficient is less than 1, rpr will be negative. Hence,
D-l
kPxm
^
if we let j— j- = -m, the particular function takes the form X= ,
obviously indicating that as P x , factor price, increases the level of X

will decrease. The factor demand function for this particular algebraic
form of production function will have negative slope, as will hold true
for any production function with an elasticity coefficient less than 1.0.
But, as is suggested later in Chapter 3, the factor demand curve for this
particular function is asymptotic to the P x or factor price axis. In con-
trast, the factor demand curve for the production function in equation
2.11 is
-1 -1 1
(2.79) X = .5c b - (.5c Py )PX ,
a straight line which intersects the factor price or Px axis. The mag-
nitude of the short-run “demand function* for factor again will de- X
pend on the magnitude at which Z in equation 2.68 is fixed. For a
greater fixed level of Z, the derived “demand curve” will be further to
the right.
Long-Run Cost Functions
Long-run cost functions also can be derived from production func-

tions, although the algebraic manipulations are more difficult than for
short-run cost functions. We can use equation 2.68 to illustrate the
First, we express factor requirements for X and Z as in equa-
10
steps.
tions 2.80 and 2.81, respectively. With total cost equal to the sum of
1 d 1
b b
(2.80) X = a Z Y^
i b x
(2.81) Z = a
_d
X_d Y¥
factor prices times factor quantities as in equation 2.82, we can
(2.82) T = Px X + Pz Z
substitute the values of X in equation 2.80 into equation 2.82 to obtain

equation 2.83. If there are fixed costs not attached to quantities of X
i d i
(2.83) T = Px (a’
¥ Y¥ )
+ Pz Z
and Z used, we add K, the sum of such fixed costs, to equation 2.83.
Total cost is now expressed as a function of Z and output, Y. For the
long-run cost curve, we first need to determine the magnitude of Z
which will allow a minimum cost for any specific output, Y. Hence, we
can set Y at the desired levels in equation 2.83, then take the derivative
of total cost with respect to input Z and set it to equal zero as in equa-
tion 2.84.
_i d_ l
-1 b _b_l
(2.84) * -db a" Z Y b Px + Pz = 0
10
For this function, an alternative method might be preferable: First derive the equation
of marginal rate of factor substitution. Set it to equal k, to define an isocline, and solve for
X in terms of Z. Substitute this expression in Z into the production function (2.68) and the
cost function. Now solve for Z in terms of Y. Substitute this expression in into the total Y
cost function just defined. Cost then is expressed alone as a function of output.
With the derivative equated to zero, we can solve for Z, indicating the
quantities of this input which minimize cost for the particular output*
Setting the derivative equal to zero and solving for Z, we obtain
b
f '1 u -f V d+b
1
(2.85) Z = Vd ba P" PZ Y )
which expresses input, Z, of one factor as a function of the price of the

two factors and output. Having obtained this value of Z which mini-
mizes costs for the given output, we can compute the corresponding
magnitude of X. It can be computed directly from equation 2.80, or it
can be computed from the isocline equation for the production function
in equation 2.68 as X = bkd™ Z where we know k, the price ratio of the
1
factors, as well as the magnitude of Z from equation 2.85. Substituting

the quantities of X and Z so computed into the long-run cost equation of
equation 2.82, we obtain the total cost for the specified output, as a
point on the long-run total cost function* If we wish to obtain long-run
average cost, A, it can be computed by dividing equation 2.86 by Y or
output, to obtain
_d j. _1
( 2 86 )
. A = a^P x Z b
Yb + P z ZY
Setting the derivative of A, average costs, with respect to Z equal to

zero and solving for Z, the similar long-run average cost quantities
can be computed by use of equations 2.80 and 2.82.
PRODUCTIVITY COMPARISONS FOR ECONOMIC DEVELOPMENT
The principles outlined above apply to use of production function

quantities in estimating allocations and mixes of resources for profit
maximization of the farm as an individual firm. Other uses also can be
made of production function estimates. Problems in economic develop-
ment might be tackled partly through data derived from production
function studies if these data could be provided in appropriate quanti-
ties and algebraic forms. As an example, we may take an underde-
veloped country which needs to increase food output. It has limited
facilities for producing or importing fertilizer. Of the given quantity of
fertilizer available, what quantities should be allocated to different re-
gions or soil types? If there is a single main crop such as rice and its
supply is extremely short in most regions so that transportation costs
are not important, the problem is approximately that of maximizing
total physical production from the given supply of fertilizer. Hence, if
fertilizer production functions could be estimated for a sufficient num-
ber of soil regions, the general principles outlined above could be used
in establishing the quantity of the resource to allocate to each region.
For a single type of fertilizer, the procedure might be as follows where
production functions are estimated on a per acre basis: the number of

acres in each region would be multiplied times the per acre production
function to give the regional production function. The task is to allo-
cate a quantity of fertilizer to each region so that the marginal physical
products are equal among the n regions. Given knowledge of the pro-
duction functions, we could use the method outlined previously in equa-
tion 2.36. For the particular problem, it now appears as system 2.87
, dYi
where ^7 is the marginal product, derivative of output in respect to
fertilizer input, for the i-th region. This quantity in each region is set
to equal m, so that the marginal productivity of the resource is equal in
the n regions. There are n such differential equations. In addition, an

equation of resource restraints is added, so that the total of input used
in each region, Xi , does not exceed the total available, X Hence, there
.
are n + 1 equations and n + 1 unknowns: n unknowns to represent the

amount of resource to be allocated to each of the n regions and one un-
known to represent the magnitude of marginal product, m, to be attained
in all regions. Arranging the systems of equations into the form AX = S
where A is a matrix of known coefficients from the production functions,
X is a column vector with elements representing the unknowns Xi, and
m and S is a vector of known constants from the production function and
resource supplies, the procedures of equations 2.40 and system 2.41
could be applied to determine the quantity of fertilizer to allocate to
each soil region. An alternative procedure would be one paralleling the
method of equation 2.54 and illustrated in equation 2.88 for the present
Yi + X( 2 Xi - X)
problem. We wish to maximize Y t , total output, from the total re-

sources available. It is the sum of the outputs, Yi from the n regions.
,
But we must add the restraint represented by the last term of equation
n
2.88 where X is a Lagrange multiplier, 2 Xj is the sum of the re-
_ i=i
source used in the n regions and X is the total quantity available. Now
substituting the actual regional production functions Yj = fj(Xi) into

Y
equation 2.88 for the Yi , we take the partial derivatives of t in respect
to all Xj and X and for functions where this derivative exists, equate
them to zero as:
6Yt
= 0
ax x
8Y*
= 0
6X 2
(2.89) :
5Yt
= 0
6X n
6Yt
6\
= 2 Xi - X = 0
i=i
Now, solving as for equation 2.55, we again obtain the quantity of re-
source to be allocated to each region. The fertilization rates per acre
and per farm would then be established accordingly. The resulting
quantities are those which would maximize output from the total re-
source supply.
If two or more resources with predetermined fixed supplies are
available, the same general procedures apply. In applying the method
of equation 2.87 for two nutrients or input categories, we would now
have n differential equations for each nutrient and two equations for
resource restrictions, or a total of 2n + 2 equations. We would also
have 2n + 2 unknowns, with 2n of these representing the amounts of the
two resources to allocate to each region and two representing the un-
known marginal products to be equated among regions for each re-
source. (Since stocks of each are separately predetermined, marginal
products of the two factors need not be equal.)
Some underdeveloped countries have food problems somewhat ap-
proaching that outlined above. However, the allocative problem is not
this simple in others because deficit and surplus regions, transporta-
tion costs of factors and products and manufacturing costs of factors
must be considered — even where food is short and the goal is to pro-
duce a specified quantity of a particular crop. Once the problem is ex-
tended even to include only discrete quantities required for consumption
at different localities, opportunities for production in different regions,
and transportation costs between origin and destination, the problem
can become computationally complex. Models of continuous functions
similar to those for equations 2.36 and 2.54 can be outlined, to solve for
quantities to be produced and consumed and resources to be used in
each region if costs are to be minimized. But linear programming or
transportation models, using ^discrete” quantities derived from pro-
duction function estimates, may provide ample refinement. These
models would be set up with restrictions requiring that consumption

requirements be “exactly met” in each region and that planted acreage
be equal to or less than acreage available in each region. Different
crop producing activities would be used in each region to represent dif-
ferent mixes or levels of fertilizer which could be used and fertilizer
purchasing activities might be added for each region.
Productivity Comparisons
Estimates of firm production functions usually are used for diag-

nostic purposes, in judging the degree of factor disequilibrium under
different degrees of economic and market development. If appropriate
samples and algebraic models could be devised, these functions also
could be used to estimate equilibrium quantities and proportions of re-
sources. However, practically all such studies have employed Cobb-
Douglas forms of functions. As outlined in the next chapter, functions
of this algebraic form which assume constant elasticity, generally
overestimate equilibrium input quantities under the criterion of profit
maximization for a firm. Consequently, firm production functions
based on the power -type equation have been used mainly to estimate
marginal value productivities for mean inputs of resources based on
farm samples. Marginal value productivities computed at the mean
then may be used to indicate whether disequilibrium in resource use is
great or small. One criterion is the magnitude of the marginal value
productivity of a unit of a particular resource as compared to the per
unit price of the resource. Since farm production functions typically
are derived with value rather than physical output representing the de-
pendent variable, the comparison is made simply as follows: all other
resources are set at a magnitude equal to their mean. The marginal
value product of the resource being studied, computed as the derivative
of output in respect to input, is then derived. For example, suppose we
have the production function in equation 2.90 where V is value of output
and Xi represents different resources such as labor, land, and capital
of different types. With Xi considered to be the variable resource in
question, its marginal product at its mean is that in equation 2.91 where
b bo b« b 4a
(2.90) V = aX 4 x2 2 x3 3 x4
6V bi ,
__bi _b 2 bs b4 bi
(2.91) = (aXj.
l X* X, Xj )
= -± V
5Xi Xi Xi
the bar indicates inputs at mean levels. With V computed from equation
2.90 when all factors are fixed at their sample mean, we multiply this
quantity times bj, the coefficient of Xi, and divide by Xi , to obtain the
marginal product of the factor, Xi, when input is at mean level. In turn,
these marginal productivities could be compared with the respective
factor prices to suggest the degree of disequilibrium in resource use.
Also, from the computed marginal value product, we can derive the
quantity, with other inputs at mean necessary to cause produc-
level,
tivity to equal factor price. For equation 2.91, this is accomplished by
setting the derivative or marginal product to equal the factor price Pi ,
and solving for Xi where the latter is not at mean level. The magnitude
of Xi the input of the particular variable factor to equate marginal
,
value productivity and price when other resource inputs are at mean
magnitude, is that of equation 2.92.
1
(2.92) Xi = biVPf
Of course, if appropriate procedures could be employed in sampling,

resource classification, and aggregation and algebraic form of equation,
the equilibrium quantities of all resources could be specified. This
would be accomplished by procedures outlined previously; namely, by
taking derivatives such as equation 2.91 for all factors, equating them
to their respective prices and then solving the system of simultaneous
equations for the value of the Xj
Or, if it is supposed that the allocative problem is more nearly one
of the distribution of resources within the agricultural economy rather
than in relation to the total economy as reflected in factor and product
markets, these marginal value productivities might be computed for
different samples of farms, representing different sectors in respect to
tenure, soil, climate, location, or product specialization. They then
might be compared, to determine the degree of resource imbalance
among the sectors being compared. If interest were in maximizing the
value of product from the collection of resources already existing in
agriculture, the differences would suggest the relative degree of re-
source adjustment required. If the data and empirical method were ap-
propriate, a system of equations could be derived, in which the mar-
ginal value product, m^, of the i-th resource is equal for all regions.
We could then solve for the magnitude of the i-th resource to be allo-
cated to each region in a manner paralleling that for equation 2.87. If
we simply wish to know by how much a specific mobile resource such
as labor for one farm sector should be increased or decreased to cause
its marginal productivity to equal that of the same resource in another
sector, we could use the following procedure. First, compute the mar-
ginal value product of the factor for each region, as in equation 2.91
where we suppose that other factors are fixed in quantity. Then equate
these two marginal value products as in equation 2.93 where subscript i
buVi b i2 V 2
_
(2.93)
refers to the i-th resource and subscripts 1 and 2 refer to two regions.
Input in region 2 is held at mean level while input in region 1 is varia-
ble. We wish to determine the magnitude of ijL ,
X
if the equality of equa-
tion 2.92 is to hold true. The value is that of equation 2.94, derived
kji ^1^x2
(2.94) Xi, =
from equation 2.93, indicating the input or X ix level if the marginal

value productivity of the factor in region 1 is to equal the marginal
value productivity of the same factor in region 2, when all other inputs
are at mean value and the particular factor also is at mean level in re-
gion 2. But this procedure is useful only where the form of function
employed, the specification of input categories and sampling variance
do not lead to serious errors for estimates of productivities from non-
mean quantities of inputs. Farm production functions of the Cobb-
Douglas type derived under usual circumstances are not very appropri-
ate for these purposes.
SCALE RETURNS AND PRODUCT IMPUTATION
Farm or firm production functions also may be computed to deter-

mine the magnitude of returns to scale. Scale returns are measured by
the ratio, per cent increase in output divided by per cent increase in in-
put, under the condition that all factors be increased by the same pro-
portion. For this measure to be applicable, there must not be omission
of important input categories or any serious aggregation problems.
For the Cobb -Douglas function, the elasticities of the individual fac-
tors are their exponents in the production function or regression equa-
tion. These individual elasticities can be summed to determine the
scale coefficient, e , indicating the percentage by which output will be
increased if all factors are increased by 1 per cent. For a function
such as equation 2.90, constant returns to scale hold true if b x + b 2 + b 3
+ b 4 = 1.0. A 1 per cent increase in input of all factors will cause output
n
to increase by 1 per cent. If 2 b*< 1.0, decreasing returns to scale
i=i
n
hold true while for 2 > 1.0, increasing returns to scale hold true.
i=i
Thus for equation 2.68, suppose that b = 0.3 and d = 0.4. We increase
input of each factor by the proportion R and the function becomes
A
(2.95) Y = a(RX) *® (RZ )
which by the rule of exponents is
K4 ZA aR^^X* 3 Z"4
-3 3
(2.96) Y = aR X* or Y =
7
and output only increases by the proportion of R' , rather than by R as
in the case of factors, because the sum of individual elasticities is only
0.7. Had the individual elasticities been b = 0.2 and d = 0.8, then output
would have increased by the proportion R, as input of both factors is
increased by this proportion. As indicated in the next chapter, other

forms of functions must be employed to estimate scale returns when
elasticity coefficients are not constant over all ranges of inputs. How-
ever, the general statements on scale returns and their measurement
hold true for all algebraic forms of production functions.
Imputation
Estimation of production functions by Douglas and his associates

centered around imputation problems; the division of a total output
among the factors used in producing it. Agricultural economists long
have divided total outputs in this manner, although somewhat arbitrarily
and without knowledge of the underlying production functions. Examples
present themselves in land appraisal where net income imputed to land
is capitalized into a land value and division of income between landlord
and tenant, on the basis of the share of inputs contributed by each.
Hence, this section is devoted to a brief summary of the conditions al-
lowing imputation of a product among the factors producing it, so as to
just exactly exhaust the product.
Of course, a total product can be allocated to its factors in an infi-
nite number of ways, to just exhaust the total output. But the econo-
mists’ concern has been in allocating the product in proportion to the
marginal products and the quantities of resources. It can be shown that
under competitive conditions an allocation of factors rewarded in this
manner maximizes output from resources available. Hence, we exam-
ine the conditions under which an imputational procedure based on mar-
ginal resource productivities can be accomplished. Again, using the
production function in equation 2.68, we determine the magnitude of out-
put necessary to reward the two factors according to their marginal
productivities. This is determined by computing the marginal products
of the two factors from equation 2.68 and multiplying them by the quan-
tities, X and Z, of the two resources, respectively, as in expression
2.97. Multiplying out these products, we obtain the sum in expression
2.98, which can be simplified to (b + d)aX b Z d Substituting the Y value
.
b"1 d
(2.97) (baX Z )X + (daX b Z d_1 )Z
(2.98) baX b Z d + daX b Z d
from the production function, equation 2.68 in this case, this becomes
(b + d)Y. Hence, if the sum of b and d is less than 1 for equation 2.68,
then (b + d)Y is less than Y, the total product, and the quantity imputed
to each factor in expression 2.97 will not exhaust output. If b + d is
greater than 1.0, the quantity of expression 2.97 will be greater than y,
and the amount of product is in deficit relative to the amount allocated
to the factors. With b + d = 1, then the quantity in expression 2.97 is
equal to Y and the marginal productivity method of rewarding factors
just exhausts the total product. While a power function has been used
to illustrate these conditions, the same principles apply to other forms
of functions.
CHOICE UNDER UNCERTAINTY
The general principles discussed thus far assume the parameters of

production functions and prices of products to be known with certainty.
This condition, of course, is unreal: variance attaches to the estimates
of productivity coefficients even for a single year and a given weather
condition. Weather itself varies in different years; and prices, espe-
cially those of products, are uncertain. Various decision-making prin-
ciples can be outlined for these situations and range from informal pre-
cautionary measures to the more formalized mathematical models of
game theory and decision models. No attempt will be made to outline
these principles, however, because an entire treatise would be required
11
for this purpose. Explanations of them are given elsewhere Prob- .
lems of variance quantities related to estimating productivity coeffi-

cients are discussed in later chapters.
Presence of uncertainty does not eliminate the need for use of ap-
propriate and correct economic principle. Physical scientists have
long made recommendations as if price and production parameters
were known with certainty. Examples are recommendations to ferti-
lize at particular levels or to use specific mixes or rations or feeds
for livestock. These recommendations not only suppose that the rele-
vant quantities are known without error, but also that all farmers have
the same profit goals, capital availability, or tenure situation. Too,
economists have used crude average productivity comparisons drawn
from census or farm sample summaries, to suggest resource disequi-
librium or allocation needs. The data are applied within the framework
of principles as if the quantities were known with certainty. Quite ob-
viously, problems of uncertainty are of no less magnitude when "rule
of thumb” principles and discrete or point estimates of input to output
ratios are used, as compared to correct economic principle and formal
knowledge of production functions. The purpose of this chapter has
been to outline and summarize the correct economic principles which
can be applied to production function estimates.
As pointed out previously, the cost per unit of knowledge for physi-
cal phenomenon is probably no greater, and often less, when complete
production functions are estimated, as compared to a similar number
of treatments directed toward a few point estimates. Presence of un-
certainty perhaps best applies to the degree of refinement needed in
11
See Heady, Earl O. and Candler, Wilfred V. Linear programming methods. Iowa State
University Press, Ames. 1958. Pp. 499-527; Luce, R. D. and Raiffa, H. Games and deci-
sions. John Wiley and Sons, New York. 1957. Pp. 278-308; Dillon, J. L. Theoretical and
empirical approaches to program selection within the feeder cattle enterprise. Journal of
Farm Economics. 40:1921-31. 1958.
estimating the production functions and in applying the relevant princi-

ples. Because price and productivity coefficients are uncertain, a pre-
cise definition of input equilibrium is impossible. Hence, once a rea-
sonably accurate estimate of the production function has been
completed, a sufficient number of points may be selected from it, with
the appropriate principle applied accordingly. In this case, we may
simply compute average marginal products between discrete inputs as
—
AY
where AY is increase in total output and AX* is the increase in
between the two discrete input levels. Input is then in-

total input,
creased as long as the marginal product so defined is greater than the
price ratio, or
AY p*
2 99 )
. ^
(
AXj Py •
Similarly, the substitution rate between two factors X and Z may be de -

fined as ~~ between two input levels of Z. Input of Z then should be
substituted for X as long as the substitution ratio is greater than the
price ratio, or
2 100 )
.
AX . Pz
'
(
AZ Px
Similar modifications can be made for other principles of choice out-

lined on previous pages. The only modification is substitution of the
marginal ratios (as averages between discrete input levels), and

^^for the respective derivatives and The choice then is be-
tween the discrete input levels and ratios relative to which the marginal
ratios are defined. If extended to sufficiently complex choice or alloca-
tion problems, programming procedures are probably most convenient
for specifying input and output mixes and magnitudes. If the relevant
quantities can be provided in sufficient detail through experiments and
samples designed to provide point estimates, there is no reason why
they should not be used as substitutes for continuous functions.
CHAPTER 3
Forms of Production Functions
N UMEROUS ALGEBRAIC equation forms can be used in deriving

production functions. Some were reviewed in Chapter 1. No sin-
gle form can be used to characterize agricultural production
under all environmental conditions. The algebraic form of the function
and the magnitudes of its coefficients will vary with soil, climate, type
and variety of crop or livestock, resources being varied, state of
mechanization, magnitude of other inputs in "fixed quantity” for the
firm, etc. Hence, a problem in each study is selection of an algebraic
form of function which appears or is known to be consistent with the
phenomena under investigation. Guides on appropriate algebraic forms
may come from previous investigations and the theories of the sciences
involved. Selection of any specific type of equation to express produc-
tion phenomena automatically imposes certain restraints or assump-
tions in respect to the relationships involved and the optimum resource
quantities which will be specified. However, some equations are more
flexible than others. To illustrate this fact, algebraic evaluation is
made in this chapter of a few general types of production functions.
Suggestions also are made of phenomena where these might be empiri-
cally appropriate. An infinite number of functional forms are possible
in productivity studies but those considered in following sections either
(a) have logical implications which cause them to "stand out” from the
others or (b) have been widely used in production function studies.
Some of the equations examined have not had wide application because
their parameters or coefficients are difficult to derive in statistical
treatment of data. Some have terms which cannot be transformed, or
are not readily transformed into linear regression equations and,
hence, can be estimated only by iterative processes. However, a few
of these are examined to illustrate the range in technical coefficients
and conditions which are allowed by different types of functions.
SINGLE VARIABLE EQUATIONS
Numerous research studies in agriculture revolve around produc-

tion functions with a single resource or treatment applied at different
levels. Different levels of applying nitrogen or insecticides to crops
73
74 FORMS OF PRODUCTION FUNCTIONS
are examples. Hence, we will use production functions of this nature
as a step toward a more general analysis of functional forms. Equa-
tions with a single variable defining the input can be used to evaluate
certain properties of production functions and their marginal products.
These same properties generally apply when there are n variables.
Obviously, however, output is never greater than zero when only one
production factor is used. Appropriately, the production function should
be represented as
(3.1) Y = fpq, Xa ,
... Xn )
where Y is the output and the X* are the inputs. In general, given the
existence of the production function, the following quantities can be de-
rived and are of direct importance in economic application:
5Y
(3.2) = f'
Xi (Xi,
X 29 •• xn )
SXi
fo
SXi fL(x x x a, ,
••• Xn )
"
6x
j
"
% (x,, x2 ,
••• xn )
(3.4) x i
= f"(Y, X 15 ... Xn )
SXi
(3.5) = -k
SXj
SX,
(3.6) = 0
“i
These physical quantities represent the data necessary for economic
application of results from production functions. We wish to evaluate
the characteristics of these relationships as they are derived from
various functional forms. In review, they are as follows: equation 3.2
is the equation of marginal physical products for the i-th resource,
equation 3.3 is the equation of marginal substitution rates between the
i-th and j-th resources, equation 3.4 is the equation of isoquants, equa-
tion 3.5 is the equation of isoclines, and equation 3.6 is the equation of
ridgelines. If equation 3.3 is substituted into equations 3.5 and 3.6, it
is obvious that any of these quantities for one resource depend on the
entire range of resources which can be used in the particular produc-
tion process. It is with this qualification that we first turn to evalua-
tion of production functions which involve a single variable resource.
It is, of course, possible to hold certain categories of inputs at
fixed levels while others are variable. However, certain resources or
input categories are exogenous in the sense that they are the “result of
outside forces," and are not subject to control by the decision-maker or
research worker. Thus where only resources X , X 62 , * . X 0 can be .
i.
g
,
FORMS OF PRODUCTION FUNCTIONS 75
controlled in magnitude, then a random disturbance or effect also is

associated with the variables Xg +1 , X g+2 • •
X n . Expressing this
,
•
random disturbance as e , the production function, without control of

some factors, is
(3.7) Y = g(X x , X2 ,
... Xg ) + € .
We may use a vertical bar as below to indicate that only one factor, X*
is variable while others are held fixed at some predetermined level.
(3.8) Y = 0(Xij • • •
Xg )
+ e
Similar conventions could be used for two or three variable factors.

However, to simplify presentation, and with the above recognition of the
existence of variance or random disturbances which attach to predic-
tions, we will present functions in the form Y = f(X) where a single -
variable production function is under consideration, or Y = f(X1? X 2 )
where two factors are variable in quantity while others are in fixed
magnitude or are exogenous. Too, while there are few cases where a
single input category is variable, we employ the convention used in
some agricultural experiments and suppose that it is possible. It may
be possible to add labor in a production process, without varying the
quantity of other production factors. However, even a simple capital re-
source such as nitrogenous fertilizer cannot be varied, without simul-
taneous variation in labor or other capital inputs. But regardless of
this, useful interpretation can be made from a single input equation of
the algebraic assumptions implicit in the use of various types of equa-
tions for estimating production functions.
Cobb -Douglas
The Cobb-Douglas or power function, in the form generally used, is
(3.9) Y = aX b
where X is the variable resource measured, Y is output, a is a constant,

and b defines the transformation ratio when X is at different magni-
tudes. The exponent or b coefficient is the elasticity of production and
can be used directly. (The equation is estimated in logarithmic form.)
This function allows either constant, increasing, or decreasing marginal
productivity. It does not allow an input -output curve embracing all
three. With all other inputs held in fixed magnitude, the marginal prod-
uct is expected to decline. The marginal product equation is
dY baX
(3.10) baX b-x _
dX X
indicating that if b = 1, the marginal product (and also the average
product) will be constant at the level a. Where b > 1, the magnitude of
marginal products will increase as X increases, depending on the mag-
nitude of b. If b = 2, for example, the marginal products are ba, 2ba,
3ba, and 4ba when Xhas the respective values 1, 2, 3, and 4. Where
b < 1, the magnitude of marginal products will decline as X increases
since Xb <
X.
This function assumes a constant elasticity of production, E p over
,
the entire input-output curve, or that
(3.H)
£*,5^
dX1 Yj
= ^
dXj
Xj_
Y2
= '" =
dYn
dXn
Xn
Yn
where the subscripts refer to marginal products and total outputs cor-
responding to various magnitudes of X. This condition of the equation,
that successive equal increments of input add the same percentage to
total output, can be proved by multiplying the derivative or marginal
product equation 3.10 by the inverse of the average product (the defi-
nition of production elasticity) as shown below.
b_1
(3.12) Ep = (baX =
)
| •
|
Now substituting the value of Y of equation 3.9 into equation 3.12 we ob-
tain
(3.13)
, _ bY X
P X ‘
Y
and since the Y’s and X’s cancel, we have E = b, or the elasticity is a
p
constant equal to the exponent of X in equation 3.9.
Given the mathematical properties of the equation, this function
cannot be used satisfactorily for data where there are ranges of both
increasing and decreasing marginal productivity. Neither can the func-
tion be used satisfactorily for data which might have both positive and
negative marginal products. Also, since the rate of decline in the mar-
ginal product decreases with input magnitude, the power function pro-
vides a curve of the nature indicated in Figure 3.1. The curve “flattens
out” as input increases and a maximum product is not defined. Unless
an economic optimum is defined for small magnitudes of input, the
power function especially may overestimate the input of X which
equates marginal revenue and marginal cost.
When Y measures total output, the equation assumes the resource
and that output is zero when input is of zero magni-
to be limitational
tude. With Y measuring yield added by the variable resource, the fac-
tor is not assumed to be limitational.
Figure 3.1. Graphic illustration of function computed from equation

log Y = log a + b log X.
Spillman Function
The exponential -type function suggested by Spillman will be evalu-

ated in the form of equation 3.14
(3.14) Y = M - AR X
where Y again measures total output and X measures total input. The
other coefficients have this meaning: M
is the maximum total output
which can be attained by use of the variable resource, A is the total in-
crease in output which can be attained by increasing X, - A is the M
level of output defined by fixed resources and a zero input of the varia-
ble resource, R is a constant defining the ratio of successive incre-
ments to total product. Hence, R also defines the magnitude of the
marginal product of input level Xj relative to that of input level Xj_ x .
From equation 3.14 the equation of marginal products is
X
(3.15) -g = -AR log e R
and the marginal product of the i-th input bears the following relation-
ship to the marginal product of the previous input:
(3.16) - *
dXj vdXj.j/
In equation 3.16, the subscripts on the terms of the derivatives again
define levels of input of resource X
and the marginal products associ-
ated with each. Hence if R = 0.8 and the marginal product for i _ 1 is X
10, themarginal product of X t is (.8) (10) = 8.
The total product curve of equation 3.14 is asymptotic to M. The
marginal product curve is asymptotic to the zero axis, never becoming
negative as might be the case of fertilizer used in excess. For this
reason, the function is not appropriate for samples drawn from experi-
ments or surveys where input magnitudes are great enough to cause a
decline in total product.
The Spillman -type curve in Figure 3.2 illustrates these properties.
The increments in product due to increments in inputs are indicated as
A i . The product curve approaches the maximum as X increases and M
AR X decreases. If used as a response function, the equation is
X
(3.17) Y = A(1 - R )
and the input axis has origin at M

- A in Figure 3.2. Similarly, yield
then becomes asymptotic to A, rather than M.
Quadratic Forms
The simple quadratic equation in equation 3.18, with a minus before

c to denote diminishing marginal returns, does not impose such strict
restraints on the production function as does the Cobb-Douglas or
(3.18) Y = a + bX - cX2
Spillman equation. It allows both a declining and negative marginal

productivity, but not both increasing and decreasing marginal products.
A maximum total product is defined where input magnitude or is equal X
1
to .Sbc” . The elasticity is not constant as in the power function, but
declines with input magnitude as indicated by the elasticity equation:
2
bX - 2cX
(3.19) 2 ‘
P " a + bX - cX
The marginal products do not bear a fixed ratio to each other as in the
case of the Spillman function. However, as indicated by equation 3.20,
(3.20) ^- = b - 2cX
use of the quadratic equation does assume a particular characteristic

in relationship between marginal products; namely, that they decline by
a constant absolute amount or mj = m^_ 1 - k where m^ is the marginal
product of the i-th input, m
i_ l
is the marginal product of the i-1 input
FORMS OF PRODUCTION FUNCTIONS
Figure 3.2. Nature of Spillman-type exponential function.
and k (k = 2c) is the constant by which successive marginal products

decline. In other words, the marginal product curve is linear (the sec-
ond derivative being a constant equal to -2c) and the total product forms
a “mirror" curve, with the portion on the right of the maximum being
the “mirror" of the portion on the left.
The constant a in equation 3.18 represents the product forthcoming
from the mix of fixed resources. If Y is used to measure response or
output due to the variable resource only, we expect that a is zero.
This is also true for other polynomial forms where the exponent or de-
gree of X
differs from 1 and 2.
Modifications of the polynomial equation in equation 3.18 may be
used to relax the restraint that marginal products decline by a constant
amount. For example, using the cube rather than the square for in X
equation 3.18 will cause the marginal productivity of the variable re-
source to decline at an increasing rate (i.e., the second derivative is
-6cX).
6
(3.21) Y = a - bX + cX* ,
This square root equation 3.21 provides a simple compromise be-

tween the power or exponential equations and the quadratic form in
equation 3.18. Supposing that all express diminishing marginal produc-
tivity, equations 3.9 and 3.14 have marginal products which decline at a
diminishing rate, but never allow a declining total product. The equa-
tion 3.18 allows a diminishing total product but assumes a constant de-
cline in marginal products. The square root equation 3.21 allows a di-
minishing total product but also, as equation 3.22 indicates, has
marginal products which decline at a diminishing rate. This particular
(3.22)
H = .5cX"
-5
- b
relationship may be found under certain biological conditions. The

marginal product may be great at low input levels, and decline at
slower rates as input level increases; although a sufficiently large dos-
age may depress total yields. This equation has a maximum total prod-
2 2
uct when X equals .25c b . Its elasticity declines with increases in
input and output magnitudes.
The appropriate functional form for estimating production relation-
ships will, of course, depend on the type of phenomena being examined.
Ordinarily, those represented by crop response to fertilizer will have
diminishing marginal products for all inputs greater than zero. But
there are isolated instances where increasing marginal products have
been encountered for initial inputs of nutrients. Production functions
relating feed input to live weight output of livestock evidently have di-
minishing marginal products for all input levels. A function allowing
some range marginal products may be needed, however,
of increasing
in estimating output of edible dressed products. In meat production,
one does not expect negative marginal product with age of animal and
total feed input during the growing and fattening period and may select
an equation accordingly. Under many circumstances, except for potash
or the single variable under certain soil and climatic conditions, a
function allowing negative marginal products would be needed for ferti-
lizer application. However, under normal conditions, one would never
expect a single input category such as labor or machine capital to lead
to a depressed total output where an entire farm represents the fixed
plant for which the production function is being derived.
Resistance Formula
A modified statement of the equation proposed by Balmukand, based

on MaskelFs “resistance formula," is expressed in equation 3.23. 1
1
Balmukand, B. Studies in crop variation. V. The relation between yield and soil
nutrients. Journal of Agricultural Science, 18: 602-27. 1928.
(3.23) Y- 1 = a(b + X)’ 1 + c
Here a, b, and c are constants and, in the case of a fertilizer function,

b might be the amount of nutrient in the soil while X is the amount
added. This is an asymptotic function with a maximum yield of c” 1 and
supposes only positive marginal products, as suggested by the marginal
product equation 3.24 which has an asymptote of zero. This equation,
2
= a[a + c(b + X)]"
(3.24) -^
like that of Spillman and Mitscherlich, assumes that different input fac-
tors have effects independent of each other but assumes a marginal re-
lationship such that there are differences among successive responses.
Hyperbolic and Other Functions
Numerous forms of hyperbolic equations have been proposed to de-

scribe crop -fertilizer production functions. Most of these are more
difficult to fit empirically than quadratic equations, although they
possess somewhat similar properties. For example, hyperbolic equa-
tion 3.25 has a marginal product curve similar in nature to that of the
(3.25) Y = aX(b + X)'

1
- cX
square root equation in 3.21. It has the marginal product equation

shown in equation 3.26.
(3.26) g- = ab(b + X)“

2
- c.
The marginal product declines at a decreasing rate with a maximum

-1 -
total product at the input level where X equals V abc b.
The hyperbolic equation suggested by Thilau in equation 3.27 has
the nonlinear marginal product equation 3.28.
(3.27) Y = Y aX + bX 2
dY _ a + 2bX
(3.28)
^ 2/aX + bX 2
It has a marginal product which becomes zero at an X value of -

2b
Increasing Marginal Productivity
A production function embracing both increasing and decreasing

marginal productivity is seldom needed for a single variable resource.
Where needed, a function allowing this condition is the polynomial
3
(3.29) Y = a + bX + cX2 - dX .
1
Ithas increasing marginal products until X is equal to .3333cd ,
then
diminishing but positive marginal products until
_1 2 5
(3.30) X = .3333d [c + (3db + c )’
]
which is the value of X that maximizes total product. Supposing a to be

1
zero, the elasticity is greater than 1 (stage 1) for 0 < X < .Scd" , equal
to 1 (constant returns) at X = .Scd” , and less than 1 but greater than
1
1
zero (stage 2) for X > .Scd"* . Marginal products decrease at an in-
creasing rate in the later stage.
Several other types of quadratic equations allow these various
stages. Other algebraic forms serve similarly. One proposed by A. N.
2
Halter, et al ,, is somewhat a hybrid or combination of the power and
exponential equations. It is
(3.31) Y=cXa ebX

where Y and X measure output and
input respectively, a, b, and c are
coefficients to be estimated and e is the base of natural logarithms.
5
This function has increasing marginal products up to X = b” 1 (-a t a’ ),
the input defining the point of inflection on the input -output curve. It
1
then has declining but positive marginal products to X = -ab" the input ,
defining the maximum total product. The elasticity of production is 1.0

at X= a)b“ 1 . Hence, stage 1, with marginal products greater than
(1 -
average products, extends over the input range 0 < X < (1 - ajb’' 1 and
stage 2, with marginal products less than average products but greater
than zero, extends over the range (1 - ajb” 1 < X < -ab” 1 . The logistic
function in equation 3.32 has both increasing and decreasing marginal
productivity of the variable resource X. The constants a, b, and c
a
(3.32) -cX
1 + be
are parameters to be estimated from sample observations and e is the

base of natural logarithms. Again, this is an asymptotic type curve,
with upper asymptote at a, and would not be appropriate for observa-
tions characterizing a diminishing total product. The corresponding
marginal product equation is equation 3.33.
A note on the transcendental production function. Jour. Farm Econ., 39: 966-74. 1957.
abce~
cX cX )^2
(3.33) = (l + be‘
Marginal product is a maximum where X is

1
equal to c" log b, the in-
flection point of the total product curve. It has changing elasticity, with
the elasticity coefficient equal to unity at
(3.34) X = b“ l c“ l (b + e cX ) .
For values of X less than this, production is in stage 1, with an elas-

production greater than 1; for all values of X greater than this,
ticity of
production is in stage 2 since the elasticity coefficient never becomes
negative.
n RESOURCES
Where parameters are to be estimated for more than one variable

resource, selection of an equation consistent with known empirical re-
lationships or relevant theories also is involved. As in the case of a
single resource, selection of a particular equation implies some as-
sumption about the algebraic nature of the production process. To fur-
ther illustrate the types of assumptions or restraints imposed by differ-
ent functions, we evaluate a few of the more commonly used equations
in this section. The same algebraic restraints exist for n variables,
but we restrict the analysis to two input variables, X L and X 2 , for pur-
poses of simplicity.
Power Functions
The power function in equation 3.35 has the same mathematical

characteristics discussed for equation 3.9, for either resource, when
input-output curves are derived for one resource with the other held
1 2
(3.35) Y = aX^ X^
constant. The assumptions of constant elasticity and marginal products

with only a plus or minus sign, regardless of input or output magni-
tudes, are retained. Again, the regression coefficients, b x and b 2 de-
rived with the observations in logarithms, are the production elastici-
ties of the individual resources. Their sum indicates the nature of
X
returns to scale, provided x and X 2 are the only relevant inputs. With
the sum b x + b 2 = 1, a given percentage increase in both inputs will re-
sult in an equal percentage increase in output. With elasticity sums
being more or less than 1, output will increase by a greater or smaller
percentage, respectively, than inputs.
If the sum of elasticities is less than 1, the power function provides
a surface which has no distinct peak. It slopes down to the input axes
and tends to form a broad ridge in output space. It never forms a prec-
ipice over an input axis, as would be the case where a factor is nonlim-
itational. Because the isoclines "fan out” in linear fashion from the
origin, the point of the wedge representing the surface comes at zero
input of both factors, with the wedge broadening along the input axes as
the surface rises to greater heights. As mentioned for the one factor
input -output line, the surface has only positive slopes when the elastic-
ities are greater than zero (and only a negative slope when they are
less than zero).
The isoquants of this production function are asymptotic to the input
axes, as suggested in isoquant equation 3.36 where X x is expressed as a
(3.36)
function of X 2 . Any one isoquant can be derived from another simply by

multiplying the input quantities by the appropriate ratio. Because of
this property, and as suggested in equation 3.35, each resource serves
as a limitational input: No output is forthcoming if either Xj or X 2 is
zero. The marginal rate of substitution is, as indicated in equation
3.37, a linear function of the ratio in which X x and X 2 are combined. If
(3.37) El. '

VEl
8Xz
inputs of Xx and X 2 are increased in constant proportions, the marginal

rate of substitution remains constant at the ratio b2 /b 1 , even though
level of output changes. This condition is unrealistic for two classes of
inputs such as carbohydrate and protein feeds for a growing and fatten-
ing animal. At earlier ages, it is known that the marginal rate of sub-
stitution of high protein for high carbohydrate feeds is highest when the
animal is young, and declines as the animal ages and the growing stage
merges into the fattening stage. Hence, the marginal rate of substitu-
tion will not remain constant as inputs of two feeds are increased in
constant proportion (a specific ration) as the animal gains in weight.
Evidence also suggests that the marginal rate of nutrient substitution
declines with yield level on a given land area. However, it is entirely
possible that rates of factor substitution may remain constant as all in-
puts of the production process are increased in the same proportion for
a firm. While the power function may characterize the production
process for certain firm conditions, it is less appropriate for a fixed
producing unit represented by an acre, animal, or bird.
The isocline equation, derived from substitution rate equation 3.37
by setting the latter derivative to equal a constant -k (to represent a
given marginal rate of substitution or price ratio) and solving for one
input quantity in terms of the other, is:
Figure 3.3. Geometric nature of isoquants and isoclines

for Cobb-Douglas function.
(3.38) Xx = b^kX .
The quantity of Xx thus expressed as a function of X2 . Hence equa-

is
tion 3,38 indicates the quantity of Xx necessary to provide a rate of sub-
stitution of k magnitude when X 2 is given various values. Again, this is
a linear equation, indicating that the isoclines are straight lines passing
through the origin. A map indicating the geometric nature of isoquants
and isoclines “forced” by the power function is included in Figure 3.3.
Since the isoclines are straight lines passing through the origin, they
also are scale lines, indicating a fixed proportion or mix of the two in-
puts used at different levels. Because of these characteristics, the
power function denotes that the ratio in which the two resources are
combined should remain the same regardless of the level of output. The
optimum magnitude of input and output changes as the price of the prod-
uct changes relative to the price of the inputs, but the optimum input
ratio does not change if the factor price ratio remains constant. The
optimum ratio of factors does change, however, as the factor price
ratio changes. Ridgelines for this function do not converge at a factor
combination defining maximum output. (No maximum is defined by the
function.) In fact, the ridgelines, denoting zero rates of factor substi-
tution, are identical with the input axes in a two -dimension contour map
of the surface.
The power function in equation 3.35 can be modified to relax the re-
straints of constant elasticity over the surface and isoclines that denote
the same ratio of factor inputs regardless of output level. One modifi-
cation is accomplished in equation 3.39 by adding a constant to each in-
put magnitude.
In logarithms, this function does not reduce to a linear equation.
Hence, the added constants may be estimated, with numerous iterations
used to determine which magnitudes give smallest deviations from re-
gression. The corresponding equation of marginal substitution rates
bl
(3.39) Y = a(h +
x
X L)
(h 2 + x/ 2
can be defined as in equation 3.40. When equated to -k, equation 3.40
5X,
=
b2 (h x + XJ
(3.40)
6X 2 (h 2 + X2 )
allows derivation of the isocline equation in equation 3.41. In this iso-

cline equation, X x is still a linear function of Xjj , with the slope being
(3.41) Xt = kb^j'hg - h1 + kb^X.,

1
kbjL b^T However, the isoclines do not pass through the origin but
.
through the X x axis if kb 1 b 2 h2 - h x is positive and through the Xg axis

1
if itis negative. Under this modification, the optimum ratio of factors

changes as different levels of output are attained.
Since the elasticity coefficients are constants in the Cobb -Douglas
production function, the scale returns so indicated simply represent the
“average condition” for the sample. If we sum the exponents of the re-
source variables (regression coefficients with variables measured in
n
logarithms) and the quantity S bi is less than 1.0, we can only say
i=i
that, as an average for the sample, decreasing returns to scale hold

true provided no relevant inputs are excluded as discussed in Chapter 6.
It is possible that increasing returns to scale may hold true for a mod-
est range of inputs, but that the portion of the sample drawn from input
ranges characterized by decreasing returns dominates. Or the opposite
may hold true, with observations from the range of increasing returns
dominating those from the range of decreasing returns.
Spillman Function
In two variables the Mitscherlich -Spillman response function can be

evaluated in the form of equation 3.42 where Y is output resulting from
the resources being varied, X is the quantity of one factor and Z is
( 3 . 42 ) Y = A(1 - R*)(l - Rf)

QUANTITY OF X
Figure 3.4. Production surface for Spillman or
Mitscherlich function.
quantity of another factor. R x indicates the ratio by which marginal

products of X decline while R z has the same meaning for Z. In this
case, A is the maximum response from increasing both factors. The
surface has height asymptotic at level A and extends to a long broad
“ridge” over the input plane, rather than to a peak at a single point
over the input plane. If input of either or both factors is zero, output
also is zero. An example of the function is shown in Figure 3.4. The
isoquant equation is equation 3.43, indicating that the isoquants are
~X
(3.43) X = log [l - —^—2-1 (log Rx )
L -
A(1 R~)J
asymptotic to the axes. Hence, one factor can never substitute com-
pletely for the other.The isoquants do not, however, maintain a
constant slope (substitution rate) at points of intersection with a scale
line extending from the origin. A set of isoquants, the negative sloped
curves, illustrating their nature for the Spillman function, is included
in Figure 3,5.
The equation of marginal rates of substitution is equation 3.44. Ex-
pansion of equation 3.44 indicates, of course, interaction between fac-
tors in the sense that the marginal products of the one which is varia-
ble depends upon the magnitude in which the other is held fixed. By
/c
5X (1 - Ry )(Rz loge Rz)
=
'
5Z (1 - Rz )(Rx loge Rx)
setting equation 3.44 to equal -k, a given price or substitution ratio,

the isocline equation becomes equation 3.45 where w has the value
_ log w
(3.45)
log Rx
Rz
(3.46) w =
7 7
k(l ~ Rf) loge Rx+ Rz loge Rz
From equation 3.45, it is evident that if Z is zero, X also is zero indi-
cating that the isoclines pass through the origin. Too, it is evident
from this equation that they are not straight lines as in the case of the
Cobb -Douglas function. Figure 3.5 indicates that, while the isoclines —
the curves with positive slope — are not straight lines, their slope is of
"asymptotic nature,” and they generally approach linearity out over the
input plane. They do not converge because the surface forms a ridge,
rather than a peak. Because the isoclines "bow out” from the origin,
even if they do approach linearity, they specify that a different input
mix should be used for each level of output. (The isoclines are not also
scale lines.)
Quadratic Forms
Extension of the quadratic form in equation 3.18 to two resources

results in the production surface equation in equation 3.47.
(3.47) Y = a + bjX, + b 2 X2 - b 3 X, - b 4 X3 + b^Xj

Diminishing marginal returns exist for either factor alone but there
3
is positive interaction between the two factors (Negative or zero .
3
The equation might be presented with only positive signs for the b/s. However, for
fixed plants, it is logical that the signs are negative for b3 and b4 and either positive or
,
negative for b 5 Hence, this framework is used in the text, although certain general algebraic
.
relationships are the same regardless of the signs used.

OF
QUANTITY
Figure 3.5. Isoquants for Spillman function.
interaction also may exist where diminishing marginal returns hold

true for both factors.) The corresponding isoquant equation is equation
3.48. The isoquants are not asymptotic to the input axes, as was true
2
bj + b,X, t [(b, + b 5 X2 ) - 4b3 (Y + b4 x| - b,X, - ayf
(3.48) Xx
2b *
for the Cobb -Douglas function. Certain output levels can be attained
from input of Xx alone, with X 2 at zero level, depending on the magni-
tudes of a, b 1? and b 3 . Similarly, certain levels of output can be at-
tained with zero input of X x and isoquants are allowed to intersect the
input axes of a contour map of the surface. Because of these character-
istics, the surface need not slope to zero level at the input axes, but
may form a precipice over them. In contrast to the surface formed by
the power function, the surface formed by the quadratic equation can
have a distinct peak, denoting a maximum output for a single limita-
tional combination of the factors. This limitational combination repre-
sents the point in the input plane where the family of isoclines converge
and the output contour, representing the level of maximum output, re-
duces to a single point.
The equation of substitution rates corresponding to equation 3.48 is
^ - 2b 4 X 2 + b 5 X,
(3 *
' 49 ) El _
" '
'
5X 2 b, - Zb3 X1 + b 5X a *
By setting equation 3.49 to equal the constant price ratio or substitution

rate, the isocline equation 3.50 is obtained. While the isoclines again
kb, - b2 /kb^+jbA
(3.50) X2
V b5 + 2kb s /
are linear, they do not impose the same conditions on the production
surface and the economic optima in factor combination as do those of
the power function. This is true since the isoclines, although linear,
are not forced through the origin of the input plane. Only a single iso-
cline, the one representing a substitution rate of k such that
kb, - b 2
(3.51) i 2_ _
0
b 5 + 2kb 3
'
does so. In other words, k must be equal to b l b If the ratio in .
1 2
equation 3.51 is greater than zero, the isoclines intersect the X, axis.
If it is less than zero, they intersect the X axis. In general, since
2 they
intersect an input axis at quantities greater than zero, they are not also
scale lines. As expansion paths, they denote a changing proportion of
resources, if the substitution rate is to remain constant at k magn itude
as higher levels of output are attained. For this same reason, they in-
dicate that the least -cost path of outputs is attained only if the mix of
inputs is changed as the factor to product price ratio changes.
The isocline representing a ridgeline, with a zero marginal rate of
substitution of X 2 for X,, can be derived when the partial derivative
shown in equation 3.49 is equated to zero. It has the value
and intersects the X2 axis where the later resource is of input magni-
tude Xj = .5b 2 b 4\ Similarly, the ridgeline defining a zero rate of sub-
stitution of X,for X2 intersects the input axes at X, = .5b bj 1
1 . The
ridgelines generally have a positive slope when the interaction term
in
equation 3.47 is positive and intersect each other, as is true for
all
other isoclines, at input values corresponding to the peak of the
produc-
tion surface. If there is no interaction between inputs, isoclines
denot-
ing substitution rates greater than zero are positively sloped.
However,
those isoclines representing zero rates of factor substitution,
the
ridgelines, form a 90 degree angle at the point of intersection
of the
Figure 3.6. Isoquants and isoclines for

quadratic function.
isocline family if there is no interaction between inputs. The isocline

representing a zero rate of substitution of X 2 for X x is parallel to the
X x axis at the value X 2 = ,5b 2 b 4 \ The one representing a zero rate of
substitution of X x for X 2 is parallel to the X 2 axis at the value X x
1
= .Sb^ . If interaction between resources is negative, isoclines rep-
resenting nonzero substitution rates can have both positive and negative
slopes. The ridgelines have negative slope as defined by the ridgeline
equations above when the substitution rates between X x and X 2 are zero.
A map of isoquants and isoclines for a quadratic equation with posi-
tive interaction between factors is included in Figure 3.6. It provides a
geometric illustration of some of the characteristics described above.
The isoquants extend only over the range where their slopes are zero
and infinite, defining the path of the ridgelines. In this form, with ex-
tension of the inputs so that isoclines converge at a “Von Leibig” point,
the function would hardly be appropriate for a livestock gain equation
where the surface slopes up to a broad plateau, rather than a definite
peak.
Square Root
The square root equation in equation 3.53 with positive interaction

term allows isoclines which are curved and pass through the origin. In
this sense, it is a compromise between such functions as equations 3.35
and 3.47. It does not specify a fixed mix of resources for attaining
s
(3.53) Y = a - - b 2 X 2 + b 3 X 1* + b 4 Xf + b sxfxf
different output levels, as does equation 3.35, and it does not impose
linear isoclines as equation 3.47. The corresponding isoquant equation
is that of equation 3.54.
(3.54) X,
b3 + b 5 Xf t ^(a - Y - b 3 Xa b 4 Xf) + (b, + bjXj
2b,
It allows product contours nearest the origin in the input plane to inter-
sect the factor axes. The derivative of equation 3.54, the marginal rate
of substitution of X 2 for x,
X
is provided in equation 3.55.
5
5X, -b, + -5b4 X-- + .Sb^Xfx ;-5
(3.55) "*5 3
8 X3 -bx + .5b
3
X1
+ .5b
5
X"‘ X^
The corresponding isocline equation is equation 3.56.
2
• 5h 4 r
2
'5
+ /2b 5 X“'
5
(.5kb 3 + .5kb 5 X2 '5
) + (b 2 - kb, - ,5b4 x"'
5
f
(3.56)
From equation 3.56, it is obvious that the isoclines pass through the
origin and have positive slope. However, because of the numerous
square root terms, the isoclines are not linear and converge to a point
in the input plane corresponding to the maximum point on the produc-
tion surface. For k = 0, denoting the zero rates of factor substitution,
the ridgelines are curved but intersect the axes at
(3.57) X2 = -25 b4 b 2 2
for a zero rate of substitution of X for X and
“2
(3.58) Xx = • 25 bf b
for a zero rate of substitution of Xx for X 2 .

A map of isoquants and isoclines for the square root function with
positive interaction between the factors is included in Figure 3.7. Again
the isoclines converge at an input combination corresponding to the
maximum output, the peak of the production surface. As the curved iso-
clines or expansion paths indicate, the least-cost path to higher outputs
denotes a changing mix or ratio of inputs. The input mix should change
as product price increases or decreases and output level increases or
decreases accordingly.
Other Forms
Other polynomial forms provide a compromise between equations

3.47 and 3.53 in respect to nature of isoclines. One is the function in
equation 3.59 where the x
X
are raised to the 1.5 power. The marginal
forms of production functions 93
Figure 3.7. Isoquants and isoclines for

square root function.
(3.59) Y = a + b^t b 2 X2 -
1 *5
b 3X 1 - b4 X^ + b 5 X t X 2
rate of substitution equation, defining slopes of isoquants is equation

3.60.
_b 2 - 1.5b 4 Xf + b X? 1
(3.60) =
6X 2 bx - l.Sb.X* + b 5X 2
Equating this to -k, the isocline equation in equation 3.61 results.

“ 1.5kb 2
s + V 4b 5 (kb 1 - b 2 + kb 5 X 2 + 1.5b4
/
X2 ) + (1.5kb 3 )
(3.61) Xx
2b 5
Obviously, the value of X x can have a nonzero value even when X 2 is

zero and the isoclines do not pass through the origin as in equation 3.56.
Still they are not linear as in equation 3.50.
A production function combining characteristics of the power and
exponential functions is the transcendental function extended to two re-
sources in equation 3.62.
b c X1 b c X
(3.62) Y = aX x e X2 e
This function assumes factors are limitational and that Y is zero if

either X x or X 2 is zero. It has a maximum point at the single peak
where one limitational combination of inputs defines the largest output
*
attainable. The isoquant equation in logarithmic form is equation 3.63.
1
-Sx 2
(3.63) log + c 1 b^ X1 = b"
1
log(a X2 e Y)
It has the disadvantage that solutions can only be ascertained by iter-

ative procedures. The corresponding equation of substitution rates is
equation 3.64.
6X (b 2 + c 2 X 2 )X 1
(3.64)
i=
5X 2 (b x + c 1 X 1 )X 2
Setting equation 3.64 to equal -k, the isocline equation in equation 3.65
is derived. As it indicates, the isoclines pass through the origin, are
curved and converge in factor space, defining a maximum physical
product, at inputs where the partial derivatives of X 1 and X2 are zero.
kb,x2
(3.65) X,
(c
2
- kc x )X2 + b 2
The function in equation 3.62 allows the surface to display both in-
creasing and diminishing marginal products. Equation 3.47 can be ex-
tended to allow this condition by adding a cubed term, for example, for
each X x . The same may be accomplished for equations 3.53 and 3.59
by adding terms of higher order for each resource. The logistic func-
tion also allows for increasing and decreasing (but not negative) mar-
ginal products.
The a resistance * or Balmukand type production function can be
written in the form of equation 3.66 when two input categories are used.
A representative surface is included in Figure 3.8.
-1 1 -1
(3.66) Y = a(b + XJ" + d(f + X2 )
+ c
This function does not allow negative marginal products but provides a
surface asymptotic to the maximum yield of It has the isoquant

equation in equation 3.67.
aY(f + XJ
(3.67) Xl = ' b
(f + Xj) - Yd - Yc(f + XJ
The equation of marginal substitution rates is
2
5X1
= "
d(b + XJ
(3.68) 2 '
6X7 a(f + X,)
Setting it to equal -k, we obtain isocline equation 3.69.
_1
(3.69) X l = (
/
| kad )(f + XJ - b
QUANTITY OF X,
Figure 3,8. Surface representing resistance function.
The isoclines axe linear but do not pass through the origin as indicated
in Figure 3.9. Hence, the optimum proportion of factors changes as
product price increases or decreases relative to factor prices. In
other words, different proportions of factors give a constant substitu-
tion rate as output is increased to higher levels. The isoclines do not
converge to a maximum point.
From the characteristics pointed out above, it becomes apparent
that a function of this general nature might well describe a livestock
gain function where, with both input categories increasing over the
complete factor plane, the function forms a surface which "slopes to-
ward a plateau* or long ridge, rather than to a peak. But equation 3.66
would not be appropriate for a fertilizer response surface where
negative marginal products are encountered and the sample observa-
tions include product isoquants which have ranges of both positive
slopes (complementary), as well as negative slopes (competition).
Various modifications of the resistance functions are possible. One
is shown in equation 3.70
(3.70) + d
X,X 2
where a term has been added and the Xx and

to allow factor interaction
X2 variables are not with another constant. A function such as
summed
equation 3.70 may apply well where one need not account for that por-
tion of output due to other fixed factors, such as nutrients already in
the soil or feed to an animal prior to initiation of the experiment. How-
ever, equation 3.70 does not allow negative marginal returns and has
1
an output asymptotic to d” Hence, it may be particularly applicable
.
for production functions describing meat production of birds and ani-

mals. The isoquant equation corresponding to equation 3.70 is shown
in equation 3.71.
aX2 + c
(3.71) 1
(Y" - d)X 2 - b
The equation of substitution rates for equation 3.71 is equation 3.72.
6X x (bXj + c)X 1
(3.72) ~ "
6X 2 (aX 2 + c)X 2
The isoquants intersect the X axis where X x is

x equal to ~cb~
1
They
.
intersect the X 2 axis where X 2 is equal to -ca" 1

. The isocline equation
is equation 3.73.
2
-c t Vc + 4bkX 2 (aX 2 + c)
(3.73) Xx
2b
The isoclines are not linear and do not pass through the origin of the
input plane. A particular advantage of equation 3.70 is its computational
convenience. It can be estimated directly as a linear regression equa-
tion in the form of
(3.74) Y" 1 = ax*" *

1
bX^ 1 + cX^X' 1
+ d .
APPROPRIATE SURFACE FUNCTIONS
A few of the infinite number of possible production function equa-

tions have been examined to illustrate that each supposes some
particular property in respect to marginal products, isoquants, mar-

ginal rates of substitution, and isoclines. If the production logic, the
correct mathematical form, were fully known and data were available
for a segment of the surface, the logical function could be fitted to the
data. This fitted function then could be used to estimate or predict the
whole surface. But if the production logic is totally unknown, the fitted
function can only be a “statistical” function, and not a logical function.
Hence, it can be used mainly to predict within the range of the observed
data. Within the data range, it serves as an approximation to the basic
structural phenomena of the process under study. The fitted parame-
ters for the form of the equation may not correspond, variable by vari-
able, to the true structural representation; but taken in toto the fitted
coefficients have effects that approximate those of the true but unknown
parameters, at least within the range of data fitted.
The appropriate form of function to be fitted to a production surface
should be selected in terms of the environment and nature of the pro-
ducing unit for which it is estimated. Biological and production knowl-
edge should provide some basis in selecting forms of functions which
are appropriate. A power function with constant elasticities and iso-
clines which are linear through the origin of the input plane may be
useful for certain firm or industry estimates. This might be particu-
larly true where interest revolves only around knowledge of resource
productivities at the mean of inputs. However, if the goal is to predict
the entire surface and if factor substitution rates change, as well might
hold true for labor and capital increased for a fixed firm plant, with
larger inputs, other algebraic forms would be more appropriate. Func-
tions estimated from farm samples ordinarily have been of power form
because of the smaller number of degrees of freedom involved in
estimating the parameters, and partly because a multiplicative model
has seemed logically appropriate. A quadratic or similar form fitted
to a farm sample ordinarily proves to have many regression coeffi-
cients which are not significant in a probability sense. Usually, only
the coefficients for linear terms are significant.
Again, it is quite obvious that if the sample or experiment examined
includes a range of increasing marginal products, algebraic forms
should be selected accordingly. A modification of the power function to
transcendental form as in equation 3.62, or addition of terms of higher
power as in equations 3.47 and 3.53, would allow this condition. Simi-
larly, if predictions are to be made for an experiment which indicates
a decline in total physical product and hence implies negative marginal
productivity, some algebraic forms are not appropriate while others
are. A Cobb -Douglas or Spillman function allows only an increasing
total product (unless the coefficients are negative, denoting only a de-
creasing function) If fitted to data of this nature they would generally
.
underestimate the slope of the surface and the magnitudes of marginal

products for smaller inputs. If there are reasons for believing that
constant elasticity does not prevail, and if knowledge of elasticity and
related magnitudes over various parts of the function are important, the
power function should not be used. Similarly, the Spillman function
should not be used where it is believed that successive marginal prod-
ucts do not bear a constant proportion to each other; nor is the conven-
tional quadratic form, as in equation 3.47, appropriate if a linear mar-
ginal product curve seems unrealistic. The same reasoning would
apply to use of the “resistance* formula, in finding situations appropri-
ate to the fixed relationships supposed for reciprocals of marginal
products.
The surface for meat production from an animal or bird is expected
to differ from one for milk production by a dairy cow or fertilizer re-
sponse for land, where inputs are expanded to the point of maximum
physical output from the “fixed* animal or acre. Yet, if the input range
does not approach the level defining maximum output, the same or a
similar algebraic form might serve nearly as well for meat production
as for milk or crop production. Or, a form of function which is appro-
priate in predicting a fertilizer response surface which rises to a dis-
tinct peak might also serve appropriately for meat production where
the surface slopes up to a long ridge. The coefficients for the latter,
used only to predict over the range of observations, may lead to a sur-
face sloping upward with growing width as is expected to be the case.
The production surface appropriate for total gains of meat from a bird
or animal ordinarily will not, in relevant growing and fattening periods,
have a positively sloped area converging to a peak between ridgelines. '
Rather, in most cases, the ridgelines will not converge to a point of in-
tersection and the surface will tend to form a long plateau. This pla-
teau widens over the factor plane and along the input axes. These state-
ments apply to total gain or production per bird or animal. On the
other hand, if the production period is defined as a day, and output is
measured as daily gain, the resulting surface appears to be defined with

converging ridgelines and a definite peak or maximum. The peak of the
surface defines the maximum daily gain and ordinarily is possible only
with a single limitational ration or resource combination. This surface
has certain characteristics paralleling that outlined later for milk.
A fertilizer production function which includes both applied and soil
nutrients in the variables, or one relating only to applied nutrients for
a pure sand containing no growth elements undoubtedly would have
characteristics such as those represented by the isocline family for the
square root function, or perhaps for a Spillman function if sample ob-
servations do not extend into ranges of negative marginal products. The
isoclines, perhaps other than those representing a zero replacement or
substitution rate, would necessarily pass through the origin. No yield
is possible without some quantity of nutrients required for chemical
and physiological processes for growth. To the extent that with non-zero
inputs of the various nutrients, the same yield level can be attained with
a number of combinations or proportions of nutrients, the isoclines and
ridgelines will "spring apart.” (Otherwise the surface would reduce to
a knife’s edge, passing through the origin and indicating that higher
yields can be attained only by using more nutrients in the single, limi-
tational ratio.) Then, the isoclines and ridgelines must converge and
intersect at inputs corresponding to the maximum height or peak of the
yield surface for nutrients which have toxic or yield depressing effects
when applied in large quantities. Hence, a production function with
characteristics paralleling those of the square root function or the
modified power function might be selected for a soil highly deficient in
two or more nutrients. But in most cases, except for a pure sand,
some quantity of the nutrients ordinarily will be available in the soil.
In other words, the basic isocline family, for applied and soil nutrients,
might have the nature of those intersecting the origin and point in M
Figure 3.10. But, as might be indicated by a soil test, the soil may al-
ready contain a quantity of A and B as indicated by point S in Figure
3.10. The soil would still be relatively limited in the two nutrients,
however, and the family of isoclines would likely be curved. The square
root production function equation allows this characteristic because of
the shape of its isoclines.
Suppose, however, that the quantity of nutrients A and B available in
the soil is that indicated at Q before addition of fertilizer for the plant.
Then a new origin is relevant. This is denoted as Q and the relevant
axes are the broken lines. In other words, the available amount of nu-
trient A is OL and the available amount of nutrient B is ON. A quanti-
tatively accurate soil test would indicate these quantities, denoting that
the fertilizer response surface, predicted from fertilizer application,
starts from point Q, rather than from point O, in the nutrient plane. In
a regression equation including soil-test quantities as observations, co-
efficients for nutrients already in the soil would be based on quantities
ON of B and OL of A. (By including nutrients already in the soil as
variables in the regression equation, better knowledge of the production
surface and isocline family should develop.) In cases where the nutri-
ent content of the soil is relatively high, however, and soil -contained
nutrients are not employed in the estimates, the production function
used for estimating may not provide nonlinear isoclines for the added
nutrients. The relevant isoclines to be predicted for the added nutri-
ents are those above and to the right of the dashed axes (with origin Q)
in Figure 3.10. Although these portions of the isoclines are curved, the
curvature probably would not be great enough to show up in the produc-
tion function equations. The best -fitting production function may then
provide a family of linear isoclines such as those expressed in equation
3.50 for the quadratic form. They intersect at M, the maximum yield
per acre. Also, since they are linear and do not intersect the origin,
they denote, like those expressed in equations 3.56 and 3.65, that the
least-cost nutrient ratio changes as higher yields are attained.
If the response surface estimated has actual origin at S in Figure
3.10 and applied inputs are SC of A and SD of B, a function of the nature
in equation 3.70 might be appropriate. It provides isoclines which are
nonlinear and which do not converge at either the origin or a maximum
yield. While the entire isoclines over the input plane with origin at S
are curved, those falling in the observed rectangle SCQD may have so
little curvature that it cannot be effectively estimated.
Occasionally, the research worker encounters special properties of
production functions which must be considered in his estimates. An ex-
ample is the milk production surface for dairy cows. While its exact
algebraic nature has not yet been established, it is possible to offer
some hypotheses about certain of its characteristics. Generally, the
limit of a cow’s stomach prescribes the maximum amount of an input
such as hay which can be fed in a day, month, or year. On the other
hand, physiological conditions may specify a minimum amount of hay
which she can consume in an extended time period, particularly where
the substitute feed is corn or other concentrate. The milk production
surface, although not yet fully established, may appear somewhat like
Figure 3.11. Diminishing returns exist for transforming any one ration
into milk since the surface declines in slope at higher levels of output.
Also, diminishing substitution ratios are indicated between feeds since
the slope of the milk isoquants changes with feed ratios or rations.
Maximum milk production per cow is at level M. Only one ration and
level of feeding will result in maximum milk production and is denoted
at point G in the feed plane.
Milk is forthcoming if a hay ration alone is fed (i.e., if level of feed-
ing is extended along the forage axis OF) . However, the maximum level
of milk production, FH, under a pure forage ration is lower than
the
maximum level GM attainable under the ration following line OG in the
feed plane. With the cow fed hay only, OF represents the quantity de-
noting her stomach capacity. Supposedly, if a cow were fed to the high-
est milk level possible from a pure hay ration, and if milk level FH
were attained, the amount of hay could be reduced along a stomach limit
line. By replacing it with successive quantities of grain, milk
Figure 3.10. Alternative isocline families for

production surfaces.
production could be raised to higher levels by following the line on the

surface denoted as HM until the maximum milk level of GM is attained.
The milk level line denoted as HM parallels the feed line FG in the feed
plane. Line FG also represents the limit of the cow’s stomach capacity
and indicates that, to attain higher milk levels along HM, the amount of
forage must be decreased as grain intake and milk production are in-
creased. The extreme grain ration denoted as OG does not follow along
the grain axis OC, under the postulate that a physiological minimum of
forage must be included in the ration if lactation is to extend over a
long period. Hence, the area OFG in the feed plane represents the
limits of rations with respect to (1) the grain to hay ratios which may
be fed and (2) the maximum possible intake of any particular grain to
hay ratio or ration. Over the top of the production or milk surface
(i.e., the area OMH in Figure 3.11) are milk contours or isoquants such
as a" through f", each indicating a particular milk level. If lines cor-
responding to these milk isoquants are drawn in the relevant feed plane,
OFG, they indicate all of the possible rations which allow attainment of
a particular milk level.
The two-dimensional isoquant map corresponding to the model of
Figure 3.11 is shown in Figure 3.12. Line H'M' in Figure 3.12 is the
stomach limit line and parallels line FG (HM) in Figure 3.11; O'M' is
the physiological limit line corresponding to OM in Figure 3.11. The
isoquants a" through i" are the counterparts of those shown in Figure
3.11 and, again, indicate all possible rations or feed combinations which
allow attainment of the specified milk output. The slope of these
isoquants indicates the grain to hay substitution ratio. No particular
function may readily describe this particular production situation.
However, if rations extend to the point of isocline convergence, quad-
ratic forms may serve better than exponential forms of functions.
However, if sample observations do not approach the convergence point
in Figure 3.12, these and other types of functions may serve satisfac-
torily for a particular experiment.
BASIS FOR SELECTION

The sections above suggest that knowledge of biological, economic,
or other environmental factors may exist to provide some guide or
basis for selecting a production function. Frequently, however, previ-
ous knowledge as a basis for selection may not exist and several alge-
braic forms may be used initially, with various empirical criteria used
for selecting among them. Different research workers emphasize var-
ious empirical extremes or approaches in selecting a form to be used.
At one extreme, some prefer to start out with an initial hypothesis
about the algebraic form of the regression equation.. They then use
this form and retain all coefficients even if their standard errors are
relatively large and even if the term has small numerical importance
in predictions. Having selected the particular function on a priori
bases, they would retain it unless it appears entirely inconsistent with
the numerical observations. At the other extremes, some research
workers would prefer selection of a polynomial with sufficient coeffi-
cients and powers for the variables and retain them only if the coeffi-
cients are significant at the conventional 5 per cent probability level.
Or, two other extremes are sometimes followed. In the one case, basic
conditions or laws of production are supposed and an algebraic form is
selected, with the data fitted to this single form. In the other case, no
assumption or hypothesis about the form of the function is formulated
and the experimental or sample data are subjected to statistical analy-
sis, leading to selection of an algebraic model which evidently provides
the “best fit.” Numerous empirical criteria may be used in this se-
lection: one is to select several different algebraic models and fit the
observations to each of them. One of “best fit” then may be designated
by the magnitude of the coefficient of variation, R2 , if the assumption of
normally and independently distributed errors is not violated. With R 2
indicating the proportion of variance in the dependent variable ac-
counted for by a particular type of equation, the larger value might be
taken to indicate the form which is most appropriate for estimates.
Other related statistics which might be used as empirical criteria in-
clude the F ratio and the mean square of deviations from regression
(lack of fit). A larger F ratio (the mean square due to regression di-
vided by the mean square of experimental error) or the smaller mean
square of deviations from regression (i.e., the lack of fit) is taken to
indicate a model most appropriate for the particular set of experimental
Figure 3.11. Production surface with

limits of substitution.
or sample observations. Or, a general polynomial form can be selected,

Including constants for various powers or degrees of the independent
variables and an appropriate number of interaction terms. An analysis
of variance may be made for these, with a mean square and F ratio
computed for each. Terms can be dropped if the corresponding F ratio
is not significant at an acceptable probability level. Or, the constants
in the regression equation may be computed directly, with the “t” test
(the null hypothesis of regression coefficients at zero level) used to in-
dicate terms which should be dropped because they are not significant
at acceptable probability levels. (A procedure sometimes preferred is
to select initially a simple polynomial form and add terms one at a
time, retaining those which account for a significant incremental por-
tion of variance in output.)
It is not necessary that the research worker confine himself to one
or the other of these extremes. In some cases he may have sufficient
Figure 3.12. Isoquants and isoclines for Figure 3.11.

theoretical and empirical bases for proposing a single initial model and
fitting the data to it. This approach has appeal in logic and may be es-
pecially appropriate for laboratory experiments in chemistry and
physics where control of magnitude for all variables is possible or
nearly so. However, for biologic and economic phenomena, it appears
unlikely that a single mathematical form of production function is most
appropriate for all environmental and resource situations. Too, differ-
ent individuals may be able to give equally valid reasons for selecting
alternative types of functions. Errors in design, measurement, and
specification also may cause a function, other than the one initially se-
lected, to be more appropriate. Still, the research worker usually does
have some basis for selecting, among the very many algebraic forms
possible, one or a few which are appropriate for the phenomenon under
consideration. If he feels that the initial hypothesis of functional form
should be retained, he may follow this procedure, after applying a sim-
ple criterion such as that the lack of fit term is no larger than experi-
mental error for the regression equation. Some of the mathematical
properties outlined on previous pages provide guides in selection. For
example, the conventional power function would be inappropriate in fit-
ting a surface for a fertilizer experiment where isoclines converge
within the range of observations. Yet, while a general type of function
may be appropriate for a particular set of observations, the research
worker may not have empirical or biological knowledge for specifying
a single modification of it. He may then use various empirical tests to
determine the particular modification which gives the “best fit.” In
case previous knowledge and theory are nil, he might be forced to use
this approach from the outset, even going so far as to select functions
with contrasting algebraic properties and subjecting them to tests of
“best fit.”
Few experiments and samples exist which do not require exercise
of experience and judgment by the analyst. Even if he rejects the
somewhat “rigid” approach of using a priori assumptions of the alge-
braic form, and fitting the data to this single selection, he seldom can
let choice rest alone on objective empirical selection of the one “best-
fitting” form. In the first place, he is faced with limitations in funds
and time. He must reject some types or modifications of equations be-
cause it is too costly to compute regression coefficients or estimate the
parameters for the entire range of possible functions. Some equations
may be rejected simply because their algebraic form makes computa-
tions and solution of economic or other optima too difficult. In other
cases, the probability level at which a regression coefficient of a poly-
nomial equation will be accepted as nonzero differs among research
workers and is a matter of experience and judgment. Some insist that
the conventional probability level of .05 must be attained if the coeffi-
cient for a particular degree of an independent variable is to be re-
tained in the production function. Others would insist that this pure
empirical appeal is too rigid, as perhaps it is. For example, a “t” test
for the quadratic equation in equation 3.18 may indicate that the
regression coefficient for the squared term derived from a fertilizer

experiment can be considered to differ from zero only at a probability
level of .20. If we retain the term in the equation only if its coefficient
is significant at a .05 probability level, we assume that the input -output
relationship is linear. In general, this assumption is inconsistent with
known biological conditions, since most empirical research indicates
diminishing marginal productivity when one nutrient is varied and all
other resources are fixed in magnitude. Too, it would imply that if
fertilizer is profitable at all, the nutrient should be applied in unlimited
quantities. The research worker has two alternatives: he can accept
the term on the basis of theory and previous experience, and perhaps
as a basis for conservatism in recommendations He can refine the ex- .
periment (through improved design, a larger sample, and better meas-

urements) and repeat it under similar environmental conditions. Per-
haps both alternatives should be used, with the first serving as a basis
for immediate decisions and recommendations and the second used to
increase precision of later estimates. Infrequently, the phenomenon
being examined does not allow the latter type of refinement. The re-
search worker may lack funds and facilities or the observations may be
time series generated from variables of the national economy. Some
have suggested the following criterion on decision to include or delete
a variable: omit the variable if the standard error of the regression
coefficient exceeds the coefficient itself.
Selection among algebraic forms of equations often is no less diffi-
cult than decision in respect to the probability level at which variables
will be omitted from a polynomial model. There are the logical criteria
mentioned previously (such as functions allowing negative marginal
products, or converging isoclines for a fertilizer experiment, noncon-
verging isoclines and an asymptotic maximum for meat production, etc.)
which, along with such statistical criteria (as magnitude of R 2 or devi-
ations from regression) do provide guides. Even then, however, the
optimum function is not always apparent. For example, the three func-
tions below have been derived from a corn fertilization experiment in
Iowa with corn yield, Y, measured in bushels per acre and nitrogen
level, N, measured in pounds. 4 Only observations where P 0 is con-
2 5
stant at 160 pounds per acre are considered. The magnitude of variance
quantities provide little basis for selecting among the three functions.
The power function in equation 3.75, while it gives the highest R 2 , might
be rejected because of the apparent possibility of negative marginal
(3.75) Y = 33.01N’ 2288 R2 = .95
(3.76) Y = 42.60 + .80N - .0018N 2

to
il 00
o
Y 5 N
(3.77) = 23.67 + 12.82N' - 13.37 R2 = .91
4
Heady, Earl O., Pesek, John T., and Brown, W. G. Crop response surfaces and eco-
nomic optima m fertilizer use. Iowa Agr. Exp. Sta. Bui. 424. Ames. March, 1955.
response at large N levels and the equation does not define a maximum.
The quadratic 76 ) and square root ( 3 77 ) equations are similar in the
(3 . .
sense that both allow changing elasticity, diminishing marginal prod-

ucts, and a defined maximum. With their derivatives equated to nitro-
gen to corn price ratios of . 8 , 12 16 , and 20 , the profit maximizing
.
,
. .
rates of fertilization are, respectively, 206 , 195 , 184 , and 172 pounds
of N for the quadratic and 183 , 156 , 134 , and 117 pounds for the square
root equation. (The corresponding rates for the power function are,
respectively, 361 , 213 , 147 , and 110 pounds of N.) Generally, the
square root function, as compared to the quadratic function, will call
for lower input rates with larger price ratios and higher input rates
with smaller price ratios because its marginal product curve is not
linear (i.e., the second derivative is not constant). Agronomists have
not yet established laws of marginal crop products, although some would
suggest that they do not “ decline in straight line fashion.”
In a pragmatic sense, small differences in estimates by alternative
equations can be unimportant because price, or even weather underlay-
ing the production function, is so uncertain that exact profit maximizing
input combinations can never be selected. Extension specialists, with
some rightful basis, would be prone to ask, “Why should extreme effort
be invested in refining functions for such inputs as fertilizer because
farmers discount future production greatly due to uncertainty and use
less than recommended rates (which, in turn, fall short of the level
equating marginal revenue and cost)?” Generally, farmers can only
guess on future revenues because price and weather are unknown.
However, modifications in empirical procedures may be used to lessen
differences in optima specified by different functions. Anderson has
suggested elimination of observations of zero inputs for variable re-
sources as one method of attaining greater predictional consistency be-
tween models. 5 When observations with N = 0 are eliminated for the
Iowa data used in the equations above, the following quadratic and
square root functions result:
(3.78) Y = 80.77 + 14.22N - 1.0413N

2
R2 = .70
(3.79) Y = 32.32 - 13.37N + 17.6693N*5 R2 = .70
In terms of magnitude of R2 the two functions are equally acceptable.

,
The estimated inputs where the marginal products equal the price ra-
tios of .8, .12, .16, and .20 now are, respectively, 212,
181, 150, and 120
for the quadratic and 188, 156, 131, and 113 for the square root equation.
Within the range of price ratios historically realized, some greater
consistency is attained but there is still considerable variance in rec-
ommendations which might be based on the two equations over reasona-
ble price ratios.
5
Anderson, R. L. Some problems in the analysis of fertilizer response data,
statistical
*
in Baum, E. L., et al. (eds.), Economic and technical analysis of fertilizer innovations and
resource use. Iowa State University Press, Ames. 1957. P. 192.
Another criterion of general use in deciding between alternative

functions is to compare predicted and observed yields over the relevant
range of the production surface. Assuming the other criteria men-
tioned above are adequately satisfied, the best function will be that for
which the scattering of positive and negative residuals over the surface
is random. Should there be a systematic tendency for predicted yields
to exceed observed yields (or vice versa) over parts of the surface,
this is an indication that the fitted function does not adequately char-
acterize the observed data. Not unexpectedly, when there are 3 or more
input factors involved, the search for systematic tendencies in the re-
siduals becomes messy unless some orderly approach, such as sug-
6
gested by Valavanis, is used.
Difficulties encountered in selection among continuous models may
be eliminated in form free models such as those suggested by Hildreth. 7
In estimating points on the production surface, the investigator is not
forced to specify a continuous algebraic form. However, some would
argue that if this general procedure is to be used, agronomists might as
well use more replicates to obtain better mean yield estimates of the
different treatments, and then assume the function to be linear between
these points. Again, it is possible that, if sufficient points are esti-
mated, farmers need no more refinement than this. Because of uncer-
tainty and imperfect knowledge, they would not be able to equate the
slope of a continuous function with a price ratio even if the algebraic
form were known with absolute certainty.
6
Valavanis, S. Econometrics. McGraw-Hill, New York. 1959. P. 152.
7
Hildreth, C. G. Point estimates of ordinates of concave functions. Journal of the
American Stat. Assoc., 49: 598-619. 1954.
CHAPTER 4
Data Analysis
for Production Function Estimation
REVIOUS CHAPTERS have shown the usefulness of production

P functions as decision-making guides. Still, unless empirical esti-
mates are available, knowledge of how production functions may be
algebraically manipulated to give meaningful economic information is
of no practical use. A skeleton of theory is lifeless without empirical
flesh; to be more than playthings in an ivory tower, production func-
tions must be specified in terms of real-world situations. In deriving
these empirical specifications, statistical procedures are generally
used complemented by the biological, physical, and economic logic rel-
evant to the particular production process being examined.
An introductory outline of the more important statistical techniques
and problems pertinent to production function estimation is presented
in the current chapter. The aim is not to present a treatise on tech-
niques of estimation. Rather, it is to give the non- statistician working
on production functions knowledge of the relevant concepts. He can
then make better use of the statistical services and references avail-
able to him. Also, they provide the statistical background for the em-
pirical work discussed in later chapters.
There are two phases in the fitting of production functions: (a) the
collection of data and (b) its analysis. Data may be collected from
either experimental or nonexperimental sources. The source used will
generally depend on the particular production process being examined,
the stage of the investigation, and the research resources available.
However, once a choice has been made between experimental and non-
experimental data, the techniques and problems of analysis largely
dictate the particular methods of data collection to be used. On these
grounds, the techniques and problems of analyzing data are considered
in this chapter; procedures and problems of data collection are dis-
cussed in Chapter 5. Difficulties involved in the “economic specifica-
tion^ of production functions are discussed in Chapter 6. Their appre-
ciation requires some understanding of the statistical procedures
relevant to the collection and analysis of data. It must be emphasized
that the statistical andeconomic aspects of specification are very
closely interrelated; indeed, so closely that the line of demarcation be-
tween them is somewhat arbitrary. They are considered separately for
expository convenience.
DATA ANALYSIS FOR ESTIMATION 109
SINGLE EQUATION APPROACH
basic approaches to production function estimation are possi-

Two
ble. the one hand, the production function may be regarded as but
On
one of a number of interdependent relationships constituting a model of
the economic situation being studied. Such is the simultaneous equa-
tions approach Alternatively, the single equation approach may be
.
adopted: the production function being considered as a single inde-

pendent relationship, uninfluenced by other relationships relevant to
the economic milieu enveloping the production process. Which of these
approaches is strictly appropriate is a question of fact. It can only be
answered in terms of the logic — economic, physical, or biological —
underlying the production process being examined. Such problems of
specification are taken up in Chapter 6. From a computational view-
point, however, a single equation model is often regarded as an alter-
native to a more strictly correct but computationally irksome simul-
taneous equations model. The estimational procedure for the two
approaches, while overlapping, are not identical. Since the single
equation approach has generally been used in deriving empirical pro-
duction functions, the statistical procedures and problems pertinent to
it will be outlined first. Essentially, this implies a discussion of the
1
technique of least squares multiple regression.
Multiple Regression Analysis with Linear Parameters
Suppose the production relationship between an output Y and inputs

X lf X 2
,
•• X k has, on a priori grounds of economic, biological, or
*
,
physical logic, been postulated as the following single equation model:
(4.1) Y = /3 0 + fi2 X2 + + /3
k X k+ e.
Regarding equation 4.1 as a regression law ,

Y is termed the dependent
variable. The explanatory factors Xx
are the independent
,
Xg, • •
•
,
Xk
variables in the sense that the values that they take arise independently
of equation 4.1. e is the error due to the fact that the postulated inde-
pendent variables do not completely explain Y; some input factors of
minor importance have not been taken into account. The parameters
Po> Piy •
0k 3X6
•
• population regression coefficients . From a
,
mathematical viewpoint, given the correctness of the single equation

model, equation 4.1 may be described as a continuous unilateral causal
relationship. In such terms Y is the effect variable and X x , X 2 , •
,
• •
X k the causal factors. It is desired to specify equation 4.1 empirically,

i.e., to determine the values of the parameters. If these values are
1
Full discussion is to be found in: Williams, E. J. Regression analysis. John Wiley
and Sons, Inc., New York. 1959. See also Bradley, R. A. Fitting response surfaces.
Industrial Quality Control, 15: 16-24. 1958.
110 DATA ANALYSIS FOR ESTIMATION
known, the economic implications of the production function may be as-
certained and applied in the real world. Thus interest is in the causal
relation as a whole, especially the j3 s, and not just in the values of Y
?
that may be estimated by using the equation. The causal mechanism is

of economic relevance in itself — as was explained in Chapter 2. Pro-
vided that some basic assumptions hold true, the fact that equation 4.1
is linear in the parameters makes estimation of these parameters by
multiple regression rather straightforward. All that is required is a
sample of n > k + 1 sets of data showing the value of Y for various
levels of each of the independent variables. Suppose such data are
available. They might be listed as follows:
X21 xkl
V -%r V
^22
... , . . , -V"
Xk2
T2
V
yn y
Am "5T ... « . a • Y
^kn
where Yj (j = 1, 2, •, n) is the level of Y attained when the i-th of the
• •
k inputs is at the level Xy. The estimate of equation 4.1 to be derived

from the given data will be of the form:
(4.2) Y = b 0 + b x X1 + b 2 X 2 + • • • + b k Xk
where bi, termed a sample regression coefficient, is an estimate of ,
and Y estimates Y. Since equation 4.2 is but an estimate of equation

4.1, the relation between Y and Y is of the form:
(«)
In other words, it is not to be expected that equation 4.2 will exactly

predict the level of output, Yj , forthcoming from a given set of input
quantities, Xy , X 2 , * * • , X
kj There will generally be some discrep -
.
j
ancy or residual between the observed and predicted value of Y. This
residual is the term e of equation 4.3. It is an estimate of the theo-
retical error, 6 , of equation 4.1. For convenience in algebraic manip-
ulation, equation 4.2 is often written as:
(4*4) Y = b 0X 0 + b X + • • •
+ bk Xk
x x
where X 0 is a dummy variable always equal to one.
Estimation of the regression coefficients
From the given data it is desired to calculate suitable values of the

coefficients b 0 b 1}
,
• • •, b k of equation 4.4. Such estimates of the /3 ’s
are regarded as suitable if, over all possible samples of data, they are
correct estimates on average and have the smallest possible variation.

In statistical terms, it is desired that the estimates be unbiased and
have minimum variance. By unbiased it Is meant, metaphorically, that
the estimates do not lie consistently to the right or left of the target. A
brief explanation of the concept of variance is perhaps necessary. Con-
sider a variable x. The expected, mean, or average value of x in a
2
population of x’s is written E(x). The expected value of [x - E(x)] ,
i.e., the expected value of the square of the deviations of the x values
from their mean, is the variance of x. It is an indication of the varia-
tion that exists among the values of x in the population. Usually this
population variance is denoted by a x . Thus
2 2
(4.5)
cr
x = E[x - E(x)] .
The variance found among a sample of x values drawn from the popu-
lation will generally not equal the population variance. However, for
practical purposes, this sample variance provides an estimate of the
2
population variance. The sample variance is usually denoted by s .
The covariance two variables x and y is a measure of the extent to

of
which they vary together. For a population of x -and y values, it is
denoted by (r xy where
(4.6) or
xy = E[x - EM] [y - E(y)].
A variance is the standard deviation of a variable.

statistic related to
It isthe positive square root of the variance of the variable. The con-
cepts of variance and covariance are important in the assumptions that
must be fulfilled if least squares estimates of the regression coeffi-
cients are to be unbiased and of minimum variance. The assumptions
are:
(a) the expected value of the error e is zero
(b) the covariance between the error associated with one value of Y
and that associated with any other value of Y is zero
(c) the variance of the error associated with one value of Y is the
same as the variance of the error associated with any other
value of Y
(d) the covariance between the error and each of the independent
2
variables is zero
(e) the observations of the independent variables are measured

without error.
2
Assumptions (a) to (d) might be expressed more succinctly as follows: (a) E(e) = 0.
2
a = a 2 for = 1 = 0
= 0 for $ p and p = 1 , 2, or
* * •
(b) j j, *
,
n. (c) j , 2, •
•
,
n. (d) cr
J P j
i
for i = 0, 1, • • k.
Consideration will later be given to the procedures that might be
adopted if the above assumptions are not fulfilled.
Least squares principle

Using least squares, the aim is to minimize the sum of the squared
deviations (SSD) between the observed Y values, Y x Y 2 , Yn , and ,
• •
•
,
the corresponding estimated values of Y, Y x , Y 2 ,

Y n , calculated ,
• *
*
from the data by using equation 4.2. Thus it is desired to minimize
2
(4.7) SSD = 2 (Yj - Yj)
j=i
which, by virtue of equation 4.4, may be written as
2
(4.8) SSD = .1 (^ - b0 X - b lXlj b kXkj ) .
0j
Notice that this application of the least squares principle is centered

upon Y, the dependent variable. Such an orientation implies that the
derived estimates of the parameters will not be the same as the esti-
mates obtained if, say, not Y but one of the X’s had been made the de-
pendent variable. Thus the estimated regression equation, equation 4.2,
is an irreversible relationship in terms of causal influences. However,
the direction of inference may be reversed. Hence, given values of
bk and a set of levels for Y, X x , • , X
b0 , bj ,
• •
•
,
k -!, equation 4.2 • •
could be used to “predict” the level at which Xk was set in producing

the given level of Y.
Normal equations
To minimize the sum of the squared deviations, the partial deriva-
tives of equation 4.8 with respect to b 0 , b x , b must be equated to • •
•
, k
zero. By doing this, the following set of (k + 1) equations are obtained;
terms involving Y having been moved to the right-hand side. These are
the normal equations expressed in terms of the raw data.
b0 s X oj X oj + bx s X ojXlj + • •
. + b
kS X oj Xkj = 2 X oj
Yj
(4.9) y Xoj + Xlj Xlj + • • • + b k S Xlj X

kj = 2 Xy Yj
b0 2 XkjXoj + blS XkjXlj + • • • + b k2 X kjX kj = 2 Xkj Yj

where 2 implies summation over j = 1, 2, n. Since there are • •
•
,
(k + 1) equations and (k + 1) unknowns, b bk the system of • •

•
0 , , ,
normal equations 4.9 may usuaUy be solved for these unknowns. How-
ever, for computational purposes and as a theoretical simplification, it
is more convenient to express the normal equations in other
ways. One
such is in terms of the deviations of the observations on each variable

from their average value. Consider the first equation of system 4.9.
Since X„j = 1, it may be written as in equation 4.10.
(4.10) b 0 n + bL 2 X tj + • • • + bk 2 Xkj - = 2 Yj
Dividing through by n and writing X { and Y for the mean of Xi and Y,

respectively, equation 4.10 becomes
(4.11) b0 + bî + • • •
+ bk Xk = Y
and thus
(4.12) b0 = Y - b xX t bk X k .
Hence it is really only necessary to find b x , ,

b k . Substituting equa- • • •
tion 4.12 into equation 4.8 and rearranging terms, we have
(4.13) SSD = .2 [(Yj - Y) - b 1 (Xlj - XJ b k (X kj - X k )] 2
(4.14) = ,2 (yj
- b^j b k xkj)
2
where y, = (Yj - Y) and Xij = (Xjj - Xj ). The system of k normal equa-

tions obtained by taking the partial derivatives of equation 4.14 with re-
spect to bx , ,
b k is then
• •
•
bl S *uj + b2 S Xlj X 2j + * • • + b k 2 X X kj = 2 Xlj y

1 j j
(4.15) ?i Sx aj
x ij +b 2 S<j + ••• +b k 2x 2j x kj =2x. 7j
bl 2 xkj x lj + b 2 2 x kj x 2 + + b
k 2 x kj = 2 X kj y •
j j
n n
Each term £ x^ x hi or £ x- y,
J
,
(i, h = 1, 2, • •
•
,
k) in equation 4.15
J=i j=i
is a corrected sum of cross products Of course, a term in X*

. if i = h,
is a corrected sum of squares; these are found on main diagonal of the
the left side of the normal equations. Each cross product term might be
spelled out as exemplified in equations 4.16 and 4.17.
(4.16) Sxijyj = xu yj, + x12 y2 + ... + x m y n
(4.17) = (Xu - XJ (Yx - Y) + (X12 - XJ (Y2 - Y) + ... + (X ln - Xx ) (Yn - Y)

Moreover it can be shown, by rearranging equations such as equation
4.17, that
SXijSX
- hj
iJ
(4.18) S Xij x hj = Z XijXhj
SX M E Yj
(4.19) S
¥j =SX ijYj
These relationships provide an easy method of calculating the cor-
rected sums of squares and cross products. So far as solving equation
4.15 to obtain the bj values is concerned, the easiest way of visualizing
the problem — and of understanding the methods used — is as a matrix
3
equation. If we let
Sx *j S X ij X :
Sx ljXkj
(4.20) A =
Sx 2jXlj s X 2j x kj
Sx kj xij S x kj x S x kj
(4.21) G' =[Sxlj y j

Sx 2j yj
S x kj yj ]
(4.22) B' = [
bA b2 ••• bk ]
then the set of normal equations, 4.15, may be written as
(4.23) AB = G.
The matrix B is obtained as
(4.24) B = CG
-1
following the usual notation of denoting A by C. It is noteworthy that
A, and hence C also, is a symmetric matrix — a fact which simplifies
the computational effort involved in solving the normal equations. Usu-
ally, when formulated in terms of deviations of the variables from their
means, the normal equations are solved by the forward solution of the
4
abbreviated Doolittle method.
Frequently, the normal equations are expressed in terms of the
standard partial regression coefficients b' b' , ,*>'k, and the
• •
*
, ,
3
For an extremely simple exposition of the matrix concepts involved see Heady, E. O.
and Candler, W. V. Linear programming methods. Iowa State University Press, Ames.
1958. Pp. 378-407.
4
An easily followed example of the abbreviated Doolittle method is to be found Ander- m
son, E. L. and Bancroft, T. A. Statistical theory in research. McGraw-Hill Book Company,
Inc., New York. 1952. Pp. 197-99.
coefficients of correlation ry (i, j = 19 X 2 Xk,Y), between pairs

,
X ,
of variables. In such form, the normal equations are:
bjL r u + b'2 r 12 + + b'k r = r

ik ly
bi r 2i b 2 r 22 + + b' r
k 2k = r 2y
(4.25)
b'x rkl + b'rk2 + k kk = r ky

+ b r
where, for i = 1, 2, • •
•
,
k,
.2 xfj
j=i
(4.26) b'i = bi
and the correlation between and Y is given by
j=i
(4.27) L
iy
The correlation between Xi and Xh h = 1, 2, k) is calculated as(i, • •

•
,
in equation 4.27, replacing yj by xhj.These correlation coefficients,

which by construction lie between plus and minus one, indicate the de-
gree of association between the relevant variables. They have no
causal implications. If negative, the correlation coefficient indicates
that as one of the pair of variables increases, the other tends to de-
crease. the correlation coefficient is positive, both variables tend
If
to move same direction. The closer the coefficient is to plus or
in the
minus one, the greater is the degree of association between the two
variables. Thus rii , the correlation of a variable with itself, is neces-
sarily unity. Hence, the main diagonal of equation 4.25 really consists
only of b< values. If the correlation coefficient between a pair of inde-
pendent variables is, roughly speaking, greater than 10.8 , the problem I
of multicollinearity may arise. It is discussed later in this chapter.

As shown by equation 4.26, the standard partial regression coeffi-
cients are related to the sample regression coefficients. In fact, they
are the sample partial regression coefficients standardized in terms of
5
For an illustrative example of an efficient, widely used calculation procedure based on
the b', rij formulation of the normal equations, see Ostle, B. Statistics in research. Iowa
State University Press, Ames. 1956. Pp. 207-14.
the standard deviations of Y and X i? as equation 4.26 indicates. Being
standardized, the various b' values are comparable. They enable com-
parisons to be made of the relative influence on Y of the various inde-
pendent variables. Such comparisons cannot be made via the sample
regression coefficients because, in general, the independent variables
will not be measured in comparable units. To illustrate, from equation
4.2, it can be seen that
5Y
(4.28) = 1-1, 2,
Hence, if Xj changes one unit, all the other independent variables

Jjy
remaining constant, Y will change on average by bj units; alternatively,
Y is estimated to change by bj units. Now suppose X L is measured in
tons and X 2 in acres. Clearly, b L and b 2 tell nothing of the relative ef-
fects of Xj and X 2 on Y. However, a comparison of b'L with b'2 indicates
the number of standard deviations by which Y would change if X or X ,
t 2
considered separately, were each changed by one standard deviation. In
such fashion, the relative influence of the independent variables may be
gauged.
Variances and covariances of the sample regression coefficients
The matrix C of equation 4.24 plays an important role in estimating

the variances and covariances of the sample regression coefficients.
Denoting the general element of C by Cy (i, j = 1, 2, . ,
k), the vari- • •
J
ance of bi is given by
2
(4.29) Var(bi) = cucr
where a 2 is the true (but unknown) variance of the error, e Of course, .
c is a diagonal element of C. As an estimate of <r 2 s 2 is used where

£i ,
2
S (Yj - Y: )
(4.30)
n - k - 1
The covariance between pairs of regression coefficients is given by
(4.31) CovarCbjbj) = c^cr2 , i^j and i, j = 1, 2, • •

• k
,
The variance of thesum or difference of two of the sample regression

coefficients may be calculated as follows:
(4.32) V&r(b i ± b.) = Var(b, ) + Var(b-) ± 2 Covar(b.b.).

J A 1
J J'
DATA ANALYSIS OF ESTIMATION 117
Likewise, using the relationship given in equation 4.12, the variance of

b 0 may be estimated as in equation 4.33:
— k _ j-i k _ _
(4.33) Var(b 0 ) = Var(Y) + 2 X|Var(bi) + 2 2 S XjXj CovarO^bj)
i=i 1=1 j=i
2
where Var(Y) equals s /n.
Significance tests of the sample regression coefficients
Having solved the normal equations to ascertain the b values, it is

desirable to ascertain the reliability of these estimates. This may be
done by calculating confidence limits for the estimates or by testing to
see if the estimates are significantly different from some selected
value. For such procedures to be accurate, it is necessary that the
distribution of the error term e be known so that a test statistic may
be devised. In general, the derivation of such a test is extremely com-
plicated. Because of this, it is desirable that the errors be normally
distributed; in which case the a t” statistic, for which tables are avail-
able, may be used. Assuming that the errors are normally distributed
2
with a mean of zero and variance a the value of t to test whether the
,
sample regression coefficient bj is significantly different from zero at

some probability level a is given by
bj
(4.34) ‘
Vvar(bi)
Ifthis t is larger than the tabled level of t^ with (n - k - 1) degrees of

freedom, then bjis significantly different from zero at the a level of
probability. The confidence limits on b^re given by
(4.35) L = + t
a s^cT7
where isread from the “t” table with (n - k - 1) degrees of freedom. 5 1 *
There is then a probability of (1 - a) that the value of /3 ± actually lies

within these limits.The value of t to test the significance of the differ-
ence between two of the sample regression coefficients is given by
bi - b j
(4.36)
VVar(bi - bj)
where t a is obtained from the “t” table with 2(n - k - 1) degrees of

freedom.
^Methods of deriving confidence limits for such estimates as the location of the maxima,
etc., of a fitted functionare given in Williams, E. J. Regression analysis. John Wiley and
Sons, Inc., New York. 1959. Pp. 90-116.
118 DATA ANALYSIS FOE ESTIMATION
Variances and confidence intervals for predictions
Using the estimated regression equation 4.2, estimates of Y may be

made. Such estimates will be valid so long as they do not involve levels
of the independent variables outside the range found in the data used to
derive equation 4.2. No estimates of the production parameters are
available outside of this range. That this is so is explicitly shown in
equation 4.24; the matrices C and G relate only to the range of data ex-
amined. If extrapolations are made beyond this range, they must be in-
terpreted extremely cautiously.
Two types of estimate of Y may be made. Consider the case when
it is desired to estimate the average Y that might occur from many
trials with a given set of values of the independent variables. In such

circumstances, the variance of Y is given by
(4.37) Var(Y) = (J +
xn
I c iiXi + 2
J
S 2 c^xAs
J J/
2
.
i=i i=i j=i
It is noteworthy that the variance of Y depends on the actual level at

which the independent levels are set. JThe larger the deviation between
the level of X* studied and its mean, the greater will be the vari-
,
ance of Y. These comments also apply for the other type of prediction
of Y: the estimate of a particular Y from a given set of values of the
independent variables. The variance of this prediction of a specific Y
value, denoted Var(Y *), as might be expected, is larger than for the
first type of prediction, as is shown by comparison of equations 4.37
and 4.38.
(4.38) Var(Y*) = (l
'
+ -
n
+ 2
i=i
^+2? i=i
2 c-oqx^s
j*i
J J
2
'
The fact that Var(Y) depends upon the level of X* (i = 1, , k),

coupled • •
•
with the aim of minimizing Var(Y), has some important implications

for experimental design relevant to production function estimation, as
is indicated in Chapter 5.
Confidence intervals for Y and Y* are given by Y± t# VVar(Y) and
Y* ±t a Vvar(Y*), respectively, where t a has (n - k - 1) degrees of

freedom.
Coefficient of multiple determination
The correlation between the n observed values of Y and the corre-

sponding Y values is shown by the coefficient of multiple correlation .
2
It is denoted by R. The coefficient of multiple determination R indi- , ,
cates the percentage of the variation in the n observed Y values that is

explained by the fitted regression equation 4.2. Thus it is a measure of
j
the goodness of fit of the estimated regression equation. As such, it is

of use in choosing between alternative single equation models of a pro-
duction process. Of course, this is only true in the absence of strong a
priori evidence as to what exact form the model should take. It can be
shown that
2
s yf - | (Yi - Yi )
jri J=1
(4.39) R2 =
S yij
k n
s bi S Xij y
i=i j=i
(4,40)
?=i y/
j
When the number of parameters to be estimated is large or, as

often happens in production function estimation, the number of sets of
2
observation is small, the above calculations tend to overestimate R .
2
To take account of this, R , the adjusted coefficient of multiple deter-
6
mination may be used. It is calculated as
,
(4.41) R2 = 1 - (1 - R2 ) ({*-=-£) .
In equation 4.39, the denominator is the corrected sum of squares of

the n values of Y. The numerator is the amount of the corrected sum
of squares of Y accounted for or explained by the fitted regression
equation. As is indicated in equation 4.39, this sum of squares attribut-
able to regression (SSR) equals the total corrected sum of squares
minus the sum of squares of the deviations of Y from Y. If equation 4.2
fitted the data perfectly, i -
Y %
2
would be zero and R would be one.
Analysis of variance for multiple regression
The relationships between the sums of squares can be shown con-

veniently in an analysis of variance table, as illustrated in Table 4.1.
The mean -square values are obtained by dividing the sums of
squares by their associated degrees of freedom. The value of the “F*
statistic given by
Regression mean square SSR/k
(4 42) *
F __
Error mean square "

s
2
6
See Ezekiel, M. and Fox, K. A. Methods of correlation and regression analysis. John
Wiley and Sons, Inc., New York. 1959. P. 300.
, .

where F has k and (n - k - 1) degrees of freedom, provides an over-all
test of the significance of the regression. Equivalently, it provides a
test of the null hypothesis that /3 ± = /3 2 = . . .
= ^= 0. Thus if F from
equation 4.42 is larger than the tabled value of F at the desired proba-
bility level, the null hypothesis is probably not true.
Curvilinear regression with linear parameters
So far in this chapter, concern has been with the estimation of the
production function in equation 4.1 which was linear in both its variables
and parameters. Least squares multiple regression may also be ap-
plied directly to estimate the parameters of production functions non-
linear in the variables; the only proviso being that the parameters enter
linearly. If they do not, some indirect or “ round about” estimation pro-
cedure has to be used. Assuming that the parameters enter linearly,
all that is necessary is that some functional transformation of the vari-
ables be available which will transform the production function model
into a regression law linear in the transformed variables. To date, two
types of transformations have been used extensively in empirical work:
a logarithmic transformation for multiplicative Cobb-Douglas type pro-
duction models; and a simple “relabeling” of the variables in polynomial
type functions.
Consider the ordinary Cobb-Douglas production function for an out-
put P from inputs Zx ,
Z2 ,
• .
. Zk It is
2
(4.43) P = aZ^Zz ••• zf
k q
where q a proportional error due to the exclusion of some relevant

is
inputs. By expressing each
of the variables, and also a, in terms of
logarithms (either natural or to the base 10), a linear function is ob-
tained. Thus equation 4.43 becomes
(4.44) log P = log a + ^ log Z^ + • • •

+ % log Z k + log q.
If log P is denoted by Y, log a by /3

0 ,
log Z * by and log q by e ,
equa-
tion 4.44 is identical to equation 4.1. To wit:
(4- 1 ) Y = j3
0 + jS, Xx +
• • •
+ j3 k Xk + e.
Thus, while equation 4.43 represents a curved surface in the (k + 1) di-

mensional space with axes P, Z x , • Zk, equation 4.44 represents a •
. ,
flat surface or plane in the (k + 1) dimensional space with axes log P,

log Z x •••, log Z k
, .
As an example of a production function postulated as a polynomial,

consider the following function of second degree in two explanatory in-
puts:
c
Table 4.1. Analysis of Variance for Multiple Regression
Degrees of
Source of Variation in Y Freedom Sum of Squares Mean Square
Regression k SSR SSR/k
2 2
Deviations from regression n - k - 1 2 (Y - f ) s
J
=1
n
Total n - 1 s y,*
j=i
(4.45) Y = 0O + + /3
2
Z2 + /3
U ZJ + 0 22 Z22 + ^ 12 Z l Z 2 + e.
By a simple relabeling of the terms involving Z, it becomes
(4.46) Y = j
8
0 + /3.X, + + £3 X3 + /3 4 X4 + j3 5 X5 + €
where, for instance, /3

4 X 4 corresponds to the quadratic effect P 22 Z\ and
P 5 X 3 to the interaction effect P 12 ZfL 2 . While equation 4.45 denotes a
curved surface in the three dimensional space with axes Y, Z , and Z ,
x 2
equation 4.46 specifies a plane in the six -dimensional space with axes
Y,X t ,X a> -,X 5 .
When transformations
of the above type have been carried out, care
must be taken in interpreting the estimated regression coefficients.
Thus if the derived estimate of equation 4.46 is
(4.47) Y - b 0 + bjXjL + • • •
+ b 5X
(4.48) = b0 + b^ + + b gZ X Z 2
then the partial derivative of Y with respect to X is meaningless. X

x
cannot vary with X 2 •, X 5 held constant; because of equation 3^5^*.
,
• •
when X x varies, X 3 and Xg must also vary. However, we cahêrive *
- J * “
partial derivatives from equation 4.48 as shown in equ^fm Hon. 4
4 49 .
r
(4.49)
6 Y b i + 2b Z + b Z
6Z 1 s i 5 2
Orthogonal polynomial transformations

J
/— -ft..*;.-
First, a word of explanation is necessary. Two polynomfzdîiay^" ’
I* and i|
2 , where
2
(4.50) = a0 + ax X + a2X + • • •
+ ak Xk
(4.51) |2 = b 0 + bjX + b2 X2 + • • • + bk X k
.

are orthogonal if
k
(4.52) S aibi = 0.
i= o
Equivalently, and £ 2 are uncorrelated. The latter fact makes the

orthogonal transformation of polynomial functions useful for regression
analysis, at least under certain circumstances. These circumstances
are that (a) the levels of each explanatory variable be equally spaced;
and (b) there are an equal number of observations for each set of levels
of the explanatory variables. Obviously, these restrictions will only be
met if specialized techniques of data collection are used. Such tech-
niques will usually only be possible under experimental conditions. If
the restrictions are not met, the use of orthogonal polynomials is still
possible — but not conveniently enough to be worthwhile.
Consider the k degree polynomial production function in a single
explanatory variable
(4.53) Y = 0O + j3
x
X + P2 X2 + • • •
+ /3
k Xk
Further suppose that n observations of different X values and corre-

sponding y values are available. The orthogonal polynomial transfor-
mation of equation 4.53 consists of a function of the form
(4.54) Y = ot
0 + al |1 + a2 12 + • • • + ak | k
where a 0 equals Y and the other ol’s are known functions of the j3 ’s of
equation 4.53. The £ ’s are orthogonal polynomials in (X - X). In par-
ticular, corresponding to X 1 of equation 4.53 will be a polynomial of
degree i (X - X) Moreover, there will be a set of n
in . values, £- ,
(j
= 1, 2, ,
.
n), corresponding to the observed values X. , (j =1, 2,
*
•
7
• • n), of X.
• , For equally spaced levels of the explanatory variables
or input factors, tables of ^ values are available. 8 They do not have to
be computed by the researcher.
Since any pair of the i-’s are uncorrelated, their sums of cross
products are zero. Thus the normal equations for equation 4.54, cor-
responding to equation 4.15, are:
7
Simple numerical examples of the use of an orthogonal polynomial transformation are
given by Ostle, B. Statistics in reserarch. Iowa State University Press, Ames. 1956.
Pp. 142-46; and Davies, O. L. (ed.), Statistical methods in research and production. Oliver
and Boyd, London. 1957. Pp. 245-49.
8
Such tables, together with illustrative examples of the use of orthogonal polynomials,
are to be found in Pearson, E. S. and Hartley, H. O. Biometrika tables for statisticians.
Cambridge University Press. 1958. Voi. 1, Pp. 91-95. Also Anderson, R. L. and House-
man, E. E. Tables of orthogonal polynomial values extended to n = 104. Iowa State Univer-
sity Research Bui. 297. Ames. 1942.
^ S Sij
a2 L | 2j
(4.55)
ak ^ ^kj
~ ^ ^kj yj
where the a’s are estimates of the at’s of equation 4.54. Since all the
cross product terms have vanished, each equation of system 4.55 may
be solved independently. Thus
n
S Hj y j
j=i
(4.56) for i = 1, 2,
• • •
,
k.
n 2
£.
P *U
.
Also, again because the |’s are uncorrelated and lead to independent
estimates of the a’s, the estimate of a i of equation 4.54 remains un-
altered when variables are added to or taken out of equation 4.53. For
instance, if a term in X k+1 were added to equation 4.53, direct estima-
tion procedures would involve recalculating all of the estimates of the
Fb. By using the indirect orthogonal polynomial approach of adding
lk+i to equation 4.54, the estimate of a k+1 can be calculated without
having to recompute the estimates of a 0 , oq , • • , ak . It is then a •
simple matter to derive the new set of sample regression coefficients

estimating j9 0 , /3X , • •
, /3 k+1
• .
The orthogonal polynomial transformation can be applied to poly-

nomials in more than one explanatory variable. All that is necessary
is a simple extension of the procedures outlined above. For example,
suppose it was desired to estimate the following second degree poly-
nomial in three variables.
(4.57) Y = /3 O +0 X 1 1 + /3 2 X2 + $3 X3
**
0 lA "**
0 22^2 ^ ^ 33 X 3
+ $ 12 X X2 +
1 $ 13 X X3
1 + $23-^2-^3
The required transformation would be of the form:
(4.58) Y = ao +
«i li + ^1 + ^3 wi
+ °ui ^2 + ^22 ^2 + ^33 ^2
^12 H 1^1 )
+ ^13 l^l) * ^23 C 7? 1 ^1 )
124 data analysis for estimation
where a 0 equals Y and the other a’s are known functions of the 0’s and
the |’s, r]’s and w’s are pairs of orthogonal polynomials in (X x - X 3 ),
(X 2 - 2 ) and (X 3 - X 3 ), respectively.
X
As shown by equation system 4.55, the sample regression coeffi-
cients in an orthogonal polynomial regression are independent. This
fact makes it possible to use analysis of variance directly to test the
significance of the contribution to the sum of squares due to regression
of each term of the transformed equation; and hence of individual terms
of the original equation. For instance, analysis of variance relevant to
the estimate of system 4.58 may be carried out as shown in Table 4.2
in which the a’s are estimates of system 4.58. It should be contrasted
with the analysis of variance possible for ordinary multiple regression
as shown in Table 4.1.
As shown in Table 4.2, the mean square for regression can be par-
titioned into a component for each effect when orthogonal polynomials
are used. The significance of each of these individual effects may be
tested by an F test. Thus, from Table 4.2, the significance of the lin-
ear interaction between X x and X 3 may be ascertained from equation
2
4.59, s being the mean square of the deviations from regression.
af 3 S
(4.59)
Ifthe value of F from equation 4.59 is larger than the tabled F value
with one and (n - 10) degrees of freedom at the desired probability
level, the X x X interaction is significant. Similar tests apply to the
other effects. It should be mentioned that, reading down the analysis
of variance table, nonsignificant lower order effects may precede sig-
nificant higher order effects. In the estimation of production functions,
therefore, the significance or nonsignificance of lower order effects
provides no guide as to the possible role of higher order effects and
interactions.
Multiple Regression Analysis with Nonlinear Parameters
So far, least squares multiple regression has only been considered

as applies to functions linear in the parameters. Of course, such a
it
function may be a transformation of a production function that is non-

linear in the parameters to be estimated. For instance, a logarithmic
transformation of a Cobb-Douglas function leads to linearity in the pa-
rameters (as well as in the transformed variables). Nothing has been
said of the estimation of production functions nonlinear in the parame-
ters for which no suitable transformation to linear form is available.
The commonest group of such production relations are those known to
biological researchers as “growth functions.” Their common feature
-
Table 4.2. Analysis of Variance for a Second Degree Orthogonal Polynomial

Multiple Regression in Three Explanatory Variables
Source of Variation mY Degrees of Freedom Mean Square
Regression 9 SSR/9
Linear effect of X 1
1 4 s iij
Linear effect of X 2
1 *a SnJj
Linear effect of X3
1 4
Quadratic effect of X L
1
4= €!j
Quadratic effect of X2 1
4s vij
Quadratic effect of Xg 1 ^33 ^ W2j
Linear Xx by linear X2 1
Linear X L by linear X3
1
4 S Ui«i)j
Linear X 2 by linear X3 1
4 s “i)/ (Vt
2
Deviations from regression n - 10 S (Yj - Yj ) /(n - 10)
Total n - 1 sy
;
9
is thatthey are exponential relationships of one type or another.
Growth functions that have been used empirically are the Spillman
Mitscherlich functions of the form
_/3x
(4.60) Y = A(1 - 6e ) ,
the Gompertz function

_i3x
(4.61) log e Y = \(1 - ,6e ) ,
and the logistic growth function
(4.62) Y = :o— .
1 + 5e~
fc
Production function researchers will, in general, not be interested in

these functions with only one explanatory variable. To be useful, pro-
duction functions should include a number of explanatory variables.
Extension of the above functional forms to include k explanatory inputs
leads to production functions of the following types.
9
An interesting account of the derivation of growth functions is to be found in Hald, A.
Statistical theory with engineering applications. John Wiley and Sons, Inc., New York.
1952. Pp. 658-62.
X2
(4.60.1) Y X(1 - 5 ie
-i3lXl
)(l - 5 2 e'^ )
••• (1 - 6
k e^ kXk )
(4.60.11) Y X(1 - - 6 2 e'^

2X2
8 k e"^
kXk
)
(4.62.i) Y
+ 6 1 e”^ lXl + 5 2 e~Pz
X2 ”^ kXk
1 + • • •
+ ê
(4.62. ii) Y
! + se-'
3
^! -
^k x k
For functions such as these, no simple transformations to a form lin-

ear in the parameters, ^9 P ±9 > & k’
are available. Indeed, iter-
* *
*
ative or successive approximation procedures for estimating such func-

tions are largely unexplored. An exception is an iterative procedure
for estimating Spillman -Mitscherlich functions such as equation 4.60.L
10
that has been developed by Ibach. His approach utilizes a set of
standard values that markedly reduce computational procedures. A
number of possible approaches to the estimation of logistic functions
involving a single explanatory variable have been summarized by Nair. 11
In general, experience to date indicates that cumbersome iterative pro-
cedures, as have to be used for most exponential functions of the above
types, are not desirable. Especially is this so when account is taken of
the fact that research resources are not limitless. An additional disad-
vantage is that the estimates obtained by such iterative procedures are
not readily amenable to statistical significance tests. Since any pro-
duction surface can be fitted reasonably well by an easily calculated
polynomial -type function, there seems little justification for persever-
ing with functions requiring complex iterative procedures.
Functions of the following algebraic form,
(4.63)
a + 3X
1
1 + 6X ’
have also sometimes been suggested for production functions. To fit

equation 4.63, which is nonlinear in both its variables and parameters,
the method of internal least squares regression developed by Hartley
may be used . 12 However, computational procedures become extremely
complicated when such algebraic functions are extended to more than
one explanatory variable, as in “resistance functions” like equation 4.64.
10
Ibach, D. B. Use of standard exponential yield curves. ARS. 43-69. USDA, Washing-
ton, D. C.1958. Also, A graphic method of interpreting response to fertilizer. Handbook
No. 93. USDA, Washington, D. C. 1956.
11
Nair, K. R. The fitting of growth curves, in Kempthorne, O. et at. (eds.), Statistics
and mathematics in biology. Iowa State University Press, Ames. 1954. Pp. 119-32.
“See Hartley, H. O. The estimation of non-linear parameters by “internal least
squares.* Biometrika, 35: 32-45. 1948.
a.
~
(4.64) = a0
\ + 0 A 52 + ^ 2 X2 + ^kX k
As for the exponential type functions, the computational disadvantages

of equation 4.64 outweigh any advantages resistance functions may have
as logical expressions of the production process under consideration.
For this reason, it will be assumed In the following discussion that if a
single equation approach is taken, the regression function being esti-
mated is linear in the parameters. The restriction on the algebraic
form of the production function is that it be capable of being trans-
formed to a form linear in the parameters.
Statistical Problems in Least Squares Estimation

of Production Functions
Early in this chapter, the assumptions of the least squares approach

were listed. If these assumptions hold true, least squares estimates of
the production function parameters will be unbiased and of minimum
variance. In general, the assumptions will not be satisfied exactly.
Problems of estimation and interpretation then arise. Some of these
difficulties are discussed here, insofar as they commonly arise in pro-
duction function research. It should be emphasized, as a rider to the
following discussion, that statistical techniques, no matter how "high
powered,” cannot be used to derive satisfactory estimates from bad
data.
Errors of observation in the independent variables
In the prior discussion of multiple regression, it was assumed that

the data contained no errors. The only error allowed was an error in
the equation due to the omission of some input factors. Of course, for a
model such as equation 4.1 to be satisfactory, any excluded inputs
should only be of minor importance. In point of fact, the assumption of
no errors whatsoever in the measurement of the input variables is ex-
tremely strong. Errors of observation, of some magnitude, will always
be present. Such errors may arise for a variety of reasons, including
the obvious one that a continuous variable can rarely be measured with
absolute exactness. Other causes of observational errors are the
human element involved— for instance, mistakes may occur in the col-
lection and transcription of data -
and, most importantly, the fact that
in many situations a satisfactory conceptual basis of measurement is
not available. The latter, of course, is the essence of the problem of
aggregation. It arises whenever the input or output under consideration
is not homogeneous, either within or between observations. Insofar as
aggregation is a problem of economic specification, it will be discussed
in Chapter 6. However, some explanation of the problem is necessary
here. A typical example of the aggregation problem occurs in the
measurement of land, labor and capital, say, as inputs for the deriva-
tion of a farm production function from cross-sectional data. Varia-
tions in quality and mode of use will occur in all these variables from
one observation to another. Consider the land input. Not only may
there be variations in soil type within the one observation and between
the cross-sectional observations, there may even be variations in
quality within the same soil type. To measure the land inputs simply
in terms of acres would ignore these differences in quality. What is
needed, in such circumstances, is some basic unit in which the obser-
vations may be measured so that they are strictly comparable. To at-
tempt to formulate such a basic unit involves conceptual difficulties of
a high order, as the reader will quickly realize should he try to solve
the problem. Similar difficulties arise with regard to any input or out-
put category that is not strictly homogeneous both within and between
observations. If the observed values of the variables are not comparable
because of a lack of homogeneity, use of the raw observations means
that observational errors are necessarily present in the data. If some
method of adjusting the data to take account of heterogeneity is used,
the adjusted data will contain observational errors insofar as the ad-
justment process or aggregation procedure is imperfect. Moreover, in
general, no aggregation procedure will be perfect because of the con-
ceptual difficulties involved. On the other hand, if the relevant varia-
bles are homogeneous, the degree to which errors of observation occur
is simply a question of the mechanics of data collection. Provided suf-
ficient care is exercised, mechanical errors of observation can be re-
duced to negligible proportions. On these grounds, in fitting empirical
production functions, the most important source of errors in the ob-
servations is the lack of homogeneity.
Given the fact that observational errors will generally be present in
the data, the researcher may take a number of approaches. If the er-
rors are thought to be negligible — averaging no more than 1 or 2 per
cent, say, and with no gross errors present — then the ordinary least
squares multiple regression approach may be used. If the errors can-
not be regarded as negligible, the alternative approach of weighted re-
gression may be used. However, before discussing this technique,
mention should be made of a procedure sometimes used to determine
whether or not the observational errors may be regarded as negligible.
It is known as confluence analysis.
Confluence analysis
Briefly, the steps involved in confluence analysis are as follows.

Suppose equation 4.65 is the production model being examined, Xx being
the output, X2 and X 3 the inputs.
(4.65) Xx = +0 2x 2 +^3X3 + e
The sample regression coefficient of X 1 on X 2 is b 2 of equation 4.66. It

is obtained by excluding X 3 from the analysis and minimizing the stun
of squares in the direction of X r The latter procedure implies the as-
sumption that the only errors of observation are in X1 .
(4.66) X, = b
,
+ b2 X 2
If the stun of squares is minimized with respect to ,

again excluding
Xs ,
an equation such as 4.67 results.
(4.67) X 2
= c0 + c.X,
Ifequation 4.67 is put in the same form as equation 4.66, as shown in

equation 4.68
1
(4.68) Xi = -c 0 cl + cf X 2
be regarded as the sample regression coefficient of X x

1
then cf may
on X 2 when errors of observation are in X 2
all the .
The procedures typified by equations 4.66, 4.67, and 4.68 are then
carried through in similar fashion for the set X^ y and X 3 the sum
, ;
of squares being minimized in the direction of each variable in turn.

The next step is to convert each of these “estimates* of b 2 to the stand-
ardized partial regression coefficient form. The procedure for doing
this is given by equation 4.26. The standardized coefficients of X x on
X 2 obtained with X 3 excluded, are then compared with those obtained
,
with X 3 included. For convenience, this comparison is usually made

graphically by way of bunch maps Consider first the situation when
.
the inclusion of X 3 leads to the standardized coefficients becoming

more similar, i.e., a bunching together. It may then be concluded that
the effect of any observational errors in X 3 is negligible. Also, that X 3
is a “useful* variable to include in the model. On the other hand, the
inclusion of X 3 may lead to an explosion of the standardized coefficients.
The researcher then has an indication that if any errors are present in
X 3 , they will probably induce nonsensical estimates of the parameters
of equation 4.65. In such circumstances, it would be wise to exclude X 3
from the model — unless the logic underlying the production process
strongly supported its inclusion.
For guidance as to whether the inclusion of X 2 would be useful or
detrimental, the same analysis as above must be carried through; the
roles of X 3 and X 2 being interchanged. In similar vein, a model con-
taining k inputs can be checked. 13 However, as k increases, the
13
Detailed accounts and illustrations of the method are to be found in Frisch, R. Statis-
tical confluence analysis by means of complete regression systems. Publication No. 5.
Oslo University Institute of Economics. 1934; and Stone, R. The measurement of con-
sumers’ expenditure and behaviour in the United Kingdom, 1920-1938. Cambridge Univer-
sity Press. 1954. Vol. I, pp. 300-1 and 342 ff. For an application oi confluence analysis in
the fitting of production functions, see Antill, A. G. Towards a production function for dairy
farms. The Farm Economist. 8: 1-11. 1955. Also, Valavanis, S. Econometrics.
McGraw-Hill, New York. 1959. P.146.
number combinations of variables that must be examined increases
of
rapidly.The necessary computations quickly become immense. The
method also has other disadvantages. One, perhaps the most important,
is that confluence analysis is a subjective technique, there being no ob-
jective manner of deciding whether the standardized partial regression
coefficients are sufficiently similar to imply negligible errors. Nor
whether an explosion really implies nonnegligible errors. Also, it as-
sumes that no relevant variables have been omitted from the model.
Obviously, in most production function research, this latter assumption
will not be true.
Weighted regression
With errors of observation present, each of the observed values X^-
(i = 1, 2, • •
*
,
k; j
= 1, 2,
• n) may be considered to be made up of a
•
•
,
systematic component M*j which is the true quantity of the variable, and
an error of observation 77^. Thus
(4.69) Xij = My + 77ij

.
Under the assumptions that (a) there are no relevant variables omitted
from the single equation model being used, i.e., no errors in the equa-
tion, and (b) the errors 77 are noncorrelated, random, and normally
.
j
distributed the method of weighted regression may be used. The essen-
weighted regression technique is that the variances
tial feature of the
and covariances of the observational errors are used as weights in es-
timating the regression coefficients. By the nature of the problem, the
errors and their variance -covariance matrices are generally unknown.
These variances and covariances have to be estimated in some fashion
— either from past experience, by guessing, or by an iterative pro-
cedure.
Adetailed exposition of the technique of weighted regression, and
some examples of its application, has been given by Tintner. A varia-
tion of the method, in a production function context, is illustrated in
14
some work of Antill.
There are a number of disadvantages to using weighted regression.
The computations involved are cumbersome. It may be difficult to ob-
tain satisfactory estimates of the variances and covariances of the er-
rors. The assumptions that the observational errors are noncorre-
lated, random and normally distributed may be just as false as the
assumption of no observational errors. Indeed, when the observational
errors arise because of an imperfect aggregation procedure, it is
rather unlikely that these errors are normally distributed and arise at
random. In addition, the assumption that there are no errors in the
equation, i.e., that no relevant variables have been left out, will not be
14
Tintner, G. Econometrics. John Wiley and Sons, New York. 1952. Pp. 121-43; Antili,
G. A. Towards a production function for dairy farms. The Farm Economist, 8:1-11. 1955.
true in general. Against the latter criticism, it must be noted that no

method has so far been developed by which account may be taken of both
errors in the variables and errors in the equation. Still, given the
complexity of the weighted regression approach and the assumptions
that it makes, it may be that the use of weighted regression raises
more problems than it solves. The researcher who does not have neg-
ligible errors of observation in his data might be best advised to seek
better data rather than try to develop reliable estimates out of unrelia-
ble data.
Autocorrelation
By autocorrelation is meant correlation between successive items

in a time series of observations. Since time series data are frequently
used in production function estimation, the difficulties raised by auto-
correlation are pertinent to this discussion. Two facets of the problem
may be distinguished: autocorrelation in the observations on one or
more of the variables, say Y, X2 ,
X k of equation 4.1, and auto-
,
• -
•
correlation of the errors e., (i = 1, 2, • . , k; j = 1, 2, •

. ,
n) of equa- . .
15 J
tion 4.1.
Autocorrelation in the variables
If autocorrelation is present in the observations of one or more of

the variables, the estimates of the regression coefficients obtained via
least squares are still unbiased and of minimum variance. However, if
the variable is autocorrelated, the variance of its regression coeffi-
cient bi will be affected so that equation 4.29 does not hold true. In
consequence, there is no need to worry about autocorrelation among the
variables unless it is desired to calculate confidence limits and signifi-
cance tests for the regression coefficients. Such calculations are gen-
erally necessary in production function research. Hence any time se-
ries observations to be used in estimating a production function should
be tested for autocorrelation unless the observations are obviously in-
dependent. Some explanation of “obviously independent” is required.
Suppose 40 readings of hog weight, each reading taken from a different
hog at a different point in time, are to be used in estimation. There
would be no reason to expect these 40 readings to be autocorrelated. On
the other hand, it is very likely that 40 readings of hog weight at differ-
ent points in time, all made on the same hog, would be autocorrelated.
The reason, of course, is that the readings are not independent - they
all relate to the same hog.
The simplest test for autocorrelation is the von Neumann ratio test
The first step is to calculate the ratio a 2 /s 2 where, for the n observa-
,
tions on the variable Xj

15
See Anderson, R. L. The problem of autocorrelation m regression analysis. Journal
of the American Statistical Association, 49: 113-29. 1954. Also, Beach, E. F. Economic
models: an exposition. John Wiley and Sons, Inc., New York. 1957. Pp. 176-80.
The calculated von Neumann ratio is then checked against tabled values
of the ratio, as presented by Hart,
to see if the series is significantly
autocorrelated.
If autocorrelation is present in the observations of Xi , the variance
of bi may be calculated by the following modification of equation 4.29
17
due to Wold.
2 2
(4.72) Var(bi) = c ijL
s f
2
where f is estimated by the first few terms of
2
(4.73) f = 1 + 2r
il
P 1 + 2r
i2
P 2 + ..- + 2r in Prf i= 1,2,.-., k .
The rxj andPj (j = 1, 2, • , n) values of equation 4.73 are the noncir cu-
•
•
ter autocorrelation coefficients for Xj and the errors, e of equation 4.1

18
estimated by e of equation 4.3, respectively. Approximate tests of
significance and confidence limits may then be calculated as in equation
4.34 by using the estimate of the regression coefficient’s variance ob-
tained from equation 4.72. Little reliability can be attached to these
procedures if the number of observations, n, is less than 30. More-
over, equation 4.72 assumes that the regression is strictly linear in the
variables. If the regression is curvilinear, that is, nonlinear in the un-
transformed variables, the procedure is only approximate. Insofar as
the majority of production functions are nonlinear in the original varia-
bles, this is unsatisfactory.
An
alternative approach may be taken to the problem of autocorrela-
tion the autocorrelated series is replicated. The procedure is based
if
on the fact that the effect of autocorrelation is to, in a sense, reduce

the number of effective observations to which the regression function is
fitted. In other words, the number of degrees of freedom used in enter-
ing tables for tests of significance relevant to an autocorrelated variable
16
Hart, B. I. Significance levels for the ratio of the mean square successive difference
to the variance. Annals of Mathematical Statistics. 13: 445-47. 1942.
17
Wold, H. Demand analysis. John Wiley and Sons, New York. 1953. Pp. 44 and 209-12.
18
The rather complicated formulae by which these autocorrelation coefficients may be
estimated, together with a detailed discussion of autocorrelation, are to be found in Tintner,
G. Econometrics. John Wiley and Sons, New York. 1952. Pp. 240-55, esp. 243. See also
Ezekiel, M. and Fox, K. A. Methods of correlation and regression analysis. John Wiley and
Sons, Inc., New York. 1959. Pp. 325-47.
is smaller than it would be if the observations of the variable were not

autocorr elated. Procedures are available for approximating the effec-
19
tive number of degrees of freedom In autocorrelated series. How-
ever, the necessity of calculating the autocorrelation coefficient and
approximating the effective number of observations may be avoided.
How? By basing the tests of significance on the minimum number of
effective observations to which the series could be reduced by autocor-
relation. To give an example, 20 observations of weight taken on one
hog could, through strong autocorrelation, be equivalent to a minimum
of one effective observation. However, a series of 100 hog weight ob-
servations made up of 20 observations on each of five hogs could, at
worst, be reduced to five effective readings since the experimental
units (hogs) are independent. If significance tests of the regression co-
efficients calculated in the ordinary fashion of equation 4.34 but with
degrees of freedom based on the minimum number of effective observa-
tions are significant, the null hypothesis that the regression coefficient
is zero may be rejected. Should the test be nonsignificant, the null hy-
pothesis cannot be accepted without further testing. The test must then
be conducted on the basis of the actual number of observations, disre-
garding the presence of autocorrelation. If the test is still not signifi-
cant at an acceptable probability level with the greater number of de-
grees of freedom, the null hypothesis that the regression coefficient is
not significantly different from zero may be accepted. The effective
number of observations need be estimated only if the tests are nonsig-
nificant on the basis of the minimum effective number of observations,
but significant on the basis of the actual number of observations taken.
Autocorrelation in the errors
K the errors e j , (j =1, 2, ,

•
n), relevant to equation 4.1 are auto-
•
•
correlated, the least squares assumption that E(ei e ) = 0, i / j, does

j
not hold true. Under such circumstances, ordinary least squares esti-
mates of the regression coefficients will not be unbiased, nor will they
have minimum variance. However, if the variances and covariances of
the errors are known, a modified least squares procedure leading to
20
suitable estimates may be used. Insofar as the error variances and
covariances usually have to be estimated from the data, or guessed, this
adaption of least squares may not, in practice, lead to reliable esti-
mates of the production parameters. In general, the errors might be
expected to be autocorrelated when (a) the algebraic form of the esti-
mated production function is not very satisfactory in terms of fitting the
data, (b) there are errors of observation in the data, and (c) important
explanatory variables have been omitted from the analysis. In estimat-
ing production functions from nonexperimental time series data, all
19
See Orcutt, G. H. and James, S. F. Testing the significance of correlations between
time series. Biometrika, 35: 397-413. 1949.
20
See Tintner, G. Econometrics. John Wiley and Sons, New York. 1952. Pp. 279-283.
three of these causes of autocorrelation may occur. However, there is
no reason why these sources of error autocorrelation should not be
minimized under experimental conditions, especially if care is taken in
designing the experiment and collecting the data. Insofar as most agri-
cultural production function estimation involving time series uses ex-
perimental data, the problem of autocorrelation of the residuals is not
too frequent. A number of tests for testing if the errors are autocor-
related are possible. Perhaps most useful is that developed by Durbin
and Watson. 21
It should be pointed out that the lack of independence — known as
autocorrelation in time series data — between observations of a given
variable, and between the errors in the regression model, may also oc-
cur between observations adjacent in space. To guard against this is
one of the reasons why randomization procedures are adopted in de-
signing experiments. Should spatial observations be “autocor related”
— which may often occxir with nonexperimental data — little account can
be taken of the lack of independence since, while time series are gener-
ally equally spaced in time, spatial data are usually irregularly spaced
in terms of distance between observations.
Multicollinearity
By multicollinearity is meant the situation when there is more than

one relationship among the variables being considered. For purposes
of exposition, two types of multicollinearity may be distinguished. The
first relates to the problem of whether the single equation model is
valid or whether the production function should be estimated as one of
a number of simultaneous relations. Since this question is largely one
of economic specification, it is deferred until Chapter 6. The other
type of multicollinearity relates to the single equation model when two
or more of the independent variables are so highly correlated that there
is, for all intents, one or more linear relations between some or all of
the independent variables. Such cases of multicollinearity are statisti-
cal in emphasis and will be considered here. They are closely related
to the problem of errors of observation in the independent variables.
As a simple example consider the problem of estimating the pro-
duction function
(4.74) Y = j3„ + + i3
2
X2 + e
when X x and X 2 are perfectly correlated so that
( 4 75 )
-
^x, = 1 =r x x
2 t
21
Durbin, J. and Watson, G. S. Testing for serial correlation in least squares regres-
sion. Biometrika, 37; 409-28. 1950.
The normal equations, expressed in terms of the partial correlation

coefficients, as in equation 4.25, are then as follows:
ru + b'2 r i2 = r iy
(4.76)
K ru + ^2 r 22 it
%
Since all the partial correlation coefficients are equal to one, the two
normal equations are identical. In other words, there is only one equa-
tion from which to estimate the two unknowns, and b 2 Obviously,
.
the values of b \ and b'2 — and hence of b x and b 2 — are indeterminate.

Now suppose that X1 and X 2 are, in truth, perfectly correlated but be-
cause of observational errors in the data, the partial correlation be-
tween the observed values of Xx and X 2 happens to be 0.98. The normal
equations would be:
b\ + 0.98b'2 = r xy
(4.77)
0.98b'x +b'2
Since these two normal equations are not identical, they could be solved
for b[ and b'2 . However, these values of b[ and b' 2 would have no rele-
vance in terms of the production process in equation 4.74. Being
merely the consequence of errors in the data, they are meaningless.
Referring to equation 4.69, it can be said that the sample regression
coefficients computed from equation 4.77 are determined completely by
the unsystematic errors of observation The estimated coefficients
.
do not reflect a relationship between the systematic portions of the My

observations on the variables.
A geometric interpretation of multicollinearity is also possible.
Suppose Y, X± , and X2 were measured without error, X x and X2 being
perfectly correlated. Plotting the observed values of Y against the ob-
served values of X x and X 2 would give a straight line in the three-
dimensional space with axes Y, Xl9 and X 2 Since an infinite number of
.
planes can satisfy the condition of passing through one straight line, re-
gression analysis using such data would lead to indeterminate results.
However, since the observations will always be subject to some error,
plotting the observed Y values would not give a straight line. The re-
sult is an apparently determinate regression plane.
If, instead of there being one or more exact linear relationships
among the independent variables, there is a high degree of correlation

between some of the independent variables, the collinearity will be only
approximate. Such a situation results in the estimates of the regres-
sion coefficients having large variances because the errors in the ob-
servations tend to play a dominant role in determining the values of the
parameters. Moreover, the estimated values of the sample regression
coefficients tend to be relatively unstable due to their sensitivity to ob-
servational errors in the data.
A number of approaches to multicollinearity are possible. Since the
correlation coefficients between the independent variables serve as in-
dicators of the possible presence of linear or near linear relations
among the independent variables, it is wise to calculate and examine all
these coefficients for the sample data. If any of these coefficients are
close to plus or minus one, say greater than 0.81 the regression
I
,
analysis should be carried through with one of the highly correlated

variables omitted. Which variable(s) to omit and which to retain should
be decided on the basis of the logic — physical, biological, or economic
— relevant to the production process being examined. If the R 2 is satis-
factory and the logic of the production situation does not dictate that the
excluded variable(s) must be included, the new regression estimates
may serve satisfactorily. The method of confluence analysis, discussed
above with regard to errors in the observations, may be of some use as
a guide to deciding whether nuisance relations are present among the
independent variables. Also, bunch maps may help in deciding whether
a particular variable should be included in the analysis or whether its
inclusion might lead to meaningless estimates of the regression coeffi-
cients. Frisch has outlined the use of confluence analysis as a guide to
22
the problem of multicollinearity . An example of the use of confluence
22
analysis in production function estimation has been given by Ant ill .
An alternative procedure is to attempt to test statistically whether

there are any other linear relations among the variables apart from
24
that postulated by the production model These test procedures are
.
rather closely related to the technique of weighted regression and may

conveniently be used when weighted regression is used. Such comple-
mentarity exists because weighted regression is aimed at overcoming
the effect of errors in the observations while the effects of multicollin-
earity, of the type discussed here, are also closely related to the pres-
ence of errors in the variables. The uninitiated reader must be warned
though that such testing procedures, like weighted regression analysis,
involve formidable computations and a number of assumptions that may
not be well founded. For instance, it is assumed that the 77 s of equa- ?
tion 4.69 are known. So far as possible, therefore, it is very worth-

while to try and overcome the possibility of multicollinearity in the
sample data — where it may be present by fortuitous circumstances —
by the use of appropriate sampling methods. Such methods are dis-
cussed in Chapter 5.
Should the researcher have a priori knowledge of the value of any of
the structural parameters, the normal equations may be made determi-
nate by utilizing these known values. Thus, referring to equation 4.76,
if b were known a priori, there would be no problem of indeterminancy.
x
The value of b 2 could be ascertained by inserting the
'
known value of b'
x
in equation 4.76 and solving for b' .
2
22
Frisch, R. Statistical confluence analysis by means of complete regression systems.
Publication No. 5. Oslo University Institute of Economics. 1934.
23
Antill, A. G. Towards a production function for dairy farms. The Farm Economist,
8: 1-11. 1955.
24
Such a test is illustrated in Tintner, G. Econometrics. John Wiley and Sons, New
York. 1952. Pp. 127-38.
Unequal error variances
One of the assumptions under which the least squares principle

gives unbiased estimates with minimum variance is that the e*s, arising
2
as in equation 4.1, have the same variance. In short, that E(ef) = o for
j
= 1,2, •
•
,
n. In production function estimation, this assumption may
•
be doubted if there is thought to be a relation between any of the inde-

pendent variables studied and the omitted variables accounted for by the
error term, e Consider the following hypothetical example. Suppose
.
the crop response function to a fertilizer mixture is being studied. It

may be that the effect of the omitted variable sunlight, measured exclu-
sively by e say, depends on some crucial enzyme whose production in
turn depends on the quantity of fertilizer applied. At low levels of fer-
tilizer, the effect of sunlight may be inhibited so that e is small and
relatively constant, having a small variance. At high levels of ferti-
lizer, because the effect of sunlight is not inhibited but varies in re-
sponse to other features of the crop, e may be large and have a large
variance. The assumption of constant error variance would, therefore,
be violated.
When it is thought that the error variances may not be constant, the
data may be partitioned into a number of sets to obtain an estimate of
the error variance in each range of the data. Comparison of these var-
iance estimates then indicates whether the assumption of constant error
variance is violated. If it is, the variance for the j-th range of the data
may be weighted by a factor, wj , satisfying equation 4.78.
2
(4.78) Wj s j = cr
An arbitrary value is selected for or 2 , say a 2 = 1, and equation 4.78

solved to determine Wj s 2 having been determined for each range of
;
the data by way of equation 4.30. Using these Wj values as weights on

the cross product sums in the normal equations, unbiased estimates of
the regression coefficients with minimum variance may be derived. An
exposition of the method and some illustrative examples are given by
Anderson and Bancroft. 25 Unlike the procedures necessary when many
of the other assumptions of least squares are violated, the above
tech-
nique for handling unequal error variances is quite straightforward.
All that is necessary is that a sufficient range of data and number of
observations be available to calculate the s 2 and Wj values.
SIMULTANEOUS EQUATIONS APPROACH

So far in this chapter, production function estimation has been con-
sidered within the framework of a single equation model. If the
25
Anderson, R. L. and Bancroft, T. A. Statistical theory in
research. McGraw-Hill
Book Company, Inc., New York. 1952. Pp. 182-86.
production relationship is in fact a unilateral causal relation with output
dependent upon a number of predetermined input variables, the single
equation model is logically appropriate. Under such circumstances, the
least squares multiple regression procedure provides the best esti-
mates of the production function parameters. However, as mentioned
at the beginning of the current chapter, the production relationship may
be but one of a number of simultaneously determined relationships in-
volving output and inputs, and other variables also, not as dependent
26
and independent variables but as mutually determined variables. If
the economic, biological, or physical logic relevant to the production

process dictates that such is the case, then, ideally, the parameters of
the production relationship should be estimated in terms of the com-
plete set of simultaneous equations in which the production relation is
embedded. Leaving questions of model formulation till Chapter 6, the
estimation procedures relevant to the simultaneous approach are dis-
cussed briefly below. No attempt is made to elaborate the mechanics
of the estimational procedures. With a knowledge of the statistical and
mathematical concepts elaborated in the prior sections of this chapter,
the researcher should be able to follow the simpler expositions of the
simultaneous equations techniques to be found in the references noted.
Also, many of the problems that arise in dealing with a system of equa-
27
tions are similar to those that occur with single equation models.
Some such problems are the choice of variables, selection of the alge-
braic form of the functions, the validity of the assumptions made and
the interpretation of results.
Types of Variables
In a model consisting of a system of simultaneous equations, two

types of variables may be distinguished. Endogenous variables are
those whose current values are determined simultaneously by the sys-
tem. Contrasting with these are the predetermined variables; their
current values may be taken as given. Predetermined variables may be
of two types: exogenous if their values are determined outside the sys-
tem of equations, and lagged endogenous variables. Which variables
are classified as endogenous and which as exogenous depends upon the
logic of the situation being examined. To be complete, a model with G
endogenous variables should consist of G equations.
26
For examples of simultaneous equation models oriented to production function estima-
tion, see the following: Marschak, J. and Andrews, W. H. Random simultaneous equations
and the theory of production. Econometrica, 12: 143-63. 1944; French, B. L. Simultaneous
economic relationships and derivation of the production function, in Heady, E. O. et al. (eds.),
Resource productivity, returns to scale, and farm size. Iowa State University Press, Ames.
1956. Pp. 97-105. Also, Jarrett, F. G. Estimation of resource productivities as illustrated
by a survey of the Lower Murray Valley. Australian Journal of Statistics, 1: 3-11. 1959.
27
See Foote, R. J. A comparison of single and simultaneous equation techniques. Journal
of Farm Economics. 37: 975-90. Also, Valavanis, S. Econometrics. McGraw-Hill, New
York. 1959.
Assumptions
The assumptions made in using simultaneous equation techniques of

estimation are rather similar to the assumptions of least squares.
There are assumed to be no errors of observation in the variables; only
errors in the equations are permitted. Like the e of equation 4.1, these
errors are supposed to arise because some relevant variables are
omitted. The expected value of the error in each equation should be
zero. Moreover, the error in a given equation should be independent of
the predetermined variables in the equation and should not be autocor-
related. Obviously, this list of assumptions is just as formidable as
those for least squares estimation. If tests of significance and confi-
dence intervals are to be derived for the estimates of the parameters,
the additional necessary assumptions are generally stronger than those
made in the least squares case.
Identification
In order to estimate the parameters of a production relation em-

bedded in a system of simultaneous equations, it is necessary that the
production equation be either just -identified or over -identified. For a
system of G equations with G endogenous variables, the following rules
of thumb can be used to decide whether a given equation is identified or
not. Let D be the number of predetermined variables appearing in the
system but not in the equation at hand; and H the number of endogenous
variables that appear in the equation being studied. The given equation
is just-identified if D = H - 1; over -identified if D > H - 1; and under -
identified if D < H - 1. The parameters of an under -identified equation
cannot be estimated satisfactorily. If an equation is just -identified or
over -identified, a number of statistical techniques are available for es-
timating the parameters.
Estimation of Just -Identified Equations
A model consisting of G interdependent equations in G endogenous

variables may, by algebraic manipulation, be written so that each en-
dogenous variable appears as a function only of the predetermined vari-
ables in the system. In such equations, known as reduced form equa-
tions, the coefficients of the predetermined variables are combinations
of the coefficients in the original equations of the model. Each reduced
form equation may be considered as a regression law, the endogenous
variable being taken as dependent upon the predetermined or independ-
ent variables in the reduced form equation. Hence the coefficients of a
reduced form equation may be estimated by least squares multiple re-
gression. If the production relation is just -identified, estimates of its
parameters can be derived from the least squares estimates of the
28
parameters reduced form equations
in the It should be noted that, in
.
themselves, the estimates of the parameters of the reduced form equa-

tions mean nothing. What is wanted are estimates of the parameters of
the original equations of the model. These parameters depict the struc-
ture — economic, biological or physical - of the situation being studied.
Hence they are known as structural coefficients. As emphasized previ-
ously, it is knowledge of the structure of the process that is pertinent
to the economic analysis of a production relationship.
Estimation of Over -Identified Equations
A number of approaches are possible if the production relation is

over -identified. From the point of view of obtaining the most satisfac-
tory estimates of the structural coefficients, the full information maxi-
mum likelihood method is best. This technique involves solving all of
the equations in the system simultaneously. As might be expected, the
procedure involves an extremely large number of complicated compu-
tations. For this reason, it has seldom been used. Instead, the single
equation limited information maximum likelihood procedure has gener-
ally been followed. Using this algorithm, each equation of interest in
the model is estimated separately; utilizing, however, the information
that is available about the role of the predetermined variables in the
other structural equations. Statistically, the limited information ap-
proach does not give estimates of the structural coefficients that are as
reliable as those obtained by the full information approach. But it is
very much simpler in terms of computational procedures; and gives,
relevant to the amount of information that it uses, estimates that are
7
29
statistically satisfactory.
Mention should also be made of a newer single equation technique,
known as the Theil-Basmann method, for estimating the structural co-
efficients in a simultaneous equations model. The procedure is rather
similar to the single equation limited information approach but is much
simpler computationally 30 .
28
The procedures involved have been explained and illustrated by Tintner, G. Econo-
metrics. John Wiley and Sons, New York. 1952. Pp. 166-72. See also: Koopmans, T. C.
(ed.), Statistical inference in dynamic planning. Cowles Commission for Research in Eco-
nomics, Monograph No. 10. John Wiley and Sons, New York. Pp. 153-237; and Hood, W. C.
and Koopmans, T. C. Studies in econometric method. Cowles Commission for Research in
Economics, Monograph No. 14. John Wiley and Sons, New York. 1953. Pp. 112-99.
29
For a simple, complete explanation of limited information estimation procedures, see
Friedman, J. and Foote, R. J. Computational methods for handling systems of simultaneous
equations with applications to agriculture. Agriculture Handbook No. 94. USDA, Washing-
ton, D. C. 1955.
30
An
explanation of the Theil-Basmann approach is to be found in Wallace, T. D. and
Judge, G. G. Econometric analysis of the beef and pork sectors of the economy. Tech. Bui.
T-75. Oklahoma State University, Stillwater. 1958. See also Basmann, R. L. The compu-
tation of generalized classical estimates of coefficients ma structural equation. Econo-
metrica. 27: 72-81. 1959.
Recursive Systems
A particular type of simultaneous equations model is that known as

31
a recursive causal chain system. Each equation in such a system
specifies a unilateral causal relationship between a dependent variable
and a number of independent variables, as in equation 4.1. The equa-
tions of the model must be read in sequence, the dependent variable of
an earlier equation entering the system again as an independent varia-
ble in a later equation. Because it contains only a single dependent
variable, each equation in such a system may be estimated by least
squares. As for the general simultaneous equations approach, whether
or not a recursive causal chain model should be used depends on the
underlying logic of the particular production process being examined.
CHOICE BETWEEN ESTIMATION PROCEDURES
research resources were free goods, the choice between estimat-

If
ing production parameters from a single equation or a simultaneous
equations model could be decided in terms of the logic underlying the
production process being examined; provided, of course, that this basic
logic is known. But research personnel, funds, equipment and time are
not free goods. Some account has to be taken of their availability. As
previously mentioned, least squares estimation is generally least ex-
pensive and simplest to carry out. Moreover, it is the logically appro-
priate technique when a single equation model is correct or when a re-
cursive causal chain model applies. Also, least squares may be used
indirectly via the reduced form equations if the production equation in a
simultaneous model is just -identified. If the production relation in a
simultaneous equations model is over -identified, least squares esti-
mates — at least theoretically — are not as good as those derived by si-
multaneous equation techniques. The commonest such techniques are
the full information maximum likelihood, limited information single
equation maximum likelihood and the Theil-Basmann limited informa-
tion procedures. For most practical purposes, the full information pro-
cedure must be dismissed as too expensive and complicated. The lim-
ited information procedures are also far more expensive than if least
squares is applied to the production relation. The practical question
that arises is whether the additional reliability of the estimates derived
by single equation limited information procedures is worth their addi-
tional expense. No general answer can be given to this question. It is
noteworthy, though, that in the United States, where the relevant re-
search resources are far less restricted than in any other country, the
majority of agricultural production function estimates have been based
on least squares. Still, that was before the advent of electronic com-
puters.
31
For exhaustive references and major recent developments, see Wold, H. Ends and
means, in Grenander, U. (ed.), Probability and statistics. Almgrist and Wicksell, Stock-
holm. 1960. Pp. 355-434.
CHAPTER 5
Data Collection
for Production Function Estimation
N OUTLINING the statistical procedures and problems relevant to

the fitting of empirical production functions, it was assumed that data
I were available. Other than noting that the techniques of data collec-
tion are in part dictated by the estimation procedures to be used, no
mention was made of the possible methods of data collection nor of the
difficulties inherent in data collection. Still, it must be strongly em-
phasized that the collection of data — whether it be by experiment, sur-
vey, or transcription — is an integral part of research aimed at the em-
pirical specification of production functions. The fitted function can
only be correct to the extent that the data behind it reliably represent
the production process. If the data are wrong, the estimated production
parameters must also be wrong. Likewise, if the data are incomplete,
only incomplete implications can be drawn from the fitted function. In
short: shabby data can only lead to shabby results.
The present chapter considers only the more general techniques
and problems of data collection. Those procedures and difficulties that
are pertinent only to a particular product or particular circumstances
will be considered later in Chapter 7 and the empirical chapters rele-
vant to particular production processes.
The primary problem of data collection is to decide which variables
should be observed. Only then does the question arise of how best these
variables might be studied and in what form the observations should be
made. So far as determining which variables are pertinent, the re-
searcher must have a fair degree of familiarity with the production
process under study. He must, at least, be sufficiently familiar with
the situation to sketch a model indicating, perhaps tentatively, the vari-
ables involved. Such a tentative model, for the single equation case,
might be written as in equation 5.1 with output merely specified as
some unknown function of inputs Xx , X 2 , . . X^. .
(5.1) Y = f(Xx , X2 , ..., Xk )
So far as possible, any a priori logic relevant to the process should be

utilized; irrespective ofwhether it be based on theoretical considera-
tions or previous empirical study. Such logic is essential in deciding
whether a single equation or a simultaneous equations' model is
DATA COLLECTION FOR ESTIMATION 143
appropriate; and, if the latter, which variables are endogenous and

which predetermined. Moreover, as discussed in Chapter 3, familiarity
with the production process and Its underlying logic provides a basis
for delimiting the possible algebraic forms the production function
might take. If a simultaneous equations model appears relevant and
computationally feasible, it may be necessary to collect data on one or
more variables that enter the system of equations but do not enter the
production relationship itself. The sketching of a model enables such
variables to be easily pinpointed. Also, from an examination of the
model and the practical circumstances surrounding the production
process, it may be decided that satisfactory observations cannot be
made of some of the variables. Such a judgment might be necessary if
some variables have no scale of measurement or if the surmounting of
mechanical difficulties of observation would be too expensive.
For all the above reasons, model construction and practical famili-
arity with the production process are primary requirements for mean-
ingful data collection. They are essential. Oftentimes, too, the re-
searcher must take account of institutional restrictions on the type and
range of data he may collect. In this category, perhaps the most impor-
tant class of restrictions are those that arise in co-operative research.
ALTERNATIVE APPROACHES TO DATA COLLECTION
There are two broad approaches to data collection. The researcher

may sometimes have a choice between the two. More generally, the
availability of research funds and the circumstances of the production
process will be such that it is feasible or advisable to utilize only one
of these approaches to data collection. The first approach is to use ex-
perimental data. Contrasting with this is the use of real-world or non-
experimental data. For both experimental and nonexperimental mate-
rial it is pertinent to make a distinction between time series and
cross-sectional observations.
Time Series and Cross-Sectional Data
As the words imply, time series data consist of a series of obser-

vations on a particular unit — such as an animal or a farm —
made at
different points in time. Typically, these sequential observations
will
be equidistant in time. Cross-sectional data, on the other hand,
consist
of observations made on a number of units, one reading
per relevant
variable being taken from each unit. Frequently a given
set of data will
consist of both cross-sectional and time series observations
. For ex-
ample, most livestock production functions are based on
such data; ob-
servations being made relative to each of a group of animals
at a num-
ber of points in time. As discussed in Chapter
6, both time series and
cross-sectional data may lead to problems of economic
specification.
144 DATA COLLECTION FOR ESTIMATION
In addition, time series data may introduce estimational difficulties if
it is autocorrelated.
Occasionally, a choice may be possible between time series and
cross-sectional data. More often, only one or the other type of obser-
vation will be feasible in terms of the circumstances surrounding a
given production process. For example, whole -farm production func-
tions generally have to be fitted from cross-sectional data because time
series observations of sufficient length are unavailable; or, if available,
may relate to not one but a number of production functions because of
changes in technology over the period. Thus time series data over the
period 1900-1955, for a United States farm producing corn, might re-
flect at least four basic technological patterns or production functions,
as follows: 1900-1917 horsepower; 1918-1932 tractor power; 1933-1949
tractor power and hybrid corn; 1950-1955 tractor power, hybrid corn,
weedicides, and insecticides. Only a mongrel production function could
be fitted to observations covering such a range of technologies, unless
the data were adjusted in some fashion to allow for shifts in technology.
A suggestion of Solov as to how this might be done is discussed in
1
Chapter 7.
Just as for whole-farm functions, crop response functions relative
to fertilizer, rainfall, etc., generally have to be estimated from cross-
sectional data. Satisfactory time series observations on yield from a
given plot are difficult to obtain. Chiefly due to the effects of variations
in unmeasurable micro and macro climatic factors, year to year obser-
vations are usually not comparable. However, time series observations
are generally used in deriving animal production functions. With ani-
mals, the observations are usually made at short intervals of a few
days or weeks over a period of months. Over such a short period, suf-
ficient control can often be exercised to reduce to negligible propor -
tions variations caused by extraneous factors. The drawback, of
course, is that observations on the same animal will generally be auto-
correlated. Still, such autocorrelation may lead to fewer difficulties
than intrinsic variations between animals might cause if only cross-
sectional observations were utilized.
Experimental and Nonexperimental Data
Experimental data, in contrast with nonexperimental data, is char-

acterized by the fact that it is largely generated under the researcher’s
control. In large measure, the experimenter decides which variables
are to be held constant and which allowed to vary. He also has the op-
portunity of setting the levels at which the controlled variables will be
allowed to operate, and of deciding which particular combinations of
levels he will examine. Moreover, by virtue of his controlling influence,
1
So low, R. M. Technical change and the aggregate production function. Review of Eco-
nomics and Statistics, 39: 312-20. 1957. Also Aukrust, O. Investment and economic growth.
Prod. Meas. Rev., 16: 35-53. 1959.
, ,
the experimenter can generally ensure that there are no significant er-
rors of observation in the input factor data. This feature of experimen-
tation is most important in production function research since it fulfills
one of the basic assumptions of least squares multiple regression.
Still, it must be emphasized that, in practice, the generation of experi-
mental data can never be controlled absolutely by the researcher. 2 Ex-
traneous influences will always be present. Especially is this true with
respect to experiments aimed at the estimation of agricultural produc-
tion functions. Whether the function to be fitted relates to a single tech-
nology, such as fertilizer use, or an aggregation of technologies, as
found on a farm, there will always be some factors that cannot be con-
trolled by the experimenter. Unlike experimental data, nonexperimental
data is generated independently of the researcher; the only control he
may exercise is to use some purposive method of data collection ex
post Such control, being ex post cannot be nearly so effective as the
.
,
ex ante control that is possible over experimental data. For instance,

real-world data generally being collected “after the event/ errors of
observation are to be expected in the explanatory variables. This state
of affairs is particularly apt to be true with survey data where the re-
searcher has to rely on the respondent’s recall of his past actions.
In terms of the control exercised by the researcher, the distinction
between experimental and nonexperimental data, might be formalized
as follows. Suppose it were desired to estimate the production function
in equation 5.2, Xx to X k being the totality of input factors relevant to
output Y.
(5-2) Y= f(X„ X2 , ..., Xk )
Under nonexperimental conditions it might be possible to gather obser-

vations on the first g input factors. The estimated production function
would be equation 5.3,
(5-3) Y = f(Xx , X 2, . .
.
Xg ) + 6
the errore being due to the omission of the input factors X„

+1 to X k ,
assuming no errors of observation on Xx to X Among the factors X g+1 .
g
to Xk, some will be fixed and some variable. Thus, if the
researcher
knew which of the factors Xg +1 to X k were fixed, equation 5.3 might be
written more explicitly in the manner of equation 5.4.
(5.4) Y= f (X, ,
X 2 , . .
.
Xg / X g+1 , — , Xh / X h+l , ...,X k ) + e
In this equation the factors X to Xg are represented as variable and

x
observed, X g+1 to X h as fixed at some known or unknown levels and
an interestin g discussion of the distinction between
^°f experimental and nonexpert
mental data, and their relative roles in physical and social
i“ c
e ' r ° m observational ^ta. Journal of the Royal
119 : 28 “ 50 .
science, see Wold H Causal
Statistical Society, Series A.
Xh+ito X&, occurring after the double bar, as variable and unobserved.
So far as its real-world implications are concerned, the empirical pro-
duction function corresponding to equation 5.4 is more or less unrelia-
ble depending on the importance of the factors X k+1 to X k Also, to the .
extent that factors X g+1 to X k are fixed at atypical levels, the useful-
ness of the fitted function is lessened. Consider now the situation when
experimental observations are used. Assuming that normal experi-
mental procedures aimed at curbing the effects of “extraneous” varia-
tions are used, the estimated production function should tend to be of
the form of equation 5.5 where, relative to equation 5.4, g is greater
than or equal to g and h is greater than or equal to h. In other words,
(5.5) Y = f(X u X2 ,
...,X
g
/ X i+l ,
...,X E // X E+1 ,...,X k )
+ e
by using experimental procedures it is generally possible to increase

the number of variable factors observed and to increase the number of
additional factors held fixed. The net effect tends to be a reduction in
the number of unobserved variable factors. To the extent that g is
larger than g, and h larger than h, production functions fitted from ex-
perimental data are more reliable and tell more of the workings of the
production process than do functions based on nonexperimental data. In
so far as g and h may always be made as large as g and h, respec-
tively, production functions based on experimental data can always be
as reliable and useful as those derived from real-world observations.
The above statements do not imply that data collection should always be
based on experimentation. For many production processes, experi-
mental procedures are either mechanically infeasible or not worthwhile
in terms of their costs and benefits relative to nonexperimental data
collection.
Relative to the use of real-world observations, the other major ad-
vantage of experimentation is that it allows the individual input factors
that are controlled to be varied at the researcher’s discretion. Con-
sider a production process involving two inputs and a single output.
The output response might be represented by a production surface in
three dimensional space. By tradition and other institutional circum-
stances, it may be that in the real-world the two inputs are always used
in the same proportions and over but a small range of values. From a
bird’s-eye view of the production surface, real-world observations of
input quantities would always lie on a straight line across a small part
of the surface. Such a circumstance is illustrated in Figure 5.1. a.
Since an infinite number of planes can be fitted to pass through a
straight line, the researcher using such real-world data would be un-
able to fit a plane that might be identified as the production surface. 3
3
Such a criticism has been leveled at the pioneering production function research of
Douglas. See Mendershausen, H. On the significance of Professor Douglas' production
function. Econometrica, 6; 143-53. 1938.
data collection for estimation 147
X2
Figure 5.1. Bird’s-eye view of observed

points on production surface
for Y = f(X lf X 2 ).
Or, as discussed in Chapter 4, the multieollinearity present among the

input data may lead to the fitting of a completely irrelevant production
surface, determined not by the structure of the production process but
by errors of observation in the input data. Moreover, even if the real-
world input combinations are not strictly proportional, it is very likely
that they will tend to conglomerate in certain sections of the surface,
assuming a random selection of observations. Figure 5.1 .b illustrates
such a situation. While a determinate surface may be estimated for the
observed portion of the surface, no information is available relative to
the major part of the surface. In this respect the implications of fig-
ures 5.1.a and 5.1.b are similar. Sometimes, of course, real-world
observations may be scattered over all parts of the surface that are
thought relevant by the researcher. Still, if the scatter of observations
148 data collection for estimation
in the reliability
is markedly uneven, there will be great differences
with which the surface is estimated over its different regions. Such a
lack of reliability reduces the usefulness of the empirical estimates.
The advantage of an experimental approach to data collection is il-
lustrated by the contrast between the scattering of observations in Fig-
ure S.l.c and that in figures 5.1.a and 5.1.b. As shown in Figure 5.1.c,
the researcher using experimental procedures may select the input
combinations studied so as to give an even scattering of observations
over the section of the surface that he wishes to study. Of course, any
distribution of the observations deemed desirable could be arranged at
the discretion of the experimenter by virtue of his ex ante control over
the generation of the experimental data.
While the above discussion has been in terms of only two input fac-
tors, it applies equally well to the estimation of production functions in-
volving more than two inputs. Over-all, therefore, experimental data
collection has two highly desirable features relative to the collection of
data from real-world sources. First, under experimental conditions it
is generally possible to control a greater number of the relevant varia-
bles. Secondly, the experimenter may purposively arrange the obser-
vations on each variable input so as to cover all sections of the produc-
tion surface that are of interest. In addition, he may select the levels
at which the input factors are allowed to operate so as to simplify com-
putational procedures in fitting the production function. For instance,
estimation of a single equation production function model is greatly
simplified if orthogonal polynomial regression can be used. As indi-
cated in Chapter 4, this technique requires that the observations on the
independent variables be equally spaced. Only under extremely special
circumstances would real-world observations meet this restriction.
Conversely, it is a simple matter to arrange an experiment so that the
factor levels are equally spaced.
It has been argued that production function research based on ex-
perimental data has a major disadvantage relative to the use of real-
world observations. The criticism is that experimental conditions are
not typical of the uncertainty -ridden environment with its multitudinous
4
range of factors in which the farmer has to operate. In terms of equa-
tion 5.5, the farmer may be able to control factors X
x to Xg but will
have little or no control over Xg +1 to X£. Consequently, from the func-
tion estimated experimentally, the farmer will have no idea of the vari-
ation that may occur in Y due to real-world variations in any of the fac-
tors Xg+ x to X£. He only knows the values Y may take for certain fixed
values of Xg+ x to X£; assuming the factors Xh +1 to Xkto be of negligible
importance. Such a criticism of the experimental approach is not fatu-
ous. Still, strictly speaking, it is not a criticism of experimental data
collection per se but of the inadequacy of research funds. Given
4
Johnson, G. L.
Planning agronomic- economic research in view of results to date, in
Baum, E. Economic and technical analysis of fertilizer innovations and re-
L. et al. (eds.),
source use. Iowa State University Press, Ames. 1957. Pp. 217-25.
sufficient research resources, it would be possible to estimate the pro-

duction function 5.6 in which, unlike equation 5.5, all the controllable in-
puts are variable.
(5.6) Y = f(X 1? X2 , ..., X E // X E+l , ..., Xk )

+ e
Data for such a function need not necessarily be collected

fitting
from a single large experiment. That would be rather infeasible. It
might be obtained from a drawing together of the analyses and results
of experiments conducted and reported by many different authorities.
In this regard, a striking example of such an amalgamation process
formed the basis for Britain’s World War II fertilizer policy. Despite
the fact that experiments on crop -fertilizer relations had been carried
on for many years in Britain, at war’s start there was no clear basis
for devising a policy aimed at optimal allocation of the scarce fertilizer
supplies. By sifting and amalgamating meaningfully all the experi-
mental information available, Crowther and Yates produced a clear and
consistent picture of fertilizer responses that served as the corner-
stone of Britain’s wartime fertilizer policy. 5
Alternative to such an amalgamation process, a co-ordinated pro-
gram of sequential experimentation might be carried out. Initially, in-
terest may lie in determining the general shape of the production sur-
face and in deciding which inputs are most important. At later stages
stress might be laid increasingly on more applied analysis oriented
more and more to a controlled simulation of the farmer’s operating
conditions. An extension of the experimental program in this manner
appears preferable to using real-world data in an attempt to introduce
“extraneous” variations of the type typically encountered by farmers.
Real-world data does not incorporate these variations in any systematic
fashion. Under experimental conditions these effects can, in large
measure, be introduced systematically; so allowing meaningful inter-
pretation of their real-world implications.
Choice between Experimental and Nonexperimental Data
The two modes of data collection — experimental and nonexperi-

mental -are not entirely competitive. Within certain ranges they are
complementary. The extent of such complementarity depends on the
particular production process being studied, the region of interest on
the surface, and the estimation procedures to be used. In general,
some real-world information is necessary in order to plan experimental
procedures. For the major part, however, the collection of real-world
and experimental data are competitive activities in terms of available
research resources and the benefits to be derived therefrom. The
5
Crowther, E. M. and Yates, F. Fertilizer policy in war-time. The fertilizer require-
ments of arable crops. Empire Journal of Experimental Agriculture, 9; 77-97. 1941.
researcher has to make a choice between them. Ideally, this choice
should be made in terms of the marginal costs and benefits associated
with each type of data collection. In actuality, the decision cannot be
made in such terms since research benefits are never known ex ante .
However, the researcher does at least know that with a sufficiently

complete experiment he can always generate data as suitable, if not
better, than any real-world data that might be collected. The choice,
therefore, hinges primarily on relative costs. It is usually readily ap-
parent which procedure would be most expensive in terms of collecting
a specified quantity of data. Indeed, within the range of research re-
sources usually available, one or the other method of data collection
will often be infeasible although not impossible. Thus it would be pos-
sible, but surely infeasible, to set up a farm or a region as an experi-
mental unit from which production function data could be generated at
the researcher’s discretion. Conversely, biological production func-
tions often have to be fitted from experimental data because satisfac-
tory nonexperimental data could not be collected without great expense.
In cases where the choice is not so clear cut, account has to be taken of
the research resources required for the other phases of the investiga-
tion such as analysis, interpretation, and presentation of the results.
How much of the limited research resources should be allocated to
each of the research phases? In such a frame of reference, assuming
both experimental and real-world data are feasible, the choice may lie
between (a) expending a small quantity of resources on real-world data
collection with a large quantity of resources left for analysis and in-
terpretation, and (b) conducting an expensive experiment which would
leave only a modicum of resources for other phases of the research.
Given the uncertainties inherent in research, precise rules for deciding
such choices are impossible. The researcher can only attempt to make
an informed guess as to what pattern of allocation of the scarce re-
search resources might serve best in the light of the information avail-
able to him.
EXPERIMENTAL DESIGN
Relative to real-world data, the major advantage of experimental
data is its generation under the researcher’s control. Not only may he
decide which variables are to be examined. He also has the opportunity
of controlling the levels at which many of the variables will be allowed
to operate, and of deciding which particular combinations of levels he
will examine. The problem of experimental design lies in deciding how
this control should be exercised . 6 For instance, in what particular
6
See Cochran, W. G. and Cox, G. M. Experimental Designs. John Wiley and Sons, Inc.,
New York. 1957. Chs. 1, 2, and 8A; Box, G. E. P. Fitting empirical data. Annals New
York Academy of Sciences, 86: 792-816. 1960; Dykstra, O. Response Surface Designs.
Technometrics, 2: 185-96. 1960.
pattern should the observations shown in Figure 5.1.c be distributed

over the production surface? In making such decisions there are a
number of factors the researcher must bear in mind: the limitations
on available research resources; the purposes for which knowledge of
the surface is required; the current state of «knowledge about the sur-
face and its functional representation; the peculiarities of the particu-
lar production process being examined; the estimation techniques to be
used; and, if pertinent, the sequential nature of the experimentation.
As well, the researcher must take account of the accuracy with which it
is desired to estimate the production function: should a small section
of the surface be fitted with a high degree of accuracy or should a
larger portion be fitted roughly? Obviously, for an experiment to be
fruitful, intensive planning is essential. Advantage should be taken of
whatever competent advice is available. In particular, if a statistician
is available, he should be consulted during the planning stage and not,
7
as so often happens, after the experiment has been carried out. In-
deed, the justification for the discussion of experimental design given
here is not to show the production function researcher how he might
plan experiments independently of a statistician. Rather, it is to give
him a sufficient awareness of the possibilities and problems so that he
might derive maximum benefit from the statistical services at his dis-
posal.
Basic Concepts
For an appreciation of the experimental techniques useful in pro-

duction function research, some understanding is necessary of a few
concepts that are basic to experimental design. In any experiment,
treatments are applied to experimental units . If the experiment is a
cross-sectional study, there will be one treatment per experimental
unit. Time series experiments, on the other hand, imply the applica-
tion of a number of treatments, each at a different point in time, to
each experimental unit. Some experiments may involve both cross-
sectional and sequential treatment applications. The treatments ap-
plied, and the response they induce, are recorded as the observations .
Thus in a fertilizer -crop response experiment, one of the many treat-

ments used might consist of 80 pounds of nitrogenous fertilizer and 100
pounds of phosphatic fertilizer per acre. The experimental units might
be 0.1 acre plots, the response being the amount of crop grown on each
experimental unit converted to a per acre equivalent. With such a
cross-sectional study there would be a single set of observations for
each experimental unit. In the present instance, the set of observations
7
For some comments on the statistician’s role, see Finney, D. J. The' statistician and
the planning of field research. Journal of the Royal Statistical Society, Series A, 119: 1-17.
Also, Ostle, B. Statistics in research. Iowa State University Press, Ames. 1956. Pp. 421-
26; and Box, G. E. P. The exploration and exploitation of response surfaces: some general
considerations and examples. Biometrics, 10: 26-7. 1954.
would be a triplet showing the amount of each fertilizer applied to the
experimental unit and the response obtained. Just what form the ex-
perimental unit takes depends on the design of the experiment. It may
be a single object such as a plot of ground, a pot of soil, a plant or an
animal; or it may be a group of such objects. The essential feature is
that the experimental unit is the subject of a treatment, and that each
set of observations refers to a particular experimental unit. As dis-
cussed in Chapter 4, sets of observations made on the same experi-
mental unit are likely to be autocorrelated. Such observations will also
contain an experimental error insofar as there is any lack of exactness
or uniformity in the experimental technique used. When a number of
experimental units are involved, as in cross-sectional studies; experi-
mental error will also arise to the extent that the experimental units
are not exactly similar. If they are not identical, as is usually the
case, variations in response will not be wholly attributable to differ-
ences between treatments; some of the variation will be due to the dif-
ferences between the experimental units. Since the effects of the vari-
ous treatments are the object of study, it is extremely desirable to
have the experimental units as similar as possible. However, experi-
mental error can never be completely eliminated in agricultural pro-
duction function research. There will always be some lack of uniform-
ity in technique and dissimilarities between the experimental units,
either intrinsically or in terms of the environment surrounding the ex-
perimental unit.
Since experimental error is always present, the technique of ran-
domization has been devised to prevent this error from affecting the
observations systematically. Randomization implies the allocation of
the treatments to the experimental units in a random fashion. It is an
insurance against experimental error causing systematic bias in the
data; it does not remove the error. To gain some idea of the experi-
mental error involved in an experiment, and hence of how reliable the
data may be, replication is often practiced. By replication it is meant
that the experiment is repeated a number of times; in part or in full;
the variations in response between the same treatment applied to a
number of experimental units giving an indication of the experimental
error. Replication does not reduce the experimental error. It merely
enables the researcher to take better account of the error.
Response Surface Designs
Historically, the experimental examination of response phenomena

has usually been based on designs oriented to discrete data classifica-
8
tion. These studies have commonly failed to determine the functional
8
See Anderson, R. L. A comparison of discrete and continuous models m agricultural
production analysis, in Baum, E. L . et al. (eds.), Methodological procedures in the economic
analysis of fertilizer use data. Iowa State University Press, Ames. 1956. Pp. 39-47.
interrelationships between the input and output factors examined. Nor

do they adequately emphasize the “fixed” conditions under which the
experiments were carried out; important factors frequently being
omitted from consideration altogether. Generally only one or two input
factors have been examined, each at a small number of levels. More-
over, these two or three factor levels have usually covered but a small
part of the range over which a response might have been elicited. Pro-
viding estimates of only two, three, or four points on a limited section
of the production surface, these studies are markedly — if not totally —
inadequate in terms of the generation of data for production function es-
timation. Similar strictures apply to current response studies of simi-
lar design.
Instead of attempting to estimate production functions such as equa-
tion 5.1, researchers have most often simply tried to ascertain such
qualitative facts as whether (a) output responded to some hypothesized
input, (b) one level of output induced a greater response than another,
(c) any interaction occurred between inputs. Such questions were an-
swered by way of analysis of variance, care being taken to replicate
the experiments sufficiently to assure correct answers to the questions
posed. In terms of the questions asked, such procedures are correct.
However, from an applied research viewpoint, the questions are inade-
quate. The farmer needs more than qualitative information or infor-
mation about one or two factors at two or three levels. He requires
quantitative information of the type derivable only from empirical
knowledge of the over -all production surface. For maximum benefit to
the farmer, applied research on input -output phenomena should lead to
the estimation of the relevant production function. Given empirical es-
timates, the implications of the response function may then be drawn out
as illustrated in chapters 2 and 3.
That most input-output research by agricultural scientists has failed
to recognize the existence of a continuous production surface, or at
least the implications of such a surface, is not surprising. Why should
agronomists,, animal nutritionists, and their confreres, have appreci-
ated such concepts as marginal productivities, marginal rates of sub-
stitution, marginal costs, and marginal revenues in a multifactor set-
ting when the majority of agricultural economists did not! 9 Moreover,
more than a modicum of the less recent work on response phenomena
by agricultural scientists was in the nature of fundamental research.
Those who carried out this research cannot be wholly blamed for the
fact that it was often misrepresented as applied research and
seized as
ready made grist for the extension mill.
9
There were a few early agricultural economists who appreciated
some of the implica-
tions of production function analysis. See, for example,
Tolley, H. R. et at. Input as related
to output in farm organization and cost of production
studies. Bui. 1277.
USDA, Washing-
ton, D. C. 1924; and Black, J. D. Introduction to production economics. Henry Holt and
Co. New York. 1926. Ch. 12. For an appreciation of the slowness with
which the concep-
tual framework and empirical specification of the production
function was accepted, see:
Case, H. C. M. and Williams, D. B. Fifty years of farm
management. University of Illinois
Press, Urbana. 1957; also, Black, J. D. Dr. Schultz on farm
management research. Jour-
nal of Farm Economics, 22: 570-80. 1940.
Presaged by recognition of the inadequacy of the historical ap-
proach to input -output phenomena, experimental designs for production
function estimation have been the subject of much discussion over the
10
last few years . A number of designs specifically oriented to the fit-
ting of production functions have been developed. Some of the older de-
signs have also been adapted by simply increasing the number of factors
and treatments considered so that the data might be amenable to multi-
ple regression analysis. In contrast to the experimental layouts used
in discrete, “analysis of variance” type studies, the newer response de-
signs generally involve a larger number of input factors; always with a
more than proportionate increase in the number of factor levels and
combinations examined. In other words, a greater number of observa-
tions are made per replicate. Moreover, these newer designs acknowl-
edge that for satisfactory estimation the observations must be located
methodically over the relevant section of the production surface. Be-
cause of the increased number of treatments and observations, experi-
ments based on these response surface designs tend to be more expen-
sive than their inadequate predecessors — at least in toto Especially .
is this true for cross-sectional studies since they necessitate the use
of additional experimental units. Research resources being limited,
compromise becomes necessary. The usual adjustment is to reduce the
extent of replication rather than to cut down the per unit size of the ex-
periment. Of course, it is also possible to reduce the number of input
factors that are allowed to vary under the experimenter’s control. It
must be recognized, though, that any reduction in the number of factors
considered tends to decrease the practical usefulness of the subsequent
analysis.
Broadly speaking, there are four classes of experimental designs
that are especially pertinent to the generation of data for production
function estimation. They are the complete factorials, the fractional
factorials, the composite, and the rotatable designs. In contradistinc-
tion to the classical experimental designs, such as randomized blocks,
Latin squares, etc., the factorial, composite, and rotatable designs re-
late to the arrangement of the treatments relative to one another and
not to the positioning of the experimental units. Such orientation to the
treatments, and hence to the observations, is what makes these designs
relevant for the fitting of production functions. Of course, in using
these surface -oriented designs, the concept of randomization is usually
employed in allocating the treatments to the experimental units. Since
10
For example, see Box, G. E. P. and Wilson, K. B. On the experimental attainment of
optimum conditions. Journal of the Royal Statistical Society, Series B, 13: 1-45. 1951; An-
derson, R. L. Recent advances in finding best operating conditions. Journal of the Ameri-
can Statistical Association, 48: 789-98. 1953; Fellows, I. F. Production functions in farm
management, in Halcrow, H. G. (ed.). Contemporary readings in agricultural economics.
Prentice -Hall, Inc., New York. 1955. Pp. 74-80; Baum, E. L. et aL (eds.) Methodological
procedures in the economic analysis of fertilizer use data. Iowa State University Press,
Ames. 1956. Pp. 37-98; Chew, V. (ed.) Experimental designs in industry. John Wiley and
Sons, Inc., New York. 1958; Box, G. E. P. and Lucas, H. L. Design of experiments in non-
linear situations. Biometrika, 46: 77-90. 1959.
the factorial, composite, and rotatable designs relate to the arrange-

ment of the treatments, it is possible within limits to superimpose
them on some of the classical experimental designs. For instance, the
treatments in a randomized block design might be arranged to form a
fractional factorial. Some of the classical designs have been used for
response surface estimation without the super imposition of planned ar-
rangements of the treatments. In this regard, completely randomized,
randomized block, and split-plot design's have been most popular.
Together with their broad advantages and disadvantages, the gen-
eral features of the above-mentioned designs are outlined below so far
as they relate to production function research. Particular stress is
given to the factorial designs and the rotatable designs because of their
greater relevance for production function estimation. Except where
otherwise stated, it is assumed that a single equation model satisfac-
torily represents the production surface and that estimation is to be
via least squares multiple regression. If a simultaneous equations
model were pertinent, the experimental procedures would not neces-
sarily be drastically altered. In most cases, the only change required
would be an extension of the experiment’s scope to enable measurement
of the effects of those variables present in the system of equations but
not entering the production relation embedded in the system.
Randomized Block Designs
Probably no design has been more frequently used in production

function research than the randomized block. In this design, as usually
used in cross-sectional studies, the selected treatments are allocated
at random among the experimental units. Under such circumstances,
the number of experimental units required per replicate is equal to the
number of treatments being studied. In a time series study, where the
interest centers on a number of treatment sequences, the treatment se-
quences would be allocated at random among the experimental units.
Each replicate would require a number of experimental units equal to
the number of treatment sequences being studied. An advantageous
fea-
ture of the randomized block design is that it allows the
experimental
units to be allocated so as to make those within each replicate
(block)
as homogeneous as possible. Such grouping of the experimental
units
tends to minimize the experimental error within each block.
If only a
single block of experimental units is used, the randomized
block design
reduces to a completely randomized design. For discrete
“analysis of
variance” type studies based on randomized blocks, only a
few treat-
ments consisting of two or three levels of a single factor are
usually
examined. Such a restricted layout is of limited use for
production
function research. This fault, however, is not an
intrinsic feature of
the randomized block design. It can easily be
remedied by extending
the experiment to allow continuous multifactor
analysis oriented to the
fitting of a response surface. All that is
required is an increase in the
number of factors and factor levels studied so as to provide sufficient
observations for the fitting of whatever production function model is
hypothesized.
As a simple example, suppose a randomized block design is to be
used in a fertilizer -crop response study. Assume two fertilizers, P
and K, are to be studied. Each of these factors is to be held once at
each of five levels: P at 0, 20, 40, 60, and 80 lbs. per acre and K at
0, 30, 60, 90, and 120 lbs. per acre. Such a range of factor levels im-
plies that the surface region of interest is bounded by 0 < P « 80 and
0 « K « 120. All told, there would be 25 possible treatments. One set
of five drawn at random might be as follows:
Treatment Level of P Level of K

1 0 60
2 20 30
3 40 0
4 60 120
5 80 90
These (P, K)treatments would be allocated at random within blocks of

five experimental units. The units might be pots, fields, or plots, de-
pending on the scope of the experiment. Suppose there were only a sin-
gle replicate of the experiment. It would generate five sets of cross-
sectional observations for estimation of the surface. For estimational
purposes, five is a very small number of observations. Only models
containing four or less parameters could be fitted, otherwise there
would be no degrees of freedom available for statistical testing of the
estimated parameters. For instance, a five parameter polynomial
model would be forced to fit the data perfectly but the resultant esti-
mates of the parameters could not be tested statistically. Still, models
of quite diverse logical implications, such as equations 5.7, 5.8, or 5.9
to give a few examples, could be fitted and tested via such a small ex-
periment.
(5.7) Y = j3 0 P^1 2
(5.8) Y = ft, + ft x P + 02 KP
(5.9) Y= ft, + ft P + 02 K
2
+ /3 3 KP
With no replication, would be impossible to obtain an estimate of
it
the experimental error from the experiment

itself. If the error were
known from previous experiments of the same type, assuming the ex-
perimental error to be fairly constant between experiments, this a
Table 5.1. Analysis of Variance for Multiple Regression

Based on a Replicated Experiment
Degrees of
Source of Variation in Y Freedom Sum of Squares
‘ '
/t
Replicates r-1
[A
<
i^ )
]-[ I
< 5
5 , /,r
]
Regression k SSR
tr A ?
2
Deviations from regression 7T 1
T—
i M 2 <Y -Yj )
J
j=i
Lack of fit t-k-1 SST-SSR

Experimental error (t-l) (r-l) Found by subtraction
tr 2
Total tr-1 S yf
j=i
priori estimate of the error could be used. Without knowledge of the

experimental error, the researcher could not satisfactorily compare
the goodness of fit of the various hypothesized functions he might fit.
He would have no way of knowing what proportion of the deviations from
regression were due to experimental error and what to the intrinsic un-
suitability of the fitted function. Analysis of variance for multiple re-
gression could only be carried out as illustrated in Table 4.1. Now
suppose that the experiment had been replicated twice. This would en-
tail the use of ten experimental units split into two blocks of five units
each. Completion of the experiment would give ten sets of observa-
tions, made up of two response readings for each of the five treatments
used. The number of degrees of freedom available for estimational and
testing purposes would be doubled. Just as importantly, an estimate
could now be made of the experimental error; making it possible to
Table 5.2. Alternative Presentation of Analysis of Variance for Multiple

Regression Based on a Replicated Experiment
Degrees of
Source of Variation in Y Freedom Sum of Squares
’
Replications r-l 7" 7 *]-['& Y‘ f/H
Cl'i 1
Treatments t-l SST

Regression k SSR
Lack of fit t-k-1 SST-SSR
Experimental error (t-l) (r-l) Found by subtraction
2
Total tr-1 S h
J=i
carry out an F test of the goodness of fit to the data of the fitted func-
tional form. Assuming a k + 1 parameter production function model
such as equation 4.1 or 4.46, the analysis of variance for multiple re-
gression of Table 4.1 could be extended as shown in Table 5.1 for the
general case with r replications and t treatments. In format, this
table corresponds to Table 4.1. An alternative method of presenting
the same material that is frequently used is shown in Table 5.2. If or-
thogonal polynomial regression is used, the regression sum of squares
of tables 5.1 and 5.2 can be further subdivided as shown in Table 4.2.
As Table 5.2 implies, the treatment sum of squares (SST) equals
the sum of squares explained by regression (SSR) plus the lack of fit
sum of squares. Moreover, SST can be calculated as shown in equation
5.10. Hence the lack of fit sum of squares can be obtained by subtract-
ing SSR from SST. SSR is the numerator of equation 4.39.
SST £ f /tr]
(5.10) = £.1 ( Yjj )* / r
] - [( ^ Yj
The experimental error sum of squares can be found by subtracting the

replicate sum of squares and the treatment sum of squares from the
total sum of squares. As is implied in Table 5.1, the deviations from
regression are made up of deviations due to the failure of the fitted
model to correctly represent the surface and to experimental errors.
Mean squares are not shown in tables 5.1 and 5.2. They may be ob-
tained by dividing the sums of squares by their respective degrees of
freedom. An F test of the regression mean square against the error
mean square provides an over-all test of the significance of the regres-
sion analysis, as discussed in relation to Table 4.1. The lack of fit
mean square may also be tested against the error mean square. If this
F value is significant at the selected probability level, it indicates that
the model does not adequately describe the production relationship.
Mason has suggested a stronger criterion: if the lack of fit mean square
is of the same order of magnitude (or less) than the error mean square,
then the fitted functional form adequately characterizes the data. 11 Of
course, if the lack of fit mean square were zero it would imply that the
function used fitted the surface exactly. When alternative functional
forms are fitted to the data, the only changes that occur in the analysis
of variance are in the regression and lack of fit mean squares. The
total, replicate, treatment, and error mean squares are independent of
the regression analysis. Hence the lack of fit mean squares of alterna-
tive functions can be compared as a statistical guide to the adequacy of
the various functions. Such a procedure is, of course, only a rule of
thumb to be used in conjunction with other criteria for assessing the
correctness of a particular model. These other criteria may be
u Mason, D. D. Functional models and experimental designs for characterizing response

curves and surfaces, in Baum, E. L. et al. (eds.) Methodological procedures in the economic
analysis of fertilizer use data. Iowa State University Press, Ames. 1956. P. 80.
Figure 5.2. Bird’s-eye view of observed points

on production surface for Y = f(P,K)
with a randomized block design.
statistical in nature or they may be based on the underlying logic of the

production process being studied.
In the above hypothetical experiment, which of the available levels
of P and K entered each treatment was decided at random. The result-
ant pattern of observed points is shown, from a bird's-eye view, in
Figure 5. 2. a. It is but one of the 51 observational patterns that could
have arisen. As the figure shows, it is not an intuitively satisfactory
pattern since the observations are scattered rather unevenly over the
surface. For this reason it would be advisable for the researcher to
select the treatments — and hence the pattern of observations — purpos-
ively rather than at random. Thus, assuming only five treatments are
possible, a systematic selection of factor combinations as illustrated in
Figure 5.2.b might be satisfactory. Certainly it gives a much more
evenly distributed scattering of the observations over the surface area
bounded byO^P<80, 120.
Figure 5.2.c indicates the worst possible distribution of the five ex-
perimental points. Because of the equal spacings between factor levels,
treatments made up of corresponding rankings of the factor levels are
multicollinear. Of course, multicollinearity would not have resulted
had the factor levels been unequally spaced. However, unequal spacings
should not be used without a particular reason, for they imply greater
interest in some parts of the surface than in others. Moreover, equal
spacings permit orthogonal polynomial regression to be used conven-
iently and hence can lead to computational simplicity. Even without us-
ing orthogonal polynomials, a very worthwhile computational saving can
be introduced by coding the input data if it is equally spaced. Thus, for
the experiment considered above, a unit of P might be taken as 20 lbs.
and a unit of K as 30 lbs. The five (P, K) treatments of Figure 5. 2.
might then be written as (0,2), (1,1), (2,0), (3,4) and (4,3). An alterna-
tive coding of the factor levels that is still more convenient computa-
tionally is that given in equation 5.11,
where X'y the coded value of the j-th level of the i-th factor, X £ is the
is
mean level of X^nd
h is the spacing between one level of X^nd the
next. With this system of coding, the levels of P and K used in the ran-
domized block experiment would be -2, -1, 0, +1, and +2. For instance,
the 30 lb. level of K becomes (30-60)/3Q or -1 units. Use of the above
coding procedures is not limited to randomized block designs. They
may be used with any design.
As an example of an experiment planned as a randomized block de-
sign with both cross-sectional and time series features, consider the
following Iowa study. It was aimed at estimating the production function
for liveweight beef from feed consisting only of corn grain and brome-
alfalfa forage (plus a constant negligible quantity of essential feed ele-
ments not present in the corn or forage) The production function
.
model might be represented as in equation 5.12 where Y is the cumula-

tive quantity of liveweight beef produced from the cumulative quantity of
(5.12) Y = f(F, C) + e
forage, F, and corn, C, consumed. The error, e , is assumed to be neg-

ligible and arises because of the noninclusion of other relevant variable
factors such as temperature, humidity, water consumption, etc.
Eighty -four steers of initial weight between 500-600 lbs. were avail-
able for the experiment. They were divided into two blocks of 42 ani-
mals each, providing for two replicates. Each block was divided into
six pens of seven animals. An attempt was made to allocate the animals
so that they were as similar as possible within each pen and block.
Figure 5.3. Bird’s-eye view of forage-corn

rations followed in a beef
production study.
Each pen constituted an experimental unit. Forage and corn inputs

were controlled by using six rations with F/C ratios as follows: 1:0;
20:1; 10:1; 5:1; 3:1; and 2:1. These six rations were allocated at ran-
dom to the six pens in each replicate, each pen being fullfed the one ra-
tion throughout the course of the experiment. By keeping each group of
animals on a constant feed mixture the possibility of carry-over ef-
fects, aswould occur if the animals had to adjust to varying feed mix-
tures,was precluded. From a bird’s-eye view, as shown in Figure 5.3,
each pen followed a particular F/ C ration line out from the origin over
the period of the experiment. Rations containing more than one -third
corn were not considered since more corn than this could not be han-
dled by the experimental animals. Of course, with conditioning over a
suitable period of time, cattle could be fed a higher proportion of corn
than one -third. However, under the particular conditions of the experi-
ment, it was not possible to condition the animals. For this reason, the
feeding of a fixed feed mixture to each experimental unit was a defi-
ciency in the design. It precluded the taking of observations over part
of the surface. Observations were recorded every two weeks for 20
weeks. Counting the initial observation of zero feed intake and zero
gain, there were thus 11 sets of observations on each unit or 66 sets of
observations per replicate in all. These observations consisted of the
cumulative weight gain of the cattle in each pen and the cumulative
quantity of corn and forage consumed by each pen, tabulated for analy-
sis on a per head basis. From a production function research view-
point, the treatments are not the six rations that were used, but the ac-
tual feed quantities consumed. Since these intake quantities were
determined by the animals, this experiment illustrates a situation
where the treatments (except the initial one of zero feed) are not known
in advance. Given the design of the experiment, the treatments were
only partially under the researcher’s control; their location was deter-
mined only insofar as they were forced to lie on the ration lines of Fig-
ure 5.3. Nonetheless, given the necessity for the livestock to eat, the
feeding of a constant mixture to each pen at least guaranteed a perhaps
not too unsatisfactory scattering of observations over the relevant sec-
tion of the surface. At the same time, the feeding procedure tended to
intensify the number of observations close to the origin. Most impor-
tantly, with the treatments being determined ty the animals, there is a
difficulty in that true replication of such an experiment is well-nigh im-
possible despite the fact that two “replications” were carried out. The
experiment was therefore not a randomized block design with two repli-
cates but a completely randomized design using 12 experimental units.
It would be virtually impossible to have groups of fullfed animals con-
sume the same cumulative quantities of various feed types within the
same time periods. For an experiment such as the above it is, there-
fore, not possible to carry out analysis of variance oriented to compar-
ison of the “lack of fit” and experimental error mean squares as illus-
trated in tables 5.1 and 5.2. Of course, an approximate replication
might be possible; the data between blocks being adjusted in some fash-
ion so as to simulate replication. To do this, however, without strong
grounds for the particular method of adjustment used, would be unsatis-
factory. An alternative procedure would be to weigh the animals after
they have consumed preselected quantities of feed. To some extent this
would lead to mechanical difficulties in conducting the experiment since
the time to consume a given quantity would vary between animals. Sta-
tistical analysis would also be complicated by the necessity to take ac-
count of the time factor.
The misinterpretation between ration types and feed quantities as
treatments is not uncommon. Basically, it stems from a failure to
recognize (a) that a continuous response surface is involved, and (b) that
the marginal rates of substitution between feed components are not con-
stant. An example of such misinterpretation is the suggestion that rep-
licated cross-over designs be used in livestock response studies. 12
These designs are closely related to the randomized block design.
Their essential feature is that the rations (mistakenly regarded as
treatments) are interchanged among the experimental units during the
experiment. For discrete “analysis of variance* type studies, such in-
terchange makes it possible to take better account of the variation be-
tween experimental emits; thereby reducing the experimental error.
However, the benefits for continuous multifactor production surface es-
timation are slight, given the infeasibility of replicating the feed
treatments.
Just as with any experimental design, the decision to use random-
ized blocks poses three questions. They are: How many treatments
should be used? What particular treatments should be studied? How
12
See Hoglund, C. R. et al. (eds.) Nutritional and economic aspects of feed utilization by
dairy cows. Iowa State University Press, Ames. 1959. pp. 167-92.
many replications should be carried out? Given that research re-

sources are not free goods, these questions are related. For instance,
with only a limited number of experimental units available, any increase
in the number of simultaneous replications must be at the expense of
the number of experimental units in each replicate. The result is a re-
duction in the number of treatments that might be studied per replicate.
Concomitantly, the question of which particular combinations of factor
levels should be used as treatments becomes far more important.
There are no clearly defined answers to these questions. Still, the
number of treatments, replications and types of treatments that should
be used can be partly decided in terms of the minimal requirements to
be met. Thus, to determine the experimental error and lack of fit mean
squares, at least two replicates (blocks) are necessary —unless the re-
searcher has prior knowledge of the experimental error to be expected.
So far as treatments are concerned, there should always be at least
sufficient to provide a few more sets of observations than there are pa-
rameters to be estimated in the production function model. Also, while
replication generally tends to display rapidly diminishing returns (in
terms of the satisfactoriness of the estimate of the error), increasing
the number of treatments studied does not lead too quickly to diminish-
ing returns — at least within the range of experimental resources usu-
ally available. With a larger number of treatments it is possible to ob-
tain a more complete distribution of observed points over the response
surface, as well as the provision of additional degrees of freedom for
statistical testing of the estimated parameters. It is therefore best to
have only two or, at most, three blocks so that as many different treat-
ments as are necessary for estimation of the production function pa-
rameters may be studied. Of course, if the experimental error is ex-
tremely variable from block to block, it may be necessary to use more
than three replicates. The question of which particular treatments
should be studied when the randomized block is used has already been
discussed in part with reference to Figure 5.2. The minimal require-
ment is that the selected treatments should not be multicollinear. So
far as possible, an even distribution of the treatments over the surface
section of interest is desirable — assuming the researcher is equally
interested in all parts of this portion of the surface. If he is not, the
scatter of the observed points should roughly correspond to the intensity
of his interest in different sections of the surface. With respect to the
factor levels to be used, a handy rule of thumb is to arrange the treat-
ments so that the mean level of each factor approximates what is
thought to be its most profitable level. The advantage of such a proce-
dure lies in the fact that regression analysis provides estimates of
highest precision for points closest to the means of the explanatory var-
iables. If the range of each factor is fairly large, this procedure is to
be highly recommended.
When a sequential program of experimentation using randomized
blocks is being followed, which treatments to use at each phase will in
large part be determined by the stage of the study. Thus, in the early
phases, a small number of treatments involving a large number of fac-
tors and covering a wide range of the surface might be examined. After
ascertaining in this way the broad shape of the surface, as well as
screening out unimportant factors, subsequent experiments might in-
vestigate particular sections of the surface in more detail. For in-
stance, once the approximate positions of the surface’s ridge lines have
been ascertained, treatments for any experimental design should be
confined to the section of the surface between the ridgelines. Outside
of this region, the surface is of no economic interest. Response there
is always inefficient, regardless of input -output price ratios.
Complete Factorial Designs
In studying Figure 5.2, the reader may have noticed that the distri-
bution of the experimental points over the surface was limited because
each input level was only studied once. Had each level been used more
than once, in combination with a number of different levels of the other
factor, it would have been possible to obtain a more complete scattering
of observations over the surface. Perhaps the most intuitively appeal-
ing distribution of the experimental points is that which would have re-
sulted if each level of one input had been studied in conjunction with
each level of the other input. The resultant distribution of the observa-
tions would have been akin to that of Figure 5.1.c. Such experimental
layouts, with treatments consisting of all possible combinations between
factors of the within-factor levels selected for study, are known as
complete factorial designs. They are frequently used to generate data
for response surface estimation. The factorial design of an experiment
says nothing of the way in which the experimental units should be organ-
ized. The word “factorial” refers only to the arrangement of the treat-
ments. Thus a factorial design may be superimposed on any of the or-
dinary experimental designs. In response surface experiments, the
factorial arrangement is most often used with a completely randomized,
randomized block, or split -plot arrangement of the experimental units.
Of course, when the factor levels entering the treatments are not fully
determined by the researcher — as is often the case in livestock feeding
experiments — a factorial arrangement is infeasible because factor
levels cannot be conveniently replicated.
Had the fertilizer -crop randomized block experiment relative to
Figure 5.2 been arranged as a complete factorial, the necessary (P, K)
treatments would have been as listed below. The treatments are coded
by way of equation 5.11.
(-2,-2) (-1,-2) (0,-2) (+1,-2) (+2,-2)
(-2,-1) (-1,-1) (0,-1) (+1,-1) (+2,-1)
(-2, 0) (-1, 0) (0, 0) (+1, 0) (+2, 0)
(-2,+l) (-1,+1) (0,+l) (+1,+1) M,+l)

(-2, +2) (-l,+2) (0,+2) (+l,+2) (+2, +2)
The design would be described as a randomized block with the treat-

2
ments arranged as a 5x5 or 5 factorial, there being 25 treatments or
experimental points per block. Likewise, if three factors were being
investigated in a complete factorial design, one at three levels, one at
four levels and the other at five levels, the experiment would be de-
scribed as a 3x4x5 factorial. It would involve 60 treatments. The lo-
cation over the surface of the experimental points of a two factor fac-
torial may be shown graphically in two dimensions. Such a planar
view can only be given when two inputs are involved. With three fac-
tors, a three dimensional diagram would be necessary. For more than
three inputs, geometric representation of all the experimental points is
impossible so that it is extremely difficult to conceptualize the location
of suitable experimental points. This being so, it is useful to have
rules of thumb — such as factor ializat ion provides — for deriving factor
levels to enter the various treatments.
So far as they are oriented to the simultaneous examination of mul-
tifactor effects and provide sufficient observations for the fitting of a
continuous function to the response surface, factorial designs are
13
suited to production function research The symmetric arrangement
.
of the treatments guarantees a relatively satisfactory distribution of

experimental points over the surface, especially if equal spacings are
used between factor levels. Of course, the researcher using a complete
factorial still has to decide on how many levels of each factor to use
and the size of these factor levels, as well as the extent of replication
to be carried out. While there are no strict rules for answering these
questions, the situation is no more difficult than for a nonfactorial ran-
domized block design except insofar as more treatments and experi-
mental units are required per trial. Nor does the factorial arrange-
ment complicate the estimational procedures in any way. Still, there
are a number of disadvantages to complete factorial designs. From an
estimational point of view, they are unsatisfactory in that they lead to
estimation of pure higher order effects, such as 0U of equation 5.13,
with less precision than the interaction effects of the same order, such
as ft.3 of equation 5.13. Also, and most importantly when research ma-
terials are limited, they require too many treatments, and hence too
many experimental units, relative to the number of parameters to be
estimated. As the number of factors or factor levels is increased,
complete factorials become increasingly wasteful of research re-
sources. In the estimation of response surfaces by polynomial type
functions, as frequently used in crop and livestock studies, it has rarely
been found worthwhile to use a function involving terms of higher
13
Historically, factorial experiments have been analyzed m discrete terms, with the ef-
fects of each factor alone (pure or main effects) and the joint effects of a number of factors
(interaction effects) considered as single-valued phenomena. The significance of these var-
ious effects were tested by way of analysis of variance. Such an approach is, of course,
misdirected in a continuous response surface is relevant. For a contrast of the two ap-
proaches to factorial analysis, and the extent of their congruity, see Anderson, E. L. A
comparison of discrete and continuous models m agricultural production analysis, in Baum,
E. L. et al. (eds.) Methodological procedures in the economic analysis of fertilizer use
data. Iowa State University Press, Ames. 1956.
degree than the second. Such a quadratic polynomial for three factors
is given in equation 5.13. For the fitting of a quadratic polynomial, at
least three levels of each factor must be studied. If four levels are ex-
amined, a cubic polynomial could be fitted; with five levels a quintic
polynomial; and so on as the number of factor levels is increased. Thus
with more than three levels, the use of a complete factorial arrange-
ment provides for the estimation of higher order effects that the re-
searcher generally knows to be negligible. This emphasis on effects
that are of little interest could be reduced if the number of replications
of each factor level were decreased, substituting an incomplete facto-
rial arrangement for complete factorialization. Such designs are dis-
cussed later.
Even when only three levels of each factor are used, complete fac-
torialization is an inefficient method of generating data. Consider
Table 5.3. It shows the number of experimental points or sets of ob-
servations implied by 3 n and 4 n complete factorials relative to the
number of parameters to be estimated in an n factor quadratic poly-
nomial for 2 « n « 7. With two or three factors at three levels in a sin-
gle replicate, a 3 n experiment would not be too large. For instance,
with three factors there would be 27 sets of observations to estimate
the 10 parameters of equation 5.13. However, with more than three
(5.13) Y = 0O + 0, X x + 0, X, + 0, X, + 0U X2 + @ 22 X* + ,S
33 X 2
3
+
/3
12 X, X 2 + @ X,X
i3 S + @23 X2X3 + e
factors the number of observed points becomes increasingly dispropor-

tionate to the number of parameters to be estimated. Even if sufficient
research resources were available to conduct such large experiments,
it is simply not necessary to have so many degrees of freedom
availa-
ble for statistical testing; quite apartfrom the well-nigh impossible
mechanical complexity conducting such large experiments. With
of
more than three levels of each factor, these effects are increased more
than proportionately, as the figures of Table 5.3 for 4 n designs indicate.
Since even a single replicate of a complete factorial involving more
than three factors or factor levels requires a rather impractical num-
ber of treatments and experimental units, it is undesirable to obtain an
estimate of the experimental error by replicating the whole experiment.
For instance, a 3 s complete factorial with two replicates would require
54 treatments and experimental units, generating 54 sets of observa-
tions for the estimation of 10 parameters. Obviously, to obtain an esti-
mate of experimental error in this way compounds the wastefulness of
the complete factorial design. If an estimate of the error is available
from previous experimentation, assuming the experimental error is
relatively stable between experiments, it may be compared with the
mean square of the deviations from regression relevant to the unrepli-
cated factorial data. If the deviations from regression mean square
is
larger than the error, it is an indication that the fitted function probably
does not adequately describe the surface. An alternative that is far
,
n 11
Table 5.3. Number of Experimental Points in 3 and 4 (n= 2, . . 7) Complete .
Factorials Contrasted With the Number of Parameters To Be

Estimated in an n Factor Quadratic Polynomial
Number of Number of Parameters

Number of Factors
Experimental Points To Be Estimated
n n
(n) (3 ) (4 ) ( 2n +l + nîi)
2 9 16 6
3 27 64 10
4 81 256 15
5 243 1,024 21
6 729 4,096 28
7 2,187 16,384 36
more economical than complete duplication of the experiment is also

available. It is to replicate only the treatments at the center of the
experiment, i.e., those treatments in which all the factors have a coded
value of 0. If this treatment is replicated r times, an estimate of the
experimental error mean square is given by the sum of squares of the
deviations of these r central responses from their mean, Z

J
=1 J
Uv r
)
I 2
,
divided by their degrees of freedom, r-1.
mate of the lack of fit mean square for a first order function is given
An esti-
2
by qr(Y* - Yj ) / (q + r) where Y* is the mean response for the q
noncentral treatments and Yj is the mean response of the r applica-
tions of the central treatment. These lack of fit and error mean
squares may then be compared as a guide to the adequacy of the fitted
function, as discussed previously.
Another approach to the estimation of experimental error in unrep-
licated factorial designs is possible. It is to use the apparent effects of
those effects known a priori to be zero or negligible as the basis of an
estimate of the experimental error. In the absence of prior knowledge,
this procedure is dangerous in that misleading results will be given if
the effects assumed to be negligible do, in fact, play some role. 14 Gen-
erally it is the higher order interactions that are so used, but any effect
that is justifiably thought to be really nonexistent may be used. For ex-
2
ample, with a 4 factorial it would be possible to estimate a third degree
polynomial such as equation 5.14.
(5.14) Y = ft, + ft Xx + 0,X, + 0U X 2

+ 0 22 X 2
2
+ 0 12 Xx X 2 + 0 m X*
2
+ 0222 X| + 0 U2 X X2 + 0 122 X,X 2 + 6
14
See Davies, O. L. (ed.) The design and analysis of industrial experiments. Oliver and
Boyd, London. 1956. Pp. 286-89.
Suppose /3112 and /3 122 coefficients could be assumed to be zero. Then
the sums of squares associated with the X^X 2 and x
X
X 2 effects in the
analysis of variance for multiple regression could be summed and di-
vided by the sum of their degrees of freedom, which is two, to give an
estimate of the error mean square. These interaction sums of squares
could only be conveniently obtained if orthogonal polynomial regression
were used as in Table 4.2. The resultant analysis of variance table
would be as shown in Table 5.4 which, except for the absence of repli-
cation, corresponds to Table 5.1. Note that with the third order inter-
action coefficients of equation 5.14 assumed to be zero, there are only
eight parameters of equation 5.14 to be estimated, including the con-
stant term /3 0 The sum of squares for regression therefore has seven
.
degrees of freedom. Given the information of Table 5.4, comparison of

the lack of fit and error mean squares may be made in the usual way.
A concept related to the use of factorial designs is that known as
confounding. Because of the large number of experimental units nec-
essary for complete factor ializat ion, it may be impossible to study all
of the treatments under uniform conditions. The most efficient method
of overcoming this difficulty, granted the unlikely assumption that a
complete factorial has to be used, is to divide the experimental units
into blocks in such a manner that some or all of the less important in-
teraction effects are confounded or “lost.” Such confounding then al-
lows the more important effects, either pure or interaction, to be in-
vestigated under uniform conditions. The construction and implications
of such designs will not be discussed here. 15 The factor levels and the
treatment arrangements required for the more useful factorial designs
with interaction effects confounded have been listed by Cochran and
Cox. 16 It is also possible to confound main effects. Such confounding
occurs when split-plot designs are used. In general, this fact makes
split -plot designs rather unsuitable for production function estimation
since the main effects in response studies are usually not of negligible
importance.
Fractional Factorial Designs
Complete factorials of greater size than 3 2 3 3 , and 42 , even with-

,
out replication, generally tend to be very wasteful of research re-

sources. This fact has stimulated the development of a number of de-
signs related to the factorial but specifically oriented to the fitting of
continuous response or production functions. Before considering these
response surface designs, however, mention should be made of two
compromise factorial designs that have frequently been used in produc-
tion function research. The first of these is the rather informal
15
For a full discussion of the confounding of interaction effects, see Davies, O. L. (ed.)
Thedesign and analysis of industrial experiments. Oliver and Boyd, London. 1956.
Ch. 9.
Cochran, W. G. and Cox, G. M. Experimental designs. John Wiley and Sons,
Inc.,
New York. 1957. Pp. 234-43.
Table 5.4. Analysis of Variance for the Multiple Regression Estimate of a Second
Degree Polynomial From a 4 2 Factorial With Experimental Error
Estimated From Third Order Interaction Effects
Source of Variation in Y Degrees of Freedom Sum of Squares
Regression 7 SSR
16 A 2
Deviations from regression 8 £ (Yj - Yj)
J
J=i
Lack of fit 6 Found by subtraction

Experimental error 2 From third order interactions
16 2
Total 15 £ yj
j=i
selection by the researcher of some of the experimental points of a

complete factorial. Instead of studying all possible combinations of in-
put factor levels, only some would be studied. Thus if 24 experimental
s
units were available they might be used in an incomplete 4 design by
selecting for study only 24 of the 64 experimental points in the complete
design.
Presumably, researchers using this procedure have been guided by
the desire to obtain observations adequate both in terms of their cover-
age of the relevant surface section, and in terms of having sufficient
observations to enable estimation of the hypothesized response model.
From a statistical point of view, such an informal approach to the de-
sign of an experiment for production function estimation is extremely
naive. It takes no account of the precision or efficiency of estimation,
nor of untoward confounding that might occur. Still, such an informal
selection of treatments does recognize the major aspects of quantitative
response phenomena. It is obviously a more satisfactory procedure
than the classical experimental approach of considering but one or two
factors at a few levels with analysis aimed at the derivation of one or
two point estimates of supposedly discrete phenomena.
The other method of circumventing the large size of a complete fac-
torial when five or more factors are studied is to use fractional facto-
rials. In essence, a fractional factorial design is equivalent to a single
block of a confounded complete factorial. The effect(s) confounded
should be one(s) that the researcher believes to be negligible. Such de-
signs make it possible to gather sufficient data to investigate the more
important factor effects from an experiment a fraction of the size of a
4
complete factorial. Thus a 3 complete factorial would involve 81 treat-
ments and sets of observations. By confounding the higher order inter-
actions, blocks of 27 units constituting a one -third replicate could be
used. One such block would generate 27 sets of observations for the es-
timation of the 15 parameters in a second degree four factor polynomial
model. To gain maximum benefit from such fractional factorials it is
essential that the particular treatments necessary for correct
confounding be used. The researcher does not have to work these out
for himself; he may obtain them from the tables of designs presented in
17
the standard texts on experimental design .
From a statistical viewpoint, the use of fractional factorials for re-

sponse surface estimation is not entirely satisfactory. Like complete
factorials, they lead to the estimation of higher order pure effects with
relatively low precision. In addition, if the effects that are confounded
are in reality not negligible, fractional factorials may lead to a false
picture of the structure of the production process. The parameter es-
timates will be biased by the omission of the significant higher order
effects that are lost by confounding. Nor, without replication, do frac-
tional factorials directly provide an estimate of the experimental error;
although they do provide some degrees of freedom for lack of fit analy-
sis. However, an estimate of the experimental error may be obtained
by replicating the central treatment as outlined above for the case of a
complete factorial. Alternatively some of the unconfounded higher
order interaction effects thought to be of no importance may be used to
estimate the error, as previously explained. Or, again, any a priori
knowledge of the error may be utilized. In these ways it may be possi-
ble to carry out lack of fit analysis.
To summarize, while fractional factorials are not as satisfactory
as the more specialized response surface designs, they do have a large
order of merit. This is especially true insofar as they can often be
used in co-operative research with other scientists not interested in
the fitting of a surface function. For these other scientists the use of a
specialized response surface design of the type discussed below may be
quite unsatisfactory. In other words, fractional factorials may often
enable a compromise to be reached in the generation of data for a va-
riety of research purposes.
Composite Designs
Although fractional factorials are far less cumbersome than com-

plete factorials, they may still require an inordinate number of treat-
ments and experimental units. Moreover, if a polynomial production
function model is to be fitted, there appears to be no compelling reason
why the factorial approach should be adopted. Indeed, other designs are
possible which (a) require considerably fewer treatments than frac-
tional factorials in order to estimate a given number of polynomial co-
efficients and (b) provide estimates that are statistically just as good,
or better, in terms of the number of observations on which they are
based. The designs constructed so far that have these characteristics
are the composite and the rotatable designs. Both these designs were
17
See, for instance, Cochran, W. G. and Cox, G. M. Experimental designs. John Wiley
and Sons, Inc., New York. 1957. Pp. 244-92 and 341. Also Davies, O. L. (ed.) The design
and analysis of industrial experiments. Oliver and Boyd, London. 1956. Pp. 431-90; De
Baun, R. M. Response surface designs for three factors at three levels. Technometrics
’
1: 1-8. 1959.
,
developed for use in industrial chemical experimentation. For such re-

search, the experimental conditions differ markedly from those gener-
ally found in biological and agricultural research. First, the experi-
mental error in chemical response studies is generally small and
relatively constant for experiments of the same type. Secondly, se-
quential experimentation is far more feasible in chemical studies since
the time required for a single trial is but a few days, at most. Conse-
quently, both the composite and the rotatable designs are oriented to
sequential experimentation. Nonetheless, both are eminently suited to
the generation of data for the fitting of agricultural production functions
from either a single experiment or a sequence of experiments. The
most important proviso to their use is that the hypothesized model
should be a polynomial. As will be pointed out later, this restriction is
not too severe. Except where otherwise noted, a polynomial model of
the production process will be assumed in the following discussion of
the composite and rotatable designs. Equivalently, it might be said that
it is assumed that the researcher is satisfied with a polynomial approx-
imation of the true response function. These assumptions relate to the

economic specification of the production function and will be discussed
in Chapter 6.
Composite designs, as most frequently used, consist of a complete
or fractional 2 k factorial supplemented with additional experimental
points or treatments in such manner that all the coefficients of a second
order polynomial in k factors may be estimated. If these extra treat-
ments are arranged symmetrically about the center of the factorial, a
central composite design results. By adjoining the additional experi-
mental points at a distance from a “corner” of the 2 k factorial, a non-
central composite design is obtained.
To construct a central composite design, 2k+l supplementary treat-
ments have to be added to the basic 2 k factorial. One of these additional
treatments is placed at the center of the design and its coded value, in
terms of equation 5.11, is therefore (0, 0, ., 0). The other 2k addi-
. .
tional treatments are placed in pairs along the coordinate axes at dis-
tances of ± ax t a 2 ,
.
., t Q! k from the center, respectively.
, . They
thus have the coded values (t a x 0, 0, . , 0), (0, ± a 2 0,
,
.
. ,
. .
. 0),
,
. .
.
(0, 0, 0,.
., ta k ). For convenience it is desirable that the a values
.
be equal or, what is the same thing, that the 2k additional noncentral ex-
perimental points be equidistant from the center of the design. Such an
arrangement is assumed in the following discussion. As an example,
consider the central composite design for the study of three factors, Xx ,
X 2 and X 3 shown in Table 5.5.
, ,
3
The first 8 treatments are those of a 2 complete factorial. Treat-
ments 9 to 15 augment this factorial, the ninth being at its center while
treatments 10 to 15 are the noncentral supplementary experimental
points. A single replicate of the experiment would give 15 sets of obser-
vations which would be sufficient to enable estimation of the 10 parame-
ters in a three factor polynomial of second degree. If a 3 s complete
factorial were used to estimate such a polynomial, 27 experimental
units and treatments would be required. In this instance the composite
design is thus far less demanding of research resources than the fac-
torial. As Table 5.6 shows, this fact is always true when more than
two factors are involved. Moreover, as k increases the saving that
results from using a central composite design becomes greater.
So far, nothing has been said of the a values that might be used in a
central composite design. They may take any value except zero. More-
over a can be chosen to give any desired compromise between preci-
sion and bias since both the bias that results if the true representation
of the surface is not quadratic, and the precision of estimation, in-
18
crease as a becomes larger. As well, the value of a can be chosen
so as to make the design orthogonal. This, of course, leads to great
computational savings and enables the sum of squares due to regression
to be conveniently separated into a component due to each effect as in
Table 4.2. To obtain orthogonality, a should be 1.000, 1.215, 1.414, or
1.547 for the case of 2, 3, 4, or 5 factors respectively; each factor be-
ing held at levels oft 1 in the basic 2 k factorial. These a values do
not make all the regression estimates orthogonal directly. To take ad-
vantage of the orthogonality, the production function parameters have to
be estimated by fitting a function such as equation 5.15 which corre-
sponds to the three factor case.
(5.15) Y = 77 + + /3
2 X2 + /3
3
X3 + J3 U XU + /3
22 X 22 + /3 33 X 33
+ &* Xx X2 + ^XiX, + /3 2S X2 X3
This equation corresponds to equation 5.13 where X u of equation 5.15
is related to xf of equation 5.13 as shown in equation 5.16, N being the
number of sets of observations.
(5.16) Xu = X? -
The estimated /3 values of equation 5.15 correspond to those of equation
5.13 except for the, rj term. To obtain b 0 the estimate of /30 of equation
,
5.13, the relation shown in equation 5.17 is used.
2 X? + b !2 SX; + b,,SX
(5.17) -)
N
The same general procedure holds for any number of factors.
Another alternative choice of a values
which gives the de- is that
sign the property of rotatability; the estimated response at any point on
the surface having the same variance as any other estimated response
18
Discussion of the statistical technicalities of composite designs is to be found Box, m
G. E. P. and Wilson, K. B. On the experimental attainment of optimum conditions. Journal
of the Royal Statistical Society, Series B, 13: 1-45. 1951.
Table 5.5. Central Composite Design for Three Factors
Treatment X. x2 X 3
1 1 1 1
2 1 1 -1
3 1 -1 1
4 1 -1 -1
5 -1 1 1
6 -1 1 -1
7 -1 -1 1
8 -1 -1 -1
9 0 0 0
10 a 0 0
11 -a 0 0
12 0 a 0
13 0 -a 0
14 0 0 a
15 0 0 -a
the same distance from the center of the design. Such designs are dis-
cussed in the next section.
In addition to alternative selections of the value of a, various ex-
tensions of the central composite design are possible. One modifica-
tion is that obtained by using a number of 2 k factorials instead of a sin-
gle 2 k factorial as the nucleus of the design. Such designs have been
19
studied by Tramel. He found an increase in the precision of estima-
tion resulted from using a nest of three 2 k factorials with each factor
held at levels of ± 1 in the first, ± 2 in the second, and ± 3 in the third
Table 5.6. Number of Experimental Points per Replicate Required for Estimation
of a k Factor Second Degree Polynomial Using Factorial
and Central Composite Designs
k
Design
2 3 4 5 6
3 k factorial 9 27 81 243 729

k factorial 27 243
(1/3) 3 81
Central composite 9 15 25 27* 43*
*Based on augmentation of a k factorial

(1/2) 2
19
Tramel, T. E. A
suggested procedure for agronomic- economic fertilizer experiments,
in Baum, E. L. et Economic and technical analysis of fertilizer innovations and
al. (eds.)
resource use. Iowa State University Press, Ames. 1957. Pp. 168-75.
factorial and an a value of four. Relative to the central composite de-
sign having factor levels of 1 1 and a equal to two, Tramel found that
the triple 2 k design gave worthwhile decreases in the variance of the
estimate of the constant term, i.e., /30 of equation 5.13 for instance, and
of the quadratic and interaction coefficients. The disadvantage of this
design relative to the ordinary central composite design is the number
of experimental units and treatments it requires. Thus while the cen-
tral composite for k factors requires only 2k + 2k + 1 experimental
points, the triple 2 k design requires 3 (2k) + 2k + 1 points, an increase
of 2 k+1 . To this extent the triple 2 k central composite designs are
self-defeating. Relative to the ordinary central composite design, they
do not markedly reduce the size of experiment required to estimate
second order polynomial approximations to the production surface as
compared with complete or fractional 3 k factorials.
The estimation of surface models by multiple regression from cen-
tral composite design data is quite straightforward. For lack of fit
analysis of variance, an a priori estimate of the experimental error
may be utilized, or an estimate of the error may be obtained by repli-
cating the central treatment a number of times. Alternatively, the
whole experiment may be replicated. Whichever procedure is used, the
lack of fit analysis proceeds as illustrated previously for the complete
factorial design.
Noncentral composite designs are obtained by adding k additional
points at a distance from a “corner” of the basic 2 k factorial. They
are most suitable for sequential experimentation when the researcher
has no prior knowledge of the response surface. 20 Of course, in agri-
cultural or biological response studies, sequential experimentation is
only possible under closely controlled conditions. Used in this way, the
first stage experiment would consist of the basic 2 k factorial. In the
second stage, k supplementary treatments would be added about the
“corner” of the original factorial at which the greatest response was
obtained. Thus if there were three factors under consideration and the
first stage factorial experiment indicated that the surface region of
most interest extended away from the treatment (-1, +1, -1), then three
additional treatments, (-1 -a,+l, -1), (-1, 1+ a , -1) and (-1, +1, -1 -a),
would be carried out. In all, the researcher would then have 11 sets of
observations from which a three factor polynomial of second degree
could be fitted. If no a priori estimate of experimental error were
available, the central treatment might be replicated a number of times
at either stage of the experiment to give an estimate of error for lack
of fit analysis.
Noncentral composite designs may also be of some use when
20
Noncentral composite designs were developed for the experimental determination (by
the sequential “method of steepest ascent”) of the point on a response surface at which
some feature, such as response or cost or profit, attains its maximum (or minimum) value,
For such purposes it is not necessary to estimate a response function. However, in order
to draw out the economic implications of a given response process under varying
input-
output price ratios it is essential that a production function be estimated.
sequential experimentation is not possible. For instance, if the approx-

imate location of the ridge lines is known, the noncentral composite de-
sign provides an algorithm for locating relatively efficient treatment
locations within the ridge lines. However, noncentral composite de-
signs are not to be recommended for nonsequential experiments when
the researcher is equally interested in all surface points within some
range of each factor that is independent of the range of other factors.
Rotatable Designs
There are a number of criteria that might be used as a basis of

comparison between response surface designs. In a practical research
context, the most important feature of a design is perhaps the quantity
of research resources required to derive satisfactory estimates of the
production parameters; the definition of satisfactory estimates varying
at the discretion of the researcher. Another criterion is the amount of
information used in estimating the less important parameters of the
surface function. It would be desirable to “allocate” the information
used in estimation among the parameters in proportions corresponding
to their importance. On the above grounds, the composite, fractional,
and complete factorial designs are generally to be preferred, in that
order. Two riders must be attached to this statement: the designs
must be feasible in terms of both (a) the treatments possible and,
roughly speaking, (b) the model to be fitted.
A third principle that a good response surface design should fulfill
has been enunciated by Box and Hunter 21 It is that of rotatability: the
.
variance of an estimated response should depend only on its distance

from the center of the design and not on its direction from the center.
Stated another way, rotatable designs are such that all estimated re-
sponses an equal distance from the design’s center lie on the same var-
iance contour. This principle is intuitively appealing, especially in ex-
ploratory response surface studies where the researcher has a uniform
intensity of interest in all parts of the surface section within the range
of factor levels being studied. Moreover, rotatable designs do not nec-
essarily involve more experimental points than nonrotatable designs.
To date, the concept of rotatability has only been developed in relation
to the estimation of polynomial approximations of the response surface.
This is a disadvantage in that a design that is rotatable for response
estimates based on a polynomial function will not be rotatable in terms
of a nonpolynomial production function. However, since the use of
mul-
tiple regression in production function research generally implies
the
use of a polynomial in one disguise or another, this defect is not too
serious.
21
Box, G. E. P. and Hunter, J. S. Multi-factor experimental
designs for exploring re-
sponse surfaces. Annals of Mathematical Statistics, 28: 195-241.
1957. For a less techni-
cal discussion, see Box, G. E. P. and Hunter, J. S.
Experimental designs for the explora-
tion and exploitation of response surfaces, in Chew,
V. (ed.) Experimental designs in
industry. John Wiley and Sons, Inc., New York. 1958.
Pp. 138-90.
For the fitting of production surfaces, the rotatable designs of most
importance are those of second order, i.e., rotatable designs from
which all quadratic factor effects can be estimated. To a lesser extent,
third order rotatable designs may also be of consequence. It is of in-
terest to note that complete or fractional 2k factorials, with or without
the inclusion of central treatments, are rotatable first order designs.
The construction of rotatable designs of any order is based on the se-
lection of treatments located at the vertices and center of certain regu-
lar figures or combinations of figures in the space of k dimensions
where k is the number of factors being studied. The finer details of
construction are of no concern here; they lie within the province of the
statistician rather than of the production function researcher. It should
be noted, though, that rotatable designs of any order can be arranged in
blocks so that the estimates of the parameters of the polynomial fitted
to the total array of data are uninfluenced by block effects. As might
be expected, these arrangements have to be specially planned; the ap-
propriate details are outlined below. The possibility of such orthogo-
nality between block effects and the production parameters increases
the attractiveness of the rotatable designs. It enables rotatable de-
signs to be used in sequential experimentation without worrying about
time trends. Also, for single stage experiments, such blocking gives
the researcher the opportunity to segregate his experimental units into
more homogeneous groupings either in terms of the units per se or of
the experimental environment; so reducing the experimental error with-
out introducing added effects due to blocks. With or without blocking,
the treatments in rotatable designs should be allocated at random
among the experimental units to be used so as to reduce the biasing ef-
fect of unidentified factors.
Second order rotatable designs

Consider first the second order rotatable designs. A relatively ef-
ficient group of such designs is obtained by using the central composite
design with a , the distance of the supplementary points from the center
of the design, selected so as to make the design rotatable. Using a
complete 2 k factorial as the nucleus of the central composite design, a
rotatable arrangement results if a is chosen to satisfy equation 5.18.
k/4
(5.18) a = 2
Thus for two, three, four, five, or six factors, with a complete 2 k de-
sign as core, rotatability would result with a values of 1.414, 1.682,
k
2.000, 2.378, or 2.828, respectively. If a fractional 2 factorial is used
as nucleus, a must be selected to satisfy equation 5.19 in order to
achieve rotatability
(5.19) a
p k
where (i) a complete 2 factorial used as the nu-
is the fraction of
5
cleus. Thus with five factors and using a (i)2 factorial as nucleus, a
rotatable central composite design would result if a were set equal to
2.000. To avoid confounding of second order effects, k should be
greater than four if a (|)2 k nucleus is used and greater than seven if a
(i)2 k nucleus is used.
A general feature of second order rotatable designs that should be
noted at this point is the role of central treatments. With none or few
replications of the central treatment, the variance of the estimated re-
sponse tends to be larger at the center than at points adjacent to the
center. With many replications of the central treatment, the variance
profile of the estimated response has a deep trough at the center. As a
compromise between these two extremes, it has been suggested that the
central treatment be replicated that number of times which would make
the variance of the estimated response uniform within a radius of one
unit, in terms of equation 5.11, from the center. Designs arranged in
such fashion are known as uniform information rotatable designs. The
replications at the center are also advantageous in themselves. If there
are n c such replications of the basal treatment, they provide n c -1 de-
grees of freedom for the estimation of experimental error. Lack of fit
analysis may then be carried out. Plans for two to six factor rotatable
central composite designs, with the required number of central points
added to give a constant variance of estimation within a distance of one
unit from the center of the design, have been tabled by Cochran and
22
Cox. They also list second order designs in blocks arranged so that
the estimated parameters of the fitted parameters are independent of
block effects.
In addition to the rotatable designs obtained by appropriate selection
of a in the central composite arrangement, numerous other rotatable
designs of second order are possible. With only two factors any dis-
tribution of treatments equally spaced on a circle around the central
treatment is rotatable, provided that the number of experimental points
on the circle is greater than four and that there is at least one observa-
tion using the central treatment. With three factors, a rotatable design
that requires only 12 noncentral treatments in addition to replicates of
the basal treatment is available. The design matrix for this arrange-
ment is shown in Table 5.7. In using this design, a must be chosen to
satisfy equation 5.20, N being the total number of experimental points
including replications of the central treatment. It compares favorably
with the three factor rotatable central composite design which requires
14 noncentral treatments.
(5.20) 2 X\- = 4o- 2 (1.176 J + 1.902

2
= N i = 1, 2, or 3
) ;
j
=l J
22
Cochran, W. G. and Cox, G. M. Experimental designs. John Wiley and Sons, Inc.,
New York. 1957. Pp. 348-50 and 370-75; also, Box, G. E. P. and Hunter, J. S. Experimental
designs for the exploration and exploitation of response surfaces, in Chew, V. (ed.). Experi-
mental designs in industry. John Wiley and Sons, Inc., New York. 1958. Pp. 187-89.
X ,

For more than three factors, the simplest rotatable designs are
those obtained by using the central composite design with an appropri-
ate a value based on equation 5.18 or 5.19.
Ordinary least squares multiple regression procedures may be used
to fit a production function to data generated by second order rotatable
23
designs. Because of the symmetrical arrangement of the treatments,
the regression computations are greatly simplified. This is especially
true for the rotatable central composite designs where the factorial nu-
cleus has equal spacings between factor levels. A short-cut procedure
for estimating the coefficients of the second degree polynomial relevant
24
to second order rotatable designs is also available. This procedure
involves the following calculations where Yj is the observed response
for the j -th treatment; this treatment being denoted by (X^, X 2j, • *
*
X kj) for k factors.

-if N k N ",
(5.21) b0 = AN 2
2
(k + 2) (25 Tj )
- 2Xc 2 2 Xfj Yj
L j=i i=i j=i J
1
(5.22) bi - cN” 2 Xij Yj
j=l
N k N
x 2 2
(5.23) = AN“ J[(k + 2) a- k] c s Xjj Yj - (1 - x c 2 2 Xfj Yj - 2Xc 2 Yj
bxx
j=i
)
x=i j=i j=i

j
J
1
_1 l 2
(5.24) b. = X N" c S X.jX^Yj i / h = 1, 2, . .
. ,
k
h ;
where
1
(5.25) A = [2\{(k + 2)X -k}]'
N
1
(5.26) c = N( £ X?,)"
j=i
(5.27) X = kN[(k + 2)(N - nc )]
The b values are the estimates of the 0 coefficients in the second order
polynomial representation of the surface. N is the total number of
treatments, including n c replications of the central treatment.
The analysis of variance for multiple regression may also be sim-
plified when a second order rotatable design is used; Table 5.8 shows
the appropriate paradigm. The sum of squares, S 0 attributable to fit- ,
ting a zero order polynomial is calculated as indicated in equation 5.28.
23
See, for example, Hader, R. J. et al. An investigation of some of the relationships be-
tween copper, iron and molybdenum in the growth and nutrition of lettuce: I. Soil Science
Society of America Proceedings, 21: 59-75. 1957. This reference also contains a valuable
discussion of the general experimental procedures relevant to composite and rotatable de-
signs.
24
Hunter, J. S. Applications of statistics to experimentation. Statistical Techniques
Research Group, Princeton University. 1958.
Table 5.7. Rotatable Design of Second Order for Three Factors
Treatment X, X 2 X3
1 0 1.902a 1.176a
2 0 1.902a -1.176a
3 0 -1.902a 1.176a
4 0 -1.902a -1.176a
5 1.176a 0 1.902a
6 1.176a 0 -1.902a
7 -1.176a 0 1.902a
8 -1.176a 0 -1.902a
9 1.902a 1.176a 0
10 1.902a -1.176a 0
11 -1.902a 1.176a 0
12 -1.902a -1.176a 0
13 0 0 0
14 0 0 0
N 0 0 0
Table 5.8. Analysis of Variance for Multiple Regression of Data From a

Second Order Rotatable Design With k Factors
Source of Variation in Y Degrees of Freedom Sum of Squares
Regression (k + 2) (k + l)/2 S 0 + S x + S2
b0 term 1 So
bi terms k s3
b^ and b ih terms k(k + l)/2 s2
N 2
Deviations from regression N - (k + 2) (k + l)/2 s Yj - s0 - S -
x
S2
j=i
Lack of fit N - (k + 2) (k + l)/2 - (n c - 1) By subtraction

nc
2
Experimental error nc - 1 S (Yc - Y*)
C=1
N 2
Total N
N 2
(5.28) S0 = ( 2 Y;
J
) /N
J
=1
The extra sum of squares attributable to the fitting of a first order

polynomial, i.e., associated with the b i terms, is given by equation 5.29.
k N
(5.29) Si S bi 2 X-Y.
J J
1=1 j=i
Likewise the additional sum of squares explained by fitting the second

order polynomial, i.e., associated with the squared and interaction
terms, is given by equation 5.30.
N k k N N
(5.30) S 2 = b0 2 Y: + 2 S b ih 2 XyXhjYj
J J
- (2 Y-f/NJ
j=i J 1=1 h=i J
j=i j=i
The experimental error sum of squares is obtained as the sum of the

squared deviations of the n c central treatment responses from their
mean, Y*, as indicated in Table 5.8. Relative to tables 5.1 and 5.2, the
sums of squares of Table 5.8 are in uncorrected terms. However, this
does not interfere with the subsequent analysis. Mean squares for each
of the listed sources of variation may be obtained by dividing each sum
of squares by its associated degrees of freedom. An F test for lack of
fit of the second degree polynomial model may be carried out in the
usual way. The order of the model may also be tested by calculating
the F value with k(k + l)/2 and (n c - 1) degrees of freedom for the mean
square of the second order coefficients, bn and bih, as in equation 5.31.
2
(5.31) F = 2 S 2 (n
C
- l)/k(k + 1) 2 (Y c - Y*)
C=1
If this F value is significant at the desired probability level, it is an in-

dication that the second degree model provides a better approximation
to the surface than a first degree model. Should the F value of equation
5.31 be nonsignificant, the mean square, S^/k, of the first order coeffi-
cients may be tested to give an indication of the adequacy of a first
order model.
Third order rotatable designs

A number of third order rotatable designs have been devised by
Gardiner, Grandage, and Hader. 25 Since these designs are not tabled in
the standard texts on experimental design, some of the more useful ar-
rangements will be outlined below. The use of third order designs pre-
sumes, of course, that the researcher believes a third degree
25
Gardiner, D. A. et al. Some third order rotatable designs. Mimeo Series No. 149.
Institute of Statistics, Raleigh, N. C. 1956. Gardiner et al. Third order rotatable designs.
Ann. Math. Stat., 30: 1082-96. 1959.
polynomial would fit the production surface more advantageously than a

second degree model. In the past, production function researchers have
generally been satisfied with functions of second order. However, as
they become more facile with the experimental procedures and prob-
lems involved, and with the advent of better experimental facilities
making possible greater control of the experimental conditions, it is
26
likely that third order designs will become increasingly useful.
With two factors, the simplest set of third order rotatable designs
consist of experimental arrangements with seven or more treatments
spaced equally around each of two concentric circles. In all, a mini-
mum of 14 treatments is therefore required. The central treatment
may be replicated a number of times to provide an estimate of the ex-
perimental error for lack of fit analysis. For instance, if it were de-
sired to estimate all the 10 parameters of a two factor third order
polynomial such as equation 5.14, a design might be used with seven
points on each of two concentric circles lying in the surface area of in-
terest. If the (0, 0) treatment were replicated three times, two degrees
of freedom would be available for estimating the experimental error.
It is possible to arrange a third order two factor rotatable design in
two blocks such that the first block constitutes a second order rotatable
design. Such an arrangement is extremely attractive for sequential ex-
perimentation. If the second order polynomial fitted to the first block
is felt to be unsatisfactory, the researcher may proceed to carry out
the second block of treatments. A third order approximation to the
surface could then be fitted using the data from a third order rotatable
design constituted by the two blocks. For such a sequential design, the
first block should consist of seven or more points equally spaced
around the circumference of a circle within the surface region of inter-
est and having its center at the experimental point corresponding to the
coded treatment (0, 0). By replicating this central treatment one or
more times, a rotatable second order design is obtained. The second
block should have seven or more treatments arranged on a circle of
different radius but with the same center as that used in the first block.
It is desirable that the estimates of the parameters of the third order
polynomial be free of block effects. For such to be the case, the radius,
r 2 of the circle used in the second block should be chosen to satisfy
,
equation 5.32. In this equation rx is the radius of the circle used in the
first block, n x is the total number of treatments used in the first block
and n ci is the number of treatments at the center of the first block.
(5.32) r* = rX - n ci )/n x
26
In this regard, it is interesting to note that the Australian Commonwealth Scientific
and Industrial Research Organization has recently installed a Phytotron consisting of 300
six feet by three feet cabinets. Each of these experimental units is under separate control
in terms of simulated climatic conditions, in addition to the normal factors such as ferti-
lizer, soil type, etc., that a researcher might introduce. Under such conditions the use of
third order designs could not be necessarily classed as impractical; nor could the simulta-
neous investigation of seven or eight factors in a series of sequential experiments be dis-
missed as too unwieldy.
Thus it would be possible to carry out a sequential two factor third
order rotatable design with a minimum of 15 treatments. Of course, to

estimate experimental error a few replicates of the central treatment
would be required, assuming a lack of a priori information about the
error.
For three factors, perhaps the most useful design is that shown in
Table 5.9 where a is chosen to satisfy equation 5.33 for a total of N
treatments.
N 2 2 2
(5.33) 2 X xj = [2(1.82969* + 1.16343 )
+ 8(1.25992*) ]a = N ;
i = 1, 2, or 3
Itrequires 32 noncentral treatments and, for best results, should be

supplemented by at least four or five replicates of the central treat-
ment. Five such central replicates would give a total of 37 sets of ob-
servations from which efficient estimates of the 20 parameters of equa-
tion 5.34 could be obtained. Four degrees of freedom would be available
for estimation of the experimental error.
(5.34) Y = 0O + b 0! X i
+ S /SijXiXj + , .
2 0ijk x i
X, Xv + e
1=1 J^l=l
Three factor rotatable designs of third order that may be carried

out sequentially in two blocks with the first block constituting a second
order rotatable design are available. They require an over -all larger
number of treatments and experimental units than does the nonsequen-
tialdesign of Table 5.9. However, fewer treatments are required within
each stage of the experiment. Table 5.10 gives the design matrix for
such a three factor rotatable arrangement of third order for a satisfy-
ing equation 5.35.
N
(5.35) S X?=
J
= N ;
i = 1, 2, or 3
j=i
The values of d, e, f, and g in Table 5.10 are 1.414214, 0.341564,

1.286527, and 1.985406, respectively. The first stage requires 20 non-
central treatments while the second block has 30. An arbitrary number
of replications of the central treatment may be used in each block to
give estimates of the experimental error. To make the estimated coef-
ficients of the fitted third order polynomial free of block effects, the
number of central treatments used in each block, denoted by n ci and
n C2 , respectively, should approximately satisfy equation 5.36.
(5.36) n c2 = 2.2n cl + 14.1
Thus for n C1 equal to two, three, or four, n c2 should be 19, 21, or 23, re-
spectively. The total number of treatments would then be 71, 74, or 77,
respectively. Using the two blocks sequentially, the researcher would
a
Table 5.9. Rotatable Design of Third Order for Three Factors
Treatment x2 X 3
1 a a a
2 a a -a
3 a -a a
4 a -a -a
5 -a a a
6 - a a -a
7 -a -a a
8 -a -a -a
9 1.82969 0 0
10 -1.82969a 0 0
11 0 1.82969a 0
12 0 -1.82969a 0
13 0 0 1.82969a
14 0 0 -1.82969a
15 1.25992a 0 1.25992a
16 0 -1.25992a 1.25992a
17 -1.25992a 0 1.25992a
18 0 1.25992a 1.25992a
19 1.25992a 0 -1.25992a
20 0 -1.25992a -1.25992a
21 -1.25992a 0 -1.25992a
22 0 1.25992a -1.25992a
23 1.25992a 1.25992a 0
24 1.25992a -1.25992a 0
25 -1.25992a -1.25992a 0
26 -1.25992a 1.25992a 0
27 1.16343a 0 0
28 -1.16343a 0 0
29 0 1.16343a 0
30 0 -1.16343a 0
31 0 0 1.16343a
32 0 0 -1.16343a
33 0 0 0
34 0 0 0
N 0 0 0
Table 5.10. Noncentral Treatments for Rotatable Design of Third Order
for Three Factors in Two Blocks
Block Treatment X L
X2 X 3
1 a a a
2 a a -a
3 CL -a a
4 a -a -a
5 -a a a
6 -a a -a
7 -a -a a
8 -CL -a -a
9 da 0 0
10 da 0 0
11 -da 0 0
12 -da 0 0
13 0 da 0
14 0 da 0
15 0 -da 0
16 0 -da 0
17 0 0 da
18 0 0 da
19 0 0 -da
20 0 0 -da
21 ea fa fa
22 ea fa -fa
23 ea -fa fa
24 ea -fa -fa
25 -ea fa fa
26 -ea fa -fa
27 -ea -fa fa
28 -ea -fa -fa

29 fa ea fa
30 fa ea -fa
31 -fa ea fa
32 -fa ea -fa
33 fa -ea fa
34 fa -ea -fa
35 -fa -ea fa
Table 5.10 (Continued)
Block Treatment x2 X 3
2 36 -fa -ea -fa

37 fa fa ea
38 -fa fa ea
39 fa -fa ea
40 -fa -fa ea
41 fa fa -ea
42 -fa fa -ea
43 fa -fa -ea
44 -fa -fa -ea
45 ga 0 0
46 -ga 0 0
47 0 got 0
48 0 -ga 0
49 0 0 gat
50 0 0 -ga
not need to carry out the second block of treatments if the second order
polynomial fitted to the first stage block should prove satisfactory.
With four factors the most practicable third order rotatable design
is that shown in Table 5.11. Replications of the central treatment are
not listed. The values of d, e, and f are 1.200919, 0.256303, and
1.736604. Again a must be chosen so as to satisfy equation 5.35 for a
total of N experimental points. As indicated in the table, the design
matrix may be split into two blocks. The first block is a second order
rotatable design. Hence, this experimental arrangement may be used
effectively for sequential experimentation. To eliminate block effects,
n ci and n c2 should be selected to satisfy equation 5.37 with, of course,
(5.37) n c2 = 3.3n ci - 26.7
Thus with ncl equal to nine, the central treatment of the second block
should be replicated four times. Such an experiment would provide 141
sets of observations from which each of the 35 effects in a third order
four factor polynomial could be estimated.
For more than four factors, third order rotatable designs require
far too many experimental points to be useful. While some degree of
fractional replication is possible with rotatable central -composite de-
signs for five or more factors, it is not feasible with third order de-
signs for less than seven factors. Fractional replication with less than
seven factors is impossible without confounding some of the second
Table 5.11. Noncentral Treatments for Rotatable Design of Third Order
for Four Factors in Two Blocks
Block Treatment x* X3 X,
1 1 a a a a
2 -a a a a
. . . . .
• • • . .
* • • * -
16 -QL -a -a -a
17 2a 0 0 0
18 -2a 0 0 0
• • . . .
• •
23 0 0 0 2a
24 0 0 0 -2a
2 25 da da ea ea
26 da da ea -ea
40 -da -da -ea -ea

41 da ea da ea
56 -da -ea -da -ea

57 da ea ea da
• • • . .
• • • . .
* • * •
120 -ea -ea -da -da

121 fa 0 0 0
122 -fa 0 0 0
. . .
r
• - •
.
- * . . .
127 0 0 0 fa
128 0 0 0 -fa
and third order effects. Still, if the researcher believed some of the
third order effects to be negligible, he could confound them and so re-
duce the size of the experiment by fractional replication. Assuming in-
terest in all effects up to and including third order terms, a five factor
rotatable design would require nearly 400 treatments. For six factors,
approximately 1,000 treatments would be required. On the other hand,
5 6
complete 4 and 4 factorials would necessitate 1,024 and 4,096 treat-
ments respectively! By comparison, therefore, the rotatable designs
of third order are still relatively efficient. Likewise, rotatable designs
of higher than the third order would appear attractive in comparison
with the corresponding factorial design. Nonetheless, they would be
impractical in terms of the experimental resources required. More-
over, it appears most unlikely that polynomials of higher order than
three will ever be warranted as approximations to the production sur-
faces found in agriculture.
So far as estimational procedures are concerned, data from third
order rotatable designs may be fitted by ordinary least squares proce-
dures. Analysis of variance oriented to lack of fit analysis may be car-
ried out in the usual way. If desired, the additional sum of squares ex-
plained toy fitting a third degree polynomial instead of a second order
function may be calculated by subtraction and tested for significance.
COLLECTION OF NONEXPERIMENT AL DATA
Under experimental conditions the researcher has the opportunity of

exercising ex ante control over the generation of data. The decisions
as to which factors will be allowed to vary, and the extent of their vari-
ation, are made by the experimentalist. Moreover, he can choose the
experimental units through which the effects of these factors and their
variation are to be studied. Such direct control over the generation of
data is impossible under nonexperimental conditions. With real-world
data the researcher can generally only exercise ex post procedures in
purposively selecting his data. The primary decisions as to input factor
levels are made by the real-world controllers of the production process
and not by the researcher. Still, there is no sharp line of demarcation
between experimental and nonexperimental data. By exercising a suffi-
cient degree otex post control, a researcher might sometimes be able
to obtain a set of real-world data comparable to experimental data in
terms of the factor levels and production units to which the data relates.
Conversely, some experimental data is only partially generated under
the researcher's control. For instance, the factor levels studied in
many livestock response experiments are determined by the animals
and not by the researcher.
As previously stressed, experimental data is in many ways more
preferable for the fitting of production functions than nonexperimental
material. Paradoxically, the major disadvantage of experimental data
lies in the rigid control that it implies. Such control often leads to
estimates having much lower variances than the variance ruling in the
real-world, thereby reducing the usefulness of the fitted function for
extension purposes. Still, this disadvantage could be overcome if ex-
periments of sufficient size to take account of all sources of real-world
variation were used. Such experiments, however, tend to be infeasible
because of restrictions on research resource availability. Real-world
data may then have to be used, balancing its probable inaccurate speci-
fication of the function against the enhanced usefulness of the estimates
in an extension context. Also, under many circumstances, an experi-
mental approach even on a small scale is impossible because of the ex-
penses involved. Thus the estimation of firm production functions, un-
like the estimation of functions for individual physical technologies,
necessitates the use of real-world data. No research program would
have sufficient resources to set up an experiment with firms as experi-
mental units let alone provide inputs on the scale required for a range
of experimental treatments. Hence it is not surprising that the most
common use of real-world data in production function estimation occurs
in the fitting of functions for whole firms or enterprises within a firm.
In discussing nonexperimental data, therefore, the general orientation
will be towards the estimation of farm -firm production functions. Ex-
amples used will relate to these situations. It must be pointed out,
however, that the procedures and problems mentioned also relate to the
estimation of production functions for particular technologies from
27
real-world data .
Just as with experimental data, nonexperimental observations may

be either time series or cross-sectional. Sometimes a mixture of both
types may be used. So far as collecting nonexperimental data, two ap-
proaches are possible. Use may be made of data already recorded for
some other purpose. Alternatively, data may be collected directly from
the real-world decision makers associated with the production process.
Either of these procedures may be used with both the time series and
cross-sectional approaches. For obvious reasons, time series data
have usually been transcribed from data already recorded in such form
as census tabulations and farm management accounts and records.
Conversely, cross-sectional data have most frequently been obtained by
survey procedures, usually of the personal interview type 28 Inherently, .
both these methods of collecting data are similar. Both imply the sam-
pling of some population relevant to the production process under study.
Whichever method of collecting real-world data is used, the re-
corded observations must satisfy a number of criteria. Chief among
these is that the data be relevant to the production function that it is
27
For examples of the use of real-world data in the derivation of production relationships
for particular technologies, see Clark, J. and Bessell, J. E. Profits from dairy farming.
Bulletin No. 7. Imperial Chemical Industries Ltd., Central Agricultural Control, London.
1956.
2S
If worthwhile data is to be obtained, it is essential that the survey be conducted with
care. For a discussion of the mechanics of survey procedures, see Cannell, C. F. and Kahn,
R. L. Collection of data by interviewing, in Festinger, L. and Katz, D. (eds.) Research
methods in the social sciences. Dryden Press. New York. 1953. Pp. 327-80.
, ,
desired to estimate. For a given production function, there will be a

population of situations where the function is to be found in operation.
Only data from these situations should be used to estimate the produc-
tion function. Quite obviously, to use data relevant to some other pro-
duction function can only vitiate the analysis. Stated in such fashion,
this primary requirement of the data appears rather trite. As a prac-
tical matter, however, it is quite a stumbling block; rarely, if ever, can
it be fully satisfied empirically. The difficulty stems from the myriad
of input factors relevant to any real-world response phenomena. Sup-

pose it were possible to take account in a single continuous function of
all the inputs X
x to X k relevant to some product Y. Provided that each
factor was recorded at a sufficient number of levels, an over -all pro-
duction function could be fitted. Such a function is given as equation
5.38.
(5.38) Y =f(X ,X 2 1 ,
Xk )
As a practical matter, though, account cannot be taken of the totality of

input factors. Not only are they too numerous, but some may as yet
have no satisfactory scale of measurement. So far as the number of
inputs is concerned, researchers have frequently aggregated various
types of factors into a small number of categories so as to simplify the
computational problem. However, as shown in Chapter 6, aggregation
is not without danger. It may lead to a meaningless specification of the
production function. For the moment, to simplify the exposition, we
assume no aggregation of the inputs. Since account cannot be taken of
5.40
all the k inputs of equation 5.38, the researcher has to compromise by
attempting to fit a function such as equation 5.39 with factors X g+1 to X k
fixed at known or unknown levels.
(5.39) X
Y = f(X 1? 2 , ..., g /X g+1 , ..., Xk) X
5.40
5.40
But the function that is actually fitted will always be akin to equation
because the researcher using real-world data can never exercise
sufficient control to prevent some of the unrecorded factors, say X^+i
to Xk, from varying.
(5.40) Y = f(X 1? X2 , ..., X g /X g+1 , ..., Xk //X h+1 , ..., Xk )

+ €
While equation 5.39 specifies a particular production function, equation

does not. There are as many functions corresponding to equation
as there are possible levels |of the variable but unrecorded factors
X h+l to X k , i.e., an infinite number of functions assuming these factors
to be continuous. In a given sample of real-world data, either time se-
ries or cross-sectional, each vector of observations (Y,Xr ,X 2 Xg ) ,
. .
.
will be associated with some particular set of values of the factors X h+1
to X k Each such set of values locates a different production surface in
.
the space with axes ,X^Y^ X Fitting a function to the sets of

,
. .
.
g
.
,

observations on Y,X 19 . X g therefore results in the fitting of a hy-
.
.
brid surface. The same result would ensue if some of the factors X 1 to
X had been aggregated.
g
Figure 5.4 illustrates the problem of hybridity when only variations
in a single input, X
1? are recorded and analyzed. For simplicity, sup-
X X
pose there are only two other inputs, 2 and 3 relevant to Y. Assume
,
X 2
X
remains fixed between observations on Y and x , but that 3 varies X
between observations. If three sets of observations were taken, they
might locate the points D, E, and F in the YX 1 plane, as shown in the
figure. Fitting a function to these observations gives the curve DEF as
an estimate of the production function. However, since X 3 varies be-
tween observations, the points D, E, and F lie on three different produc-
tion functions in the YX x plane. These are shown as the curves OA, OB,
and OC. The fitted function is a hybrid incorporating points lying not on
a single production function but on a number of them. Of course, if the
levels of X 3 were recorded and incorporated in the analysis, there would
be no difficulty. The sets of observations would all lie on the same pro-
duction surface in the space with axes Y, X l9 and X 3 .
The extent to which the fitted hybrid function misinterprets the

structure of the production process depends upon the importance of the
unobserved variable factors relative to the observed factors, and also
on the extent of variation in the unrecorded variables. If these factors
vary but little, or are only of negligible import, then the production sur-
faces associated with the sets of observations will lie close together.
The hybrid function might then be taken as providing a satisfactory esti-
mate. 2 ® But if the influence of the factors Xh+x to X k is not negligible,
then the true surfaces associated with the vectors of observations will
not lie close together and the fitted hybrid may be quite atypical of the
true response process. The problem of hybridity also arises with ex-
perimental data. However, the essence of the experimental approach is
the element of control that the researcher may exercise. The experi-
mentalist can, in large measure, ensure that only unimportant factors
occur in the subset Xk+1 to X k of unrecorded variable factors. Such
control cannot be conveniently exercised in the collection of real-world
data. Especially is this true when the cross-sectional approach is used.
With time series data the situation is not so bad because within a single
production unit there is less tendency for the “fixed” factors to vary
over time, at least in terms of the periods usually covered by time se-
ries observations. Also, it should be noted that even if the function of
equation 5.39 could be estimated, or if a satisfactory hybrid function
corresponding to equation 5.40 was estimated, these functions would not
be of wide applicability. They refer only to the levels at which the var-
iables X g+1 to Xk are fixed or approximately fixed. If these variables
have nonnegligible influence when varied, then the fitted function would
29
For an empirical study in which such an interpretation is used consciously, see
Jawetz, M. B. Farm size, farming intensity and the input- output relationships of some
Welsh and West of England dairy farms. University College of Wales, Aberystwyth.
1957.
Figure 5.4. Relation between time

and hybrid production
functions for Y = f(X / ).
x
not serve satisfactorily for estimates involving markedly different

levels of the variables X h+1 to X
k . For this reason, it is desirable that
in collecting real-world data attempts should be made, so far as possi-
ble, to have the fixed variables at levels close to their modal values.
The fitted function might then be applied more extensively in a policy or
extension context than would otherwise be the case.
Another problem pertinent to the collection of real-world response
data is that of ensuring that information is recorded on all the variables
thought relevant. Like the previous problem, this one also appears
rather trite. Most often, the difficulty is not that the more important
input factors are unknown, but that their level of use is not available in
sufficiently disaggregated form. The researcher may be forced to use
aggregated input categories rather than analyzing response in terms of
the individual factors. This problem is most acute with data abstracted
from such sources as census tabulations or farm records. With survey
data gathered especially for production function estimation it may be
overcome, at least to some extent. It is only necessary that a suffi-
ciently detailed survey be undertaken and that the respondents be capa-
ble of supplying the data required. However, even if information is
available on all the important factors taken individually, there may be
too many of them from a computational viewpoint. The researcher may
then have to aggregate the data, just as if disaggregated information
were not available. Of course, it is preferable for the researcher to be
able to aggregate the data at his own discretion rather than to have it
forced upon him, in perhaps undesirable fashion, by the data itself.
The third criterion that nonexperimental data must meet is that it
adequately cover the section of the production surface in which the re-
searcher is interested. Certainly, the observations should be scattered
.

sufficiently to avoid problems of multicollinearity Should some of the
.
inputs be collinear per se, then the redundant variables must be omitted.
If the correlation between certain inputs is extremely high throughout
the relevant population, there is no means of selecting a meaningful
sample so that the correlation is low. However, it is far more likely to
find multicolline ar ity arising because of institutional factors — such as
the traditional use of inputs in fixed proportions. The researcher must
then seek observations to include in his data from production units free
of the institutional factors causing multicollinearity. More usually, the
problem will be one of approximate rather than exact multicollinearity
For instance, suppose a random sample were taken of the population of
firms relevant to a given production function. It is likely that a scatter-
ing of input observations akin to those of Figure S.l.b would result. To
avoid such concentration of the observations within a small segment of
the production surface, sampling must be purposive rather than random.
The best procedure is to decide on the range of interest for each factor,
divide this range into a number of equal segments and for each combi-
nation of factor ranges obtain an equal number of sets of observations.
The procedure is illustrated schematically in Table 5.12 for the case of
three factors A, B, and C with each factor range divided into three seg-
ments. These sub-ranges are denoted by the subscripts 1, 2, and 3.
For each section of the surface corresponding to the factor range
A^j C^, (i,j,k = 1,2,3), an equal number of observations would be
sought.
Such procedures are obviously related to the arrangement of a fac-
torial experiment. The difference is that instead of using definite pre-
selected input quantities as the experimental treatments, the real-world
“treatments* are chosen so as to lie within some range of combinations
of the input factors. The result of such a procedure, known as an ana-
lytic survey, is to produce a pattern of observations roughly correspond-
ing to Figure 5.1.c. Just as in the case of experimentation, the design
of an analytic survey may be modified to give a pattern of observations
analogous to the fractional factorials. The essential feature of the de-
sign of an analytic survey is the emphasis on all parts of the possible
range of the variables examined and not just on point estimates of the
means of the variables. 30 The procedure recognizes that a sample which
is designed to estimate mean values is not an efficient one for the esti-
mation of relationships covering the entire relevant range of the varia-
bles. Controls, analogous to those used in experiments, may also be in-
troduced in survey procedures. For instance, a further restriction
might be introduced into the design of Table 5.12 that the additional fac-
tors D and E be at a constant level throughout the sample. Such con-
trolled sampling leads to far more meaningful estimates of the produc-
tion parameters. However, they have the disadvantage of reducing the
30
See Anderson, R. L. Proper planning reduces research errors. Journal of Farm
Economics, 35: 572-81. 1953. For an example of the collection of data from previously
recorded material by means of an analytic survey, see Antill, A. G. Towards a production
function for dairy farms. The Farm Economist, 8: 1-11. 1955.
Table 5.12. Design for an Analytic Survey Involving

Three Factors - A, B, and C
A x A2
Bi B 2 B3 Bl B2 Bs
Ci
C2
C3
size of the population to which the estimates are relevant. From this
point of view, the rougher estimates from an “uncontrolled" analytical
sample may be of more use in a policy or extension context. Whether
or not this is so, depends on the importance of the factors such as D
31
and E above .
Analytic survey procedures are, of course, infeasible for the collec-

tion of time series data unless a number of production units are to be
observed through time. For a single production unit it is difficult to ob-
tain a satisfactory scattering of time series observations since input
levels in succeeding production periods tend to be highly correlated;
leading, of course, to the estimational problems associated with serial
and autocorrelation. In addition, there may be time trends in the re-
corded input levels that have to be taken into account 32 .
With real-world data, the researcher has also to be particularly

wary of errors in the data. Unlike the experimentalist, he has no direct
record of the input and output quantities. The data come to him in
second- or even third-hand fashion. As emphasized in discussing esti-
mational procedures, error -ridden data can only lead to shabby esti-
mates. Moreover there may be errors of measurement due to lack of
homogeneity, i.e., quality differences between and within observations
on what is regarded as a single input or output factor. The within -
observation problem of quality differences is really a forced type of ag-
gregation —
ideally each separate quality of a factor should be regarded
as a separate factor. Aggregation of inputs, if it is necessary, should
then ideally be based on an index of the comparative production contri-
bution of each quality level. Between observations, each observation
might likewise be adjusted in terms of some standard quality of the fac-
tor, whether it be an input or an output. The problem, of course, is that
the required standards are not available; nor, in most instances, are the
quality differences ascertainable — either within or between observa-
tions of a “single" factor. Such problems of equating the itheoretical
31
An example of controlled sampling is to be found m Heady, E. O. Resource productiv-
ity and returns on 160-acre farms in north central Iowa. Iowa State University Research
Bulletin 412. Ames, July, 1954.
32
See Phelps Brown, E. H. The meaning of the fitted Cobb-Douglas function. Quarterly
Journal of Economics, 71: 546-60. 1957.
product or production factors with observed outputs or inputs arise
most frequently in the estimation of production functions for farms or
for enterprises within farms — especially for the land, machinery, and
labor inputs. Since these are really problems of economic specifica-
tion, further discussion of the influence of such measurement problems
on the derived estimates is deferred till the next chapter. It might be
noted, though, that as elaborated in Chapter 4, the use of weighted re-
gression procedures may enable these problems to be overcome; pro-
vided that the additional necessary assumptions can be made.
Another problem that arises in the collection of real-world data re-
lates to the basis of measurement to be used for inputs and outputs.
The best approach is to gather observations in terms of physical quan-
tities. If necessary, and feasible, these may then be adjusted in terms
of standard physical indices to take account of quality differences within
factors. However, real-world data is frequently only available in
money terms. The recorded observations are then relative to the
prices ruling at the time the product or factors were sold or bought. If
the data cannot be converted back into physical terms, the researcher
may have no alternative but to base his estimates on money values.
Use of the derived estimates is then limited to situations in which com-
parable price relationships exist. Moreover, if the researcher does not
know the original quantity -price relationship, he will never know when
a comparable price relationship exists between the variables!
Finally, the number of real-world observations to be taken must be
determined with regard paid to the number of parameters to be esti-
mated. At least this many sets of observations should be obtained, and
preferably more so that sufficient degrees of freedom will be available
for testing purposes.
CHAPTER 6
Economic Specification
of the Production Function
N PLANNING production function research, the first step must be an

attempt to specify the production function in an economic sense. Only
I then is it appropriate to consider the questions of data collection and
statistical estimation, all the while cycling back to the economic speci-
fication hypotheses to adjust them to what is feasible in terms of data
collection and statistical estimation. Indeed, none of these aspects of
production function research — economic and statistical specification
and data collection — is independent; each influences the others. It is
only for expository convenience that we have rather arbitrarily segre-
gated them. That we have left the discussion of economic specification
till last may seem rather perverse. However, such treatment is justi-
fied— to the uninitiated reader the problems of economic specification
are probably the hardest with which to grapple, unless he has some un-
derstanding of the associated problems involved in statistical estima-
tion and data collection.
The problem of economic specification in econometric models is a
broad one. For production function research, some of the questions it
raises are the following: What specific production function is it desired
to estimate? Is this function independent of other economically mean-
ingful relations in the milieu of the production process? Or is it em-
bedded in a system of economic relations? If the latter, do the relative
time relationships imply a causal chain type representation or a set of
simultaneous relations? What variables are relevant to the production
function? Must some of these variables be aggregated and how might
such aggregation be carried out? What is the algebraic form of the pro-
duction function — continuous or discontinuous, linear or nonlinear?
What economic implications should the function sustain? Is there any
a priori information available about the economic and physical logic of
the production process that may help to answer the above questions?
Such are some of the questions pertinent to the economic specification
of the production function. To the extent that they are answered incor-
rectly, the estimated production function will be biased. It will tend not
to reflect the true structure, in an analytical economic sense, of the
production process.
196 ECONOMIC SPECIFICATION
INTER AND INTRA PRODUCTION UNIT FUNCTIONS
Perhaps the commonest problem of economic specification in em-

pirical production function research is that of hybridity. The major ele-
ments of this problem have already been discussed with reference to
the collection of real-world data in Chapter 5. To reiterate, the prob-
lem of hybridity arises because of the impossibility of measuring all of
the input variables relevant to a given production process. Relative to
a sample of observations, either time series or cross-sectional, there
will nearly always be a number of relevant factors that are unmeasured
but variable throughout the sample. Because of his neglect of these fac-
tors, the researcher cannot help but fit a hybrid function. Of course,
the neglect of these variable but unobserved factors is generally forced
upon the researcher because of the limited availability of research re-
sources.
The problem of hybridity in empirical production function research
was the subject of debate, in the economic literature, involving the va-
1
lidity of Douglas’ pioneering research on macro -production functions.
To a degree, this literature is rather confusing, the discussion being in
terms of inter and intra firm functions. In the present context, it is
better to think in terms of production units rather than firms. The pro-
duction unit might be a firm, a region, or a glasshouse pot — or any
other vehicle for the production process. “Intra” or “inter” then refer
respectively to the production function estimated from time series ob-
servations on a given production unit or to the function estimated from
cross-sectional observations on a number of units. More importantly,
the discussion in the literature is blighted by the fact that only very
broad aggregates of input factors — capital and labor — are considered.
In reality, both of these input categories encompass a large number of
essentially different input factors, especially the capital category. Be-
cause of this level of aggregation, it appears meaningful to postulate
that each of the firms in a cross-sectional study is most likely to be
operating on a different production function. However, so long as the
output and inputs are each measured in some homogeneous terms, there
are no grounds for saying that cross -sectionally estimated production
functions must be hybrid unless unobserved variations occur in some
relevant factors. Hybridity will not be caused by imperfect aggregation
in the absence of unobserved variable factors. Rather, imperfect aggre-
gation per se might be said more strictly to lead to other types of bias
in the derived estimates.
The implication in the literature that time series data may not lead
to the problem of hybridity is also not entirely true. For each set of
time series data that is used, the researcher must decide whether there
has been significant variation over time in any of the nonnegligible
x
See Reder, M. W. An alternative interpretation of the Cobb-Douglas function. Econo-
metrica, 11; 259-64. 1943; and Bronfenbrenner, M. Production functions: Cobb-Douglas,
mterfirm, intrafirm. Econometrica, 12: 35-44. 1944.
ECONOMIC SPECIFICATION 197
unobserved factors, at least insofar as he can, granted that the factors

are unobservedl
It is perhaps worth noting that the estimated production function of
a firm, if hybrid, may still be of interest. Intuitively, it seems that

firms are probably far more likely to tend to move along the hybrid
surface as time passes than they are to move along the theoretically
correct production surface. In terms of policy implications, therefore,
a fitted hybrid function may be of use insofar as it depicts the growth
path over time of firms' input-output relationships. Still, the hybrid
function is of no use in terms of the usual implications that may be
drawn from correctly estimated production functions, unless the hybrid
function is sufficiently akin to the correct functions. If the correct
functions do depict surfaces lying close together, then the implication is
that the time path of firms' input-output relations tends to follow a
movement rather similar to that along a single production surface.
As noted previously, the problem of hybridity may also arise under
experimental conditions. There, however, it is not so important since
the researcher has greater opportunity to ensure that the uncontrolled
factors are only unimportant ones.
SPECIFICATION OF THE MODEL
In formulating an economic model of the production process he

wishes to study, the researcher faces three main tasks. First, he has
to decide whether a single equation or a system of equations is appro-
priate. Secondly, he has to choose the set of variables that are relevant
to the model. Lastly, hypotheses have to be made, and tested, as to the
most appropriate algebraic form of the equation(s). The ideally correct
answers to these questions lie in the logic — economic, biological, or
physical — underlying the production process. In such terms, a given
production process has some "true" functional representation, perhaps
of a stochastic nature, involving some definite set of variables either in
a single equation or a set of equations. However, in an empirical re-
search context, it is generally impossible to answer these questions of
economic specification. On the one hand, the economic, physical, and
biological logic of the production process is usually in large degree un-
known. This is especially true of biological processes where the un-
derlying logic involves such phenomena as cell mechanisms, photosyn-
thesis, nutrient absorption, hormone, and enzymatic processes. For
purely mechanical production processes involving no biological ele-
ments, the problem of economic specification is not so great — at least
ideally. As a practical matter, however, difficulties arise from a lack
of knowledge of the logic of entrepreneurial decision-making processes.
Apart from such general lacunae in understanding the production
logic, there is another difficulty that besets the researcher in his at-
tempt to formulate a model of the production process. For empirical
research, the model has to be not only logically sound but also
computationally feasible. Otherwise, the model can only be an ivory

tower plaything. Generally, the construction of a computationally feasi-
ble model implies that the researcher has to compromise with the the-
oretically ideal model - two types of adjustment usually being neces-
sary. First, the number of separate variables that may be considered
has to be determined in terms of data availability, and also with regard
to the resources available for estimation. As the number of variables
increases, the work involved in computation grows more than propor-
tionately. Secondly, it is necessary to use a functional representation
that is statistically manageable, both in terms of estimation and of test-
ing. For these reasons, the construction of an economic model of a
given production phenomenon is not independent of the problems of data
collection and statistical estimation.
Single and Multi -Equation Models
A little has already been said in Chapter 4, from an estimational

viewpoint, of the question of single versus multi -equation models in
production function research. The crux of the problem is a question of
fact. Can the production process be satisfactorily represented by a
single unilateral causal relationship between output and inputs or are
these variables so mutually interdependent that the process can only be
satisfactorily represented by a system of simultaneous equations? Or
should the production process be depicted as a chain of causal relation-
ships, the dependent variable in each prior equation entering later
2
equations as a predetermined causal factor? In a sense, the latter in-
terpretation lies between the single equation and the simultaneous equa-
tions representations. If the levels of the explanatory variables gener-
ated by the prior equations of a causal chain system are taken as given,
then only the final equation depicting output as a function of the predeter-
mined input variables remains to be estimated. This final equation cor-
responds to the single equation used in the single equation model.
Should the independent variables of this single equation be generated
exogenously relative to the production unit under study, it would be
rather irrelevant to attempt to specify the prior equations. To do so
would imply specifying a causal chain model when a single equation
model would do just as well. However, should the causal factors in the
final production function be determined within the production system,
the causal chain model is to be preferred — at least theoretically if not
in terms of data availability. Whether a causal chain or a simulta-
neous system of equations should be used depends largely on the unit
time period to which the production function estimation is oriented.
Figures 6.1 and 6.2 illustrate how this is so. Consider Figure 6.1. It
2
See Wold, H. Causal inference from observational data. Journal of the Royal Statisti-
cal Society, Series A, 119: 41-49. Also Wright, S. The interpretation of multi -variate sys-
tems. In Kempthorne, O. et at. (eds.) Statistics and mathematics in biology, Iowa State Uni-
versity Press, Ames, 1954. Pp. 11-33; Wold, H. Ends and means. In Grenander, U. (ed.)
Probability and statistics. Almgrist and Wiksell, Stockholm, 1959. Pp. 355-434.
t . . . ! 2 3 4
X,. . .
X3
Y . . .
Figure 6.1. Arrow scheme for the causal chain

model of equations 6.1 to 6.5.
depicts a hypothetical causal network between four “inputs/ X19 X 2 ,
X and X 4 and an “output” Y over a period of four time units, t(l), t(2),
3,
t(3), and t(4). These time units are basic relative to the production
process. As the arrows indicating causation show, a complete cycle of
the process requires three basic time units. It will not necessarily be
true that the three time periods within each cycle will be equal. How-
ever, corresponding periods within different cycles of the system would
be of equal length.
Figure 6.1 may be represented by equations 6.1 to 6.5. In these
equations the second subscript denotes the time period in which the
value of the variable is determined, except for Y which has only a time
subscript. Each of the equations represents a unilateral causal chain
with the variable on the left-hand side determined by the variables on
the right. Moreover, the values of all the right-hand side variables are
predetermined in a prior time period and are therefore exogenous.
After choosing the appropriate algebraic form and provided that it be
suited to least squares analysis, unbiased estimates of the parameters
may be obtained by the ordinary multiple regression procedures.
(6.1) Xx.t = f(X 2)t - 1 )
(6.2) x 2>t = f(X3 ,t-i)
(6.3) XS)t = KXx.t-!, X^t-x)
(6.4) X4 ,t = f(Xs,t-i, Yt -x)

(6.5) Yt = f(X 4>t - 1 . X S(t -x>
Such a model is perfectly satisfactory from an estimational view-

point so long as the observations that are used relate to the basic time
periods, t(l), t(2) .... However, data for production function estimation
is usually not available in such form but only for some aggregation of
the basic time units. For instance, for a firm, data is usually only
available for monthly or yearly periods while the basic time units might
be but a few days. Likewise, for biological phenomena, the basic time
periods may be only a matter of seconds whereas the only mechanically
feasible observations might relate to months in the life of the plant or
animal. Thus, suppose that for the production process depicted by Fig-
ure 6.1 it is only possible to gather observations for a time span T
where T covers a sequence of basic time units. For instance, T might
span t(l), t(2), and t(3). To depict the production process diagrammati-
cally in terms of the time unit T, it is then necessary to aggregate the
basic periods t(l), t(2), and t(3) into a single period. Within this single
period the variables then become simultaneously determined. They can
only be interpreted as being mutually interdependent, as indicated by
the arrow network of Figure 6.2. This figure depicts the periods T(l),
T(2), and T(3) and therefore corresponds to nine of the basic time units
of the type depicted inFigure 6.1. As in Figure 6.1, the arrows con-
necting different time periods in Figure 6.2 denote the influence of pre-
determined exogenous variables.
Figure 6.2 may be represented by equations 6.6 to 6.10, the last
equation being the production function of interest. As perusal of the
equations indicates, they constitute a system of simultaneous equations,
with both endogenous and exogenous variables appearing in each equa-
tion. The identification of these equations does not concern us here,
although, as indicated in Chapter 4, it must be taken into account in de-
ciding on the appropriate estimational procedures to be used.
(6.6) X l T = f(X
2)T _ 1 ,
X 2(X )
(6.7) X 2fT = f(X

3iX _ x , X 3>x )
(®‘®) X 3 ,T
= *( X 1,T-1> X 1,T> X 4,T-i> X '4
,t)
( 6 -9) X4,T = f( YT-l> Y T> X 3,T-1> X 3,t)

yt = f( X 3,T-l> X 4,T-1> X 3,T> X4 ,t)
It should bo pointed out that all of the relations in a causal

chain or
a simultaneous equations model need not be between inputs or input
and
output. They may also relate to such phenomena as entrepreneurial de-
cision processes and supply and demand relations. Such phenomena
are, of course, particularly relevant for models incorporating
the pro-
duction function for a firm. Another point that must be noted is
that
some production phenomena may only be representable by a simultane-
ous system, it being impossible to postulate any meaningful basic time
periods in terms of which a causal chain model might be hypothesized.

In large degree, whether or not such is the case is a philosophical
question. It suffices to note here that so far as empirical research is
concerned, simultaneous models are theoretically correct whenever
data cannot be collected in terms of the basic time units. It does not
matter whether the data cannot be collected because the basic time pe-
riods do not in fact exist or whether the collection of such data is me-
chanically infeasible.
Obviously, the simplest type of production process — at least in eco-
nomic terms — is one that can be fully represented by a single unilateral
causal relation. However, few real-world production processes of im-
port can be truly represented by a single equation model. Probably only
processes of a purely physical -engineering type that encompass only a
single technology and a single product can be so represented. Despite
this, it is a fact that most production function research, whether it has
involved plants, animals, firms, or enterprises within a firm, has been
based on single equation models. To a large extent, the single equation
approach has been used because of its computational simplicity, the im-
plicit hope of the researcher being that the single equation estimates
are not greatly biased. Whether or not they are depends on the me-
chanics of the process being studied. In general, however, the re-
searchers have had no idea of the extent of the bias introduced by not
using a multi -equation model. To a lesser extent, the single equation
approach has been used without any appreciation by the researcher that
a system of equations might be more appropriate — at least theoreti-
cally (although perhaps not computationally. Such a situation is evident
with regard to nearly all of the work that has been carried out with
plants and animals. 3 Thus with plants, researchers have estimated nu-
merous production functions based on a single equation model involving
such inputs as various types of fertilizers and moisture with grain or
forage as output. No explicit attention has been paid to the fact that the
plant produces many things apart from the particular part of it that is
harvested. For instance, the relation between fertilizer and corn grain
production is in reality but part of a simultaneous system of equations
representing, say, the production of grain, leaves, stalks, husks, and
roots. Likewise with animals, feed intakes enter a system of interde-
pendent production relations involving muscle, bone, excreta, and so on,
all of which react with one another in some causal or simultaneous net-
work. Indeed, insofar as these plant and animal production processes
are amenable to experimental control, it is likely that the use of the
more correct multi -equation models would be fruitful, at least in com-
parison with the mediocre success they have attained when used in
macro -economic studies in which nonexperimental data has had to be
used by necessity.
From a theoretical viewpoint, there is no doubt that the use of a
single equation approach when a system of equations model is appropri-
ate must, under usual circumstances, lead to biased estimates of the
4
production parameters. While the single equation approach allows the-
oretically satisfactory predictions of output to be made when the process
is allowed to operate in its normal fashion, it need not allow predictions
to be made of the effect of deliberate interventions with the production
process. In other words, if the single equation approach is used instead
of the multi -equation approach, the underlying structure of the produc-
tion process will most likely remain unknown. As a practical matter,
however, as pointed out in Chapter 4, single equation estimates have
generally been found to be just as logical and meaningful in an economic
sense as those derived at much more expense by the use of simultaneous
equations models. Relative to recursive systems, too few studies have
been made to justify an opinion.
Choice of Variables
Whatever type of model is used to depict the production process, the

researcher has to decide on the variables to be used. Should any rele-
vant variables be omitted, the fitted model will be biased in an eco-
nomic sense; it could not be expected to truly depict the production
process, either structurally or predictively. Likewise, the unwarranted
inclusion of variables will lead to bias.
3
For an exception, see Lucas, H. L. Experimental designs and analyses for feeding
efficiency trials with dairy cattle, in Hoglund, C. R. et al. (eds.) Nutritional and economic
aspects of feed utilization by dairy cows. Iowa State University Press, Ames, 1959. Pp.
187-91.
4
See Marschak, J. and Andrews, W. H. Random simultaneous equations and the theory
of production. Econometrica, 12: 143-205. 1944. Also, Haavelmo, T. The statistical im-
plications of a system of simultaneous equations. Econometrica, 11: 1-12. 1943; Wold, H.
Interdependent versus causal chain systems. International Review of Statistics, 26: 5-25.
1958.
Ideally, the choice of variables should be made in terms of the un-

derlying mechanics of the production process. Given such knowledge
and sufficient research resources, the researcher could estimate a
function with all the relevant variables included. Such is never the
case, however. Research resources are always limited and the under-
lying production logic can only be hypothesized. In consequence,
(a) some of the relevant variables may be unknown and may, perhaps,
only be discovered through fundamental research and (b) some of those
variables known to be relevant may be unobservable. Still, the re-
searcher must draw on all aspects of his knowledge of the production
process in attempting to enumerate as many as possible of the relevant
variables, compromising where necessary in terms of the resources
available for collecting observations and carrying out analyses. Such
compromises must be made with reference to what is known, from
prior experience, of the relative importance of the variables known to
be relevant. So far as possible, the variables omitted because of lim-
ited research resources should be those of least importance.
To some extent, also, the researcher may be able to use trial and
error methods in deciding on the variables that are relevant. A given
algebraic form of the production function may be tried with a variety of
combinations of variables. That combination which best accounts for
the observed output may be selected — provided that the influence of the
included variables is not contrary to any of the physical, biological, or
economic logic known to underlie the production process. The latter
criterion also especially applies to the problem of choosing an alge-
braic form for the production function. Indeed, the selection of varia-
bles and the choice of a functional form are closely related topics. The
following discussion of alternative algebraic forms is therefore also
pertinent to the selection of variables.
Choice of Algebraic Form
In choosing an algebraic form for the production function to be esti-

mated, the researcher must attempt to take account of whatever is
known of the logic or basic mechanics of the production process. Also,
the selected function must be computationally manageable, both for es-
timation and testing, as stressed in Chapter 4. The true algebraic form
will never be known. In general, the only knowledge of the production
process available to the researcher will be broad information on the
shape of the surface. Oftentimes, this data will be sufficient to indicate
the type of returns to scale to be expected, and the approximate location
and shape of the isoquants and isoclines. The choice of an appropriate
algebraic form when such information is available has been discussed
at length in Chapter 3.®
5
See also: Heady, E. O. Organization activities and criteria in obtaining and fitting
technical production functions. Journal of Farm Economics,) 39: 360-69; Baum, E. L. et al.
(eds.) Methodological procedures in the economic analysis of fertilizer use data. Iowa State
University Press, Ames, 1956. Pp. 39-98; and Hald, A. Statistical theory with engineering
applications. John Wiley and Sons, Inc., New York, 1952. Pp. 658-62.
Still, there will often be no strong guides as to what algebraic form

might be appropriate. In such circumstances, the ideal procedure would
6
be to use the method proposed by Hildreth. He suggests the use of a
discrete model, account being taken of the qualitative restrictions, such
as diminishing returns, that are known to exist. With such an approach,
no attempt is made to specify the production function in complete alge-
braic detail as is done in attempting to estimate a continuous model. As
yet, these form -free procedures are not sufficiently developed to be of
any general use. This is especially true for situations involving more
than one input factor. The alternative approach is to attempt to ap-
proximate the response surface by means of a polynomial.
Polynomial approximation
Frequently, production functions are estimated as polynomials with-
out appreciation of the reason why the response surface may be approx-
imated by a polynomial. The reason is that if the algebraic form of a
function is unknown, it may be approximated over the range of interest
by the standard mathematical procedure known as a Taylor series ex-
pansion. Such expansions can be reduced to a polynomial form. While
the approximation will differ in algebraic form from the true function,
7
its implications will be quite similar over the relevant range.
Thus, suppose y = f(x), the function being a smooth curve of unknown
algebraic form. Via a Taylor series expansion, it is possible to esti-
mate the value of y for x values in the neighborhood of any point x = a.
The actual expansion is shown in equation 6.11 where f'(a) denotes the
value of the first derivative of f(x) at x = a, f " (a) the second derivative,
and so on.
(6.11) y = f(x) = f(a) + f'(a) + f"(a) + • • •
The more terms evaluated in this expansion, the better the values
of y can be estimated for x values near x = a. As inspection shows,
grouping of x terms of like power reduces the Taylor series of equa-
tion (6.11) to a polynomial in x as in equation (6.12).
2 3
(6.12) y = f(x) = p0 + ft x + (32 x + j3
a X + • • •
6
Hildreth, C. G. Discrete models with qualitative restrictions. In Baum, E. L. et al .
(eds.) Economic analysis of fertilizer use data. Iowa State University Press, Ames, 1956.
Ch. 4. Also, Stemberger, A. P. North Carolina Agr. Exp. Sta. Bui. 126, 1957.
7
If the aim is to study the basic logic as well as the implications of a response process,
a differential equations approach should be used. See Box, G. E. P. The elucidation of basic
mechanisms. Bui. Inst. Int. Statist., 36: 215-25, 1958,
An analogous result occurs when y is a function of more than one

variable. y = f(x x , x 2 ), the polynomial resulting from a
Thus if
Taylor’s series expansion would be as shown in equation 6.13.
2 2
(6.13) y = 0 O + (01 Xj. + 02 x2) + (011 Xx + 02 2 x2 + ^XiXjj)
(0111 X 1 0222 X2 0 112 X 1 X 2 ^ A.22 X 1 X 2 )**"•• *
In consequence, the fitting of a polynomial type production function is

nothing other than the evaluation of the first few powers in a Taylor se-
ries expansion of the unknown true production function. Because the
Taylor series expansion only holds for a region in the neighborhood of
the selected point on the surface, such as x =a in equation 6.11, the es-
timated polynomial will only adequately fit the production surface within
a limited region of the surface. Within this region, a satisfactory fit
can generally be obtained by evaluating only the first few powers in the
Taylor series. For this reason, a first or second degree polynomial
usually fits the production surface adequately. Sometimes it will prove
worthwhile to estimate terms of the third degree. Such will tend to be
the case the greater the precision required and the larger the area of
the production surface that the researcher is trying to fit.
How well a polynomial of given degree fits the production surface
also depends on the choice of scale used for the response and for the
independent variables. Thus instead of using the simple first degree
polynomial shown in equation 6.14, it has often been found better to use
a log transformation of the variables as in equation 6.15.
(6.14) y = 0O + /3
1
x1 + 02 x2 + ... + j3
kxk
(6.15) log y = log 0 O + /3

X
log xx + 0 2 log x2 + ... + (3 k log x k
Equation 6.15 is simply the ordinary Cobb-Douglas function. Probably

few who have used it have recognized that it may be regarded as a very
simple Taylor series estimate. An interesting point is that recogni-
tion of the Cobb-Douglas function in these terms leads to consideration
of higher degree polynomials in the logarithms. Thus, for the two input
variable case, a second degree polynomial of the form shown in equa-
tion 6.16 might be estimated by conventional least-squares procedures.
2
(6.16) log y = log 0 O + 0 X log x x + 0 2 log x 2 + 0 U (l°g x i)
2
+ 0 22 (log x 2) + 0 12 log xt log x 2
In more conventional fashion, equation 6.16 might be written as in equa-

tion 6.17,
+ A
011 l °s xl 0 22 log X, log Xl

(6.17) y = £o xi x 02„
which reduces to
(6.18) y = 0O *^
+ '
3 “ 1° 8 x
*
xf 2
+
^ log X
*
+ ^ l0g X
^
There are innumerable transformations that might be applied to the

variables. One type that has been used extensively is the square root
transformation of the independent variables. It has proved very suc-
cessful in fertilizer studies. Polynomials of first and second degree
using the square root transformation for two variables are shown in
equations 6.19 and 6.20.
(6.19) y = 0 0 + 0i Vx7 + 02
2 2
(6.20) y = 0o +0i VxT + 02 V^2 + 0ii ( Yxi) + 022 ( Yx a)
+ 0i2 Yx7 VxJ
The latter equation reduces to equation 6.21 in which form it is not so

easily recognized as a Taylor series expansion.
(6.21) y = 0 O + 0
ux4 + 0 22 x2 + 02 + 02 Vxl + 812 Vxl
Such square root transformations are members of a far larger family
of functions obtained by using the general transformation shown in
equation 6.22.
(6.22) x* = x
m/n
;
For the square root transformation m is 1 and n is 2. For m = 3 and

n = 5, the two factor second degree polynomial would be as shown in
equation 6.23; with m = -1 and n = 1, equation 6.24 results.
5
(6.23) y = 0 O + 0X x*/ + 0 )X y* x®/ 5 + 0.22 a 28 s/s
01 @2 ft.1 022
(6.24) y = t t +
Xl X2 Xy X/ X,X,
The square
root transformation having proved useful, it would seem
that many of the other functions derivable by
use of the power transfor-
mation of equation 6.22 would also have utility. However, they have not
been tried extensively. One difficulty in their use is that they lead to
extremely complicated equations for the derivation of isoquants and
isoclines.
Another type of transformation that may sometimes prove useful is
to use the logarithm of the new variable obtained by adding a
constant to
8
the independent variables. The resultant function for two variables in
a first degree polynomial is given in equation 6.25.
(6.25) y = log p 0 + j3 1 log (a x + x x ) + /32 log (a 2 + x 2 )
With such a function, the isoclines are straight lines not passing
through the origin. In consequence, this function allows the marginal
rate of substitution between inputs to change as the level of output
changes. However a difficulty exists in that the constants, such as ax
and a 2 of equation 6.25, which allow the best fit can only be ascertained
iteratively.
From a mathematical point of view, there, is no reason why more
than one transformation should not be used within the same polynomial
expansion. By using a variety of transformations it is very easy to
build odd looking functions. Thus if a logarithmic and a power trans-
formation are used on x 1 and Xj> respectively, the second degree poly-
nomial appears as shown in equation 6.26.
log x, + 0 2 x 2
m/n 2
(6.26) y = /3 0 + + (3
U (log xl )
+ i
3 22 X 22m//n + /3i 2 x 2m//n log x x
In part, the choice as to which type of transformation to use in a

polynomial approximation to the surface may be decided with reference
to the known logic of the production process. For example, it is known
with respect to animals that as they grow a greater proportion of the
feed intake is used for maintenance and less for growth. The produc-
tion function must therefore allow decreasing productivity to each unit
of feed input, either on an individual feed type basis or over -all. It is
also known that as the proportion of protein concentrate in the ration is
increased, it replaces fewer units of low -protein carbohydrate. Thus a
second requirement of the fitted function is that it must allow diminish-
ing marginal rates of substitution between, say, corn and soybean oil-
meal. Another biological fact is that, for rapid growth, young animals
generally require a greater percentage of protein in the ration than
older animals. Therefore, a third requirement of the production func-
tion is that it must allow substitution rates between feed types to change
as the animals gain in weight. Equivalently, the isoclines must either
be curved, or be linear and not pass through the origin. Functions 6.27
and 6.28 below fulfill all of the requirements outlined above, the sym-
bols C, S, and Y referring to, say, pounds of corn, soybean oilmeal, and
animal liveweight gain respectively.
2 2
(6.27) Y = p 0 + 0, C + S +
U C
/3 + 0 22 S + /3
12
CS
(6.28) Y = /3
0 + j3
uC + 0M S + VC~+ & Vs"+ j8 u ^CS
8
See Heady, E. O. Organization activities and criteria in obtaining and fitting technical
production functions. Journal of Farm Economics. 39:366-67. 1957.
The Cobb-Douglas function, equation 6.29, permits decreasing produc-

tivity to the feeds and diminishing rates of substitution between feed
types, but does not allow substitution rates to change for a particular
type of ration as the animals increase in weight. However, the logarith-
mic function, equation 6.30, does satisfy all three requirements of the
biological logic.
(6.29) log Y = j3 0 + /3
X
log C + 0 2 log S
(6.30) log Y = 0 O + 0i log (a, + C) + 0 2 log (a* + S)
Still, the logarithmic function, equation 6.29, may be useful. For ex-
ample, an animal feeder may wish to feed only one ration which aver-
ages lowest in cost over the entire production period. Function 6.29 is
appropriate for this purpose because the substitution rates between
feeds along any ration line are then constant. In other words, the iso-
clines are then regarded as linear and passing through the origin of the
input axes. Equating the feed price ratio with the slope of the gain iso-
quants gives an average least -cost ration for the whole growing period.
Should the producer be interested in frequent changes of rations during
the production period, however, equations 6.27, 6.28, and 6.30 allow
better estimates of the least-cost rations than does the logarithmic
function 6.29.
The use of polynomial functions in production function research is
generally beset by two difficulties. The first stems from the fact that
while the consideration of higher terms in the Taylor series generally
leads to a better fit to the surface, it also leads to greater difficulties
in estimating isoquants and isoclines. This difficulty is generally com-
pounded by the use of a transformation of the variables, and by any in-
crease in the number of input factors considered. Thus the derivation
of isoquants and isoclines for polynomials involving other than square
root and squared factors in addition to first degree terms is very time
consuming. As mentioned previously this difficulty arises when by use
of the power transformation in equation 6.22, the resultant function has
terms raised to peculiar powers, as for instance in equations 6.23 and
6.26. Moreover, it must be stressed that to draw out the economic im-
plications of the production surface, the researcher has to do more than
simply fit a function to the surface — as can easily be done by way of a
Taylor’s expansion over the relevant ranges of the inputs. For the eco-
nomic implications of the production surface to be recognized and acted
upon, the isoquants and isoclines have to be calculated. It is only
through their evaluation that account is taken of the price and cost con-
ditions ruling in the real world. The manipulations involved in deriving
isoquants and isoclines and their subsequent interpretation have been
outlined in Chapter 3.
The second difficulty with the use of polynomial functions is that as
the degree of the polynomial increases, and as the number of factors
considered rises, the number of terms involved in the expansion rapidly
Table 6.1. Number of Coefficients To Be Estimated

for Polynomials of Various Degree
Number of Input Factors

Degree of Polynomial
1 2 3 4 5 6
1 2 3 4 5 6 7
2 3 6 10 15 21 28
3 4 10 20 35 56 84
4 5 15 35 70 126 210
increases. In consequence, the number of degrees of freedom available

for statistical testing is decreased. The reason is that for each term in
the polynomial there is a coefficient that has to be estimated at the cost
of a degree of freedom. The number of coefficients, say c, in a poly-
nomial of degree d for k factors is given by the combinatorial formula
(k + d) 1
(6.31)
d!k!
The number of coefficients to be estimated for polynomials of various

degrees involving various numbers of factors are shown in Table 6.1.
Perusal of this table indicates how rapidly the number of degrees of
freedom available for testing, out of a given total, may be diminished as
the polynomial is extended.
ADEQUACY OF THE FITTED FUNCTION

Having fitted a particular function, two questions arise. First, in an
over -all sense, does the function adequately characterize the data?
Secondly, are some of the individual terms in the function of such neg-
ligible import that they should be excluded? These questions are, of
course, related. Moreover, it must be stressed that, granted the rele-
vance of the input factors considered, there is no sure answer to these
questions. Researchers may quite rationally differ in the weight they
attach to the various criteria used in assessing the suitability of a given
or alternative functions.
Consider first the question of the adequacy of the over -all function.
The major question to be asked is: Does the function adequately char-
acterize the data both in terms of the known logic about the' production
process and in statistical terms (which may or may not be the same
thing) ? The former question will generally have been raised and an-
swered at the initial stage of choosing the function. For testing the
statistical adequacy, although not necessarily the basic logical ade-
quacy, of the fitted function, three procedures are generally available
when experimental data has been fitted by regression procedures. They
are based on the size of the coefficient of multiple determination, an F

test of the regression mean square and a comparison of the lack of fit
and error mean squares. For nonexperimental data, only the first two
approaches are feasible, it not being possible to calculate a lack of fit
term from nonexperimental data. Replications are needed to segregate
the deviations from regression into lack of fit and error components.
2
It is desirable that the coefficient of multiple determination, R ,
be
2
as close to unity as possible. Generally, quite low values of R will be
statistically significant. However, such statistical significance merely
denotes that there is some relation between the independent and depend-
ent variables. In production f -unction research we desire more than
this; we want to explain the major part of the response in terms of the
considered inputs. If we cannot do so, then the fitted function will be of
little operational use for real-world decision making. If the number of
parameters estimated is large relative to the number of observations
used, then the appropriate term for looking at the percentage of the
2
variation in output explained by the regression model is R ,
the ad-
2
justed coefficient of multiple determination. The derivations of R and
2
It are given in equations 4.40 and 4.41 of Chapter 4.
An over -all test of the significance of the fitted regression model
may be carried out by calculating the F ratio, regression mean square
divided by error mean square, as shown in equation 4.42 of Chapter 4.
This ratio provides a test of the null hypothesis that all the regression
coefficients are equal to zero. If the F value is larger than the tabled
value of F at the desired probability level, the null hypothesis is prob-
ably not true.
With experimental data it is also possible to calculate a lack of fit
mean square. This term can be compared with the error mean square
to test whether the variation of the observations about the fitted model
are like those that might be expected given the errors natural to the ex-
perimental environment. The latter errors are estimated by the error
mean square and are independent of the fitted model. They are ascer-
tained by repeated observation, i.e., by replication. A statistically sig-
nificant lack of fit F ratio testifies that the variation around the fitted
function is larger than that which might be expected from experimental
error alone. In such cases, the model cannot be said to adequately fit
the data. A stronger criterion sometimes used is that the fitted func-
tion should only be regarded as providing an adequate fit if the lack of
fit mean square is of the same order of magnitude (or less) than the
error mean square. The calculation of the lack of fit mean square for
alternative experimental designs is discussed in Chapter 4.
When these various tests indicate that the fitted model is statisti-
cally satisfactory, and the model does not infringe the basic logic of the
production process, no problem arises. However, having fitted a num-
ber of functions, it may be found that some satisfy some of the criteria
better than others and vice versa. The function chosen as best will then
depend on the weight the researcher attaches to the various criteria,
statistical and logical. At such a stage, the selection of a function is
more of an art than a science.
As for the over -all function, a number of criteria provide guides as

to whether or not individual terms should be dropped from the fitted
function. One criteria used is the testing of the statistical significance
of the individual regression coefficients. The method is shown in equa-
tions 4.34, 4.35, and 4.36 of Chapter 4. However, the dropping of indi-
vidual nonsignificant terms is not to be recommended; by its nature a
significance test merely takes account of the strength of evidence
against the worst possible result. Even if the evidence against the re-
gression coefficient being zero is slight, the best estimate of its size is
still that obtained from the data. Indeed it is most unlikely that the
9
true value of the coefficient is exactly zero. Hence Anderson has sug-
gested a more lenient criterion: a variable should be dropped only if
the standard error of the regression coefficient exceeds the absolute
size of the estimated coefficient, and then only if there are no strong
logical grounds for including the variable.
When a polynomial function has been fitted, it is possible to calcu-
late F ratios whereby the contribution of each group of terms — linear,
quadratic, cubic, etc. — to the regression sum of squares can be tested
for significance. If orthogonal polynomial regression has been used,
this testing can be done very conveniently as shown in relation to Table
4.2 of Chapter 4. If the data is not amenable to orthogonal polynomial
regression, more tedious procedures have to be used. 10 Suppose the
fitted function is of second degree in three variables, as in equation
6.32, and that an estimate of the error mean square is available from
X 2 + b 33 X
2 2
(6.32) Y = b 0 + b, X, + b 2 X2 + b3 X3 + b u X, + b 22 2 3
+ b I2 X Xx 2 + b 13 X, X 3 + b 23 X2 X3
replication or prior experience. Working in terms of the raw data, as
in equation 4.9 of Chapter 4, the sum of squares associated with the
linear regression on Xx X2 , ,
and X 3 may be calculated as follows. First
fitequation 6.33.
(6.33) Y = b 0* + b* X, + b* X2 + b* X 3
With the b*’s of equation 6.33, calculate the linear sum of squares as
shown in equation 6.34
2
(SYj)
(6.34) Linear S.S. = S b* 2 X^ J
Y,-
J
-
R
i=o j
where X 0 is one and Z X^-Yj is the sum of crossproducts over the n

corresponding observations on X^nd Y. The linear mean square,
9
Anderson, R. L. Some
statistical problems in the analysis of fertilizer response data.
In Baum, E. L. Economic and technical analysis of fertilizer innovations and
et al. (eds.)
resource use. Iowa State University Press, Ames, 1957. Pp. 187-206.
10
See Anderson, R. L. and Bancroft, T. A. Statistical Theory in Research. McGraw-
Hill, Inc., New York. 1952. P. 172.
obtained by dividing the linear sum of squares by its degrees of free-

dom, may be tested against the error mean square as an F ratio. If the
F value is significant, then the linear regression portion of equation
6.32 has made a worthwhile contribution.
The additional variation in Y accounted for by the quadratic terms
2 2 2
of equation 6.32, i.e., by X. , X 2 , 3^ , X^, X^,
and X2 X3 , may be
found as shown in equation 6.35.
3 3
(6.35) Quadratic S.S. 2 bi S XijYj s bih ^XyXyYj

i=o j
=1 i^h^l
n 2
‘ ( Yj) Linear S.S.
?
The quadratic contribution may then be tested by an F ratio in the same

way as suggested for the linear mean square. The calculations involved
in such testing of groups of variables of like degree become quite sim-
ple for experimental data from factorial, composite, or rotatable de-
signs because of their semiorthogonal nature.
The majority of the criteria discussed above for assessing the ade-
quacy of a particular function, and the contribution of individual terms
in the function, are statistical. Nonetheless, the selection of a particu-
lar fitted function is the final act of economic specification. Once a
particular fitted function is selected, the economic implications of the
relevant surface region are set. They follow automatically from the
fitted function, for any given set of price conditions. This fact is the
justification for the inclusion of these statistical tests under the topic
of economic specification. Moreover, it is worth emphasizing again
that most weight should be attached to whatever information is availa-
ble about the logic underlying the production process. Given such in-
formation, and being sure of its reliability, only secondary importance
need be attached to the statistical tests. Still, more often than not, in-
formation about the basic logic will be lacking. The role of the statisti-
cal tests must then be dominant.
SPECIFICATION ERRORS
It is never possible to specify and fit the true production function

that is relevant to a given production process. Generally the true func-
tionalform and the complete range of variables that it should contain
are unknown. Some variables may be known to be relevant but be im-
possible to include because data about them is not available. As well,
approximations and abridgements may have to be made due to limited
research resources. All these exclusions, approximations, and other
empirical compromises result in specification errors. The fitted func-
tion cannot help but be biased to some extent. Indeed, it might be said
that a large part of chapters 4 and 5 and the preceding portion of the
current chapter are largely concerned with various procedures oriented
to the minimization of these specification errors* We now elaborate a
method whereby the bias that does exist may be formally recognized
11
and, to some extent, its consequences evaluated. There is a proviso,
however. The framework of analysis assumes (a) there is no error in
specifying or measuring the output, (b) that the input variables are non-
stochastic, and (c) that the production function is linear in the parame-
ters. Insofar as least-squares procedures are most frequently used in
estimating production functions, the framework will be outlined in rela-
tion to the ordinary regression procedures.
Suppose the true production function is
(6.36) Y = I3 0 X0 + 13, X, + ...+ 0k Xk .
The general case will be that the X’s are not specified correctly. The
actual function estimated will be
(6.37) Y = p0 X0 + &X, +...+ 0 h X k
where Xj may or may not be exactly the same as X i and h may be

greater or less than k. What we would like to do is estimate equation
6.36. Instead, we are forced by the exigencies of the real world to esti-
mate equation 6.37. The question that arises is what is the relationship
between the 0 estimates of equation 6.37 and the true production coeffi-
cients which are the 0’s of equation 6.36. Following the matrix notation
outlined in equations 4.20 to 4.24 of Chapter 4, the 0 estimates are
given by
(6.38) [0] = ([X]' [XP- [X]' [Y] .
the true regression model is
(6.39) [Y] = [X] [0] + [€] .
Substituting equation 6.39 in equation 6.38 and taking the expectation of

[0] we have, since E[e] is zero,
(6.40) E[j9] = [P] [0]
where
1
(6.41) [P] = ([X]' [X])- [X]' [X] .
In doing so, we mainly follow the development of an idea of Theil’s as ably presented
11
by Griliches. See Griliches, Z. Specification bias in estimates of production functions.

Journal of Farm Economics, 39; 8-20, 1957; Theil, H. Linear Aggregation of economic re-
lations. North Holland Publishing Company, Amsterdam, 1954.
In equation 6.40, the expected value of the estimates actually derived

are expressed in terms of the true parameters, [0], and the available
data, [X], Each 0£ is influenced by all the 0’s. Now equation 6.41 is
analogous to equation 6.38 or 4.24, Hence the elements of [P] can be
regarded as regression coefficients of the regression of each all X^n
of the included X’s. For example, the i-th column of [P] is composed of
the regression coefficients p oi to p hi that would be obtained if equation
(6.42)could be calculated.
6.42
X. = p 0i X0 + p .X + ...+ P
i i M X h+ e. ;
i = 0, 1 k
The set of regressions denoted by equation 6.42 has been termed “aux-
iliary* regressions. Together, equations 6.40 and 6.42 indicate that the
effect of the true parameters on the actually estimated coefficient
for a particular X
£ will depend
on the sign and size of the p’s attached
to thatXi in the various auxiliary regressions. If, in some fashion, the
signs and magnitudes of the p’s were known, it would be possible_to
ascertain the size and direction of bias in the actually estimated 0
values.
Omission of Variables
Supposeall of the X’s, the true inputs, are measurable except one,
say Xk So long as this variable is uncorrelated with any of the other
.
X’s, its omission will not bias the estimates of 0 O , 0 1? . ., 0 k-1 Still, . ,
the assumption of zero correlation between the excluded factor and any
of the other inputs is not likely to hold in the real world. In general, a
certain degree of positive correlation is to be expected. The result will
be a tendency to overestimate one or more of the coefficients of the in-
cluded variables. Such an effect can be deduced from equation 6.43
which gives the expected value of 0£ when X k is excluded.
(6.43) E (0.) = 0.+ p ik 0 k ;

i = 0, 1, ..., k-1
The p’s of equation 6.43 are from equation 6.44 which

is the only non-
trivial auxiliary regression. All the other auxiliary regressions are
trivial since, in the case under consideration, X £ = X£.
(6.44) Xk - p ok X 0 + p lk X 1 + ...+ p k - l k + e
If X k should be negatively associated with some of the other inputs, un-

likely though be, the p ik term of equation 6.43 will be negative and
it
E (0i) will be less than 0£ there would be a bias towards underestima-

;
tion.
To generalize, if the last K input factors are omitted from the esti-
mation, the expected value of 0£ is given by equation 6.45.
(6.45) E (jSi) = 13. + S Pi J

.
i3. ; i = 0, 1, ..., k-K
J
j=k-K+i
Again, it is generally to be expected that fii will be biased upwards,

given the fact that positive correlations among inputs are most common.
The economic implications of the bias in the p values will be traced
out through the economically oriented manipulations and interpretations
of the production function. For instance, such biases will be carried
over into the isocline and isoquant derivations. Moreover, it should be
noted that the above discussion of the omission of relevant variables
applies equally to the derivation of production functions from experi-
mental and nonexperimental data.
Aggregation Within Inputs
Within a single input category, no matter how finely we define the

category, there will usually be quality differences. Only under experi-
mental conditions where the inputs can be made of uniform quality
through the application of chemical and other standards are these
quality differences likely to be negligible. But with factors such as
land and labor, quite large quality differentials will be the rule. Gen-
erally, little account is taken of these differences — one acre of land is
regarded as being much the same as any other acre. If adjustments
are made, they can only be approximate in the absence of precise
knowledge of the quality differentials. How does the failure to take ac-
count of the true quality differences affect the specification of the pro-
duction function?
Suppose quality differences occur only in X k of equation 6.36. The
quantity of X k used in estimation is taken as Xk But, in truth, we may .
say X k is made up of m
portions each differing in quality, the j-th such
portion being denoted by Xk . The use of X k without taking account of
j
these quality differences assumes that equation 6.46 holds.
(6.46) Xk =
x
X k+ X k+ 2
... +
mfk
Denote by g the weight that should be attached to X k to transform it
j j
into equivalent units. Hence X k may be expressed in terms of X k as •
(6.47) Xk — gi x
Xk + g2 2 ^k + + Sm m^k
Thus, for the case under consideration, equation 6.36 may be rewritten
as in equation 6.48.
m
(6.48) Y = /3
0
X0 + j3 lXl + ... + P HX , + k gj jX k
The ignoring one factor is thus equiva-

of quality differences within the
lent to the omission a number of variables together with the inclu-
of
sion of an additional variable, X k X k is an imperfect aggregate, being
.
based on the assumption that gj Is always unity. _ The exclusion of the

variables gj Xk will generally tend to bias the jS values upward, as
j
outlined in relation to equations 6.43 and 6.45. However, the net result
will depend on whether the Influence of the inclusion of the imperfect
aggregate X k is to complement or counteract the bias resulting from
the omissions.
Aggregation Over Inputs
The more general case of the situation examined above is that in

which several distinctive inputs are aggregated into a single input cate-
gory. Suppose all inputs are aggregated into two categories X x and X 2 ,
and the estimated function is as in equation 6.49.
(6.49) Y = £ 0+ ^X 1+ j3
2
X2
while equation 6.50 denotes the true production function.
(6.50) Y = P0 + £ Xu + ? /3
2j
X 2j ,
i + j
= k
Hence the aggregation is such that
(6.51) Xx = 2 Xa and X2 = S X 2i
.
i j
By virtue of equation 6.40, the expected values of the /3’s of equation

6.49 are given, in scalar notation, by equations 6.52 and 6.53.
(6.52) E (fi ± ) = ? Pil 0* + S Pjl i3 2j
(6.53) E (j§ 3 ) = ? p i2 + 5 p j2 £ 2j
Each estimated j§ thus consists of a weighted average of the true pa-
rameters associated with each of the basic unaggregated inputs that are
included in the corresponding aggregate plus a weighted sum of the pa-
rameters not included in the aggregate relevant to the given /3 The .
weights, as before, are the auxiliary regression coefficients. If the

auxiliary regression coefficients and the true production function pa-
rameters, i.e., the 0’s, are uncorrelated, there will be no bias due to
aggregation. Such a situation is not to be expected.
The problems associated with aggregation over inputs are most
likely to arise in the estimation of production functions for firms. In
such cases the number of input categories is large and quality differ-
ences in inputs between and within firms are to be expected. To mini-
mize specification bias due to aggregation in such cases, two working
rules should be used. First, perfect complements, i.e., resource cate-

gories that have to be used in fixed proportions, should be treated as a
single input. The use of one such resource implies the use of its com-
plements. To include each of the complementary categories would lead
to multicollinearity (see Chapter 4) because of the perfect correlation
between levels of the complementary inputs. Secondly, perfect substi-
tutes should also be aggregated into a single input category. Ideally
this can only, be done by the use of standard units as in equation 6.47.
The difficulty is that the standardizing weights will generally be un-
known, Should equation 6.46 hold true, the aggregation problem is
trivial — the input categories are perfect substitutes simply because
they are but portions of identically the same input!
EQUILIBRIUM PROBLEMS IN SURVEY DATA
It has been argued that if producers maximize profits, identification

of the production function from single period survey data will be im-
12
possible.
Thus if entrepreneurs maximize profits, and all use the same pro-
duction function and there is no resource fixity, all sample firms would
have the same input -output array and the production function could not
be estimated. No real-world data has ever evidenced such concurrence.
However, suppose entrepreneurs are only prevented from maximizing
profits by resource fixity and there is a common proportionate element
of resource fixity across the sample. Then all observations will lie on
the one isocline (assuming a common production function) and the data
will (usually) be multicollinear. However, the assumptions of certainty,
omniscience, proportionate resource fixity, and a common production
13
function are too strong to apply to real-world data. As Konijn has
shown, if they do not hold true, it is possible to derive a meaningful
production function estimate.
12
See Soper, C.S. Production functions and cross-section surveys. Economic Record,
34: 111-17, 1958. Also, Marschak, J. and Andrews, W. H. Random simultaneous equations
and the theory of production. Econometrica, 12: 143-205, 1944.
13
Konijn, H. S. Estimation of an average production function from surveys. Economic
Record, 35; 118-25, 1959. Also, Soper, C. S. Production functions: A reply to Dr. Konijn.
Economic Record, 35: 434-35, 1959.
CHAPTER 7
Miscellaneous Empirical Problems
Relating to Estimation
of Production Functions
A ctual estimation of production functions and eliciting of their

real-world implications is an empirical task. A variety of un-
avoidable empirical problems thus arise. These do not fit neatly
under the theoretical labels discussed in Chapters 3, 4, 5, and 6. We
consider the more important of these miscellaneous empirical prob-
lems in the present chapter. Emphasis is on the compromises and ap-
proximations that may be expected to serve best in data collection and
analysis. The research worker is faced with the necessity of striking
a balance between the theoretically desirable and the empirically fea-
sible, in terms of data availability and computational resources. Fre-
quently, it is possible that a worthwhile compromise cannot be attained.
The majority of the problems discussed relate to the estimation of
farm-firm production functions. Farm-firm problems predominate in
the discussion because their estimation is more difficult than the deri-
vation of production functions for individual technologies. 1 The latter
can be studied under experimental conditions while the farm- firm pro-
duction process is complicated by its relation to entrepreneurial de-
cision making.
CLASSIFICATION AND MEASUREMENT OF

INPUTS AND OUTPUTS
The
applicability of an empirically derived function depends on the
way which the input and output factors are defined and measured, and
in
on the use to which the fitted function is to be put. If a high degree of
aggregation is used, the implications of the resultant function may be of
1
For general discussion of the estimation of farm-firm functions, see Heady, Earl O.
Elementary models in farm production economics research. Jour. Farm Econ., 30: 1-18;
Plaxico, J. S. Problems of factor-product aggregation in Cobb-Douglas productivity analy-
sis. Jour. Farm Econ., 37: 664-75; Parish, R. M. and Dillon, J. L. Recent applications of
the production function in farm management research. Review of Marketing and Agricultural
Economics, 23: 215-36; Jarrett, F. G. Resource productivities and production functions.
Australian Jour. Agr. Econ., 1: 67-78; and Konijn, H. S. Estimation of an average produc-
tion function from surveys. Econ. Record, 35; 118-25; Wolfson, R. J. An econometric
investigation of regional differentials in American agricultural wages. Econometrica,
26: 225-56.
PROBLEMS RELATING TO ESTIMATION 219
littlerelevance in decision making. For the farm operator, knowledge

that the marginal return to the broad category * capital* exceeds its
marginal cost is insufficient. Returns may not exceed costs for some
capital items within the aggregate; for others, the opposite will be true.
On the other hand, the information derived from a production function
based on aggregative input and output categories may be quite useful to
a government policy maker.
Ideally, input and output variables should be measured in physical
units of a homogeneous nature; quality differences in a given factor
being absolved by a standardization procedure based on the relative
contribution to production of each quality level of the particular factor
Largely, this requirement is attained in experimental studies of partic-
ular technologies involving individual fertilizer and feed elements. The
situation differs for farm-firm studies based on either time series or
cross-sectional from the world of decision making. Then it is impos-
sible to measure all items in physical terms, especially capital goods
and services. Heterogeneous capital forms have no common physical
unit. They must be aggregated to some extent and measured in value
terms for computational purposes. While a tractor gives inputs of
work hours at a given intensity, this physical input cannot be added to
that from, say, a welding kit. Faced with a myriad of different types of
these inputs, the research worker must aggregate to some extent. The
practical basis of aggregation ordinarily is in value terms. Similarly,
there are various grades of output which can be aggregated feasibly
only in value terms. Consequently, the theoretical distinction between
a physical production function and a value function is generally blurred
in farm-firm analysis. Also, the generality of the fitted function is re-
duced since it strictly applies only to the particular price regime on
2
which the value estimates are based .
The major broad resource categories are land, labor, capital, and
management. The peculiarities associated with classification and meas-
urement for farm -firms of these categories will now be outlined.
Capital
A statement that capital inputs have, based on farm sample esti-

mates, a, marginal return of so many dollars indicates nothing about the
productivity of particular forms of capital inputs. The statement is
meaningful only under the assumption that the composition of capital
increments corresponds to the average capital composition employed
in the sample firms. This problem can be minimized, however, by
classifying capital inputs into a number of value categories. The par-
ticular categories chosen will depend on the investigator’s assessment
2
For a procedure oriented to overcoming these difficulties see Trant, G. I. Adjusting
for price levels in production function studies, in Heady, E. O. et al. (eds.) Resource pro-
ductivity, returns to scale, and farm size. Iowa State University Press, Ames. 1956. Pp.
162-67.
220 PROBLEMS RELATING TO ESTIMATION
of the strategic inputs in the production process, thebookkeeping habits
offarmers, and the purpose of the investigation. Two general rules
should be followed in deciding on the various individual items to be
placed in each input category. They are:
1. The inputs within an individual category should be as

nearly perfect substitutes or perfect complements as
possible,
2. Relative to each other, the categories of inputs should
be neither perfect substitutes nor perfect complements.
Quite apart from the fact that these two rules lead to theoretical
correctness, they are quite functional in that they tend to specify the
production problem in a fashion meaningful to farmers. The more im-
portant real-world problems of resource allocation fall under the sec-
ond rule.
To demonstrate the variety of ways in which the capital factor has
been approached, the classification of factors used in a small number
of published farm-function analyses are listed below in historical se-
3
quence.
4
Kamiya (1941) : Land; labor.
5
Tintner (1944) Land; labor; farm improvements (buildings, fences, etc.); liquid
:
assets (livestock, feed, seed, fertilizers); working assets (machinery, breeding

stock, equipment); cash operating expenses (repairs, fuel, oil, purchased feed).
6
Heady (1946) Real estate (land and improvements); labor; machinery and equip-
;
ment (inventory value plus value of repairs; fuel and lubricants); livestock and
feed (stock on hand and purchased, livestock expense and feed fed); miscellaneous
operating expenses.
7
Johnson (1952) Land; labor; machinery investment (inventory value); livestock
:
(inventory value) and forage production investment (replacement value of hay and
pasture stands plus investment in structures or land clearing necessary to estab-
lish such crops); cash operating expenses.
Heady and Shaw (1954) 8 Crop function: Cropland; crop labor; capital services
:
used on crops plus annual cash expenses (including depreciation on all items used
directly or indirectly in crop production). Livestock function: Labor; capital
services used on livestock (including depreciation on buildings, etc., and depre-
ciation on breeding stock, and purchase value of feeding stock).
3
In all cases, land is measured in acres, real estate in dollars or pounds, labor in man-
months or man-years, all capital inputs in dollars or pounds.
4
Kamiya, K, On productivity of labor. Jour. Rur. Econ., 17: 89-98.
5
Tintner, G. A note on the derivation of production functions from farm records.
Econometrica, 12: 26-34.
8
Heady, E. O. Production functions from a random sample of farms. Jour. Farm Econ.,
28: 989-1004.
7
Bradford, L. A. and Johnson, G. L. Farm management analysis. John Wiley and Sons,
Inc.,New York. 1953. Pp. 145-47.
8
Heady, E. O. and Shaw, R. Resource returns and productivity coefficients in selected
farming areas. Jour. Farm Econ., 36: 243-57.
9
Heady (1954) : Land fixed. Crop function:Labor, machinery expenses (depre-
ciation, repairs, fuel, oil, etc.); annual crop expenses (seed, fertilizer, lime
seed treatment, etc.).
10
AntiU (1955) Land; labor; purchased feeds; other capital inputs (including in-
:
terest on crops, livestock, machinery and equipment inventories, depreciation on

machinery and equipment, rent or rental value of real estate, cost of salaried
management).
Schapper and Mauldon (19 57) 11 Labor (excluding part-time and contract); pur-
:
chased feed; fertilizer purchased; miscellaneous operating expenses (part-time

labor, fuel, oil, machinery repairs and maintenance, freight).
12
Wang (1958) : Cropland; labor; capital services (seed, fertilizers, rent for
working cattle and equipment, miscellaneous crop outlays).
This listing illustrates the diversity of approaches that have been

used in classifying capital inputs. Neglecting Kamiya’s study (1941)
which did not consider capital, the historical tendency has been (a) to
reduce the number of capital input categories and (b) to measure the
input of durable assets by the actual maintenance and depreciation costs
associated with their use rather than by their capital value on an inven-
tory basis. For the reasons of interpretation already stressed, the
grouping of capital items into a single category must be regarded as
unwise unless they are good substitutes or complements. On the other
hand, some level of aggregation is required for computational feasibility.
However, caution must be exercised in inferences since the diverse
micro- inputs may have varying productivity magnitudes.
As usually oriented, the production function study for firms is con-
cerned with the short run of a single production period or year. Meas-
urement of the input of durable assets by maintenance and depreciation
costs is certainly to be recommended over that of capital investment or
inventory value. Durable items such as machinery constitute a stock
resource which gives rise to a flow of input services over the life of the
asset. Only the flow of services within the selected production period
is relevant to the production function for that period. The actual input
flow relevant to the period under consideration should be used. Meas-
urement of this flow as a depreciation rate based on an arbitrary book-
keeping basis often is not satisfactory. For instance, on one farm a
substantial amount of depreciation might be allowed as capital service
input on a relatively new tractor used but little during the production
period. On another farm in the same sample, a small amount of depre-
ciation might be taken as the capital service input on an old tractor used
9
Heady, E. O. Resource productivity and returns on 160-acre farms in north central
Iowa. Iowa Agr. Exp. Sta. Bui. 412. Ames. 1954.
10
Antill, A. G. Towards a production function for dairy farms. The Farm Economist,
8 : 1 - 11 .
11
Schapper, H. P. and Mauldon, R. G. A production function from farms in the whole-
milk region of Western Australia. Econ. Record, 33: 52-59.
12
Wang, Y. Resource returns and productivity coefficients for selected crop systems in
Taiwan area. Proceedings of Agr. Econ. Seminars. National Taiwan University, Taipei.
Sept. 16-20, 1958. Pp. 90-98.
extensively. In each case, a substantial observation error may occur.
In the absence of detailed records on the actual use of capital items,
however, little can be done to overcome these measurement deficiencies.
Still, they should be recognized and necessary caution exercised in in-
terpretation. The effect of such measurement errors can be ascertained
via the method outlined in Chapter 6, relative to errors of specification,
provided that the necessary auxiliary information is available.
The classifications used by Antill, and Schapper and Mauldon, as
listed above, illustrate separate treatment of particular categories of
capital input (purchased feed and fertilizer) which are considered to be
of key importance in the particular production process being examined.
Heady’s study (1946) illustrates a method of treatment for complemen-
tary factors: a preliminary analysis showed high correlations between
machinery and equipment inputs and fuel and lubricants. These items
were therefore grouped together to form a single input category. Live-
stock input and feed expenses were lumped together for the same rea-
son.
Care needs to be taken in delimiting the cash operating expenses
category. Items of input which are determined directly by the volume
of production should not be included. Selling charges, where they con-
stitute a fixed commission on the amount of the sale, are an example.
Included as a separate input they would have an elasticity of production
of unity and a very high marginal productivity. But the results would
be meaningless; selling expenses do not determine farm production but
are determined by production. Other examples are packing and han-
dling costs. Perhaps such expenses substantially explain the high mar-
ginal productivity of cash operating expenses derived in some early
farm-firm studies. This possibility is indicated by the results of an
Australian study. In a wool growing area, annual capital expenditure
including shearing and wool selling expenses was found to have a pro-
duction elasticity of 0.55 and a marginal productivity of $1.73 per $1
spent. When shearing and wool selling expenses were excluded, the
production elasticity of capital fell to 0.43 and its marginal productiv-
13
ity to $1.59. Of course, the extent to which particular cost items are
determined by output will vary by industry and region, and judgment is
required to determine whether the decision makers vary particular in-
14
puts independently of output.
Labor
In measuring the labor input, two factors should be borne in mind.

First, what is required for estimational purposes is a measure of the
13
Parish, R. M. and Dillon, J. L. Recent applications of the production function in farm
management research. Review of Marketing and Agricultural Economics, 23: 215-36.
14
See Jawetz, M. B. Farm size, farming intensity and the input-output relationships of
some Welsh and West of England dairy farms. University College of Wales, Aberystwyth.
1957. P. 9.
labor input actually used in deriving the given production, not a measure
and unutilized, available during the production
of total labor, utilized
period. (Unless detailed records have been kept by the farmer, the re-
quired information may not be available.) Second, account should be
taken, so far as possible, of variations in labor quality. Work perform-
ance not only varies with age but also with other factors. The above
comments apply especially to farm family labor. Hired employees
work under supervision and receive wages presumably corresponding
to their input. Where hired labor is a substantial portion of total in-
put, it might best be measured in a separate category to make full use
of the more reliable measurements. But if it is used directly as a sub-
stitute for family labor, the two might best be grouped together.
Land
The difficulties associated with measuring the land input are not so
intractable as those associated with labor and capital. First, the sample
of observations can be confined to farms that are relatively homogeneous
with respect to land quality 15 Secondly, the flow of services from the
.
land input seems to be reasonably indexed, both between and within

farms, by measurement in acreage or market value terms so long as
only land actually used in production is included as input. Thirdly, dif-
ferences in land quality are likely to be reasonably reflected in market
values so that a standardized measure of the land input may be used to
at least take account of gross differences in land quality. Still, the ef-
fects of differences in land quality may be serious. They are likely to
be systematically related to differences in other inputs since, usually,
the better quality land tends to be farmed more intensively. The re-
sulting marginal productivity estimates derived from a fitted production
function suggest that labor and capital, if employed in larger quantities
. on the poorer land, would give returns similar to those achieved on the
better land. Similarly, the production elasticities may be overesti-
mated. Failure to standardize the land input observations provides
16
fitted functions which are hybrids relating to different land qualities .
Management
The most difficult input to measure in farm-firm studies is the

management or entrepreneurial factor. Unlike experimental studies
where the managerial function is exercised by the investigator and re-
mains fixed, the farm-firm research worker is faced with observations
whose genesis rests not with him but with many decision makers, each
15
See Heady, E. O. Productivity and income of labor and capital on Marshall silt loam
farms in relation to conservation farming. Iowa Agr. Exp. Sta. Bui. 401. Ames.
1953.
16
Heady, Earl O. Elementary models in production economics research. Jour. Farm
Econ., 30: 1-18.
possessing a different degree of entrepreneurial ability. Such differ-
ences would be unimportant if management could be measured directly
in physical terms. It could then be handled in the manner of other phys-
ical inputs. However, a generally applicable scale for measuring man-
agement has not yet been devised. Likely it never will be, except in
terms of interaction of such factors as time and intensity of cerebra-
tion, intelligence level, problem handling ability, etc. Still, it may be
feasible to scale the management factor as it relates to a given produc-
tion process. Such scaling techniques have sometimes proved fruitful
in handling nonphysical variables in psychological and sociological
analyses.
Granted that differences in management input are likely to exist in
a cross-sectional sample of farm data, danger exists in the possibility
that differences in inputs of management are systematically related to
variations in the input of other factors. However, perfect correlations
are not to be expected between management and other inputs. While a
large scale of operation may often be associated with a high level of
management, it may also be a reflection of waste. Likewise, a small
scale of production due to lack of knowledge or initiative reflects a low
level of management. Still, a small scale of enterprise often results
from lack of resources due to circumstances beyond the control of a
capable entrepreneur. Nonetheless, it is reasonable to hypothesize that
a significant positive correlation exists between managerial and capital
inputs. Also, increased nonmanagerial inputs probably can be effec-
tively utilized with less than proportionate increase in management. If
these hypotheses do hold, then, as Griliches has shown, the omission of
management may lead to underestimation of returns to scale and over-
17
estimation of capital returns Should these tendencies be regarded as
.
serious disadvantages, the researcher may attempt to minimize them

by (a) attempting to restrict the farm sample to a group relatively homo-
geneous with respect to the level of management, for instance, to good
or poor managers on some subjective rating scale; (b) attempting to
select a farm sample that minimizes correlations between management
and other factors; or (c) by introducing some measure of management
18
into the production function analysis .
A few attempts have been made to include management as a factor
in empirically fitted production functions. None of these has been re-
ported fully in the literature, so it must be assumed that little success
was attained in devising an adequate managerial scale. In such studies
the usual procedure has been to rate the sample entrepreneurs on a
management index relative to their knowledge of farming practices and
techniques and the degree of economic rationality thought to be shown
17
Griliches, Z. Specification bias in estimates of production functions. Jour. Farm
Econ., 39: 8-20.
18
See Johnson, G. L. Problems m studying resource productivity and size of business,
arising from managerial processes, m Heady, E. O. et al. (eds.) Resource productivity,
returns to scale and farm size. Iowa State University Press, Ames. 1956. Pp. 16-23.
Also, Westermarck, N. Management and success in farming. Acta Agriculturae Scandi-
navica. 8: 375-403; 9: 164-80.
by their current managerial decisions relative to the use of recom-

mended practices. Such indices have three major disadvantages.
First, they may not adequately distinguish between knowledge and
entrepreneurial logic. Secondly, there is a danger that the management
index may tend to measure the managerial potential or capability of the
entrepreneur rather than his actual management input over the produc-
tion period being analyzed. Lastly, such indices suffer from the fact
that they incorporate subjective elements. What one assessor might
regard as indicative of a large input of management might typify a low
input to another assessor. Still, such difficulties of subjective judg-
ment are not likely to be too important. For instance, in an unpublished
study (1948) by Heady, a management index was used to measure man-
agerial inputs; the managerial input ratings based on this index and
scored independently by two people had a correlation coefficient of .91.
With management incorporated, the Cobb- Douglas function estimated in
this study is given in equation 7.1 where Y is output and X x , X 2 X 3 and , ,
X 4 are land, labor, capital services, and management, respectively.

(7.1) Y = .038X*
273
X*
3
^ 3.383 ^-1.002
A 4
Obviously, the estimate of production elasticity with respect to manage-

ment is untrustworthy. Still, it is statistically true that for the farm
sample studied, increasing returns to scale did appear to prevail when
management was included. With management excluded, constant returns
to scale seemed to prevail. The sample of farms to which equation 7.1
refers were Iowa farms which had successfully repaid and utilized a
loan from the Farmers’ Home Administration. This restriction may in
part explain the high elasticity of production for the management factor
since it is to be expected that returns on resource inputs obtained from
loans may depend to a greater extent on management than on any other
single factor.
Given the difficulties inherent in measuring management, a reverse
approach has sometimes been suggested. It is that the residuals between
production levels estimated from the fitted function and the actually ob-
served production levels be attributed to management. These residuals
would then be used as the basis for an objective management rating. 19
Thus in Figure 7.1, using the single input simplification, the fitted func-
tion might be AB. The actually observed output-input combinations
might be those marked. Those lying above AB would be given a positive
management rating and those below a negative rating, each rating being
proportionate to the size of the residual. Such an approach has a num-
ber of disadvantages. The residuals may not be related to management
but to other factors. Also, this procedure only considers one aspect
of the management question — that typified by wasteful or inefficient use
of a given resource bundle. The residual between estimated and
19
See Heady, E. O. Production functions from a random sample of farms. Jour. Farm
Econ., 28: 1003. Swanson, E. R. A measure of economic success in farming. Acta Agr.
Scand. 9: 485-96.
Figure 7.1. Residuals between input-output observations

and the fitted production function.
observed output levels, where the observed is the smaller, indicates

how much more production might, on average, be expected from the
observed input levels. The other contribution of management, pertinent
in farm situations when decreasing returns prevail, is to drive the en-
trepreneur to the optimum position on the production function as deter-
mined by the prevailing price regime. Due to exogenous resource re-
strictions, however, the "distance” between observed and optimal input
combinations may give no indication of the management input that is
being used. Relative to the management factor as it relates to wasteful
resource use, it might be noted that resource waste may also be re-
garded as equivalent to an error of observation in the data. Thus, sup-
pose a farmer used 15 units of a resource in such a fashion that they
were only equivalent to 10 units applied efficiently. The true input
would then be 10 units and not 15 units. However, by force of circum-
stances associated with data collection, such discrepancies cannot be
detected. They remain as errors of observation in the data, closely
akin to errors arising from quality differences within a single resource.
OUTPUT
Aggregation problems generally exist with respect to output in
farm-firm studies. The difficulty arises because most farms produce
either more than one type of product or several qualities of the same
product. Where particular input quantities can be associated with each
type or grade of output, distortions caused by aggregation may be
avoided by fitting a separate function for each product. To this end,
separate functions have been fitted for crop and livestock enterprises
20
in a number of Iowa cross-sectional studies. Iowa studies, using
See Heady, E. O. Resource productivity and returns on 160 acre farms in north cen-
simultaneous equations under the assumption of joint relationships,

have not proved direct interdependence between broadly grouped out-
puts such as crops and livestock. Provided that the productivity of re-
sources is independent between the different products, the production
function for each enterprise can be validly used to estimate marginal
resource productivities, isoquants, and isoclines relative to each prod-
uct and, hence, to ascertain the optimum allocation of resources be-
tween the various enterprises. However, such estimates must be in-
terpreted cautiously if resource productivities are not independent
between the different enterprises; the fitting of a separate function for
each product assumes that the productivity of a resource relative to a
specific type of output is uninfluenced by the level of resource use as-
sociated with the other products produced by the firm. Still, except in
cases of gross interdependence between enterprises, estimates derived
from individual output functions are likely to be more reliable and use-
ful than recommendations derived from a single aggregative function
with all outputs lumped together. With products aggregated on a value
basis, nothing can be said relative to investments in individual enter-
prises.
In many instances, the researcher has no alternative but to use an
aggregative measure of value output. Estimation of functions for an in-
dividual enterprise is infeasible unless records are available indicating
the quantity of each input associated with each output. Such detailed
records are rarely available for a cross-sectional sample although they
may be available in time series form for some individual farms. If an
aggregative measure of output has to be used, the distortions caused by
output aggregation may be minimized by deriving separate functions
for groups of firms producing the various outputs in approximately the
same proportions. This approach requires a fairly large analytical
sample to ensure a sufficient number of farms in each group for which
a function is fitted. Still, the disadvantages of using large purposive
samples will generally be outweighed by the greater usefulness of the
estimates, compared to the “rough and ready” function obtained by
fitting a single over-all function to a group of multi- enterprise farms.
The justification of the grouping procedure is that the outputs within
each group of farms are treated as if they were joint products produced
in fixed proportions. For such products, aggregation would be war-
ranted since their mix is fixed and cannot be varied by the farmer.
Another approach to relationships in multi- enterprise situations
has been suggested by Plaxico: A function of the form of equation 7.2
21
might be used. Marginal
tral Iowa. Iowa Agr. Exp. St a. Bui. 412. Ames. 1954; Heady, E. O. Productivity and in-
come and capital on Marshall silt loam farms in relation to conservation farming.
of labor
Iowa Agr. Exp. Sta. Bui. 401. Ames. 1953; also Heady, E. O. and Shaw, R. Resource re-
turns and productivity coefficients in selected farming areas. Jour. Farm Econ., 36: 243-
57.
21
Plaxico, J. S. Problems of factor-product aggregation in Cobb-Douglas value produc-
tivity analysis. Jour. Farm Econ., 37: 664-75. See also Waugh, F. V. Regression between
sets of variates. Econometrica, 10: 290-310.
(7.2) X k0k
productivities, isoquants, and isoclines relevant to the i-th output, Y
i?
could than be estimated with other products held constant at specified
levels. However, unless resource productivities are strongly interde-
pendent among products, functions such as equation 7.2 may not be
satisfactory in terms of the economic, biological, or physical logic
underlying the production process. Moreover, the fitting of such func-
tions is computationally irksome relative to the estimation of production
functions involving but a single independent variable.
USE OF COBB-DOUGLAS TYPE FUNCTIONS
Of possible algebraic forms, Cobb- Douglas functions have been the

most popular in farm- firm analyses. This algebraic model provides a
compromise between (a) adequate fit of the data, (b) computational fea-
sibility, and (c) sufficient degrees of freedom unused to allow for sta-
tistical testing. In other words, the Cobb-Douglas is a relatively “effi-
cient user" of degrees of freedom. Such efficiency is important where
research resources are limited and collection of farm- firm data is ex-
pensive. However, it probably has greatest use in diagnostic analyses,
reflecting marginal resource productivities at mean levels of inputs.
Because of the elasticity restraints outlined in Chapter 3, estimates
departing above the mean may well overestimate productivities. Too,
under the aggregation procedures frequently necessary, the fitted func-
tions can tell the farm little about returns from specific investments.
The Cobb-Douglas function has been used less extensively for individual
technologies such as fertilizer and feed relationships. More reliable
response estimates have usually been attainable from nonmultiplicative
polynomial type functions. However, whenever a Cobb-Douglas type
function is to be used, certain procedures should be followed in assem-
bling and analyzing the data. If these adaptions are not made, the conclu-
sions drawn from the analysis will not be as reliable as they might be.
Geometric Versus Arithmetic Aggregation
Since some degree of aggregation is necessary in farm-firm studies,

machinery services arising from, say, tractors and milking machines
may be aggregated together as a single input. The usual procedure has
been to simply add the money value of such micro- inputs to provide an
aggregative input. However, the arithmetic sum of the micro- inputs
introduces bias in the resultant estimates, except when the micro- inputs
22
that are summed are always used in fixed proportions. Bias can be
22
See Shephard, R. W. Cost and production functions. Princeton University Press,
,
reduced by using as the aggregated input, not the arithmetic sum of the
micro- inputs, but their geometric sum, i.e., their product. Thus, if
micro-inputs x x x 2 ,
x n are to be aggregated into the single cate-
,
. .
.
gory X for use in a Cobb- Douglas function, aggregation as specified by

equation 7.3 must be regarded as less satisfactory than aggregation as
in equation 7.4. Intuitive justification for the multiplicative approach
toaggregation in Cobb-Douglas analyses is provided by equation 7.5
which corresponds to equation 7.4. It is obviously anomalous to use
equation 7.3 when the basic estimational feature of the Cobb-Douglas
function is a logarithmic transformation into a linear form.
(7.3) X = xx + Xg . . . + xn
(7.4) X = x1 x 2 . . . xn
(7.5) Log X = log x + log x

x 2 . . . + log x n
As Griliches has shown, the possible bias arising from using the geo-
metric sum can be determineda little auxiliary information is avail-
if
23
able. The procedure has been outlined in Chapter 6. However, if the
arithmetic sum method of aggregation is followed, it is virtually im-
possible to estimate the possible biases. Moreover, the multiplicative
basis of aggregation approaches more closely the ideal, but impractical,
method of aggregation for Cobb-Douglas functions. The perfect method
would be to aggregate the micro- inputs multiplicatively with each
micro- input weighted proportionately to its (unknown) elasticity of pro-
duction.
Zero Input Levels
The Cobb-Douglas function implies that at least some quantity of

each input must be used if output is to be nonzero. In the real world,
such a condition does not hold true; indeed, most samples of data,
either real world or experimental, will contain observations with one
or more of the inputs at zero level. The practical difficulty arises in
the conversion of the raw data to logarithmic form, the logarithm of
zero being minus infinity. To overcome this problem, assuming there
are strong grounds for using a Cobb-Douglas model, the zero observa-
tions may be replaced by some figure of arbitrary small size, (Or, a
constant can be added to all observations for the particular input cate-
gory.) For instance, if a fertilizer has been applied at rates of zero,
Princeton. 1953. Pp. 61-71. (This work is an important classical study of the production
function, although much of the text involves relatively abstract mathematics.) See also
Solow, R. M. The production function and the theory of capital. Review of Economic Stud-
ies, 23: 101-8.
23
Griliches, Z. Specification bias in estimates of production functions. Jour. Farm
Econ., 39: 8-20.
80, and 160 pounds per acre, the zero application might be taken as
being one pound per acre for the purpose of fitting a Cobb- Douglas
function. Too, when aggregating, the geometric aggregate would very
often be zero unless the adjustment to some arbitrary nonzero level is
made. Moreover, if zero observations occur quite frequently in the
sample, their piecemeal adjustment is unsatisfactory and may induce
nonnegligible errors. Alternative algebraic functions should be then
investigated.
Returns to Scale
Historically, in Cobb- Douglas analyses of farm-firms, the sum of

the estimated input coefficients has been taken as an indication of the
returns to scale. Thus, relative to equation 7.6, Zb* < 1 has been taken
as implying decreasing returns to scale and Zbi > 1 increasing returns
to scale.
h
(7.6) Y = aX^ 1 X 2 2 ... X^k
The procedure usually adopted is to perform a t test to ascertain if Zbi

from unity at the desired probability level.
is significantly different
However, such estimates of scale returns will be biased unless all input
24
factors are included in the production function. Returns to scale will
be underestimated if the excluded inputs vary less than proportionately
with changes in the included factors over the sample of observations.
If the reverse situation holds true the elasticity of production or returns
to scale will be overestimated. Thus omission of management from

farm-firm functions can take on importance. Intuitively it seems true
that as the scale of operation of a farm increases, the management fac-
tor does not need to increase to the same extent as other factors. The
exclusion of management may therefore be expected to lead to under-
estimation of returns to scale. Likewise, neglecting quality differences
in other input factors will lead to biased estimates of scale returns.
For instance, it seems most likely that the ignoring of quality differ-
ences in labor and land will generally lead to overestimation of the re-
turns to capital, underestimation of the returns to labor and land, and
underestimation of returns to scale.
If the research worker believes that constant returns to scale must
prevail if all input factors are included, then it is logical to test the
divergence between Zbi and unity as an indication of the importance of
omitted input factors. It may be argued that such an approach is
M See Jarrett, F. G. Estimation of resource productivities as iUustrated by a survey of

the lower Murray Valley dairying area. Australian Jour, of Stat., 1; 3-11; Heady, Earl O.
Production functions from a random sample of farms. Jour. Farm Econ., 27; 995-97; and
Griliches, Z. Specification bias in estimates of production functions. Jour. Farm Econ.,
39: 8-20.
PROBLEMS RELATING TO' ESTIMATION 231
logicallymore attractive than the procedure of regarding Ebi as an

estimate of.the true returns to scale.
Variance of Marginal Productivities
The formal derivation of marginal productivities has been noted in

Chapter 3. To reiterate, the marginal productivity of Xj
, the i-th input
in equation 7.6 is given by equation 7.7.
6Y = b
(7.7) i
6Xi .
The most reliable, and perhaps the most useful, estimate of marginal
productivity is obtained by taking X^ at its geometric mean; i.e., at the
value where log Xi assumes its arithmetic mean. Also, Y should be the
estimated level of output when each input is held at its geometric mean.
For further implications of the analysis, it may be desirable to estimate
marginal productivities when inputs are held at levels other than their
geometric means. However, the qualifications mentioned earlier should
then be kept in mind.
Given the estimate of marginal productivity from equation 7.7, the
next step is to calculate its variance. Any discrepancy between the esti-
mated marginal productivity and the existing opportunity cost of the re-
source may then be tested for statistical significance. Historically, the
variance of the marginal productivity has generally beeji derived by as-
suming Y and Xi to be constants. The variance of (bi Y/Xi) is then
given by equation 7.8 where Var(bi ) is calculated as in equations 4.29
and 4.30.
2
(7.8) Var(bj Y/Xi) = (Y/Xi) Var(bi)
However, the assumption that Y is a constant will seldom hold; Y

will vary over alternative samples since it is based on b values which
are only estimates of true parameters. Nonetheless, for estimates of
the marginal productivities with the inputs at their geometric means,
equation 7.8 leads to negligible errors in the variance estimate. These
errors, however, increase rapidly as marginal productivity estimates
are made further away from the geometric mean levels. The implica-
25
tions of nonconstancy of Y have been discussed by Carter and Hartley.
They derived a more accurate expression of the variance of the mar-
ginal productivity estimates. Their formula is given in equation 7.9
where Var(Y) is as shown in equation 4.37, m is the vector bi (log Xj -
log Xj) + u where u is one if i = j and zero otherwise, and C is the
matrix of equation 4.24 in terms of log X.
2*
Carter, H. O. and Hartley, H. O. A variance formula for marginal productivity esti-
mates using the Cobb- Douglas function. Econometrica, 26: 306-13.
)
(7.9) Var(biY/Xi) = (Y/Xj

2
(bf/n + mCm' ) Var(Y)
In equation 7.9, it is assumed that the logarithmic transformation used

in the least squares estimation is to the base e. More normally, a
transformation to the base 10 will be used, in which case the term
2
(b- / n) in equation 7.9 must be multiplied by 2.3026 .
RETURNS TO SCALE
Mention must now be made of scale returns in relation to the general

production function depicted by equation 7.10 where it is assumed, for
the moment, that no
(7.10) Y = f(X x , X2 ... ,

Xk )
relevant inputs have been excluded. Increasing, constant, or decreasing

returns to scale prevail at a given point on the surface depicted by equa-
tion 7.10 depending on whether a small proportional increase in all in-
puts leads to a more than proportionate, proportionate, or less than
proportionate increase in output.
Physical and Economic Returns to Scale
Practical statements about returns to scale can be made only if the

entrepreneur can actually make proportionate changes in all the inputs
considered. Where inputs are not under his control, he cannot make a
proportionate change in every input factor. There is then little point in
telling him that more profit can be attained by increasing or decreasing
his scale of operation if uncontrollable factors are included in the rec-
ommendation. Needed is an estimate of the returns to scale or produc-
tion elasticity for inputs under the control of the entrepreneur. The
noncontrollable inputs should be held fixed at known levels. Knowledge
of a crop farmer that he is operating under constant returns to scale
relative to land, labor, capital, moisture, and sunlight has little utility.
He cannot make decisions about the amount of weather. But he would
have useful knowledge if he knew the range of weather under which de-
creasing or increasing returns to scale hold relative to land, labor, and
various capital items. Distinction therefore may be made between
“physical returns to scale,* incorporating all inputs, and what might be
called “economic returns to scale.* The latter would include only those
inputs under the control of the entrepreneur. Empirical difficulty
arises, however, since exclusion of the uncontrollable inputs from the
analysis tend to bias estimates of the economic returns to scale. Such
biases will be discussed later.
Euler’s Theorem
We now examine how returns to scale at a given point on the general

production function indicated by equation 7.10 may be estimated. The
procedure is based on the relationship found to prevail between the two
sides, LHS and RHS, of equation 7.11 where Y is the estimated output
with the inputs X x to X k held at the levels used in evaluating the RHS. 28
(7.11) 6Y/6X 1+ X 2 5Y/5X 2 + ... +X k 6Y/6X k

Thus Y is the estimated output at the point on the production surface
defined by the specified levels of X x to X k On the RHS, 5 Y/ 6X* is the .
partial derivative of the estimate of equation 7.10 with respect to Xi.

If, on evaluating equation 7.11, it is found that the LHS equals the RHS,
then constant returns to scale prevail. Equivalently, by Euler’s Theo-

rem, equation 7.10 is homogeneous of degree one at the specified values
of X x , X 2 , .
.
,
X k . If the LHS is greater than the RHS, then decreasing
.
returns to scale prevail at the specified point. If the LHS is less than
the RHS, increasing returns to scale prevail. The actual size of the re-
turns to scale prevailing at the surface point under consideration is
given by dividing the RHS of equation 7.11 by the LHS. Since equation
7.11 can only be evaluated in terms of the estimate of equation 7.10
derived via statistical procedures, statements about returns to scale
must be regarded as probability statements and tested accordingly.
As a simple example of the use of Euler’s Theorem, consider equa-
tion 7.12. It is a two- factor Cobb- Douglas function, Y, a, b 15 and b 2
being empirical estimates.
b 1l b o2
(7.12) Y = aX, X 2
Corresponding to equation 7.11 we have
'1 2_1
2
(7.13) Y ^XiâX^ 1 X 2 )+ X 2 (b2 aX^ X 2 )
which reduces to
(7.14) 1 > bj. + b2 .
Thus if (^ + b 2 ) is greater than unity, increasing returns to scale pre-

vail, and so on. It is interesting to note that equation 7.14 is independent
of Xj. and X 2 In other words, returns to scale in equation 7.12 do not
.
depend on the level of the inputs at the point under examination. By ex-
tending equation 7.12 to a model with k input factors, it is easily seen,
26
For the derivation of equation 7.11 and its justification relative to returns to scale, see
Allen, R. G. D. Mathe matical analysis for economists. Macmillan and Co., Ltd., London.
1956. Pp. 317-22. For an interesting discussion, see also Wicksell, K. Selected papers on
economic theory. Harvard University Press, Cambridge. 1958. Pp. 93-100.
as noted in Chapter 3, that the Cobb- Douglas model implies an unchang-
ing elasticity of production over the whole of the production surface.
Bias in Estimates of Returns to Scale
As mentioned previously, bias will occur in the estimate of physical

returns to scale if any input factors are excluded. Moreover, the bias
will not be constant over different sets of observations from the same
population but will vary as the levels of the excluded inputs vary over
the sample of observations to which the function is fitted. It is possible
to estimate the bias in scale returns estimates resulting from the ex-
clusion of variable factors which are under the entrepreneur’s control
but which cannot be included in the analysis. The procedure is outlined
in equations 6.36 to 6.44. As an example, suppose equation 7.15 depicts
the true production function with all factors included. Alternatively, it
may be the production function with all excluded factors held at some
constant level throughout the sample observations. Corresponding to
equation 7.15, the function actually fitted might be equation 7.16, X3
having been excluded.
(7.15) Y = 0 O +^ 1 X 1 + |3 2
X2 + 03 X 3
(7.16) Y = b0 + bjXj + b 2 X 2
Denote by e the true returns to scale, either physical or economic, cor-

responding to equation 7.15 and by § the returns to scale estimated from
equation 7.16. By application of equation 7.11 we have:
(7.17) e = O x X4 + f3 2 X2 + 03 X 3 )/Y
(7.18) S = (b^ + bsXj/Y
The expected bias in e is given by equation 7.19.
(7.19) E(e - e) = E[(bl X +b 2 X 2 )/t-

l (^ 1 X1 +02 X2 + j3 3 X 3 )/Y]
Assuming that Y equals Y and making use of equation 6.43, equation 7.19
simplifies to equation 7.20 where the p’s are the auxiliary regression
coefficients of equation 6.42.
(7.20) E (e- e) = + p 23 X 2 - X 3 )/Y

Hence, relative to equation 7.15 we can say that the size of the bias in
e resulting from the omission of X 3 varies as the level of X 1? X 2 , and
X 3 vary. Also, for the most likely case with j3 3 positive and absolutely
large relative to p 13 and p^, underestimation is the type of bias to be
expected.
TECHNOLOGICAL CHANGE
Farm input- output data, either time series or cross-sectional, is
generally confounded by technological change. Roughly speaking, the
production surface relevant to an individual firm tends to rise over time
it becomes possible to produce the same quantity of a given product
with less effort, or to produce a greater quantity of a given product with
the same effort. These effects may occur for one or both of two rea-
sons. First, new types of inputs may be added to the input array used
in the production process and old types may be discarded. An example
would be addition of a welding plant to the machinery services used.
More obvious, but rather more subtle in its implications, would be a
shift from animal to tractor power. The second source of technological
change is improvement in the quality of the inputs used. For instance,
the quality of the labor input on farms may improve over time under
the influence of higher educational standards and an increasing inventory
of experience. Ideally, of course, the use of improved quality inputs
might be regarded as the substitution of new inputs for old. However,
it is convenient to consider quality changes and distinctive changes in
the input array as separate phenomena. The justification is that quality

changes are not easily discernible in the short run. The use of inputs
of improved quality may often occur without any conscious decision on
the part of the entrepreneur. In contrast, changes in the input cate-
gories used are quite easily discerned. Moreover, they generally occur
as the result of a definite decision by the entrepreneur.
Technological change resulting from both the use of new inputs and
improvements in input quality is reflected directly in time series data.
Over time, given sufficient data, it is easy to pinpoint the addition of
new types of inputs to the production process. To delimit changes in in-
put quality is a more difficult task unless supplementary information is
available. Such supplementary information might come from semiex-
perimental studies of individual inputs. For instance, improvements in
particular types of machinery may be gauged from engineering reports
on the performance of the machinery. In cross-sectional data, techno-
logical advance due to changes in the input array is only reflected in-
directly by way of the differential rates of adoption prevailing among
farms. Cross-sectional surveys usually disclose some firms operating
with an “old” input array, some with a “new* array, and some in a
transitional stage. Input quality differences also occur in cross-
sectional data. However, such differences largely represent rather
fixed discrepancies in quality inherent in the fixed resources of the
firms. Probably only to a minor extent do cross-sectional quality dif-
ferences reflect the positioning of firms at different stages in the se-
quence of technological advance. Still, exceptions would occur if the
cross-sectional sample "related to a population whose elements were
not relatively contiguous.
So far as the use of new input types is concerned, it would seem pos-
sible to take account of such technological change by extending the input
framework of the production function analysis. The problem of techno-
logical change due to improvements in input quality is, however, rather
intractable in an operational sense. We now consider shifts in the pro-
duction function for a given product arising from distinctive changes in
the set of inputs used.
New Input Types
When new types of inputs are added to the production process, some
of the input types previously used may be discarded. Thus, the old and
new processes might correspond to equations 7.21 and 7.22, respec-
tively.
(7.21) Y = f(X x , X 2 ,
Xk )
(7.22) Y = f(X 1? X2 ,
X k _ m X k+1 , ,
..., X k+h )
As depicted in equation 7.22, the new production process involves the

exclusion of m inputs, X
k - m+1 to X
k and the addition of h new inputs,
,
Xk+1 to X k+k
However, the old and new production processes corre-
.
sponding to equations 7.21 and 7.22 may be equally well depicted by the
single function shown in equation 7.23. Relative to the old production
process, X k+1 to X k+k are always at zero level in equation 7.23. Rela-
tive to the new production process, X k _ m+1 to X k are always at zero
level in equation 7.23.
(7.23) Y = f(X 1? X2 , ..., X k+h )
For transitional stages between the old and of produc- new techniques
tion, most or all would be at nonzero
of the variables in equation 7.23
levels. The fact that some inputs are zero at different times does not
affect the estimation of equation 7.23 by normal least squares proce-
dures. The only proviso, and it is a major one, is that sufficient data
be available for computational purposes. It might be noted that the sub-
suming of the old and new production functions under the broader frame-
work of equation 7.23 is related to the discussion of hybrid production
functions in Chapters 5 and 6, To attempt to fit either equation 7.21 or
7.22 to a set of data encompassing both the old and new processes
would result in the fitting of a hybrid function. But to fit equation 7.23
would not lead to hybridity, since equation 7.23 recognizes that X k+1 to
Xk+h are zero in equation 7.21 while X k - m+1 to X k are zero in equation
7.22.
Changes in Input Quality
Reference has already been made in Chapter 6 to the problem of

variation in input quality, at least insofar as it relates to aggregation.

Here we are not concerned with quality variations within a given obser-
vation but with variations in quality between observations taken over
time. These changes in quality may be either positive or negative. For
convenience, we speak of the positive case — an improvement in quality.
Ifquality changes are not taken into account, the effects on the derived
estimates are equivalent to the biases resulting from errors of obser-
vation in the data. The situation is perhaps best illustrated by a simple
numerical example.
Suppose that an input is such that of every 10 units used in the pro-
duction process, only 6 units are actually utilized; the remaining 4 units
lost due to frictions beyond the control of the entrepreneur. Further
suppose the quality of the input is improved so a 10 per cent increase in
utilization results. Then, for every 10 usable units, 6.6 units would now
be actually utilized and only 3.4 units lost. For the purpose of estimat-
ing a production function, the relevant data for “before” and “after” the
quality change are the figures of 6.0 and 6.6 units, respectively. To use
the figure of 10 units for both situations would introduce a sizeable error
of observation. While the figure of 10 would serve as a satisfactory and
legitimate index so long as quality remained unchanged, it becomes an
inaccurate index as soon as quality alters significantly. For the two
situations discussed, the two readings of 10 units are not comparable.
Corresponding to the figures of 6.0 and 6.6, the error-free index values
would be 10 and 11, respectively. The practical difficulty is that only
the raw data, corresponding to observations of 10 for both cases, are
generally available. If data on quality changes were available, it would
be possible to index all observations on a particular input in terms of a
standardized unit as outlined in relation to equations 6.46 and 6.47. The
production function with all inputs evaluated in standardized units might
then be depicted as in equation 7.24.
(7 . 24) Y = F (X i X 2
, j
. . • , Xk , qx , q2 ?
• • • ? )
In equation 7.24, qij would be a weight used to convert the j-th observa-
tion on Xi to units of a standard quality. However, due to lack of knowl-
edge about the qij values, evaluation of equation 7.24 is generally in-
feasible in an empirical context. Still, an alternative procedure exists
by which approximate account may be taken of such technological
27
change. If quality changes are regarded as equivalent to the introduc-
tion of new inputs as they should in a strict theoretical sense, then a

new production function, such as equation 7.22 instead of equation 7.21,
becomes relevant. The situation is exemplified in Figure 7.2. The
27
It is worth noting that changes in input quality that occur without any change in the
basic input array, imply a movement along the production function estimated in terms of
standardized input units. So long as quality changes are accounted for by some standardi-
zation procedure, the position of the surface will only be altered when new inputs enter the
production process. The relevant definition of a new input is that it be one that cannot be
standardized in terms of any of the inputs already in use. A new input must be of a differ-
ent generic nature to any of the inputs already in use.
,
Figure 7,2. Shifts in the production function

due to technological change.
curve A
corresponds to the production function of equation 7.25 with
only Xallowed to vary. The B curve also corresponds to equation
x
7.25, but implies a proportionate increase in the quality of each of the

inputs. The input of X x is measured along the horizontal axis in un-
standardized units, no adjustment being made for the improvement in
quality. Strictly, the B curve corresponds to equation 7.26 where the
prime on the X’s denotes an improvement in quality relative to equation
7.25.
(7.25) Y = f(Xi
|
X2 , ..., Xk )
(7.26) Y = f(X u
|
X 21 ,
• « • >
Because of the assumption of a proportionate quality improvement in

each input, the B curve corresponds to the A curve with every ordinate
shifted upward by the same proportionate amount. Thus the correspond-
ence between the A and B curves, or between equations 7.25 and 7.26
might be expressed as in equation 7.27 where z is the shift factor. Since
curve B lies above A, z is greater than unity. (In drawing the B curve
of Figure 7.2, z was taken as equal to 1.5.)
(7.27) f(X u |
X 2l ,
. .
.
X kl ) =zf(X l |
X 2> Xk )
As equation 7.27 implies, the shift from function A to function B is such

that the marginal rates of substitution between inputs remains unal-
tered. Technological change of this type has been termed “neutral.” It
is, of course, a special case. Still, it may serve as a working approxi-
mation in empirical studies involving aggregative input categories. If
quality improvements occur in only some of the inputs, or if all im-
prove but in different proportions, then a curve such as C might result.
It is also possible to look at distinctive changes in the input array
in terms of Figure 7.2. For instance, the curve A might correspond to

equation 7.25 and the curve C to equation 7.28 which involves an addi-
tional h input factors over and above those used in equation 7.25.
(7.28) Y = f(X,
|
X 2 ...,X k+h
, )
Quantification of Technological Change
As indicated in discussing Figure 7.2, technological change resulting

from either the use of new inputs or quality improvement in old inputs
may be regarded — at least approximately — as causing an upward shift
in the originalproduction surface. Such a shift may be formally incor-
porated in the production function. For instance, consider equation 7.29
in which the inputs X x to X k are assumed to be mildly aggregative
rather than micro- inputs of highly specialized nature. If the inputs
were not aggregated to some extent, a formulation akin to that of equa-
tions 7.23 or 7.24 would be pertinent. The variable, t for time, is in-
serted in equation 7.29 to allow for the influence of technological change.
(7.29) Y = F(X1? X2 ...,X k ;t)

,
Using such a framework, Solow has suggested (and Illustrated empiri-

cally) how an approximate assessment of the extent of technical change
28
in time series data may be made. The measure that he suggests is
AF/F where this statistic is calculated by way of equation 7.30.
k
(7.30) AF/F = AY/Y- Z Wj AX^X^^
where
(7.31) AY = Yt - Yt _ x ;
AX £
= X i>t - X ift-1
and Wi is the relative share of in the total product. The ideal deri-
vation of Wj would be as in equation 7.32 which, however, presupposes
knowledge of equation 7.29, the function to be estimated.
5Y Xj
(7.32) w
6Xi Y
4
An estimate of w* might be obtained by assuming, perhaps heroically.
28
Solow, R. M. Technical change and the aggregate production function. Review of Eco-
nomics and Statistics, 39: 312-20. See also Leser, C. Production functions and British coal
mining. Econometrica, 23: 442-6; and Aukrust, O. Investment and economic growth.
Productivity Meas. Rev. 16; 35-53.
that the farm-firm is operating at or near its equilibrium position. At
equilibrium, the outlay on each factor corresponds to its share in the
total product. Hence, denoting the price of Y and Xi by p and p i? re-
y
spectively, w i might be approximated via equation 7.33.
(7.33) Wj = PiXi/p y Y
In the special case of neutral technological change, equation 7.29

takes the form
(7.34) Y = Z(t)f(X x , X2 ,
Xk )
where the factor Z(t) corresponds to z of equation 7.27. Z(t) denotes the
value of z for the set of observations for the t-th time period. AF/F is
then equal to AZ/Z and the latter, calculated by way of equation 7.30,
may be used to estimate Z(t) as in equation 7.35. For convenience,
Z(t = 1) may be set equal to unity.
}
Z(t + 1) = Z(t)
(7.35) (1
+
^| )
Dividing Yt by Z(t), an adjusted series of output data is obtained. These

data may be used to estimate the basic or original production function
denoted by equation 7.36.
(7.36) Y = f(X 1# X2 , Xk)
It is this function which prevails at

t = 1. The shift factor Z(t) specifies
the upward production surface at time t relative to the orig-
shift in the
inal position of the surface at t = 1. Thus if curve A of Figure 7.2 is
specified as the original production surface, curve B corresponds to a
neutral shift with Z equal to 1.5.
Whether or not neutral technological change has occurred can be
assessed empirically, at least in an approximate way. If the technical
change is neutral, the AF/F values will be independent of the Xi values.
A plot of AF/F against Xi should suffice to ascertain whether the tech-
nical change can be regarded as neutral. Unless a high degree of con-
fidence can be placed in the Wi values used in calculating AF/F, the
use of more abstruse statistical tests would seem unwarranted.
PROBLEMS IN LIVESTOCK STUDIES
As noted above, there is no dearth of problems in the estimation

and interpretation of farm-firm production functions. In large measure,
these difficulties arise because the researcher has no ex ante control
over the generation of the firms’ inputs. The farm-firm research
worker can only exercise ex post discretion over the data to be
analyzed. For instance, purposive sampling procedures may be used.

Animal or crop input- output relationships, on the other hand, are usu-
ally studied under experimental conditions; the research worker is able
to exercise a fair measure of control over the genesis of the data. Still,
experimental control with animals is not absolute. They do not behave
in a strictly passive way when faced with an array of possible inputs.
In consequence, some special difficulties arise in the estimation of
livestock production functions. Chief among these are the problems of
free versus controlled feeding, physiological restraints on the feed mix
vector, changes in the nutritional requirements over time, and forage
intake under grazing conditions. Note also needs to be taken of the im-
portant distinction between the animal itself as product (as in meat pro-
duction) and the situation where the product is a flow over time from the
animal (as in milk, egg, wool, and power production).
As well, there are some other aspects of livestock input- output anal-
29
ysis that should be mentioned. First, as experimental units, livestock
are short-lived and expensive. Research resources being limited, the
researcher has to seek a compromise between (a) the number of ani-
mals to be penned together as a single experimental unit within a repli-
cate, (b) the size of a replicate in terms of treatments, and (c) the num-
ber of replicates to be studied. It is important that sufficient treatments
be examined to provide an estimate of the relevant sector of the produc-
tion surface. Still, excessive experimental units (i. e. pens of animals)
,
should not be used to provide more information than is necessary.

These additional experimental units might be better utilized to increase
the number of animals within each pen; or to provide a sufficient degree
of replication to enable estimation of experimental error. As outlined
in Chapter 5, lack of fit analysis may then be carried out. Another gen-
eral problem is that of autocorrelation in successive observations on
the same animal. This difficulty could be circumvented if sufficient
research resources were available to enable each treatment to be ob-
served on a different animal or pen of animals; or to space the observa-
tions on each animal sufficiently far apart in time to minimize autocor-
relation. Even so, such approaches may not be wholly desirable in
terms of the finesse of the experiment itself: the effects of variations
among animals may be worse than the autocorrelation effects. As noted
in Chapter 4, autocorrelation presents problems mostly for probability
statements and fiducial limits, rather than in predictions of production
and substitution rates. In other words, it mainly raises problems in
making tests of significance. Experimental error due to variations be-
tween animals may, on the other hand, lead to false estimates of the
true production function. Of course, as usually conducted, livestock
experiments will contain some errors from both these sources. An
indication of how they may be assessed has already been given in chap-
ters 4 and 5.
29
For a general reference to livestock input-output considerations, see Hoglund, C. R.
et al. (eds.)Nutritional and economic aspects of feed utilization by dairy cows. Iowa State
University Press, Ames. 1959.
Feed Mix Restraints
The treatments applied, in sequence, to an individual animal gen-

erally need to be relatively consistent in terms of their feed mix make-
up. This point can be illustrated by means of Figure 7.3. It represents
a bird’s-eye view of the feed plane for the overly simple case of two
feed inputs, say harvested forage and grain. The lines R 1? R 2 , R3 , and
R4 are ration lines; along each such line the forage-grain ratio remains
constant. Suppose the points ty are taken as defining the desired or
observed treatments, the treatment being the pounds of grain and forage
corresponding to the point ty The points ty have been located purely
.
for e3q)ositional convenience. They are not meant to represent an ideal

experimental layout. Each treatment consists of the cumulative total of
grain and forage utilized by the animal up until the time of the observa-
tion. For the observations to be comparable, each of the animals should
be as similar as possible in terms of age, weight, nutritional back-
ground, breed, etc., at the start of the experiment.
Since the treatments of Figure 7.3 are cumulative feed totals, there
would be no difficulty in observing a sequence such as (t31 , t 32 , t33 ) on
an individual animal. But the sequence (tu tj 2 t 43 ) would be clearly
, ,
impossible on mechanical grounds since t 32 implies a larger forage in-

take than More important, the sequence of treatments followed by
an individual animal should not diverge too rapidly from a ration line.
If the proportions of roughage and concentrate in the ration are suddenly
altered markedly, the physiological processes of the animal may be up-

set; especially if the change is in the direction of more concentrate.
While such effects are not likely to be severe in the case of smaller
Figure 7.3. Schematic representation of ration

lines and treatment locations in the
feed plane for a dry-lot experiment.
animals such as poultry, the effects may be severe for larger- bodied
animals such as cattle and hogs. 30 With such animals a negative re-
sponse, even death, may result. Thus, the sequence (t 1JL t 12 t 23 ) may , ,
be feasible by following a feed path such as dotted between t 12 and t 23 .
In contrast, the sequence (t 21 t43 ) is rather infeasible for a large-bodied

,
animal. It would imply immediate change to a ration of nearly 100 per

cent grain after reaching t 21 While the animal might survive such a
.
31
swift shift in the feed mix, it would probably lose weight for a period.
Output increments would be negative. In contrast, an animal following
the path along R 4 to t 43 would show positive output increments. Total
production at t 43 would most likely be greatest if t43 were reached by
following the R 4 line rather than by following the path (t n t^). Hence, ,
the latter is an inefficient feed path; by moving along R 4 a greater out-

put can be obtained from the same quantity of input. Thus the produc-
tion surface that is probably most efficient in a physical sense and most
relevant to farmer decisions is best determined by moving each animal
along a ration line. Interpretations of this production surface in terms
of optimal movements over the feed plane as feed and product price
change should be qualified. We need to remember that recommenda-
tions for large-bodied animals, implying a substantial shift from the
current feed mix, may lead to a point on another surface; most likely
32
to a point on a lower surface. Recent research has shown, however,
that if the adjustment is not too severe in magnitude, the animal may
respond after a short period of adjustment, as if it had always been on
the same ration. In other words, the effect of the previous ration may
not last long.
These points can be illustrated by Figure 7.4 for a two-input situa-
tion. Suppose an animal is fed along each of the ration lines Rj and R 2 .
Each animal will operate on a production surface relative to the ration
line it is following. Because of its physiological adjustment to this ra-
tion, sudden and large movements of the animal s feed mix off the ration
5
line may immediately lead to less efficient use of the feed input than by
an animal continuously on the new ration line. Hence, product contours
indicating the effect of sudden and large movements away from the ra-
tion line may be drawn for each animal following a particular ration
line. For the animal on ration line Ri two such isoquants are shown
,
as Mn and i2 . M Corresponding isoquants for an animal on R 2 are 21

M
and M
22 . Mu
and M
2i each refer to the same amount of product. Like-
wise, M
X2 and M
22 refer to an identical level of output but one which is
30
See Heady, E. O. et al. Least-cost rations and optimum marketing weights for turkeys.
Iowa Agr. Exp. Sta. Res. Bui. 443. Ames. 1956. P. 870; Jacobson, N. L. Problems in de-
signing feeding experiments from a nutritional standpoint, in Hoglund, C. R. et al (eds.)
.
Nutritional and economic aspects of feed utilization by dairy cows. Iowa State University
Press, Ames. 1959. Pp. 206-12.
31
Relative to cattle, see Southcott, W. H. Drought feeding of beef cattle. Australian Vet.
Jour., 35: 126-29. For sheep, see Briggs, P. K. et al. The performance of adult metino
wethers fed weekly on all-grain rations. Australian Vet. Jour., 32: 299-304.
32
For an indirect indication of this point, see Heady, Earl O. A production function and
marginal rates of substitution in the utilization of feed resources by dairy cows. Jour.
Farm Econ., 33: 485-98.
-J
Figure 7.4. Schematic representation in the

feed plane of product contours
around ration lines.
larger than the level corresponding to M M

u and 2 i* Comparison of
points such as A and B shows each animal to be more efficient along
its current ration line than on another ration line to which it might
suddenly be shifted. The same output on the “new” ration line may be
obtained with less input from an animal always following the “new” ra-
tion line.
experiments to derive a response surface for production deci-
In
sions, care must be taken to ensure that there are sufficient treatment
responses. In recognition of this fact, livestock production surfaces
usually are estimated by studying response along an array of ration
lines covering the surface portion of interest. The isoquants derived
from most livestock experiments therefore correspond to “isoquants”
obtained by connecting up series of points such as A and C, or D and E.
The procedure is reasonable for large-bodied animals if it is remem-
bered that such isoquants do not always relate to production possibil-
ities from a single animal once it has started on its feeding program,
particularly if wide shifts are to be made from a current ration line.
Often it is necessary to move gradually to a desired ration line,
especially by large-bodied ruminants for feed mixes high in proportion
of grain or concentrates. Thus in Figure 7.5, responses along the ra-
tion line R 4 may be quite different, depending on whether it is approached
directly (i.e., always using the feed mix ratio implied by R 4 ) or as it is
reached gradually by, say, the dotted path. The use of such an approach
gives an animal the chance to gradually adapt its physiology to the ex-
treme feed mix implied by R 4 Of course, the use of the dotted path
.
implies that the surface area lying between this path and the grain axis
t
i0 POUNDS OF GRAIN PER ANIMAL
Figure 7.5. Schematic representation in the
feed plane of a free feeding dry-
lot livestock experiment
cannot be estimated. So long as this surface section is not relevant to

real-world decisions, as might be the case in meat production, the use
of the indirect path is reasonable. However, in circumstances where
points on the surface below the dotted path are relevant, as may be the
case for drought feeding conditions, the use of a gradual approach to R 4
is not justifiable.
Treatment Characterization
treatments used in livestock studies should be specified

Ideally, the
in terms time over which the input
of (a) the actual feed input, (b) the
acts, (c) the animal receiving this input, (d) the previous nutritional
history of the animal, and (e) the nonnutritional environment of the ani-
mal. Obviously, the treatments cannot be specified so exactly. The re-
searcher mainly can attempt to control aspects (c), (d), and (e) so as to
enable valid analysis over responses for various feed inputs and time
periods. There remains a problem in characterizing the feed inputs
used. Often no more can be done than to specify the inputs in such ge-
neric forms as alfalfa hay, sorghum silage, corn grain, etc., with an
attempt being made to ensure that the feed within each such category is
of some specified uniform quality. Such a procedure has an obvious
disadvantage from the point of view of making recommendations to real-
world decision makers. There are too many different qualities and
types of feeds for a production surface to be evaluated for each of the
possible combinations among feed types and livestock. It would be far
better if feed inputs could be characterized in terms of a basic stand-
ard referring to the chemical and physical character of feed 33 As yet, .
such standards are not available. If they were, experiments with a

small number of feed types might be used as the basis for developing
production surfaces of relevance to a wide range of feeds. The cur-
rently available methods of expressing nutritive values (such as TDN,
net energy, digestible protein, etc.) are not suited for use as basic
standards in the estimation of livestock production surfaces. They gen-
erally assume that a given feed quantity has a constant nutritive value
independent of both the quantity of feed utilized and of the character of
34
the feed mix, and also of the condition of the animal In other words,
.
they assume constant marginal rates of substitution between feed types

and constant marginal productivities of feed types. Such a linear
framework of comparison between and within feeds assumes away the
possibility of a production surface showing increasing or diminishing
returns.
Free Versus Controlled Feeding
Compared to such production units as machines, soils, and plants,

animals are relatively “free” agents. They exercise choice in obvious
ways. In consequence, it is more difficult for the researcher to exer-
cise control over the treatments used in livestock experiments. This
difficulty is enhanced in production function experiments since they re-
quire a large number of distinct treatments.
Two extreme possibilities face the researcher designing a drylot
livestock experiment. He may either (a) control the quantity of feed to
be supplied in a given time period or (b) he may specify the time inter-
val but not the quantity of feed to be consumed. Intermediate positions
between these extremes are possible: Some elements of the ration may
be supplied ad libitum while others may be controlled; or, within one
experiment, identical feed mixes might be fed under both controlled and
free-feeding conditions. Whichever approach is used, certain elemen-
tary precautions need to be taken. So far as possible, the various feed
types in the over- all ration, or in the segregated parts of the ration
under a mixture of controlled and free feeding should be mixed. Other-
wise, an animal may exercise its preference strongly, gorging itself
on one type of feed and ignoring others. Again, the supply of a given feed
type should be as homogeneous as possible in terms of quality and other
characteristics during the experiment. In the initial stages of an ex-
periment, allowance should also be made for the adjustment of the
53
See Lucas, H. L. Experimental designs and analyses for feeding efficiency trials with
dairy cattle, in Hoglund, C. R. et al. (eds.) Nutritional and economic aspects of feed utiliza-
tion by dairy cows. Iowa State University Press, Ames. 1959. Pp. 189-92.
34
Some biological researchers have attempted to remedy these defects. See Blaxter, K.
L. and Wainman, F. W. Some observations on the digestibility of food by sheep and on re-
lated problems. British Jour, of Nutrition, 10: 69-91; also Brown, W. G. and Arscott, G. H.
Animal production functions and optimum ration specifications. Jour. Farm Econ., 42; 69-78.
,
animals’ tastes and physiologies to the feeds and feed mix being studied.
Precautions must be taken against (or allowances made for) discrepan-
cies between the quantity of feed placed before an animal and the amount
actually consumed. For instance, some of the feed supply may be tram-
pled underfoot or made unusable in other ways. For the biological in-
vestigator, the above comments are quite mundane. Still, they are fea-
tures of the experimental situation that must be recognized by the
cooperating economist.
With free or ad libitum feeding, an animal is always faced with a
quantity of feed in excess of the amount it can immediately consume.
An animal will tend to utilize feed at the limit of its stomach capacity,
provided that the ration is not unattractive or distasteful. In either
case, we cannot specify in advance the feed quantities that will be uti-
lized in a given time period. We can only specify the proportion of dif-
ferent feeds in the ration placed before the animal and the times at
which input and output will be measured. The treatments observed with
free feeding are largely determined by the animal. The situation may
be illustrated by way of Figure 7.5. Suppose the investigator decides
to evaluate the beef production surface by studying the response of cat-
tle along the ration lines R 0,
Rx , ,
R 3 and R 4 with each of these feed
,
mixes being fed ad libitum. (Because of the high grain intake implied
by R4 it may be necessary to approach R 4 indirectly. For instance, R 4
,
might be eventually attained by following the dotted feed path.) If ob-

servations of the response along each ration line are taken at equal time
intervals, say weekly or monthly, these observations might correspond
to the positions marked tij (i = 0, 1, 2, 3, 4; j = 0, 1, 2, 3, 4) in the feed
plane. The fact that the observations are not spaced equally along each
ration line indicates that the feed consumption varies from one time
period to another. Each observed point tij specifies, through its forage-
grain co-ordinates, a treatment in terms of the cumulative quantity of
feed fed up to the time of the observation. As a consequence of free
feeding, these treatments are only definable ex post Before the event,
.
the researcher could not specify their locations in the feed plane. This
situation may now be contrasted with that under controlled feeding con-
ditions, assuming handfeeding is the normal mode of production.
With controlled feeding, the researcher decides how much of a given
feed mix is to be made available to the animals in a given time period.
Thus, within the specified time periods, it might be decided to feed
quantities determined by the intersection of ration lines and a contour
defining, for example, total feed poundage or energy intake per animal.
With such a system, the observational points or treatments t£ 0 t^,
,
. .
.
ti 4 of Figure 7.5 could be purposively spaced over the i-th ration line.
An approach of this nature would, of course, have to recognize any re-
straints imposed by the animals’ stomach capacity or physiology.
The choice between free and controlled feeding procedures depends
in large part on the type of output being studied. If the livestock prod-
uct is a flow (such as milk, eggs, or power) which is harvested in a
semicontinuous manner over time, then controlled feeding must be
, ,

strongly recommended. As an important example, consider the case of
milk production where daily collection of the product is the general
35
rule. If dairy cows are free fed they will consume feed quantities that
are determined by their stomach capacities. Such quantities are de-

picted by the points t 1? t 2 t 6 of Figure 7.6.
,
. .
.
They lie on the stom-
ach line SL. With such an experimental design the observed treatments
will not allow estimation of the milk production surface. The treatments
would be the same, day by day. Apart from variations over the lactation
period and uncontrolled influences, response would only relate to the
surface points lying above the points t L t 2 .
t6 ,
The only relation-
,
. . .
ships these observations can reveal are (a) the location in the feed plane
of the stomach line and (b) maximum milk obtainable from each of the
six treatments. The treatments would give no indication of marginal
rates of substitution between feed components, nor of marginal yields
either along a particular ration line or for a particular type of feed.
In short, the experiment of Figure 7.6 does not allow estimation of the
production surface. It merely locates a line across the surface. For
estimation of the whole surface relevant to either a single day’s produc-
tion or production over the entire lactation period, it would be necessary
to have observations scattered purposively throughout the area lying
below the stomach line SL of Figure 7.6.
When the livestock product is a stock resource, such as meat har-
vested upon slaughter of the animal, the analytical limitations imposed
Figure 7.6. Schematic representation in the feed

plane of the effect of free feeding in
a milk production experiment
35
See Heady, E. O. Problems m
designing dairy feeding experiments for economic anal-
ysis, m
Hoglund, C. R. et al. Nutritional and economic aspects of feed utilization by dairy
cows. Iowa State University Press, Ames. 1959. Pp. 193-205. Also see Orton, F. J. The
economy of feed input in milk production. The Farm Economist, 9: 11-25.
by free feeding to stomach capacity are not very severe. They tend to
be cancelled out by the longer term nature of the production process.
In such instances, the observations required to estimate marginal pro-
ductivities and marginal rates of substitution relate not so much to
daily intake and daily production but to the cumulative quantity of inputs
utilized and output produced at each stage of the production process.
These major stages in the production process may be periods of weeks
or months, depending on the type of animal. In general, for meat pro-
duction under handfeeding, the farmer’s main problem is to know the
least- cost ration at different weights of the animal. He may then feed
at over- all least- cost to the final weight expected to give the highest
market grade and price. Such a procedure is necessary because, in
contrast to crop production where the optimal fertilizer input can be
applied in one dose, the meat producer has to supply inputs continuously
over time to a production unit whose capabilities change over the pro-
duction period. For instance, changes occur in an animal’s nutritional
requirements between the growth and fattening stages of the life cycle.
As a consequence, the marginal rates of substitution between various
feeds are not constant over the production period (a fact which should
be taken into account in deciding on the appropriate algebraic form of
the production function). Ideally, a least- cost ration should be deter-
mined for each weight to marketing. For practical reasons, however,
the farmer could not be expected to change rations more than a few
times, say once for each major stage in the animals progression to
market weight. The researcher, therefore, needs to ascertain rations
which average least- cost over selected weight intervals, preferably
those commonly used by livestock producers. For instance, with hand-
fed hogs, separate production functions might be fitted for the weight
intervals: weaning to 75 pounds, 75 to 150 pounds, and 150 to 250 pounds.
Rations averaging least- cost over each of these production stages could
then be determined. Free feeding experimental procedures would be
appropriate for such analyses.
If the date of marketing was an important factor, account would have
to be taken of time in the production function, and hence in the experi-

mental design. For instance, a farmer may purchase or breed animals
over a range of months, but desire to market them all in a single month.
For such situations, experiments would need to be designed with feeding
along each ration line controlled at various levels per unit of time. The
treatments would be the cumulative weight of each feed mix constituent
and the cumulative time from the start of the experiment up to the time
of each observation. One set of free feeding treatments would need to
be included for completeness. Given a certain amount of time available
for attaining the desired market weight, the economically optimum ra-
tion for reaching this weight within the available time could be ascer-
tained. As before, recognition would have to be given to restrictions
imposed by stomach capacity and physiological factors.
Similar experimental arrangements would be most appropriate for
production function studies of semiflow products produced under
handfeeding conditions. Consider the case of wool which is customarily
harvested on an annual basis. With only ad libitum feeding treatments,
no account could be taken of variations in the extent to which sheep chan-
nel various feed types into wool production versus fat production at al-
ternative levels of feed availability. Such considerations are important
because optimal handfeeding under certain price regimes might be at a
low level of daily feed intake relative to what the animal would consume
if free fed.
Other considerations may also necessitate the use of controlled feed-
ing in experiments for semiflow or stock products. Such would be the
case were it desired to examine the relationships over time between the
various elements of an animaPs growth and maintenance processes. For
instance, the proportion of the feed mix going to bone, fat, and meat pro-
duction may vary with the constitution and quantity of the ration received
36
per unit of time. Controlled feeding would be necessary to determine
such relationships and their associated marginal quantities. It is also
pertinent to note that such experiments would tend to be extremely de-
manding of research resources. Unless relatively accurate means were
available to measure an animal’s constituents, without its slaughter,
each animal would need to be killed as it reached the desired treatment.
Moreover, as noted in Chapter 4, the most appropriate model by which
to analyze an animal’s growth and maintenance interrelationships may
be one using simultaneous equations, also an expensive procedure.
Nonetheless, all of these considerations are important; both from the
standpoint of basic science in predicting the physiological relationships
in livestock nutrition, and for the making of economic decisions.
For flow type products harvested semicontinuously, experiments
with some constituents of the ration fed ad libitum and others controlled
may suffer from the disadvantage outlined in relation to Figure 7.6.
Free feeding will lead to treatment locations along the stomach line. A
mixture of free and controlled feed components in an experiment relating
to a nonflow product may also tend to be disadvantageous. This is the
case where animals gorge themselves on ad libitum feed components and
other treatments are not scattered satisfactorily across the feed plane.
Production Functions for Grazing
So far, in discussing livestock problems we have only considered

production relations for full handfeeding. But, in actuality, few large-
bodied animals are handfed all the year around. This is especially true
for the tropical and temperate regions of the globe. The main types of
livestock in these regions, cattle and sheep, obtain the bulk of their feed
by grazing. And even where climatic conditions necessitate full hand-
feeding in the fall and winter, grazing is important in the spring and
summer.
38
See Heady, E. O. Economics of agricultural production and resource use. Prentice-
For the production economist, the essential features of pasture are

thatit is produced (a) at a low cost, (b) in a marked seasonal pattern,
and (c) in amounts that are largely beyond the farmer’s control in the
short run of a single climatic season. In other words, given some
prior fertilizer and utilization program, the quantity of pasture feed at
a farmer’s disposal is largely fixed by climatic factors beyond his con-
trol. But, relative to this rather fixed bulk of pasture, there is one im-
portant livestock production input that is under the farmer’s control.
To wit: the animals themselves. It is the intensity of stocking, set by
the farmer, that determines the grazing input and the output per animal
from a given bulk of pasture. 37 The production function relevant to a
farmer’s utilization of pasture is therefore broadly of the type depicted
in equation 7.37 where, assuming homogeneity within each resource
category, Y is the livestock product (e.g., meat, milk or wool); P is the
amount of pasture; A is the number of animals; and S is the quantity of
supplementary feed used by the animals. For convenience, we first as-
sume equation 7.37 relates to the period from start to finish of a single
cycle of the animal in pasture.
(7.37) Y = f(P, A, S)
As required empirically, each of the input categories of equation 7.37

might be expanded into a number of variables. The quantities of differ-
ent types of pasture and other grazing crops could be considered as dis-
tinct inputs; likewise, different types of animals producing the same
product could be segregated. As an example, ewes and wethers could
be inserted separately in a wool production function. Again, the function
might be expanded to include other inputs; for instance, fertilizer and
irrigation water entering the production process via the pasture. Alter-
natively, the production function for pasture might be substituted in
equation 7.37 for the variable P, production of pasture being an inter-
mediate step in the grazing production process. Such a function might
be of the general form depicted in equation 7.38 where A and S are as
in equation 7.37, L is the area of pasture land, F is the quantity of fer-
tilizer applied to the land area, W
is the quantity of irrigation water
used, and C is an index of the favorability of the climate to pasture
growth during the pasture cycle.
(7.38) Y = f (A, S, L, F, W, C)
Again, if desired and feasible, a function such as equation 7.38 might be

further e>q?anded. Ignoring such possibilities, suppose equation 7.38 is
to be fitted. The optimal resource allocation for any particular value
of the climatic factor could then be decided by marginal analysis, in the
usual way and with all the usual provisos. Similarly, for any fixed area
37
See Candler, W. V. Wool and wethers. Farm Management Report No. 1. University
of New England, Armidale. June, 1958.
of land, the optimum number of animals, and hence the best stocking
rate, could be decided. such calculations the cost of animals must
In
be taken inclusive of their “running” costs involving labor, fencing, and
other services that may have to be varied with the stocking rate but are
not included in equation 7.38 as direct Inputs.
Assuming a satisfactory index of climatic favorability to be avail-
able, an experiment to estimate equation 7.38 would have to be conducted
over a sufficient number of years to give a satisfactory range of obser-
vations on the climatic factor. Alternatively, equation 7.38 might be
estimated from an experiment conducted in a year of “average” cli-
mate. The fitted function would then correspond to equation 7.39.
(7.39) Y = f (A, S, L, F, W, |
C)
Normal procedures for response surface designs, as outlined in

Chapter 5,should be used for the estimation of grazing input- output
functions like equation 7.39. Since the inputs involved are relatively
expensive, the rotatable designs are to be strongly recommended, pro-
vided that the researcher has some idea of the location of the relevant
section of the surface'. If he does not, a fractional factorial design may
be best. Should the researcher be prepared to assume that constant re-
turns to scale prevail to the controllable factors, the experiment may
be simplified, and cheapened, by examining only a single level of the
land input. The production surface, inclusive of land as an input, could
then be estimated from the constant correspondence between output and
proportionate increases in all inputs (including land) as implied by con-
stant returns to scale. Thus, suppose A, S, F, and of equation 7.39 W
were observed at various levels with L fixed at some level g. Under
constant returns to A, S, L, F, and W, the response at any level of L,
say kg, in combination with any observed treatment (A, S, F, W) could
be obtained by multiplying the response for the treatment (A, S, F, W)
by the factor k. Equation 7.39 might then be fitted to a set of data on
Y, A, S, L, F, and W
even though L had been held to a single level in the
experiment. Alternatively, equation 7.40 might be fitted and equation
7.39 derived by making use of Euler’s Theorem as in equation 7.41.
(7.40) Y = f (A, S, F, W, | L, C)
(741)J
«IA JJS 5YL 6YF 8IW
K
5A Y 5S Y 5LY 6FY + 5WY
Functions such as equation 7.37, 7.38, and 7.39 are, however, only
broadly applicable to farmers’ decisions. They have a number of dis-
advantages. First, they do not take specific account of seasonal varia-
tions in the quality, quantity, and composition of a pasture. Such varia-
tions are subsumed within the variable area of pasture. In doing so, the
possibility of conserving feed from one stage of the pasture cycle for
use at a later stage is ignored. Of course, this criticism could not be
made where there was no marked cycle in pasture growth

in situations
and composition. The second difficulty with such functions is that their
usefulness depends, in large measure, on the extent to which a farmer
can vary his stocking rate. On an annual basis, this criticism might
not be severe. But it could negate the usefulness of grazing functions
that related to each seasonal stage of the pasture cycle; especially if
the farmer’s livestock breeding program was an important element in
his farm operation. The third criticism of output- stocking rate func-
tions is that they make no allowance for risk or uncertainty. Due to the
strong dependence of pasture growth on climate, grazing tends to be a
very risk- prone operation. As the stocking rate rises, the associated
risk increases more than proportionately since drought is not only a
function of rainfall but also of stocking rate. The conservation of some
pasture growth to serve both as an inter-year drought reserve and as
an intra-year supplement therefore becomes an important considera-
38
tion. As well, so far as the farmer is concerned, an important ele-
ment of the situation might be the pasture management system to be
39
adopted and its relation to the cropping rotation. To incorporate such
factors in the production function analysis would require extremely ex-
tensive experiments. Indeed, the most suitable approach might be
through the use of real-world data.
The above criticisms basically relate to the fact that such functions
as equation 7.39 do not provide a full basis for decisions during the pe-
riod of pasture availability. Of the resources included, land and ani-
mals will generally be fixed in the short run, climate is not controllable,
and fertilizer applications will usually have been made in the preceding
period. Only the irrigation water and supplementary feed inputs will
be currently at the farmer’s discretion, and in most instances both
these inputs are not used. In contrast, under handfeeding conditions,
the farmer can exercise day to day control over the inputs used.
These criticisms lead to the conclusion that the most appropriate
approach to the economics of pasture may be a (dynamic) programming
40
one. The mechanics of such procedures do not concern us here. Still,
38
For a conceptual analysis of the interrelationships between stocking intensities, fodder
conservation, and risk see Lloyd, A. G. Fodder conservation m the southern tablelands wool
industry. Review of Marketing and Agricultural Economics, 27: 5-50. Insofar as move-
ments in the pasture cycle and drouth occurrences may be characterized in probability
terms, the techniques of inventory analysis are also pertinent to the economics of pasture
use. See Mauldon, R. G. and Dillon, J. L. Droughts, fodder reserves and stocking rates.
Australian Jour, of Agr. Econ., 3(2): 45-57.
39
See Olson, R. O. and Heady, E. O. Economic use of forages in livestock production on
Com Belt farms. Circular No. 905. USDA, Washington, D.C. 1952; Blaser, R. E. et al.
Animal performance and yields with methods of utilizing pasturage. Agronomy Jour., 51:
238-42; also Willoughby, W. M. Limitations to animal production imposed by seasonal fluc-
tuations in pasture and by management procedures. Australian Jour. Agr. Res., 10: 248-88.
40
Basic programming techniques, as related to grazing and storage for later use, have
been outlined by Heady, E. O. Economic concepts m directing and designing research for
programming of range resources. Jour. Farm Econ., 38: 1604-16. See also Candler, W. V.
A new look at budgeting from the viewpoint of linear programming. Australian Jour. Agr.
Econ., 3: 46-56; Nelson, M. The production function and linear programming in valuation of
intermediate products. Land Econ., 33: 257-61; Nelson, M. and Castle, E. Profitable use of
fertilizer on native meadows. Jour. Range Management, 11: 80-3.
itshould be noted that estimates of output- stocking rate functions on a
seasonal basis would be extremely useful in any programming approach.
Too, estimates of such functions are relevant (a) in choosing between
grazing and the feeding of harvested forage where the use of grazing is
not obviously advantageous and (b) as a guide to public planners in land
settlement and the allocation of grazing rights.
A noteworthy feature of the approach of equation 7.37 is that it com-
pletely avoids the problem of the amount of pasture actually consumed
by an animal. This question is well-nigh irrelevant to the “on the spot”
grazing decisions which a farmer must make. Still, specific knowledge
of pasture intake is important; for instance, in the consideration of such
problems as the physiological effects of different sward types, the effi-
ciency with which different types of animals utilize pasture, and the ef-
fects of the grazing animal upon pasture growth and composition. Most
such questions are best answered in terms of their relevant response
surface since they will rarely depend on a single direct influence. In the
past, response surface experiments involving the measurement of graz-
ing intake have not been feasible due to the lack of any reliable measure-
ment technique. Today, however, a sufficient backlog of pasture intake
research is available to indicate more reliable measurement procedures
41
that may be followed.
RECOMMENDATIONS FROM FITTED

PRODUCTION FUNCTIONS
The estimation of production functions is not an end in itself. They

simply provide the framework through which input- output relations may
be evaluated in economic terms. Moreover, production phenomena are
essentially physical, although they may include some nonphysical ele-
ments incapable of direct quantification. From an economics standpoint,
the usual criterion of evaluation is the maximization of net revenue from
the production process, given the relevant price regime associated with
the inputs and outputs. Too, other criteria may sometimes be used by
the economist: for instance, the highest monetary return per unit of a
42
given input. Indeed, any production process which is consummated in
the market place is most rationally controlled by economic criteria
covering both the monetary and other goals of the entrepreneur. Of
course, due to a lack of information about the nonmonetary goals, the
economist is generally forced to work in monetary terms only. Physical
criteria, such as the maximization of physical production, are in gen-
eral quite irrelevant; except insofar as (a) it is always better to produce
41
J. T. Pasture evaluation in relation to efficiency, in Hoglund, C. R. et al.
See Reid,
(eds.) Nutritional and economic aspects of feed utilization by dairy cows. Iowa State Uni-
versity Press, Ames. 1959. Pp. 135-51. Also Reid, J. T. et al. Symposium on forage
evaluation.Agronomy Jour., 51: 212-38.
42
See Pesek, J. and Heady, E. 'O. Derivation and application of a method for determining
minimum necessary rates of fertilization. Soil Science Society of America Proc., 22: 419-
23.
a greater amount of product from a given quantity of input, other factors

remaining unchanged, and (b) it is never rational to add inputs when out-
put increments are negative.
The use of estimated production functions as a guide to the eco-
nomic allocation of resources is fraught with problems. The conditions
under which a fitted function will serve as an error-free guide are ex-
tremely severe. To be a perfect guide, the estimated function should
(a) account for all inputs involved in the production process, (b) cover
the relevant range of input and output quantities, and (c) specify the
production parameters without error. As well, it is necessary (d) that
the function be supplemented by similar information about other produc-
tion alternatives, (e) that all the relevant prices be known, and (f) that
the levels of the inputs not under the entrepreneur’s control he known
ex ante In addition, even if the fitted function is accurate, it cannot
.
serve as a reliable indicator outside of the statistical population to

which it refers.
Obviously, no estimated production function can satisfy all the above
conditions. Indeed, this chapter and the three preceding ones have been
concerned with methods of accounting for as many of the relevant inputs
as possible, and of reducing errors in the estimated parameters.
Bluntly, it can be said that no estimated function will ever contain all
the relevant inputs or contain error-free estimates of the parameters.
Still it is not necessary for such perfection. With reasonable precau-
tions in the collection and analysis of data, complemented by discretion
in interpretation, production functions that can serve as reasonable
guides in resource allocation may be estimated. This is so even if no
information is available about alternative production possibilities; some
information of reasonable validity is better than none. But it is cer-
tainly better if production functions are available for alternative lines
of investment. If resources are limited they may then be allocated
among these alternatives in profit maximizing fashion 43 The last two
.
conditions for perfection noted above, ex ante knowledge of future prices

and uncontrollable input levels, also are unlikely to be satisfied. In
other words, the economic interpretation of a fitted production function
will always be shrouded with elements of risk (if objective probability
statements can be made about the unknown factors) or of uncertainty
(if probability statements cannot be made). It is possible, however, to
handle risk and uncertainty in an economically rational fashion.
Another frictional feature in the evaluation of fitted production func-
tions is the time factor. No real-world production process is instanta-
neous. Some processes will have a natural time length. For others,
the length of the production period may be at the discretion of the en-
trepreneur; if so, time may be included in the function as an input fac-
tor. In other cases, where production is relatively continuous over
time, the estimated function may refer to some arbitrarily selected
43
See Henderson, J. M. and Quandt, R. E. Micro- economic theory. McGraw-Hill, New
York. 1958. Pp. 72-73. Also Heady, E. O. Economics of agricultural production and re-
source use. Prentice-Hall, Inc., New York. 1952. Pp. 237-75.
time span; for example, an accounting period. Whatever time period
the fitted function refers to,it should take account of all inputs relevant
to production during that time period. Thus, some of the inputs may-
have entered the production process during a prior time period. Still,
however time is handled, production functions as estimated empirically
are essentially static 44 Their use as guides for resource allocation
.
can refer only to the length of the production period covered by the
function. The resultant resource allocation for profit maximization
over this period may be quite different from the allocative pattern
necessary to maximize profit over a sequence of such periods. More-
over, as previously noted, the production function for a given process
is not invariant over time. It will tend to be affected by technological
change, or more accurately, it will tend to be replaced by a new pro-
duction function.
In providing production function estimates as guides to production
decisions, and in making recommendations based on these estimates,
the economist or farm adviser assumes that the real-world decision
makers desire to maximize profits. This assumption is reasonable
even though, in the absence of production function information, entre-
preneurs and other planners may follow some simpler motive in their
resource allocation. For instance, they may merely search for a pat-
tern of resource use that is satisfactory in terms of their current as-
45
pirations; behaving as satisficers rather than maximizers Given .
precise recommendations or decision-making guides derived from

input-output analyses, however, these real-world decision makers
might be expected to shift their aspirations to a level commensurate
with profit maximization.
With these broad comments in mind, we may now consider the prob-
lems that arise in using farm-firm functions and micro-functions as
guides for real-world decisions. Due to the mechanics of data collec-
tion, such a division is tantamount to a distinction between functions
based on survey data and those from experimental data. To a major
degree, the possibility of making useful recommendations depends upon
the scope of the fitted function, its reliability in terms of statistical and
economic logic, and the quality of the data on which it is based. Some
consideration has already been given to these topics in this and pre-
ceding chapters. Consequently, these aspects will not be treated in
detail here.
Farm- Firm Functions
Typically, farm-firm production functions are estimated from

cross-sectional survey data. Estimates so derived are not of extreme
usefulness as guides to decision making on individual farms unless the
44
See Schneider, E. Pricing and equilibrium. William Hodge and Co., London. 1952.
Pp. 182-88.
45
See Simon, H. A. Models of man. John Wiley and Sons, New York. 1957. Pp. 241-60.
functions are relatively disaggregated, both in inputs and outputs. A

farmer can be told that he should use more * capital.” But the recom-
mendation has little specific meaning where he has dozens of avenues
for the use of additional capital, some of which are profitable and others
which are not. Likewise if “output” corresponds to a variety of prod-
ucts, the farmer needs specific information about each alternative
rather than some broad recommendation in aggregate. Still, a measure
of compromise is possible. He might be advised to allocate additional
increments of a broad resource category between specific alternatives,
in proportions similar to those found on the “average farm” of the sam-
ple. Even so, the fitted function must be reliable in terms of the data
used and the homogeneity of the production process. A function fitted
to bad data can never serve any useful purpose, although it must be
recognized that different research workers may, in good faith, have
different conclusions about the efficiency of the data. Still, even if
there are slight variations among farms in the “true” production func-
46
tion, an average production function of some utility may be calculated.
Other factors also restrict the usefulness of farm-firm functions as
guides for individual firms. Satisfactory recommendations may be fea-
sible only for individual farms of an average character. The possible
errors associated with recommendations for farms of a sample using
atypical input mixes may be too large to warrant recommendations to
them. Account also needs to be taken of climatic uncertainty. The re-
source productivities derived from a cross-sectional farm sample will,
in part, depend on the climatic conditions prevailing in the survey year.
If this year was extremely atypical, the usefulness of the fitted produc-
tion function is limited. Many farm resources are relatively fixed in

the short run of a few years. The most obvious example is land. To
tell a farmer that under current price conditions he should be using
another 100 acres of land, when the land may not be available or he may
not be able to afford it, is quite meaningless. Such restraints are, of
course, best handled by programming procedures. Production function
estimates may contribute to these by providing estimates of the input-
output coefficients at varying levels of production. Noneconomic con-
straints may also be important. For instance, a farmer may restrict
his input of hired labor because of the psychological burden involved in
directing additional labor. Another disadvantage of empirical produc-
tion functions is that they usually fail to take account of discontinuities
that may exist on the input side. 47 It is, for instance, impossible to
vary the input of machinery services in the continuous manner possible
for fertilizer. Still, machinery services may be hired on a contract
basis. Or a three-furrow plow might be used as a compromise between
a two-furrow plow and two such plows; alternatively, the two-furrow
46
See Konijn, H. S. Estimation of an average production function from surveys. Econ.
Record, 35: 1X8-25.
47 Functional forms
allowing for discontinuities are possible. See Hildreth, C. G. Dis-
crete models with qualitative restrictions, in Baum, E. L. et al. (eds.) Methodological pro-
cedures in the economic analysis of fertilizer use data. Iowa State University Press, Ames,
1956. Pp. 62-74.
plow might be used one and a half times as much. Likewise, while half
a man cannot be hired, a man might be hired half-time, or a less pro-
ductive man employed full-time.
Lack of ex ante knowledge about future prices also restrains use of
farm-firm functions from cross-sectional data. However, it is more of
a marginal disadvantage than any of the criticisms noted above. Price
risk or uncertainty will always be present regardless of the decision
guide used. Programming procedures while eminently suited to indi-
vidual farm planning, must also face up to the problem of uncertain or
unknown future prices.
On all these grounds, cross-sectional production functions must be
used warily as guides to individual farmers. In contrast, they may
serve more usefully as a guide to problems or policies in broad alloca-
tion of resources among regions or between industries, even when based
on such aggregates as land, labor, and capital 48 Although land is im-
.
mobile, the switching of other inputs such as capital and labor from one
region or industry to another is possible. Thus, on the basis of current
marginal returns to capital, derived from production functions in two
regions A and B, a bank or a government might decide that credit be
restricted in region A and increased in B. For capital, such decisions
are even feasible in an international context. With labor, however,
transfers cannot be made so easily in the short run unless government
planning is all supreme and perhaps if little consideration is given to
the personal values of the transferred workers.
For production functions to be used as policy guides, discretion
must be used in selecting the sample population. If the population is
defined with ideal strictness, a cross-sectional sample would never be
possible. Variations among individual farms are so many that only in-
dividual farms can represent a homogeneous population. Compromise
thus has to be made between the degree of homogeneity in the sampled
population and the range of applicability of the resultant function. Cer-
tainly, given limited research resources, it is desirable that the fitted
49
function be applicable to as large a population as possible .
48 For example,
see Heady, E. O. and Du Toit, S. Marginal resource productivity for
agriculture in selected areas of South Africa and the United States. Jour. Pol. Econ., 62:
494-505; Dillon, J. L. Marginal productivities of resources in two farm areas of N.S.W.
Economic Society of Australia and New Zealand. Economic Monograph No. 168. Sydney.
1956; Wang, Y. Resource returns and productivity coefficients for selected crop systems in
Taiwan area. Proc. of Agr. Econ. Seminars, Sept. 16-20, 1958. National Taiwan University,
Taipei. 1959. Pp. 90-98; Heady, Earl O. Marginal productivity of resources and imputa-
tion of shares for rental farms. Iowa Agr. Exp. Sta. Bui. 433. Ames; and Miller, W. G.
Comparative efficiency of farm tenure classes in the combination of resources. Agr. Econ.
Res., 11: 6-16.
49 Although farm-firm
functions are rarely estimated from time series data, some com-
ments might be made on the use of such functions as guides to resource allocation. If the
data refers only to a single farm, its application relative to other farms will be unwarranted
unless these other farms are rather similar to the farm studied. This factor will also de-
termine the usefulness of the time series function as a help to interregional resource allo-
cation. Should there have been significant technological changes over the time series pe-
riod, of course, the fitted function will be of little use on either an inter- or intra-farm
basis, or interregionally, unless account is taken of these changes.
Inference From Micro- Analyses
We now outline the difficulties in using production function esti-

mates for individual technologies as guides to real-world decisions. By
individual technologies we refer to such phenomena as crop, animal,
and machinery input- output- relationships. Relative to the over-all farm
operation, they are of a micro- nature In contrast to farm-firm func-
.
tions which cover a host of micro- production processes within a single

aggregative framework, the micro-functions are far more useful as
guides for individual farmers. On the other hand, their potential rele-
vance as guides to public planners is not so direct. Still, the foundation
50
of any macro-policy must be in terms of micro-phenomena For in- .
stance, economic growth, in the final analysis, must be based on the

development of new micro-production processes and the better utiliza-
tion of existing processes. Likewise, the use made of micro-processes
determines the aggregative supply response. Understanding of supply
influences cannot be divorced from micro- phenomena. Insofar, there-
fore, as micro-functions serve as guides to more efficient resource al-
location and hence to welfare gains for the whole of society, they im-
pinge upon public planning. Moreover, micro- and macro- analyses of
production are complementary: without some knowledge of both, plan-
ning may be inefficient. Micro- recommendations may be valid for in-
dividual farmers only as long as masses of farmers do not follow them.
Finally, the same basic economic principles apply in analyses relating
to micro- and macro- functions and problems. The principles for spec-
ifying the optimum resource allocation are essentially the same within
a firm, between firms, within industries and regions, and between in-
dustries and regions.
A production function based on physical phenomena cannot take di-
rect account of a farmer’s nonmaterial goals. Moreover, these farm
family goals will vary from farm to farm so that each decision relates
unique features. Given knowledge of these nonmonetary aspects, cor-
rect recommendation about use of individual technologies might involve
resource allocations differing greatly from that implied by profit max-
imization. It is for this reason that recommendations to individual
farmers based on production function analyses should be couched in
terms of “do this if you wish to maximize your profits” rather than in
dogmatic statements of the type “you should do this.”
Also, since the farmer is not an expert in the differential calculus,
production function analyses need to be interpreted for him. The re-
searcher’s job is not finished when he has estimated a particular pro-
duction function. He has a responsibility to interpret the analysis in a
80
For an elaboration of the role of microanalysis in macro-agricultural policy see:
Johnson, G. L. Some contributions of microanalysis to agricultural policy, in Baum, E. L.
et al. (eds.) Economic and technical analysis of fertilizer innovations and resource use.
Iowa State University Press, Ames. 1957. Pp. 362-74j and Heady, E. O. Need for produc-
tion economics research in solving policy problems, in Baum, E. L. et al, (eds.) Economic
and technical analysis of fertilizer innovations and resource use. Iowa State University
Press, Ames. 1957. Pp. 348-61.
260 PROBLE MS RELATING TO ESTIMATION
farmer or his go-between can understand and to ensure the

way the
farmer. Otherwise the research
channeling of such information to the
To end, such communication devices as price
input may be wasted. this
51
maps, nomograms, and costulators might be used .
of micro-
Relative to farm-firm analysis, the greater usefulness
functions as farmer guides stems from the fact that (a) they are usually
determined under experimental conditions where the researcher can ex-
is far easier to
ercise some control and (b) being of a micro- nature, it
and outputs in a meaningful way. Still, the range of
specify the inputs
application of these micro- estimates is always limited by the experi-
mental framework used. For instance, a fertilizer- crop production
function estimated from experiments on a given soil type cannot be used
indiscriminately for other soil types. Likewise, if the level of technical
management or other background factors relevant to the response ex-
periment are too different from those prevailing on farms, the useful-
52
ness of the fitted function will be curtailed Indeed, the extent to which
.
a response function for a given technology in a particular environment

can be used with reference to the same technology in a different envi-
ronment is perhaps the major problem in planning and using production
function research. It is obvious that research resources are too lim-
ited and technological advance too rapid for production function esti-
mates to be derived for all possible production processes in all possible
environments. At the same time, many real-world environments are
not so dissimilar as to necessitate a separate function for each of them.
Moreover, no matter how bad the error may be in extrapolating a pro-
duction function from one environment to another, such a procedure at
least has the salutary effect of forcing those concerned to reconsider
their positions. The extent to which such extrapolations can be reason-
ably made, however, is largely a matter of discretion on the part of
those familiar with both the experimental and real-world environments.
In this regard, one avenue of worthwhile exploration is the development
of adjustment factors. These could be used to adapt results for a par-
53
ticular environment to other environments of real-world significance.
Even within the same basic environment (i.e., soil type, breed of
animal, etc.), there are important problems of inference from the ex-
perimentally derived function to the real-world situation. These relate,
first, to the levels at which the controlled factors were held in conduct-
ing the experiment; and, secondly, to the influence of the uncontrolled
81
Such devices are illustrated in Profitable use of fertilizer in the Midwest. Wisconsin
Agr. Exp. Sta. Bui. 508. Madison. 1954; and Heady, E. O. Integration of physical sciences
and agricultural economics. Canadian Jour. Agr. Econ., 4: 1-15.
52 Such an effect appears to have occurred in some milk production experiments. See
Antill, A. G. and Clark, C. The overfeeding of dairy cows. Westminster Bank Review.
Feb. 1958. Pp. 13-15.
83
A start in this direction has been made with respect to initial soil fertility conditions
in crop and pasture production by Jensen, D. and Pesek, J. Generalization of yield equa-
tions in two or more variables. I. Theoretical considerations, n. Application to yield
data. Agronomy Jour., 51: 255-63. Also Rust, R. H. and Odell, R. T. Methods used in
evaluating the productivity of some Illinois soils. Proceedings of the Soil Science Society of
America, 21: 171-75.
factors operating in the experiment from which the function was esti-
mated. 54 Distinction might also be made between problems of inference
arising (a) from inaccurate statistical estimation due to the omission of
relevant variables or the inclusion of incorrectly measured variables
and (b) from inaccurate economic interpretation due to the omission of
input factors which have interaction effects with the included variables.
Difficulties from all these sources can be discussed by way of equation
7,42 where X x to X g are the controlled factors that were allowed to
vary; X g + x to X^ are the controlled factors held fixed at some level;
and Xh+i to X^ denote uncontrollable factors that operated at some
level (either fixed or variable) over the experiment. As noted in Chap-
ters 5 and 6, all of the unstudied factors X g + x to X^ must be expected to
condition the estimated function. For this reason, g should be as large
as possible and h - g as small as possible.
(7.42) Y = f(X 1? X 2, ..., Xg , I X g+l ,

Xh , II X h+l
,
. . .
Xk )
Of course, as estimated empirically, equation 7.42 would only be written

in terms of X t to X the levels of the other variables being noted in the
,
g
commentary on the experiment. We will first consider problems of in-
ference relating to the controlled factors.
Influence of controlled factors
The levels at which the variable factors X x to X g are allowed to

operate raise no difficulties so long as the use of the fitted function is
restricted to the surface section studied experimentally. The estimated
function cannot be validly extrapolated outside this region. Recommen-
dations involving levels of X 1 to X g beyond the studied range must be
advanced with extreme caution. This difficulty is best overcome by en-
deavoring to study as much as possible of the surface that may be of
economic significance, given the quantity of research resources avail-
able.
Perhaps more important are the problems associated with the fac-
tors Xg+i to Xh of equation 7.42. If these controllable factors are in
fact held fixed over the experiment, the fitted function strictly applies
only to real-world situations where these factors operate at the same
levels. For instance, in a fertilizer- crop response study, X g+1 to X h
might refer to the slope of the experimental plots, the degree of weed
infestation, the level of management, the type of cultivation used, and
so on. If the experimental level of these factors does not correspond
to the farm situation, the fitted function would, in all strictness, not be
relevant. The researcher has two alternatives in overcoming this prob-
lem, assuming that he has included as many as possible of the controlled
factors as variables. The first alternative is to repeat the experiment
54
See Swanson, E. R. Problems of applying experimental results to commercial prac-
tice. Jour. Farm Econ., 39: 382-89.
either in time or space with the fixed controllable factors held at a dif-
ferent level in each experiment. Such a procedure is, of course, equiv-
alent to conducting one large experiment with the “farm” factors (slope,
weed infestation, management, cultivation procedure, etc.) treated as
variables in the statistical analysis. In general, however, it would seem
to be a waste of research resources to carry out such a large experi-
ment or sequence of experiments, even if it were computationally fea-
sible. Research resources might be better used to look at a number of
production processes instead of examining a single one in minute detail.
The other alternative, advocated by Johnson, is to set up the experiment
with X 1 to Xg of equation 7.42 as the treatment combinations but with
X g+1 to X h varied throughout the experiment in the fashion prevailing
on farms, no attempt being made to incorporate these “farm* factors in
the statistical analysis. 55 The fitted function would then be an average
function with respect to the factors X
g+1 to Xh- Such an approach is
attractive insofar as these factors are not major determinants in the
production process. Moreover, as the array of estimated micro-
functions increases, experience can be gained on the degree of impor-
tance of these lesser influences. Rules of thumb might then be devel-
oped for making inferences from a fitted function to individual farm
situations.
Influence of uncontrollable factors
In any experimental situation and its farm counterpart, a host of un-

controllable factors will operate. Chief among these are climatic ef-
fects. The researcher, just like the farmer, has to accept these factors
as they are stipulated by Nature. For instance, suppose equation 7.42
relates to fertilizer- crop response. The experiment might be carried
out, by chance, in a climatically poor crop year. The factors
Xh+1 to
denoting such inputs as rain, heat, wind, etc., might all be at levels
inimical to crop production. Obviously, the fitted function would not
be
relevant to crop production in a good crop year. In situations where
the
uncontrollable factors were relatively stable, such problems would only
be of minor import. Still, most farmers have to face climatic and
other
risks and uncertainties. But if sufficient research resources
are avail-
able, variations in the uncontrollable factors can be
evaluated in such
fashion that reasonable recommendations are possible.
Thus, a
fertilizer- crop response function might be estimated for
a range of
climatic situations. For the single variable input case,
Figure 7.7 pro-
vides a simple illustration. Curves A, B, and C might
correspond to
the production function for excellent, average, and
poor climatic condi-
tions respectively. Alternatively, a single production
function might be
estimated with climate or other uncontrollable factors
included as
55
Johnson, G. L. Planning agronomic -economic research
in view of results to date in
Baum, E. L. et ah (eds.) Economic and technical analysis of
fertilizer innovations and re-
source use. Iowa State University Press, Ames.
1957. Pp. 217-25.
Figure 7.7. Influence on the production function

of the level of uncontrollable factors.
56
variables. How might reasonable recommendations be based on such
functions? The answer depends on whether the uncontrollable factors
are characterized by risk or by uncertainty. Risk implies that proba-
bilities can be associated with alternative levels of the uncontrolled
factors. If probabilities cannot be assigned, then uncertainty prevails.
Risk
Using the situation of Figure 7.7 as our example, suppose the cli-
57
mate can be specified in terms of risk. There might be, say, a proba-
bility of 0.6 of average climate, and of 0.3 and 0.1 for poor and excellent
climate respectively. Recommendations could then be made in two ways.
The farmer might be advised to operate on functions A, B, and C for 1,
6, and 3 years out of 10, respectively.
Alternatively, each year he
might follow B with 60 per cent of his resources, C with 30 per cent,
and A with 10 per cent. Such an approach would allow the farmer to be
partly right each year, although he could never be completely correct.
The “years out of 10” approach, on the other hand, while it would prob-
ably lead to the farmer being right some of the time, could lead to a
serious run of inefficient resource allocation patterns from an ex post
view. To the extent that the making of a whole range of recommendations
in probability terms may be too confusing for farmers, an average rec-
ommendation might be made with a rider attached giving the probability
58
of the average recommendation being profitable in a given year.
Such
“Such an approach is to be found in Parks, W. L. and Knetsch, J. L. Corn yields as

363-64. Also Parks,
influenced by nitrogen level and drouth intensity. Agronomy Jour., 51:
moisture
W. L. Methodological problems in agronomic research involving fertilizer and
variables, in Baum, E. L. et al. (eds.) Methodological procedures in the
economic analysis
of fertilizer use data. Iowa State University Press, Ames. 1956.
Pp. 113-33.
“See Heady, E. O. Economics of agricultural production and resource use. Prentice-
ss Such a procedure has been followed by Hildreth, H. J. Influence of rainfall on ferti-
an approach would be of most value in situations where a fair degree of

stability prevailed.
It should be noted that Figure 7,7 reflects a rather simple situation,
involving only three possibilities. In real-world situations, there will

generally be an infinite array of possibilities involving continuous prob-
ability distributions. Moreover, whether the probability distribution is
discrete or continuous, its higher moments, especially the variance and
skewness, may be extremely important to farmer decisions. For in-
stance, these higher moments are relevant to any consideration of the
possibility of extremely large losses or gains resulting from a given
course of action within the probabilistic environment. However, without
knowledge of the decision maker’s preferences about these higher mo-
59
ments of the distribution, their evaluation is impossible.
Uncertainty
uncertainty prevails with respect to the uncontrollable factors,

If
the appropriate approach would be that of a game against Nature. Thus,

even if no probabilities could be assigned to the alternative functions of
Figure 7.7, a payoff matrix might be calculated. Suppose equation 7.43
depicts such a matrix where Na Nb, and N c are Nature’s possible
,
strategies of excellent, average, and poor weather; and F a F b , and F c ,
are the farmer’s pure alternatives of following functions A, B, or C.

As well, the farmer may follow a mixture of these pure alternatives.
By definition, no probabilities can be assigned to Nature’s possible
strategies.
Na Nb Nc
Fa 7 2-1
(7.43) Fb 4 5 0
Fc 2 1 3
The numbers in this payoff matrix are hypothetical examples of the

farmer would make if he followed a given strategy when the
profit the
corresponding state of Nature occurred. Thus, if the farmer selected
function B and Nature “chose” Nc
poor weather, the farmer’s
,
i.e.,
profit would be zero. For any such payoff matrix
it is possible to de-
cide on courses of action that have strong logical appeal; making use of
decision criteria that have been developed for uncertainty situations of
the type depicted by Plan 7.43. The best known of such decision-making
lizer profits. Jour. Farm Eeon., 39: 522-24. See also Lloyd, A. G. Agricultural experi-
ments and their economic significance. Review of Marketing and Agricultural Economics,
26; 185-209.
59
In this regard, see Tintner, G. A contribution to the non-static theory of choice.
Quarterly Jour. Econ„, 56: 274-306; and Dorfman, R. et al. Linear programming and eco-
nomic analysis. McGraw-Hill, New York. 1958. P. 431,
algorithms are those of Wald, Laplace, Hurwicz, and Savage. No at-

tempt will be made to outline these decision criteria here. Their use
60
is rather marginal to the general topic of production function research,
We now pass, in succeeding chapters, to a consideration of actual
empirical production function research.
Raiffa, H. Games and de-

60
For a detailed outline of these criteria see Luce, R. D. and
cisions. John Wiley and Sons, New York. 1957. Pp. 278-86. For some examples of their
function context, see: Dillon, J. L. and Heady, E. O. Decision
application in a production
Australian Jour. Agr. Econ., 2: 113-20; and Swanson, E. R *prob-
criteria for innovation.
lems of applying experimental results to commercial practice. Jour. Farm Econ., 39: 382-
CHAPTER 8
EarS O. Heady
Damon V. Catron Pork Production Functions
Roger Woodworth
Gordon C. Ashton for Hogs Fed in Drylot
STUDY was
T HIS
economics and swine
conducted jointly by staff members
employs principles from both
nutrition. It
sciences to obtain more information than conventional research
procedures and to provide more complete data for economic analysis
in production
of protein levels for hogs. The investigation was designed particularly

to provide information allowing specification of optimum protein rations
for hogs of different weights. It was made in recognition of the fact that
the price of feeds which are high in protein and carbohydrates change
relative to each other. Hence, the ration which provides least-cost gains
or the most profit per hog is not the same under all price conditions.
Hence, experiments were designed which would allow estimation of the
pork production surface and the equations of isoquants, isoclines and
marginal transformations which relate to them. The chief objective
was to determine average rates of substitution over several weight in-
tervals in pork production. The weight interval substitution rates were
emphasized because farmers do not change rations each day. As a
matter of practicality, they buy or mix and feed the same ration over
an interval of time. Thus emphasis was on establishing the average
substitution rates between corn and a high protein feed over these weight
intervals; rather than to emphasize estimation of functions which
allow
derivation of substitution ratios at all levels of gain. The power function
provides a simple means for estimating average substitution rates over
gain intervals and is used accordingly. However, other functional
forms
also have been derived and are presented in this chapter for other
uses,
such as estimating optimum marketing weight or other purposes. A
few alternative forms of functions were estimated but are not
presented
(a) because of space limitations and (b) because
they provide estimates
which appeared no more or less efficient than those presented.
ECONOMY IN RATIONS
Economy in the hog ration depends especially on the manner in
which feeds are combined. Many feeds can be used in pork
production
and they can be combined in many ways. A number of
carbohydrate
feeds such as corn, barley, or wheat can be substituted
for each other
266
FUNCTIONS FOR HOGS 3N DRYLOT 267
and used in combination. On the other hand, vegetable proteins such as

soybean oilmeal, linseed oilmeal, and cottonseed oilmeal, and animal
proteins can also be substituted for each other or used in combination.
Finally, supplementary proteins and carbohydrate feeds can be substi-
tuted for each other and used in many combinations. One of the main
economic problems in hog rations is this: What combination of corn
and protein should be used in producing 100 pounds of pork? While
other carbohydrate feeds sometimes have a more favorable price and
supply situation, corn — aside from oats — has had the traditional ad-
vantage, both in price and proximity.
Aside from legume pasture, tankage was formerly the important
protein source because it supplied vitamin B 12 and other factors. The
absence of the then unknown vitamin in vegetable proteins limited
their use in pork production up to moderate hog weights. A main eco-
nomic problem then was deciding (a) the combination of corn and animal
proteins to use in producing lightweight hogs and (b) the combination of
corn, animal proteins, and vegetable proteins to use in producing heavier
hogs. (Minerals, pasture, other feeds, and labor and capital costs, of
course, also play a part in profit calculations.)
New technologies in production and recent developments in swine
nutrition, however, have caused older notions on hog rations to become
obsolete. The discovery that vegetable proteins, balanced or fortified
with trace minerals, vitamin B 12 and antibiotics, can combine with grain
as effectively as animal proteins in producing pork changed previous
protein level recommendations. Because of the relatively low cost of
small amounts of vitamin B 12 aureomycin and trace elements, a main
,
problem in the pork ration now appears to be how corn and a vegetable
protein such as soybean oilmeal — containing the other feed elements
mentioned — can combine with or substitute for each other to give the
least-cost ration.1 Animal proteins are, of course, still an economic
consideration in the hog ration.
DESIGN OF EXPERIMENT AND FEEDING METHODS
Data from three hog feeding trials were used in this study. The
first protein levelexperiment — no. 506 — was already completed when
the decision was made to analyze the economy of different rations.
This experiment included four protein levels of 14, 16, 18, and 20 per
cent. These percentages of protein were fed to pigs weighing from 34
to 75 pounds. At weights of 75 pounds, and again at 150 pounds, the
protein levels were lowered three percentage points. Under each of the
protein levels there were two replicates containing aureomycin and two
without. Each replicate had eight pigs per lot. This experiment was
not completely suited for the substitution analysis since there were only
four protein levels. Also, the shifting of protein levels at different
*For purposes of brevity, the term meal is used to replace oilmeal in part of this report.
268 FUNCTIONS FOR HOGS IN DRYLOT
weights caused difficulty in cumulating weekly gains and feed observa-

tions. However, these data are included in part of the analysis although
the remaining two experiments provide the main basis for the empirical
analysis.
Experiment no. 536 was designed especially for the requirements of
this study. It included treatments representing six protein levels and
three antibiotic arrangements. The six levels contained 10, 12, 14, 16,
18, and 20 per cent crude protein. There were three replicates without
antibiotics, three replicates with aureomycin, and two replicates with
terramycin, all using four pigs per lot. A single protein level was fed
to each lot throughout the feeding period — because the study was de-
signed for regression analysis, rather than analysis of variance. The
replicated lots were retained to fulfill requirements for other studies
using analysis of variance. The pigs were weighed each two weeks and
feed inputs were accumulated to provide input-output observations at
varying points over the growth period. The pigs were taken off the ex-
periment at 200 pounds for use in a study dealing with the effect of pro-
tein level on carcass quality. Some of the pigs in this experiment con-
tracted a skin disease. The disease increased the experimental error
and reduced feed efficiency.
Experiment no. 554 was the same as experiment no. 536 except that
it employed three replicates with aureomycin and three
without. In the
empirical analysis which follows, the observations for terramycin have
not been used in deriving production functions and substitution quantities.
In all three experiments, corn was the main source of carbohydrate
and soybean oilmeal was the main source of protein. The soybean oil-
meal contained about 45 per cent crude protein. The per cent protein
in the corn, however, varied from 8.4 in experiment no. 554
and 8 in
experiment no. 506 to 7.2 in experiment no. 536. Each pound of the
corn and soybean oilmeal rations was supplemented by .04 pounds of
minerals (including steamed bonemeal, iodized calcium carbonate,
iodized salt, and trace minerals) and .01 grams of vitamins
(including
vitamins A, D 2 5 meg. of B
, 12 niacin, pyridoxine, riboflavin, and
,
thiamin). The lots with antibiotics received 5 mg. of

aureomycin or
terramycin per pound of feed. In all experiments the complete
ration
was ground, mixed, and self -fed. The breed of hogs was not the
same
in all cases - Duroc in experiment no.
506, Duroc and Poland China x
Landrace x Duroc in experiments no. 536 and 554. Pigs
were randomly
allotted for all three experiments.
STATISTICAL ANALYSIS
Some background information of grain-protein substitution in

pork
production is already available. Previous experiments
suggest that for
light hogs, supplementary protein
ordinarily is a limitational feed —
some amount of it is necessary to attain a given gain. With
this infor-
mation, a particular problem in grain to protein
replacement (or protein
FUNCTIONS FOR HOGS IN DRYLOT 269
levels) for pigs becomes more nearly one of estimating the substitution
rates. This experiment was designed accordingly, and the design and
analysis allow the testing of alternative hypotheses with respect to the
nature of pork isoquants and grain to protein substitution ratios. While
an experimental design based on many individual pigs for different
treatments scattered over the contour lines and the ration lines would
have been better, if each hog had been used as a single observation,
funds did not allow this approach. In the analysis which follows, feed
and weight were accumulated over the complete feeding period with
measurements taken at stated intervals. Each of these measurements
then was used as an observation in the regression analysis. This ap-
proach gives rise to the problems of autocorrelation discussed in
Chapter 6. Mainly the procedure tempers statements which can be
made about confidence intervals and does not bias the estimates of
mean gain and substitution quantities. The observations between pigs
are independent, but the several observations for each pig are not.
Thus, given the inclusion of all observations for each pig, the usual
coefficients of correlation coefficients and standard errors must be
qualified. In general, the regression coefficients used in making esti-
2
mates of substitution rates and R values were significant even if the
number of degrees of freedom were considered to be equal only to the
number of hogs, rather than the number of hogs times the number of
observations on each. However, the probability levels indicated on
pages which follow are based on the total number of observations.
Several types of functions have been fitted to the data. However,
most use is made of the power function because, for the purposes
mentioned earlier, it allows estimation of the average rate of substitu-
tion over a weight interval. The major reason for making this study
was to allow determination of minimum cost feed combinations, as an
average over the weight intervals stated. Emphasis in United States
pork science is now on producing hogs to around 225 pounds of weight,
to avoid production of surplus lard and fatty cuts. This is becoming the
common practice. Within this framework, the major feeding problem
is to determine the least-cost rations in attaining this weight and the
emphasis on this study is made accordingly. As mentioned above,
average rates of substitution over three weight intervals are estimated
by power functions, to conform with certain practical considerations in
pork production. In the regression equations which follow, we use the
following symbols: Y is gain in pounds per pig after weaning, C is
pounds of corn, and P is pounds of soybean meal fed beyond weaning.
Regression Equations
Regression equations were fitted both to gain data over the weight
range 34-200 pounds and for each of the weight intervals, 34-75 pounds,
75-150 pounds, and 150-200 pounds. Hereafter, the regressions for
individual weight ranges will be referred to as interval functions; those
for the 34-200 range will be referred to as over-all functions. A dif-

ferent function was fitted for three weight intervals to allow greater
flexibility in substitution rates. This statement applies to the loga-
rithmic or Cobb-Douglas function. Although it allows diminishing rates
of substitution, and diminishing productivity of both feeds increased in
constant proportions or one increased by itself, derived as a single
regression equation it does not allow the substitution rate to change
with hog weight. In other words for a ration or scale line all contour
,
lines will have the same slope or rate of substitution at each point they
intersect the ration line. Quadratic, ratio, square root, and other
equations allow substitution rates to change along a scale or ration line.
However, the power functions for different intervals allow different
rates of substitution within each. While a different substitution rate at
each pound-change in hog weight is theoretically and logically possible,
substitution rates which serve as "averages” over a few weight intervals
are sufficient for farmer decisions since the same ration is fed for
several days, rather than changed every day. While the Cobb-Douglas
functions for the intervals cause the intersection points of all weight
contours and ration lines to have the same substitution rate within an
interval, it does allow estimation of three different substitution rates,
as averages, for the three weight intervals.
The regression equations (production functions) derived from the
experimental observations are included below. Regression equations
are provided separately for the two major drylot experiments used as
a part of this study. The observations from these two experiments
have then been pooled as a further basis for estimating interval and
over-all functions. It is likely that the pooled results more nearly
parallel typical farm conditions than either experiment by itself.
Also, separate regression coefficients have been estimated for the
protein level study of experiment 506. The rations and other conditions
in this experiment were identical with the others except that
each lot
was not fed the same ration over the complete weight range; protein
levels were lowered as successive weights were reached. Finally,
the
observations from this study have been pooled with those from the two
studies mentioned above for further estimates of the productivity
and
substitution coefficients. (Regression equations for the single
experi-
ment, 506, are recorded elsewhere.)
Many production function equations were derived in the course of
this study. Since space prohibits presentation of
all these, only enough
will be presented to illustrate those which gave
great similarity or
difference in prediction.
2
Production Functions
A. Production functions from experiment 536.

I. With aureomycin
(a) Cobb-Douglas
(8.1) 34-75 lbs.: Y = 1.445C'*47 P* 289

(8.2) 75-150 lbs.: Y = .555C'793 P 151 •
(8.3) 150-200 lbs.: Y = ^C'798 ?*181

(8.4) over-all: y = l.seoc^p-201
(b) Quadratic
(8.5) over-all: Y = 2.032 + .324C + .464P

2 2
- .0G0129C - .000917P
- .000111CP
n. Without aureomycin
(a) Cobb-Douglas
(8.6) 34-75 lbs.: Y = l^SSC^P*357
(8.7) 75-150 lbs.: Y = .342C*938 P*083 T
(8.8) 150-200 lbs.: Y = .345C’88S P*096
(8.9) over-all: Y = 1.174C"837 P"179
(b) Quadratic
(8.10) over -all: Y = 2.815 + .318C +

2 2
- .000108C - .000665P
- .000118CPT
B. Production functions from experiment 554.

I. With aureomycin
(a) Cobb-Douglas
(8.11) 34-75 lbs.: Y = 1.874P ,323 C'193
(8.12) 75-150 lbs.: Y = .SISP^C-735
(8.13) 150-200 lbs.: Y = .523P" 082 C* 884
(8.14) over -all: Y = 1.412P*218 C*828
In terms of degrees of freedom figured on the basis of all observations, the symbols
2
used have these meanings: t Regression coefficients significant at a probability level of

more than 20 per cent; t significant at 10-20 per cent; * significant at 5-10 per cent; no
asterisk means a regression coefficient significant at a probability level of 5 per cent or
less.
(b) Quadratic
2
(8.15) over-all: Y = .125 + .995P + .282C - .000865P
+ .000Q46C 2 - .00144PC
II. Without aureomycin

(a) Cobb- Douglas
n C*529
(8.16) 34-75 lbs.: Y = 1.820P'*
P* 218
(8.17) 75-150 lbs.: Y = .8 14 C891
089 ,S77
(8.18) 150-200 lbs.: Y= .428P' C
229 592
(8.19) over-all: Y = 1.577P* C*
(b) Quadratic
2
(g.20) over-all: Y = 6.441 + .692P + .258C - .003 07P
2
- .00Q077C + .000615PC
C. Production functions from experiments 536 and 554 pooled.

I, With aureomycin
(a) Cobb-Douglas
297 533
(8.21) 34-75 lbs.: Y = 1.600P’ C'
WZ 787
(8.22) 75-150 lbs.: Y= .714P' C-
092 858
(8.23) 150-200 lbs.: Y= .439P- C*
(8.24) over-all: Y = l^P^C* 836
(b) Quadratic
(8.25) over-all: Y = 3.004 + .582P + .314C - .00144P 2

- .000099C 2 - .000118PCJ

(a) Cobb-Douglas
(8.26) 34-75 lbs.: Y = 1.662P’ 287 c *529

183 770
(8.27) 75-150 lbs.: Y = .614 P' C'
0S1 918
(8.28) 150-200 lbs.: Y = .343P* C’
208 818
(8.29) over-all: Y = 1.422P’ C‘
(b) Quadratic
2
(8.30) Y = 3.778 + .620P + .288C - .00165P
2
- .000085C - .00009PCJ
In addition to the regression equations provided above, certain others

were also derived. A few of these are given below for experiment 554.
Table 8.1. Number of Observations, Standard Errors, and Values

of t for Cobb -Douglas Equations
Standard Error (s) and Values of t^

Treatment N*
S b(p) S b(c) t
P
A. Experiment 536
I. With aureomycin
34-75 lbs 34 .0366 .0471 7.91 11.61
75-150 lbs 52 .0232 .0305 6.52 26.05
150-200 lbs 56 .0264 .0416 6.08 19.15
over -all 142 .0090 .0103 22.46 61.08
II . Without aureomycin
34-75 lbs 30 .0407 .0502 8.77 9.23
75-150 lbs 42 .0409 .0568 1.53 16.52
150-200 lbs 62 .0316 .0457 3.03 19.41
over -all 119 .0162 .0186 11.06 35.26
B. Experiment 554
I. With aureomycin
34-75 lbs 42 ,0184 .0256 17.59 19.26
75-150 lbs 60 .0150 .0238 11.61 30J39
150-200 lbs 55 .0196 .0422 2.71 20.48
over -all 161 .0084 .0104 26.03 60.26
34-75 lbs 46 .0216 .0310 12.85 17.06
75-150 lbs 59 .0164 .0271 13.3 25.5
150-200 lbs 57 .0247 .0513 3.59 17.1
over -all 161 .0087 .0109 26.38 54.42
C. Experiments 536 and 554 pooled

I. With aureomycin
34-75 lbs 76 .0273 .0279 14.31 19.11
75-150 lbs 112 .0163 .0244 8.70 31.41
150-200 lbs 111 .0174 .0318 5.30 26.94
over -all 303 .0130 .0158 15.34 40.35
34-75 lbs 76 .0230 .0311 12.52 16.99
75-150 lbs 102 .0221 .0343 7.38 22.49
150-200 lbs 119 .0204 .0342 3.98 26.77
over -all 280 .00917 .0118 21.41 52.32
^Number of observations based on all measurement for all hogs.

^The p subscript refers to soybean meal and the c subscript refers to corn.
D. Other production functions derived for experiment 554.

I. With aureomycin; over -all
2
(8.31) Y = 12.24C + .372P + .364C - .000977P
2
- .000126C - 1.010C/P
(8.32) Y = - 8.782 - .794P + .0754C + 4.966 ylT

+ 1.196 YCT + .750 ypcT
EL Without aureomycin; over -all
2 2
(8.33) Y = 12.659 + .454P + .329C - .00120P - .0001Q7C
- .981C/P
(8.34) Y = -13.302 - .190P + .206C + 8.77 VF

+ .477 ]/Ct + .033 VVCt
Standard errors and t values for selected Cobb-Douglas and quad-

ratic equations are provided in tables 8.1 and 8.2. All of the equations
2
for which these equations are presented had R values of .94 or greater.
In fact none of the equations from 8.1 through 8.34 had R 2 values smaller
than .90.
Fromthese statistics, and considering the relation of measurements
to problems of autocorrelation, it was decided that the regression coef-
ficients provided useful estimates for the major objective of this study.
Table 8.2. Standard Errors and t Values for Regression Coefficients

and Number of Observations for Quadratic Cross-Product
Equations, Pooled Data from Experiments 536 and 554
Standard Errors
N s b(p) s b(c) s b(p)

2
s b(c) 2 s b(pc)
With aureomycin 303 .0461 .0154 .0002 .00003 .00013

Without aureomycin 280 .0565 .0176 .0002 .00027 .00003
t Values
*P
tC
V t
c 2 ^pc
With aureomycin 12.61 20.39 7.30 3.63 0.93

Without aureomycin 10.97 16.34 6.22 2.93 0.59
Product Isoquants
Selected equations of pork isoquants have been derived

from the
production 'functions estimated above. These have been
derived only for
pooled data from experiments 536 and 554. The equations
for the in-
terval Cobb-Douglas functions provide isoquant
families which all have
the same slope within an interval along a ration line
intersecting the
origin. However, the slopes for the isoquants
differ among intervals.
The quadratic function provides an isoquant family with each
product
contour having a different slope along a ration line.
I. With aureomycin
(a) Cobb -Douglas
(8.35) 34-75 lbs.: C
(8.36) 75-150 lbs.: C
(8.37) 150-200 lbs.: C
(8.38) over -all: C
(b) Quadratic
(8.39) C = 1587.384 - .596P
+ 5050.505 /.0000005561P 2 - .000304P + .0991 - .000396Y

(a) Cobb -Douglas
i
(8.40) 34-75 lbs.: C
(8.41) 75-150 lbs.: C
,916
(8.42) 150-200 lbs.: C

/___ Y_ \ 081
\.343P" /
i
616
(8.43) over -all: C
(b) Quadratic
(8.44) C = 1694.232 - .529P
t 5882.353 Y~- .000000552P 2 - .000307P + 8.297 - .000340Y
The isoquant equations above allow derivation of marginal substitu-

tion rates. The Cobb-Douglas equations give rise to substitution
coefficients which are constant, within a weight interval, along

a ration
line. Substitution ratios for the quadratic function change at each point
along a ration line. The derivatives below correspond to the isoquant
equations above for pooled data.
I. With aureomycin
(a) Cobb-Douglas
34-75 = .557 (8.47) 150-200 lbs.: = .108

(8.45) lbs.:
| ^
(8.46) 75-150 lbs.: -||-= .185
^ (8.48) over -all: |^= .314
^
(b) Quadratic
-39.419 + .293P
-.596 t-
2
/.000000556P + .9991 - .G00304P - .000396Y
EC. Without aureomycin

(a) Cobb-Douglas
(8.50) 34-75 lbs.: ||- = .543

^ (8.52) 150-200 lbs.: — = .089
^
fir r
(8.51) 75-150 lbs.: = .211 (8.53) over -all— = .338
^ ^
(b) Quadratic
-90.465 - .00325P
.5294 t-
2
/-.000000552P - .000307P + 8.297 - .000340Y
Comparison of Estimates From Different Functions
The main purpose of this study was to derive average rates of sub-
stitution of specified weight intervals in pork production. It was not to
establish the most efficient equation for estimating all specified quanti-
ties which might be derived from a pork production surface. However,
the data do provide a basis for some comparisons. A few comparisons
of product contours and input -output curves are provided below.
Figures 8.1 and 8.2 show outputs predicted for 16 and 14 per cent
protein rations, respectively, by the Cobb-Douglas and quadratic pro-
duction function equations for experiments 536 and 554. Figures 8.3
and 8.4 allow comparisons between four forms of functions for these
two rations under experiment 554. Because of its particular algebraic
0 100 200 3 00 4 00 500 6 00 700

POUNDS OF FEED
Figure 8.1. Feed-gain relationships for a 16 per cent protein ration

derived from experiments 536 and 554 with aureomycin.
restraints, the quadratic cross-product equation tends to underestimate

gains at heavy weight intervals. Otherwise, the remaining equations
are quite similar. The Cobb-Douglas equation appears to fit the data as
well as any of the other forms, probably because marginal feed produc-
tivity was relatively constant at heavy weights and the hogs were not
carried to levels approaching maximum weights and sharply declining
feed productivity.
Table 8.3 provides comparisons of pork gains predicted from Cobb-
Douglas and quadratic production functions under experiments 536 and
554 pooled with aureomycin. While they are not presented in the text,
interval functions were derived for the quadratic equation as well as

derived from experiments 536 and 554 with aureomycin.
100 300 400 600

POUNDS OF FEED

derived from experiment 554 with aureomycin.
for the Cobb-Douglas equation. Estimates from both over -all and in-
terval functions in attaining pig gains of 60, 110, and 175 pounds are
given in Table 8.3.
As these data indicate, the interval quadratic equations provide es-
timates which are quite similar to those of the Cobb-Douglas interval
function, especially for gain levels of 60 and 175 pounds. The predic-
tions by interval functions were even more similar for the Cobb-
Douglas as compared to the ratio and square root equations of 8.31
through 8.34. The Cobb-Douglas and quadratic over -all equations

derived from experiment 554 with aureomycin.
17 5 -pound pigs
Figure 8.5. Comparison of 60 pound pig contours (showing corn-

soybean meal combinations for producing 100 pounds
of gain) predicted from six different production
functions. Experiment 554 with aureomycin.
provide similar estimates for the 60-pound gain level, but have much
greater differences for the 110- and 175-pound gain levels. Again
over-all equations by the Cobb-Douglas function gave gain predictions
which were much more similar to those of equations 8.31 through 8.34,
Figure 8.6. Comparison of 110 pound pig contours (showing corn and
soybean meal combinations for 100 pounds of gain)
predicted from six different production functions.
Experiment 554 with aureomycin.
Pounds of Soybeen Meal
Figure 8.7. Comparison of 60 pound pig contours (showing corn-

soybean meal combinations for producing 100 pounds
of gain) predicted from four production functions.
Experiments 536 and 554 pooled, with aureomycin.
Some comparisons of gain isoquants predicted from different equa-

tions are provided in figures 8.5, 8.6, and 8.7. The interval Cobb-
Douglas functions provide gain isoquants for all three weight levels
which are, on the average, more similar to selected other functions es-
timated in an over -all manner, than are those from the over -all Cobb-
Douglas function. (The data shown in the figures are the means of lot
treatments within the earlier stated weight intervals, from which the
regression equations were fitted. However, the regression equations
were fitted to the individual observations, rather than to the means of
these.) The comparisons in these figures are only for experiment 554
with aureomycin. The Cobb-Douglas interval functions and the other
over -all functions correspond even more closely for experiments 536
and 544 pooled. Based on these and other comparisons, it was decided
that the Cobb-Douglas interval functions might serve best for estimat-
ing average marginal rates of substitution over the three intervals.
Again, average rates of substitution over intervals were desired be-
cause farmers typically feed the same ration over a period of time;
rather than changing it with each unit gain of pigs. Hence, the main ob-
jectives of this study are continued using the Cobb-Douglas interval
functions. Also, certain other comparisons between rations are made
using either these functions or the over -all Cobb-Douglas function.
This function is used, not because it is most appropriate for all phases
of analysis, but because (a) the function does appear to conform with
the data about as well as other over -all functions and (b) use of the
same type of function throughout provides a “standard” in comparisons.
COMPARISON OF RATIONS
Pigs fed aureomycin evidently had somewhat higher marginal feed

gains than those without this additive, on the basis of
production func-
tions estimated by Cobb-Douglas and other over -all production func-
tions. Gain levels relative to feed input levels for three
levels of pro-
tein are indicated in Table 8.4. These comparisons are from the
over -all power function when aureomycin rations are compared. Values
of t, based on a number of degrees of freedom representing all obser-
vations in the study, are included in Table 8.5 for comparisons of re-
gression coefficients between interval production functions. Even con-
sidering a much smaller number of degrees of freedom, these differ
significantly. However, based on the number of degrees of freedom
representing all observations in the study , it is not possible to say that
regression coefficients for the production functions differ significantly.
(See Table 8.6.) Since most protein feed supplements for hogs now con-
tain antibiotics, the functions with aureomycin will be emphasized in
certain of the analyses and comparisons which follow. At the time of
the study, the cost of the antibiotic added was about 60 cents per 1,000
pounds of feed.
Table 8.4. Comparison of Total Product From Weaning From 50-Pound Feed
Increments With and Without Aureomycin, Derived From Over-all
Cobb-Douglas Equations for Experiments 536 and 554 Pooled*
12 Per Cent Protein 14 Per Cent Protein 16 Per Cent Protein

.
Table 8.5. Values of t for Difference between Over-all Cobb-

Douglas Regression Coefficients With and
Without Aureomycin, Experiments
536 and 554.
Weight Range Protein Coefficient Corn Coefficient
34-75 .37 .13

75-150 1.08* .13
150-200 .59 1.82*
over- all .71 1.43*
*5 < p < 10;

tl0<p<20;
*p >40; based on all observations in study.
Marginal Feed Productivity
The equations of marginal feed productivity for the Cobb-Douglas

over -all function applied to rations including aureomycin are:
6Y
(8.55) Corn: = .871P .200 £-.364
(8.56) Soybean meal: —5Y

= .274C 3
p-.E
Based on the Cobb-Douglas over -all function, the total and marginal
gains for aureomycin are presented in Table 8.7. In these data the ra-
tion has been held constant in the sense that corn and soybean meal are
always fed in the same proportions. Only data for aureomycin are
shown in the text figures since it appears that this antibiotic is eco-
nomical to include in the ration (in the quantities of this study) at any
prospective price level. Even with no increase in an “average year,"
this small investment is likely good insurance against certain diseases.
The data show that the marginal productivity of feed declines as
more of it is fed, in a constant -proportion ration. For example, the
Table 8.6. Values of t for Testing Interval Regression Coefficients

Against Each Other — Experiments 536 and 554
Pooled— Cobb-Douglas Functions*
With Aureomycin Without Aureomycin

Experiment and Comparison
Protein Corn Protein Corn
34-75 lbs. vs. 75-150 lbs. . . 6.67 9.09 5.56 7.14

34-75 lbs. vs. 150-200 lbs.. . 9.09 11.11 10.00 11.11
75-150 lbs. vs. 150-200 lbs. 2.94 2.78 3.85 5.88
Significance level for all figures is 0 < p < 5, based on all observations
in study.
Table 8,7. Total Gain Beyond Weaning and Marginal Gam for Different Feed Levels
per Pig With Rations m
Fixed Proportions; (Gam Over Beginning Weight
of 34 Pounds) Experiments 536 and 554 Pooled With Aureomycm
Marginal or Additional Gam

Total Gam per Pound Added Feed
Pounds of
Feed After Per cent protein m ration Per cent protein m ration
Weaning 20 10 12 14 16 18
10 12 14 16 18 20
50 18.41 20.77 21.77 22.12 22.12 21.83 .3682 .4154 .4353 .4425 .4423 .4367
100 32.85 37.06 38.84 39.47 39.46 38.96 .2888 .3258 .3414 .3470 |
.3469 .3425
150 46.10 52.00 54.49 55.39 55.37 54.66 .2649 .2988 .3131 .3182 .3182 .3141
200 58.62 66.13 69.30 70.43 70.41 69.51 .2504 .2825 .2961 |
.3009 !
.3008 .2970
250 70.63 70.68 83.50 84.87 84.84 83.76 .2402 .2710 .2840 .2886 .2886 .2849
300 82,25 92.79 97.23 98.83 98.80 97.54 .2324 .2622 .2748 .2792 .2792 .2753
350 93.55 105.54 110.59 112.41 112.38 110.94 .2261 .2551 .2672 .2716 .2716 .2680
400 104.59 117.99 123.65 125.68 125.64 124.04 .2207 .2490 .2612 .2654 .2652 .2620
450 115.41 130.19 136.43 138.67 138.63 136.86 .2164 .2440 .2556 .2598 .2598 .2564
500 126.03 142.17 148.98 151.43 151.38 149.45 .2124 .2396 .2510 .2552 .2550 .2518
550 136.47 153.95 161.33 163.98 163.93 161.84 .2088 .2356 .2470 .2510 .2510 .2478
600 146.76 165.56 173.49 176.34 176.29 174.04 .2058 .2322 .2432 .2472 .2472 .2440
650 156.91 177.01 185.49 188.54 188.48 186.08 .2030 .2290 .2400 .2440 .2438 .2408
700 166.93 188.31 197.34 200.58 200.52 197.96 .2004 .2260 .2370 .2400 .2408 .2376
750 176.84 199.48 209.04 212.48 212.41 209.70 .1982 .2234 .2340 .2380 .2378 .2348
800 186.63 210.53 220.62 224.25 224.18 221.32 .1958 .2210 .2316 .2354 .2354 .2324
“50th pound” of feed under a 10 per cent ration adds .368 pound to hog
weight; at 150 pounds of total feed, “one more pound” adds .265 pound to
hog weight; at 400 pounds the added gain is only .221; while at 800
pounds it is only .196 pound. With a ration of 14 per cent protein the
added or marginal gain per added pound of feed is .435, .313, .261, and
.232 pound for total feeding levels of 50, 150, 400, and 800 pounds, re-
spectively.
The same relationship is also shown in the total gain figures. The
total weight increases by smaller and smaller increments as feed in-
take per hog is increased. Of particular importance, however, is the
difference between protein rations in total gain for a given feed Intake.
With a feed input of 400 pounds, total gain is 105 pounds under a 10 per
cent ration but 124 pounds under a 14 per cent ration and 126 pounds
under a 16 per cent ration. Total gain for a given feed input under an
18 per cent ration is practically the same, or slightly lower, than for a
16 per cent ration. The predicted total gain is less for a 20 per cent
ration than for a 16 per cent ration. Since soybean meal costs more
than corn, it cannot be fed profitably in quantities giving 18 or
20 per
cent protein in the ration: for any given total feed intake per hog, total
gain and value of hog decreases, and total cost of feed increases. These
problems can be better solved, however, from the substitution data of
later sections. How far the ration should be extended between
14 and
16 per cent protein depends on the substitution and price ratios for the
two feeds.
Average Elasticities
Diminishing marginal transformation of feed into pork is apparent

in the exponents or regression coefficients in the Cobb -Douglas pro-
duction functions. The exponent for either corn or protein in these
equations (i.e., the regression coefficients for the data in logarithmic
form) can be termed the production elasticities.
They are average elasticities since they indicate, as constants over
the surface when estimated by a Cobb-Douglas function, the percentage
increase in weight for each 1 per cent increase in feed intake per pig.
If these quantities are 1.0, they indicate that a 1 per cent increase in
the amount of feed increases hog weight by 1 per cent —feed has a con-
stant added productivity. In all cases, however, these quantities are
less than 1.0 for either corn or protein denoting (1) that if each feed is
increased alone by 1 per cent, weight will increase by less than 1 per
cent, and hence (2) the marginal production (the amount added to total
weight or production from an added pound of feed) becomes smaller and
smaller as feed intake per hog increases.
The 536 and 554 pooled production functions show that for 34-75
pound pigs with aureomycin, corn has an elasticity of .533 — a 1 per
cent increase in the amount of corn fed will increase pig weight by only
.533 per cent. Similarly, the elasticity for soybean oilmeal is .297, in-
dicating that this percentage increase in weight can be expected from a
1 per cent feed increase. For all of the Cobb-Douglas regression coef-
ficients presented, the quantity is less than 1.0 for either corn or soy-
bean meal alone.
The sums of the exponents for corn and soybean meal suggest the
general nature of the input -output or feed transformation ratio as they
are fed in a ration of fixed proportions (i.e., a ration of say 16 per cent
protein fed from one weight to another). A sum of 1.0 indicates that a
1 per cent increase in the two feeds, always held in constant propor-
tions, will result in a 1 per cent increase in weight or pork output. For
all of the logarithmic functions derived, however, the sum of the expo-
nents is less than 1.0. This indicates that on the average, as more feed
is fed, always with corn and protein in the same proportion, the hog will
gain weight — but at a diminishing marginal rate. In the over -all func-
tion for 536-554 pooled with aureomycin, the sum of the elasticities is
.836 (.200 + .636) denoting that a 1 per cent increase in feed intake in-
creases hog weight only .836 per cent.
In the Cobb-Douglas equation, and given the average nature of the
elasticities so defined, three tendencies are evident and were expected
from previous nutritional research:
1. The elasticity coefficient for protein declines with higher weight
intervals. This phenomenon is expected since a small pig is
more dependent on protein for growth than a heavier one.
2. The elasticity for corn increases between weight intervals; a
phenomenon also expected since a mature hog needs more car-
bohydrate feeds in fattening.
of Elasticity Coefficients From Pooled Data for Logarithmic Functions*

Table 8.8. Summary
Corn Protein Coefficient of
Interval and
Elasticity Elasticity Substitution
Experiment
1. 34-75 pounds
.557
536-554 pooled WA .533 .297
.287 .543
536-554 pooled WOA .529
536- 554-506 pooled WA .530 .303 .572
536-554-506 pooled WOA .514 .289 .562

.523 .294 .557
Average of 4 for 34-75 pounds . .
2. 75-150 pounds
536-554 pooled WA .767 .142 .185
536- 554 pooled WOA .770 .163 .211
536-554-506 pooled WA .762 .136 .178
536- 554-506 pooled WOA .805 .158 .197
Average of 4 for 75-150 pounds. . .776 .150 .193
3. 150-200 pounds
536-554 pooled WA .856 .092 .108
536-554 pooled WOA .916 .081 .089
536-554-506 pooled WA .849 .110 .130
536-554-506 pooled WOA .907 .090 .099
Average of 4 for 150-200 pounds . .882 .093 .107
*WA indicates “with aureomycin” and WOA means “without aureomycin.*
3. The sum of the elasticities increases between weight intervals.

A suggestion of declining feed productivity is the decline in the
production function constants between weight intervals.
The clear tendency for gain elasticities of oilmeal to decline with

weight are shown in Table 8.8. For example, over the weights 34-75,
75-150, and 150-200 pounds, the average of all elasticities computed
were, respectively, .294, .150, and .093. In contrast, the corresponding
elasticities for corn were .523, .776, and .882. (See Table 8.8.) Simi-
larly, the coefficients of substitution of soybean meal for corn decline
consistently between weight intervals, as estimated by the interval
functions. As an average for all interval functions estimated, these co-
efficients (Table 8.8) are .557, .193, and .107 for the weight ranges
34-75, 75-150, and 150-200 pounds. (The coefficients in column 3 of
Table 8.8 are not the complete derivatives of the isoquant equations,
but are only the multipliers of the feed ratios as indicated in equations
8.45 through 8.53.) Similar changes in elasticity and substitution coef-
ficients are apparent from other algebraic forms of over -all equations.
However, those for the Cobb-Douglas interval equations provide quan-
tities which might be looked upon as averages for these weight ranges.
Previous nutritional research has indicated that the rate at which
protein replaces corn decreases as the weight of the hog increases.
This condition is, of course, inherent in (a) the coefficients of the loga-
rithmic functions presented above and (b) the signs of the coefficients in
the quadratic equations for the pooled data. That protein and corn sub-
stitute for each other at a diminishing marginal rate, with hog weight at
a fixed level, is inherent in the derived equations for the same reason;
it is obvious in the Cobb-Douglas equations, equations 8.35 through 8.43,
since a constant value of Y is divided by an increasingly larger quantity

relating to protein (as protein is raised to the relevant power). These
same conditions are reflected in the coefficients for the equations of

marginal substitution rates. The decline in the substitution rate (of
soybean meal for corn) is greater, of course, with higher weight inter-
vals than these interval coefficients suggest.
The reason is this. As the hog grows to heavier weights, a greater
amount of feed is required to produce a given gain; for any one level of
protein feeding, more corn is required to produce the given gain. Hence
the C/P ratio, as well as the substitution coefficient itself, declines for
heavier hogs in the equations of marginal substitution rates.
CORN AND SOYBEAN MEAL SUBSTITUTION RATES
We now turn to the major objective of this study, namely, the esti-
mation and use of average rates of feed substitution over selected
weight intervals. The equations of isoquants and marginal substitution
rates, based on the interval Cobb-Douglas functions, have been used to
derive the relevant quantities, expressed per 100 pounds of gain, in
Table 8.9. The first two columns under each weight show the various
combinations of soybean oilmeal and corn which will produce 100 pounds
3
of pork. They may be looked upon as providing the substitution possi-
bilities over the three weight ranges. More specifically they are calcu-
lated as the substitution possibilities within the ranges, but at exactly
60, 110, and 175 pounds.
The figures again show that the amount of corn required to produce
100 pounds of gain decreases as more protein is fed and vice versa. In
other words, they show the many possible combinations of feed which
will produce a given amount of pork for pigs of three different weights.
The marginal substitution rates show the amount of corn replaced
by 1 pound of soybean oilmeal at each of the combinations shown in the
first two columns. These substitution rates are of a diminishing mar-
ginal nature; each added pound of protein substitutes for less corn than
the previous pound in producing a given amount of gain. For 60 -pound
pigs (pigs in the 34-75 pound range) 1 pound of soybean oilmeal substi-
tutes for 12.5 pounds of corn when the ration includes 15 pounds of soy-
bean oilmeal and 337 pounds of corn; it substitutes for only 4.3 pounds
of corn when the feed combination is 30 pounds of soybean oilmeal and
229 of corn; with 75 pounds of soybean oilmeal and 137 pounds of corn,
1 pound of protein replaces only a pound of corn.
Diminishing substitution rates also hold true for weights of 110 and
175 pounds. Thus the rate of replacement of corn by protein declines
as the growing stage merges into the fattening stage. In Table 8.9 with
30 pounds of soybean oilmeal in the ration, 1 pound of soybean oilmeal
for a 60-pound pig replaces 4.3 pounds of corn; it replaces only 1.8
pounds of corn for a 110-pound pig, and 1.2 pounds for a 175 -pound pig.
s
The feed quantities were first derived for the gains necessary to take a pig from the
beginning of the specific interval to the weight indicated. Next, these were transformed to
the equivalent of a 100-pound gain since this is the manner in which many persons consider
rations.
Figure 8.8. Pork isoquants showing the combination of corn and soybean meal
which will produce 100 pounds of gain for pigs of 60, 110 and 175
pounds. Derived from interval Cobb-Douglas equations. (The lot
observations are used only to suggest the nature of the fit for
average gain data over each interval. The functions were not
fitted to these single lot observations but to many more observations
scattered over each specific interval.)
These differences in substitution rates are also reflected in the

pork isoquants shown in Figure 8.8. These curves, drawn from the
data in Table 8.9, show that, for a given gain, as the rate at which soy-
bean oilmeal substitutes for corn increases, the slope of the isoproduct
curve also increases, and vice versa (i.e., the slope of the curve de-
clines as more protein and less corn is fed to get a given gain). The
slopes of the curves also decline as the pig becomes heavier. This in-
dicates that the rate at which soybean oilmeal substitutes for corn de-
creases as the pig increases in weight.
The isoquants in Figure 8.8 are not all measured from the 34-pound
weight at the beginning of the experiment. Only the lower contour is so
measured and represented in the figure. The other two are measured
relative to the beginning of their respective weight intervals, with feed
inputs representing those beyond this level. In this sense, the three
isoquants do not express a production function represented by a single
“continuous” family of product contours.
;
total
cent)
CMC— C\3D— CSJCDtHCO O LO 05 Tt*
Protein
ration
OJ 05 oOH N M CO ^
rt
rHrHr-<>-HrHWrHWiHrH
N PO
(per
in
03 meal
P
6 by SBOM
corn
bfl
of
t-cDT^THT^ôcbcÔitr-cq
£
fa Marginal substitution
corn
HcocjiniMOQt- 00 ioin
1 rate
soybean
pounds
replaced
-pound CM rH rH ?H rH
§
& for
1
m i
I>
o
rH 0
£© c
co
OWMC-OOJMH^CJO^
fa
5 ^ *1
pounds
t* O
O CO « 05 U5 N 05 D
ffl
COC-WIO^COCO(NCS1(MHH
N
tog)
corn
coooeococococococooooooo
2 d co
I £1x3 X3
&o a
X5 % 5 oifloinotfloifloiflom
1 0 a a
o pounds
HHCCIlMCOW’Cf^lflin©^
a r
ô
fa +»
SBOM
&
0
a
0 cent)
WOOÔCONC'Wffl'ÔibO
^ a Protein
total
ration
05a o h h w «i co w in 0 rj? to*
*2 (per
o in
0 a 1
fa B
d w
CO meal
C
6
-^=-;
OP by SBOM
corn
bo
of
fa
rH W U5 t- CO 05
COrHOSCQCOlDCOH
OWO
O O) OM t—
0 Marginal
substitution CO 00 Eh CD
n
s3 rate
corn
pounds
replaced
-pound CO M? CM* CM rH rt HHH
soybean
0
fa
for
,
1
*3 1
§
Is r-i
r-i
0
CO
ODWOOOWCOtDN^NÔl
2 © -
O
SH CO 05 UO |> O rj? 05* oa
pounds
CO* CO* CD*
*1 corn
incOrHOCDCbCOC-COCOCOlOlO
** cococooocmcmcmcmcmcmcmcmcm
0 5 |2 S3 co
S 3
o S
X3 X3
| ’0^3
2
s ao0 pounds OtOOtOOtOOlOOlOOlOO
HHNNCOn^^LOLOOO)>
® o a SBOM
fa +»
o £
w S
g 5 cent)
H CO CO W Tf Tt<HMMCONU5005
.2 fa total
1
4-»
aj O Protein
ration
05 05* o’ rH O
hhhhhhhhhhnn<M* CO* lO CD* CO 00* C)
(S CD (per
in
—
d
5 if
C P
* }
'O
a>
W CO
meal
6 0 -r-rr
corn
by SBOM
,
bO
0 2 of HOO5ÎOrJfH0(MlO^tr*^lM
fa £ fa Marginal
ininoiownMMOicD^NHO
N ^ IO « ci W H H H H H H
substitution
XJ rate
corn
replaced CO Tl?
pound
05 poimds CM rH
soybean
§
for 1-
-
ft
3 OI
CO
0CO
H
1> LO t" O rH M CO CO CM
C“ 00 CO "Cf 1
S s .s pounds
corn
CM CO
rH CD CO* CO 00* 05 M<
00 CM I>
MMCOlO£NOO5COt-0IO^^CO CM* lf5
i5 s3 »
a rfCOMNNNHHHHHHHH
SLxs xj “
ô fl
X3 3 OWOtOOlflOlflOWOWOlfl
0 a oa
pounds
SBOM
® o
fa +»
[ 289 ]
-
Minimum Cost Rations
The substitution data from Table 8.9 provide the necessary infor-
mation for specifying the ration which will give the lowest cost for any
given gain. Since the substitution or replacement rates are of a declin-
ing rather than a constant nature, a different ration is required with
each change in the corn and soybean meal price ratio to give the least
cost gains. Feed costs for a given gain are lowest when the substitution
ratio equals the price ratio. Hence with a soybean oilmeal price of 4.52
cents per pound and a corn price of 2.0 cents, a price ratio of 4.52/2.0
or 2.26, specifies that a combination of 45 pounds of soybean meal and
182 pounds of corn gives the lowest feed cost for 60-pound pigs; the
marginal rate of substitution in Table 8.9 is also 2.26 for this feed
combination. With a soybean meal price of 3.34 cents per pound and a
com price of 1.0 cent, the price ratio of 3.34/1.0 or 3.34 specifies that
35 pounds of soybean oilmeal and 210 pounds of corn gives the lowest
feed cost for pigs of this weight. At this point the marginal rate of sub-
stitution of soybean oilmeal for corn is also 3.34.
Table 8.10 has been provided to simplify computation of the combi-
nation which will give the lowest feed cost for a given gain. In effect, it
specifies the feed combination at which the marginal rate of substitution
equals the price ratio being examined. One need only divide the soybean
oilmeal price or cost per pound (including vitamins, antibiotics, and
other ingredients) by the price of corn per pound. This ratio (or the
ratio nearest to it) can be located in the first column of Table
8.10.
Then the least-cost ration can be found along the same line, under the
appropriate weight. Suppose that soybean oilmeal costs 4.4 cents
per
pound and corn is 2 cents per pound. Since the ratio is 2.2, a ration
with 45.8 pounds of soybean oilmeal and 180.7 pounds of corn or
15.5
per cent protein will give the lowest feed cost for 60-pound
pigs; an
11.1 per cent protein ration — for this same price ratio —
will give low-
est feed costs for 110-pound pigs and a 10.0 per
cent protein ration will
give lowest costs for 175-pound pigs. (Actually, the
rations can be ap-
plied throughout the weight ranges.)
From 1936-52, Iowa prices averaged $1.74 and $3.40 per 100 pounds
of corn and soybean meal, respectively. Under this
average yearly
price ratio of 1 .96 the rations which give lowest
feed costs for 60-
H0-, 175-pound hogs have 16.0, 11.3, and 10.1 per cent proteiil,
re-
spectively If a single ration were fed from
.
weaning to marketing
weight (34-225 pounds) under this price ratio,
one having 14 per cent
protein would give lowest feed costs per
100 pounds of gain. The price
ra Ctl teS between years, of course.
In the 15 years between 1938
înio
and f'
1953, the ratio was as high as 2.8 and as low as
1.3 in Iowa.
TIME OF GAINS
Minimum feed cost, while an important consideration
in maximizing
pork profits, is only one facet of the economic problem.
Also to be
Table 8.10. Least-Cost Ration and Feed Quantities for Stated Price Ratios; Based on Cobb-DougLas
Interval Functions With Aureomycin — Pounds Feed per 100 Pounds Gain
Price ratio:
60- Pound Pig 110-Pound Pig 175-Pound Pig
Price per Pound SBOM (34 to 75 Pounds) (75 to 150 Pounds) (150 to 200 Pounds)
Price per Pound Corn
Per Cent Per Cent Per Cent

Corn SBOM Protein* Corn SBOM Protein* Corn SBOM Protein*
1.0 136.3 75.9 21.0 269.3 49.9 13.9 336.9 36.3 11.8
1.1 141.0 71.4 20.3 273.3 46.0 13.4 340.1 33.3 11.5
1.2 145.5 67.5 19.6 277.1 42.8 13.1 342.9 30.8 11.2
1.3 149.7 64.2 19.0 280.5 40.0 12.7 345.6 28.6 11.0
1.4 153.7 61.2 18.4 283.9 37.6 12.5 348.1 26.8 10.8
1.5 157.5 58.5 17.9 296.9 35.4 12.2 350.5 25.2 10.7
1.6 161.3 * 56.2 17.5 289.9 33.6 12.0 352.7 23.7 10.6
1.7 164.8 54.0 17.1 292.6 31.9 11.8 354.7 22.5 10.4
1.8 168.1 52.0 16.7 295.1 30.4 11.6 356.7 21.3 10.3
1.9 171.4 50.3 16.4 297.7 29.0 11.5 358.6 20.3 10.2
2.0 174.7 48.7 16.0 300.1 27.8 11.3 360.4 19.4 10.1
2.1 177.7 47.2 15.8 I 302.4 26.7 11.2 362.1 19.6 10.0
2.2 180.7 45.8 15.5 304.6 25.7 11.1 363.7 17.8 10.0
|
2.3 183.6 44.5 15.2 306.8 24.7 11.0 365.3 17.1 9.9
2.4 186.4 43.3 15.0 308.8 23.8 10.9 366.8 16.5 9.8
2.5 189.2 42.2 14.8 310.8 23.0 10.8 .368.3 15.9 9.8
2.6 191.9 41.1 14.6 312.7 22.3 10.7 369.7 15.3 9.7
2.7 194.4 40.1 14.4 314.6 21.6 10.6 371.1 14.8 9.7
2.8 197.0 39.2 14.2 316.3 20.9 10.5 372.3 14.3 9.6
2.9 199.5 38.3 14.0 318.1 20.3 10.4 373.7 13.9 9.6
3.0 201.9 37.5 13.9 319.8 19.7 10.4 374.9 13.5 9.5
3.1 204.3 36.7 13.7 321.4 19.2 10.3 376.1 13.1 9.5
3.2 206.6 36.0 13.6 323.0 18.7 10.3 377.3 12.7 9.5
3.5 213.4 34.0 13.2 327.5 17.3 10.1 380.6 11.7 9.4
4.0 223.8 31.2 12.7 334.5 15.5 10.0 385.5 10.4 9.3
To determine per cent protein in the total ration, soybean oilmeal of 45 per cent protein and corn of
8.4 per cent protein have been used.
considered are labor and other capital costs. However, as has been
pointed out elsewhere, feed represents such a major portion of total
pork production costs that any reduction in feed costs usually lowers
total expenses. An additional part of the feeding problem is that of
time. Gain comes slowly for small pigs which receive only small
amounts or no supplemental protein. Because of low daily gains, the
production period is extended accordingly and the selling price may be
increased or decreased from different rations depending on whether
faster gains put the pigs on the market in a higher or lower price period.
Hence, the farmer must consider the price to be received for his
hogs as well as the feed cost per pound of gain in deciding which ration
is most profitable. If a higher percentage of protein in the ration in-
creases cost per 100 pounds of gain but allows an even greater increase
in the price received for pork, the most profitable ration will not be the
one which gives the minimum feed cost.
Time Functions and Daily Gains
Since time and rates of gain are necessary along with substitution
rates and feed costs —
to specify the most profitable ration, the three
total time functions shown below have been derived.
2 2
T = 6.958 + .038Pt + .291C + .00197P - .000093C - .001073PC
(8.57)
2 2
T = 3.969 + .083Pt + .246C + .000078P t - .000123C + .735C/P
(8.58)
(8.59) T = -9.837 + .690P + .167C - .432 ]/P + 4.264 j/C~- .590 YVC
In these equations, total time (T) to consume a specified quantity of4

feed is considered a function of the amount of corn and protein fed.
Total time rather than daily rate of gain has been predicted since we
are interested in estimating the time required to reach marketing
weights under different rations.
With knowledge of the weights attained from different rations and
amounts of feeds, information on the total time required to attain these
weights and consume the corresponding amount of feed makes possible
the determination of the effect of different rations in putting hog mar-
ketings into periods of high or low prices. An equation for predicting
the change in the daily rate of gain in relation to the amount of protein
fed was computed as follows where D refers to daily rate of gain in
pounds:
(8.60) D =
-9.837 + .690P + .167C
1.369P
- .432
#200
—
C *
j/pl- 4.264
636
—
YcT- .590 |/PC
—
Using these two over -all functions, the total gains in Table 8.7 divided
by the total time in Table 8.11 gives the daily gain rates in Table 8.12.
Where gains over different intervals are being analyzed, the production
function for the particular interval, rather than for the over -all func-
tion, is divided by the corresponding time function. Total time as a
function of feeds fed, rather than daily rate of gain, was predicted in
the equations since our main interest is in predicting the time to reach
market weights under different rations. From the daily gain equation it
is possible to specify the change in the daily gain (the marginal daily
gain) associated with change in the amount of protein fed— with total
corn and total protein inputs at specified levels. The equation is shown
below where the derivative specifies the change in the daily gain for
each 1 pound change in the amount of protein fed:
(8.61) IP _
6P
636 S 8 7 3 5 3 2
1.369C* (.0334C - 1.967 + .853C* )P"* + .138P"* + .13P’* + .177C* P"* - .69P*
2
(.69P + .167C - .432 Vl? + 4.264 Vc - .59 YVC - .9837)
After examining the regression coefficients and their standard

errors and after plotting the functions against the observations, it was
4
In the equations the coefficients indicated by t are not significant at the 20 per cent
level of probability; those without asterisks are significant at the 5 or 1 per cent probability
levels on the basis of all weight observations included in the study. The coefficients of de-
termination for the three equations are .971, .966, and .973, respectively.
Table 8.11. Total Timeto Consume Various Quantities of Feed for Rations of 10, 12,
14, 16, 18,20 Per Cent Protein (Experiment 554 With Aureomycin and
Equation 8,59 With Square Root Terms)
Total Days (To Feed Out Different Quantities of Feed)

With Protein Levels of t
Pounds
of Feed* 10 12 14 16 18 20
per cent per cent per cent per cent per cent per cent
50 22.4 19.3 17.9 16.9 16.2 15.8
100 36.8 32.6 30.3 29.0 28.3 28.2
150 49.0 43.2 40.3 38.7 38.2 38.4
200 59.9 52.5 48.9 47.2 46.8 47.5
250 69.8 60.9 56.7 54.9 54.7 55.8
300 79.1 68.7 63.9 62.0 62.1 63.7
350 87.9 76.1 70.7 68.7 69.0 71.1
400 96.4 83.0 77.1 75.0 75.6 78.2
450 104.5 89.7 83.3 81.1 82.0 85.1
500 112.3 96.1 89.1 86.9 88.1 91.8
550 120.0 102.3 94.8 92.6 94.1 98.3
600 127.4 108.3 100.3 98.1 99.8 104.7
650 134.6 114.2 105.7 103.4 !

105.5 110.9
700 141.8 119.9 110.9 108.6 111.0 117.0
750 148.7 125.5 116.0 113.7 116.4 122.9
800 155.6 130.9 120.9 118.6 121.7 128.8
^Pounds of feed fed beyond weaning.

tDays beyond weaning.
decided that the square root function, equation 8.59, fits the data best.
This function allows prediction of the amount of time associated with
consumption of different quantities of feed under the various corn-
protein rations. The figures in Table 8.11 indicate that time is de-
creased in the ration up to slightly beyond 16 per cent protein; further
increases in protein lengthen the time period because a smaller pro-
portion of carbohydrates is available for the fattening stage of the pro-
duction process.
Using the total gain and the time functions from above, the figures
of Table 8.12 can be computed. These show the average daily rate of
gain for 50-pound feed intervals. These figures show greater daily
rates of gain for small pigs with rations containing as much as 18 per
cent protein. For pigs weighing 85 pounds, the higher protein percent-
age lowers the average daily rate of gain; a 16 per cent protein ration
gives a greater average daily rate of gain beyond this weight. In other
words, pigs receiving 18 per cent protein are predicted to have the
greatest daily rate of gain at the outset but eventually those with a
Table 8.12. Average Daily Rates of Gain at Various Total Levels of Feed Intake
for Different Protein Rations* (With Aureomycin)
Average Daily Gain up to Time of Feed Level

Shown for Protein Levels of
Pounds
of Feed 10 12 14 16 18 20
100 .894 1.137 1.281 1.362 1.392 1.382

150 .940 1.203 1.357 1.431 1.450 1.424
200 .979 1.259 1.417 1.493 1.504 1.465
250 1.012 1.308 1.473 1.547 1.551 1.501
300 1.040 1.350 1.521 1.595 1.592 1.532
350 1.064 1.387 1.564 1.637 1.629 1.560
400 1.085 1.421 1.603 1.675 1.661 1.585
450 1.105 1.451 1.639 1.710 1.691 1.608
500 1.122 1.479 1.671 1.742 1.718 1.628
550 1.138 1.505 1.701 1.771 1.743 1.646
600 1.152 1.528 1.729 1.798 1.766 1.663
650 1.165 1.550 1.755 1.824 1.787 1.678
700 1.178 1.571 1.780 1.847 1.807 1.693
i i
Daily gains are averages from weaning to the time necessary to consume the feed
shown. They are not daily rates of gain at exactly the feed levels indicated.
16 per cent ration are expected to catch up-ending the production

pe-
riod with a higher average daily rate of gain. 5
Table 8.13 shows the predicted daily rates of gain between
50-pound
feed intervals. In contrast to Table 8.12 which showed the
cumulative
average, these data show the average gain rate only
within each feed
interval. In this form, daily gain rates for a
16 per cent ration surpass
those of an 18 per cent ration except for the first
100 pounds of feed.
Those for a 14 per cent protein ration are only slightly
below those for
a 16 per cent ration as feed intake and hog weights
reach higher levels.
Of more direct importance to the farmer is the
total time required
to add different amounts of gain on hogs of specified
weights. These
figures are presented in Table 8.14 and are
rounded to the nearest day,
since the farmer does not refine his feeding
and marketing decisions to
fractions of days. For 34-75 pound pigs a
gain from the beginning to
the end of the weight interval is attained
with minimum time when the
ration reaches 18 per cent protein; over
the weight interval 75-150
Anearlier experiment at Iowa State

University (Catron, D. V. et al Re-evaluation of
^ ,I a, An Science > 2(2 > : 221 ' 32 - 1952.) even showed
ênt ^to ™ ^ tT\
e lag^ in ^ behind ^
that pigs on a 12
ose with higher protein rations at the outset,
eventuaUv increased°thM>
more TroUn
as those getUng^lfêrênt^rofefn.
1
to a ?° int where « surpassed the
*
reMhed a Weight °f 200 “
rate for pigs receiving
s as
.
Table 8.13. Average Daily Gain Within Feed Intervals for Different Percentages
of Protein (With Aureomycin)
Daily Gain Computed as an Average Between 50-Pound Feed

Increments With Protein Levels of
irUUilUB
of Feed 10 12 14 16 18 20
1.229 1.371 1.432 1.431 1.383
150 1.575 1.638 1.618 1.540
200 1.154 1.522 1.713 1.775 1.637
250 1.818 1.879 1.829 1.706
300 1.249 1.678 1.962 1.899 1.758
350 1.282 1.740 1.972 1.956 1.799
450 1.333
1.353
1.787
1.830
1.868
2.034
2.131
2.138
2.183
SB 2.079
1.834
1.861
1.885
550 1.369 2.173 2.222 2.111 1.906
1.385 1.931 2.258 2.138 1.923
650 1.398 1.958 2.244 2.291 2.162 1.939
700 1.409 1.981 2.275 2.311 2.185 1.951
pounds, minimum time is attained when the ration reaches 13 per cent
protein; the corresponding figure is 13 per cent for the weight interval
150-225 pounds. For a single ration over the entire weight interval,
6
minimum time is reached with a 15 per cent ration.
Returns Considering Feed Costs and Time of Marketing
Maximum daily gain or minimum time to attain a given marketing

weight is not a sufficient criterion for the most profitable hog ration. A
shorter production period which throws marketings into a higher price
period will lower profits if it increases the cost of feed more than the
increased profit from a higher selling price. On the other hand, a
quicker gain will lower profits if (1) it increases costs by more than
price or (2) it moves hog marketings back into a period of lower prices
while leaving costs unchanged (or higher)
6
While other rations give time as low as some of those mentioned, interest generally is
in the lowest protein rationwhich gets hogs to market in the shortest time period. The
number of days for a specified gain are rounded in Table 8.14, to the nearest day. The fol-
lowing rations give a shorter time period than any other ration:
34-75 lbs. 27.8 days 20.3 per cent protein
Table 8.14. Total Days to Attain Specified Gains (With Aureomycin)
Days Required Beyond Weaning To Take Pigs Between Weights of

Per Cent
Protein 34-225 lbs. 34-75 lbs. 75-150 lbs. 150-225 lbs.
In Ration (191 lbs. gam) (41 lbs. gain) (75 lbs. gain) (75 lbs. gain)
10 151 51 53 46
11 127 41 46 40
12 120 38 44 38
13 113 34 42 36
14 111 33 42 36
15 108 31 42 36
16 108 30 42 36
17 110 29 43 38
18 112 28 44 40
19 118 28 47 44
20 122 28 48 46
i
Since seasonal prices are affected by movements in the general

price level, the farmer should use price outlook material and balance
the economy of timeliness against the cost of the ration. H national eco-
nomic outlook suggests a break in prices, he may want to feed more
protein and rush his pigs to market before the price decline, even if
feed costs per 100 pounds of gain are increased. If continued inflation
and rising prices are in prospect, he will wish to hold down daily rate
of gain, particularly if a lower protein content ration results
in lower
feed costs.
Using the average prices for pork, corn and soybean oilmeal over
the period 1937-52 and without considering expectations of falling
or
increasing prices due to national inflation or deflation, the data of Ta-
ble 8.15 have been computed to show the relationship of
prices re-
ceived, time of gain, and feed cost to net return per hog.
These rfgfrg
are for fall pigs weaned on November 1 and November 15.
The following figures result for a farmer who makes his decision
for future marketings and the most profitable ration
on the basis of
corn and protein prices at weaning time; the average SBOM
to corn
price ratio on November 15 (for the period 1937-52)
was 2.1. Pigs
weaned on November 15 receiving a 16 per cent protein ration
require
the least time to attain a marketing weight of
225 pounds. The earlier
marketing date of March 3 would not, however, have
given the highest
price. A 14 per cent ration with a marketing
date of March 5 would
have given the highest price per hundredweight.
A 12 per cent protein
would have given the highest net return above the
cost of feed after
weaning.
For pigs weaned on November 1, a 16 per cent protein
ration again
wou d have given the most rapid gains and the earliest
marketing dates;
a 12 per cent ration would have given the lowest
cost for the gain after
weaning, and a 20 per cent ration would have
given the highest price for
pork. The 12 per cent ration would have
given the greatest return
above feed costs after weaning. Hence, it is
apparent that whether
Table 8.15. Days to Attain Market Weight of 225 Pounds, Price per Hundredweight and Costs and
Returns From Different Protein Rations. Based on 1937-52 Average Prices for
Pork at Marketing Date and Feed at Weaning Dates (With Aureomycin)
Days to Gain Net Income

>er Cent Pounds of Feed To 191 Pounds Average Gross Feed Over Feed
Protein Gain 191 Pounds* Over 34-225 Marketing Price Per Income Cost Costs Per
n Ration Pound Weight Date Hundred- Per Hog Per Hog* Hog After
Corn Protein Interval* weight Weaning*
pigs weaned November 1
10 705 34 150 Mar. 31 14.53 32.68 13.44 19.24

12 610 70 120 Mar. 1 |
14.57 32.79 12.99 19.80
14 561 107 j
110 Feb. 19 14.52 32.67 13.39 19.28
16 529 146 !
108 Feb. 17 14.37 32.33 14.16 18.17
18 504 189 112 Feb. 21 14.50 32.63 15.18 17.45
20 483 !
237 122 Mar. 3 14.71 33.11 16.46 16.65
pigs weaned November 15
10 705 34 150 Apr. 14 14.26 32.08 13.44 18.64

12 610 70 120 Mar. 15 14.61 32.86 12.99 19.87
14 561 107 110 Mar. 5 14.74 33.17 13.39 19.78
16 529 146 108 Mar. 3 14.71 33.11 14.16 18.95
18 504 189 112 Mar. 7 14.73 33.15 15.18 19.97
20 483 237 122 Mar. 17 14.61 32.86 16.46 16.40
These data are based on a single low-cost ration over the entire weight range.
jriority should be given to rations which maximize daily gains or to

hose which minimize feed costs depends upon the weaning date of the
rigs, and the feed price ratio in relation to likely pork prices at mar-
keting. Quite a range of rations can be used without changing feed
iosts per 100 pounds of gain.
Usually the farmer will want to consider different rations for hogs
>f different weights as suggested by the data in tables 8.9 and 8.10.
These rations then can be related to weaning date and time in the man-
Ler indicated in Table 8.14. (Table 8.15 uses the same ration through-
>ut the production period.) The farmer also has the opportunity to ad-
ust his weaning and marketing dates by selecting different dates for
seeding and farrowing.
The calculations in Table 8.15 are for average selling prices over
he period 1937-52. Table 8.16 not only emphasizes that in different
r
ears pigs weaned at the same time may need different rations, but
s also of some interest in appraising the risks involved in different
ations. Since low protein rations require a longer period for growing
nd fattening a market hog, it might be said that more risk and uncer-
ainty is involved. The longer period gives more time for prices to
hange and the further into the future the farmer must extend his price
xpectations, the less accurate he may be in gauging the likelihood of
Afferent price outcomes. An examination of Table 8.16 suggests that
his situation does hold true for hogs weaned on November 15 and fed a
0 per cent protein ration; this ration gives a marketing date with the
3 west price in 11 out of 14 years and the highest price in 2 years. In
his same sense there is a lower probability of obtaining the lowest

rice of the six marketing dates when higher protein rations are fed,
nd a slightly higher probability of obtaining the highest price. For the
Table 8.16. Number of Years in Which Prices Were Highest or Lowest

for Six Marketing Dates for Pigs Weaned November
15 and Fed Different Rations, 1937-52*
Number of Number of
Per Cent
Days Reach
to Marketing Years Highest Years Lowest
Protein
225 Pounds Date Price Paid Price Paid
In Ration
At This Time^ At This Time^
November 1 weaning
10 159 March 31 6 5
12 121 March 1 1 4
14 108 February 19 5 4
16 105 February 17 5 6
18 107 February 21 2 4
20 113 March 3 3 2
November 15 weaning
i
10 159 April 14 2 11
12 121 March 15 4 0
14 108 March 5 5 "1
16 105 March 3 4 3
18 107 March 7 5 1
20 113 March 17 4 0
*The war years 1945 and 1946 were omitted because the price paid was the same for
all marketing dates included here.
'Computed relative to the six marketing dates shown in column 3 and not in terms of
prices over the entire year.
earlier -weaning date of November 1, the marketing date for a 10 per

cent protein ration is 15 days earlier and the “price risk” is less pro-
nounced.
However, in 14 out of 16 years for pigs weaned November 15 the
least -cost ration would have resulted in greater profits. In other words
the price gain was greater than cost increase for the least -time ration
in only two years. The ration which gave the lowest cost of gain would,
of course, have differed between years.
LEAST -COST RATIONS AND OPTIMUM MARKETING WEIGHTS
In the end the farmer must make two decisions in livestock feeding:
(1) the ration to be fed for any one level of production or output (the
feed substitution problem) and (2) the amount of the best ration to feed
the animal and the marketing weight or level of production to be at-
tained. The appropriate procedure is to decide first the least -cost ra-
tion and then the amount to be fed (i.e., the marketing weight
or pro-
duction level)
Usually the farmer knows the current feed prices but must estimate
the future livestock price. In order to facilitate these decisions
in pork
.
production, tables have been computed for these purposes. They then
show the price per 100 pounds which must be received to "break even"
on feed costs if hogs are sold at various weights under different feed
7
prices.
These procedures can be used for determining either least-cost or
least -time rations.First, the farmer may decide on the amount of pro-
tein and corn which will allow the lowest feed costs per pound of gain.
He then can compute the cost per 100 pounds of this ration and use it
along with the hog price data to estimate the necessary selling price to
"break even." Or, he can select a least -time ration and apply the same
procedure
PRACTICAL APPLICATIONS
Previous sections have been concerned largely with estimation of

substitution coefficientswhich might have use in practical application.
A few tables have been presented which summarize the data and appli-
cation of principles to simple forms. For example, Table 8.10 could be
understood and used by most high school and college students and by
most commerical hog producers. However, since some readers may
complain that the algebra and calculus presented is too complex and
confusing to be used by farmers, we illustrate one further step in sim-
plification of findings and in application of economic principles.
Communication of Findings
The communication of results stemming from complex research

should not be confused with the research effort itself. Research and
communication are two entirely different steps in scientific and educa-
tional activity. In most cases, the two can be separated, although it is
certainly true that the research worker does not always excel in com-
municating results to the person who will use them. Often the research
worker has no specialist in education or communication methods at
hand and must complete the "translation step" by himself, if the find-
ings are to be put to use. But even in this case, he should consider the
two steps to be more or less separable. He should employ the research
techniques which are most efficient for the predictions at hand. Then,
he should use the communication media which are most effective in al-
lowing operational use of basic data and principles.
7
See tables 14 and 15 in the publication: Heady, Earl O. et al Replacement rates of
.
corn and soybean oilmeal in fortified rations for growing-fattening swine. Iowa Agr. Exp.
Sta. BuL 409. Ames. 1954.
Pork Costulator
It is known, of course, that few hog farmers

understand calculus or
the application of basic economic principles. They cannot go out to the
pig pen and equate the appropriate partial derivatives and price ratios,
to specify the optimum ration or marketing weight. However,
it is pos-
sible to develop mechanical aids to let a farmer make

calculations
and decisions as if he understood the basic mathematical and economic
principles. To illustrate this point, the pork costulator developed by
the authors is illustrated in Figure 8.9.
The costulator is a mechanical device constructed from three sheets
of heavy metal. The center or square sheet is about 12 inches square.
Disks of a slightly smaller diameter were cut from the other two
sheets, with one riveted to each side of the square piece. Prices of
corn were printed on the center square, near the perimeter of each
disk. Prices of protein supplement were printed at the edge of each
disk. A slit, through which is read the least-cost and least-time ra-
tions for particular price ratios, was cut in the disk on one side for 35-
75 pound and 75-150 pound hogs. A similar slit was provided for ISO-
225 pound hogs on the disk on the other side. The ration data is printed
on the center square. By matching the current protein price printed on
the edge of the disk with the current corn price printed on the center
square, the appropriate price ratio is specified. By glancing through
the slit for hogs of a particular weight, the producer can read off the
ration which minimizes feed costs per pound of gain, or the ration
which minimizes time required for marketing at 225 pounds. Certain
auxiliary information also is indicated and includes the components of
these rations, the range price ratios over which it would be profitable
to provide self -feeding and free choice of rations by hogs, rates of gain
and time to marketing. In effect, the farmer can use the costulator as
his “mathematical brain.” In matching up the feed prices and reading
off the appropriate information, he is accomplishing the equivalent of
(a) equating partial derivatives with price ratios and (b) simultaneously
solving a set of equations for magnitude of feed variables for attaining
least-cost or least-time rations. He need not know mathematics to
apply and use the principle.
The costulator in Figure 8.9 represents one practical mechanism
for encouraging use of appropriate principle and data. Others could be
suggested. For example, the price maps illustrated in Chapter 9 rep-
resent an alternative.
Figure 8.9. Pork Costulator derived from production functions for farmer use of
economic principles.
C H APTER 9
Damon V. Catron Pork Production Functions

Earl O, Heady
Dean E. McKee
Gordon Ashton
C.
and Substitution Coefficients
Vaughn C. Speer
for Hogs on Pasture
CHAPTER reports a study similar to the drylot experiments

T HIS
explained in Chapter 8, except that it refers to hogs produced on
pasture. Like the drylot study, the objectives of the pasture ex-
periment were to estimate (1) the production function, (2) the substitu-
tion rate between corn and soybean oilmeal at different points on the
production surface, (3) the least-cost ration for different soybean oil-
meal to corn price ratios, (4) the relationship between the rate of hog
gains and the input of corn and soybean oilmeal, and (5) the proportion
of the years in which a least- cost feeding system results in greater prof-
its than a least-time feeding system. Substitution between major classes
of feed such as corn and soybean oilmeal is possible mainly where the
rations are fortified with appropriate quantities of trace minerals (as
well as antibiotics in the case of drylot feeding). These fortifying ele-
ments have been included in the rations of this study.
DESCRIPTION OF THE EXPERIMENTS
Two experiments were conducted co-operatively by research per-

sonnel in swine nutrition and production economics to obtain the neces-
sary data. The first experiment, A. H. 597, was conducted during the
summer of 1953. The second experiment, A. H. 597A, was conducted
during the summer of 1954. Both experiments were conducted on an
alfalfa pasture. The data from the two experiments were combined for
the purposes of this study.
Both experiments were randomized complete block designs and in-
cluded 12 treatment combinations with three replications each. Treat-
ment combinations consisted of six rations, an antibiotic treatment and
an antibiotic check. The rations were: 8 per cent, 10 per cent, 12 per
cent, 14 per cent, 16 per cent, and 18 per cent protein. The antibiotic
treatment consisted of crystalline chlorotetracyline (aureomycin) fed at
the rate of 5 mg. per pound of ration. The rations were composed of
ground yellow corn and solvent- extracted soybean oilmeal fortified with
dicalcium phosphate, calcium carbonate, salt, trace minerals, and vita-
mins. Aside from corn and soybean oilmeal, other feed ingredients were
retained as a constant amount per 100 pounds of feed. The experimental
302
FUNCTIONS FOR HOGS ON PASTURE 303
unit was an individual hog, and each hog received the same ration
throughout the entire experiment. Forage intake was not measured, al-
though it is a resource in pork production just as are corn and soybean
oilmeal. However, farmers generally put their hogs on pasture and then
decide on the best ration when hogs already have access to free choice
of forage. Hence, the functions are estimated within this framework.
They do not allow evaluation directly of the quantitative effects of forage
but allow computation of least- cost rations for hogs with free choice of
pasture.
The hogs were fed individually in portable field units. Each field
unit consisted of three pens equipped with individual self-feeders and
waterers. The units were aligned side by side on pasture. They were
moved each Monday, Wednesday, and Friday during the experiment.
The original order of the units on the field was maintained at all times.
The pasture sward for experiment 597 was composed of a mixture
of alfalfa and bromegrass. Mower clipping was used to maintain a
maximum herbage height of about 8 inches; the pasture area was clipped
several times during the trial to prevent excessive growth. Moisture
was sufficient in both seasons so that the herbage remained of good
quality over the experimental period.
Treatment combinations and the pigs were randomly assigned to
pens within a block. Of the three replications, one included females,
and the other two included males. The hogs were weighed every second
week while they were on the experiment and were removed from the ex-
periment as each hog reached 200 pounds.
The breeding of the hogs used in experiment A. H. 597 was Duroc x
Poland China x Landrace x Duroc and Poland China x Landrace x Duroc.
A Poland China x Landrace x Duroc cross was used in experiment A. H.
597A. Thirty-six hogs were required for each experiment, a total of 72
hogs for both experiments.
ESTIMATION OF THE PRODUCTION FUNCTION
Two steps have been followed in estimating the production function.

First, three alternative types of functions have been fitted to all obser-
vations of the two experiments. These functions are denoted as over-all
functions . Second, each of the three types of functions have been fitted
to the observations on each of the six rations separately. The latter
functions are called individual ration functions . Interest is mainly in
the over-all functions; they express the relationship between hog gains
and the input of any one of many combinations of the feeds. Individual
ration functions express the relationship between hog gains and feed in-
put when feeds are held in fixed proportions. The input-output curves
for different rations varied in fixed proportions can be readily obtained
from the over-all function. Therefore, comparisons of the feed-gain
relationship estimated by the over- all function with that estimated from
the individual ration function provides a simple means of checking the
reliability of the over- all function.
304 FUNCTIONS FOR HOGS ON PASTURE
The production functions express total gain beyond weaning as a

function of total feed consumption beyond weaning. Experimental obser-
vations were taken on the consumption of feed and the amount of gain
over 2-week intervals. The interval observations were progressively
totaled over the entire feeding period to obtain a series of cumulative
summations of gain, corn consumption, and soybean oilmeal consump-
tion beyond weaning for each hog. The over-all production functions
were then fitted to the 72 series of observations. Each individual ration
function is fitted to 12 such series of observations. Analyses of variance
were made for both experiments and for two weight intervals. They in-
dicated no significant antibiotic effects for daily gain or feed consump-
tion under conditions given. For feed consumption per pound of gain,
antibiotic effects were significant only for the initial-to-200-pound
weight interval in experiment 597. Antibiotic effects upon feed per 100
pounds gain are significant at the 5 per cent probability level for only
the linear term for the initial-to-the- 200-pound weight interval. In
general, the analyses of variance do not support the hypothesis that the
antibiotic treatment and the check lots constitute separate populations.
Consequently, data from treatment and check lots are pooled for esti-
mation of the production function and each protein level includes obser-
vations from 12 hogs.
Autocorrelation
Fitting of the functions for the cumulative series again introduces

a
problem of autocorrelation. The different observations for each hog
are
not independent since (a) the second observation taken on a hog is
the
sum of the feed consumption and gain over the first and second 2-week
intervals, (b) the third observation is the sum of the feed
consumption
and gains in each of three 2-week periods, etc. Although the series of
observations taken on a hog is itself autocorrelated, it is independent
of
the series of observations taken on other hogs. Since the
over- all pro-
duction function is fitted to all observations in each series, the
auto-
correlation coefficient for the entire collection of data is likely
to have
a value greater than zero.
The presence of autocorrelation in the observations does not present
problems in predicting the relationship between the dependent and
the
independent variables, but it does introduce problems in
making tests of
significance. The
effect of autocorrelation is to reduce the number
of
effective observations to which the function is fitted.
In other words, the
number of degrees of freedom used for tests of significance of uncor-
related series is reduced when autocorrelation is present.
Procedures are available for approximating the effective
number of
degrees of freedom in autocorrelated series. However, the
necessity of
calculating the autocorrelation coefficient and
approximating the effec-
tive number of observations may be avoided by basing the tests of sig-
nificance on a minimum number of effective
observations to which the
series would be reduced by autocorrelation.
Since the observations taken on different animals are independent,

the minimum number of effective observations may be regarded as
equal to the number of hogs from which observations were taken. The
minimum number of effective observations is 72 for the over-all func-
tion and 12 for the individual ration functions. If the tests are signifi-
cant on the basis of the minimum number of effective observations, the
nullhypothesis may be rejected. If the tests are not significant, the
null hypothesis cannot be accepted without further testing. In the latter
case, the test must be conducted on the basis of the actual number of
observations used, disregarding autocorrelation for the moment. If the
test still is not significant at an acceptable probability level with the
greater number of degrees of freedom, the null hypothesis may then
1
be accepted.
Over- all Functions
In the functions which follow, C refers to pounds of corn, P refers

topounds of soybean oilmeal and Y refers to pounds of gain, all meas-
ured beyond weaning.
Quadratic:
2 2
(9.1) y = -1.7536 + .2988C + .9828P - .00003012C - .003880P - .0001684CP
Square root:
(9.2) Y = -17.4939 + .2472C + .03568P + 1.4249 Yc + 6.6133 Vp - .08138 Vc" Yv
Cobb-Douglas:
&426 ieo4
(9.3) Y = .5493C* P *
The quadratic function equation 9.1 explains 98.3 per cent of the
variance in hog gains (Table 9.1), the square root function explains 98.1
per cent, while the power function explains only 94.2 per cent of gain
variance. Using the minimum number of degrees of freedom, the linear
and squared terms of equation 9.1 are significant at the .01 and .05
probability levels. The cross-product term is acceptable at a proba-
bility level between .10 and .15. The linear term for P in equation 9.2
is significant only at a probability level greater than .30. Both terms
for the Cobb-Douglas function, equation 9.3, are significant at a .01
probability level.
1
The effective number of observations need to be approximated only if the tests are not
significant on the basis of theminimum effective number of observations but are significant
on the basis of the actual number of observations taken. For example, if the calculated “t”
for a regression coefficient in the over- all function were 2.616, the regression coefficient
would be significant at the .01 level of probability on the basis of 500 degrees of freedom.
On the basis of the minimum number of effective observations for the over- all function, 72,
the regression coefficient would not be significant. If the autocorrelation reduces the num-
ber of effective observations to less than 125, the null hypothesis would be accepted at the
.01 level of probability. If the effective number of observations is 126 or greater, the null
hypothesis would be accepted.
Table 9.1. Correlation Coefficients, Standard Errors and t Values for Over -All Functions.
Stan dar d Errors and t Values in Order Given in Equations
Standard Errors Value of t

Equation n R2
S b4 S t
s bx s ba s b3 fc6 tbi ^2 bs t
bs
1 521 .983 .0091 .0348 .00001 .00026 .00011 32.89* 28.25* 3.00* 15.06* 1.55*
2 521 .0146 .0353 .3760 .5107 .0339 16.89* 1.01§ 3.79* 12.95* 2.40^
3 !
521 !
.981
.942 .0213 .0072 68.78* 22.41* — —

*p < *01
f.05 >p> .01
+ .15 >p> .10
§p> ,30
The sum of the elasticities for corn and soybean oilmeal in the
power function is equal to 1.003, indicating slightly increasing returns
to proportional increases in the input of the two feeds. This relation-
ship appears unlikely in pork production, and, for the pasture data, the
function appears to overestimate gains for higher levels of feed inputs,
(This characteristic holds true only for the particular observations of
this study and is not a characteristic of the same function fitted to other
data.) The quadratic and the quadratic root functions express decreas-
ing returns to proportional increases in both feeds.
Comparisons Between Over- all Functions
Relationships among predictions from the three functions are shown

in Figure 9.1 for a 12 per cent protein ration. Similar estimates are
obtained from all three functions up to a feed input of about 250 pounds.
Beyond 250 pounds of feed, the curve estimated by the Cobb-Douglas
function rises above the curves estimated by the other functions. The
quadratic and quadratic root functions give very similar results through-
out the entire range of the curves. Below feed inputs of 350 pounds, the
curve for the quadratic function lies below the curve for the square root
function. Beyond feed inputs of 350 pounds, the positions of the two
curves are reversed.
The relationships between the three functions at other protein levels
are similar. At lower protein levels, the curves from the two quadratic-
type functions are more nearly linear and correspond more closely to
the estimates obtained from the Cobb-Douglas function. At higher pro-
tein levels, the curves for the quadratic -type functions have greater
curvature and fall away from the Cobb-Douglas curve more rapidly.
The quadratic functions produce curves that are most consistent with
the scatter diagrams at all protein levels.
Individual Ration Functions
Since the over- all functions have been fitted to all observations on
functions.
over-all
three
the
by
estimated
as
ration
protein
cent
per
12
for
curve
Growth
9.1.
Figure
9NINV3M GN0A39 NIV9 30 S91
[
307 ]
Table 9.2. Correlation Coefficients, Standard Errors of Regression Coefficients and

tValues for Individual Ration Functions
Standard Errors t Values

Equation R2 t
Sbl ®b2 bi *b2
2
Y = a + b.C + b s
,c
8 per cent ration . . . . . .965 .0288 .000048 10.06 .16

10 per cent ration . . . . . .983 .0187 .000031 17.94 1.30
12 per cent ration . . . . . .992 .0136 .000025 31.95 5.60
14 per cent ration . . . . . .991 .0160 .000032 31.27 7.43
16 per cent ration . . . . . .976 .0997 .000210 5.72 1.87
18 per cent ration . . . . . .983 .0232 .000052 25.28 8.08
Y = a + biC + b 2 Vc
8 per cent ration . . . . . .965 .0412 1.2257 7.77 .64
10 per cent ration . . . .
. . .982 .0268 .8159 10.35 1.26
12 per cent ration . . . . . .992 .0199 .5691 13.04 5.06
14 per cent ration . . . . . .991 .0242 .6614 9.20 6.78
16 per cent ration . . . . . .975 .0406 1.1010 3.13 6.54
18 per cent ration . . . . . .987 .0298 .7772 3.24 10.47
J3
it
3
8 per cent ration . . . . . .932 .0424 — 27.24
10 per cent ration . . . ,
. . .978 .0165 — 62.22 —
12 per cent ration . . . . . .947 .0256 — 41.63 —
14 per cent ration . . . . . .987 .0122 — 79.58
16 per cent ration . . . . . .939 .0252 -- 37.73 —
18 per cent ration . . . . . .971 .0154 — 56.19 —
the production surface, they might result in “abnormal” predictions for

individual rations. Single-variable functions express the result of a
single ration without encountering some of the “joint relationships” in-
herent in the over- all functions, and are, therefore, compared with the
over-all functions. These comparisons show that “spurious” predictions
do not arise from the over-ail functions. In these graphic comparisons,
corn alone is the independent variable. In any one ration, the ratio of
corn to soybean oilmeal is fixed: An increase in the consumption of
corn must be accompanied by a constant proportion of soybean oilmeal.
There is no necessity for measurements to include both feeds in the
individual ration function.
The individual ration functions, paralleling the three over-all func-
tions, with gain as the dependent variable and corn
as the independent
variable, are as follows:
Quadratic functions:
(9.4) 8 per cent protein ration Y = -5.102 + .290C + .000009C

(9.5) 10 per cent protein ration Y = -1.200 + .335C - .00004C
2
(9.6) 12 per cent protein ration Y = -3.062 + .433C - .0001C 2
(9.7) 14 per cent protein ration Y = -1.982 + .500C - .0002C 2

(9.8) 16 per cent protein ration Y « -1.664 + .570C - .0004C

2
(9.9) 18 per cent protein ration Y = 1.260 + .586C - .0004C

2
Square root functions:
(9.10) 8 per cent protein ration Y = - .189 +.320C- .784 Vc

(9.11) 10 per cent protein ration Y = - 5.854 + .278C + 1.028 Vc
(9.12) 12 per cent protein ration Y = -14.831 + .260C + 2.880 Vc
Cobb-Douglas functions:
(9.16) 8 per cent protein ration y = .me 1
1
(9.17) 10 per cent protein ration Y = .272C
1
*
!

*’
(9.21) 18 per cent protein ration Y = 1.000C
For all three equations, estimates for the 8 per cent ration show an
increasing marginal productivity of feed. The quadratic and square root
functions show a decreasing marginal productivity for rations with 10
per cent or more of protein. The Cobb-Douglas function shows increas-
2
ing marginal productivity through the 12 per cent protein ration. In-
creasing marginal feed productivity for low protein rations may be an
effect of pasture. Young pigs consume very little forage, but, as they
mature, they consume increasingly greater amounts. Hence, with the
low palatability of a low-protein ration, small pigs may obtain insuffi-
cient amounts of protein from forage. As they grow, however, forage
intake and hence gain per pound of concentrates may increase sharply,
even for low-protein rations. Forage then becomes a substitute source
of protein for hogs obtaining a small proportion of soybean oilmeal in the
concentrate ration. However, this substitution is possible mainly as the
hog grows. The tendency to substitute forage protein for concentrate
protein is less with rations high in protein because of their greater
palatability and nutritional “completeness.” This phenomena would not
have been expressed if feed value of forages could have been measured
and used in predictions.
Comparison of Over- all and Single -Variable Estimates

After examination of the various statistics for the three over- all
Only equation 9.17 of the single ration power functions has an elasticity which does not
2
differ significantly from one at a probability level of 5 per cent.

functions, the quadratic equation, equation 9.1, was selected as the best
estimator for the production surface. The Cobb-Douglas over-all func-
tion, equation 9.3, was eliminated because of the smaller proportion of
the gain variance explained and the greater algebraic restrictions im-
posed by its logarithmic form. The square root over-all function,
equation 9.2, provides estimates highly similar to the quadratic func-
tion. However, since it explains a slightly lower portion of variance in
gains and has a relatively greater standard error for the P terms, it
was rejected in favor of the quadratic function. 3 Hence, the text com-
parisons which follow compare estimates of single-line, input-output
curves derived from over-all and single -variable equations for the
latter functions.
"Growth curves” for six rations estimated by the single- variable
and the over-all quadratic functions are shown in Figures 9.2 to 9.7.
Similar curves are obtained from the estimates of the two types of
quadratic functions. At the 8 per cent, 10 per cent, and 12 per cent
protein levels, the curve estimated from the over-all quadratic function
is almost identical to the curve estimated from the function fitted to
each ration separately. The curves for the 14 per cent, 16 per cent,
and 18 per cent protein levels are similar up to feed inputs of about 500
pounds. Beyond this, the curve estimated by the over- all function has
slightly greater slope than the curve estimated by the individual ration
function for 14 per cent and 16 per cent protein levels; the reverse is
true for the 18 per cent protein level.
Interval Estimates from Cobb-Douglas
Since the over-all Cobb-Douglas equation appeared less satisfactory

than other functions for the over-all surface, an attempt was made
to
predict three interval functions; and hence, to eliminate “overestimates”
of gain at high feed inputs. As mentioned in Chapter
8, a reason for
this attempt was to provide “average rations” for
three gain intervals
and to conform with the normal practice of changing rations two
or three
times during the growing- fattening period. The estimated relations
are:
Weaning 9087 27M
(9.22) to 75 pounds: Y = .3350C' P'
(9.23) 75 pounds to 150 pounds: 8072 wos
Y = .6543C* P'
9875
(9.24) 150 pounds to 200 pounds: Y = .3127C* p* 0270
The three equations have R 2
values of .82, .89, and .73, respectively.
All regression coefficients are significant at a
1 per cent level of prob-
ability, except for the coefficient of soybean
oilmeal in the last equation.
It is significant at a 20 per cent
level. The elasticity of production for
3
While the regression coefficient for the cross-product term was significant at
a prob-
ability level greater than .10 but less than .15, it has
been retained in the over-all quadratic
function (9.1) since it adds some precision to estimates.
o 100 200 300 400 300 600 700

POUNDS PEED PER PIG BEYOND WEANING POUNDS OF FEED PER PIG BEYOND WEANING
Figure 9.2. Growth curves for the 8 per cent protein Figure 9.5. Growth curves for the 14 per cent protein
ration estimated from individual ration ration estimated from the individual ration
and the several quadratic functions. and over-all quadratic functions.
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
POUNDS PEED PER PIG BEYOND WEANING BEYOND WEANING
POUNDS OF FEED PER PIG
ration estimated from the individual ration ration estimated from the individual ration
and over-all quadratic functions. and over-all quadratic functions.
100 200 300 400 500 600 700

0 100 200 300 400 900 600 700
POUNOS FEED PER PIG BEYOND WEANING POUNDS OF FEED PER PIG BEYOND WEANING
ration estimated from the individual ration ration estimated from the individual ration
and over-all quadratic functions. and over-ail quadratic functions.
”
soybean oilmeal declines, from lower weight to higher weights, as ex-

pected. However, the elasticity of production for corn falls and then
rises, instead of consistently rising from low weights to high weights
as expected. The sum of the elasticities of production for the two feeds
are 1.17 for the first interval, .95 for the second, and 1.01 for the third.
This relationship, increasing feed productivity followed by decreasing
feed productivity and then increasing feed productivity, is inconsistent
with known biological conditions.
While the interval Cobb-Douglas approach gave satisfactory results
”
in estimating “average rations to be fed over a gain interval in the
drylot study, it does not appear to be appropriate for the pasture data.
The quadratic over-all function again appears to be the best choice
among the various alternative functions examined, although modifica-
tions must be made in its use for determining rations to be used as
“averages over gain intervals.
PRODUCTION SURFACE ESTIMATES
The pork production surface for corn and soybean oilmeal, based
on equation 9.1, is shown in Figure 9.8. Consumption of corn and soybear
oilmeal is measured by the vertical distance of the surface. The gains
in hog weight, between weaning and market weight, follow a path over
the face of the surface. The location of the path upon the surface is de-
termined by the ration fed. A ration consists of a fixed combination of
corn and soybean oilmeal and represents a vertical slice of the surface
through the origin. The ration is represented by a straight line drawn
in the horizontal or feed plane of the surface passing through the origin
of the graph.
Marginal Productivity of Feeds in Fixed Proportions
Predicted total gains beyond weaning and the marginal productivity

of feed at several levels of total feed
consumption are shown in Table 9.3
for six rations. Marginal productivities are calculated on the basis of
a proportional increase in the consumption of both corn and soybean
oilmeal. In other words, the marginal productivities show the increase
in hog weight resulting from a 1-pound increase in the quantity of the
ration consumed.
The change requirements as the hog approaches maturity
in nutrient
is partly reflected in the total
gains for each ration (Table 9.3). Fifty
pounds of an 8 per cent protein ration produce 12.5 pounds of gain; 50
pounds of an 18 per cent protein ration produce 20.6 pounds of gain.
The higher protein ration supplies more of the protein necessary for
tissue building and growth at low weights. However, 700 pounds of the 8
per cent protein ration produce 185.8 pounds of gain while the same
amount of 18 per cent protein ration produces only 176 pounds of gain.
Figure 9.8. Production surface for hogs on pasture (lines or contours on surface
are gain isoquants).
The high protein ration does not supply a sufficient amount of carbohy-
drate for production of fat at later stages of growth. These differences
are brought out even more clearly by the marginal productivity figures.
Up to a total feed intake of 250 pounds, marginal feed productivity is
highest with an 18 per cent ration. At the 300-pound feed level, an ad-
ditional pound of the 16 per cent ration has the same marginal produc-
tivity as an 18 per cent ration. The marginal productivity of the 18 per
cent protein ration declines from .61 at the 50-pound feed level to .06
at the 700-pound feed level because of the decline in the protein re-
quirements as the hog matures.
Iso- Product Contours
Figure 9.9, a contour or pork isoquant map corresponding to the

production surface in Figure 9.8, is estimated from the quadratic
of a Pound of Ration
Table 9.3. Total Gain Beyond Weaning and Marginal Productivities
for Various Protein Rations
Total Gain Beyond Weaning Marginal Productivity of Feed*

Pounds
ofFeed m the ration Per cent protein in the ration
Per cent protein
Consumed
10 12 14 16 18 8 10 12 14 16 18
Beyond Weaning 8
50 12.54 13.67 15.76 17.30 18.93 20.64 .296 .334 .399 .461 .529 .611
100 !
26.70 28.90 32.91 35.72 38.64 41.55 .293 .330 .390 .447 .503 .568
150 40.72 43.96 49.70 53.54 57.36 60.96 .290 .326 .382 .433 .477 .526
200 54.60 58.84 66.13 70.74 75.10 78.88 .287 .322 .374 .418 .452 .484
250 68.34 73.54 82.20 87.32 91.87 95.31 .284 .318 .365 .405 .426 .422
300 81.95 88.06 97.90 103.30 107.65 110.24 .282 .314 .357 .390 .400 .400
350 95.41 102.39 113.25 118.65 122.45 123.68 .279 .311 .349 .376 .375 .358
400 108.74 116.55 128.24 133.38 136.27 135.63 .276 .309 .340 .362 .349 .316
450 121.92 130.52 142.86 147.50 149.11 146.09 .273 .303 .332 .348 .323 .274
500 134.97 144.32 157.13 161.01 160.97 155.06 .270 .299 .324 .333 .298 .232
'
550 147.88 157.93 171.03 173.90 171.85 162.53 .267 .295 .316 .319 .272 .190
600 160.65 171.36 184.57 186.18 181.75 168.51 .264 .291 .307 .305 .246 .148
650 173.29 184.61 197.76 197.84 190.67 173.00 .261 .287 .299 .291 .221 .105
700 185.78 197.68 210.58 208.88 198.61 176.00 .258 .283 .291 .276 .195 .063
*Added gain resulting from an added pound of ration. All figures predicted as derivatives from equa-
tion 9. 1.
equation 9.1, for 26, 76, and 141 pounds of gain beyond weaning. The
isoquant equation is:
2
(9.25) C = 4960.36 - 2.7961P ± (-16,600.2656) [-.00000044P
5
+ .00001774P + .08907733 - .00012048Y]* .
The quantity of corn and soybean oilmeal required to produce 26,

76, and 141 pounds of gain beyond weaning also has been determined
from the single- variable equations 9.4 through 9.9 for each of the six
rations. These feed quantities have been plotted in Figure 9.9. (In
every instance, the quantities estimated from the individual ration func-
tions fall very close to the contours estimated from the over- all equa-
tion.) The close agreement between the estimates from the individual
ration functions and the over-all function is further proof that, the over-
all equation provides reliable predictions of the relationship expressed
within the experimental data.
The experiments conducted did not include rations beyond the 18
per cent protein level. Therefore, the portion of the iso-product con-
tours lying below the 18 per cent protein ration line is an extrapolation
beyond the range of the data. The present study provides no informa-
tion on the shape of the contours in that section of the production sur-
face. However, with increasing levels of protein, the contours should
flatten out and eventually approach a zero slope.
LBS. OF SBOM
Figure 9.9. Iso-product contours for pasture fed swine.
Rates of Substitution
The rate at which soybean oilmeal substitutes for corn in the hog
ration at a given level of output is indicated by the slope at a particular
point on the iso-product contour. The iso-product contours in Figure
9.9 are curved. Consequently, the rate of substitution of soybean oilmeal
for corn declines as the ration includes a greater percentage of protein.
Marginal rates of substitution of soybean oilmeal for corn can be
computed from equation 9.26.
J5C = -.982769 + .Q07760P + .000168C

(9 26)
V
'
6P -.298812 + .000060C + .000168P
Table 9.4. Corn and Soybean Oilmeal Quantities and Substitution Rates Along the 26-, 76-, and
141 -Pound Gain Isoquants Derived from Equations 9.25 and 9.26
;
26 Pounds of Gam 76 Pounds of Gain 141 Pounds of Gam

Feed Feed Feed
Proportion
Per Cent required Marginal required Marginal required Marginal
of SBOM Protein t rate of rate of rate of
m Ration* corn SBOM substitution I 1
corn SBOM substitution t
corn SBOM
substitution
(lbs.) (lbs.) (lbs.) (lbs.) (lbs.) (lbs.)
.02 8.7 88.0 1.76 3.25 251.1 5.02 3.19 472.8 9.46 3.09
.04 9.4 82.9 3.32 3.22 237.1 9.48 3.07 447.6 17.90 2.86
.06 10.0 78.5 4.71 3.18 225.0 13.50 2.97 426.5 25.59 2.65
.08 10.6 74.5 5.96 3.15 214.3 17.15 2.87 408.4 32.67 2.46
.10 11.2 71.0 7.10 3.12 204.9 20.49 2.79 392.6 39.28 2.28
.12 11.8 67.8 8.13 3.10 196.4 23.57 2.71 379.1 45.49 2.11
.14 12.4 64.8 9.08 3.07 188.7 26.42 2.64 367.1 51.40 1.95
.16 12.9 62.2 9.95 3.05 181.8 29.09 2.57 356.6 57.05 1.79
.18 13.5 59.7 10.75 3.03 175.5 31.59 2.50 347.2 62.50 1.64
.20 14.0 57.5 11.50 3.01 169.7 33.94 2.44 338.9 67.78 1.50
.22 14.5 55.4 12.19 2.99 164.3 36.15 2.38 331.6 72.94 1.35
.24 14.9 53.5 12.83 2.98 159.4 38.25 2.33 325.0 78.01 1.21
.26 15.4 51.7 13.44 2.96 154.8 40.24 2.28 319.3 83.02 1.07
.28 |
15.8 50.0 14.00 2.95 150.5 42.14 2.23 314.3 88.01 .93
.30 16.3 48.4 14.53 2.94 146.5 43.95 2.18 310.0 93.01 .79
.32 16.7 47.0 15.04 2.92 142.8 45.69 2.14 306.4 98.05 .65
.34 17.1 45.6 15.51 2.91 139.3 47.35 2.10 303.5 103.18 .50
.36 17.5 44.3 15.95 2.90 135.9 48.94 2.05 301.3 108.45 .35
.38 17.9 43.1 16.37 2.89 132.8 50.47 2.01 299.8 113.93 .18
.40 18.2 41.9 16.77 2.88 129.9 51.95 1.98 299.2 119.68 .01
The figures show the pounds of soybean oilmeal for each pound of corn. Hence, the figure .20 refers
to .2 pounds of soybean oilmeal for each 1 pound of corn.
tBased upon a protein content of 45 per cent for soybean oilmeal and 8.2 per cent for corn.
fThe negative signs have been omitted from the substitution rates. The substitution ratios are the
derivatives in equation 9.26. The feed combinations for specified gams have been derived from
equation 9.25.
Table 9.4 includes prediction of the pork isoquants and the marginal
rates of substitution associated with them. With a corn to SBOM ratio
of .20, a 14 per cent protein ration, 57.5 pounds of corn, and 11.5 pounds
of soybean oilmeal are required to produce 26 pounds of gain. The rate
of substitution on the 26-pound gain contour with a 14 per cent protein
ration is 3.01 (i.e., a pound of soybean oilmeal replaces 3.01 pounds of
corn at the particular point on the 26-pound contour). For the same
ration, 169.7 pounds of corn and 33.94 pounds of soybean oilmeal are
required to produce a 76-pound gain. However, the quantity of corn re-
placed by a 1-pound increase in soybean oilmeal drops to 2.44 pounds
for this level of gain. For a gain of 141 pounds, 338.9 pounds of corn
and 67.78 pounds of soybean oilmeal are required and the rate of sub-
stitution drops to 1.50. Hence, the substitution rate and relative feed
value of soybean oilmeal declines as the hog increases in weight. While
this point has been illustrated for a 14 per cent ration only, it also holds
true for rations containing other percentages of protein.
Changes in Substitution Rates for a Given Gain
Not only do the substitution rates for protein decline as the hog at-
tains greater weight, but also they decline as the proportion of protein
increases for growth to a given weight. For example, substitution rates

vary from 3.25 to 2.88 over the range for which the 26-pound gain con-
tour has been predicted. In a ration containing .02 pounds of soybean
oilmeal per pound of corn, 1 pound of soybean oilmeal replaces 3.25
pounds of corn. A pound of soybean oilmeal replaces only 2.88 pounds
of corn at the same gain level in a ration with .40 pounds of soybean
oilmeal per pound of corn. One pound of the former replaces only 2.99
pounds of the latter when the ratio of soybean oilmeal is .22.
For the 76-pound gain contour, the marginal rate of substitution
varies from 3.19 with soybean oilmeal to corn ratio of .02 to 1.98 with
a soybean oilmeal to corn ratio of .40. The rate of substitution declines
much more rapidly along the 76-pound gain contour than along the 26-
pound gain contour because of the greater hog weight. The range in
magnitude of substitution rates is even greater along the 141-pound
gain contour. A pound of soybean oilmeal replaces 3.09 pounds of corn
with a soybean oilmeal to corn ratio of .20, but only .01 pounds of corn
with a ratio of .40.
The production function, equation 9.1, predicts the gain resulting
from various total amounts of feed consumed beyond weaning. Hence,
each contour, such as in Figure 9.9 and Table 9.4, is derived with refer-
ence to the origin or weaning weight. The predictions suppose that the
quantities of corn and soybean oilmeal specified by the co-ordinates of
a point on a contour are fed in that proportion from weaning to the level
of gain represented by the contour.
LEAST- COST RATIONS
The least-cost ration can be defined by equation 9.26 where the

derivative or marginal rate of substitution of corn for soybean oilmeal
is equated to the price ratio of soybean oilmeal and corn. However,
least-cost rations determined by use of substitution rates estimated
from the over- all production function must be carefully interpreted.
The hogs in these experiments were each fed a constant ration throughf-
out the entire course of the experiments. The production function
therefore expresses, under a constant ration system of feeding, the re-
lationship of total weight gain to total consumption of corn and soybean
oilmeal from weaning weight. The iso-product contours (Figure 9.10)
derived from the over-all production function, therefore, show the pos-
sible combinations of corn and soybean oilmeal to produce various
levels of gain under a single ration technique of feeding. The ration
which is determined by equating the substitution rate on these contours
to the feed price ratio gives the total quantity of corn and soybean oil-
meal which will produce the amount of gain represented by the respec-
tive contour and not over-all gain contours, with the lowest outlay for
,
the feed under a system of feeding a single ration throughout the feeding
period. For example, point G on the 225-pound contour in Figure 9.10
is the locus, on that contour, where soybean oilmeal substitutes for corn
at the rate of 2.5.Therefore, the co-ordinates of point G are the quan-

tities ofcorn and soybean oilmeal which will produce 194 pounds of
gain beyond weaning under a constant ration system of feeding. With a
soybean oilmeal price of 2.5 times the price of corn, corn and soybean
oilmeal would be fed, from weaning to the 225-weight level, in the pro-
portions represented by the line OG.
POUNDS OF SOYBEAN OILMEAL

Figure 9. 10. Expansion path of the least cost ration with a price ratio of 2. 5,
2.25,and 2.0.
A single ration fed throughout the entire feeding period is one pos-
sible system to follow; but it is obvious from the relationships shown
in Figure 9.10 that this system does not result in the lowest possible
feed cost Feed costs can be further reduced by adjusting the propor-
tion of corn and soybean oilmeal fed at intermediate points throughout
the feeding period. The line AFG in Figure 9.10 is an isocline joining
all points on successive contours having a slope of 2.5 (i.e., a substitu-
tion rate of 2.5). Comparing line AFG to line OG, it is obvious that be-
low 194 pounds least-cost gains are not attained if the same ration is
fed throughout the entire feeding period. One hundred and nineteen
pounds of gain can be produced at lower cost by feeding corn and soy-
bean oilmeal in the proportions represented by the feed quantity at
point F. This ration has a higher proportion of soybean oilmeal than
the ration represented by the line OG. Similarly, the ration for pro-
ducing 44 pounds of gain has a higher proportion of soybean oilmeal
than the ration for producing 119 pounds of gain under a single ration
feeding system. The rate of substitution between corn and soybean oil-
meal is changing continuously as the hog gains in weight. Consequently,
if feed costs are to be minimized, a different ration should be fed for
each successive pound of gain produced.
Particular Interpretation of Isoclines
The isocline equation is:
/q .299K + ,Q077P - .00Q16PK - .98

C -
.0006K - .00017
The isoclines in Figure 9.10 might appear to be interpreted as indicat-

ing that the least- cost ration toreach 75 pounds includes the amount of
protein shown at F, with some of this taken away from the pig as he
attained 119 and then 225 pounds. However, this is not the case. Point
A shows only the feed combination, fed as a single ration, which would
be used to attain minimum cost if the pig were to be taken only to 75
pounds; point F shows the least- cost feed combination if the pig were to
be fed to exactly 150 pounds; point G shows the feed combination for
lowest cost if the hog were fed a single ration and taken to exactly 225
pounds. These isoclines do not directly show how the corn and soybean
oilmeal proportion are to be adjusted throughout the period from wean-
ing to market weight in order to minimize feed cost per each successive
pound of gain produced. The isoclines refer only to a constant ration
feeding system, to the gain level indicated. They are, however, indica-
tive of the necessity for feeding a lower protein ration for each succes-
sive pound of gain produced.
Generally, isoclines have a positive slope. In Figure 9.10 they ap-
pear to be vertical, or to have a slight negative slope. This phenomenon
arises mainly because of the nature of the experiment and measure-
ments. The quantities measured in the feed plane are corn and soybean
oilmeaL The third feed, forage, is not measured. If the protein in

forage were added to that in soybean oilmeal, the isoclines would have
a positive slope. Small pigs eat very little forage because their diges-
tive organs cannot handle it. However, as hogs progress in weight, they
can and do consume much more forage relative to concentrates. Thus
for a 225-pound hog, the feed equivalent of soybean oilmeal in forage
would, if added to the soybean oilmeal measured in the study, fall at a
point to the right of G in Figure 9.10.
In a study designed to relate the gain surface physically with feed
input, forage should be measured and introduced into the production
function. The current study did not, however, have this objective. It
was designed to allow specification of least- cost rations under condi-
tions representing the environment in which most farmers make their
decisions. Most farmers turn their pigs on pasture as a disease con-
trol precaution, as well as to obtain some feed advantage. Yet hog
pasture usually includes an abundance of forage and no attempt is made
to fully utilize it in matching costs of forage against concentrate feeds.
The farmer is concerned, given an ample supply of forage and the quan-
tity of it consumed when different concentrate rations are fed, of bal-
ancing corn and protein supplement feeds in a manner to minimize con-
centrate costs.
Interval Rations
A fixed ratio of corn and soybean oilmeal fed over an interval of

gains does not result in the lowest possible feed costs for the entire
gain interval because substitution rates change continuously as the hog
increases in weight. Therefore, if feed costs are to be minimized, the
proportions in which the corn and soybean oilmeal are fed should be
changed for each unit of gain produced (i.e., should follow an isocline).
In practice, it is impossible to make such extremely small changes in
the ration. To adjust the ration for gains even as small as a pound, the
hogs would have to be fed individually and the rations changed daily.
Farmers are concerned with the least-cost ration for rather wide in-
tervals of gain. From a practical standpoint, they may consider chang-
ing the ration only two or three times in the course of the entire feeding
period from weaning to market weight.
Hence, in providing practical figures for farmer recommendations,
the production surface has been divided into three weight intervals.
The three weight intervals are: weaning to 75 pounds liveweight, 75 to
150 pounds liveweight and 150 to 225 pounds liveweight (i.e., the total
weight contours shown in Figure 9.10). For a given ratio of the feed
prices, a constant ration is selected for each interval. The ration
selected should produce the gain at a lower feed cost than any other
constant ratio n fed over the same interval. 4
^Greater refinement in the least- cost feeding system can be achieved by dividing the
surface into a greater number of weight intervals. The ration then can be altered more fre-
quently over the total production period.
Ratios
Price
Corn
to
Oilmeal
Soybean
Various
for
Pasture
Alfalfa
on
Rations
Least-Cost
9.5.
Table
[ 321 ]
The computed least-cost ration for each interval is given in Table

9.5 for a series of soybean oilmeal to corn price ratios. These rations
were approximated from the over- all quadratic production function in
this manner: A series of ration lines was projected through the surface
from weaning to a liveweight of 225 pounds. The total feed requirements
beyond weaning then were computed for producing a 75-, 150-, and 225-
pound hog along each of the ration lines. In other words, the various
quantities of corn and soybean oilmeal which can be used for producing
44, 119, and 194 pounds of gain beyond weaning were determined. The
difference for a particular ration in respect to protein percentage, be-
tween the total feed requirements at the beginning and the end of each
weight interval was used as the feed requirements for the particular
interval. The least-cost ration for the interval is then determined by
summing the value of corn and soybean oilmeal, for the numerous ra-
tions. While the procedure is an approximation (in contrast to the
equation of the derivative of the isoquant equation with the price ratio),
it gives estimates of the least- cost ration, accurate within a few tenths
of a per cent in protein and sufficiently accurate for practical uses.
Indication of Rations
In Table 9.5, least-cost rations can be determined as follows: With

a price of 4.05 cents per pound for soybean oilmeal and 1.5 cents per
pound for corn, the price ratio is 2.7. At this price ratio, the least- cost
ration over the weaning to the 75-pound interval is 15.8 per cent protein.
The ration will include 85 pounds of corn and 24 pounds of soybean oil-
meal for this amount of gain, plus the 2.5 pounds of the minerals indi-
cated on page 302 for each 100 pounds of feed. For gains in the 75-150
pound interval for a price of $1.12 per bushel for corn and $4.00 per
hundredweight for soybean oilmeal, a price ratio of 2.0 on a per-pound
basis, the least-cost ration includes 204 pounds of corn and 23 pounds
of soybean oilmeal, plus 2.5 pounds of minerals per 100 pounds of feed.
With 11.7 per cent protein, growth over this gain interval requires 43 days
The described above procedure for finding the least- cost ration can
be contrasted to other concepts of feeding in Figure 9.10. With a price
ratio of 2.5, the rations for each of the three weight intervals would be
represented by line OABCDE. From weaning to 75 pounds liveweight,
corn and soybean oilmeal would be fed in the proportion represented by
the line OA. At the 75-pound contour, a shift is made along the contour
to point B. From 75 to 150 pounds, corn and soybean oilmeal would be
fed in the proportion represented by BC. The ration is shifted again at
the 150-pound contour to the proportions represented by DE. The line
OABCDE might be called the * practical expansion path* of rations for a
price ratio of 2,5. It represents a compromise between feeding of a dif-
ferent ration for each pound of gain (line AG) and feeding the same ra-
tion over the entire feeding period (line OG). The first alternative is
impractical while the second does not minimize feed costs.
It can be seen by comparing the line OG with line O ABODE that the
constant ration over the entire production period results in underfeed-
ing of soybean oilmeal throughout the first and second weight intervals
and overfeeding of soybean oilmeal throughout the third weight interval.
Cost of feed is not minimized in any of the three weight intervals by
feeding the constant ration. The same condition may hold true for
OABCDE within each weight interval.
Least-Cost Graph
Figure 9.11, based on the rations in Table 9.5, provides a graph for
PRICE PER 100 LBS. OF SOYBEAN

OILMEAL ($)
Figure 9.11. Least-cost ration graph.

calculation of least- cost rations under different prices for corn and
soybean oilmeal. The series of iso-price ratio lines in Figure 9.11
show allcombinations of corn and soybean oilmeal prices giving the
same price ratio. Rather than to consider an infinite series of price
ratio lines with minute changes in the proportions of corn and soybean
oilmeal in the ration, only eight lines have been drawn. This procedure
amounts to assuming that the gain isoquant is made up of a series of
linear segments (rather than of continuous points on a smooth curve).
The least- cost ration for a given set of prices for corn and soybean
oilmeal is found by reading up the corn axis of Figure 9.11 to the given
price of corn, and then reading across in a horizontal direction until a
point is reached directly above the given price of soybean oilmeal. The
area of the graph in which the point lies determines which of the rations
in Table 9.6 is the least- cost ration to be fed for the gain intervals of
75-150 and 150-225 for growing -fattening hogs on alfalfa pasture. For
example, if the price of corn is $1.40 per bushel and the price of soy-
bean oilmeal is $6.50 per hundredweight, these prices form the co-
ordinates of point X on the graph. Point X falls in area G. The least-
cost ration for hogs between 75 and 150 pounds liveweight is given
opposite line G in section I of Table 9.6. For hogs between 150 and 225
pounds liveweight, the least- cost ration is given opposite line G in sec-
tion II. The rations in Table 9.6 are in terms of feed requirements per
hundred pounds of gain rather than in terms of the quantity of feed re-
quired for a single hog to produce the amount of gain for each weight
interval.
Table 9.6. Least-Cost Rations for Pasture Fed Hogs in Terms of Feed
Per 100 Pounds of Gain
Per cent Average Number

Corn SBOM
protein daily gain of days
I. 75 to 150 pounds
A 232 56 14.9 1.74 43
B 237 51 14.3 1.75 43
C 243 48 13.9 1.76 43
D 261 36 12.4 1.75 43
E 279 27 11.0 1.71 44
F 291 21 10.5 1.68 44
G 297 19 10.1 1.64 46
H 315 12 9,3 1.54 48
I 323 11 9.1 1.52 49
H. 150 to 225 pounds

A 309 31 11.2 2.33 32
B 317 25 10.6 2.33 32
C 321 21 10.2 2.31 33
D 331 17 9.8 2.29 33
E 332 16 9.6 2.27 34
F 335 15 9.5 2.25 34
G 339 13 9.4 2.23 34
H 341 12 9.2 2.20 34
I 345 11 9.0 2.18 34
RATE OF GAIN
To allow prediction of the effect of rations on rate of gain, two types

of functions are examined as alternatives for expressing the relationship
between the inputs of corn (C) and soybean oilmeal (P) and the number
of days (T) required to consume various quantities of the two feeds. The
two functions are the quadratic, equation 9.28, and the square root, equa-
tion 9.29. Each function has been fitted over the observations from all
six rations in both experiments.
(9.28) T = 4.2477 + .4414C - .3673P - ,0003C

S 2
+ .0047P - .0010CP
(9.29) T = -23.0421 - .0064C + .5304P + 7.9090 V'c - 4.0347 Yp - .2120 VC V'p
The statistics for the two functions are presented in Table 9.7.
Equation 9.29, the quadratic root, explains only a slightly greater pro-
portion of the variation in the dependent variable, (T), than does equa-
tion 9.28. Equation 9.28 gives a relation showing time as a maximum
at a level of feed consumption within the range of experimental obser-
vations. For total time to reach a maximum would mean that the hog
would have to “die off” and cease feed intake, an unrealistic sit uati on.
It appears more logical that the slope of the total time function should
fall off rapidly at low total feed input and approach linearity as total
feed consumption reaches a high level and the hog approaches maturity.
A mature hog has nearly a constant daily feed intake upon reaching
maturity. Conceivably, the hog could live several years, with total
time continuing to increase with age and continued feed consumption.
Table 9.7. Correlation Coefficients, Standard Errors, and t Values for Time
Functions in Order of Regression Coefficients in Equations
9.28 and 9.29
Standard Errors t Values *

Equation R2 S S S
S S tbi t
b2 tbs tbs
bi b2 b3 bi b5 tfcH
9.28 .912 .0141 .0539 .000027 .00040 .00018 31.35 6.81 12.04 11.83 5.60
9.29 .918 .0021 .0503 .5358 .7283 .0483 3.05 10.55 14.76 5.54 4.39
*A11 1 values exceed those for a probability level of .05 or less.
Equation 9.29, the square root function, more nearly allows the lat-
ter conditions. The relations obtained from the two types of equations
are plotted in Figures 9.12 and 9.13 for two rations. In terms of com-
parisons, equation 9.29 has been used as the basis for estimating rate
of gain.
Estimates of the total time required to consume various quantities
of feed for six different rations are shown in Table 9.8. At low levels
of feed input, the least time required to consume a given quantity of
Figure 9.12. Time relation for 10 per cent

protein ration.
Quodrt tic
Quo dr he ftoct
....
. ^
0 100 200 300 400 500 600 700

LBS OF FEED CONSUMED BEYOND WEANING
Figure 9.13. Time relation for 12 per cent

protein ration.
feed is obtained with the 18 per cent protein ration. As the pigs become
older and heavier, the advantage of the extremely high protein ration
becomes increasingly smaller and finally gives slower gains than a
ration with somewhat less protein.
An equation of daily rates of gain can be expressed as in equation
9.30 where D is gain per day, T (the equation in the denominator) is total
time to consume a given amount of feed, Y (the equation in the numerator)
is gain forthcoming from the same feed, and C and P refer
to corn and
soybean oilmeal consumption per pig.
+ .2988C + .9828P 8 - 2
(9.30) D = - .00003C .0039P - .0Q017 CP
-23.0421 - .0064C + .5304P + 7.909 Vc - 4.035 V?" - .212 Vc Yv
The average daily rate of gain, between 50-pound feed increments,

of the 18 per cent protein ration rises up to a total feed
input of 250
pounds (Table 9.9). Between inputs of 200 and 250 pounds, a maximum
average daily rate of gain of 1.685 pounds is reached. Beyond the 250-
pound level of feed consumption, each additional 50-pound increment of
the ration results in lower daily gains. The 16 per cent protein
ration
produces gains at a slower rate than the 18 per cent protein ration up to
Table 9.8. Total Time Required to Consume a Given Amount of Feed

in Rations of Various Protein Levels
Total Days To Consume Specified Feed Quantity for

Pounds Protein Percentages of
of Feed
10 12 14 16 18
50 31.5 23.8 19.5 16.5 14.0 11.7
100 53.9 42.4 36.6 32.8 29.9 27.2
150 71.0 56.5 49.6 45.3 42.3 39.7
200 85.4 68.1 60.4 55.9 52.9 50.5
250 98.0 78.2 69.8 65.2 62.3 60.2

300 109.4 87.2 78.3 73.6 70.9 69.2
350 119.9 95.4 86.0 81.3 78.9 77.7
400 129.6 102.9 93.1 88.5 86.4 85.7
450 138.7 109.9 99.7 95.2 93.5 93.3
500 147.3 116.4 106.0 101.6 100.3 100.6
550 155.4 122.6 111.9 107.7 106.7 107.7
600 163.2 128.4 117.4 113.4 113.0 114.6
650 170.7 133.9 122.8 119.0 119.0 121.2
700 177.8 139.2 127.9 124.3 124.8 127.7
750 184.7 144.2 132.8 129.5 130.5 134.0
800 191.4 149.1 137.5 134.5 135.9 140.2
850 197.8 153.7 142.0 139.3 141.3 146.3
900 204.0 158.2 146.4 144.0 146.5 152.2
an input of 200 pounds of feed. The same comparison holds true be-
tween the 14 per cent and the 16 per cent protein rations. As the hog
consumes more feed and increases in weight, the rate of gain falls off
with the higher protein ration. It again is evident from the data in Table
9.9 that the protein content of the ration must be decreased over the
production period if the rates of gain are to be kept at a maximum.
LEAST-COST VERSUS LEAST-TIME RATIONS
Rations for least-cost and least-time rations are much more similar
on pasture than on drylot (see Chapter 8). This condition holds true on
pasture because of the availability of protein in the forage. If the price
ratio is one favoring the use of a small percentage of protein in the con-
centrate mix, the hog can supplement the protein intake by consuming
more forage. Accordingly, the rate of gain is not decreased much when
the concentrate ration is adjusted to include less protein, for hogs on
pasture. In drylot, a shift in ration to meet a higher protein to corn
price ratio cannot be offset by a greater intake of protein from forage.
Table 9.9. Average Daily Gain Between Feed Intervals for Various Protein Levels
Daily Gain Computed as an Average Between 50-Pound Feed

Pounds Increments With Protein Percentages of
of Feed 10 12 14 16 18
8
100 ,632 .817 1.000 1.128 1.241 1.343
150 .819 1.074 1.294 1.422 1.511 1.563
200 .965 1.280 1.520 1.630 1.676 1.661
250 1.088 1.457 1.704 1.784 1.777 1.685
300 1.194 1.613 1.860 1.901 1.833 1.658
350 1.288 1.754 1.992 1.988 1.852 1.591
400 1.373 1.882 2.106 2.051 1.843 1.493
450 1.449 2.000 2.205 2.094 1.809 1.369
500 1.519 2.110 2.289 2.118 1.754 1.222
550 1.583 2.211 2.362 2.126 1.680 1.056
600 1.642 2.305 2.423 2.119 1.590 .873
650 1.696 2.394 2.474 2.099 1.484 .675
700 1.746 2.476 2.516 2.067 1.365 .463
Historic Outcomes
The effect of the seasonal fluctuation in hog prices on returns

over feed cost has been examined over the 20-year period from 1935
through 1954 for spring hogs farrowed at four different dates. These
figures indicate the number of years in which either the least- time or
least- cost rations would have been most profitable. The market price
at which the hogs would have been sold under each system has been de-
termined by taking into account the time required to produce a 225-
pound hog with least- cost and least- time rations. The feed prices used
were the average annual price of soybean oilmeal in each of the years
and the price of corn in the month at which the hogs reach weaning
weight, 75 pounds live weight, and 150 pounds liveweight. The price of
corn was assumed constant for the duration of each weight interval. It
was further assumed that 6 weeks would be required to raise pigs from
farrowing to weaning weight. The farrowing dates considered were
February 1, March 1, April 1, and May 1.
In the 20 years, with hogs farrowed on February 1, the least- cost
rations would have resulted in the greater return over feed costs in 15
of the years. The least-time ration would have given the greater re-
turns over feed cost in only 5 years. With hogs farrowed on March 1,
the least-cost ration would have given the greatest return over feed
cost in 19 years; the least- time ration would have given greater returns
in only 1 out of the 20 years. For hogs farrowed on April 1, the least- cost
rations would have been more profitable in 13 years, while the least-
time ration would have been more profitable in 7 out of the 20 years.
For a May 1 farrowing date, the least- cost ration was more profitable
in 9 years, while the least-time ration was more profitable in 11 of the
20 years.
Rate of gain is of lesser importance with hogs farrowed early in the
season. Hogs farrowed in February, March, and April can be produced
more profitably on the least- cost ration a greater proportion of the time
than on the least-time ration. Rate of gain is of much greater impor-
tance for hogs farrowed late in the season, because of sharp seasonal
price declines in October. The lower market price often more than
offsets the feed economies obtained by feeding the least-cost ration.
The average feed costs for the 20-year period in producing a 225-
pound market hog on pasture differ only slightly for least- cost and least-
time rations. The small difference is due to the fact that the hogs are
on pasture, with no costs figured for the latter. Protein from legumes
replaces some of that which would otherwise be obtained at a cost from
soybean oilmeal in drylot. The rates at which soybean oilmeal substi-
tutes for corn in the hog ration under a pasture feeding system are such
that the least- cost ration deviates only slightly from the rations which
maximize the rate of gain. The modal gain for the least-cost ration was
less than 50 cents over the 20 years examined. The largest difference
was for the February 1 farrowing date. In 2 of the 20 years, the returns
for the least- cost ration exceed the returns for the least- time ration by
more than $2.50. In 10 of the years, the returns with the least-cost ra-
tion were within $1 of the returns with the least- time ration. The re-
turns from the least- cost and least-time rations differed by $1 or less
for all 20 of the years under a March 1 farrowing date. On the average,
over a period of years, little gain is forthcoming from feeding the least-
cost, as compared to the least-time, ration on pasture. However, in a
few years the economic advantage of the least- cost ration with a pasture
feeding system is quite large.
Grinding and Mixing Versus Free Choice
Feeding either least-time or least- cost rations required grinding

and mixing of concentrate feeds. These extra steps add to the cost of
the rations. Hence, an additional study is needed for comparison be-
tween (a) the costs of rations fed free choice and (b) the costs of rations
plus the costs of grinding and mixing, for least-time and least-cost
rations. These comparisons are not possible from the data of this study.
CHAPTER 10'
Production Functions,
Ear! O. Heady
Stanley Balloon
Least-cost Rations, and Optimum
Robert McAlexander
Marketing Weights for Broilers
eed is the major cost in broiler production. Hence, one of the main
F opportunities for increasing profit is to lower feed costs per unit

of gain. Great progress has been made in recent years in develop-
ing high energy feeds which lessen feed inputs for producing birds of a
given weight. However, the types of cost- minimizing and profit-
maximizing principles explained elsewhere in this book have not been
widely applied. This void has existed because of lack of data in appro-
priate forms. This study was designed to provide data adapted for these
purposes.
Ordinarily, broiler feeds are made up of two major categories of
feeds, along with the proper vitamins and minerals. These two catego-
ries include feeds high in carbohydrate such as corn and feeds high in
protein such as soybean oilmeal. If prices of these feeds did not
change, the least- cost ration determined at one point in time also would
be the least- cost ration at all later points in time. However, the prices
of these major feed sources do change. In recent years the price of
corn has been as low as 1.8 cents per pound with soybean oilmeal as
high as 4.5 cents per pound, a SBOM to corn price ratio of 2.5; in other
years the price of corn has been as high as 4.5 cents per pound with
soybean oilmeal as low as 3.5 cents per pound, a SBOM to corn price
ratio of .8. The ration or combination of these two feeds which mini-
mizes costs of gains under one of these price ratios will not also mini-
mize costs under the other ratio. The least- cost ration can be deter-
mined by relating the prices of the feed sources to the rates of substi-
tution of the feeds.
OBJECTIVES
The specific purposes of this study are: To predict the broiler pro-
duction surface (function) of gains in relation to' two feed categories; to
predict input- output relationships of gain in terms of a fixed combination
of the two feeds; to predict gain or growth isoquants indicating the pos-
sible combinations of two feeds which will result in a fixed gain level;
to predict the marginal rates at which high- carbohydrate and high-
protein feeds substitute for each other in producing a particular level
330
FUNCTIONS FOR BROILERS 331
of gain; to predict isoclines indicating feed combinations for particular

gain levels which have the same rate of substitution; to indicate rations
which will minimize costs of gains under various price relationships;
to predict rations which minimize the time required for attaining a
particular marketing weight and to indicate marketing weights which
will maximize profits above feed costs for various price relationships.
However, the study is partly of a methodological nature. Since
broiler production functions have not been estimated previously, several
types of equations have been fitted to the data. For the practical use of
the data in specifying rations which average least- cost over two weight
ranges, a power function is employed although it does not serve most
efficiently in predicting the entire surface. However, as mentioned in
Chapter 9 for hogs, broiler producers are interested in rations appli-
cable for a feeding period rather than in daily recalculating the least-
cost ration. Finally, comparisons are made in prediction of rations
and marketing weights when several types of equations are used.
DESIGN OF EXPERIMENT AND FEEDING METHODS
Data for this study were obtained from an experiment conducted by

the Department of Poultry Husbandry at Iowa State University. Six
hundred New Hampshire chicks were used in the experiment. These
chicks were randomly assigned to 30 pens (batteries) with a restriction
of having 10 cockerels and 10 pullets per pen. The broilers were self-
fed on six different rations consisting of 16, 18, 20, 22, 24, and 26 per
cent protein levels. The experiment was designed so that there were
at least two replicates on each ration. Twelve groups of broilers were
fed rations with fixed proportions of corn to protein for the entire pe-
riod. In other words, two pens each of the birds were fed the entire
period on the 16, 18, 20, 22, 24, and 26 per cent rations. The other 18
pens of birds were fed up to a weight of approximately 1.32 pounds per
bird, then changed to lower protein rations for the remainder of the
feeding experiment as shown in Table 10.1. The birds were weighed
each week and corresponding feed inputs were determined to provide
observations for regression analysis. The birds were taken off the ex-
periment at the end of 11 weeks. The experimental unit was a pen, with
each weighing becoming an observation.
Corn was the main source of carbohydrates and soybean oilmeal
was the main source of protein. The soybean oilmeal contained approx-
imately 45 per cent crude protein while the corn contained approximately
8.4 per cent. (See Table 10.2.)
ESTIMATION OF PRODUCTION FUNCTIONS
Four over- all equations were fitted to the experimental observations

as one step in the objectives of this study. They are quadratic equation
332 FUNCTIONS FOR BROILERS
Table 10.1. Design of Experiment for Broiler Study
Pen- Numbers Per Cent Protein

Per Cent Protein Rations Fed Broilers
Replicates Rations Fed Broilers From Weight of
From Weight of 1.32 Pounds to End
I II .09 to 1.32 Pounds of Feeding Period
22 25 16 16
27 20 18 18
16 29 18 16
2 10 20 20
3 8 20 18
28 19 20 16
16 24 22 22
17 5 22 18
6 13 22 16
18 14 24 24
11 9 24 20
7 30 24 16
1 4 26 26
12 23 26 22
21 15 26 18
10.1, a square root, equation 10.2, a Cobb- Douglas, equation 10.3, a re-
sistance, equation 10.4, and a Spillman- type, equation 10.5, functions.
These functions are based on the 12 pens (two pens per ration) of ra-
tions which were continued from initiation of the experiment up to an
average liveweight of 3.13 pounds, with 11 to 13 weighings per pen for
1
a total of 146 observations. The functions for over-all production sur-
face estimates are as follows:
Table 10.2. Pounds of Ingredients Used per 100 Pounds

of Feed in Broiler Experiment
Per Cent Protein in Ration

Ingredients
16 18 20 22 24 26
Ground yellow corn 71.0 65.5 59.6 53.9 48.2 42.5

Wheat middlings 5.0 5.0 5.0 5.0 5.0 5.0
Dehydrated alfalfa
meal (17%) 2.5 2.5 2.5 2.5 2.5 2.5
Soybean oilmeal 15.0 20.5 26.0 31.5 37.0 42.5
Fishmeal 2.5 2.5 2.5 2.5 2.5 2.5
Steamed bonemeal 2.0 2.0 2.0 2.0 2.0 2.0
Ground oyster shells .5 .5 .5 .5 .5 .5
Iodized salt .5 .5 .5 .5 .5 .5
C-2054 (premix) 1.0 1.0 1.0 1.0 1.0 1.0
Soybean Oil ~ .2 .4 .6 .8 1.0
1
The number of observations were not the same for each pen since observations at 600
and 1,300 gram weights also were obtained for all pens; for some pens these weights oc-
curred at the same time as the regular weekly weighings, resulting in fewer observations
in these particular pens.
. . .
2 2
(10 JL) G = .0331 + .4823C + .6415S - .0183C - .0497S - .0232CS
(10.2) G = 10-1730 - .2300C - .1775S + .3314 Yc + .5004 Ys + .0200 Yc~S

5537 3371
(10.3) G = .0922C* S*
1
(10.4) G “" = -l 532 + -2700C" + .4512S""
1
(10.5) G = 17.908 - 11.442(.9238 c )

- 14.343(.9257 s ) + 8.323(.9238 c )(.9257 s )
G refers to gain in pounds per broiler, C refers to pounds of corn per

bird, and S refers to pounds of soybean oilmeal per bird. Statistics for
these equations of the over- all production surfaces are presented in
Table 10.3. Aside from the interaction term in the square root equa-
tion, the regression coefficients are all highly significant. A problem
of autocorrelation arises in estimating the regression coefficients of
the over-all production surface for this reason: The 12 and 13 observa-
tions for each pen are not independent. Hence, it can be claimed that
the total degrees of freedom (df) is something less than the 140 remain-
ing after estimating the regression coefficients. However, if the total
number of degrees of freedom remaining after estimation of regression
coefficients is considered to be only 6 (i.e., to correspond to 12 pens or
independent observations), the regression coefficients for the quadratic,
equation 10.1, power, equation 10.3, and resistance, equation 10.4, equa-
tions are still acceptable at the .01 and .05 probability levels. In all of
the equations, the feed inputs account for over 99 per cent of the vari-
ance in gains. The over- all quadratic and power equations are used for
estimates of this study in prescribing least- cost "average* rations and
optimum marketing weights.
These particular functions were selected for estimating a broiler
production surface because of their logical basis. Other studies for
meat production indicate that output tends to increase at a decreasing
rate (i.e., each additional input of feed usually results in less and less
gain in weight), and feeds tend to substitute for each other at diminishing
Table 10.3. Values of R and t for Equations 10.1 to 10.5
Values of t for Regression Coefficients in Order Shown

Equation R
bx b2 b3 bs
10.1 .999* 39.82* 29.69* 7.42* 7.44* 3.28^
10.2 .999* 6.68* S.Ol* 6.10* 6.89* .24 §
10.3 .999* 43.72* 26.97* -- --
10.4 .979* 23.48* 20.99* -- --
10.5** .910* 2.09* 4.19^ 3.52^ --
*p < both 140 and 6 df

.01 for
'p < 140 df and p < .02 for 6 df
.01 for
tp < .01 for 140 df and .01 < p < .05 for 6 df
§p > .5 for 140 df.
**Because of the nonlinear form of this equation, the listed statistics are somewhat
inexact.
rates. These conditions are expected for broilers. As birds increase

in size, more and more of each pound of feed is required for mainte-
nance. Also, as broilers increase in size, the composition of the body-
changes. The changes in body composition are expected to cause
changes in the rate of substitution between feeds. The structural por-
tions of the body and the organs develop first, followed by muscle, tis-
sue, and fat. Thus, rations of a higher protein content are required
during the early periods of growth. As birds increase in size and body
weight, less protein is required, and feeds containing more carbohy-
drates may be substituted for high protein feeds as the fattening stage
is approached.
Production functions which permit estimation of the above relation-
ships were desired. That is, they should permit (a) decreasing produc-
tivity per pound of a fixed ration (the two feeds increased in fixed pro-
portion) as well as diminishing productivity of either feed alone, (b)
diminishing rates of substitution along a particular gain isoquant, and
(c) changing substitution rates along a ration line (i.e., changing substi-
tution ratios for a particular ration) as the bird progresses in weight
and higher gain isoquants are attained. Functions 10.1, 10.2, and 10.5
meet all of these qualifications. The Spillman- type function has mar-
ginal products for corn bearing a ratio of .9238 to each other; those for
soybean oilmeal bear a ratio of .9257 to each other. Production func-
tions 10.3 and 10.4 do not permit substitution rates along a ration line
to change as the broilers increase in size. That is, they do not account
for the fact that protein, in relation to carbohydrates, is of greater value
to the young birds than to older birds. They force into the analysis the
relationship that the rate of substitution must be constant along a ration
line.
The difference between the functions such as the resistance, Spill-
man, Cobb-Douglas and quadratic equations with respect to conditions
of substitution rates along a fixed ration line can be illustrated by Fig-
ures 10.1 to 10.3. (The relationships in these illustrations are assumed
and do not represent actual predictions in this study.) In Figure 10.1,
the negatively sloped curves are gain isoquants or contours, indicating
all of the possible combinations of the two feeds which produce, respec-
tively, a 1.0-, 2.0-, and 3.0-pound gain on broilers. The solid and posi-
tively sloped (straight) lines are isoclines.
The isoclines are expansion paths showing the path of rations which
should be followed as a bird gains in weight if profits are to be maxi-
mized (i.e., if the least- cost ration is to be used for each particular
level of gain). Thus, if the price of protein feed were twice the price
of carbohydrate feed, the price ratio would be 2.0. Using equation 10.1
or 10.5 to express the necessary condition, the rations (proportions of
the two feeds) along isocline 2.0 should be followed in Figure 10.1.
Since the isoclines in Figure 10.1 do not intersect the origin, a different
ration (i.e., a different proportion of the two feeds as read off the two
axes) would be required for each fractional pound of change in gain.
This path is biologically logical since the proportion of carbohydrates
Figure 10.1. Isoclines for over-all quadratic and

Spillman functions. (assumed).
to protein feed should increase with weight. (Rations higher in the plane
along the 2.0 isocline include a greater proportion of carbohydrates.)
However, producers cannot practically change rations with each
fractional pound of gain. Generally they feed the same ration, or
change it only once, throughout the production period. If the optimum
ration for a 3.0- pound gain were selected, through equating the substi-
tution ratio with the price ratio as in equations 10.2 or 10.5, the opti-
mum ration would include OC of the carbohydrate feed and OP of the
protein feed. If the ration with the proportion of feeds at OC/OP, were
2
fed throughout, the "feed path* would be OM. This line does not indi-
cate the least- cost ration. (Isocline 2.0 does.) Hence, another ration
such as B, with feeds combined in the proportions read from the axes,
could be less costly than the ration indicated by line OM, if fed over the
entire production period. (Ration B would not equate substitution and
price ratios for any particular levels of gain.) Thus, while the path
traced by isocline 2.0 indicates the least- cost ration for all individual
weights, the ration indicated by OB may be more practical and less
costly than the ration indicated by OM.
The functions 10.3 and 10.4 provide isoclines of the nature of the
positively sloped lines labeled 2.5 and 2.0 in Figure 10.2. Since they
are linear and pass through the origin, these isoclines suggest that the
rates of substitution of the two feeds do not change along a fixed ration
line as a bird progresses to heavier weights. If they are used for de-
cisions, they indicate that the same ration should be used from the
intersects the
diagonal line could be drawn from each point when the 2.0 isocline
2
A
rations which
1.0- , 2.0-,and 3.0- pound isoquants to the origin. The slopes would indicate
could be fed in each of the three weight ranges: 0 to 1.0, 1.01 to 2.0,
and 2.01 to 3.0 pounds,
respectively. These rations would be “averages” for the intervals
and would not equate sub-
stitution and price ratios for each gain isoquant within an interval.
( 2 5)
.
Figure 10.2. Isoclines for over-all resistance and

Cobb-Douglas functions (assumed).
beginning to the end of the production period. Biologically, it is ex-

pected that rates of substitution do change along a ration line. There-
fore, a linear isocline should not pass through the origin. However,
while these equations may not provide the greatest degree of biological
accuracy, they may provide a practical basis for selecting the one
“average” ration to be fed over the entire production period for pro-
ducers who use this method of feeding. Hence, with a price of protein
feed twice as great as the price of carbohydrate feed, the single ration
(proportion of the two feeds) indicated by isocline 2.0 in Figure 10.2
would be fed over the entire production period.
An alternative combining the advantages of the two types of functions
has been used in this study. It includes the estimation of Cobb-Douglas
functions for different weight intervals or segments of the production
function. The effect on isoclines is that shown in Figure 10.3. The
Cobb-Douglas functions of the various weight intervals provide iso-
clines for a particular segment of the production surface. They indi-
cate the one best “average” ration to be fed over the particular weight
interval; the “average” ration for the interval can be determined by
equating the price and substitution ratios. The isoclines for each in-
terval can be combined to give isoclines with linear segments as in
Figure 10.3. In this case, with a price ratio of 2.0, the lower linear
segment of isocline 2.0 indicates the “average,” least-cost ration to be
fed to 2 pounds of weight, and the middle segment indicates the “aver-
age” least- cost ration to be fed between 2 and 3 pounds of weight. Since
the second segment, based on the second interval function, has a greater
slope than the first segment, a ration containing a greater proportion of
carbohydrate feed would be used for heavier weights. 3 Hence, on
3
Actually all segments of the “combined isocline” in Figure 10.3 originate at zero for
(2.5)
functions (assumed)*
practical grounds, the producer could use two or more rations over the
production period (without changing feed proportions for each fraction
of gain as indicated along the isoclines of Figure 10.1).
For the reasons outlined above, the over-all power and resistance
functions are used to predict optimum feed quantities when a single ra-
tion is to be fed over the entire production period. Two interval power
equations, presented later, have been used to provide data allowing one
change in the ration where this practice is preferred. (Producers sel-
dom use more than two rations.) However, since they assume constant
elasticity over the entire production surface, the over-all power equa-
tions tend to overestimate the gains associated with total feed consump-
tion during the latter part of the production period. Hence, they tend to
overestimate the optimum marketing weight for a particular broiler
price. For this reason the quadratic equation is used in predicting
most profitable marketing weights. In effect, this procedure is one of
”
using Cobb- Douglas functions to predict “average rations to be used
as practical alternatives up to weights of nearly 3 pounds. Beyond this
weight, the quadratic function can be used to specify (a) “exact” feed
combinations as marketing time approaches and (b) optimum marketing
weights.
Production Surfaces From Over-all Functions
Production surfaces based on equations 10.1, 10.3, 10.4, and 10.5

are presented, respectively, in figures 10.4 to 10.7. Lines OD and OB
in the feed plane are ration lines indicating 16 and 26 per cent protein
their particular weight interval. However, they can be spliced together as indicated to rep-
resent total gains and contours over the entire production surface, rather than gains within
a single interval.
o
Figure 10.4, Production surface showing feed-gain relationships
predicted from over-all quadratic function 10.1 for
broilers fed 16- to 26-per cent protein rations.
rations, respectively. Curve OE above the 16 per cent ration line of

OD indicates gain levels when various amounts of this particular ration
are fed per bird. Curve OC above the 26 per cent ration line of OB in-
dicates gain levels when various amounts of a 26 per cent ration are
fed. Other ration lines (i.e., 18, 20, 22, etc.) such as OD and OB could
be drawn in the feed plane, and each would have above it an input- output
curve for the particular ration. These quantities, corresponding to
equation 10.1 provide the later basis for determining the optimum mar-
keting weight when a particular ration is fed; they are not used for pre-
dicting the optimum ration. Figures 10.6 and 10.7 are drawn to provide
a somewhat different perspective of the surface than are figures 10.4
and 10.5.
Figure 10.5. Production surface showing feed-gain relationships

predicted from over-all Cobb- Douglas function 10.3
for broilers fed 16- to 26-per cent protein rations.
Figure 10.6. Production surface showing feed-gain relationships

predicted from resistance formula 10.4 for broilers
fed 16- to 26-per cent protein ration.
The contours on the surfaces are gain isoquants. For all functions,
the gain contours are curved indicating that, as a greater proportion of
one feed is used, the amount replaced of the other feed declines. The
nature of the surfaces indicates diminishing marginal productivity (i.e.,
each pound of feed adds less to total weight than the previous pound) for
particular rations.
Individual Growth Functions
It has been suggested that under certain conditions input- output

curves and isoquants derived from a single equation estimate of the
production surface might be spurious. Supposedly, this situation might
occur where one portion of the surface drops discretely down to a ledge,
or a u canyon” exists on one part of the function. The over- all equation
Figure 10.7. Production surface showing feed- gain relationships

predicted from Spillman function 10. 5 for broilers
fed 16- to 26 -per cent protein ration.
would be affected equally by all observations over the surface and would
not allow prediction of this discrete depression in gains. Also, if the
joint relationships involved were sufficient (gain depending on feed input
and feed intake depending on weight), estimation by simultaneous equa-
tions might be appropriate. As a basis of comparison of the predictions
made from the over- all functions, single- variable equations of the quad-
ratic and Cobb-Douglas types were estimated for each ration included
in the study. An input- output curve for each ration, independent of those
for all other rations, was then predicted from the single- variable equa-
tions and compared with a similar estimate from the over- all function.
The single- variable functions have been derived with gain, G, as the
dependent variable and corn, C, as the independent variable. This pro-
cedure can be used since the proportion of soybean oilmeal to corn is
fixed for any one of the six different protein levels.
The derived polynomial functions, hereafter designated as quadratic
single- variable functions for the indicated per cent protein levels, are:
2
(10.6) G = -.0296 + .5984C - .0244C (16%)
2
(10.7) G = .0370 + .6886C - .0323C (18%)
2
(10.8) G = .0444 + .7183C - .0305C (20%)
2
(10.9) G = .0256 + .8726 - .0520C (22%)
2
(10.10) G = .0319 + 1.0030C - .0680C (24%)
2
(10.11) G = .0377 + 1.0983C - .Q868C (26%)
where G is pounds gain in weight per broiler, and C represents the

pounds of corn fed in the various broiler rations. Implicit in each pound
of C are other feed inputs as described earlier.
Single- variable ration functions of the Cobb-Douglas type for the
various protein levels are:
9029
(10.12) G = .5878C* (16%)
8905
(10.13) G = .6669C* (18%)
Figure 10.8. Comparison of input-output curves for broilers

on a 16 per cent protein ration as predicted by
quadratic over-all function 10.1 and quadratic
single-variable function 10.6.
8786
(10.14) G = .7240C' (20%)
8172
(10.15) G = 9.7997C* (22%)
8624
(10.16) G = .9422C* (24%)
8723
(10.17) G = 1.0048C* (26%)
The input- output curves for particular rations derived from the
single- variable ration functions were all plotted on scatter diagrams
for comparison with their respective over-all functions. Comparisons
of input-output curves derived from the single- variable and over-all
quadratic functions are shown in figures 10.8 through 10.13 for the six
rations. The similarity of these two sets of curves indicates that the
over-all function does not give spurious results for any particular level
of protein. Similar comparability existed for estimates from single-
variable and over- all Cobb- Douglas equations.
Interval Functions
The interval functions of the Cobb- Douglas type used for predictions

on an 18 per cent protein ration as predicted by

single -variable function 10.8.
of "average” least- cost rations over two weight intervals are provided
in equations 10,18 and 10.19. Equation 10.18 has been fitted to observa-
tions in the weight interval of 1.3 pounds (600 grams) or less while
equation 10.19 has been fitted to observations in the interval of weights
greater than 1.3 pounds.
M25 3838
pounds liveweight)
(10.18) G = 1.0754C* S* (up to 1.3
6163 2944
(10.19) G = .7021C* S* (over 1.3 pounds liveweight)
Statistics for thetwo interval functions are given in Table 10.4.

Even for a numberof individual pens (rather than of pens by number of
weighings) the statistics are significant at the 1 per cent level of prob-
ability. As mentioned earlier, some pens were fed the same ration
throughout the experiment. (These are the observations upon which
the over-all functions and the single- variable functions are based.)
Since some pens were switched to different rations at a liveweight level
of 1.3 pounds, an analysis of variance was made for gains of birds in
the second interval in relation to gains on rations fed in the first inter-
val. It was found that gains in the second period did not differ signifi-
cantly in terms of the ration fed in the first period. Gains in the second
period did not appear to be associated with protein level in the first
Feed per Broiler in L.bs
Figure 10.11. Comparison curves for broilers

of input-output
on a 22 per cent protein ration as predicted by-
period. Hence, data for the “straight through” and “switched” pens
were pooled, and each of the interval functions is based on observations
for 30 pens averaging slightly over 6 weighings each (189 observations).
The constant elasticity of the over-all Cobb-Douglas function causes
it to overestimate the gains associated with particular feed inputs as
birds approach maturity as illustrated in Figure 10.14. The curves for
the Cobb-Douglas over- all function fit the gain observations poorly at
high feed inputs.
Table 10.4. Multiple Correlation Coefficients and Values

of t for Interval Cobb-Douglas
Functions 10.18 and 10.19
Value of t for Regression

R Coefficients in Order
Interval Values Given in Equation
hi *
b2
.09-1.32 lbs.
liveweight .9956* 33.10* 46.49*
Above 1.32 lbs.
liveweight .9885* 38.09* 17.95*
*p < .01 for 186 or 27 degrees of freedom.


quadratic over -all function 10.1 and quadratic
single- variable function 10.10.
Input-output curves for the same three protein levels are provided
in figures 10.15 and 10.16 when the estimates are made by “splicing*
together the two interval Cobb-Douglas equations. The “spliced* input-
output curves represent the portion for the second interval added (at the
end of the first interval) to the portion for the first interval. Obviously,
the problem of overestimation through use of Cobb-Douglas functions has
been lessened by “splicing* together the two interval functions; “aver-
age* least- cost rations can be estimated as practical measures for the
two intervals without a problem of overestimating gains.
ESTIMATION OF MARGINAL QUANTITIES

AND OPTIMUM RATIONS
For the practical purposes of this study, for estimation of “average*
least- cost rations over weight intervals and best marketing weights, we
use estimates from the Cobb-Douglas and quadratic functions in this
section. For the purpose of thosewho may prefer to use quantities
from the Spillman- typeand resistance formulas, we present added de-
tails on the latter two equations in a following section. Also, we later
make some comparisons between these two functions and the quadratic.

on a 26 per cent protein ration as predicted by-
quadratic over -all function 10.1 and quadratic
Total and Marginal Gains
Marginal product functions with corn and soybean oilmeal fixed are
listed in equations 10.20and 10. 21, respectively, for the over-all quad-
ratic production surface and in equations 10.22 and 10.23, respectively,
for the over-all Cobb-Douglas function.
(10.20) = .4823 - .0366C - .0232S
(10.21) = .6415 - .0994S - .0232C

0S
- *4463 3371
(10.22) If = .5494C S'
0b
5537 6829
(10.23) = .3345C' s""*
o S
Table 10.5 includes total weights per bird and marginal gains per
pound of feed when total feed input per bird is at specified levels for
various rations. Diminishing productivity of feed is indicated in total
Feed per Broiler in Lbs
Figure 10.14. Comparison various over-all functions for

of
estimating input-output curves for an 18 per
cent protein ration. Quadratic equation 10.1
and Cobb-Douglas equation 10.3.
Table 10.5. Total Weight per Bird and Marginal Gains for Various Levels of Feed Inputs per Bird,
With Feed in Fixed Proportions for Specified Rations;
Over -all Quadratic Function 10.1
Pounds
Total Weight m Pounds Marginal Gain in Pounds*
of Per cent protein levels Per cent protein levels

Feed 16 18 20 22 24 26
16 18 20 22 24
1 .547 .556 .562 .569 .576 .582 .415 .421 .427 .434 .440 .445
2 .947 .964 .967 .989 .999 1.010 .392 .395 .400 .405 .409 .412
3 1.322 1.346 1.363 1.380 1.394 1.406 .368 .369 .374 .377 .379 .379
4 1.671 1.702 1.723 1.742 1.757 1.768 .344 .343 .347 .349 .348 .346
5 1.994 2.032 2.056 2.077 2.091 2.097 .321 .317 .320 .320 .318 .313
6 2.292 2.336 2.363 2.383 2.393 2.394 .297 .291 .293 .292 .288 .280
7 2.565 2.614 2.642 2.660 2.666 2.657 .273 .265 .266 .264 .257 .247
8 2.811 2.866 2.894 2.910 2.907 2.888 .250 .239 .239 .235 .227 .214
9 3.032 3.091 3.120 3.131 3.120 3.085 .226 .213 .212 .207 .196 .181
10 3.228 3.291 3.319 3.323 3.300 3.250 .203 .186 .185 .178 .166 .148
11 3.397 3.464 3.491 3.487 3.451 3.382 .179 .160 .158 .150 .136 .115
Marginal gams are computed as a derivative of the over-all quadratic function and represent the
marginal physical products at the feed quantities shown in the first column.
Figure 10.15. “Spliced” input-output curve for broilers fed

an 18 per cent protein ration predicted from
power interval functions 10.18 and 10.19.
weights; the amount added to total weight for each added pound of feed
declines with total feed inputs. The maximum weight attained with 9
pounds of feed is with a 22 per cent protein ration. Rations with a
greater percentage include relatively too much protein for greatest nu-
tritional efficiency at heavier weights; rations with a smaller percentage
include relatively too much carbohydrate for greatest efficiency at low
weights. If extrapolations are used, the 20 per cent ration gives a max-
imum weight for 11 pounds of feed. Of course, the ration which gives
maximum weight for a given total input of feed need not be the most
profitable ration. The value of the greater gain from the particular ra-
tion must be compared with the prices of the two feeds and the quantity
of each used in the ration.
The marginal gains per combined pound of feed for different rations
again reflect the relative nutritive importance of the two feeds at differ-
ent bird weights and total feed inputs. Up to a total feed input of 3 pounds,
the marginal productivity of feed is greatest for a 26 per cent protein
ration; between 3 and 5 pounds of total feed input, marginal products are
greatest for a 22 per cent ration; between 6 and 8 total pounds of feed,
a 20 per cent ration has the largest marginal products while for total
feed inputs of 9 or more pounds, the 16 per cent ration has the greatest
marginal productivity. These shifts in marginal productivity, as feed
inputs become greater, parallel the total weights shown in the left-hand
Figure 10.16, “Spliced* input-output curve for broilers fed

a 22 per cent protein ration predicted from
power interval functions 10.18 and 10.19.
portion of the table. The fact that marginal gains per pound of feed are
greatest for (a) higher protein rations at light weights and (b) lower
protein rations at high weights is illustrated graphically in Figure 10.17.
Gain Isoquants
The production functions of equations 10.1 and 10.3 are used to de-
rive equations describing the various combinations of the two feeds
which will produce a given level of gain:
2
(10.24) C = 13.1959 - .6350S ± 27.3581 V\2351 + .0245S - .00310S - .0731G
(10.25) c =
3371
[ 9922 s* ]
The gain isoquants derived from equations 10.24 and 10.25 are pre-
sented in figures 10.18 and 10.19. The contours from both equations for
a given gain fall at about the same location in the feed plane for lower
gains. However, for greater gains, the location of isoquants for the
Cobb- Douglas function fall higher in the feed plane. (It was mentioned
previously that the over- all power function tends to overestimate gains
Figure 10. 17. Marginal (additional) gains per pound of feed

for broilers on various protein rations.
Derived from over-all quadratic function 10.1.
for large feed inputs or weights per bird.) The figuration of the iso-
quants is most accurate for the quadratic function. However, since the
slope of the isoquants along a line of given percentage protein is the
same for the Cobb- Douglas function, it can serve in the practical man-
ner mentioned earlier (i.e., it can be used to suggest the “average”
least- cost over the entire growth period, although it is not best for in-
dicating the least- cost ration for a particular increment of gain). The
dots in figures 10.18 and 10.19 indicate the feed combinations and quan-
tities necessary to give the specified gains when predictions are pro-
vided by the single- variable equations representing particular rations.
Isoquants for the lower and upper interval functions (Cobb- Douglas
type) are given in equations 10.26and 10.27, respectively. It should be
remembered that within each gain interval each member of the family
of gain isoquants will
(10.26) '
5 —
U.0754S*
-1
3838
.
1.8403
(up to 1.3 pounds liveweight)
(10.27) (over 1.3 pounds liveweight)

Figure 10.18. Gain isoquants predicted from over-all quadratic function

10.1. Dots show feed quantities required for same gains
when predictions are based on quadratic single-variable
equations for particular rations.
have the same slope along a fixed ration line for these equations. Hence,
the predictions provide the basis for the practical recommendation of
the “average* least-cost ration within the particular interval.
Substitution Rates for Corn and Soybean Oilmeal
Prediction of the substitution rates of soybean oilmeal for corn

along the isoquants allow specifying least- cost feed combinations for
particular gains. The marginal substitution rate, the slope of the iso-
product curve at a particular point or for a particular feed combination,
by the over-all power function is:
- *
=
(10.28)
|| .6088-f
For producers who want to use only one ration during the production
period, the substitution rates from the above equation would provide the
“average* basis of ration selection. Where they desire to feed two ra-
tions during the production period, equations 10.29 and 10.30 can
be
used to express “average* substitution rates within the lower and upper
,16%
18%
f
20%
N/ ^
/
22%
1Vy V ,Z
26%
A In
V 1 Lbs Gain
bs Gain
fi AyK7i :2 0 Lbs
(
Coin
Percent figu ires indicate
m L >
1 0 Lb
Gam
Gai
Lbs. Gain amount
ration
ration
of
Straight hr»es
lines
Curved lines ore
isoquants
|
protein
are
gam
in
30
J4 0
0 1.0 20 5.0
Lbs. of Soybean Oilmeal per Broiler
Figure 10.19. Gain isoquants predicted from power function 10.3.

Dots show feed quantities required for same gains
when predictions are based on power single-variable
equations for particular rations.
interval, respectively. (They are based, respectively, on equations

10.26 and 10.27.)
6C f-
(10.29) .7075
6S
j5C
(10.30) .4555-^-
6S
Data in Table 10.6 derived from equations 10.18 and 10.19 show the
various combinations of the two feeds which will produce 1 pound of gain
at broiler weights of 1.32 and 3.09 pounds liveweight. Columns 4 and 7
provide the substitution quantities in tabular form and are derived from
equations 10.29 and 10.30. Since the data in Table 10.6 are for power
functions, the substitution ratio will not change between isoquants within
a gain interval (i.e., for other broiler weights) when the feeds are com-
bined in a fixed proportion to result in a given percentage of protein in
the ration. In other words, 1 pound of soybean oilmeal substitutes for
1.62 pounds of corn when .58 pound of soybean oilmeal is combined with
1.33 pounds of corn, a total of 1.91 pounds, into a 20 per cent protein
Table 10.6. Combinations of Corn and Soybean Oilmeal for Producing a Pound of Gam
and Marginal Substitution Rates for Broilers of 1.32 and 3.09 Pounds
Liveweight. Estimates Based on Interval Cobb- Douglas
Functions 10.18 and 10.19
Lbs. Feed To Lbs. Feed To

Marginal Rate Marginal Rate
Produce 1 lb. Produce 1 lb.
Per Cent of Gam*
of Substitution of Substitution
of Gaint
Protein of Soybean of Soybean
In Ration Corn Soybean Oilmeal for Corn Soybean Oilmeal for
oilmeal Corn* oilmeal Corn*
16 1.790 .378 3.349 2.456 .519 2.749

17 1.609 .418 2.720 2.301 .598 2.233
18 1.521 .476 2.260 2.171 .680 1.856
19 1.417 .527 1.903 2.087 .775 1.562
20 1.326 .578 1.622 1.957 .854 1.331
21 1.285 .631 1.396 1.867 .946 1.146
22 1.174 .686 1.211 1.786 1.044 .994
23 1.109 .744 1.054 1.710 1.148 .866
24 1.049 .805 .922 1.650 1.259 .756
25 .994 .871 .807 1.576 1.381 .662
26 .940 .940 .708 1.509 1.510 .581
Derived from equation 10.26, lower weight interval,

fDerived from equation 10.27, upper weight interval.
t Marginal rate of substitution or 5C/6P refers to the pounds of corn replaced by a pound of soybean
oilmeal at the indicated weights. Rates for 1.32-pound weights are from equation 10.29 while those for
3.09-pound weights are from equation 10.30.
ration for a pound of gain on birds weighing 1.32 pounds; it will substi-
tute at the same
rate for corn when feeds are combined in the same
proportions for other weights up to 1.32 pounds. However, it will re-
quire less of the feeds in this fixed proportion to produce a pound of
gain when broilers are at weights lighter than 1.32 pounds; it will take
more at heavier weights. This difference in feed requirements per
pound of gain, while feeds are held in fixed proportions to give a con-
stant substitution rate, comes about because of a decline in the rate at
which feed is transformed into gain. Comparisons of feed quantities to
produce a pound of gain at weights of 1.32 pounds and 3.09 pounds illus-
trate this fact. A pound of gain for birds at the latter weight, with a 20
per cent protein ration, requires 1.96 pounds of corn and .85 pound of
soybean oilmeal, a total of 2.81 pounds.
Substitution rates for corresponding rations are lower for 3.09-
pound than for 1.32- pound broilers; a pound of soybean oilmeal replaces
less corn for heavier birds than for light birds when fed the same ra-
tion. For 1.32-pound broilers on a 20 per cent protein ration, a pound
of soybean oilmeal replaces 1.62 pounds of corn, but it replaces only
1.33 pounds of corn for 3.09-pound broilers. This relationship con-
forms to the nutritional needs of broilers at different weights. At low
weights, protein is relatively more important for growth and corn is a
less efficient substitute for soybean oilmeal than at heavier weights
where maturity is approached.
One-pound gain isoquants for broilers of 1.32- and 3.09-pound live-
weights based on the data of Table 10.6 are shown in Figure 10.20. The
isoquants in this figure are to be interpreted differently than the con-
ventional isoquant maps such as shown in figures 10.18 and 10.19. The
Figure 10.20. One-pound gain isoquants for broilers at 1.32 and

3.09 pounds liveweight as determined from power
interval functions 10.26 and 10.27 respectively.
lower and upper curves in Figure 10.20 show the combinations of corn
and soybean oilmeal required for 1 pound of gain when broilers have
liveweights of 1.32 and 3.09 pounds, respectively. Conventional iso-
quant maps show accumulated gains (or weights) and feed inputs rather
than feed inputs for a pound of gain at a specified weight. The gain iso-
quants shown in Figure 10.20 illustrate graphically the preceding dis-
cussion on diminishing substitution rates between rations and between
weights. The fact that along a fixed ration line the 1-pound isoquant for
a 3.09- pound weight has less slope than for a 1.32-pound weight indicates
that soybean oilmeal substitutes at a lower rate at the heavier weight.
Least- Cost Interval Rations From Cobb- Douglas
Least- cost rations are determined by equating the marginal rate of

substitution with the inverse price ratio. The isoclines, or points of
equal substitution rates, lie on a straight line passing through the origin
for a power function. These lines are also ration lines for the particu-
lar type of function.Thus, where the need is to predict one ration which
“averages* least- cost over the entire feeding period, equating substitu-
tion rates from the over- all power function, equation 10.3, with the price
ratio provides such a ration. Where the need is to change rations be-
tween two growth periods, equating the price ratio with substitution
rates from interval equation 10.18 provides the average least- cost
ration for the first 6-7 weeks; equating the price ratio with equation
10.19 provides the least- cost average ration in the latter part of the
feeding period. (These procedures are used as practical measures
since most broiler producers change the ration not at all or only once
during the production period. The quadratic function provides “biolog-
ically more accurate* isoclines, but is less practical. The isoclines
do not indicate “average* rations to be fed over gain interv als .)
Data in Table 10.7 provide substitution rates to indicate least-cost
rations as averages over two weight intervals or over the entire
Table 10.7. Marginal Rates of Substitution of Soybean Oilmeal for

Corn for Specified Gains as Estimated by the Cobb-
Douglas Over- All and Interval Functions
1.23 Pounds Gain 3.0 Pounds Gain

(1.32Pounds Liveweight) (3.09 Pounds Liveweight)
Substitution Substitution
rates for Substitution rates for Substitution
Per Cent single ration rates for single ration rates for
Protein over entire first over entire second
in Ration production interval t production interval t
period* period*
15.0 3.666 4.259 3.666 2.742

15.5 3.234 3.758 3.234 2.420
16.0 2.882 3.349 2.881 2.156
16.5 2.589 3.008 2.589 1,937
17.0 2.341 2.720 2.341 1.751
17.5 2.129 2.474 2.130 1.593

18.0 1.945 2.261 1.945 1.455
18.5 1.782 2.071 1.782 1.333
19.0 1.638 1.903 1.638 1.225
19.5 1.510 1.755 1.510 1.130
20.0 1.396 1.622 1.396 1.044

20.5 1.294 1.504 1.294 .968
21.0 1.202 1.397 1.202 .899
21.5 1.118 1.299 1.118 .837
22.0 1.042 1.210 1.042 .779
22.5 .972 1.129 .962 .727

23.0 .908 1.054 .908 .679
23.5 .848 .986 .848 .634
24.0 .793 .922 .793 .593
24.5 .742 .862 .742 .555
25.0 .695 .807 .695 .520

25.5 .650 .756 .650 .487
26.0 .609 .708 .609 .456
Derivatives for over- all Cobb- Douglas function covering both weight
intervals. Substitution rates do not change in the different weight
intervals when the over-all function is used (see earlier discussion
on logic of estimation).
^Derivatives for Cobb-Douglas function in first interval.
•^Derivatives for Cobb-Douglas function in second interval.
t
production period. If the price of oilmeal is 3 cents per pound and of

corn is 2 cents per pound, the price ratio is 3/2 or 1.5; a 20.5 per cent
protein ration gives the least- cost ration as an average over the first
weight interval; for this price ratio, the least- cost ration falls between
17.5 and 18.0 per cent protein for the second weight interval. If the
same ration were to be fed over the entire production period, the best
“average* ration is 19.5 per cent protein. These are the rations where
the marginal rates of substitution of soybean oilmeal for corn most
nearly approximate the soybean oilmeal to corn price ratio of 1.5. If
the price of soybean oilmeal increases to 4 cents, with corn remaining
at 2 cents per pound, the price ratio becomes 2.0. An 18.5 per cent
protein ration then averages least- cost for the first weight interval.
This ration would be fed for a total gain of 1.23 pounds (1.32 pounds
liveweight), and then a ration of 16.5 per cent protein would be fed
through the second interval.
Substitution rates as averages for the over-all production period
(based on the over- all power function) fall between those for the two
Table 10.8. Rations in Percent Protein Providing Least-Cost Combinations of Corn

and Oilmeal With Different Feed Prices (for Broilers Fed a Fixed
Percentage of Protein Throughout the Feeding Period;
Cobb- Douglas Over-All Function 10.3 Used as a
Basis of Feed Combinations)
Price
of Corn Price of Soybean Oilmeal in Cents per Pound
in Cents
3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5
per Pound*
1.6 18.0 17.5 16.5 16.0 15.5 15.5 — —

1.8 19.0 18.0 17.5 16.5 16.0 15.5 15.5 15.0
2.0 19.5 18.5 18.0 17.0 16.5 16.0 16.0 15.5
2.2 20.0 19.0 18.5 17.5 17.0 16.5 16.5 16.0
2.4 20.5 19.5 19.0 18.0 17.5 17.0 16.5 16.5
2.6 21.0 20.0 19.5 18.5 18.0 17.5 17.0 16.5
2.8 22.0 20.5 20.0 19.0 18.5 18.0 17.5 17.0
3.0 22.5 21.0 20.5 19.5 19.0 18.5 18.0 17.5
3.2 22.5 21.5 20.5 20.0 19.5 18.5 18.0 18.0
3.4 23.0 22.0 21.5 20.5 19.5 19.0 18.5 18,0
3.6 23.5 22.5 21.5 20.5 20.0 19.5 19.0 18.5
3.8 24.0 23.0 22.0 21.0 20.5 20.0 19.0 18.5
4.0 24.5 23.5 22.5 21.5 20.5 20.0 19*5 19.0
*The price for corn includes the cost of grinding, mixing,and a proportionate share
of the other feed ingredients included in the feedmixture other than soybean oilmeal.
^The price of soybean oilmeal includes a charge for mixing along with a proportionate
share of the other feed ingredients included in the feed mixture other than corn.
intervals. With a 20 per cent protein ration, the rate of substitution of

soybean oilmeal for corn is 1.622 for the first interval, 1.396 for the
over-all period or function, and 1.044 for the second interval. If the ra-
tion which averages least- cost over the entire period is fed, it includes
less protein for the first interval and more protein for the second inter-
val than would be fed if separate rations averaging least- cost over the
two weight ranges were used. Hence, the cost of gains to marketing
would be greater for a single ration than for two different rations over
the growth period. This difference must be compared to the equipment,
labor and general practicality of feeding one ration throughout the period,
or of shifting the ration to conform with changes in substitution rates
with broiler growth.
Tables 10.8, 10.9, and 10.10 provide data showing the least-cost ra-
tions, respectively, (a) throughout the production period, (b) for the first
interval of growth, and (c) for the second interval of growth when Cobb-
Douglas functions are used as the basis for predicting “average” rations
over the particular periods. Hence, with a “low” price for corn at 1.7
Table 10.9. Rations in Percent Protein Providing Least-Cost Combinations of Corn

and Soybean Oilmeal With Different Feed Prices (for Broilers Fed a
Fixed Percentage of Protein From .09 to 1.32 Pounds Liveweight;
Cobb- Douglas Interval Function 10.18 Used as a Basis of
Feed Combinations)
Table 10. 10. Rations in Percent Protein Providing Least-Cost Combinations of Corn
and Soybean Oilmeal With Different Feed Prices (for Broilers Fed a
Fixed Percentage of Protein for All Weights Above 1.32 Pounds;
Cobb- Douglas Interval Function 10.19 Used as a Basis of
Feed Combinations)
Price
of Corn Price of Soybean Oilmeal in Cents per Pound 1
in Cents
3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5
per Pound*
1.6 16.5 16.0 15.5 ~ — — — --
1.8 17.0 16.5 16.0 15.5 — — — —

2.0 18.0 17.0 16.5 16.0 15.5 — —
2.2 18.5 17.5 17.0 16.0 16.0 15.5 15.0 —
2.4 19.0 18.0 17.5 16.5 16.0 15.5 15.5 15.0
2.6 19.5 18.5 17.5 17.0 16.5 16.0 15.5 15.5
2.8 20.0 19.0 18.0 17.5 17.0 16.5 16.0 15.5
3.0 20.5 19.5 18.5 18.0 17.5 17.0 16.5 16.0
3.2 20.5 19.5 19.0 18.0 17.5 17.0 16.5 16.5
3.4 21.0 20.0 19.5 18.5 18.0 17.5 17.0 16.5
3.6 21.5 20.5 19.5 19.0 18.5 17.5 17.5 17.0
3.8 22.0 21.0 20.0 19.0 18.5 18.0 17.5 17.0
4.0 22.5 21.0 20.5 19.5 19.0 18.5 18.0 17.5
*The price for corn includes the cost and a proportionate share
of grinding, mixing,
of the other feed ingredients included in the feedmixture other than soybean oilmeal.
tThe price of soybean oilmeal includes a charge for mixing along with a proportionate
share of the other feed ingredients included in the feed mixture other than corn.
cents and a * high * price for soybean oilmeal at 6 cents per pound, the
least-cost ration to be fed over the entire period includes 15.0 per cent
protein. With corn at 2 cents and soybean oilmeal at 4 cents, the least-
cost ration in the first interval is 18.5 per cent protein; the least- cost
ration for the second interval is 16.5 per cent protein. Hence, tables
10.8 to 10.10 can be used to determine the percentage of protein in
the ration which gives lowest feed costs per pound of gain for any of
the combinations of the prices shown. The rations indicated in the cells
of the table are those where the marginal rate of substitution of soybean
oilmeal for corn is equal to the price ratio obtained by dividing the soy-
bean oilmeal price at the top by the corn price in the left-hand column
of the table.
Graphic illustrations of the average least- cost rations for two dif-
ferent price ratios of soybean oilmeal and corn are given in figures
10.21 and 10.22 for the two weight intervals. Under a situation with a
price ratio of soybean oilmeal to corn of 1.6, (e.g., $4.00 and $2.50 per
hundred pounds, respectively, for the two feeds), the least- cost rations
Figure 10.21. Least-cost rations for two weight intervals

based on power functions 10.18 and 10.19 with
a soybean oilmeal to corn price ratio of 1.6.
for the two periods are as shown in Figure 10.21. A 20 per cent protein
ration provides the “average” least- cost ration until a weight of about
1.32 pounds is attained; then a 17.5 per cent ration provides the “aver-
age” least- cost ration for the remainder of the feeding period. An in-
crease in the price ratio to 1.875, caused by an increase in soybean
oilmeal prices, a decrease in corn price or a combination of these,
would cause a new set of rations to become lowest in cost, as shown in
Figure 10.22. The “average” least-cost rations now include 19.0 and
16.5 per cent protein levels for the first and second periods. Figure
10.23 shows the nature of ration paths over the two intervals when one
change is made in feed combinations over the production period. The
break in slopes of the isoclines comes at the end of the first interval.
The corresponding isocline for the second interval is “spliced” on to
indicate the “average” least- cost rations for the two weight ranges.
Hence, the isocline labeled 2.0 would be followed for prices such as 4
cents for soybean oilmeal and 2 cents for corn, etc.
In figures 10.21 to 10.23, we should remember that the slope of the
upper segment of the isocline starts from the origin of a new feed plane.
In figures 10.21 and 10.22, for example, the boundaries of the new feed
Lbs of Soybeon Oilmeol
Figure 10.22. Least-cost rations for two weight intervals

based on power functions 10.18 and 10.19 with
a soybean oilmeal to corn price ratio of 1.875,
plane are formed by the two lines which intersect at the “splice” in the
isocline. The scale for these new axes starts from zero and feeds are
measured accordingly for the second interval. Although the “new axes”
are not shown because of space limitations, a new origin actually occurs
in Figure 10.23 at the point of “splice” for each pair of segments form-
ing an isocline, and feeds must be measured accordingly.
Simplified Determination
Figure 10.24 provides a simplified basis for estimating the least-

cost ration in either weight interval, or for the total production period.
While it has been devised by relating substitution ratios to price ratios,
it considers only discrete points on the production surface and speci-
fies a single optimum ration for small ranges of price ratios. For ex-
ample, it indicates a 24.5 per cent ration over the entire growth period
for soybean oilmeal to corn price ratios between .7 and .8; for price
ratios between 2.4 and 2.5, the optimum single ration over the entire
production period is 16.5 per cent protein.
Figure 10.23. “Spliced* isoclines for two weight intervals showing

path of least- cost rations when feed combinations
are changed once during the production period.
Based on power functions 10.18 and 10.19-
The graph can be used as follows: Suppose the price of soybean oil-
meal is 6 cents per pound ($6 per hundredweight) while corn is 3 cents
per pound ($3 per hundredweight). Follow across the horizontal *$6
line” for soybean oilmeal until it intersects the “$3 line” for corn. Then
follow the diagonal line passing through this point of intersection to find
the least- cost ration. It will include 18 per cent protein if a single ra-
tion is fed; it will include 19 per cent for the first interval and 16.5 per
cent for the second interval if one change in rations is made during the
growing period.
Most Profitable Weights for Broilers
While the procedure and tables outlined above allow specification of

the optimum ration, they do not indicate the total amount of feed to be
fed per broiler and, hence, the optimum marketing weight. However,
SBOM* Opt. Opt. Interval

Corn Single Rations
Price Rot ion
Ratio
— 16.0
— 16.0
— 16.0
— 15.5 — 16.5
-- 15.5 — 16.5
— 16.0 — 16.5
—
—
—
16.0
— 17.0
16.0 17.0 15.0
— 16.5 — 17.0 15.0
— 16.5 — 17.5 15.5
— 16.5 --
--
—— 17.5 15.5
*
— 17.0 18.0 15.5

— 17.0 — 18.0 16.0
— 17.5 — 18.5 16.0
08
— 17.5 — 18.5 16.0
W .OO:
£
— IB.O " 190 16.5
6
:
— 18.5 — 19.0 16.5
:
— 19.0 — 95 t
17.0
— 19.0 —2 0 Qt 17.5
56.00E
— 19.5 — 20.5 17.5
- 20.0 — 20.5 18.0
- 20.5 — 21.0 18.5
- 21.0 — 22.0 19.0
— 21.5 -- 22.5 19.5
— 22.0
— 23.0
— 23.0 20.0
— 24.0 20.5
— 23.5 — 24.5 21.5
— 24.5 — 25.5 22.5
g.oo
1.00
E«^r~rr r.r.rn
2.00
,-rr.
3.00
1 —
4.00
1
Price of Corn per 100 Lbs. ($)

Figure 10.24. “Average” least-cost protein rations for broilers fed (a) a single
ration and (b) two different rations during the production period for
various prices of soybean oilmeal and corn. Equations 10.28, 10.29
and 10.30 are used as a basis for ration selections.
after the least- cost ration has been determined, it is possible to use the
input- output equations to determine the optimum level of feeding and the
most profitable marketing weight. The optimum market weight is deter-
mined by equating the derivative of the gain- feed function for a particu-
lar protein ration with the feed to broiler price ratio and solving for the
relevant unknowns. The marginal physical products from feed are thus
equated with the feed to broiler price ratio.
Quadratic function 10.1 has been used for obtaining the optimum
weights for protein levels. Total weights for broilers estimated for
various protein levels are shown in Table 10.11. Rations high in protein
provide the greatest gains per unit of feed used for low weights; as feed
intake increases, rations lower in protein content are more efficient.
The marginal quantities in Table 10.12 illustrate this relationship more
clearly. For the first few pounds of feed consumed, the marginal or ad-
ditional gains per unit of feed input are highest at the 26 per cent protein
level. As more feed is consumed, the rations giving the highest addi-
tional gains per unit of feed consumed are those with lower protein
levels.
Table 10.13 indicates the optimum marketing weight for various
ratios of feed and broiler prices when time and uncertainty, number of
broods and capital limitations are not of concern. The least- cost ration
might be determined first in tables 10.8, 10.9, and 10.10. Then Table
10.13 could be used to predict the total amount of the particular ration
and the optimum marketing weight per broiler. By equating the deriva-
tive of each function with the feed-broiler ratio, (column 2 of Table
10.13) the optimum quantity of feed for a particular protein ration is
obtained. (The feed to broiler ratio is the inverse of the broiler to feed
ratio.) Broiler to feed price ratios (column 1 of Table 10.13) from 3.6
to 7.0 are used as a basis for determining optimum feed quantities for
the various rations. Once the optimum quantity of feed is obtained, the
Table 10.11. Total Liveweight per Broiler for Indicated Pounds of Accumulated Feed Inputs When
Fed Various Protein Rations — Predicted From Quadratic Equation 10.1*
in Pounds 16 17 18 19 20 21 22 23 24 25 26
.5 .337 .339 .342 .343 .345 .347 .349 .350 .352 .354 ,356
1.0 .547 .551 .556 ,559 .562 .566 .569 .572 .576 .579 .582
1.5 .750 .757 .763 .768 .772 .777 .782 .787 .792 .796 .800
2.0 .947 .956 .964 .970 .976 .982 .989 .994 .999 1.005 1.010
2.5 1.138 1.148 1.158 1.166 1.173 1.180 1.188 1.194 1.201 1.207 1.212
3.0 1.322 1.334 1.346 1.355 1.363 1.372 1.380 1.387 1.394 1.400 1.406
3.5 1.500 1.514 1.527 1.537 1.546 1.556 1.565 1.572 1.579 1.586 1.591
4.0 1.671 1.687 1.702 1.713 1.723 1.733 1.742 1.750 1.757 1.763 1.768
4.5 1.836 1.854 1.870 1.882 1.893 1.904 1.913 1.921 1.928 1.933 1.937
5.0 1.994 2.014 2.032 2.045 2.056 2.067 2.077 2.085 2.091 2.095 2.097
5.5 2.147 2.168 2.187 2.201 2.213 2.224 2.233 2.241 2.246 2.249 2.250
6.0 2.292 2.315 2.336 2.351 2.363 2.376 2.383 2.389 2.393 2.395 2.394
6.5 2.432 2.456 2.478 2.493 2.506 2.517 2.525 2.531 2.533 2.533 2.530
7.0 2.565 2.591 2.614 2.630 2.642 2.653 2.660 2.665 2.666 2.663 2.657
7.5 2.691 2.719 2.743 2.759 2.771 2.782 2.789 2.791 2.790 2.785 2.777
8.0 2.811 2.841 2.866 2.882 2.894 2.904 2.910 2.911 2.907 2.900 2.888
8.5 2.925 2.956 2.982 2.999 3.011 3.020 3.024 3.023 3.017 3.006 2.991
9.0 3.032 3.064 3.091 3.108 3.120 3.128 3.131 3.128 3.120 3.105 3.085
9.5 3.133 3.167 3.194 3.212 3.223 3.230 3.230 3.225 3.213 3.196 3.172
10.0 3.228 3.263 3.291 3.308 3.319 3.324 3.323 3.315 3.300 3.279 3.250
10.5 3.306 3.352 3.381 3.398 3.408 3.412 3.409 3.473 3.379 3.353 3.320
11.0 3.397 3.435 3.464 3.481 3.491 3.493 3.487 3.473 3.451 3.420 3.382
Total liveweights obtained by adding initial weight of .09 pound for chicks, to gains estimated from
quadratic over-all function 10.1.
364 functions for broilers
Table 10.12. Marginal Gains (Pounds Gain per Added Pound of Feed) From Specified Feed Inputs per
Broiler on Various Protein Rations Estimated From Quadratic Function 10.1*
Feed
m Pounds 16 17 18 19 20 21 22 23 24 25 26
.5 .426 .430 .434 .438 .441 .444 .448 .451 .455 .458 .461
1.0 .413 .417 .421 .424 .427 .431 .434 .437 .440 .442 .445
1.5 .400 .404 .408 .411 .414 .417 .420 .422 .424 .426 .428
2.0 .388 .392 .395 .398 .401 .403 .405 .407 .409 .411 .412
2.5 .375 .379 .382 .385 .387 .389 .391 .393 .394 .395 .395
3.0 .362 .366 .369 .372 .374 .376 .377 .379 .379 .379 .379
3.5 .349 .353 .356 .358 .360 .362 .363 .363 .364 .363 .362
4.0 .336 .340 .343 .345 .347 .348 .349 .349 .348 .347 .346
4.5 .324 .327 .330 .332 .333 .334 .334 .334 .333 .332 .329
5.0 .311 .314 .317 .319 .320 .320 .320 .319 .318 .316 .313
5.5 .298 .301 .304 .306 .306 .307 .306 .305 .303 .300 .296
6.0 .285 .288 .291 .292 .293 .293 .292 .290 .288 .284 .280
6.5 .272 .276 .278 .279 .279 .279 .278 .275 .272 .268 .264
7.0 .259 .263 .265 .266 .266 .265 .264 .261 .253 .253 .247
7.5 .247 .250 .252 .253 .253 .251 .249 .246 .242 .237 .231
8.0 .234 .237 .239 .239 .239 .238 .235 .232 .227 .221 .214
8.5 .221 .224 .226 .226 .226 .224 .221 .217 .212 .205 .198
9.0 .208 .211 .213 .213 .212 .210 .207 .202 .196 .189 .181
9.5 .195 .198 .200 .200 .199 .196 .193 .188 .181 .174 .165
10.0 .183 .185 .187 .187 .185 .183 .178 .173 .166 .158 .148
10.5 ,170 .172 .173 .173 .172 .169 .167 .161 .154 .142 .132
11.0 .157 .159 .160 .160 .158 .155 .150 .144 .136 .126 .115
Figures in body of table indicate added pounds of gain from each 1-pound added unit of feed, starting
from the total feed inputs shown in the first column.
corresponding amount of gain found by substituting the feed quantity

is
into the appropriate ration function. Adding the initial weight of the
chick, or about .09 pound, provides the optimum marketing weights for
the various broiler and feed price combinations.
The predicted optimum marketing weights for broilers on rations of
protein levels ranging from 16 to 26 per cent with various broiler and
feed prices are shown in Table 10.13. These predicted weights are for
situations where (a) capital is not limiting, (b) the weights provide max-
imum returns (or minimum losses) above feed costs, (c) risk and un-
certainty are not considered, (d) time required for attaining optimum
weights is not considered, and (e) the same ration is fed throughout the
feeding period.
Using the broiler to feed price ratio of 4.8, the optimum marketing
weights for broilers according to data in Table 10.13 would range from
2.92 to 3.15 pounds, depending on the ration fed. With a 20 per cent
protein ration, the least- cost ration over this same period, a weight of
3.15 is optimum. Data from tables 10.12 and 10.13 have been used to
develop graphical guides for determining optimum weights and corre-
4
sponding feed inputs, but are not shown here.
4
See Heady, Earl O., Balloun, Stanley, and McAlexander, Robert. Least- cost rations
and optimum marketing weights for broilers. Iowa Agr. Exp. Sta. Bui. 442. Ames. 1956.
Table 10.13. Weights (in Pounds) for Maximizing Returns Above Feed Costs for Broilers on Various
Protein Rations With Specified Broiler to Feed (Feed to Broiler) Price Ratios
Predicted From Quadratic Equation 10.1
Broiler Feed
to Feed to Broiler Per Cent Protein in Ration
Price Price 20 24 25 26
16 17 18 19 21 22 23
Ratio Ratio
.278 2.38 2.43 2.48 2.51 2.52 2.53 2.52 2.51 2.48 2.45 2.39
3.6
.263 2.53 2.59 2.63 2.66 2.67 2.67 2.66 2.65 2.61 2.58 2.51
3.8
.250 2.66 2.72 2.76 2.78 2.80 2.80 2.78 2.76 2.72 2.68 2.61
4.0
.238 2.77 2.83 2.87 2.89 2.90 2.90 2.89 2.86 2.82 2.78 2.72
4.2
.227 2.87 2.93 2.97 2.99 3.00 2.99 2.97 2.95 2.90 2.86 2.80
4.4
.217 2.96 3.01 3.05 3.07 3.08 3.07 3.05 3.02 2.98 2.92 2.87
.208 3.03 3.08 3.13 3.15 3.15 3.14 3.12 3.08 3.04 2.98 2.92
4.8
.200 3.10 3.15 3.18 3.21 3.21 3.20 3.18 3.14 3.10 3.04 2.98
5.0
.192 3.15 3.20 3.25 3.27 3.27 3.26 3.23 3.19 3.15 3.08 3.02
5.2
.185 3.21 3.26 3.30 3.32 3.32 3.31 3.28 3.24 3.19 3.13 3.06
5.4
.179 3.26 3.31 3.35 3.37 3.36 3.35 3.32 3.28 3.23 3.17 3.10
5.6
.172 3.30 3.35 3.39 3.41 3.40 3.39 3.36 3.32 3.27 3.20 3.13
5.8
.167 3.34 3.39 3.42 3.44 3.44 3.42 3.40 3.35 3.29 3.23 3.16
6.0
3.37 3.42 3.46 3.48 3.47 3.46 3.42 3.38 3.32 3.26 3.19
6.2 .161
3.41 3.45 3.49 3.51 3.50 3.48 3.45 3.41 3.35 3.29 3.21
6.4 .156
3.43 3.48 3.52 3.54 3.53 3.51 3.48 3.44 3.37 3.31 3.23
6.6 .152
3.46 3.51 3.54 3.56 3.56 3.54 3.50 3.46 3.40 3.33 3.25
6.8 .147
3.48 3.53 3.57 3.58 3.58 3.56 3.52 3.48 3.42 3.35 3.27
7.0 .143
Time Requirements
Previous analyst dealt only with the cost of alternative rations.

However, the broiler producer also is interested in the time required
for gains. If he is faced with a seasonal or cyclical decline in
broiler
may wish use the least- time ration, rather than the least-
price, he to
rising
cost combination of feeds. If he is faced with the possibility of
broiler prices, he will undoubtedly want to use the least- cost
ration.
need not be identical with the least- cost ration.
The least- time ration
of the
The two will be identical under price situations where the cost
timeliness is low, so that the two
feed ingredient providing the greatest
Under other price situations the costs
types of rations are identical.
of the "time-saving” feed may be relatively
high. The least- time ration
will then be more costly than the least- cost ration.
To provide information aiding these types of decisions an additional
function showing time elapsed to consume various

amounts of these ra-
and time equations it is possi-
tions has been computed. From the gain
elapsed for a specific gain or weight level. Various
ble to compute time
the time relationship.
algebraic functions were tried as expressions of
equation 10.31 with square root terms
The best function appeared to be
where T is time in days and S and C
= .6735 + 4.7974C + 9.4576S + 21.4617 Yc + 13.6188 Ys - 12.0287 Yc S

( 10 . 31 ) T
in pounds. The property

are pounds of soybean oilmeal and corn per bird
other functions tried was that the func-
which appears to qualify it over
sharp curvature for low feed inputs but tends to
tion allows a relatively
more nearly approach linearity for high feed inputs. In other words, it
is consistent with the growth of the bird's digestive capacity at the out-
set when proportionately less time is required to consume a given addi-
tional quantity of feed; it is consistent with the tendency for a limit in
growth of the bird's digestive capacity as the bird approaches maturity.
At heavy weights, digestive capacity is limited, and a bird consumes
about a constant amount per day (i.e., each additional pound of feed is
consumed in about the same period of time as the bird approaches ma-
turity). All of the coefficients for this time function are acceptable at
a 1 per cent level of probability, and 99 per cent of the variance in time
required to consume various quantities of feed is explained by the vari-
ables in the equation.
Analysis of variance also was used to test the significance in dif-
ferences in rate of gain up to 1.23 of total gain and up to a total of 3.0
pounds of gain for the six rations of the experiment. These tests
showed the differences to be significant at the 5 per cent level of prob-
ability. However, there was no significant difference in rate of gain for
the six rations between a total gain of 1.23 and 3.0 pounds. Evidently,
the main effect of rations on rate of gain is in the earlier growing pe-
riod.
Figures 10.25 and 10.26 show the predicted relationship between
feed consumption and time for two rations of the study. Time curves
for other rations are similar. Table 10.14 indicates the predicted
amount of time required to attain weights of 1.32 pounds and 3.25 pounds
for rations of different protein levels. For the lower weight range, a
ration of slightly over 23 per cent protein is predicted to give a 1.32-
pound liveweight in the minimum amount of time. For the entire weight
range to 3.25 pounds liveweight, a protein percentage of slightly over
21.0 per cent is predicted to give most rapid gains. Rations containing
Figure 10.25. Estimated number of days required to consume

various quantities of feed for broilers fed a 16
per cent protein ration. Time estimates are
based on square root function 10.31.
a greater percentage of protein are predicted to give somewhat less

rapid gains. From the data of substitution rates in earlier tables, it
is obvious that these rations which give the most rapid gains do not
also give the least-cost rations under normal price relationships.
However, the cost of the least-time ration is only slightly above the
least-cost ration when the price of soybean oilmeal is low relative to
the prices of corn. In cases where the price of soybean oilmeal is rel-
atively high, the broiler producer needs to compare the savings in feed
from use of the least-cost ration with any possible gain in broiler
price obtained from getting to market sooner under the least-time ra-
tion. The absolute differences in profits from least-cost and least-
time rations will be very small for a few birds but can be quite large
for a large operation when corn is low in price compared to soybean
oilmeal.
Table 10.15 shows the predicted number of days for broilers to
reach the optimum marketing weights shown in Table 10. 14. The
of the two tables allow prediction of the weight which will allow maxi-
mum profit above feed costs. The sequence in considering ration
costs, marketing weights and time to market might be this: (a) Select
the least-cost ration in Table 10.9. (b) Select the most profitable
marketing weight for the least-cost ration from Table 10.13. (c) Ex-
amine Table 10. 15 for the time involved and, after considering the
prospects for prices, determine whether or not feeding plans should be
altered to fit prospects for price increases or decreases. Table 10.15
can be interpreted thus: With a broiler to feed price ratio of 4.4, and
with an 18 per cent protein ration being the least-cost one shown in
Table 10.9, the optimum marketing weight of 2.97 pounds shown in
Table 10.13 can be obtained in 74.7 days. If broilers increase in price
so that the broiler to feed price ratio becomes 5.0, with an 18 per cent
Figure 10.26. Estimated number of days required to consume

various quantities of feed for broilers fed a 24
per cent protein ration. Time estimates are
based on square root function 10.31.
ration giving lowest feed costs, the optimum marketing weight is 3.18
pounds per Table 10.13 and the time required is 79.5 days per Table
10.15.
Number of Flocks Per Year
A considerable variation exists in the number of flocks of birds

raised by broiler producers each year. Some poultrymen raise only a
flock or two each year. However, poultrymen who have broiler pro-
duction as a major source of income usually raise at least three or
more flocks each year. For producers who raise only three or less
groups per year, it usually is possible to carry the broilers to their
optimum weights without any time conflict. For broiler producers who
desire to raise four groups per year, there would be no conflict on time
for broilers fed on any of the rations when the broiler to feed price
ratio is 5.5 or less. Raising four groups under the above price ratios
would permit at least a week between each flock, depending on the
broiler to feed price ratio and the protein ration being fed. When the
broiler to feed price ratio is above 5.5, birds on the lower protein ra-
tions would require marketing at slightly less than optimum weights if
a four flock schedule were rigidly followed. With a broiler to feed
price ratio of 6.0, four flocks could be carried to optimum weights on
rations containing a 19.5 or greater percentage of protein with a week
between flocks; birds on lower protein rations could not be carried to
optimum weights without a time conflict. Data in Table 10.13 can be
Table 10.14. Estimated Feed Requirements and Number of Days

per Broiler for Specified Weights When Fed
Various Protein Rations
Starting to Starting to
Per Cent 1.32 Pounds Weight 3.25 Pounds Weight
Protein
Pounds No. Pounds No.
in Ration
feed* days t feed* days t
15.0 3.31 46.3 10.27 86.1

16.0 3.11 44.5 9.82 82.6
17.0 2.96 43.4 9.41 80.1
18.0 2.86 42.5 9.29 78.7
19.0 2.79 41.6 9.17 77.6
20.0 2.74 41.1 9.11 77.0
21.0 2.71 40.7 9.09 76.7
22.0 2.69 40.5 9.11 76.8
23.0 2.68 40.4 9.18 77.3
24.0 2.68 40.5 9.28 78.1
25.0 2.70 40.8 9.44 79.3
26.0 2.74 41.0 9.62 80.7
Predicted from equation 10.18.

t Predicted from equation 10.31.
tPredicted from equations 10.18 and 10.19.

used to determine whether birds could be held until optimum weights

are attained if four or more flocks are to be produced each year.
Risk and Uncertainty
Because of risk and uncertainty, broiler producers may not hold

their birds until they attain the optimum market weights. The uncer-
tainty of expected prices and death losses due to disease and other
hazards may result in earlier marketing. However, modern techniques
for prevention and treatment of diseases have done much to reduce this
type of uncertainty. Also, insurance against hazards such as fire tends
to reduce risk for the poultrymen. Prices usually provide the greatest
source of uncertainty except where the producers have some type of
forward pricing.
Obviously, the decision of the best marketing weight depends on
many factors including (a) input-output relationships, (b) the previous
commitments, i.e., contractual arrangements, (c) number of flocks per
year, (d) price expectations, and (e) risk preference of the individual
poultryman. However, the data provided in this study on input-output
data, selection of rations, estimation of optimum market weights, and
corresponding time required for attaining optimum weights provides
information for reducing much of the uncertainty in broiler production.
Table 10.15. Predicted Time m

Days Required for Broilers To Attain Optimum Marketing Weights
Shown in Table 10.13; Estimated From Time Function 10.31
Broiler Feed
to Feed to Broiler Per Cent Protein in Ration
Price Price
16 17 18 19 20 21 22 23 24 25 26
Ratio Ratio
(days) (days) (days) (days) (days) (days) (days) (days) (days) (days) (days)
3.6 .278 64.9 65.0 65.1 64.9 64.7 64.5 64.3 64.0 63.6 63.2 62.7
3.8 .263 68.0 68.0 67.0 67.8 67.5 67.2 66.9 66.5 66.1 65.6 65.2
4.0 .250 70.7 70.6 70.5 70.2 69.9 69.6 69.2 68.8 68.3 67.8 67.3
4.2 .238 73.1 72.9 72.7 72.4 72.0 71.6 71.2 70.8 70.3 69.8 69.2
4.4 .227 75.2 75.0 74.7 74.3 73.9 73.5 73.0 72.5 72.0 71.5 70.9
4.6 .217 77.1 76.8 76.5 76.0 75.6 75.1 74.6 74.1 73.6 73.0 72.4
4.8 .208 78.8 78.5 78.1 77.6 77.1 76.6 76.1 75.6 75.0 74.4 73.8
5.0 .200 80.4 79.9 79.5 79.0 78.5 77.9 77.4 76.8 76.3 75.6 75.0
5.2 .192 81.8 81.3 80.9 80.3 79.7 79.2 78.6 78.0 77.4 76.8 76.2
5.4 .185 83.1 82.6 82.0 81.5 80.9 80.3 79.7 79.1 78.5 77.9 77.2
5.6 .179 84.3 83.7 83.2 82.6 81.9 81.3 80.7 80.1 79.5 78.8 78.2
5.8 .172 85.4 84.8 84.2 83.5 82.9 82.3 81.7 81.0 80.4 79.7 79.1
6.0 .167 86.4 85.7 85.1 84.4 83.8 82.2 82.5 81.9 81.2 80.5 79.9
6.2 .161 87.3 86.6 86.0 85.3 84.6 84.0 83.3 82.7 82.0 81.3 80.6
6.4 .156 88.2 87.5 86.8 86.1 85.4 84.7 84.0 83.4 82.7 82.0 81.3
6.6 .152 89.0 88.3 87.6 86.8 86.1 85.4 84.7 84.1 83.4 82.7 82.0
6.8 .147 j
89.7 89.0 88.3 87.5 86.8 86.1 85.4 84.7 84.0 83.3 82.6
7.0 ,143 90.5 89.7 88.9 88.2 87.4 86.8 86.0 85.3 84.6 83.9 83.2
j
SOME FURTHER COMPARISONS
The quadratic and Cobb -Douglas equations were used in the previ-
ous section to provide certain estimates for practical use. Additional
comparisons are made in this section of estimates provided by the re-
sistance and Spillman-type functions. Some readers may be interested
in the latter two functions* in substituting them for practical recom-
mentations or in deriving added detail from them. Both differ from
other equations examined since they provide gain asymptotes, a condi-
tion which appears realistic for meat production functions. However,
since the broilers in the experiment reported were not carried to
weights approaching maturity, use of an equation allowing this condi-
tion did not seem necessary for the foregoing analysis.
Total Gains
The total gains predicted from the resistance, equation 10„4, and
Spillman, equation 10.5, equations are provided in Table 10.16. These
can be compared with similar estimates from the quadratic equation
10.1 in Table 10.11.
The resistance formula, equation 10.4, provides total gain and
marginal feed productivities more like those of the quadratic, equation
10.1, than does the particular form of the Spillman equation 10.5 used.
However, equation 10.4 predicts smaller total weights at low protein
levels and higher total weights at high protein levels than does equa-
tion 10.1. In general, equation 10.5 predicts larger total weights for a
given feed input than do either of the other two equations.
Table 10.16. Total Liveweight per Bird Predicted From Equations 10.4 and 10.5 for
Specified Feed Inputs and Rations
Resistance (10.4) Spillman (10.5)

Feed
Input in Per cent protein Per cent protein
Pounds 16 18 20 22 24 26 16 18 20 22 24 26
1 .419 .479 .547 .584 .616 .633 .633 .750 .765 .838 .806 .774
2 .787 .892 1.009 1.073 1.126 1.155 1.056 1.066 1.123 .992 1.179 1.211
3 1.113 1.252 1.405 1.487 1.555 1.591 1.370 1.234 1.441 1.504 1.522 1.586
4 !
1.404 1.569 1.748 1.843 1.921 1.962 1.684 1.638 1.827 1.869 1.677 1.864
5 1.666 1.851 2.048 2.151 2.237 2.281 1.928 2.079 2.050 2.219 2.295 2.353
6 1.902 2.101 2.313 2.422 2.512 2.559 2.293 2.398 2.432 2.610 2.533 2.649
7 2.116 2.327 2.547 2.661 2.753 2.802 2.643 2.717 2.875 2.806 3.099 3.197
8 2.312 2.536 2.758 2.874 2.968 3.018 2.940 3.065 3.050 3.354 3.448 3.540
9 2.490 2.715 2.946 3.065 3.160 3.209 3.155 3.407 3.601 3.711 3.840 3.961
|
10 2.655 2.884 3.118 3.235 3.332 3.382 3.590 3.751 3.970 4.064 4.249 4.385
11 2.806 3.037 3.273 3.391 3.487 3.536 3.822 3.935 4.315 4.451 4.485 4.552
Figure 10.29. Gain isoquants predicted from Spillman function 10.5.
Isoquants and Isoclines
Isoquants and isoclines for the resistance, equation 10.4, and

Spillman, equation 10.5, functions are illustrated in figures 10.27 to
10.30. ^They are derived from isoquant equations 10.32 and 10.33 below.
Figure 10.30. Isoclines for broilers from Spillman function 10.5.

-1
(10.32) S = (.4512CG) (C - .1532CG - .2700G)
^ 11.442(.923 8 C)1
(10.33) S - log {log ..
L 8.323(.9238 c )
- 14.343 J
The respective equations of marginal rates of substitution for these

two equations are:
2
6S = „ S
(10.34) 6112
- fi11
Jc C2
5C _ .926 s (.0772)) 14,343 - 8.323(.9238 c )|

(10.35)
5S
.9238 c (.0793)[ 11.442 - 8.323 (.9257 s )]
Isoclines have been derived for the resistance, equation 10.4, and
Spillman, equation 10.5, equations from equations 10.34 and 10.35, re-
spectively, by setting the marginal rates of substitution equal to a con-
stant, k,and solving for one feed variable in terms of the other. The
isoclines for equations 10.4 and 10.5 are respectively:
B
(10.36) S = 1.2923Ck~*
(.9238 c )(0 9069) 1

(10.37) S = log (log .9257)."
1.107k - .642k(.9238~c ) + .6597 (. 9238 c )
The isoquants for the resistance and Spillman functions are rela-
tively flat within the boundaries of the rations
used in the study. The
extrapolations outside of these boundaries in figures 10.27 and 10.29
are not used for prediction, but only to better suggest the general alge-
braic nature of these isoquants. Similarly, isoclines are included
which are outside of the experimental observations.
Marketing Weights
The marketing weights predicted by the particular Spillman function

were considerably above those for the quadratic equation. Predictions
by the resistance formula were relatively similar to the quadratic
equation, especially for broiler to feed price ratios which are reason-
able. Hence, because of these comparisons and the greater ease in
calculations, the function used in previous sections was retained as a
basis for practical recommendations.
CHAPTER 11
Least-cost Rations
Stanley Balloun
Earl O. Heady
Gerald W. Dean
and Optimum Marketing Weights
for Turkeys
URKEY PRODUCTION has become increasingly competitive and
T specialized in recent years. United States turkey production has

nearly doubled since 1940, while the number of farms raising tur-
keys has steadily decreased. An increasing number of producers are
looking toward savings in feed costs and better marketing methods as
important aids in maintaining or increasing profits. The objective of
this study was to provide turkey producers with useful predictions of
least-cost rations and most profitable, or optimum, marketing weights
for a wide range of price relationships.
As necessary information in attaining this final objective, the first
step is one of predicting the feed- gain production function. From the
production function, in turn, it is necessary to predict the gain isoquants
showing the possible combinations of feed which will produce turkeys of
a specified weight. From the gain isoquants can be predicted marginal
substitution rates between the two major feed categories included in
this study. Isoclines may also be computed, which, when related to feed
price ratios, allow indication of rations which produce each pound of
gain at lowest cost. Finally, marginal rates of transformation of feed
into gain can be used to predict optimum marketing weights.
The emphasis of this study is on feed substitution and least- cost
rations in the latter part of the production period for turkeys. Emphasis
is placed on this portion of the production period because daily and total
feed intake is greatest then; major savings in cost can be made during
this period when feed consumption is largest. Also, it is during this
period that the greatest opportunity exists for substitution between feed
categories with least restriction in the growth of poults. Accordingly,
the experiment was designed mainly to allow prediction of the production
surface for the upper weight range (based on observations in the 12-24
week portion of the turkey production period). Though only a limited
number of observations were obtained at lighter weights, it was also
possible to predict production surfaces for weight intervals based on
observations in the 0-6 week and 6-12 week periods of production. The
authors look upon the estimates for lighter weights as having limitations
which do not attach to the predictions for heavier weights. However, it
is believed that these estimates are generally of more value for feeding
recommendations than data which heretofore have been available.
LEAST- COST RATIONS FOR TURKEYS 375
Production functions and feed substitution possibilities involve corn

and soybean oilmeal as the central resources of decision. However,
opportunities for feed substitution exist primarily when the ration is
properly fortified with the vitamins, minerals, and trace ingredients
explained later.
EXPERIMENT AND BASIC DATA
This study is based on an experiment in which 600 bronze turkey

poults were fed on alternative rations for a 24- week period. At the
start of the experiment, the 600 turkeys were randomly allotted to 48
different pens of 12 or 13 birds each; individual pens contained approx-
imately half males and half females. Eight pens (or replicates) of birds
were fed on each of 6 protein rations (21, 23, 25, 27, 29, and 31 per cent)
for the first 6 weeks of the experiment. At 6 weeks, the 600 birds were
completely rerandomized into 24 pens of 24 to 25 birds each with 4 pens
of birds fed on each of 6 protein rations (15, 17, 19, 21, 23, and 25 per
cent) for the 6-12 week period. At the end of 12 weeks, the 600 birds
were again rerandomized into 24 pens, with 4 pens fed on each of 6 pro-
tein rations (10, 12, 14, 16, 18, and 20 per cent) for the 12-24 week
period of the experiment.
Previous experiments with broilers and hogs indicated that there
were no important cumulative or “carry-over” effects of previous pro-
tein rations in meat production. That is, a bird or animal fed a 1 per
cent protein diet in an early period and a different per cent protein diet
in a later period tended, after a short adjustment period, to perform in
the later period as if it had received the second protein level through-
out the entire production period. For example, in the broiler study cited,
statistical analysis indicated no significant difference occurred in gains
for later periods between (a) birds carried through the entire production
period on a single ration and (b) birds changed to this ration from one
containing another percentage of protein. A comparable analysis for
the present study was not possible, since no turkeys were fed on one
protein ration for the entire 24- week period. However, it is not ex-
pected that the switch in rations causes outcomes to differ in later periods.
The average weight per poult at the start of the experiment was .11
pound, with each bird weighed thereafter at 3, 6, 12, 16, 20, and 24 weeks
of age. The average gain per bird and the corresponding average feed
inputs per bird were computed for each treatment and pen; these quan-
tities provided the observations used in the regression analysis which
follows.
Table 11.1 indicates the ingredients included in the various protein
rations used for the 0-6 week period. Corn and soybean oilmeal were
combined in various proportions with a fixed “basic” ration of other in-
gredients to provide protein levels ranging from 21 to 31 per cent.
Table 11.2 shows the rations used for the 6-12 week period; the “basic”
ration remained the same, while the quantities of corn and soybean
t
376 LEAST- COST RATIONS FOR TURKEYS
Table 11.1. Quantities of Various Ingredients Required for 100

Pounds of Different Protein Rations*

Ingredients
21 23 25 27 29 31
Corn 45 40 35 30 25 20
Wheat middlings 10 10 10 10 10 10
Bran 5 5 5 5 5 5
Soybean oilmeal 15 20 25 30 35 40
Fish meal 5 5 5 5 5 5
Meat scraps 5 5 5 5 5 5
Alfalfa meal 5 5 5 5 5 5
Dried whey 5 5 5 5 5 5
Minerals 4 4 4 4 4 4
Vitamin mix 1 1 1 1 1 1
These rations were fed to turkeys from 0 to 6 weeks of age.
Table 11.2. Quantities of Various Ingredients Required for 100

Pounds of Different Protein Rations

Ingredients
15 17 19 21 23 25
Corn 60 55 50 45 40 35
Wheat middlings 10 10 10 10 10 10
Bran 5 5 5 5 5 5
Soybean oilmeal 0 5 10 15 20 25
Fish meal 5 5 5 5 5 5
Meat scraps 5 5 5 5 5 5
Alfalfa meal 5 5 5 5 5 5
Dried whey 5 5 5 5 5 5
Minerals 4 4 4 4 4 4
Vitamin mix 1 1 1 1 1 1
f These rations were fed to turkeys from 6 to 12 weeks of age.
oilmeal were varied to provide rations ranging from 15 to 25 per cent

in protein.
DERIVATION OF PRODUCTION FUNCTIONS
The production functions used for later predictions are regression

equations for each interval over which specific rations were fed. Be-
cause of the way in which birds were reallocated to different rations, it
was impossible to predict a single over- all production function. There
was opportunity only for predicting either gain isoquants or production
functions based on observations for each of the age intervals of 0-6,
6-12, and 12-24 weeks. Practical use of substitution data is consistent
with estimation of production functions over particular intervals because
(a) rations which average lowest in cost for the total gain in the weight
interval (but which do not necessarily represent the lowest cost for
each ounce of gain) can be so predicted and (b) producers prefer to

change the ration only a few times over the total production period (i.e.,
one ration is selected for an interval of time, then a shift is made to
another ration to be used for some time, etc.). Because of the restricted
number of weighings and the fact that there was little difference in gains
among the rations over the lighter weight ranges, an alternative method
was devised as a check on the accuracy of the production functions fitted
to the 0-6 and 6-12 week observations. However, a greater number of
weighings and considerable difference in rates of gain and total gain
among rations caused this check to be unnecessary for functions fitted
to the 12-24 week observations.
A problem of autocorrelation arises in estimating the production
functions within each weight interval where several measurements were
taken from each pen of birds. To have independent observations along
a particular ration line, it would be necessary to feed different pens of
birds on each ration, with each pen being weighed and used only once as
an observation showing gains forthcoming from particular levels of feed
input (in contrast to the method used whereby the same pen was em-
ployed in prediction of gains associated with several levels of feed input
within an interval). The autocorrelation presents problems mostly for
probability statements and fiducial limits, rather than in prediction of
mean gains and substitution rates. That is, the presence of autocorrela-
tion does not present problems of predicting the relationship between
the dependent and independent variables, but does introduce problems
in making tests of significance. The effect of autocorrelation is. to re-
duce the effective number of independent observations; the number of
degrees of freedom which can be used for tests of significance in un-
correlated series is greater than it is when autocorrelation is present.
Hence, a problem exists in specifying the number of degrees of free-
dom upon which probability statements should be based. Calculation of
the autocorrelation coefficient and approximation of the effective num-
ber of degrees of freedom can be avoided by basing significance tests on
a minimum number of observations (to which the series might be re-
duced by calculating the autocorrelation coefficient). Since the obser-
vations on different pens are independent, the number of none or related
observations generally is equal to the number of pens. Where a null
hypothesis is rejected using this minimum number of degrees of free-
dom, it would certainly be rejected for the greater number of degrees
of freedom represented by all observations in the series.
Interval Functions
Because each pen of birds was fed on a constant protein ration from
12 to 24 weeks, it was possible to fit a production surface to this par-
ticular interval with greatest confidence. The 12-24 week interval is
one in which feed consumption is great and is also the relevant period
for marketing the birds. Thus, estimates were made of optimum
marketing weights, as well as of least-cost rations, using interval

functions fitted to the observations in the 12-24 week period.
A limited amount of information was available for the period up to
12 weeks of age. Only two weighings per pen (at 3 and 6 weeks) were
made before the birds were rerandomized and the protein levels changed
at the end of 6 weeks; only the 12- week weighing was made in the 6-12
week interval before the birds were rerandomized and the protein levels
changed again at 12 weeks. Hence, because of the limited number of
observations available at lighter weights, two alternative methods were
used in obtaining estimates for the 0-6 week and 6-12 week periods.
The first method attempted to predict, in the usual manner, the entire
production surface for each interval; from these surfaces isoquants and
marginal rates of substitution between corn and soybean oilmeal were
obtained. However, since the available observations tended to be
“clustered,” a second or alternative method was devised as a check on
the production surfaces. With this alternative procedure, gain isoquants
were computed for the average turkey weights at 3, 6, and 12 weeks, i.e.,
the isoquants were computed directly from the adjusted data, rather
than being derived from a previously estimated production surface.
Marginal rates of substitution between corn and soybean oilmeal were
then obtained along the “directly computed” isoquants. The check pro-
cedure simply involved comparing particular gain isoquants derived by
the two methods for (a) consistency of slopes or marginal rates of sub-
stitution and (b) consistency with respect to the various feed combina-
tions required to produce the specified gains.
A limitation of the alternative procedure involving direct estimation
of the gain isoquants is in deciding whether to minimize sums of squares
relative to corn or protein. Generally, corn inputs have been derived
as a function of protein inputs where direct estimation of gain isoquants
is involved. Since the alternative procedure is not used for predictive
purposes, but only as a check on the reasonableness of the production
surface, this limitation is not particularly serious. The data for the
12-24 week period are adequate for obtaining a reliable estimate of the
production surface. Hence, the alternative procedure described above
is not used for this period.
Functions Fitted
Fewer regression models were applied in this study than in the

broiler study reported in Chapter 10. The data, as explained earlier,
were not so well adapted for these purposes. The various production
functions and gain isoquants fitted as regression equations are ex-
plained below.
Regression equations for 0-3 week interval

Only the observations obtained from the 3-week weighing were
available for use in predictions over the 0-3 week interval. At this
early stage in the development of the bird, very little difference (abso-
lute or relative) occurred in the gains for birds fed on various protein
rations. A regression equation which predicts gain as a function of the
two categories of feed inputs gives a low coefficient of determination
1
(R value). Therefore, the alternative procedure of fitting gain iso-
quants directly to the data was used. With this procedure, all gains for
the 0-3 week interval were adjusted to the average 3-week gain of .57
pound; total feed quantities associated with each gain were then adjusted
in the same proportion and direction as the gain adjustment. Regression
equations 11.1 and 11.2 were fitted to these adjusted data, where corn is
expressed as a function of soybean oilmeal. These equations predict
(for .57 pound of gain) the quantity of corn consumed as a function of
the quantity of soybean oilmeal fed. In equations such as 11.1 and 11.2,
where corn is predicted as a function of soybean oilmeal consumption,
the symbols C and S refer to the pounds of corn and soybean oilmeal
consumed relative to a given gain (.57 pound in this case). In later
equations where gain is predicted as a function of the feed inputs, the
symbols C and S refer to total quantities of corn and soybean oilmeal
required to produce any specified gain.
7413
(11.1) C = .1671 S"*
2
(11.2) C = .8141 - 1.5800 S + .9757 S
2
Table 11.3 shows the coefficient of determination, R and the t ,
values for the regression coefficients in equations 11.1 and 11.2. The
R 2 and t values for both equations are 2
significant at the 1 per cent level,
and little difference occurs in the R values for the two equations.
Given the limitations mentioned previously, either equation may be used,
on a probability basis, for predicting substitution rates between corn
and soybean oilmeal along the .57-pound gain isoquant. In addition to
statistical "fit,” the logic of nutritional requirements and practicality
of feeding operations become the basis on which selection of a function
is made.
Table 11.3. K* and t Values for Regression Equations 11.1 and 11.2
Using Adjusted 3 -Week Gain and Feed Data
t Values for Regression Coefficients

Equation Value of Rz in the Order Shown in
Equations 11.1 and 11.2
11.1 .9308 24.93*
11.2 .9420 6.56* 2.85*
^Significant at the 1 per cent level with 46 degrees of freedom.

^Significant at the 1 per cent level with 45 degrees of freedom.
iThe function Y = .9162 C* 2005 S*

2620
,
which was obtained for the 0-3 week interval, has
a low R value of .7230.
Figure 11.1. Comparison of 0.57-pound gain isoquants,

predicted by regression equations 11.1 and 11.2.
Figure 11.1 shows the .57-pound gain isoquants computed from re-
gression equations 11.1 and 11.2 plotted against the adjusted observa-
tions. It should be pointed out that the dots of the scatter diagram in
Figure 11.1 represent the adjusted feed quantities required to produce
.57 pound of gain for the various rations and pens. In this case both
equations, which predict corn consumption as a function of soybean oil-
meal consumption, have been fitted to these adjusted observations.
Thus, the comparison of "closeness of fit* of the isoquants to the dots
in Figure 11.1 is relevant. However, in later cases where a function is
fitted as a production surface, the isoquants computed from this surface
cannot be compared directly with the adjusted dots.
Regression equations for the 0-6 week interval

Production functions 11.3, 11.4, and 11.5 were fitted to the accumu-
lated feed and gain quantities for the 0-6 week period. Since the average
gain per poult at 6 weeks of age varied considerably with the protein
ration fed, it was possible to fit a production surface indicating gain as
a function of the corn and soybean oilmeal fed. Equation 11.3 is a Cobb-
Douglas function of this type. Functions 11.4 and 11.5 were fitted by the
same process explained for the .57-pound gain isoquants at 3 weeks of
age, the gain and feed quantities used
i.e., were adjusted to the mean
gain of 2.33 pounds for the 6 -week period.
3847
(11.3) Y = 1.7167 O'1422 S'
- ' 7506
(11.4) C = 1.9512 S
2
(11.5) C = 3.3915 - 1.6659 S + .2668 S
LEAST-COST RATIONS FOR TURKEYS 381
Table 11.4. R2 and t Values for Regression Equations 11.3, 11.4, and 11.5
Using 6 -Week Gain and Feed Data*

Equation Value of R2 in the Order Shown in
Equations 11.3, 11.4, and 11.5
11.3 .8640 12.44* 16.83*
11.4 .9324 24.95*
11.5 .9461 6.94* 3.21*
*Equation 11.3 is computed from the actual quantities of feed and gain for the 0-6 week
period; equations 11.4 and 11.5 are computed from feed and gain quantities which
have been adjusted to a gain of 2.33 pounds,
f Significant at the 1 per cent level with 45 degrees of freedom.
$ Significant at the 1 per cent level with 46 degrees of freedom.
2
Table 11.4 presents the R and t values for regression equations
2
11.3, 11.4, and 11.5. While the R values for equations 11.4 and 11.5
are larger than the value for equation 11.3, it should be noted that equa-
tion 11.3 is a production surface computed from unadjusted data while
equations 11.4 and 11.5 are gain isoquants computed from adjusted data.
2
For equation 11.4 the proper interpretation of R is that 93.2 per cent
of the sum of squares of the adjusted corn quantities is explained by the
2
soybean oilmeal variable. For equation 11.3, however, R should be
interpreted as meaning that 86.4 per cent of the sum of squares of the
true gains is explained by the corn and soybean oilmeal variables. Thus,
2
the R values of regression equations 11.4 and 11.5 should not be com-
2
pared directly with the R value of equation 11.3.
Regression equations 11.4 and 11.5 predict directly the 2.33-pound
gain isoquants (i.e., isoquants representing a total turkey weight of 2.44
pounds, including the 11-pound initial weight). However, when using
.
production function 11.3, which represents a surface or family of iso-

quants, an isoquant equation must be derived for predicting the 2.33-
pound gain isoquant. Isoquant equation 11.6 is derived from production
2. 2614
Y
( 11 6 )
. C =
3647
1.7167 S'
function 11.3; the 2.33-pound gain isoquants resulting from equations

11.4, 11.5, and 11.6 are presented in Figure 11.2. While the isoquant
computed from equation 11.6 does not appear to fit the observations as
well as the isoquants of equations 11.4 and 11.5, the scatter dots shown
in Figure 11.2 have been adjusted to a constant gain of 2.33 pounds and
thus functions 11.4 and 11.5 are partially “forced into a better fit.”
Hence, the isoquant computed from equation 11.6 should not be com-
pared, in “closeness of fit to the scatter dots,” with the isoquants from
equations 11.4 and 11.5. The relevant comparison for equation 11.6
would be that of a family of gain isoquants compared with a set of unad-
justed observations. It is expected that .the 2.33-pound gain isoquant
(i.e., an isoquant representing a total weight of 2.44 pounds, including

the .11-pound initial weight) derived from equation 11.6 fits the unad-
justed data better than either equation 11.4 or 11.5. Since equation 11.3
is a production surface based on unadjusted data, it will be used for
predictive purposes; equations 11.4 and 11.5 provide some check on the
reliability of this surface. The high degree of consistency between the
2.33-pound gain isoquants fitted by the three different equations (Figure
11.2) provides a basis for increased confidence in using equation 11.3 for
predicting marginal rates of substitution and least-cost rations for the
0-6 week interval.
Figure 11.2. Comparison of 2.33-pound gain isoquants,

predicted by regression equations 11.4 and
11.5 and by isoquant equation 11.6.

Regression equations 11.7 and 11.8 were fitted to the data obtained
at the12-week weighing, i.e., to the gains and feed quantities for the
6-12 week interval. Thus, predicted gains for the 6-12 week period are
measured from a starting weight of 2.44 pounds, the average turkey
(11.7) Y = 1.7291 C 4998 S*

2531
(11.8) C = 9.6199 - 3.3851 S + .5219 S 2
weight at 6 weeks. Equation 11.7 predicts a production surface for the

6-12 week period computed from unadjusted data, with gain the dependent
variable and corn and soybean oilmeal the independent variables. Equa-
tion 11.8,however, was computed by the alternative check procedure of
deriving a gain isoquant directly from adjusted data, i.e., the equation
uses observations adjusted to an average gain of 4.45 pounds over the
6-12 week interval. (The 4.45-pound gain isoquant for the 6-12 week
interval represents a total turkey weight of approximately 6.89 pounds;
2.44 pounds weight at 6 weeks plus 4.45 pounds gain.) Again, equation
11.8 does not predict a production surface, but only a 4.45-pound gain
isoquant in the 6-12 week period. Comparison of this isoquant with the
4.45 pound gain isoquant derived from equation 11.7 provides a check
on the production surface predicted from equation 11.7.
Observations for four pens of birds fed on a 15 per cent protein
ration were omitted in fitting the Cobb-Douglas function, equation 11.7.
The four observations were not used because, at the 15 per cent protein
level, the quantity of soybean oilmeal in the ration is zero, i.e., the ob-
servation points fall directly on the corn axis. One of the mathematical
restrictions of the Cobb-Douglas function is that the gain isoquants can-
not intersect either the corn or the soybean oilmeal axis, i.e., the iso-
2
quants must be asymptotic to both axes. The quadratic function, equa-
tion 11.8, allows the 4.45-pound gain isoquant (4.45 pounds gain in the
6-12 week period) to intersect the corn and soybean oilmeal axes. This
equation, then, using all of the observations for the 6-12 week period
(including those for the 15 per cent ration) adjusted to a common gain
of 4.45 pounds was computed as a check on function 11.7. The R and t
2
values for the regression equations 11.7 and 11.8 are given in Table 11.5.
Table 11.5. R 2 andt Values for Regression Equations 11.7 and 11.8
Using Feed and Gain Data for the 6-12 Week Period*

11.7 .9340 5.64* 14.50*
11.8 .9628 11.59* 5.44*
*Equation 11.7 was fitted to actual data for the 6-12 week period; equation 11.8 was
fitted to feed and gain quantities* adjusted to a constant gain of 4.45 pounds for the
6-12 week period.
tSignificant at the 1 per cent level with 17 degrees of freedom.
t Significant at the 1 per cent level with 21 degrees of freedom.
Equation 11.9 is the isoquant equation derived from production func-

tion 11.7.The 4.45-pound gain isoquants (average gain from 6 to 12
weeks) derived from isoquant equation 11.9 and directly from equation
2
An alternative method was devised in an attempt to use the 15-per cent protein ration
observations. A very small quantity of soybean oilmeal (1 per cent of the ration) was as-
sumed for the 15-per cent ration in order that no observation points would fall directly on
the corn axis. Because the observation points were extremely close to the corn axis, how-
ever, the shape of the gain isoquants was distorted when these observations were used.
384 LEAST-COST RATIONS FOR TURKEYS
ll.Oi
lO.Of*
Quadratic Equation(ll.8)|
isoquant Equation(il.9)
9 0
.
\ ^
x
4
\W
7.0
^
0
«
1 60
o
0-
5.0
% s
v'\
4.0f-
~03 u5 ts to £5
Pounds of Soybean Oiimeal
Figure 11.3. Comparison of 4.45 pound gain isoquants

(gains on birds weighing 2.44 pounds),
predicted by quadratic equation 11.8 and
by isoquant equation 11.9.
11.8 are plotted in Figure 11.3. These two contours have quite consistent
slopes except at the extreme upper ends (for protein rations of 15 to 17
2*0008
Y
(11.9) C =
2531
1.7291 S‘
per cent). The influence of the adjusted observations for the 15 per cent
protein ration (falling on the corn axis) forces the upper portion of the
isoquant from function 11.8 down, relative to the isoquant from equation
11.9. However, the divergence is probably exaggerated because gains
are extremely low on the 15 per cent protein ration. Hence, consider-
able inaccuracy may arise in the method of adjusting the gain and feed
data for this ration to a common gain of 4.45 pounds for the 6-12 week
period. For example, assume that 5.0 pounds of corn (with no soybean
oiimeal) is required to produce 2.225 pounds of gain on the 15 per cent
ration. Using the adjustment procedure, 10.0 pounds of corn are then
assumed to produce 4.45 pounds of gain, a doubtful conclusion. Dimin-
ishing returns to corn are more consistent with nutritional theory than
the constant returns used in the above adjustment. Thus, the observa-
tions for the 15 per cent ration should probably fall at greater quantities
on the corn axis, forcing the upper portion of the isoquant from equation
11.8 to become more consistent with the isoquant from equation 11.9.
Remember that the dots of the scatter diagram in Figure 11.3 are not
the observations to which equation 11.7 is fitted; these dots represent

only the adjusted means to which equation 11.8 was fitted. Observations
for equation 11.7 would have, if they could be presented simply, a scat-
ter more consistent with the 4.45-pound gain contour derived from iso-
quant equation 11.9. Or, again, the relevant comparison for equation
11.9 would be a family of contours related to the set of unadjusted ob-
servations. Because an entire production surface for the 6-12 week
period is given by equation 11.7, this function will serve as a basis for
prediction in the 6-12 week interval. It should be remembered, how-
ever, that this function has, starting at the origin of feed inputs for the
period, linear isoclines. Hence, the marginal substitution rates and
least-cost rations so specified are those which are “average” for the
interval. However, specification of optimum “average” rations over
weight intervals is one objective of this study.

The data for the 12-24 week period is adequate for estimation of an
interval production surface, with gain as a function of the two feed cate-
gories. Hence, the simple contour equations estimated as a check pro-
cedure for the 0-6 week and 6-12 week periods have not been computed
for the 12-24 week interval. Three different types of functions were
fitted to the gain and feed data for the 12-24 week interval: a Cobb-
Douglas function, equation 11.10, a square root function, equation 11.11,
and a quadratic function, equation 11.12. The gains predicted are those
beyond the average 12-week weight of 6.93 pounds.
5108 2517
(11.10) Y = 1.0764 c* s*
(11.11) Y = -2.8884 + .0450 C - .2966 S + .9894 Yc + 2.4592 Ys~ + .1284 VCS

2 2
(11.12) Y = .0148 + .1838 C + .8837 S + .0001C - .0214 S - .0040 CS
The 12-24 week observations for the 10 per cent protein ration (with
no soybean oilmeal included) are not used in computing the Cobb-Douglas
function, equation 11.10 for the reason given previously; use of obser-
vations falling on the corn axis distorts the gain isoquants. However,
all the data are used in computing functions 11.11 and 11.12.
2
Table 11.6 shows the t and R values for regression equations 11.10,
11.11, and 11.12. Equations 11.10 and 11.11 will be used in predicting
economic quantities at later points in the study. Quadratic crossproduct
function 11.12 is not used for later predictions because it contains one
term which is statistically nonsignificant even at the 50 per cent level
and it explains less of the deviations of the dependent variable, Y, than
equations 11.10 and 11.11, as shown by the R values of Table 11.6.
2
While one of the regression coefficients in equation 11.11 is significant

at the 20 per cent level of probability, all five terms are used, on grounds
of nutrition logic, for estimating the production surface and making pre-
dictions. Dropping the nonsignificant term and recomputing the equation
gives a slightly lower R2 value, with all terms highly significant.
However, predictions differ only by minute quantities when the term is

or is not used.
Table 11.6. R2 and t Values for Regression Equations 11.10, 11.11, and 11.12
Using Feed and Gain Data for the 12-24 Week Period

Equations 11.10, 11.11, and 11.12
11.10 .9838 29.70* 21.10*
11.11 .9936 1.40** 9.12* 3.80* 15.23* 2.87*
11.12 .9730 4.96* 14.71* 0.16* 7.36* 1.67**
*Significant at the 1 per cent level with 17 degrees of freedom,

t Significant at the 1 per cent level with 18 degrees of freedom.
^^Significant at the 20 per cent level with 18 degrees of freedom.
^Nonsignificant at the 50 per cent level with 18 degrees of freedom.
Gain isoquant equations for the three production functions 11.10,

11.11, and 11.12 are shown, respectively, in equations 11.13, 11.14, and
11.15. Equations 11.13 and 11.14 were used in predicting the gain iso-
quants of Figure 11.4, which shows three pairs of isoquants for the
average turkey gain (in the 12-24 week interval) at 16, 20, and 24 weeks
of age. The contours shown in Figure 11.4 are for gains starting from
an average weight of 6.93 pounds at 12 weeks of age.
Figure 11.4. Comparison of gain isoquants predicted

by isoquant equations 11.13 and 11.14.
1.9577
Y
(11.13) C =
* 2517
1.0764 S
(11.14) C = (
- 10.9933 - 1.4267 Vs
V.0699 S 2
+ 11.1111 - .1886 ]/s~ + .1800 Y + 1.4988)
(11.15) C = - 712.5504 + 15.4961 S

2
+ 3875.9690 V . 000027 S - .0019 S + .0005 Y + .0338
In connection with the isoquants of Figure 11.4, it should be re-

membered that the Cobb-Douglas function requires equal slopes for all
isoquants along any straight line through the origin (ration line). Hence,
this function tends to “average out” fluctuations over the input-output
surface. The square root and quadratic functions are not subject to the
restriction of constant slope along ration lines. Consequently, these
types of functions provide a closer “fit” to data which are not consistent
with the assumptions of constant slopes of isoquants at the points where
they are intersected by any one ration line. The above restriction on
the Cobb-Douglas function helps explain the difference in slopes along
the two 12.75-pound gain isoquants shown in Figure 11.4. At high pro-
tein levels, to conform to the above-mentioned restriction, the isoquant
computed from the Cobb-Douglas function is “pulled down” relative to
the isoquant computed from the square root function. The slopes of
contours from the two functions are quite similar at the lower protein
levels; least- cost rations predicted from them would also be similar.
If interest is in predicting a least- cost ration which changes with in-
creasing weight within the 12-24 week interval, the square root function
should be used since it allows the slope of the isoquants to change along
a ration line. Too, it expresses, as is generally believed to be the case,
lower rates of substitution of soybean oilmeal for corn as the bird ap-
proaches maturity.
Comprehensive tables of least-cost rations computed from Cobb-
Douglas function, equation 11.10, are given in the following section
since it is believed that the majority of turkey producers are interested
in a single “average” least- cost ration to be fed for the entire 12-24
week interval. However, because turkey production is becoming more
and more a specialized enterprise, an increasing number of producers
are interested in changing rations more frequently to obtain small sav-
ings in feed costs per bird.
Table 11.7 summarizes the marginal rates of substitution and com-
binations of corn and soybean oilmeal required for various gains in the
third weight interval, as predicted from square root function, equation
11.11. Columns (7), (8), and (9) in Table 11.7 show marginal rates of
substitution of soybean oilmeal for corn along gain isoquants of 4.50,
9.00, and' 12.75pounds in the third weight interval. Since curvature is
allowed in the isoclines of the square root function, the substitution
Table 11.7. Marginal Rates of Substitution and Combinations of Corn and Soybean Oilmeal for Various
Gains mthe Third Weight Interval — Computed from Equation 11.11
Pounds of Pounds of Pounds of

Corn and Corn and Corn and
Soybean Soybean Soybean Marginal Rates of Substitution
Per cent Protein Oilmeal for Oilmeal for Oilmeal for Along Gain Isoquants of
in the Ration 4.50 Pounds 9.00 Pounds 12.75 Pounds
Gam* Gain* Gam*
Soybean Soybean Soybean
Corn Corn Corn 4.50 lbs.* 9.00 lbs.* 12.75 lbs.*
Oilmeal Oilmeal Oilmeal
(1) (2) (3) (4) (5) (6) (7) (8) (9)
10 34.7 0 74.5 0 113.4 0 ~ —

11 18.2 0.7 41.0 1.6 64.2 2.6 8.41 7.29 6.62
12 14.8 1.2 34.1 2.8 54.3 4.5 5.39 4.42 3.86
13 12.8 1.7 30.1 3.9 48.5 6.3 4.03 3.16 2.66

14 11.4 2.1 27.2 5.0 44.5 8.1 3.23 2.43 1.95
15 10.2 2.4 25.0 6.0 41.7 9.9 2.70 1.93 1.47
16 9.4 2.8 23.3 7.0 39.1 11.7 2.28 1.56 1.13

17 8,6 3.2 21.8 8.0 37.2 13.7 1.98 1.28 0.86
18 8.0 3.6 20.6 9.2 35.5 15.8 1.72 1.05 0.64
19 7.3 3.9 19.4 10.3 34.3 18.2 1.52 0.86 0.45

20 6.8 4.3 18.5 11.5 33.3 20.8 1.33 0.70 0.30
Gains measured from an average weight of 6.93 pounds at 12 weeks.
rates along particular ration lines change as weight increases. The

least-cost ration for attaining each of the three levels of gain in Table
11.7 is determined by locating the marginal rates of substitution in
columns (7), (8), and (9) which most nearly equal the soybean oilmeal
to corn price ratio. Thus, the least- cost ration for a particular price
ratio may change three times as weight increases over the upper
weight range.
Input-output curves and choice of function
The degree of conformity of the several functions in predicting

single ration input-output curves is suggested in figures 11.5 and 11.6
for 14 and 20 per cent protein levels. The slopes of the curves are
most nearly comparable for low protein levels. At higher protein levels,
the quadratic function « rounds off” too rapidly, while the Cobb-Douglas
may “overshoot.”
The choice of a particular function to be used in making various es-
timates is based on several considerations. Foremost among these
considerations are (a) the statistical and plotted “fit” of the functions
to the data and (b) the practical aspects of applying the results to farm
conditions. No single function for the 12-24 week interval appeared to
best meet these considerations for all types of predictions. Conse-
quently, in following sections the Cobb-Douglas function is used to pre-
dict the best “average” least-cost rations over the 12-24 week interval.
Since this function has a constant slope along a particular ration line, it
gives a single least- cost ration, as an “average” over the feeding period,
for any given price ratio between corn and soybean oilmeal. Functions
LEAST -COST RATIONS FOR TURKE YS 389
17 . 5 .
-i 1 1 1 1
1 1 r-
15.0-
C0BB - DOUGLAS FUNCTION (IU0)
ll2.5(- QUADRATIC FUNCTION (11.12)

O
SQUARE ROOT FUNCTION (II.IIJ.-
o 10.01-
to
7.51-
|
o
0.
5.0-
2.5-
0 5 10 15 20
25 30 35 40 45 50 55 60
POUNDS OF FEED (CORN + SOYBEAN OILMEAL)
FED ON A 14 PERCENT RATION
Figure 11.5. Comparison of input-output curves for a 14 per cent protein ration
by three equations.
FED ON A 20 PERCENT RATION

Figure 11.6. Comparison of input-output curves for 20 per cent protein ration by
three equations.
11.11 and 11,12 provide more accurate least-cost rations than Cobb-
Douglas function 11.10 if the per cent of protein is changed several
times within the 12-24 week interval. However, because of the cost
and inconvenience of frequently adjusting the protein level within a rela-
tively short time period, many producers probably prefer to feed only
one ration in the upper weight range.
While the Cobb-Douglas function is used to predict average least-
cost rations in the 12-24 week interval, the square root function is used
in predicting optimum marketing weights: it fits the plotted input- output
data more closely than either quadratic or Cobb-Douglas functions.
While the Cobb-Douglas function appears to be satisfactory in predict-
ing the average slope or curvature of the gain isoquants, it tends to
overestimate the slope for large feed inputs.
AVERAGE MARGINAL RATES OF SUBSTITUTION

FOR COBB-DOUGLAS
Returning to one objective of this study, predicting rations which
average least- cost over a feeding interval, we use quantities from the
Cobb-Douglas equations in this section. Marginal rates of substitution
and total feed quantities to produce total gains over each of the weight
intervals are shown in tables 11.8, 11.9, and 11.10. The substitution
and isoquant data are derived from the Cobb-Douglas function for each
feeding and weight interval indicated below the tables.
Table 11.8. Combinations of Corn and Soybean Oilmeal Required to

Produce Turkeys Weighing 2.44 Pounds, and
Marginal Rates of Substitution Between
Feeds in This Weight Range*
Pounds of Feed to Produce

Turkeys Weighing Marginal Rates
Per Cent of Substitution
2.44 Pounds
Protein of Soybean
in the Soybean Oilmeal for
Ration Corn oilmeal Corn
21 2.40 0.80 2.47

22 2.18 0.90 2.00
23 2.00 1.01 1.65
24 1.84 1.10 1.37
25 1.70 1.21 1.15
26 1.57 1.33 0.97
27 1.46 1.46 0.82
28 1.35 1.60 0.70
29 1.25 1.75 0.59
30 1.16 1.93 0.49
31 1.07 2.14 0.41
The figures in this table are derived from Cobb-Douglas function 11.3,
computed from observations for the 0-6 week period.

Increase Turkeys From 2.44 Pounds to 6.93 Pounds
Liveweight and Marginal Rates of Substitution
Between Feeds in This Weight Range*
Pounds of Feed to Increase

Turkey Weight From Marginal Rates
Per Cent
2.44 Pounds to 6.93 of Substitution
Pounds of Soybean
in the
Oilmeal for
Ration Soybean
Corn
Corn oilmeal
16 10.11 0.43 11.65

17 7.86 0.72 5.57
18 6.75 0.97 3.55
19 6.03 1.21 2.53

20 5.50 1.45 1.92
21 5.08 1.69 1.52
22 4.73 1.95 1.23

23 4.43 2.23 1.01
24 4.17 2.50 0.84
25 3.93 2.81 0.71
*The figures in this table are derived from Cobb- Douglas function 11.7,

Increase Turkeys From 6.93 Pounds to 19.84 Pounds
Liveweight and Marginal Rates of Substitution
Between Feeds in This Weight Range*
Pounds of Feed to Increase

Turkey Weight From Marginal Rates
Per Cent
6.9 Pounds to 19.84 of Substitution
Protein
Pounds of Soybean
in the
Oilmeal for
Ration Soybean
Corn
Corn oilmeal
11 74.03 2.96 12.32

12 58.18 4.85 5.91
13 50.17 6.54 3.78
14 44.89 8.16 2.71

15 41.09 9.79 2.07
16 38.06 11.42 1.64
17 35.58 13.11 1.34

18 33.44 14.86 1.11
19 31.57 16.71 0.93
20 29.88 18.68 0.79
*The figures in this table are derived from Cobb- Douglas function 11.10,
Table 11.11. Combinations of Corn and Soybean Oilmeal Required To Produce

1 Pound of Gain on Turkeys Weighing 2.44, 6.93, and 19.84
Pounds
Pounds Feed for Pounds Feed for Pounds Feed for

Per Cent 1 Pound Gain on 1 Pound Gain on 1 Pound Gain on
Protein 2.44 Pound 6.93 Pound 19.84 Pound

in the Turkeys* Turkey st Turkeys t
Ration
Soybean Soybean Soybean
Corn Corn Corn
oilmeal oilmeal oilmeal
11 7.71 0.31
12 — — — -- 6.06 0.50
13 __ — — — 5.22 0.68
14 — — — -- 4.68 0.85
15 — — — -- 4.28 1.02
16 3.12 0.13 3.96 1.19

17 — — 2.43 0.22 3.71 1.37
18 — — 2.09 0.30 3.48 1.55
19 — — 1.86 0.37 3.29 1.74
20 — — 1.70 0.45 3.11 1.94
21 1.34 0.45 1.60 0.52

22 1.21 0.50 1.46 0.60 — —
23 1.11 0.56 1.37 0.68 — —
24 1.02 0.61 1.29 0.77 — —
25 0.95 0.68 1.21 0.87 — —
26 0.88 0.74
27 0.81 0.81 — —
28 0.75 0.89 — — — —
29 0.70 0.98 __
30 0.65 1.08 — —
31 0.59 1.19 — — — --
*Feed quantities predicted from Cobb-Douglas function 11.3.

^Feed quantities predicted from Cobb-Douglas function 11.7.
*Feed quantities predicted from Cobb-Douglas function 11.10.
As expected from nutritional logic and previous knowledge, pre-

is
dictions from the interval functions show that, for a given ration, the
marginal rates of substitution of soybean oilmeal for corn decline as
the bird increases in weight. This point is shown in substitution equa-
tions 11.16, 11.17, and 11.18 where the constants are -.8247, -.5065,
and -.4928 for the three successive weight intervals. A pound of soy-
bean oilmeal replaces 2.47 pounds of corn for turkeys fed on a 21 per
cent protein ration in the first weight interval (Table 11.8); on this
same ration a pound of soybean oilmeal replaces only 1.52 pounds of
corn for turkeys in the second weight interval (Table 11.9), Tables
11.9 and 11.10 indicate that with 18 per cent of protein in the ration,
one pound of soybean oilmeal replaces 3.55 pounds of corn for birds in
the second weight interval, but replaces only 1.11 pounds of corn for
birds in the third weight interval. These results occur because the bird
requires more protein relative to carbohydrates in the early growing
stages and more carbohydrates relative to protein as maturity and the

finishing period approach.
ac
(11.16) = -.825 CS"
6S
ac = -.507 CS"
(11.17)
as
ac = -.493 -3
(11.18) CS
as
Too, as each of the tables above show for the Cobb-Douglas functions,
and also as is shown in Table 11.7 for the square root function, the mar-
ginal rate of substitution of soybean oilmeal for corn declines as rela-
tively more protein is included in the ration for a particular level of
gain. (Or, conversely, the marginal rate of substitution of corn for soy-
bean oilmeal declines as the ration contains relatively less protein and
relatively more carbohydrates.) Since the marginal rates of substitu-
tion between the two feed inputs are diminishing, unique rations can be
found which minimize the cost of feed for a particular level of gain.
Table 11.11, based on Cobb-Douglas functions 11.3, 11.7, and 11.10,
shows the amount of feed to produce one pound of gain at three weight
levels. With a 22 per cent protein ration, only 1.71 pounds of feed (1.21
pounds of corn and .50 pound of soybean oilmeal) are required to produce
an additional pound of gain on birds weighing 2.44 pounds. However, to
produce an additional pound of gain on birds weighing 6.93 pounds using
a 22 per cent protein ration, 2.06 pounds of feed (1.46 pounds of corn
and .60 pound of soybean oilmeal) are required. Greater quantities of
feed per pound of additional gain are required as birds reach 19.84
pounds liveweight.
Table 11.12. Per Cent Protein in Rations Which Are Least-Cost for Turkeys From .11 Pound to
2.44 Pounds, With Various Corn and Soybean Oilmeal Prices*
Price of Corn
Price of Soybean Oilmeal in Cents per Pound
m Cents
per Pound 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 5.25 5.50 5.75 6.00 6.25 6.50
1.4 22.0 21.5 21.5 21.0

1.6 22.5 22.0 21.5 21.5 21.0 — -- -- — — — — — — —
1.8 23.0 22.5 22.0 22.0 21.5 21.5 21.0 — — — — — — — —
2.0 23.5 23.0 22.5 22.5 22.0 21.5 21.5 21.0 21.0 — — — — — —
2.2 24.0 23.5 23.0 23.0 22.5 22.0 22.0 21.5 21.5 21.0 21.0 — — — —
2.4 24.5 24.0 23.5 23.5 23.0 22.5 22.5 22.0 22.0 21.5 21.5 21.0 21.0 — —
2.6 25.0 24.5 24.0 24.0 23.5 23.0 23.0 22.5 22.0 22.0 21.5 21.5 21.5 21.0 21.0
2.8 25.5 25.0 24.5 24.5 24.0 23.5 23.0 23.0 22.5 22.5 22.0 22.0 21.5 21.5 21.5
3.0 26.0 25.5 25.0 24.5 24.0 24.0 23.5 23.5 23.0 23.0 22.5 22.5 22.0 22.0 21.5
3.2 26.0 26.0 25.5 25.0 24.5 24.5 24.0 23.5 23.5 23.0 23.0 22.5 22.5 22.0 22.0
3.4 26.5 26.0 25.5 25.5 25.0 24.5 24.5 24.0 23.5 23.5 23.0 23.0 22.5 22.5 22.0
3.6 27.0 26.5 26.0 25.5 25.0 25.0 24.5 24.5 24.0 23.5 23.5 23.0 23.0 23.0 22.5
3.8 27.5 27.0 26.5 26.0 25.5 25.0 25.0 24.5 24.5 24.0 23.5 23.5 23.0 23.0 23.0
4.0 27.5 27.0 26.5 26.5 26.0 25.5 25.0 25.0 24.5 24.5 24.0 24.0 23.5 23.5 23.0
Computed from substitution equation 6C/SS = -0.825 CS" 1 .

Table 11.13. Per Cent Protein m

Rations Which Are Least-Cost for Turkeys From 2.44 Pounds to
6.93 Pounds, With Various Corn and Soybean Oilmeal Prices*
Price of Corn Price of Soybean Oilmeal m Cents per Pound

per Pound 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 5.25 5,50 5.75 6.00 6.25 6.50
1.4 20.0 19.5 19.5 19.0 19.0 18.5 18.5 18.5 18.0 18.0 18.0 18.0 17.5 17.5 17.5
1.6 20.0 20.0 19.5 19.5 19.0 19.0 18.5 18.5 18.5 18.5 18.0 18.0 18.0 18.0 17.5
1.8 20.5 20.5 20.0 19.5 19.5 19.5 19.0 19.0 18.5 18.5 18.5 18.5 18.0 18.0 18.0
2.0 21.0 21.0 20.5 20.0 20.0 19.5 19.5 19.5 19.0 19.0 18.5 18.5 18.5 18.5 18.5
2.2 21.5 21.0 21.0 20.5 20.0 20.0 19.5 19.5 19.5 19.0 19.0 19.0 19.0 18.5 18.5
2.4 22.0 21.5 21.0 21.0 20.5 20.5 20.0 20.0 19.5 19.5 19.5 19.0 19.0 19.0 19.0
2.6 22.5 22.0 21.5 21.0 21.0 20.5 20.5 20.5 20.0 20.0 19.5 19.5 19.5 19.0 19.0
2.8 22.5 22.5 22.0 21.5 21.5 21.0 20.5 20.5 20.5 20.0 20.0 20.0 19.5 19.5 19.5
3.0 23.0 22.5 22.5 22.0 21.5 21.5 21.0 21.0 20.5 20.5 20.0 20.0 20.0 19.5 19.5
3.2 23.5 23.0 22.5 22.5 22.0 21.5 21.5 21.0 21.0 20.5 20.5 20.5 20.0 20.0 20.0
3.4 23.5 23.5 23.0 22.5 22.0 22.0 21.5 21.5 21.0 21.0 20.5 20.5 20.5 20.0 20.0
3.6 24.0 23.5 23.0 23.0 22.5 22.5 22.0 21.5 21.5 21.5 21.0 21.0 20.5 20.5 20.5
3.8 24.5 24.0 23.5 23.0 23.0 22.5 22.0 22.0 21.5 21.5 21.0 21.0 21.0 20.5 20.5
4.0 24.5 24.0 24.0 23.5 23.0 23.0 22.5 22.0 22.0 21.5 21.5 21.5 21.0 21.0 20.5
1
Computed from substitution equation SC/6S = -0.507 CS" .
Table 11.14. Per Cent Protein in Rations Which Are Least-Cost for Turkeys From 6.93 Pounds to
Finished Weight, With Various Corn and Soybean Oilmeal Prices*
Price of Soybean Oilmeal m Cents per Pound

per Pound 3.00 3.25 3.50 3.75 4.00 4.25 4.50 4.75 5.00 5.25 5.50 5.75 6.00 6.25 6.50
1.4 15.0 15.0 14.5 14.5 14.0 14.0 13.5 13.5 13.5 13.5 13.0 13.0 13.0 13.0 12.5
1.6 15.5 15.0 15.0 14.5 14.5 14.0 14.0 13.5 13.5 13.5 13.5 13.0 13.0 13.0 13.0
1.8 16.0 15.5 15.5 15.0 14.5 14.5 14.5 14.0 14.0 14.0 13.5 13.5 13.5 13.5 13.0
2.0 16.5 16.0 15.5 15.5 15.0 15.0 14 5 14.5 14.5 14.0 14.0 14.0 13.5 13.5 13.5
2.2 17.0 16.5 16.0 16,0 15.5 15.5 15.0 15.0 14.5 14.5 14.5 14.0 14.0 14.0 13.5
2.4 17.5 17,0 16.5 16.5 16.0 15.5 15.5 15.0 15.0 15.0 14.5 14.5 14.5 14.0 14.0
2.6 18.0 17.5 17.0 16.5 16.5 16.0 16.0 15.5 15.5 15.0 15.0 15.0 14.5 14.5 14.5
2.8 18.0 18.0 17.5 17.0 16.5 16.5 16.0 16.0 15.5 15.5 15.0 15.0 15.0 15.0 14.5
3.0 18.5 18.0 17.5 17.5 17.0 17.0 16.5 16.0 16.0 16.0 15.5 15.5 15.0 15.0 15.0
3.2 19.0 18.5 18.0 18.0 17.5 17.0 17.0 16.5 16.0 16.0 16.0 15.5 15.5 15.5 15.0
3.4 19.5 19.0 18.5 18.0 17.5 17.5 17.0 17.0 16.5 16.5 16.0 16.0 15.5 15.5 15.5
3.6 19.5 19.5 19.0 18.5 18.0 17.5 17.5 17.0 17.0 16.5 16.5 16.0 16.0 16.0 15.5
3.8 20.0 19.5 19.0 18.5 18.0 18.0 17.5 17.5 17.0 17.0 16.5 16.5 16.0 16.0 16.0
4.0 20.0 20.0 19.5 19.0 18.5 18.5 18.0 17.5 17.5 17.0 17.0 16.5 16.5 16.5 16.0
Computed from substitution equation 5C/6S = -0.493 CS" 1 .
LEAST- COST RATIONS
The data of tables 11.12, 11.13, and 11.14, predicted from the sub-
stitution equations derived from the Cobb-Douglas functions of the
preceding section, provide estimates of rations which “average” least
in cost (for various corn and soybean oilmeal prices) over each of the
three specified weight intervals. The figures of Table 11.12, predicted
from substitution equation 11.16, are estimates of the least-cost rations
in the .11-pound to 2.44-pound weight interval. Substitution equations
11.17 and 11.18 are used, respectively, to provide the least-cost esti-
mates given in tables 11.13 and 11.14 for the 2.44-pound to 6.93-pound
LEAST-COST RATIONS FOR TURKE YS 395
interval and the 6.93-pound to finished weight interval. The least-cost

ration in each of these weight intervals is found by equating the mar-
ginal rate of substitution over the appropriate weight range with the
inverse price ratio of the feeds, then solving for the ratio of corn to
soybean oilmeal in the ration.
Figure 11.7. Least- cost rations for three weight intervals,

predicted by Cobb-Douglas functions with a
soybean oilmeal to corn price ratio of 2.0.
The estimate of tables 11.12, 11.13, and 11.14 may be used by turkey
producers as follows: with corn at $1.23 per bushel (2.2 cents per pound)
and soybean oilmeal at $4.50 per hundred pounds (4.5 cents per pound),
3
the inverse price ratio is -4.5/2. 2 or -2. 05. With this price ratio, the
predicted least-cost ration contains 22.0 per cent protein for the first
weight interval (Table 11.12), 20.0 per cent protein for the second weight
interval (Table 11.13), and 15.0 per cent protein for the third weight
interval (Table 11.14). In the first weight interval, the producer might
choose to feed a slightly higher level of protein than given by the least-
cost ration. From a practical standpoint, savings by a least- cost ration
in the first weight interval are small, and the producer might not want
to risk slower gains from a low protein ration. However, in the second
3
As pointed out previously, a negative sign is attached to the price ratio because the
price line (on a typical two-dimensional diagram) is negatively sloping.
«
LEAST- COST RATIONS FOR TURKEYS
Pounds of Soybeon Oiimeal
Figure 11.8. Least-cost rations for three weight intervals,

predicted by Cobb-Douglas functions with
various soybean oiimeal to corn price ratios.
Soybean Optimum Estimated Quantities of Corn
Oiimeal- Percent and Soybean Oiimeal per
Protein 100 Lbs. of Feed for
in the Various Protein Rations
Soybean Oiimeal
(pounds) (pounds
($) \- 2! 0(
45 00 15.00
1-
21 0
b 21 5
Lbs.
1* 43.75 16.25
b 21 5
1 22.0
IOO -42.50 17.50
b 22 0
per
b 22 5
41 .25 18.75
} 22 5 }
V 23 0 40 oo 20.00
Oiimeal
V 23
38 75
. 21 .25
b 23 5J.
b~ 24 0 37 50 22.50
Soyean
} 24 5 36.25 23.75
L
r — oc O™. - no on
of r 25. 5 33«5 » 26*25
V 26. 0 32.50 27.50
Price
2 QQL-UJLXirJ 11,11. ,I. T I
roo 2.oo
Price of Corn per 100 Lbs. ($)
Figure 11.9, Least-cost rations for the 0.11-pound to 2. 44 -pound weight interval
predicted by Cobb-Douglas function 11.3 with various corn and
soybean oiimeal prices.
weight interval, and particularly in the third weight interval, substantial

savings in feed costs may be realized by using a least-cost ration rather
than one which produces faster gains. Under certain price relation-
ships, of course, a least- cost ration may also produce the most rapid
gains.
Soybean Optimum Estimated Quan tities of Corn
Qiimeal- Percent and Soybean Oilmeal per
Corn Protein 100 Lbs. of Feed for
Price in the Various Protein Rations
§ 10.00
JE
6 8.00
l 45.00
Price of Corn per 100 Lbs. (|)
Figure 11.10. Least-cost rations for the 2.44-pound to 6.93-pound weight

interval, predicted byCobb- Douglas function 11.7 with
various corn and soybean oilmeal prices.
The producer may wish to make further adjustments within the third
weight interval to reduce feed costs. For example, during the first few
weeks of the third weight interval, the producer may wish to feed a
slightly higher protein level than prescribed by the least- cost ration;
he may wish to decrease this protein level as the birds increase to-
wards marketing weight. Table 11.7, and equation 11.11 upon which it is
based, can be used for this purpose.
Soybean Optimum Estimated Quantities of Corn
Oilmeal- Percent and Soybean Oilmeal per
Corn Protein IOO Lbs. of Feed for
Price in the Ve nous Proteip Rations
Ratio Ration Corn Soybean Oilmeal
(pounds} (pounds)
Figure 11.11. Least-cost rations for the 6.93-pound to finished weight interval,
predicted by Cobb-Douglas function 11.10 with various corn and
soybean oilmeal prices.
Referring again to tables 11.12, 11.13, and 11.14, if the price of soy-
bean oilmeal should rise to 5.5 cents per pound, with corn remaining at
2.2 cents per pound, the price ratio becomes -2.50. Least-cost rations
then contain 21.0 per cent protein for the first interval, 19.0 per cent
for the second interval and 14.5 per cent protein for the third interval.
With a price of 2.0 cents per pound for corn and 5.0 cents for soybean
oilmeal, the least-cost rations also would contain 21.0, 19.0, and 14.5
per cent protein since the price ratio is still -2.50.
Graphic illustration of changes in average least- cost rations between
weight intervals for a price ratio of -2.0 is produced in Figure 11.7.
The line passing through the origin extending to and intersecting the
2.44-pound weight isoquant represents a 22 per cent protein ration line.
Using the intersection point on the 2.44-pound isoquant as a new origin
(circled), the least- cost ration for the second weight interval contains
20 per cent protein. Again using the intersection point of the 20 per
cent ration line with the 6.93-pound isoquant as a new origin (circled),
the least-cost ration for the third weight interval contains 15 per cent
4
protein. Figure 11.8 illustrates the least- cost ration path (expansion
path) for the entire feeding period for various price ratios. As expected
from nutritional logic, the percentage of protein in the ration consist-
ently decreases with each higher weight interval, regardless of the ex-
isting price ratio.
Turkey- Optimum Turkey- Optimum

Feed Marketing Feed Marketing
Price Weights Price Weights
Ratio ilbsl
22.7
21.2
19.8
28.5
17.2
16.0
Figure 11.12. Optimum marketing Figure 11.13. Optimum marketing

weights for turkeys fed weights for turkeys fed
on a 13 per cent protein on a 15 per cent protein
ration, predicted by ration, predicted by
square root function. square root function.
Simple Graphical Indication of Least- Cost Rations
Figures 11.9, 11.10, and 11.11 have been included to allow simple
graphical selection of least- cost rations over the three weight intervals.
These graphs assume linear segments along gain isoquants and indicate
4
The reason the slope of the 15 per cent ration line in Figure 11.7 is not steeper relative
to the 20 and 22 per cent ration lines is as follows; some of the high protein ingredients of
the “basic* ration are reduced in quantity or removed entirely at the start of the third
weight interval, thus requiring more soybean oilmeal relative to corn for a given percentage
of protein in the ration.
least-cost rations for price ratios falling within the diagonal price rays
shown. Figures 11.9, 11.10, and 11.11 may be used as follows: suppose
the price of corn is 2.5 cents per pound and the price of soybean oilmeal
is 4.1 cents per pound. These prices are located at point A in Figure
11.9. Following to the right of the diagram between the two diagonal
lines, it is found that the least-cost ration over the first weight interval
contains 23.0 per cent protein. One hundred pounds of the 23.0 per cent
ration may be formulated by mixing 40.0 pounds of corn, 20.0 pounds of
soybean oilmeal and 40.0 pounds of the basic ingredients shown in Table
11.1. The above feed prices are also found at point B, Figure 11.10, and
specify a least-cost ration containing 20.5 per cent protein for the sec-
ond weight interval. One hundred pounds of a 20.5 per cent ration con-
tains 46.25 pounds of corn, 13.75 pounds of soybean oilmeal, and 40.00
pounds of other basic ingredients shown in Table 11.2. Point C, Figure
11.11, indicates a least-cost ration of only 16.0 per cent protein for the
third weight interval, with the corn and soybean oilmeal prices assumed.
One hundred pounds of the 16.0 per cent protein ration is composed of
60.0 pounds corn, 18.0 pounds of soybean oilmeal, and 22.0 pounds of
other basic ingredients. The recommended rations resulting from use
of figures 11.9, 11.10, and 11.11 are identical with those of tables 11.12,
11.13, and 11.14 and are included only as a simple alternative method
of presenting the same results.
Throughout the analysis, the criterion for selecting rations has
been one of minimum cost. However, a ration other than the least- cost
ration for prevailing prices may be better suited for producing the most
rapid gains over a given weight range. For example, if the producer
anticipates a fall in turkey prices, he may be interested in getting the
poults to market weight as rapidly as possible, rather than in minimiz-
ing feed cost for a given gain.
OPTIMUM MARKETING WEIGHTS

The preceding sections provided estimates of the least-cost com-
binations of corn and soybean oilmeal over the three weight intervals.
Once the least- cost ration has been determined, the next question is
one of finding the most profitable, or optimum marketing weight for the
turkeys. The marketing weight which maximizes returns above feed
costs is determined by equating the marginal product of feed for the
least- cost ration with the feed to turkey price ratio. (Also, it can be
determined by equating the partial derivatives of gain with respect to
corn and soybean oilmeal with the proper price ratios. The latter is
the more refined method. However, the other method is used here as
a practical measure.) In other words, the most profitable marketing
weight above feed costs is attained under the condition of equation 11.19,
where dY/ dR is the marginal product of the particular ration, showing
the amount added to gain by each small added quantity of the ration, i.e.,
dY/dR is the derivative of gain with respect to feed inputs predicted
CD 05 CM in CO O CO CD 05 CM rtfJ> O
CM in 00 o CO in 00 O
"tf in in in
rH rH rH rH rH
CD CD CD CD* Eh
tH rH rH rH rH
tH IH co 00 00*
rH rH rH rH tH
CO* 05* 05* 05*
tH rH tH rH rH CM
05 O
E— o
"tf c— O CO CD 05 cm in CO rH "tf eh 05 CM in CO rH "tf CO
Ratios*
in in in CD
rH rH rH tH rH
CO CD CD eh eh
tH tH rH rH tH
Eh co CO CO CO
rH rH rH rH rH
05* 05* 05
,H rH rH CM CM CM
doo
05 CM CO 05 CM in 00 CM in co rH rtf 00 tH "tf Eh o
CO CD 05 CM
Price
^tf in in in CD
rH rH rH rH rH
CD CD* tH tH Eh
rH rH tH tH tH
00 C0 CO 05 05
tH tH rH tH tH
05 dddO tH
rH CM CM CM CM CM
O "tf eh rH •tf CO rH "tf C0 rH m CO CM in 00 CM in oo CM in 00

Feed
ID in in CD CD
rH rH tH rH rH
CD tH IH tH CO
rH rH tH rH tH
oo CO* 05* 05 05*
tH rH tH tH rH
odd tH tH rH
CM CM CM CM CM CM
to
rH in 05 CO CD O rtf £— rH ID oo CM CD CO o IH o rtf 00 cm in
IO in in' CD' CD
rH tH tH rH rH
d
rH
tH tH oo co
rH tH rH rH
00 05* 05*
tH tH tH CM CM
o d d rH
CM CM
rH rH CM CM
CM CM CM CM
Turkey
CM CD o
"tf 00 CM CD O
rtf CO CM CD o
CO CO CM in O
"tf 00 CM
in in CD CD CD
tH tH rH VH rH
eh tH 00 00 oo
rH tH rH rH tH
05 05 ooo
rH rH CM CM CM
TH ft CM* CM CM CO
CM CM CM CM CM CM
and
CO t- CM CD O "tf 00 CM CD O in 05 CO 00 CM CO tH in 05 "tf oo
in in' CD CD tH
tH rH rH rH tH
EH CO CO 05
rH rH rH tH tH
05 05 ©d
ft
rH tH CM CM CM
TH CM* CM CM CO* CO
CM CM CM CM CM CM
Rations
'tfCO CO t- rH CO O "tf 05 00 oo CO tH CM IH CM CD rH CD
in in CD CO* Eh
rH rH rH tH rH
EH 00* 00* 00 05
tH rH rH rH tH
05 oo rH rH
rH CM CM CM CM
CM CM CO CO
CM CM CM CM
Protem
•r 05 CO co CM Eh CM CD rH CD rH CD rH CD rH CD rH IH
in in CD CD tH
tH rH tH rH tH
IH 00 00 05 05
rH tH tH rH tH
o ©
CM CM
rH rH CM
CM CM CM
CM CO CO
CM CM CM
CO CO OO CO
•<tf CO CM Eh CM CO CO 00 CO 05 "tf O in
Various
in in CO* CO*
rH rH rH rH tH
E-*00 00 05 05
tH rH rH tH rH
OO >H rH CM
CM CM CM CM CM
CO CO
CM CM
''tf CO CO 00 CO 00 CO 05 "tf 05 in o
cd CM 00 "tf
With
in in CD CO c—
rH rH rH rH
IH CO CO 05 05
rH rH rH rH rH
O tH tH CM CM
CM CM CM CM CM
CO
CM
CO CO CO co CO 05 -tf © ID tH EH CO 05 CO CM CO
Profits
in in CO* CD* EH*
rH rH rH rH rH
eh CO* 05* 05* O*
tH rH rH rH CM
O tH rH CM* CO
CM CM CM CM CM
CO
CM
1
1
!
CM Er- CO CO "tf 05 ID rH IH CO 05 CD CM 05 CD CO
in in CD* CO Eh
rH rH rH rH rH
eh CO 05* 05
rH rH rH rH CM
O o
CM
tH CM CM CO
CM CM CM CM
"tf
CM
f
|
1
Maximum
O in O in rH tH CM 00 rtf tH 00 tH oo CD
"tf CO
in in CD* CD* IH
rH rH rH rH rH
EH 00 OO* 05*
rH tH rH TH CM
O o rH CM CM* CO
CM CM CM CM CM
rtf
CM
1
(
I
for
CD CM Er- CM CO rtf O CO CM 05 CD CM O
tH in CM
"tfin in CD* CD*
rH rH rH tH rH
EH co* oo* 05* 05
rH rH rH rH rH
O tH CM CM* co*
CM CM CM CM CM
"tf
CM
I 1 1 1
Weights
CM E- CO CO rtf ©m cm CO ‘rtf tH 00 CD CO tH 05
"tf"tf in in CD*
rH rH rH rH rH
EH EH* CO* CO* 05
rH rH tH rH rH
o o
CM CM
rH CM CO
CM CM CM
CO
CM
I ! 1 1
>>
Marketing
3 o
as
H ft E'- CD CM E-
en CD in in -tf
CO
"tf
m
o co
"tf
CM 00
CO CM
rH tH
in CM 05 CD "tf
CM CM rH rH rH
tH rH rH rH rH
rH OOOOO
rH CD CD "tf CM
©
rH tH rH tH rH tH
O CD rH rH rH rH rH
-M o
Turkey
8 ft
11.15.
TS
8 o
O ft O CM "tf CD CO o CM CD CO o CM "tf CD CO O CM "tf CD 00 o
Table
® CO CO co’ CD CD EH EH* EH tH EH 00 co co CO* CO 05 05 05 05 05 o
tH
o o
g £
Eh
[
401 ]
from the production function. In equation 11.19, Pr is price per pound of

ration and Py is price per pound of turkey.
For practical purposes, it is supposed that the least-cost ration will

be determined for each of the three weight intervals by the methods of
the previous sections. In the third weight interval, the marginal prod-
ucts for the least-cost ration will be used in determining the optimum
or most profitable marketing weight.
Square root function equation 11.11 is used in predicting optimum
marketing weights for turkeys fed on different rations, under a wide
range of feed to turkey price ratios (the feed to turkey price ratio is
the reciprocal of the turkey to feed price ratio. Table 11.15). The ,
square root function is used for these predictions because, as was men-
tioned previously, it fits the input-output observations for the various
rations more closely than the other functions for the third weight inter-
val. Table 11.15 indicates, for each ration and price ratio, the market-
ing weight which maximizes returns above feed costs. The practical
marketing weight range for female birds is 12 to 18 pounds; for males
the range is about 18 to 30 pounds. Thus, in a mixed or “straight run”
flock for which the predictions of this study apply, the practical market-
ing weight range is from approximately 15 to 24 pounds. Separate pro-
duction functions (computed from observations on all males or all fe-
males) would be required to provide a guide to optimum marketing
weights for the producer feeding a flock of predominately one sex. How-
ever, the figures in Table 11.15 should be relevant for that majority of
producers who feed “straight run” flocks.
Before using Table 11.15, the average least-cost ration for the third
weight interval is determined from Table 11.14 or Figure 11.11 in the
previous sections. Table 11.15 can then be used to predict, for any
particular least-cost ration, the marketing weight which is optimum for
a given price per pound of the ration and of the finished turkeys. Sup-
pose that the least- cost ration for the third weight interval (predicted
from Table 11.14) contains 15.0 per cent protein. If the price of turkeys
is 32 cents per pound, and the price of the 15.0 per cent protein ration is
4 cents per pound (a turkey to feed price ratio of 8.0), the predicted
optimum marketing weight for birds on this ration is 20.5 pounds (Table
11.15). If the turkey to feed price ratio is only 7.0, the optimum mar-
keting weight is reduced to 17.8 pounds (Table 11.15).
Turkey to feed price ratios outside of the 6.0 to 10.0 range shown
in Table 11.15 predict optimum marketing weights which do not fall
within the practical marketing weight range for mixed flocks. However,
the turkey to feed price ratios of recent years have been characterized
by wide fluctuations, resulting in ratios which have frequently been
higher than the upper range of 10.0 shown in Table 11.15. The average
United States ratio was 11.0 in 1949, with a January high of 16.9 in
Iowa. The average turkey to feed price ratio on the Pacific Coast was
6.4 in 1954.
Where risk and uncertainty are important, turkey producers may

wishto market their birds previous to the weight at which marginal
cost equals marginal return; use of these limited resources in the lat-
ter phases of turkey production may provide lower returns than their
use in some other alternative.
The data from Table 11.15 have been used in deriving figures 11.12
through 11.15, which provide convenient graphical approximations of
the marketing weights which maximize returns above feed costs for
various rations and turkey to feed price ratios. Figures are used as
follows: First, find the figure for which the protein level corresponds
most closely with that for the least-cost ration predicted for the third
weight interval. Second, locate the intersection of the turkey and feed
prices on the graph. Third, follow along between the diagonals to the
right side of the graph and read off the approximate optimum marketing
weight. For example, assume a least-cost ration of 13 per cent protein,
an expected turkey price of 36 cents per pound, and a feed price of 5
'
cents per pound. The intersection of the two feed prices is found at
point A, Figure 11.12. Moving to the diagram between
right side of the
the two diagonals which bracket point A, the optimum marketing weight
is found to be 18.1 pounds.
Turkey- Optimum Turkey- Optimum
Feed Marketing Feed Marketing
Price Weights Price Weights
Ratio jlbsO Ratio (lbs.)
20.8
20.1
(9.3
18.5
17.7
16.9
16.1
15.3
Figure 11.14. Optimum marketing Figure 11.15. Optimum marketing

weights for turkeys fed weights for turkeys fed
on a 17 per cent protein on a 19 per cent protein
ration, predicted by ration, predicted by
square root function. square root function.
CHAPTER 12
Earl O. Heady
Milk Production Functions and
Norman L. Jacobson
John A. Schnittker Marginal Rates of Substitution

Solomon Bloom
Between Forage and Grain
nterest in possibilities of forage-grain substitution in the dairy cow

ration has been increased by recent agricultural developments. One
I development is acreage control which allows farmers to grow forage
as a replacement crop for grain. Another is the continuing interest in
conservation; erosion control plans ordinarily require an increased
acreage of grasses and legumes and fewer acres of grains and row
crops. Both of these developments increase the supply of forages rela-
tive to grains and give rise to questions of using forage profitably. One
possibility is the substitution of forage for grain in terms of ruminants.
The feasibility of this adjustment depends, however, on the rate at which
the various classes of feeds substitute for each other.
Changes in price structures, with dairy product prices depressed
relative to feed and labor costs, also have caused farmers to examine
substitution possibilities as a means of lowering costs and increasing
profits. Then, too, yearly and geographic differentials in the costs of
concentrates relative to forages and to the prices of milk give rise to
questions of the most profitable ration under particular economic cir-
cumstances. New technologies in dairy farm management and the grow-
ing concentration of production on specialized farms also add interest
in this direction. To what extent should the grain-forage ration be var-
ied as the price ratio of grain to forage changes? To what extent
should the most profitable ration differ between grain surplus and grain
deficit areas or other areas where concentrates are priced at different
levels? These questions can be answered only if information is avail-
able on substitution ratios. The optimum ration, in terms of profit
maximization, can be determined only by relating substitution ratios to
price ratios. Finally, determining the nature of the milk production
surface with its expression of feed to milk transformation ratios and
feed substitution coefficients is a central problem in dairy cow nutrition.
OBJECTIVES
The experiment on which this report is based was designed to pro-

vide estimates of the milk production function and feed substitution ra-
tios in alternative dairy cow rations. The experiment provides
404
MILK PRODUCTION FUNCTIONS 405 *
predictions of the milk production surface and milk isoquants indicating

marginal rates of substitution between the two classes of feeds con- —
centrates and hay. The experiment reported was relatively small in
sample size, but large in terms of facilities. It was methodological in
nature but is being continued on a somewhat larger scale.
The primary objectives are to establish (1) the rates at which grains
and forages substitute under specific technical conditions and (2) the
rate at which feeds are transformed into milk for various production
levels and rations. An auxiliary objective is to investigate the economic
potential of substituting forage for grain. Hence, details are provided
for (a) explaining the models which serve as a basis for the experimen-
tal design, (b) illustrating the procedure used in predicting feed substi-
tution and transformation rates, and (c) determining the particular ration
and level of grain feeding which results in the least- cost ration and the
most profitable level of production per cow.
The study reported was restricted in magnitude because of limita-
tions in funds, cows, barri space, and other facilities. Because of its
limited magnitude, it should be looked upon as an exploratory study, to
be supplemented by later investigations now underway. The over-all
objectives of this study are of a methodological nature. The central
predictions revolve more nearly around estimation of the milk produc-
tion function and feed transformation and substitution coefficients than
around use of the particular principles in determining economic optima
in dairy rations. It is hoped that this fundamental study will provide
the basis for further work and encourage other studies which allow more
refined predictions of the milk production function and of the economic
gains in using the profit- maximizing principles outlined.
BASIC NATURE OF FUNCTIONS
Milk production is a complex process involving many resources, of

which feeds represent but one class. The milk production function is
of the general form:
(12.1) M = f (C, F, Xx X 2 X
, , 3 ,
X4 X n ).
Here, M refersto milk production per cow in a specified time period,
C refers toconcentrate intake, F refers to forage intake, X x refers to
body size, X 2 refers to inherent breed qualities of the cow, X 3 refers to
labor used, and X 4 through X n refer to unspecified resources or inputs.
While all of these resource or input categories are variable, most nu-
trition studies (and this investigation specifically) are carried on in the
framework of a production relationship such as that represented by
equation 12.2.
(12.2) M = f (C, F |
X X2 X X 4
x , , 3 ,
Xn )
406 MILK PRODUCTION FUNCTIONS
Here, only the resources or inputs to the left of the vertical bar are
considered variable. Labor required to handle cows under different
rations is necessarily varied in an experiment. For experimental pur-
poses, however, labor is assumed to be available in unlimited quantities
at no cost. While labor as a variable must be considered in terms of its
cost or price in profit decisions on the farm (e.g., more labor may be
required to feed a specific grain to hay ratio than to hand- feed grain and
self-feed hay), this step is unnecessary in technical experiments in-
volving only estimation of feed substitution rates or feed- milk trans-
formation ratios.
POSSIBLE GEOMETRIC FORM

A milk production function involving two variable categories of feed
can be represented as a three-dimensional diagram or surface. Milk
output per cow in the relevant production period is measured on the ver-
tical axis while each category of feed is measured on the respective
horizontal axis. Each point in the feed plane represents a different ra-
tion and level of feeding and will correspond to a particular level of
milk represented on the milk surface. The particular nature of this
surface, including the slope of the isoclines over it and the slope of the
contours around it, will determine (1) the forage to grain ration which
gives the lowest cost for any stated level of milk output, (2) the level of
feeding which will result in maximum profit per cow over feed costs,
and (3) the extent to which conventional ENE or TDN evaluations of the
energy or heat transformation of feed are appropriate for evaluations
concerned with milk transformation of these same feeds.
Unfortunately, little is known about the nature of the milk production
surface. Numerous feeding standards suppose that the milk production
function is homogeneous of degree 1.0; the surface is implicitly as-
sumed to be linear up to the limits of the cow’s milk producing capac-
ity. Some standards, such as the total digestible nutrient (TDN) basis
for rations consider the milk isoquants and input- output curves to be
straight lines since they are not varied to consider the proportion and
level of feeding or milk production. Other nutrition recommendations
suppose the milk surface to have nonlinear isoquants and isoclines.
This is because ration recommendations seldom include only one class
of feed (i.e., grain or forage), a recommendation which would be the
most profitable one if milk isoquants were linear. Some possible hy-
potheses about the nature of the milk production surface were outlined
in Chapter 3.
PREVIOUS RESEARCH RELATING TO NATURE

OF ISOQUANTS AND MILK SURFACE
While no previous experiment has been designed to predict the

MILK PRODUCTION FUNCTIONS 407
nature of the milk production surface, the results of other studies do

lead to hypotheses about substitution and transformation coefficients.
Huffman and Duncan discussed the possible stimulating effect of a small
amount of grain when a cow has been fed forage alone. 1 They stated
that a cow receiving forage alone will not produce as much milk (FCM)
as when a small amount of grain is substituted for an equal amount of
TDN from hay. These results lead to the supposition that the milk iso-
quant may curve rather sharply at the forage end and is, therefore, non-
linear.
Jensen et al predicted grain input-output curves which are non-
.
2
linear. Although the report provides no postulates about milk isoquants,
it appears that curved isoquants should be associated with curved input-
output curves for any mathematical function used to define the optimum
or maximum milk production per cow. Beach indicated that as more
grain is fed to a cow, the maintenance requirements (i.e., at the zero
milk isoquant) in terms of TDN become less and less. 3 As the ration
approaches an all- grain ration, TDN from grain eventually become
less efficient in maintaining health and activities of the cow. While the
Huff man- Duncan experiment suggested that input-output curves may be
nonlinear and that the milk isoquants are curved on the hay extremity,
the work of Beach suggested curved isoquants towards the grain extrem-
ity. Of course, the isoquants might have greater curvature at the ends
but be nearly linear in the middle.
In a study based on a sample of dairy farms, Ashe concluded that
the input-output curves follow a near linear relationship up to about
4,000 pounds of grain per cow; between 4,000 and 6,000 pounds of grain,
milk increases only slightly; and over 6,000 pounds of grain, milk does
4
not increase at all. Yates and others summarized several research re-
ports implying a diminishing rate of transformation of feed into milk.
These studies included both European and American data and an “eco-
nomic curve” derived from Danish experiments indicating considerable
5
decline in marginal milk yields at high feeding levels.
Recommended daily energy allowances for milk production were
summarized by the Committee on Animal Nutrition of the National Re-
6
search Council. Because of lack of appropriate input- output data, its
recommendations did not consider the concept of diminishing returns
to dairy cow rations. The committee’s approximate guide gave .32
pound TDN as the requirement for each additional pound of 4 per cent
1
Huffman, C. F. and Duncan, C. W. The nutritive value of alfalfa hay. Jour. Dairy Sci.,
32:465-71. 1949.
2
Jensen, E. et al* Input-output relationships in milk production. Tech. Bui. 815. USDA,
Washington, D.C. 1942.
3
Beach, C. L. The facility of digestion of foods as a factor in feeding. Conn. Agr. Exp.
Sta. Bui. 43. New Haven. 1906.
4
Ashe, A. J. Response of milk production to increased grain feeding. Farm Ecoru, No.
174. Cornell Univ. 1950. Pp. 4474-76.
5
Yates, F., Boyd, D., and Pettit, G. Influence of changes in level of feeding on milk pro-
duction. Jour. Agr. Sci., 32: 428-56. 1942.
6 Wash-
National Research Council. Recommended nutrient allowances for dairy cattle.
ington, D.C. 1950.
fat- corrected milk (FCM) above maintenance, but it did not indicate any
changes TDN transformation due to the combination of feeds and/or
in the
to the level of feeding. The Morrison recommendations, used exten-
sively in this country, have the same over- all recommendations for
TDN, whereas recommendations based upon the net-energy system (NE)
estimate .30 therms for each additional pound of 4 per cent FCM above
maintenance requirements. Blaxter, in a review of the energy stand-
ards for dairy cattle, stated that energy standards should express the
7
productivities of feeds as they are. However, he did not indicate that
feeding values should vary with the proportions of feed fed.
The dependence of the ruminant upon the catabolic and anabolic ac-
tivities of the rumen microflora implies that certain combinations of
hay and grain may stimulate or depress microbiological activity and,
hence, exert an effect on the nature or curvature of the isoquants. This
problem was pointed out in studies by Hamilton and Swift, et aL, who
reported a depression of ration digestibility at a hay- to- grain ratio of
8
1:1. The concepts of the “stomach capacity” and “physiological limit”
lines mentioned in Chapter 3 have further ramifications in dictating the
size and feed capacity of the cow which can be used most economically
under varying economic conditions. The problem of genetic ability or
“dairy merit,” which is defined by the percentage of consumed energy
that is converted into milk energy, and its basic interrelationships with
the plane of nutrition involved in the lactational level desired, must also
be considered in determining the milk production surface.
DESIGN OF EXPERIMENT
The basic experiment for the predictions of this study included 36

cows and is explained in detail below. This experiment, conducted with-
out the use of pasture, extended over a 17- month period in 1953 and
1954. For certain predictions and estimates, data from an earlier ex-
periment including 15 cows were also used. Both experiments are ex-
plained below. The experimental design and over-all study represent
an interdisciplinary approach by dairy nutritionists and production
economists, with the aid of specialists in statistics. Individuals involved
held exploratory seminars to discuss the logical models, based on pre-
vious knowledge in nutrition and economics. The design of the experi-
ment and the prediction methods were then selected to conform to these
models. It should be pointed out, however, that the researchers in-
volved do not hold that the design used is optimum. The experiment is
a relatively small one, but one feasible with the limited resources and
facilities available.
7
Blaster, K. L. Energy feeding standards for dairy cattle. Nutr. Abst. and Rev., 20: 1-
18. 1950.
8
S. The effect of added glucose upon the digestibility of protein and of fi-
Hamilton, T.
ber m rations for
sheep. Jour. Nutr., 23: 101-10. 1942; and Swift, R. W., Thacker, E. J.,
Black, A., Bratzler, J. W., and Jones, W. H. Digestibility of rations for ruminants as af-
fected by proportion of nutrients. Jour. Anim. Sci., 1: 447-61. 1918.
Basic 1953-54 Experiment With 36 Cows
The Holstein herd at the Iowa State University Dairy Farm was the
source of the 36 cows used from March, 1953, to September, 1954.
From the date of calving, each cow remained under experimental con-
ditions for an initial 14- day adjustment period. A fixed ratio of 7 pounds
of hay to 4 pounds of concentrates, initiated during the adjustment pe-
riod, was maintained throughout the preliminary period. 9 The prelim-
inary period provided the basis for dividing the animals into high-,
medium-, and low- producing ability groups for their subsequent random
allotment to the experimental period. In general, the production ability
ranges for the animals in terms of pounds of 4 per cent FCM were as
follows: (a) “high” — 10,500 pounds and over, (b) “medium* — 9,000 to
10,500 pounds, and (c) “low* — 9,000 pounds or less.
The design for the experimental period of 182 days is illustrated in
Table 12.1. Four of the hay- to- concentrate ratios were chosen, ranging
from a ration in which 75 per cent of the energy (ENE) was derived from
hay and 25 per cent from concentrates, to one in which 15 per cent of
the energy was derived from hay and 85 per cent from concentrates.
The four hay- to- concentrate ratios were 75:25, 55:45, 35:65, and 15:85.
Each of these hay- to- concentrate ratios was fed at high, medium, and
low levels. For each of the 12 “hay- to- concentrate ratio- feeding level”
treatments, there were three cows — one of high-, one of medium-, and
one of low- producing ability. At the start of the experimental period,
each cow, after its assignment to an ability group, was randomly allotted
to one of the 12 “hay-to-concentrate ratio-feeding level” positions.
Auxiliary Experiment
While the main analysis of this study rests on the basic experiment
Table 12.1. Nature of Allocation of Animals by Level of Feeding
and Ability (Milk Production in Pounds)
Hay-to-Concentrate Ratio
Level of
Ability
Feeding H-75:C-25 H- 55:0 45 H-35:0 65 ‘
H-15:085
High 2,553 3,157 2,710 2,982

High Medium 2,392 2,649 3,529 2,976
Low 3,632 3,263 3,160 3,538
High 3,266 2,600 2,378 3,142

Medium Medium 3,469 3,444 2,643 3,493
Low 3,440 3,597 3,432 3,516
High 3,272 3,291 2,963 3,174

Low Medium 3,450 3,483 3,128 2,606
Low 3,302 3,294 3,439 2,159
9
For a detailed explanation of the experiment, see: Bloom, Solomon. Effects of various
dietary hay-concentrate ratios on nutrient utilization and production responses of dairy
cows. Unpublished Ph.D. thesis. Iowa State University Library. Ames. 1955.
explained above, cows from the following experiment were used for pre-
dictions explained later. The auxiliary experiment is that reported by
10
Martin et at. Fifteen cows were used from this experiment to evalu-
ate four levels of alfalfa hay feeding. Rates of hay feeding during the
experimental period were at .5, 1.17, 1.83, and 2.5 pounds per day per
100 pounds of body weight. A 2- week preliminary period was used in
which the cows were fed hay at a daily rate of 2 pounds per 100 pounds
of body weight and grain to supply 100 per cent of Morrison’s recom-
mended levels. Cows were on the auxiliary experiment for 16 weeks.
BODY WEIGHT CHANGES

Table 12.2 summarizes the body weight changes from two aspects:
the total changes over the entire 182- day experimental period and the
changes from the end of the initial 4 weeks of the experimental period
to its conclusion. The results of the weight changes over the two pe-
riods are presented in the analyses of variance in tables 12.3 and 12.4.
When the entire experimental period was considered (Table 12.3), the
effects of ration, feeding level, and ability upon body weight changes
either approached significance or were significant at the 5 per cent
level. However, when the last 22 weeks of the experimental period
were considered these effects disappeared (Table 12.4). From the first
to the second case, the components of variance for ration, feeding level,
and ability decreased markedly. These results indicate that the body
weight changes, after the first 4 to 5 weeks of the experimental period,
were largely independent of the ration fed, the level of feeding, and the
ability of the animals.
Table 12.2. Body Weight Changes During Two Time Intervals of the Experimental Period
H-75:C-25* H-55:C-45* H-35:C-65* |

H-15:C- 85*
Level of Change (lbs.) Change (lbs.) Change (lbs.) Change (Zbs.)
Ability
Feeding Cow 0-26 4-26 Cow 0-26 4-26 Cow 0-26 4-26 Cow 0-26 4-26
wks. wks. wks. wks. wks. wks. wks. wks.
High 2,553 99 200 3,157 -69 -48 2,710 42 79 2,982 -80 -21
High Med. 2,392 59 125 2,649 -33 -46 3,529 92 179 2,976 -82 -38
Low 3,632 -45 -47 3,263 -24 -7 3,160 110 69 3,538 51 20
High 3,266 -22 7 2,600 -78 -31 2,378 14 64 3,142 -102 3

Med. Med. 3,469 40 57 3,444 36 26 2,643 46 22 3,493 -56 10
Low 3,440 26 39 3,597 -25 -32 3,432 55 29 3,516 58 120
High 3,272 -116 53 3,291 2 40 2,963 -153 -21 3,174 -178 -31
Low Med. -67 -24 3,483 -61 -5 3,128 23 89 2,606 -164 -91
Low 3,302 92 55 3,294 -2 19 3,439 -70 -4 2,159 21 36
*Hay-to-concentrate ratio
10
Martin, T. G., Stoddard, C. E., and Allen, R. S. Effects of varied rates of hay feeding
on body weight and production of lactating dairy cows. Jour. Dairy Sci., 37: 1233-40. 1954.
1
Table 12.3. Analysis of Variance: Body Weight Changes During

the Entire Experimental Period
Source of Sum of Mean Component of

Variation D.F. Squares Squares F Variance (Per Cent)
Ration 3 33,167 11,056 3.30 9
Level 2 30,207 15,104 4.51* 15
Ability 2 32,906 16,453 4.91* 15
R x L 6 22,949 3,825 1.14 2
R x A 6 30,826 5,138 1.53 9
A xL 4 11,792 2,948 .88 0
R x L x A 12 40,200 3,350 51
Total 35 202,047
*P <.05
t Approaches P < .05
MILK PRODUCTION FUNCTIONS FOR 36 COWS IN 26-WEEK

BASIC EXPERIMENT IN THE OVER-ALL PERIOD
Two sets of functions have been derived from the basic 36-cow ex-
perimental data: (1) those where a single time period is included and
(2) those where time is considered as a variable. While predictions for
a single and constant time period are included as methodological ma-
terials, it is believed that the production functions which include time
Table 12.4. Analysis of Variance: Body Weight Changes From the

End of the Initial Four Weeks of the Experimental
Period to Its Conclusion
Source of Sum of Mean Component of

Variation D.F. Squares Squares F Variance (Per Cent)
Ration 3 26,268 8,756 2.54 6
Level 2 8,951 4,476 1.30 0
Ability 2 705 353 .10 0
R x L 6 32,062 5,344 1.55 14
R xA 6 25,333 4,222 1.23 5
A x L 4 10,441 2,610 .76 0
Rx LxA 12 41,301 3,442 75
Total 35 145,061
as a variable provide the most efficient estimates. The time variable

allowed for some changes in body weight, as well as the normal trend
in milk output over the lactation period. As the data in Table 12.2 in-
dicate, cows attained a near- equilibrium in weight after a month of the
experiment.
Initial Equations for 36-Cow Basic Experiment

Over 26 Weeks (Time Not a Variable)
Since little is known about the milk production function, three initial
types of algebraic equations which do not include time as a variable
were fitted to the data. A power equation was selected as one of a gen-
eral type (including those such as exponential, Mitcherlich, etc.) which
does not require specification of a single maxima in milk production.
It is known that such a function may not conform adequately to a milk
production surface since it assumes constant elasticity of production,

linear isoclines, and increasingly wide ranges of rations for higher
levels of milk production. However, it was thought that the function
might allow reasonable “average” estimates of substitution ratios and
transformation coefficients in the midsection of the milk surface. The
other two equations allow specification of one ration consistent with
maximum milk production per cow and of isoclines which converge to
this point. These two functions were a quadratic and a square root
quadratic.
Variables and Regression Equations
Three variables were used in estimating the initial

milk production
functions. Two of these are the concentrate
and hay feeds discussed
earlier. The third is cow ability since this variable was found to be
highly associated with milk production in the experimental period.
Hence, the variables for the functions which follow are those explained
below where the milk output and feed input measurements are aggre-
gates extending over the 26- week experimental period of the basic ex-
periment:
M is production in pounds per cow of 4 per cent fat-corrected milk in the 26-
week experimental period.
H is pounds of alfalfa hay measured in pounds per cow over the 26 -week experi-
mental period.
G is concentrate mix, called grain hereafter, measured in pounds per cow over
the 26-week experimental period.
A is cow ability measured in pounds of 4 per cent fat-corrected milk produced in
the 50-day preliminary period when all cows were fed the same ration as detailed
earlier. 11
11
The simple correlation coefficient (r) between the ability term used and milk produc-
milk production functions 413
The 26» week initial functions for the 36- cow basic experiment are
shown in equations 12.3, 12.4, and 12.5 for the logarithmic, quadratic,
and square root functions, respectively.
1213
(12.3) M =* 15.749 H* G*
2758
g 3659
(12.4) M = 3,787.56 - .1288 H + .9842 G - 1.0991 A + .Q00042H 2 -
2
.000064 G + .000353 A2 + .000000032 HGA
(12.5) M = 19,356.40 + 1.7855 H + 1.1258 G + 3.2040 A - 300.0230 Ve
- 183.0605 VgT- 226.5986 VA+ 2.6626 YeG
The relevant statistics for these three functions are given in Table
12.5. The t values are for the regression coefficients in their respec-
tive order within the equations above. From 73 to 78.6 per cent of the
total variance in milk production is accounted for in the three variables,
depending on the function. All of the regression coefficients in the power
function are acceptable at the probability level of 1 per cent. However,
none of the individual coefficients for the quadratic or square root func-
tions are significant at this probability level, even though a larger por-
tion of the total variance in milk production is accounted for by the
variables of the latter two functions. In a pure probability sense, equa-
tion 12.3 might be accepted for prediction purposes while equations 12.4
and 12.5 might be dropped. However, even though their regression co-
efficients have relatively larger standard errors, the general forms of
equations 12.4 and 12.5 might provide predictions which conform better
to the logic of the milk production surface than does equation 12.3, par-
ticularly since the latter function (a) does not cause milk level to ap-
proach a maximum, (b) causes the milk surface to pass up a "ridge” of
wide ration latitudes, rather than to narrow to a peak of a single ration
for the maximum milk production per cow, (c) does not allow low-level
isoquants to be attained with hay or grain alone, and (d) causes linear
isoclines which "fan out,” rather than converge at the maximum milk
Table 12.5. Correlation Coefficients and t Values for

Equations 12.3, 12.4, and 12.5
Values of t in Order of b’s in Equation
Equation R b. b2 b3 b4 h5 be b.
12.3 .8545* 2.84* 5.65* 5.13* — — —

12.4 .8832* .18 .94 .60 .74 .54 .99 .64
12.5 .8864* 1.21 .45 .91 .84 .39 .66 .73
*P< .01.
tion is .64. The comparable statistic using age-corrected fat in previous lactations as the
ability term yielded a correlation coefficient of .33.
production level. As is indicated in Table 12.6, substitution ratios be-

come extremely large or small at extreme rations for the logarithmic
function. However, equation 12.3 may provide useful estimates of aver-
age marginal substitution rates over a narrow range of the surface sur-
rounding the mean level of feeding and the mean ratio of feeds in the ex-
periment.
Marginal Equations for Logarithmic Function
Equations 12.6 and 12.7, which are derived from equation 12.3, de-
fine the marginal or incremental productivity of hay and grain, respec-
tively, when one is variable and one is fixed in quantity. Both indicate
diminishing productivity since the exponent on both the H and G vari-
ables is less than 1.0 in equation 12.3. Also, since the sum of the ex-
ponents is less than 1.0, the function indicates a diminishing feed to
milk transformation as any fixed ratio of hay and grain is increased in
quantity. Equation 12.8, also derived from 12.3, provides the isoquants
paralleling those outlined in the earlier section on logic. Equation 12.9
defines the slopes of the isoquants and, therefore, indicates the mar-
ginal rate of substitution of grain for hay if equation 12.3 is used for
predictions. Obviously, the substitution ratio will be predicted to
Table 12.6. Feed Combinations and Marginal Rates of Substitution (5H/6G) Predicted
From Logarithmic Equation 12.3 for Milk Isoquants of 5,300, 6,300,
and 7,300 Pounds per Cow Over the 26 -Week Period
(Ability Fixed at Mean for the 36 Cows)
5,300 Pounds Milk* 6,300 Pounds Milk^ 7,300 Pounds Milk*

«_
Lbs. Lbs. Oj Lbs. Lbs x 5h1I Lbs. Lbs. 6hII
grain § hay§
Oil
o grain § hayS 6G grain § hay§ SG
1,000 5,815 13.22 --

1,500 2,313 3.51 1,500 9,308# 14.11 — ..
2,000 1,204 1.37 2,000 5,001 5.68 2,000 16,843j 19.15
2,500 724# .66 2,500 3,008 2.74 2,500 10,152? 9.23
3,000 478# .36 3,000 1,991 1.51 3,000 6,709* 5.08
3,500 1,401 .91 3,500 4,714 3.06
-- #
4,000 1,034 .59 4,000 3,489 1.98
... 4,500 792 # .40 4,500 2,661 1.34
-- — — -- — — 5,000 2,098 .95
* Columns 1 and 2 show feed combinations which will produce 5,300 pounds of milk;
column 3 shows marginal rates of substitution of grain for hay for these combina-
tions.
tcolumns 4 and 5 show feed combinations which will produce 6,300 pounds of milk;
column 6 shows marginal rates of substitution for these combinations,
i Columns 7 and 8 show feed combinations which will produce 7,300 pounds of milk;
column 9 shows marginal rates of substitution for these combinations.
§ Derived from equation 12.8.
II Derived from equation 12.9.
# Outside of range of observations in experiment.

change as hay to grain ratios change, if function 12.3 is used for esti-
mation.
(12.6) = 1.906 H~’ 8787 G' 2758 A' 8659

OH
‘ 1213 7242 3659

(12.7)
OL
= 4.347 H G~‘ A'
(12.8) 1
M \
8,2440
\ 15.75 G ,275b A* 3658

/
(12.9)
H (-2-2TO-H-
Isoquants for three levels of milk production predicted from the

logarithmic function (i.e., isoquant equation 12.8) are shown in Figure
12.1. The straight lines denoted as a, b, and c are isoclines indicating
all feed combinations which give a specified rate of substitution between
hay and grain, if predictions are based on equation 12.3. Line a shows,
for example, all quantities of grain and hay where 1 pound of grain sub-
stitutes for 3 pounds of hay. The slope of the isoquants does not change
greatly above isocline a with a substitution ratio of 3. The isoquants
Figure 12.1. Milk isoquants for a 26 -week period

predicted from equation 12.3 with ability
fixed at mean for 36 cows (Dots show feed
inputs yielding indicated milk quantity
from 12 cows at medium level of ability.)
bend rather sharply below line c. These same phenomena are illus-
trated in Table 12.6, the tabular counterpart of Figure 12.1, since the
substitution ratios become very large for rations with a small propor-
tion of grain and very small with a large proportion of grain. Isoquants
with sharply changing slopes (rapid changes in substitution rates) at the
extremes, such as found in the logarithmic equation, may actually be
consistent with physiological processes of milk production. However,
isoclines which fan out such as those in Figure 12.1 are inconsistent
with a maximum milk output per cow, a condition possible only with
converging isoclines.
If power functions were accepted as the best predicting equations,
the
the isoclinesshown would indicate the least- cost ration for particular
price ratios. For example, isocline b shows the points on the milk iso-
quants where 1 pound of grain substitutes for 2.5 pounds of hay. Hence,
ifthe price of grain divided by the price of milk is 2.5, the points of
intersection of the isoclines and isoquants show the least- cost rations
for milk production levels of 5,300, 6,300, and 7,300 pounds per cow in
the 26- week period. The dots in Figure 12.1 represent the 12 cows at
themedium level of ability for 4 ration combinations and 3 feed levels.
These dots provide a visual indication of whether the predicted isoquants
seem consistent with the data. The average milk production for these 12
cows was 6,463 pounds. (All 36 cows served as the basis for predicting
Figure 12.2. Milk isoquants for a 26-week period

inputs yielding indicated milk quantity from
12 cows at medium level of ability.)
the isoquants in Figure 12.1.) The marginal rate of substitution of grain

for hay at approximately the midpoint (2,800 pounds of hay and 2,600
pounds of grain) of the 6,300-pound milk isoquant is 2.45. This figure
indicates that, at the particular feed combination, one more pound of
grain would replace 2.45 pounds of hay, with milk output held constant
at 6,300pounds. Conversely, 1 pound of hay would substitute for .40
pound of grain if the particular equation were used for the predictions.
Marginal Equations for Quadratic and Square Root Estimates
The marginal milk product functions for grain and hay, as single
variables, are indicated as equations 12.10 and 12.11, respectively, for
quadratic equation 12.4; they are indicated as equations 12.14 and 12.15
respectively, for the square root function of equation 12.5.
5M
(12.10)
6H
-.1288 + .000085 H + .000000032 GA
6M .9842 - .000129 G + .000000032 HA
(12.11)
6H
(12.12) H 1,524.0 -.9556 G
+ 11,838.0 V-.5309 -.000187 G + .000000017 G 2

+ .000169 M
6H .9842 - .000129 G + .000000032 HA
(12.13)
6G -.1288 + .000085 H + .000000032 GA
6M
(12.14)
5H
1.7855 -150.0115 H"*
5
+ 1.3313 G 5
H"-
5
6M -
(12.15)
5G
1.1258 -91.5302 G '5
+ 1.3313 H -5 G _,s
(12.16) H = [84.01 -.7456 VG

2
'
+ .2800 V 7.1421 M- 290.2395 VG -.9513 G -24,469.10j
5 ,s ,s
6H 1.1258 -91,5302 G~' + 1.3313 H G~
(12.17) -' 5 5 5
6G 1.7855 -150.0115 H + 1.3313 G‘ H“-
The milk isoquant functions for the quadratic and square root func-
tions are indicated, respectively, as equations 12.12 and 12.16; the mar-
ginal rate of substitution equations are 12.13 and 12.17.
Equations 12.4 and 12.5 appear unsatisfactory for prediction pur-
poses since the signs of the H terms in equations 12.10 and 12.14 are
such that the marginal product of hay increases with higher levels of
feeding. In other words,, each pound of hay would add more to milk pro-
duction than the previous pound over all possible feeding levels. If equa-
tion 12.4 is used as the basis for predicting a milk isoquant of 6,300
pounds for the 26-week experimental period, the figures in Table 12.7
result. Similar results are forthcoming for predictions based on equa-
tion 12.5. In Table 12.7, the marginal rates of substitution of grain for
hay change slowly between the extremes of the ration and do not differ
significantly from the substitution rate of 2.3 from the linear equations
presented later.
Tab.le 12.7. Milk Isoquant of 6,300 Pounds and Marginal

Rates of Substitution From Quadratic
Equation 12.4: Ability Fixed
at Mean for 36 Cows
Feed to Produce 6,300 Pounds Milk Marginal Rate of
Lbs. grain Lbs. hay Substitution: ~lr

0G
1,000 7,711 2.45

1,500 6,490 2.43
2,000 5,279 2.41
2,500 4,083 2.37
3,000 2,908 2.32
3,500 1,769 2.23
4,000 688 2.09
Other Functions for 36 Cows
Other quadratic or square root equations similar to equations 12.4

and 12.5 appear logical in estimating a milk production surface, since
(a) the t values are low for equations 12.4 and 12.5 and (b) the marginal
product equation for hay is increasing. The t values are low in equa-
tions 12.4 and 12.5 partly because the size of the sample is small rela-
tive to the variance of milk production and because a relatively large
proportion of the degrees of freedom are exhausted in the many coeffi-
cients estimated for the equations. A larger experiment might qualify
a quadratic or square root function in estimating the surface. In a sim-
ple attempt to increase the number of degrees of freedom for the small
sample, equation 12.18 was estimated with the variables outlined ear-
2
lier. In this case. A was dropped so that ability enters into the func-
tion in linear form only. Only three regression coefficients (in contrast
to the seven coefficients of equation 12.4) were estimated, including one
2 2
for the term H - .00001 H and one for the term G -.00007 G The co- .
2 2
efficients before the H and G were simply estimated from previous
nutrition studies and the data of this study.
(12.18) M = .4304 (H - .00001 H2) + 1.5008(G - .00007 G2 ) + .9265 A -665.49

The regression coefficients for this “adjusted quadratic” equation

can be accepted in a probability sense at the 1 per cent level even if
two added degrees of freedom are dropped to compensate for direct es-
2 2
timation of the constants for H and G (see Table 12.8).
Equations 12.19 and 12.20 have been derived from equation 12.18
and are, respectively, the milk isoquant and substitution rate equations.
The milk isoquants in Figure 12.2 are based on equation 12.19; equation
12.20 indicates the slopes of the isoquants at particular points in the
feed plane.
(12.19) H = 50,000.0
2
± (-116,165.61) K.2136 + .0000258 G-.0000000018G -.0000172 M
/<* am 6H _ 1.5008 - .000210 G

{iz.M) -
gG 43Q4 _ #0000086 H
If used to estimate milk isoquants, equation 12.18 results in con-
tours which have some curvature, although the change in slope is not
great. The slope of the isoquants, and their location in the feed plane,
is almost identical with the isoquants presented later in Figures 12.3
and 12.4 for equations 12.22 and 12.23. Table 12.9 includes data show-
ing feed combinations, or the milk isoquants derived from equation 12.19,
and the marginal rates of feed substitution for three milk isoquants de-
rived from equation 12.20. The substitution rates do not change as rap-
idly as do those in Table 12.6 for the power function.
However, while the isoquants in Figure 12.2 appear to have but little
slope and substitution rates in Table 12.9 do not change as rapidly as
those in Table 12.6, a considerable change in substitution rates does
occur over the range of observations presented. These data, all derived
from equation 12.18, have a logical advantage over the estimates for
equations 12.22 and 12.23, since the former allows diminishing returns
for hay as well as grain. However, as differences between equations
12.21 and 12.22 show, there is no probability basis for retaining a hay
Table 12.8. Correlation Coefficients and t Values for Equations 12.18, 12.21,
12.22, 12.23, and 12.26
Value of t in Order of b's in Equation

Equation
R bx b, b, b4 b5
12.18 .8755* 3.98* 6.16* 5.28* — —

12.21 .8770* .99 3.57* 4.75* 1.42 .38
12.22 .8764* 3.85* 3.89* 4.91* 1.46 —

12.23 .8782* 3.76* .15 4.94* 1.61 —
12.26 .8672* 3.85* 5.82* 5.27* — —
*p<.01, based on usual number of degrees of freedom. However, fewer degrees of
freedom are probably appropriate for equation 12.18.
Table 12.9. Feed Combinations in Producing Specified Milk Levels and Marginal
Rates of Substitution of Grain for Hay Predicted From Modified
Quadratic Equation 12.18; Ability Fixed at Mean of 36 Cows
for 26-Week Basic Experiment
Feed to Produce 5,300 Feed to Produce 6,300 Feed to Produce 7,300

Pounds Milk Pounds Milk Pounds Milk
MRS of MRS of MRS of
Lbs. Lbs. grain for Lbs. Lbs. grain for Lbs. Lbs. grain for
grain hay hay grain hay hay grain hay hay
6H/6G 6H/6G 6H/6G
1,000 5,560 3.37 1,000 8,256* 3.59 1,000 11,138* 3.86
1,500 3,969 2.99 1,500 6,566 3,17 1,500 9,328* 3.39
2,000 2,558 2.65 2,000 5,074 2.79 2,000 7,738* 2.97
2,500 1,313 2.33 2,500 3,761 2.45 2,500 6,345 2.60
3,000 221* 2.03 3,000 2,612 2.13 3,000 5,130 2.25
3,500 1,616 1.84 3,500 4,080 1.94
4,000 763* 1.56 4,000 3,182 1.64
4,500 2,428 1.36
5,000 1,813 1.09
*Prediction is outside range of observations.
coefficient which results in diminishing returns for this feed category

when the 36 cows are used for an aggregate 26- week lactation period.
Equations 12.21, 12.22, and 12.23 represent estimates of the milk
production function when particular terms are dropped from the initial
quadratic equation 12.4 and the initial square root equation 12.5. In
2
equation 12.21, the A and interaction terms have been dropped from
2
12.4; in 12.22, H also is dropped. In equation 12.23, VW, j/A"and the
interaction term have been dropped from equation 12.5.
(12.21) M = .3010 H + 1.5171 G + .8978 A -.000106 G2 + .000014 H2 -459.63
(12.22) M = .4089 H + 1.4423 G + .9074 A -.000091 G2 -573.42
(12.23) M = .3977 H -.1028 G+ .8988 A + 102.9030 j/cT-2, 335.45
The R and t values for these equations were given in Table 12.8. In
equation 12.21, the regression coefficient for H 2 again is nonsignificant
and positive, suggesting the unrealistic condition of increasing marginal
productivity of hay. In 12.22, the regression coefficient for G 2 is sig-
nificant at a probability level of less than .20, and the coefficients for
the other terms are significant at a probability level of less than .01.
An isoquant map for quadratic equation 12.22 is included in Figure 12.3,
while one for square root equation 12.23 is included in Figure 12.4.
These two sets of isoquant maps are almost identical: milk contours for
both equations have only slight curvature. While the location of the
milk contours in the feed plane is similar for figures 12.2, 12.3, and
12.4, those in Figure 12.2 have somewhat greater curvature at the lower

inputs yielding indicated milk quantity from
12 cows at medium level of ability.)
end. This difference is due to the fact that the marginal rate of substi-
tution of grain for hay is lower as the proportion of grain increases for
isoquants based on equation 12.18 than for those based on equations
12.22 and 12.23. The difference is apparent in comparisons of Table
12.9, which includes substitution rates based on
equation 12.18, and Ta-
ble 12.10, which includes substitution rates based on equations 12.22
Table 12.10. Feed Combinations Producing Specified Milk Levels and Marginal Rates of
m
12.22 and 12.23;
Substitution of Gram for Hay, Predicted From Equations
Ability Fixed at Mean of 36 Cows for 26- Week Basic
Experiment
Marginal Rate of
Substitution of Grain
Pounds of Hay Required To Produce
Lbs. of
6,300 lbs. of milk 7,300 lbs. of milk
5,300 lbs. of milk Levels
Gram
Equation Equation Equation Equation Equation
Equation Equation Equation
12.23 12.22 12.23 12.22 12.23
12.22 12.23 12.22
8,157* 10,420* 10,671* 3.08 3.83

1,000 5,528 5,642 7,974*
6,447 8,935* 8,961* 2.86 3.08
1.500 4,043 3,933 6,489
5,026 7,561* 7,540* 2.64 2.63
2,000 2,669 2,512 5,115
3,789 6,299 6,304 2.41 2.33
2.500 1,407 1,275 2,853
2,684 5,148 5,198 2.19 2.10
3.000 256* 170* 2,702
4,109 4,192 1.97 1.93
3.500 1,663 1,678
750* 3,181 3,264 1.74 1.77
4.000 735*
2,365 2,401 1.52 1.67
4.500
1,660 1,591 1.30 1.57
5.000
Estimates outside range of observations.


inputs yielding indicated milk quantities
and 12.23. For example, with 3,500 pounds of grain used in combination
with hay to produce 7,300 pounds of milk, the substitution rate of grain
for hay in Table 12.9 is 1.93; it is 1.97 for equation 12.22 and 1.93 for
equation 12.23 in Table 12.10. However, with grain at 5,000 pounds the
substitution rate for equation 12.18 in Table 12.9 is only 1.09, while it
is 1.30 for equation 12.22 and 1.57 for equation 12.23 in Table 12.10.
In Table 12.10, the rates of substitution at the extremes of the iso-
quants suggest greater difference between the estimates of equations
12.22 and 12.23 than visual comparison of figures 12.3 and 12.4 would
indicate. However, between 2,000 and 4,000 pounds of grain, the two
milk production functions give almost identical slopes as indicated by
the similarity of substitution ratios. The predicted amount of hay re-
quired with a specified amount of grain also is quite similar for equa-
tions 12.22 and 12.23 for grain inputs of 2,000 and 4,000 pounds (see
Table 12.10). Similarly, the hay quantities in Table 12.9 compare fa-
vorably with those of Table 12.10 for grain inputs over the range of 2,000
to 4,000 pounds.
The substitution ratios are the same for a given grain input regard-
less of the level of milk production in equations 12.22 and 12.23. This
condition holds true because equations 12.22 and 12.23 have only a linear
term for hay. Therefore, the equations derived from them, which de-
fine the marginal ratios of substitution of grain for hay, contain no hay
term. A given grain input, regardless of the hay input per cow, will
have the same substitution rate as higher milk levels are attained. Be-
cause of this, the isoclines (lines a, b, and c) are vertical and linear, as
illustrated in Figure 12.3, for estimates based on equation 12.22 with a
squared term for G. (The isoclines for Figure 12.4 are also vertical
and linear.)
(12.24)
||
= 3.5270 - .000446 G
,s
(12.25) !|r = 129.3668 G“ -.2584
Since the isoclines in Figure 12.3 are vertical, the optimum ration
would be found by determining the quantity of grain which results in a
marginal productivity of grain, 5M/6G, equal to the price ratio
m
P g /P . Once the quantity of grain has been determined (i.e., points
on the grain axis such as those indicated at the bottom of lines a, b, and
c), the vertical isocline shows the combinations of grain and hay to be
used. Hence, if the grain to milk price ratio were 1.25, line “a” should
be followed for any prices of grain and milk giving that ratio. In other
words, using equations 12.22 and 12.23 as predictors of the milk produc-
tion surface would lead to this ration recommendation: Feed the cow an
amount of grain consistent with the prices of grain and milk and allow
her to eat hay to her stomach capacity.
Lines b and c, Figure 12.3, show estimated optimum amounts of
grain for the 26- week period when the grain to milk price ratio is 1.0
and .75. In each case, the cow should be given the specified quantity of
grain in the period and allowed to eat hay to stomach capacity. This
procedure parallels the common feeding practice or nutrition recom-
mendation where the cow is given grain in relation to her milk produc-
tion and is allowed to consume forage on a free- choice basis. Hence, it
can be said that this conventional practice is consistent with greatest
profit only if the milk production surface is characterized by equations
such as 12.22 and 12.23; namely, the functions must have a linear term
only for hay and must give vertical isoclines.
The ration solutions indicated by the isoclines corresponding to the
milk and grain prices shown in Figure 12.3 have been determined by
equating the partial derivatives of milk in respect to grain with the ap-
propriate price ratios. Since the derivative for hay is a constant equal
to .4089, the cow should be fed hay to the stomach limit after she is fed
grain in line with price ratios, as long as the hay to milk price ratio is
less than .4089. Grain feeding would, however, be adjusted to changes
in price ratios (as indicated by lines a, b, and c in Figure 12.3) if
profits are to be maximized. With a hay to milk price ratio greater
than .4089, the cow would be fed hay only at the physiological minimum.
The isoquants for equation 12.18 (Figure 12.2) do not have vertical
isoclines. However, their slopes are so great that the ration recom-
mendations mentioned above again apply; namely, feed grain in line
with the grain to milk price ratio and allow free choice of forage.
Linear Function
A final function fitted to the 26- week data for milk production by the
36 cows of the basic experiment was one with only linear terms for hay,
grain, and ability (equation 12,26). As Table 12.8 indicates, the coeffi-
cient for each of the three variables is significant at the 1 per cent proba-
bility level, a condition which also held true for the power equation (12.3)
and the modified quadratic equation (12,18). In equation 12.26, 75.2 per
cent of the variance in milk production was explained by the linear
terms. This proportion is slightly less than the 76.6 per cent for modi-
fied quadratic equation 12.18 and slightly more than for power equation
12.3, although the differences are not significant in a probability sense.
(12.26) M = .4154 H + .9560 G + .8570 A + 50.35
5M .4154
(12.27)
6H
6M .9560
(12.28)
5G
(12.29) H = 2.4071 M- 2.3017 G -2.0628 A -121.19
5H 2.3017
(12.30)
6G
If equation 12.26 were accepted as the estimate of the production
surface, the marginal rate of substitution would be predicted by equa-

tion 12.30 as 2.30 regardless of the proportions of hay and grain in the
ration. Similarly, 1 pound of hay would be predicted by equation 12.27
to yield .415 pound of milk, and 1 pound of grain would be predicted by
equation 12.28, to produce .956 pound of milk, regardless of the level of
feeding. Constant substitution and transformation rates such as these
are assumed in conventional TDN or ENE evaluations of feeds. The
substitution ratio predicted from equation 12.26 is considerably higher
than that assumed by ENE and TDN transformations. One reason why
this greater value may be predicted is that it is based on the entire area
of the milk isoquant map included in the study. Hence, the 2.301 figure
for the entire surface may be consistent with a smaller substitution value
at the convergence of isoclines (the usual point of evaluation), particu-
larly if the production surface is actually nonlinear as illustrated later.
Finally, part of the difference in substitution ratios (between the linear
estimates of this study as compared to the linear TDN and ENE esti-
mates) may grow out of experimental error in the current data.
The following nutrition recommendations would be followed for the
26- week period if linear equation 12.26 were used as the predicter of
the milk production surface: hay alone would be fed to stomach capac-
ity if the grain to hay price ratio were greater than 2.301 and the hay
to milk price ratio were less than .415. This is true since the deriva-
tive of milk with respect to hay by equation 12.27 is .415 while the de-
rivative of hay with respect to grain is 2.3017 by equation 12.28. If the
grain to hay price ratio were less than 2.301 and the grain to milk price
ratio were less than .956, grain alone would be fed to stomach capacity.
With a grain to hay price ratio greater than 2.301 and a hay to milk
price ratio greater than .415, milk production would not be profitable
from the standpoint of feed costs alone, other costs disregarded. Hence,
a linear estimate of the production function would always call for ex-
tremes in rations: cows should be fed only grain or hay (and never a
combination of grain and hay, except in the unique case where the feed
substitution ratio equals the feed price ratio) to stomach capacity, if
the feed to milk price ratio is less than the milk to feed derivative; they
should not be kept in production if the feed to milk price ratio is greater
than the milk to feed derivative.
Since the substitution ratios in Table 12.11 are constant at 2.301
(i.e., the value of the derivative in equation 12.30) all isoclines have this
same value and can be vertical, horizontal, or positively sloped. There

is, in fact, no single line representing an isocline value; rather every
point in the feed plane of Figure 12.5 has the hay to grain substitution
value of 2.301.

predicted from linear equation 12.26 with
ability fixed at mean for 36 cows (Dots show
feed inputs yielding indicated milk quantity
-
Table 12.11. Feed Combinations in Producing Specified Milk Levels and Marginal
Rates of Substitution of Grain for Hay Predicted From Linear
Equation 12.26; Ability Fixed at Mean of 36 Cows
for 26 -Week Basic Experiment
Feed to Produce Feed to Produce Feed to Produce

5,300 Pounds Milk* 6,300 Pounds Milk* 7,300 Pounds Milk*
|
Lbs. Lbs. Substitution Lbs. Lbs. Substitution Lbs. Lbs. Substitution

grain hay rate 5H/5G grain hay rate 5H/5G grain hay rate 6H/6G
1,000 5,194 2.3017 1,000 7,602'*’ 2.3017 1,000 lo.oiof 2.3017

1,500 4,043 2.3017 1,500 6,451 2.3017 1,500 8,859t 2.3017
2,000 2,892 2.3017 2,000 5,300 2.3017 2,000 7,708t 2.3017
2,500 1,741 2.3017 2,500 4,149 2.3017 2,500 6,557 2.3017
3,000 590 T 2.3017 3,000 2,998 2.3017 3,000 5,406 2.3017
— 3,500 1,847 2.3017 3,500 4,255 2.3017
-- 4,000 697 T 2.3017 4,000 3,105 2.3017
-- — 4,500 1,954 2.3017
— — — — -- — 5,000* 803 2.3017
Isoquant predicted from equation 12.29, and substitution rates predicted from equa-
tion 12.30.
tOutside range of observation.
TDN and ENE Transformations in Relation to Linear Functions
Procedures which use TDN or ENE transformations in evaluating

feeds assume a linear production function with straight-line milk iso-
quants and, hence, constant (a) hay to grain substitution coefficients and
(b) milk to feed grain transformation coefficients. This is true since 1
pound of feed is given the same TDN or ENE value regardless of the
ratio of feeds used or the level of feeding. If diminishing rates of sub-
stitution were assumed, the TDN or ENE transformation coefficients
would need to be changed as the ration changes in proportions of feeds
and level of feeding. Using the linear relationships from equation 12.26,
a pound of hay is predicted to produce .415 pound of milk and a pound of
grain to produce .956 pound of milk, regardless of the ration fed or the
level of feeding. A pound of grain is predicted to have a feeding value
2.301 greater than a pound of hay, regardless of the ration fed. If, how-
ever, equation 12.18 is used for predictions, a pound of hay is predicted
to produce .409 pound of milk, and a pound of grain is predicted to pro-
duce .953 pound of milk only at the mean level of feeding used in the ex-
periment. At this feed combination and feeding level, a pound of grain
has .9534 to .4089 or 2.3316 times as much feed value, in relation to
milk production, as does hay. But in contrast to the linear function, this
feed value relationship holds true only for the particular feed combina-
tion and level. Turning back to Table 12.9, we note that for a 6,300-
pound milk isoquant, grain has a value as great as 3.17 times that of
hay when the feed combination includes 1,500 pounds of grain and 6,566
pounds of hay over a 26- week period; it has a value of only 1.84 times
the value of hay when the ration includes 3,500 pounds of grain and
1,616 pounds of hay. With feeding at a higher level of milk production,

as denoted by the 7,300-pound isoquant, a pound of grain has a feeding
value 1.94 times that of hay when the ration includes 3,500 pounds of
grain and 4,080 pounds of hay in the 26 weeks.
The ratios of the marginal productivities or the substitution ratios
provide a basis for relative evaluation of feeds. If it can be proved that
the milk production function is linear, constants such as those assumed
in the traditional TDN and ENE evaluation of feeds are appropriate.
However, if the function is proved to be nonlinear, constant transforma-
tions are not the appropriate basis for ration evaluation. Under curved
isoquants, grain must be given a continuously lower value relative to
hay for rations which move along a milk isoquant nearer to the grain
axis; hay must be given a lower value relative to grain for rations rep-
resenting movements toward the hay axis. Actually each isocline
traces the only path of rations over which feeds can appropriately be
given a constant feeding value relative to each other; the constant value
which is appropriate will differ for each isocline, except for a milk pro-
duction function which is linear with straight-line isoquants.
Most nutritionists accept the notion that grain and hay do not sub-
stitute at constant rates over the entire milk surface. However, wide
use of TDN and ENE transformations has been continued because of
lack of other data to indicate the rate at which substitution ratios may
change. Also, some believe that the milk isoquant may be "near linear*
in the middle, with curvature especially near the ridge isocline. Be-
cause of these indications from previous research, grain alone is never
recommended and hay alone is seldom recommended.
Equation 12.26 with linear coefficients can be used to test the feeds
and results of this study against TDN and ENE transformation ratios
which implicitly assume linearity. The constant substitution rate of
2.301 (i.e., the relative feed value of grain and hay) predicted from
linear equation 12.26 compares with the constants of 1.7884 predicted
from the ENE evaluation and 1.350 for the TDN evaluation, using Mor-
rison’s standards, of feeds in the experimental ration. The ENE ratio
of 1.7884 is calculated as .7404 therms per pound of grain divided by
.4140 therms per pound of hay. The TDN ratio of 1.3550 is calculated
as .6863 pound of TDN per pound of grain divided by .5065 pound of TDN
per pound of hay.
The question posed is this: "Does the ratio predicted from equation
12.26, indicating the relative feeding value of grain and hay in producing
a given amount of milk, correspond to the conventional ratios used for
indicating relative values of grain and hay in producing energy or heat?*
Using the linear equation for the 36 cows, we examine whether the ratio
of marginal productivities of grain and hay (i.e., the substitution ratio
of the feeds in producing a given amount of milk) is similar to the ENE
and TDN ratio, when sampling or experimental error is considered for
the regression coefficients.
Table 12.12 includes the regression coefficients for grain and hay
from the linear equation and their 95 per cent fiducial limits. The
Table 12.12. Regression Coefficients From Equation 12.26, 95

Per Cent Confidence Limits and Magnitudes
to Give ENE and TDN Ratios
Item
Hay Grain
1 . Upper confidence limit .64 1.29

2. Regression coefficient .42 .96
3. Lower confidence limit .20 .62
4. Magnitude of regression coefficient to give
1.7884 ENE ratio .53* .74 t
5. Magnitude of regression coefficient to give
1.3530 TDN ratio .71* .56 t
6. testing regression coefficient
Value of t in
(2) against coefficient needed (4) to give

ENE ratio .58 1.01*
7. Value of t in testing regression coefficient
(2) against coefficient needed (5) to give
TDN ratio 1.40 § 1.86H
* If regression coefficient for grain of .956 is accepted as a parameter,

t If regression coefficient for hay of .4154 is accepted as a parameter,
t p < .30
§ p < .20
lip < .10
computed values of regression coefficients which are necessary to give

ENE and TDN ratios of feeds in the study, based on Morrison, have then
been entered on lines 4 and 5, respectively. These computed values fall
within the 95 per cent confidence limits for the ENE evaluation but not
for the TDN evaluation. Similarly, the t values testing the differences
between the regression coefficients and coefficient values necessary to
give the TDN ratio are at a higher probability level than those for the
ENE evaluation. Hence, for the feeds, cows, and data of this study, it
appears that the TDN ratio expressing the value of grain and hay in
producing a given amount of energy or number of therms is not appro-
priate for comparison of these feeds in producing a given amount of
12
milk.
Additional research is needed to study further the appropriateness
of the ENE energy evaluation of feeds. A larger study with lower stand-
ard errors also might prove differences to be significant for ENE evalu-
ations. While the comparisons above rest on the linear regression equa-
tion, it is unlikely that the milk production function is of this empirical
nature. The above analysis was made to examine differences or simi-
larities of conclusions the purely linear relationships (a condition as-
if
sumed by ENE and TDN evaluations) were accepted. The functions de-
rived with variable substitution rates in other sections of this chapter
are likely more appropriate for a fundamental analysis of feed values.
Inclusion of time as a variable in later sections of this chapter undoubt-
edly improves the estimates of feed substitution rates.
12
The test was made with the regression coefficients, but, since in the linear equation
these form the substitution coefficient for producing a given amount of milk, the statement
holds true for a linear production function.
PRODUCTION FUNCTIONS INCLUDING AUXILIARY EXPERIMENT

FOR 51 COWS IN AN OVER- ALL PERIOD
Production functions paralleling those for the 36- cow, 26- week
basic experiment were derived when data from this experiment were
pooled with the 15- cow auxiliary experiment explained previously. The
auxiliary experiment included only 16 weeks and the data for the 36- cow
basic experiment were transformed similarly. Milk output and feed in-
put include data for the first 16 weeks of the experimental period only
for the 36- cow basic experiment. Except that measurements refer to
16 weeks and 51 cows, the variables in the function explained below are
the same as those outlined earlier for the basic experiment. Time is
not included as a variable.
Quadratic- Type Functions
Since functions with squared terms and square root terms gave re-
sults which did not differ significantly for the 36- cow experiment, only
the former type of equation has been used for the 51-cow data over the
13
16- week period. The equations are listed below in the following se-
quence: 12.31, 12.34, and 12.37 refer to the basic production function
of a particular algebraic form; 12.32, 12.35, and 12.38 are the isoquant
equations derived from the production function with ability set at the
mean; 12.33, 12.36, and 12.39 are the marginal rate of substitution equa-
14
tions derived from the isoquant equation.
(12.31) M = .3449 H + 2.2711 G + .8745 A -.000264 G2 + .000072 H2 -2,264.61
(12.32) H = -2398.0 ± (6954.0) K 1610 "

-.000653 G + .000000076 G2 + .000288 M
5H 2.2711 - .000527 G
(12.33) 5G .3449 + .000144 H
(12.34) M = .8076 (H -.00005 H2 ) + 1.9895 (G -.0001 G2 ) + .9220 A - 2,472.35
(12.35) H = 10,000 + (-12,383.0) ><.6130 + .000321 G -.000000032

2
G -.000162 M
5H _ 1.9895 -.000398 G
(12.36)
6G .8076 - .000081 H
(12.37) M = .6650 H + 2.0436 G + .8860 A - .000199 G2 - 2,404.71
13
One exception was prediction of a production function with a square root term for G.
The function, M = .6601 H +.0589 G + .8763 A + 101,9959 VU- 3, 776.83, has an R value of
.9308 and t's for regression coefficients of 6.35, .17, 7.08, and 3.92, respectively. The re-
lation of this function to equation 12.37 in the text is similar to the relation of
equation
12.23 to equation 12.22 for 36 cows.
14
In each isoquant equation, ability has been set at the mean preliminary period
milk
production for 51 cows. The 2- week preliminary period data for the 15 auxiliary cows were
extrapolated to get an estimate comparable to the basic experiment.
t
(12.38) H = -3.0731 G + .000299 G2 + 1.5038 M + 393.6412
(12.39) S CKj
= 3.0731 -.000599 G
Related statistics for the production functions are given in Table

12.13. All functions account for a greater proportion of the variance in
milk production than the parallel functions for 36 cows presented earlier.
For example, equation 12.34 accounts for 86.7 per cent of the variance
in milk production while equation 12.22 accounts for only 75.7 per cent.
The t values for regression coefficients are all significant at a proba-
bility level of .05 or lower, except those for H and H in equation 12.31.
2
However, this latter equation again appears unrealistic because the co-
2
efficient for H is positive, indicating an increasing return to hay as
more of this feed is consumed. In equation 12.34, the coefficients for
H 2 and G 2 were estimated prior to prediction of the regression
'
2
equation.
(Coefficients were predicted only for the terms H - .00005 H ,
2
G -.0001 G and A.) Accepting this form for the production function
,
would allow diminishing returns to both feeds and diminishing substitu-

tion rates between them. The milk isoquants for equation 12.34 do not
have a sharp curvature. The same statement holds true for equation
12.37 which has a power term only for grain (see isoquants in Figure
12.6 for equation 12.37). Again, the isoclines for equation 12.37 are
vertical with the implications for hay feeding mentioned in respect to
equation 12.22 and Figure 12.3. Too, the substitution ratios are always
the same for a given amount of grain, regardless of the amount of hay
or the level of milk production, because the equation of substitution
rates (12.39) has only a linear grain term and does not include a term
for hay. Table 12.14 allows comparisons of rations for three levels of
milk production predicted by equations 12.34 and 12.37. For both func-
tions, the relative feeding value of hay and grain would change with the
2
proportions fed for a given milk isoquant. Since constants for G and
2
H have been “forced* into equation 12.34, it is likely that the substitution
Table 12.13. Values of R and t for 51- Cow Data Pooled for Basic
and Auxiliary Experiments for 16- Week Period
(t Values Refer to Regression Coefficients in
Order Presented in Equations 12.31,

12.34, and 12.37)
Value of t for
Equation R hi b2 bs b. b5
12.31 .9338* 1.25* 5.92* 7.10* 2.67* 1.25*

12.34 .9275* 6.35* 8.78* 7.66* —
12.37 .9314* 6.43* 6.01* 7.17* 2.28 —
12.40 .9233* 6.23* 8.40* 6.64* —
*p < .01
*p < .05
*p < .30

predicted from equation 12.37 with
ability fixed at mean for 51 cows
ratios for equation 12.37 in Table 12.14 are most appropriate for the
51 cows.
Linear Function
Equation 12.40 is a linear function fitted to the 16- week period for
51 cows. For this equation, an additional pound of hay would always
produce .6717 pound of milk, regardless of the grain to hay ratio or the
level of feeding; an additional pound of grain would always produce
1.353 pounds of milk. Grain would
(12.40) M = .6717 H + 1.3527 G + .8500 A -1,841.40
(12.41) H = 1.4888 M -2.0138 G -1.2655 A + 2,741.40
(12.42) i? = 2.0137
5 Cj
replace 2.014 pounds of hay in producing a given amount of milk.

regardless of the ration. Hay fed to the stomach limit line (or all hay)
would give the lowest cost ration for any grain to hay price ratio of
greater than 2.014; grain fed to the physiological limit (or all grain)
would give the lowest cost ration for a grain to hay price ratio less
than 2.014. The ratio, 2.014, expressing the relative feeding value of
grain and hay is less than the 2.302 ratio computed from equation 12.26
for the 36- cow basic experiment.
A comparison of (a) the substitution ratio of grain and hay in pro-
ducing a given amount of milk for the 51 -cow linear function with (b) the
ENE and TDN ratios for producing a given amount of energy, had the
same results as the same test for the 36- cow linear function; the sub-
stitution ratio could be said to differ significantly from the TDN ratio
but not the ENE ratio, supposing that a linear production function could
be accepted in predicting the milk production surface.
PRODUCTION FUNCTIONS WITH TIME AS VARIABLE

FOR 36- COW BASIC EXPERIMENT
Since time, or point of time in the lactation period, is an important
variable affecting milk production, regression equations have been com-
puted with time included and are based on the variables following. Too,
as mentioned earlier, changes in cow weights approached an equilibrium
after the outset of the experiment and allowed better prediction after
this change had been taken into account.
M is production of 4 per cent, fat-corrected milk in each 4-week period, meas-

ured in pounds.
G is quantity of grain mix (see earlier explanation), measured in pounds, con-
sumed in each 4-week period.
H is quantity of hay, measured in pounds, consumed in each 4-week period.
A is ability measured as pounds of milk produced in the preliminary period.
T is time Six such periods were included and were measured

in 4- week periods.
as the cardinal numbers, 1 through 6. The last 2 weeks of the 26-week experi-
mental period were not included.
Four production functions have been derived with variables defined

as above. These are:
(12.43) M = 4.1937 H*1506 G* 3082 a* 3716 T~ 1973
(12.44) M = 1.6302 H 2
+ 3.1309 G + .1497 A + 14.2243 T -.000388 H -.001192 G2
+ 4.3792 T 2 -.001056 HG -.1570 GT -.0865 HT -731.76
(12.45) M = .5513 H + 1.3285 G + .1488 A -110.2640 T -.000081 H 2
2 2
-.000360 G + 6.0320 T + 151.36
cs
•M 'c3
O jL,
!
oO°.£{ N N N (M
i
-a*** a
t: *s
.5 -a S* i
gL-g a 5
w £ ^
12.37
|
^ •S <2
and
12.34 CO 05 ^ lfl N
Equations
From
Predicted
o
o o o
o o
00 CM CO
Levels
Milk
Three
Producing
in
Combinations
Feed
12.14.
Table
[
433 ]
(12.46) M = .6601 H + 1.4276 G + .1553 A -.000054 H2 T -.000152 G2 T

-2.0752 T 2 -157.24
While all of the regression coefficients of equation 12.43 are accept-

able at a probability level of less than 1 per cent, this equation explains
a somewhat smaller proportion of the variation in milk production than
do the other three equations. The percentage of variation in milk pro-
duction explained by the regression equations is 74.9, 81.3, 80.2, and
80.3, respectively, for the four functions. Equation 12.44 appears to
have the greatest logical basis since (a) it allows definition of a single
ration with a single milk level maximum per cow, and (b) it allows a
more complete specification of interaction between variables than do
equations 12.45 and 12.46. Also, as indicated later, it results in an iso-
cline map which seems more consistent with nutrition logic. In equa-
tion 12.44, the crossproduct terms reduce the sum of squares of devia-
tion from regression by an amount acceptable in probability terms, as
compared with equation 12.45. Equation 12.44 gives estimates quite
similar to equation 12.46, which has t values acceptable at lower levels
of probability. Because of these several reasons, the writers believe
that equation 12.44 is as appropriate, or more appropriate, than any of
the equations in estimating the milk production surface, feed substitu-
tion rates, and profit maximizing rations. It may, however, tend to
overestimate substitution ratios near the end of the lactation period.
It is likely that, in a larger experiment with a smaller proportion of
the degrees of freedom used in estimating regression coefficients, values
would fall at probability levels as low as those for equations 12.43 and
12.46.
In evaluating thet values for the 216 observations (36 cows for 6
months) Table 12.15, it must be remembered that the six observa-
in
tions for each cow are not independent. If, however, the number of de-
grees of freedom is considered to be as low as 25 (36 minus 11) for
equation 12.44, rather than 205 (216 minus 11), the probability state-
ments indicated in the footnotes of Table 12.15 still hold true. To obtain
coifiplete independence with 205 degrees of freedom for equation 12.44,
Table 12.15. Correlation Coefficients and t Values for Equations 12.43 Through 12.46
(36 Cows With Time as Variable)
Values of t in Order of b’s in Equation

Equation R b6 b„
»>i
b. A *>. bs b.
12.43 .8654 6.69* 12.04* 9.27* 15.61* ~ — — — —

12.44 .9016 1.94t 2.60* 9.86* .34 1.14 1.6711 1.62H 1.10 3.49* 2.82*
12.45 .8953 3.77* 6.58* 9.60* 5.74* .75 1.56 11

2.24T — —
12.46 .8960 8.70* 13.72* 10.22* 4.54* 5.92* 1.65
11
- ~ -
*P < .01
Jp < .05
*p < .10
Up < .20
Table 12.16. Regression Coefficients for Quadratic Equations Estimated for

Each Month of Experimental Period
Regression Coefficient for

Month
Constant
(Equation) H G A H2 G 2
HG
1. 12.47 - 506.0 .7495 1.7177 .3075 -.000065 -.000545 -.000392
2. 12.48 - 764.0 1.1904 2.6425 .2014 -.000228 -.001096 -.000682
3. 12.49 1,851.0 -2.5569 -3.2638 .1257 .001103 .002296 .003868
4. 12.50 -1, 140.0 2.9770 4.5732 .1085 -.001188 -.002368 -.003026
5. 12.51 - 993.0 1.9343 4.1901 .1038 -.000577 -.002542 -.002402
6. 12.52 -2,313.0 4.4090 7.8840 .0591 -.001644 -.005090 -.005749
for example, it would be necessary to have 216 cows, with each one used
for a single observation.
Functions by Months
In addition to estimating production functions with time (months) as

a variable, functions of the form of equation 12.44 were estimated for
each month separately * In these six separate equations, variables in-
cluded feed input and milk output measured over the particular month.
(Time was not included as a variable, and there were only 36 observa-
tions for each month.) The regression coefficients for these monthly
observations are given in Table 12.16. Here equation 12.47 is for the
first month in the experimental period, equation 12.48 is for the second
month, etc. Table 12.17 includes the R and t values for the six monthly
functions. Because of the large proportion of degrees of freedom used
in the regression estimates and the relative within month variability of
milk production, the t values correspond to quite high probability levels,
Table 12.17. Correlation Coefficients and t Values for Equations 12.47 Through
12.52 (Presented in Table 12.16)
Value of t in Order of b's in Equation

Equation R bi b2 b3 b* b5 b6
12.47 .9614 .72 1.28 11.37 .14 .68 .34
12.48 .9464 .66 1.10 7.64 .30 .78 .33
12.49 .8396 .89 .76 3.49 .94 .88 1.14
12.50 .7935 1.30 1.30 3.04 1.26 1.06 1.07
12.51 .7379 .76 1.06 2.57 .56 .98 .72
12.52 .6720 1.64 1.92 1.32 1.40 1.80 1.60

Predictions From Functions With Time

(Month of Lactation) as a Variable
In this section results are shown using equations 12.44 and 12.46,
with time as a variable, in predicting feed combinations possible in at-
taining specified milk levels (isoquants), and in estimating the rate of
substitution between hay and grain in the various months. Equation
12.53 is the milk isoquant, based on production function equation 12.44,
where T has been set at 1 and A has been set at 2,492 (the mean of the
36 cows in the experimental period). While equation 12.53 provides es-
timates for the first month (T = 1), similar estimates can be provided
other months by assigning a particular value to T. Equation 12.54 pro-
vides estimates of substitution rates for equation 12.53. Equations
12.55 and 12.56 are, respectively, the isoquant and substitution functions
for function 12.46, with the values for T and A mentioned above.
(12.53) H = 1,989.36 -1.3608 G

+ (-1,288.66) VO 553 + .001355 G -.00000073 G2 -.001552 M
5H 2.9740 -.002384 G -.001056 H

(12.54)
5G “ 1.5437 -.000776 H -.001056 G
(12.55) H = 6,106.28 + (-9,250.69) >{4850 + .000309 G -.000000033 G2 -.000216 M
6H _ 1.4276 -.000304 G
(12.56)
6G ~ .6601 -.000108 H
Functions with time as a variable also allow estimates of the rate

of change in milk production as the lactation period progresses. Equa-
tions 12.57 and 12.58, based respectively on equations 12.44 and 12.46,
are marginal time-yield equations. They indicate the decline in milk
production associated with each unit progress of time beyond the begin-
ning of the experimental period and are used for estimates in a follow-
ing section.
(12.57) = 14.22 + 8.7584 T -.1570 G -.0865 H
(12.58) = -.000054 H2 -.000152 G2 -4.1504 T
The milk production surfaces derived from equations 12.44 and

12.46 are presented in figures 12.7 and 12.8, respectively, for the
“mean* month of the experiment (i.e., T = 3.5). 15 The surfaces are
quite similar with respect to the milk contours. Feed quantities for
15
For the “mean” month, T has been set at 3.5 to include the last half of the third month
and the first half of the fourth month.
Table 12.18. Feed Combinations in Producing Specified Milk Levels in 28 Days and
Marginal Rates of Substitution of Grain for Hay Based on Equation
12.44; Ability Fixed at Mean of Cows in Experiment — Mean
Month of Experimental Period (T = 3.5)
Marginal Rates of Substitution:

Pounds Hay Required to Maintain Pounds Hay Replaced by 1
Level of
Milk Output of Additional Pound of Grain
Grain
Along Indicated Milk Isoquants
800 900 1,000 1,110 1,190 800 900 1,000 1,110 1,190
lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs.
150 815 1,036 2.54 3.09

200 692 891 — — — 2.37 2.74
250 577 760 1,032 — — 2.24 2.50 3.41
300 468 640 877 — — 2.12 2.32 2.85
350 365 528 744 — — 2.02 2.17 2.53
400 266 423 623 973 — 1.93 2.05 2.31 4.00
450 172 323 513 806 — 1.85 1.94 2.13 2.90
500 — 229 410 674 — — 1.85 1.99 2.45
550 139 314 559 — — 1.76 1.87 2.17 —
600 — — 223 455 -- — — 1.76 1.97 —
650 — — 138 361 — — — 1.66 1.80
700 — — — 275 716 — — — 1.66 4.79
750 ~ — — 194 575 — — — 1.53 2.09
800 — — — 122 486 — — — 1.41 1.53
Table 12.19. Feed Combinations in Producing Specified Milk Levels in 28 Days and
Marginal Rates of Substitution of Grain for Hay, Based on Equation
12.46; Ability Fixed at Mean of Cows in Experiment — Mean Month
of Experimental Period (T = 3.5)
Marginal Rates of Substitution:

Pounds Hay Required to Maintain Pounds Hay Replaced by 1
Level of
Milk Output of Additional Pound of Grain Along
Grain
Indicated Milk Isoquants
(Pounds)
800 900 1,000 1,110 1,200 800 900 1,000 1,110 1,200
lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs. lbs.
150 763 1,084 3.41 5.07

200 608 870 — — — 2.82 3.67 — — —
250 477 706 1,003 — — 2.42 2.96 4.14 — —
300 363 570 822 — — 2.12 2.49 3.18 — —
350 263 454 678 965 — 1.88 2.16 2.62 3.58 —
400 174 352 557 806 — 1.69 1.90 2.23 2.82 —
450 — 262 453 677 963 — 1.69 1.94 2.35 3.21
500 — 182 362 568 820 — 1.52 1.71 2.01 2.56
550 — — 281 474 702 — — 1.52 1.75 2.14
600 — — 208 391 603 — — 1.36 1.54 1.83
650 — — 144 318 522 — — 1.13 1.36 1.59
700 — — — 254 443 — — — 1.21 1.48
750 — — — 196 377 — — — 1.07 1.21
800 — — — 145 320 — — — .95 1.07
Figure 12.7. Milk production surface and milk isoquants

estimated from equation 12.44 for mean month
of experiment, ability at mean for 36 cows
producing stated milk production levels and marginal rates of substitu-

tion derived from equation 12.44 are presented in Table 12.18 for the
“mean” month of the experimental period. Parallel quantities are pro-
vided in Table 12.19 for the mean month predictions based on equation
12.46. As data of the tables indicate, increasing inputs of hay are pre-
dicted for higher milk levels, for a given grain level, because of dimin-
ishing productivity of feed. Also, the marginal rate of substitution of
grain for hay increases (a) as the ration includes a greater proportion
of hay for given milk level and (b) as higher levels of milk are attained
with greater hay inputs, grain remaining constant. While not directly
apparent from parallel tables, the marginal rates of substitution between
grain and hay for the two equations, for a given combination of the two
feeds and time and ability fixed at the same level, are similar in early
parts of the period. However, the rate of substitution tends to widen
between the two functions as time increases. Substitution ratios for
equation 12.44 differ more from customary evaluation standards than
do those for equation 12.46.
Figures 12.9, 12.10, and 12.11 include milk isoquants for the first,
the “mean”, and the sixth month of the experimental period, based on
equations 12.44 and 12.46. The fact that the slopes of the isoquants
Figure 12.8. Milk production surface and milk isoquants

from equation 12.46 for mean month of
experiment, ability at mean for 36 cows
are similar emphasizes the point mentioned above — namely, that esti-
mates of the feed substitution rates based on equations 12.44 and 12.46
do not differ significantly. However, one condition is apparent from the
isoquant figures. The milk contours in figures 12.9, 12.10, and 12.11
take on greater curvature for higher milk levels for both equations 12.44
and 12.46 indicating that as feeding levels become greater (1) relatively
small variations in feed combinations tend to cause larger variations in
the substitution rates or (2) smaller ranges of feed combinations will
allow a specified level of milk production. Only one ration or ratio of
grain and hay would allow maximum production per cow (i.e., the point
of isocline convergence). Similarly as time progresses from Figure
12.9 to Figure 12.10 to Figure 12.11 the milk isoquants tend to take on
greater curvature, indicating a more rapid change in substitution rates
as feed proportions are varied in attaining a particular milk level. The
tendency of the milk isoquants to increase in curvature with progress of
the lactation period also is emphasized in Figure 12.12. In this figure,
1,000-pound milk isoquants based on the two functions are presented for
each of the 6 months of the experimental period. The isoquants for the
Figure 12.9. Milk isoquants for the first month of the experiment,
based on equations 12.44 and 12.46, ability fixed at
mean for 36 cows (Dots show feed inputs yielding
indicated milk quantity from 12 cows at medium
level of ability.)
two functions tend to “spread apart” with greater amounts of time be-
cause the coefficients for time, or the product of time and feeds, is
relatively greater in equation 12.46 than in equation 12.44.
Isoclines Predicted From Functions With Time Variable
Use of the functions with time as a variable would indicate that

standard ratios, such as ENE and TDN constants, are not appropriate
for evaluating feeds in relation to milk production. The basis for this
statement is the changing nature of substitution rates in tables 12.18
through 12.19. (These substitution rates have been determined relative
to a given milk level.) Since the hay to grain substitution ratio declines
with an increased proportion of grain and increases with a decreased
proportion of grain in the ration, the coefficient for evaluating feeds
should change accordingly if evaluation is relative to milk production.
The marginal substitution rate, equations 12.54 and 12.56, are the basis
for coefficients which can serve as the basis for feed evaluation when
the entire milk production surface is considered in the evaluation.
The isoclines explained earlier also provide a basis for feed evalu-
ation and recommendations: Feeds have a constant value relative to
each other only along an isocline. Isoclines, along with milk isoquants,
Figure 12.10. Milk isoquants for the mean month of

experiment, based on equations 12.44 and
12.46, ability fixed at mean for 36 cows
are presented in Figure 12.13 for milk function 12.44 and in Figure
12.14 for function 12.46. These isoclines again trace out the path of
feed combinations which result in a given substitution rate between hay
and grain as milk is taken to higher levels (i.e., is denoted by isoquants
higher in the feed plane). In other words, 1 pound of grain substitutes
for 3.00 pounds of hay for the feed combinations traced out by isocline
A. If the grain to hay price ratio were also 3, isocline A traces the
least- cost combination of feeds for the various milk levels. Similarly,
isocline C traces out the least- cost ration when the grain to hay price
ratio is 2 if equation 12.44 is used for predicting the milk production
surface. Similarly, isoclines A, B, and C in Figure 12.14 indicate the
paths of least-cost rations if equation 12.46 is used for the estimates.
While the regression coefficients for equation 12.46 have lower
standard errors than do the coefficients for equation 12.44, the slopes
of the isoclines in Figure 12.13 appear more in line with physiological
characteristics of milk production. They converge more rapidly, sug-
gesting a maximum possible level of milk production. The point at which
the isoclines converge defines a single ration consistent with maximum
milk production per cow. Convergence and definition of a maximum
milk level is not so apparent in Figure 12.14 for equation 12.46. How-
ever, the slopes of the isoclines in Figure 12.14 do approach those in
Figure 12.3 which are closer to the historic basis for recommendations
on rations. However, the historic basis is appropriate only if the
POUNDS GRAIN PER COW

Figure 12.11. Milk isoquants for the sixth month of the
experiment, based on equations 12.44 and
12.46, ability fixed at mean for 36 cows
(Dots show feed inputs yielding indicated milk
quantity from 12 cows at medium level of ability.)
marginal rate of substitution equation includes only one feed variable

with a linear coefficient, an unlikely situation unless it can be proved
that the other feed has constant marginal productivity.
Surfaces With Time and Ability Varied
The isoquant and isocline maps in figures 12.13 and 12.14 are for
the mean month (T = 3.5) of the experimental period and the mean level
of ability of the 36 cows in the experiment. The basic production func-
tions can be used to predict feed combinations, substitution rates, and
milk levels for other points of time in the lactation period and other
ability levels. Figure 12.15 includes milk production surfaces estimated
from equation 12.44 when ability is fixed at the mean of the 36 cows, for
the first, the mean, and the sixth month of the experimental period.
Parallel estimates for equation 12.46 are provided in Figure 12.16. Be-
cause of the time coefficients in the milk production functions, the milk
level declines at a decreasing rate with time. Also, the slope of the
surface declines, indicating a lower marginal productivity of either feed
as the lactation period progresses.
Figure 12.17 provides production surfaces estimated from equation
12.44 for cow ability at three different levels with time fixed at the
Figure 12.12. Feed combinations to produce 1,000 pounds milk in each

month of the experiment, ability set at mean for 36 cows
(All isoquants represent a 1,000-pound milk level.
Equation 12.44 does not permit a 1,000 pound estimate
in the sixth month within the grain input range shown.)
mean month. The slopes of these surfaces are identical, the only differ-
ence between them is the height of milk level at which the surface slopes
begin. In other words, the entire surface is moved upward as the level
of ability increases because ability is included as a linear term only in
the equation used. However, the rates of hay to grain substitution differ
between ability levels for a given milk level. This is true because a
milk contour, such as 1,000 pounds, falls lower on the sloping portion
of the surface as the level of ability increases. As is illustrated in fig-
ures 12.9 and 12.11, the curvature of the isoquants (i.e., the marginal
rates of substitution) changes for milk contours spaced further over the
feed plane. (While surfaces from equation 12.46 have not been provided
for different ability levels, they have the same basic differences illus-
trated in Figure 12.17.)
SPECIFICATIONS OF ECONOMIC OPTIMA IN RATIONS
Prediction of the milk production function or surface allows speci-

fication of the ration which will maximize returns above feed costs. The
basic conditions for profit maximization were indicated in Chapter. 2.
Using production function 12,44 for predictions, equation 12.54 can now
be set to equal any grain to hay price ratio with time and ability at
POUNDS GRAIN PER COW

Figure 12.13. Isoquants and isoclines for mean month of
experiment and for given substitution rates,
estimated from equation 12.44, ability at
mean for 36 cows
previously stated levels, as indicated in equation 12.59. (This equation

is a particular use of equation 12.54.)
2.9740 -.002384 G -.001056 H Pg

—
(12 59) =
1.5437 -.000776 H -.001056 G Ph
serving as the basis for predictions,

In this case, with equation 12.59
ability is set at the mean of cows and time is set at
in the experiment,
the first month of the experimental period. By reference to Table 12.18,
it may be seen that for any grain to hay price ratio equal to 2.5, the
least- cost combination to produce 900 pounds of milk in 28 days with a
cow of mean ability is 250 pounds of grain and 760 pounds of hay. This
is true for any level of grain and hay prices (per pound) which yields a
ratio of 2.5 for 900 pounds of output and makes no reference to the most
profitable level of production. The substitution rates in tables such as
12.18 and 12.19 allow prediction of least-cost rations for particular
milk levels in any month of the experiment (months 3 to 8 of the lacta-
tion). For example, when the price of grain is 2.45 cents per pound and
the price of hay is 1 cent per pound, the grain to hay price ratio is 2.45.
In Table 12.18, when the marginal rate of substitution of grain for hay
is 2.45, the minimum cost feed combination to produce 1,100 pounds of
MILK
POUNDS
MILK
POUNDS
FIRST MONTH MEAN MONTH SIXTH MONTH
MILK
POUNDS
Figure 12.16. Milk production surfaces estimated from equation 12.46 for first,
mean, and sixth month (28 days) of experiment, ability at mean
milk in the mean month of the experiment with a cow of mean ability is
500 pounds of grain and 674 pounds of hay.
Other price ratios can be figured similarly with interpolations made
between feed combinations. Increases in costs or sacrifices in profit
are not great for small deviations away from the feed combination
where the substitution rate is equal to the price ratio. In the case
above, for example, the cost of the optimum ration for 1,100 pounds of
A. ABILITY AT MEAN OF & ABILITY AT MEAN OF C. ABILITY AT MEAN OF

12 LOW COWS 12 MEDIUM COWS 12 HIGH COWS
Figure 12.17. Milk production surfaces estimated from equation 12.44 for “high,”
“medium,” and “low” ability cows, and for mean month of experiment
milk is $18.99. If a ration of 600 pounds of grain and 455 pounds of hay
were used, the feed cost would be $19.25; for 800 pounds of grain and
122 pounds of hay, the feed cost would be $20.82. These differences in
feed costs are perhaps not great enough to offset the added labor costs
for feeding particular hay rations. When labor costs are figured, the
least- cost milk production may be obtained when the optimum level of
grain feeding is determined in relation to the grain to milk price ratio,
with self-feeding of hay to stomach capacity as was explained in respect
to Figure 12.3.
The gain from feeding one ration rather than others along a milk
contour increases with greater curvature of the isoquant. Since the cur-
vature of the milk isoquants for equations such as 12.44 or 12.46 tends
to increase with level of milk production, gains in feeding the unique
optimum ration are greater as the level of milk production increases
(i.e., the grain to milk price ratio decreases) or as time in the lactation
period increases. While the gains from feeding unique rations are rela-
tively low from the predictions of this study, the final advantages of
particular rations for given milk levels can be determined only as the
nature of the milk surface and its isoquant family are better established.
Simultaneous Specification of Ration and Milk Level
Equation 12.44 now can be used to compute the partial derivatives.

In equation 12.60, based on equation 12.44, the partial derivative for
(12.60) = 2.9740 -.002384 C -.001056 H = ||||
grain is equated to the grain to milk price ratio when milk is 4 cents per
pound and grain is 3 cents per pound, in the first month of the experi-
ment for a cow of mean ability; in equation 12.61 (also derived from equa-
tion 12.44), the partial derivative for hay is equated to the hay to milk
price ratio when hay is 1.25 cents per pound ($25 per ton). By simulta-
neous solution of equations 12.60 and 12.61, it is estimated that the ra-
tion which will maximize return above feed costs should include 799
pounds of hay and 579 pounds of grain fed over 28 days (28 pounds of hay
and 21 pounds of grain per day) to produce 1,479 pounds of milk and a
return of $31.80 above feed costs. If a radically different ration such
as 275 pounds hay and 707 pounds grain had been fed in the first month
under these price relationships, the return above feed costs is esti-
mated at $29.63. With feeds remaining at the above prices and the milk
price rising $5 per hundred pounds, the optimum ration would include
786 pounds of hay and 648 pounds of grain, with milk production at
1,522 pounds. If the milk price were $3 per hundred pounds, the opti-
mum ration would include 820 pounds of hay and 465 pounds of grain,
with milk production at 1,387 pounds.
(12.61) = 1.5437 -.000776 H -.001056 G =
If remains at $4 per hundred pounds and hay increases

the milk price
to 1.75 cents per pound ($35 a ton), an increase in grain price to 4 cents
per pound has this effect on the optimum ration: The feed combination
should include 756 pounds of hay and 493 pounds of grain to produce
1,388 pounds of milk in the first month from a cow of medium ability.
With these same prices in the sixth month, the ration would include only
480 pounds of hay and 286 pounds of grain, producing an estimated 712
pounds of milk.
Feed and milk quantities in Table 12.20 represent optimum rations
and milk production levels derived from equation 12.44 for certain feed
and milk price situations. Estimates in Table 12.21 are based on equa-
tion 12.46. However, since the production surface for equation 12.46 is
nearly linear with time fixed at the first month (see Figure 12.16), the
feed quantities suggested for price variations in this month generally
exceed the cow’s stomach capacity. The estimates from equation 12.46
for the mean and sixth months are more similar to the estimates from
equation 12.44 in Table 12.20. The wide differences for the first month,
and the fact that equation 12.46 provides estimates far outside the range
Table 12.20. Estimated Optimum Feed Quantities and Milk Production in the First Month, Mean Month,
and Sixth Month of Experimental Period, for Various Price Ratios; Estimates From
Equation 12.44 With Ability at Mean*
Hay, Grain, and Milk Quantities (Pounds) With Milk Prices

Price
Feed Prices Per Hundredweight at
Grain Hay Grain to |

$3.00 ||
$4.00 j
$5.00
per cwt. per ton Hay |
Hay Grain Milk Hay Grain Milk Hay Grain Milk
Month 1
$2 $15 2.67 882 577 1,501 846 663 1,544 843 715 1,563
3 15 4.00 1,361 225 1,328 1,204 400 1,446 1,110 504 1,500
3 25 2.40 820 465 1,387 799 579 1,479 786 648 1,522
3 35 1.71 275 707 1,357 390 761 1,462 458 793 1,511
4 35 2.29 762 350 1,224 756 493 1,388 752 579 1,464
Mean Month
2 15 2.67 745 473 1,100 710 560 1,143 686 612 1,163
3 15 4.00 1,223 121 926 1,066 296 1,045 972 400 1,100
'
3 25 2.40 682 361 986 662 476 1,080 649 544 1,121
3 35 1.71 — t 503 — 252 657 1,062 322 690 1,111
4 35 2.29 624 247 824 618 390 988 613 476 1,064
Month 6
2 15 2.67 608 369 825 570 456 867 548 508 887
3 15 4.00 1,085 18 651 928 193 770 834 297 824
3 25 2.40 545 257 710 524 372 803 511 440 845
3 35 1.71 — t 499 — -- t 553 — 184 586 835
4 35 2.29 486 144 549 480 286 712 475 372 787
*Figures show the most profitable ration and milk level for each combination of hay, grain, and milk
prices. For example, with hay at $25.00, grain at $3.00 and milk at $3.00, the most profitable ra-
tion includes 465 pounds of grain, 820 pounds of hay and produces 1,387 pounds of milk the first m
month.
t Price per pound of grain divided by price per pound of hay.
t Physiological minimum hay quantity.
Table 12.21. Estimated Optimum Feed Quantities and Milk Production m

the First Month, Mean Month,
and Sixth Month of Experimental Period, for Various Price Ratios; Estimates From
Equation 12.46 With Ability at Mean
Hay, Gram, and Milk Quantities (Pounds) With Milk Prices

Price
Feed Prices Ratio:
per Hundredweight at t
Gram Hay Grain to

|
$3.00 $4.00 I $5.00
|
per cwt. per ton Hay*

Hay Grain Milk Hay Grain Milk Hay Grain Milk
Month 1
$2 $15 2.67 3,797 2,492 4,353 3,051 4,723 3,380

3 15 4.00 3,797 1,407 — 4,353 2,229 4,723 2,722
3 25 2.40 2,223 1,407 — 3,242 2,229 3,797 2,722
3 35 1.71 742 1,407 -- 2,038 2,229 2,871 2,722
4 35 2.29 742 321 — 2,038 1,407 2,871 2,064 —
! Mean Month
2 15 2.67 1,085 712 1,444 1,244 872 1,573 1,349 966 1,633
3 15 4.00 1,085 402 1,186 1,244 637 1,427 1,349 778 1,539
3 25 2.40 635 402 1,035 926 637 1,347 1,085 778 1,486
3 35 1.71 212 402 823 582 637 1,218 820 778 1,407
4 35 2.29 212 92 462 582 402 1,012 820 590 1,275
Month 6
2 15 2.67 633 415 878 726 508 953 787 563 988
3 15 4.00 633 234 727 726 372 868 787 454 934
3 25 2.40 370 234 639 540 372 822 633 454 903
3 35 1.71 124 234 516 340 372 747 478 454 857
4 35 2.29 124 54 306 340 234 626 478 344 780
Price per pound of gram divided by price per pound of hay.

t Pounds milk not computed for feed quantities in first month because of obvious inability of animal to
consume such quantities.
of stomach capacity, again suggests that equation 12.44 is preferable to

equation 12.46, even though the latter has lower standard errors than the
former.
It is of interest to compare the optimum rations indicated in Table
16
12.20 with those based on Jensen et ol. Estimates based on the Jensen
study indicate that for prices of $25 per ton for hay and $3 and $4 per
hundred pounds respectively, for grain and milk, the optimum ration
would include about 350 pounds of grain, leaving capacity for 800-900
pounds of hay taken free choice. This estimate is for a cow producing
40 pounds of milk per day (1,120 pounds in 28 days). Estimates for the
same prices, based on equation 12.44 of this study (Table 12.20), indi-
cate 476 pounds of grain and 662 pounds of hay to produce 1,080 pounds
of milk, for a cow of mean ability in the mean month of the experimental
period. With prices of $35 per ton for hay, $4 for grain, and $5 for
milk, the estimates based on the Jensen study include about 330 pounds
of grain and a residual of 800-900 pounds of hay; the estimates of the
current study include 476 pounds of grain and 613 pounds of hay to pro-
duce 1,064 pounds of milk in 28 days.
Predictions of the daily rate of grain and hay feeding to give maxi-
mum return above feed costs in the first, mean, and sixth months of the
16
Jensen, E. et al. Input-output relationships in milk production. Tech. Bui. 815. USDA,
experimental period can be made by dividing the quantities of Table

12.20 by 28 days. These estimates are for a cow of mean ability and
are based on equation 12.44. Estimates for other ability levels can be
predicted by adding to or subtracting from these quantities an amount
based on the coefficient for A in equation 12.44.
STOMACH CAPACITY
Dairy animal stomach capacity depends partly upon the size of the
cow and, hence, upon the breed of cow. For the large dairy breeds, es-
timates may be made from data of this and other studies. While ani-
mals in this study were not fed rations of hay or grain alone, other
studies have included such rations. The United States Department of
Agriculture conducted hay feeding trials in which animals were full-fed
17
alfalfa for 365 days. Consumption averaged 14,352 pounds per cow or
39.3 pounds per day. Maximum consumption was 47.1 pounds per day,
while the minimum taken ad lib was 30.4 pounds. Other agricultural
experiment stations reporting such esêriments include Kansas 18 ,
19 20
Nevada Oregon
,
and California. 21 Missouri also has reported feed-
,
ing only a concentrate mixture for 15 entire lactations without materially

22
lowering production. However, other trials suggest a minimum hay
requirement of 5 to 6 pounds per day. 23
Table 12.22. Estimated Maximum Grain and Hay Consumption by Large Breed Dairy
Cows in Pounds per Day, per 28 Days, and per 182 Days
Pounds per Day Pounds per 28 Days Pounds per 182 Days
Grain Hay Grain Hay Grain Hay
0 40 0 1,120 0 7,280
5 37 140 1,036 910 6,734
10 32 280 896 1,820 5,824
15 27 420 756 2,730 4,914

20 21 560 588 3,640 3,822
25 13 700 364 4,550 2,366
29 5 812 140 5,278 910
17
Groves, R. R., et al. Feeding dairy cows on alfalfa hay alone. Tech. Bui. 610. USDA,
18
Reed, 0. E.,ef al. The relation of feeding and age of calving to development of dairy
heifers. Kan. Agr. Exp. Sta. Bui. 233. Manhattan. 1924.
19
Headley, F. B. Feeding experiment with dairy cows. Nev. Agr. Exp. Sta. Bui. 119.
Reno. 1930,
20
Oregon Agr. Exp. Sta. Studies with alfalfa hay. Director’s Biennial Report. P.41-43.
Corvallis. 1924-26.
21
Wall, F. W. Alfalfa as the sole feed for dairy cows. Jour. Dairy Sci. 1: 447-61, 1918.
22
Missouri Agr. Exp. Sta, Bui. 444. Agricultural investigations. Columbia. 1942.
23
Loosli, J. K., Lucas, H. L., and Maynard, L. A. The effect of roughage intake upon the
Table 12.22 is derived from data of this study and others mentioned.
Itsuggests maximum daily, 28- day, and 182-day intakes of several ra-
tion combinations and is presented only as a guide for use with large
breed dairy animals. Figure 12.18 shows the estimates of Table 12.22
for a 28-day period, with line ab defining the estimated maximum feed
intake of various ration combinations.
0 140 280 420 560 700 840

POUNDS GRAIN
Figure 12.18. Estimates of stomach capacity of large
breed dairy cows in 28-day period
fat content of milk.Jour. Dairy Sci. 28: 147-53. 1945; and Monroe, C. F. and Krauss, W.
butterfat produc-
E. Relationship between fat content of dairy gram mixtures and milk and
tion. Ohio Agr. Exp. St a. Bui. 644. Wooster. 1943.
CHAPTER 13
Ear! O. Heady
Harold O. Carter
C. C. Culbertson
and Substitution Coefficients
for Beef
STUDY deals with prediction of production functions, iso-

T HIS
quants, isoclines, marginal substitution rates, and related eco-
nomic quantities in the rations of feeder cattle. These basic data
are important to managerial decisions by farmers since feed repre-
sents around 75 per cent of the costs involved in feeding cattle. Infor-
mation on feed replacement rates also is needed as a basis for analyz-
ing the effects of government production control programs as well as
for individual farmer decisions. It has been shown previously that
whether acreage allotments, allowing shifts of land from grain to for-
age, will reduce or increase total output depends on two things: (1) the
extent and rates at which crops substitute for each other in the crop
rotation, (2) the extent and rates at which these same crops as feed
1
substitute for each other in the livestock ration. The optimum soil
conservation plan also depends on these same substitution data.
OBJECTIVES AND DATA
Because little knowledge is available on feed substitution quantities

for beef cattle, one objective of this study is to estimate livestock-feed
production functions and the feed substitution relationships which can
be derived from them. Production functions are derived for steer
calves where the feeds include corn, legume hay, and protein supple-
ment. Time functions are derived to indicate the amount of time re-
quired for attainment of a given weight when different rations are fed.
Given the production functions and the time functions, comparisons can
be made of the monetary gain from feeding the least -cost rations with
that from feeding to produce a specific quality of beef in a minimum
amount of time.
The data upon which this study is based are drawn from experi-
ments conducted by the Animal Husbandry Department at Iowa State
1
Cf. Heady, Earl O. Resource and revenue relationships in agricultural production
control programs. Review of Economics and Statistics, 33(3): 228-40. 1951.
452
PRODUCTION FUNCTIONS FOR BEEF 453
Figure 13.1, Production surface illustrating beef grades
University over a period of 25 years. 2 Data were used from experi-

ments employing 272 choice feeder calves and including approximately
the same quality and kinds of feeds in each year. Experiments used
for estimating production functions in this study were those including
only good quality legume hay and corn. In the experiments of the 25
years, rations used included a fairly wide range of corn, hay, and pro-
tein supplement ratios. The main supplement used was linseed and
soybean oilmeal. To reduce the number of feed variables to three
(corn, hay, and protein supplement), all supplements were converted in
terms of digestible protein content to a linseed oilmeal (L.O.M.) basis.
Since the feeds from the different years are uniform generally in
respect to kind and quality, it appears that the data lend themselves to
production function analyses. Also, the quality of livestock appears
sufficiently homogeneous over the time period concerned. Generally,
a feeder calf grading choice in one year has been comparable to a calf
of the same grade in another year. Beef cattle production has not ex-
perienced breeding revolutions comparable to those for hogs and poul-
try. Too, given grades of corn and hay changed little if any over the
years from which experimental data have been drawn.
2
The experiments included are those in the Iowa State University, Animal Husbandry
Department numbered: 272 X; 287 XI; 298 HI,* 314 V, VI, VII, VTII, IX, X, XI, XII; 340 I;
vm
350 I, II, in, V, VI, TO, VIA; 352 I; 360 I, n, IV, V, VII, ;
378 I, IV, V, VII, Vin. The
initial weight of calves was approximately 400 pounds.
454 PRODUCTION FUNCTIONS FOR BEEF
This study is made in an attempt to utilize existing data in deriving

production functions and substitution coefficients for beef cattle. It
follows the recommendations provided by a joint meeting of animal nu-
tritionists and farm management workers in March, 1952. This meet-
ing was sponsored, through the facilities of the Farm Foundation, by the
North Central Farm Management Committee to outline experiments and
methods for obtaining information on feed substitution rates. The ani-
mal nutritionists present expressed the belief that existing experi-
mental data provided a sufficient basis for these estimates. The cur-
rent study was undertaken accordingly. It is, however, the belief of the
authors that existing data, while useful in obtaining information on sub-
stitution coefficients, is not ideal for this purpose. A 3 -year experi-
ment designed for prediction of beef production functions has now been
initiated at Iowa State University.
The procedure generally outlined for simultaneously specifying the
optimum marketing weight and the optimum ration must be modified in
beef production, because different rations result in different qualities of
beef. Thus P the price of gain, will change with the ration fed over a
,
g
limited range of forage to grain ratios. Thus, the models of beef gain
surfaces and isoquants may need to be modified as illustrated in Figure
13.1. The gain surface is identical to conventional ones as in Chapter 9
except iso -value lines are substituted in place of iso -gain lines and a
physiological limit line is introduced. Iso -value lines represent, for
example, all possible combinations of concentrate and forage which will
produce a given value of beef gain. Line OK divides the hypothetical
production surface into grades “good” and “choice” according to the ef-
fects of the ration fed. Figure 13.2, the two-dimensional counterpart of
Figure 13.1, shows iso -value lines at 1 ), at 2 ), and at 3 ). Iso-value lines
are discontinuous since two separate products or grades are produced.
For example, if "good” beef has a price of $15 per hundredweight while
1
"choice” has a price of $20, then at may represent a $40 iso -value
)
line following the 266 -pound gain contour through the "good” area of the
surface and then follow the 200 -pound gain contour through the choice
3
area.
The principle which specifies the least -cost methodof producing a
given gain, regardless of the grade, is the same as that for determining
the least-cost method of producing a given value of beef when grade dif-
ferentials are considered. In application to Figure 13.1, for example,
usual -procedures might be employed to specify the least-cost method
1 2 3
of producing a value of beef represented as a^ ) a( ), or a^ ) Diffi-
, .
culty arises in simple applications of the principle, however, because a

continuous value function currently cannot be derived. In most cattle
feeding operations, it is more important, at some level of gain, to feed
a ration which increases the grade but does not minimize costs of pro-
ducing a physical quantity of gain. However, within a grade range,
knowledge of substitution coefficients is important for specifying profit-
able rations.
EMPIRICAL RESULTS
As a basis for predicting the beef production surface from available

data, several algebraic forms of equations were employed. Of the func-
tions fitted, the one presented in equation 13.1 appeared logically and
statistically most appropriate.
2
(13.1) Y = .157402C + .361070P + .112332F - .00001612C
2 2
- .00090147P - .00005958F - .00003612CP - .00000637CF
+ .00054144PF + 3.58753
The variables in this equation are defined and measured as follows:
Y is total gain inpounds per steer measured from the beginning of the feeding
period to weighing date. Average starting weight of all groups of steers in-
eluded in the analysis was 400 pounds. Consecutive gain observations were
taken at 28-day intervals on a given ration line; with each observation being
the accumulated sum of previous observations on that ration line. In most
cases, the average performance of groups of 8 or 10 steers to a lot were ob-
tained.
C is total intake of corn in pounds from the time grain feeding is started to the
particular weighing dates. Hence, the first observation of corn is from be-
ginning of the experiment to the first weighing date, the second observation
is from the beginning of the experiment to the second weighing date, etc.
Corn silage was converted to corn equivalent.
3
Exaggerated price differentials were used to illustrate the point.
P protein supplement measured in pounds. Observations are

is total intake of
measured as outlined above for corn. Protein supplement is in terms of
pounds of linseed oilmeal. All other proteins used in the experiment were
converted to linseed oilmeal equivalent (L.O.M.) on the basis of per cent
digestible protein.
F good quality legume hay measured in pounds. Observations

is total intake of
are measured as outlined for corn. Grass silage, used in small amounts in
some experiments, was converted to hay equivalent.
The coefficient of determination for equation 13.1 is .87, indicating

that the of variance in beef gains on choice feeder calves
major portion
is explained by the three feed categories employed. Standard errors
for the regression coefficients are presented in Table 13.1. Variables
were left in the estimating equation where they appeared consistent with
production logic; even if they were not significant at conventional prob-
ability levels. However, as is pointed out later, the coefficients for
certain nonlinear variables are small and have only slight effect on
substitution rates. Certain equations were fitted where all coefficients
could be accepted at probability levels of .99 per cent. However, these
appeared, by inspection of scatter diagrams, and in terms of previous
nutrition and production logic, to be less in accord with available data.
Table 13.1. Standard Errors for Regression

Coefficients in Equation 13.1
Regression Coefficient
for Standard Error
C .03858
P .17660
F .07747
2
C .0002
P2 .00043
F2 .00007
CP .00016
CF .00006
FP .00023
Autocorrelation Problems
When fitting the production function with the classical method of

least squares, the assumption of independence for each observation is
explicitly made. In these livestock experiments, each observation is
the average accumulated sum of feed and gain for each pen of animals.
For example, feed and gain observations taken at the end of each month
are accumulated sums of observations taken in previous months. Hence,
the second observation is related to the first, and the third observation
in a ration is related to the first and second observation, etc. Although
the series of observations taken on a pen is correlated, it is independent
of series of observations taken on other pens or rations.
Equations Derived for Estimates
The basic production function of equation 13.1 allows derivation of

equations for isoquants, isoclines, and input-output relationships. Since
corn and hay represent the major portion of feed costs in beef fattening
and because major land use problems revolve around these crops, most
of the analysis which follows is in terms of subfunctions defining rela-
tionships between corn and forage. Accordingly, linseed oilmeal equiv-
alent, hereafter indicated as is held constant at different ratios
in respect to corn and at different absolute levels, and feed relation-
ships are derived accordingly. More specifically, for part of the follow-
ing analysis L.O.M. is held constant at (a) 25 per cent of the quantity of
corn in the ration (i.e., corn and L.O.M. are always combined in a fixed
ratio of 4 to 1), (b) 15 per cent of the quantity of corn in the ration (i.e.,
corn and L.O.M. are always combined in a fixed ratio of approximately
7 to 1), and (c) a constant amount of 175 pounds (i.e., the ratio of corn
to L.O.M. changes as the quantity of corn in the ration is varied) is
used. A different production surface, with corn and hay as variables,
exists for each ratio or absolute quantity of L.O.M. The prediction
equations for a, b, and c above in terms of corn and hay become, re-
spectively, those in equations 13.2, 13.3, and 13.4. Variables in the
following equations are measured as defined previously.
(13.2) L.O.M. held constant at 25 per cent of corn.
Y = .2476C - .00008 1C 2 + .1123F - .000059F 2 + .000129CF + 3.59
(13.3) L.O.M. held constant at 15 per cent of corn.

Y = .2116C - .000042C 2 + .1123F - .000059F2 + .000075CF + 3.59
(13.4) L.O.M. held constant at 175 pounds.

Y = .1511C - .000016C 2 + .2071F - .000059F2 - .000006CF + 39.17
Equations 13.2 to 13.4 can be used to predict a beef production func-

tion when L.O.M. is either held in a constant ratio to corn or is used in
a fixed absolute amount. Equations of gain isoquants then can be de-
rived from these production function equations. Corresponding to the
three production function equations above, the isoquant equations are:
(13.5) L.O.M. equal to 25 per cent of corn (i.e., held in constant ratio
of 4 partscorn and 1 part L.O.M.).
F = 942.699 + 1.0825C t 8,392.08(.G1347346 + .000088C
2 5
- .0000000028C - .00023 832Y)*
(13.6) L.O.M. equal to 15 per cent of corn (i.e., held in constant ratio
of approximately 7 parts corn and 1 part L.O.M.).
F= 942.699 + .6281C t 8,392.08(.01347346 + .00006724C
2 5
- .0000000044C - .00023 832)'
(13.7) L.O.M. held constant at 175 pounds (i.e., the proportion of

corn to L.O.M. changes as the quantity of corn varies).
F = 1,737.865 - .0535C ± 8,392.G8(.05221811 + .OOOQ3336C

2 5
- .00Q0000Q39C - .Q0G23832Y)*
In these isoquant equations, amounts of hay (F) required to produce

specified gains (Y) are expressed as a function of corn intake per ani-
mal. These equations can be used to derive all possible combinations
of corn and hay which will produce gains of specified amounts.
Equations defining marginal rates of substitution can be derived
from the isoquant equations 13.5 to 13.7. For example, with L.O.M.
specified as above, the marginal rates of substitution become the de-
rivatives of forage (F) in respect to corn (C) as shown below for equa-
tion 13.5.
(13.8) L.O.M. equal to 25 per cent of corn (i.e., held in a constant

ratio of 4 parts corn and 1 part L.O.M.).
a .000163C
J5F - .0Q0129F - .247669
5C .112332 - .000119F + .000129C
Equations such as 13.8 above can be used to predict the marginal

rate of substitution of corn for hay — at a given L.O.M. level — for the
various combinations of the two feeds which will produce a specified
4
gain. While the equations are expressed in terms of substitution of
corn for hay, marginal rates of substitution of hay for corn can be de-
rived as the reciprocal of 13.8.
GAIN ISOQUANTS AND SUBSTITUTION RATES
Beef gain isoquant schedules and marginal rates of substitution can

be derived from the above equations. They are presented in the tables
that follow. Isoquant equation 13.5 and substitution rate equation 13.8,
4
The marginal rate of substitution of corn for hay is empirically derived from the ratio:
6F 6Y 6F
MRS ‘
"
5C 5C 6Y
It isimportant to note, however, that with respect to the two cases where protein is fixed as
a percentage of the quantity of corn m
the ration 3Y/ 3C is slightly modified. That is, an
additional side condition is made that P = f (C). Thus, in general terms, equation 13.8 is
written:
6Y 6Y . dP
+
6F 5C 5P dC
5C _6Y
6F
6P =
For equation 13. 8, 0.25. Thus equation 13.8 provides estimates of the marginal rate
of substitution of 1 pound of corn and 0.25 pounds of L.O.M. for hay.
Table 13.2. Isoquant Schedules Showing Possible Feed Combinations and Marginal
Rates of Substitution of Corn for Forage at Four Gain Levels,
With Good -to -Choice Calves Weighing 400 Pounds at the
Outset (L.O.M. Held Constant at 25 Per Cent
of the Quantity of Corn)
Total Pounds of Pounds of Feed per

Ratio of 6F* 6C§
Feed Pound of Gain
Forage to Corn 6C 6F
Corn* Forage* Corn^ Forage
200 pounds gain
500 683 2.50 3.41 1.37 2.66 0.38

600 479 3.00 2.39 0.80 1.59 0.63
700 346 3.50 1.73 0.49 1.10 0.91
800 252 4.00 1.26 0.31 0.81 1.23
300 pounds gain
700 1,306 2.33 4.35 1.87 6.40 0.16

800 945 2.67 3.15 1.18 2.32 0.43
900 763 3.00 2.54 0.85 1.45 0.69
1,000 642 3.33 2.14 0.64 1.02 0.98
1,100 555 3.67 1.85 0.50 0.74 1.35
1,200 490 4.00 1.63 0.41 0.55 1.82
400 pounds gain
1,024 1,439 2.56 3.60 1.40 3.64 0.27

1,100 1,231 2.75 3.08 1.12 2.11 0.47
1,200 1,063 3.00 2.66 0.89 1.35 0.74
1,300 950 3.25 2.38 0.73 0.95 1.05
1,400 869 3.50 2.17 0.62 0.69 1.45
1,500 809 3.75 2.02 0.54 0.51 1.96
1,600 765 4.00 1.91 0.48 0.38 2.63
1,700 733 4.25 1.83 0.43 0.27 3.70
*Derivedfrom isoquant equation 13.5.

t An amount equal to 25 per cent of these amounts would be needed to give total
concentrates per pound of gain.
$ Derived from substitution equation 13.8.
§ Column 7 is the reciprocal of column 6.
both derived from production function 13.1, provide the basis for the
estimates in Table 13.2. For the estimates in Table 13.2, L.O.M. is
held constant at 25 per cent of the quantity of corn in the ration. Equa-
tion 13.6 and its substitution equation provide the basis of the estimates
in Table 13.3 where protein (pounds of L.O.M. equivalent) is held con-
stant at 15 per cent of the quantity of corn in the ration.
For each gain level (i.e,, 200, 300, 400, and 500 pounds) and each
L.O.M. level, the derived marginal rates of substitution between corn
and forage are at a diminishing rate. 5 In other words, based on the
5
It should be pointed out that marginal rates of substitution between feeds apply only to
a specific gain contour, i.e., 200, 300, 400 or 500 pounds of gain, rather than the gain inter-
vals, i.e., the portion of the production surface between the initial feeder calf weight of 400
pounds and a specific gain contour. However, the feed quantities shown refer to the gain in-
terval because they are derived as consecutive observations from the outset of the feeding
experiment.
Table 13.3. Isoquant Schedules Showing Possible Feed Combinations and Marginal
Rates of Substitution of Corn for Forage at Four Gain Levels,
With Good -to -Choice Calves Weighing 400 Pounds at the
Outset (L.O.M. Held Constant at 15 Per Cent
of the Quantity of Corn)
Total Pounds of Pounds of Feed per

Ratio of 8F* 5C §
Feed Pound of Gain
i
Forage to Corn 6C 6F
Corn* Forage* Corn^ Forage
200 pounds gain
600 752 3.00 3.76 1.25 3.21 0.31

700 513 3.50 2.57 0.73 1.84 0.54
800 358 4.00 1.79 0.45 1.32 0.76
900 242 4.50 1.21 0.27 1.02 0.98
300 pounds gain
950 1,175
1,000 988 3.33 3.29 0.99 2.89 0.35
1,100 769 3.67 2.56 0.70 1.71 0.58
1,200 624 4.00 2.08 0.52 1.23 0.81
1,300 516 4.33 1.72 0.40 0.95 1.05
1,400 431 4.67 1.44 0.31 0.76 1.32
400 pounds gain
1,400 1,315 3.50 3.29 0.94 3.17 0.31

1,500 1,084 3.75 2.71 0.72 1.74 0.57
1,600 938 4.00 2.35 0.59 1.22 0.82
1,700 832 4.25 2.08 0.49 0.93 1.08
1,800 749 4.50 1.87 0.42 0.74 1.35
^Derived from, isoquant equation 13.6.

tAn amount equal to 15 per cent of these amounts would be needed to give total
concentrates per pound of gain.
^Derived from substitution equation corresponding to 13.6.
§ Column 7 is the reciprocal of column 6.
equations used, the rate in which one feed substitutes for the other feed
declines as the ration is changed to include a greater proportion of the
first. For example, in Table 13.2, with 500 pounds of corn and 683
pounds of forage combined in the ration to produce 200 pounds of gain,
the marginal rate of substitution of corn (plus 0.25 pounds of L.O.M.)
for forage is 2.66; i.e., at a 600 -pound weight level per animal and with
feed combinations above, one additional pound of corn and 0.25 pounds
of L.O.M. is predicted to replace 2.66 pounds of forage. At a point on
the same 200 -pound gain contour with 700 pounds of corn and 346 pounds
of forage, one additional pound of corn and 0.25 pounds of L.O.M. is
predicted to replace only 1.10 pounds of forage.
Gain isoquants corresponding to the data in columns 1 and 2 in
tables 13.2 and 13.3 and derived from equation 13.6 are presented in
figures 13.3 and 13.4, respectively. As mentioned previously, these
isoquants specify the various feed combinations which are predicted to
produce the gain levels specified on 400 -pound choice feeder calves.
Similarly, the slope at a given point on an isoquant indicates the rate at
VW1NV
H3d
3HV1KII
191)
NHOO
Figure 13.3. Gain isoquants for 400 lb. -calves (good to choice), protein (L.O.M.)
constant at 15 per cent level of corn inputs
Figure 13.4. Gain isoquants for 400 lb.-calves (good to choice), protein (L.O.M.)
constant at 25 per cent level of corn inputs
which forage substitutes for corn, or the rate at which corn substitutes
for forage. As indicated in the figures, the curvature of the isoquants
changes only very gradually, suggesting that substitution ratios may not
depart greatly from constant rates for animals of this size, age, and
grade. It is possible that for small animals in early rapid growth
stages, or heavier and older animals in final fattening stages, the sub-
6
stitution rates may change more rapidly. Too, greater curvature of
the isoquants might have been predicted had the range of experimental
rations been greater, and had the experiments been designed for the
particular purpose. However, from the data available in the experi-
ments, it appears that cattle in the weight range 400-1,000 pounds have
relatively great physiological ability to substitute forage and corn, or
vice versa, without a rapid decline in marginal substitution rates.
Corn-Forage Substitution Rates

on Selected Ration Lines
Of further interest in decision on livestock feeding is the relative

productivity of various feeds in rations for cattle at different weight
levels. Consequently, marginal rates of substitution between corn and
forage were derived for selected forage to corn ratios (i.e., the quantity
of corn in the ration is held in a constant proportion to the quantity of
forage, at different gain levels) Isoquant equation 13.6 and its corre-
.
sponding substitution equation provides the basis for the corresponding

estimates in Table 13.4. Forage to corn ratios of 0.33, 0.50, 0.67, 0.83,
and 1.00 were selected as being representative of feed combinations
7
observed in the experimental data used.
Table 13.4 shows that with a forage to corn ratio of 0.50 (i.e., 2
pounds of corn for each 1 pound of forage), the predicted marginal rate
of substitution of corn for forage is 1.63 at the 200 -pound gain contour,
1.17 at the 300-pound gain contour, and 0.96 at the 500-pound gain con-
tour. These data would indicate that corn is less important in the ra-
tion relative to forage (given a high protein level) as the feeder calf ac-
quires additional gain. This result appears unrealistic in view of
nutritional logic and research which indicates that energy feeds (e.g.,
corn) are relatively more important than forage in the ration as the
weight of the animal increases and the fattening process takes place.
However, the “high” protein (pounds of L.O.M.) content in the ration
probably substitutes in part, on an energy basis, for corn. At high for-
age ratios, such as 1.0, the rate of substitution of corn for hay is pre-
dicted to increase with greater steer weight.
6
The linear, positively- sloped lines are included to suggest limits in rations which
might be expected to provide satisfactory growing and fattening rations. Few of the experi-
ments provided data outside of these ranges.
7
The reader is again reminded to refer to a previous footnote. Briefly, the marginal
rates of substitution in Table 13.4 refer, respectively, to 1 pound of corn and 0.25 pounds of
L.O.M. for hay.
Table 13.4. Marginal Rates of Substitution and Feed Combinations on Different

Forage to Corn Ration Lines at Four Levels of Gain With
400-Pound Good- to -Choice Calves (L.O.M. Equal to
15 Per Cent of the Quantity of Corn)
200--Pound gain 300--Pound gain 400--Pound gain

Ratio of
Forage 5F* 6F* 6F*
Corn Forage Corn Forage Corn Forage
to Corn 6C 6C 6C
0.33 860 286 1.12 1,350 450 0.82 1,950 648 0.53
0.50 775 387 1.63 1,225 610 1.17 1,690 845 0.96
0.67 727 475 1.69 1,125 740 1.59 1,550 1,015 1,46
0.83 740 455 1.61 1,150 710 1.47 1,575 980 1.84
1.00 630 630 2.43 1,000 1,000 2.96 1,400 1,400 3.92
*Derived from equation 13.6.
Corn-Protein Substitution Relationships at Various Gain Levels
Corn-forage substitution relationships, where protein is held at

constant ratios and absolute amounts were discussed in the previous
section. Of similar interest are corn-protein substitution relationships
when forage is held constant at specified levels. Estimates of this type
are presented in Table 13.5 for a 200 -pound gain. It shows the pounds
of corn necessary to produce 200 pounds of gain when the L.O.M. levels
are allowed to vary to 100, 150, and 200 pounds. Simultaneously, the
forage levels are varied between 300, 500, and 700 pounds.
The feed inputs for the various gain levels are derived from iso-
quant equations based on function 13.1, The marginal rates of substitu-
te
tion, , are derived from the equations below, also based on equation
13.1 for forage held constant at the levels indicated:
Forage constant at 300 pounds.
6C _ 0.523502 - 0.001802P - 0.000036C

(13.9)
5P " 0.155491 - 0.000032C - 0.000036P
= 0.631791
- 0.001802P - 0.000036C
(13.10)
|| 0.154217 - 0.000032C - 0.000036P
dC _ 0.740078 - 0.001802P - 0.000036C

(13.11)
6P 0.152943 - 0.000032C - 0.000036P
At all gain levels the marginal rate of substitution of protein for

corn is predicted to decrease as the quantity of protein is increased,
Table 13.5. Feed Combinations and Marginal Rates of Corn-

Protein Substitution to Produce 200 Pounds of
Gain on Good-to- Choice Beef Calves
Weighing 400 Pounds at the Outset
Feed Combinations to Produce

200 Pounds of Gain 6C* 6 pt
6P 6C
Lbs. of Lbs. of Lbs. of
forage corn L.O.M.
300 902 100 2.59 0.39

300 791 150 1.86 0.54
300 715 200 1.16 0.86
500 721 100 3.40 0.42

500 769 150 2.67 0.37
500 455 200 1.99 0.50
700 580 100 4.18 0.24

700 390 150 3.43 0.29
700 235 200 2.74 0.36
Derived as the first derivative of C in respect to P from the basic

production function equation 13.1.
t 5C
1
Reciprocal of -r^r *
oP
with hay fixed at given levels. This prediction conforms with existing
nutrition logic and previous research, namely, additional pounds of pro-
tein substitute for less corn as the quantity of protein in beef rations in-
creases.
It is of particular interest to compare the feed substitution relation-
ships on increasingly higher gain isoquants when corn and protein are
held in a constant ratio to each other. Figure 13.5 indicates iso -gain
contours and protein to corn ratio lines of 15, 25, and 33 per cent, when
hay is constant at 700 pounds. Some of the ration lines extend past the
range of the data. In all cases, however, where consecutive points (with
regard to gain levels) on the same ration line can be compared, the
6C
slope (— ) , i.e., the marginal rate of substitution of L.O.M. for corn,
is greatest for the low gains and decreases for each successively higher
gain contour. In other words, for beef calves fed a fixed corn to protein
ration, with appropriate amounts of forage, each additional pound of
protein substitutes for less corn as the animal matures and puts on ad-
ditional gain. These empirical results are consistent with the physio-
logical needs of the animal. A younger animal has a greater relative
need for protein for building body tissue and growth; a more mature
animal requires relatively more carbohydrates for adding fat and finish.
The negatively sloped and curved lines are isoquants while the linear
and positively sloped lines are ration lines. If the marginal rate of sub-
stitution between corn and L.O.M. did not change with animal weight,
gain isoquants would have equal slopes where they are intersected
by a given ration line. However, as the graph illustrates, the slope of
PROTEIN (LB.) INTAKE PER ANIMAL

Figure 13.5. Beef gain contours and corn protein ration lines for good to choice
beef calves weighing 400 lbs. at the outset (forage held constant at
700 lbs.)
isoquants representing greater gains have less slope than those repre-
senting smaller gains; indicating a declining marginal rate of substitu-
5C
tion,
-£p ,
as the animal progresses in weight.
TOTAL AND MARGINAL GAIN FOR SELECTED FIXED RATIONS

The livestock gain surface is explored further in this section by de-
riving total and marginal gain estimated for feeder calves fed selected
fixed rations. It should be pointed out that no empirical observations
were available for calves fed a single ration over the entire feeding
program. Hence, the following estimates of feed-gain relationships
with fixed rations are interpolations of the derived live stock -feed sur-
face based on the available 272 observations. Until further empirical
verification has been completed from the newly designed experiments,
great practical importance is not attached to these estimates as the
basis for feeding recommendation. However, prediction from the cur-
rent data of feeder calf gain from different “fixed” rations provides
some insights into the relative efficiency of various rations.
Gain Equations for Given Ration Lines
Using the over-all gain equation, equation 13.1, it is possible to de-

rive total and marginal gain equations for various rations (feed compo-
nents, corn, L.O.M., and forage held in fixed proportions). For each
total gain equation, the procedure used was to define algebraically all
three feed variables in terms of a new composite variable, a . Then by
substituting the redefined variables into the over -all equation, equation
13.1, a sub-function in terms of Y
(gain) and the composite variable a
is derived.
Total and marginal gain equations, derived in the maimer outlined
above, are given in Table 13.6. Eight different rations were selected to
represent a cross section of the calf gain surface from which observa-
tions were available. Equation 13.12 in Table 13.6, for example, ex-
presses the relationship between feeder calf gain in pounds and pounds
of ration A (a a). One hundred pounds of ration A (a
A ) consists of 57
pounds of corn, 14 pounds L.O.M., and 29 pounds of forage. Equation
13.20 represents the marginal gain equation corresponding to the fixed
gain equation 13.12. Equation 13.20 is derived by taking the first deriv-
ative of Y a with respect to a A. In other words, it represents the
change in feeder calf gain with an infinitely small change in the quantity
of ration A (a A ) fed.
Tables 13.7 and 13.8 are the numerical counterparts of the equations
shown in Table 13.6. Table 13.7 shows predicted total feeder calf
gain
in pounds when the quantity of each ration is varied from 100 pounds
to
4,000 pounds. For example, 100 pounds of ration A ( ct A substituted
)
into equation 13.12 gives 20.9 pounds of gain (Y
A ) indicated in Table
13.7 under “ration A” (across from 100 “pounds of feed”). The
remain-
der of Table 13.7 is constructed similarly.
Table 13.8 includes predicted marginal gains for the several ra-
tions. Substituting 100 pounds of ration A (o-
A ), for example, into equa-
tion 13.12 yields .172 pounds of gain (Y
A ), shown in Table 13.8 under
a
Table 13.6. Predicted Total and Marginal Gam Equations for Selected Rations Fed Good- to- Choice
Feeder Calves Weighing 400 Pounds at the Outset
Feed Composition Prediction Equation for

per 100 Pounds
of Ration Total Gain Marginal gain
Ration A
Corn - 57 lbs. (13.12) Ya = .174768 oA - .00001040 oA 2 (13.20) dY.
= .174768
L.O.M. - 14 lbs. d« a
Forage - 29 lbs. + 3.587530 - .00002080 a A
Ration B 1
Corn - 54 lbs. (13.13) Yb « .171393o B - .00000746 a B 2 (13.21)

L.O.M. - 14 lbs. 4^
da B
= .171393
Forage - 32 lbs. + 3.587530 - .00001492 a B
Ration C
Corn - 50 lbs. (13.14) Yc = .166964« c - .00000457a c 2 (13.22)
4^- - .166964
L.O.M. - 12 lbs. d clq
Forage - 38 lbs. i
+ 3.587530 - .00000914 a c
Ration D i
Corn - 68 lbs. (13.15) Yd = .168657a D - .00001085<* D 2 (13.23)

4^° = .168657
L.O.M. - 10 lbs. dot D
Forage - 22 lbs. + 3.587530 - .00002170 aD
Ration E
Ye = .162974a E - .00000714a E 2
Corn - 61 lbs.
L.O.M. - 9 lbs.
(13.16) (13.24)
4^
da E
= .162974
Forage - 30 lbs. + 3.587530 - .00001428 of E
Ration F
2
Corn - 57 lbs. (13.17) Yf = .160079a F - .00000594 a F (13.25) 4Ê = .160079
L.O.M. - 8 lbs. daF
Forage - 35 lbs. + 3.587530 - .00001188 a F
Ration G
2
Corn - 53 lbs. (13.18) Yg = .156301a G - .00000532a G (13.26)
= .156301
L.O.M. - 8 lbs.
Forage - 39 lbs. + 3.587530 - .00001064a G
Ration H
2
Corn - 46 lbs. (13.19) Yh - .151264a H - .00000574a H (13.27)
Sh. = .151264
L.O.M. - 8 lbs. da H
Forage - 46 lbs. + 3.587530 - .00001148 H
“ration A” and across from 100 “pounds of feed.” This means that the
predicted additional beef gain on a feeder calf for the 100th pound of ra-
tion A (a A ) is .172 pounds (i.e., .172 pounds of gain for the 100th pound
of the specified ration). The other columns in Table 13.8 are con-
structed and interpreted similarly.
Feed Efficiency
In livestock feeding, the term “feed efficiency” has many meanings.

A common definition of feed efficiency is pounds of gain per pound of
ration. Using this definition, Table 13.7 would indicate that ration C
Table 13.7. Predicted Total Gains for Different Feed Combinations (Rations) With
Good-to-Choice Feeder Calves Weighing 400 Pounds at the Outset.
Pounds
Total Gain in Pounds for Corn- Protein-Forage Rations*
of
Feed (A) (B) (C) CD) (E) (F) (G) (H)
100 20.9 20.7 20.2 20.3 19.81 19.54 19.16 18.7

200 38.1 37.6 36.8 36.9 35.90 35.37 34.63 33.6
300 55.1 54.3 53.3 53.2 51.84 51.08 50.00 48.5
400 71.8 71.0 69.6 69.3 67.63 66.67 65.26 63.2
500 88.4 87.4 85.9 85.2 83.29 82.14 80.41 77.8
600 104.7 103.7 102.1 100.9 98.80 97.50 95.45 92.3
70 0 120.8 119.9 118.2 116.3 114.17 112.73 110.39 106.7
800 136.7 135.9 134.2 131.6 120.40 127.85 125.22 120.9
900 152.5 151.8 150.2 146.6 144,48 142.85 139.95 135.1
1,000 168.0 167.5 166.0 161.4 159.42 157.73 154.57 149.1 90 days
1,100 183.2 183.1 181.7 176.0 174.22 172.49 169.08 163.0
1,200 198.3 198.5 197.4 190.4 188.87 187.13 183.49 176.8
1,300 213.2 213.8 212.9 204.5 203.39 201.65 197.79 190.5
1,400 227.9 228.9 228.4 218.4 217.76 216.06 211.98 204.1
1,500 242.3 243.9 243.8 232.2 231.98 230.34 226.07 217.6
1,600 256.6 258.7 259.0 245.7 246.07 244.51 240.05 230.9
1,700 270.6 273.4 274.2 258.9 260.01 258.55 253.92 244.1
1,800 284.5 287.9 289.3 272.0 273.81 272.48 267.69 257.3
1,900 298.1 302.3 304.3 284.9 287.46 286.29 281.35 270.3 180 days
2,000 311.5 316.5 319.2 297.5 300.98 299.98 294.91 283.2
2,100 324.7 330.6 334.1 309.9 314.34 313.56 308.36 295.9
2,200 337.7 344.5 348.8 322.1 327.57 327.01 321.70 308.6
2,300 350.5 358.3 363.4 334.1 340.66 340.35 334.94 321.1
2,400 363.1 372.0 378.0 345.9 353.60 353.56 348.07 333.6
2,500 375.5 385.4 392.4 357.4 366.40 366.66 361.09 345.9
2,600 387.7 398.8 406.8 368.7 379.05 379.64 374.01 358.1
2,700 399.6 412.0 421.1 379.9 391.57 392.50 386.82 370.2
2,800 411.4 425.0 435.3 390.8 403.94 405.24 399.52 382.1 270 days
2,900 423.0 437.9 449.3 401.4 416.16 417.86 412.12 349.0
3,000 434.3 450.6 463.3 411.9 428.25 430.36 424.61 405.7
3,100 445.4 463.2 477.3 422.2 440.19 442.75 437.00 417.3
3,200 456.3 475.7 491.1 432.2 451.99 455.01 449.27 428.9
3,300 467.1 487.9 504.8 442.0 463.65 467.16 461.45 440.3
[
3,400 477.6 500.1 518.4 451.6 475.16 479.19 473.51 451.5

3,500 487.9 512.1 532.0 461.0 486.53 491.10 485.47 462.7
|
3,600 498.0 523.9 545.4 470.1 497.76 502.89 497.32 473.7

3,700 507.9 535.6 558.8 479.1 508.84 514.56 509.07 484.7 360 days
3,800 517.5 547.2 572.1 487.8 519.79 526.11 520.71 495.5
3,900 527.0 558.6 585,2 496.3 530.59 537.55 532.24 506.2
4,000 536.3 569.8 598.3 504.6 1 541.24 548.86 543.67 516.8
*See Table 13.6 for total gain predicting equations and composition of the rations.
was the most efficient ration and ration D the least efficient. That is,
the predicted gain for feeder calves fed 4,000 pounds of ration C is 598
pounds compared to 505 pounds for calves fed ration D. Likewise, the
marginal gain for the 4,000th pound of ration C is .130 pounds compared
to .081 pounds for the 4,000th pound of ration D. This comparison of
“efficiency” of two rations is not entirely adequate. First, the gain
Table 13.8. Predicted Marginal Gains for Different Feed Combinations (Rations) With
Good-to-Choice Feeder Calves Weighing 400 Pounds at the Outset
Marginal Gain in Pounds for Corn- Protein- For age Rations*
Feed (A) (B) (C) (D) (E) CF) (G) (H)
100 0.172 0.169 0.166 0.166 0.162 0.159 0.155 0.150

200 0.170 0.168 0.165 0.164 0.160 0.158 0.154 0.148
300 0.168 0.166 0.164 0.162 0.159 0.157 0.153 0.147
400 0.166 0.165 0.163 0.159 0.157 0.155 0.152 0.146
500 0.164 0.163 0.162 0.157 0.156 0.154 0.151 0.145
600 0.162 0.162 0.161 0.155 0.154 0.153 0.150 0.144
700 0.160 0.160 0.160 0.153 0.153 0.152 0.149 0.143
800 0.158 0.159 0.159 0.151 0.152 0.151 0.148 0.142
900 0.156 0.157 0.158 0.149 0.150 0.149 0.147 0,140
1,000 0.153 0.156 0.158 0.146 0.149 0.148 0.146 0.139
1,100 0.151 0.154 0.156 0.144 0.147 0.147 0.145 0.138
1,200 0.149 0.153 0.155 0.142 0.146 0.146 0.144 0.137
1,300 0.147 0.152 0.155 0.140 0.144 0.145 0.142 0.136
1,400 0.145 0.150 0.154 0.138 0.143 0.143 0.141 0.135 200 lbs.
1,500 0.143 0.149 0.153 0.136 0.142 0.142 0.140 0.134
1,600 0.141 0.147 0.152 0.133 0.140 0.141 0.139 0.132
1,700 0.139 0.146 0.151 0.131 0.139 0.140 0.138 0.131
1,800 0.137 0.144 0.150 0.129 0.137 0.139 0.137 0.130
1,900 0.135 0.143 0.149 0.127 0.136 0.138 0.136 0.129
2,000 0.133 0.141 0.148 |
0.125 0.134 0.136 0.135 0.128
2,100 0.131 0.140 0.147 0.123 0.133 0.135 1
1
0.134 0.127 300 lbs.
2,200 0.129 0.138 0.146 0.120 0.132 0.134 0.133 0.126
2,300 0.126 0.137 0.145 0.118 0.130 0.133 0.132 0.124
2,400 0.124 0.135 0.145 0.116 0.129 0.132 0.131 0.123
2,500 0.122 0.134 0.144 0.114 0.127 0.130 0.130 0.122
2,600 0.120 0.132 0.143 0.112 0.126 0.129 0.129 0.121
2,700 0.118 ||
0.131 0.142 0.110 0.124 0.128 0.128 0.120
2,800 0.116 0.129 0.141 0.107 0.123 0.127 0.127 0.119
2,900 0.114 0.128 0.140 0.105 1
0.122 0.126 0.125 0.117
3,000 0.112 0.126 0.139 0.103 0.120 0.124 0.124 0.116 400 lbs.
3,100 0.110 0.125 0.138 0.101 0.119 0.123 0.123 0.115
3,200 0.108 0.123 0.137 0.099 0.117 0.122 0.122 0.114
3,300 0.106 0.122 0.137 0.097 0.116 0.121 0.121 0.113
'
3,400 0.104 0.120 ] 0.135 0.094 0.114 0.120 0.120 0.112

3,500 0.101 0.119 0.134 0.092 0.113 0.118 0.119 0.111
3,600 0.099 0.117 0.134 0.091 0.112 0.117 0.118 0.110
3,700 0.097 0.116 0.133 0.088 0.110 0.116 0.117 1 0.109
3,800 0.095 0.114 0.132 0.086 0.109 0.115 0.116 |
0.108 500 lbs.
3,900 0.093 0.113 0.131 0.084 0.107 0.114 0.115 0.106
4,000 0.091 0.111 0.130 0.081 0.106 0.113 0.114 0.105
*See Table 13.6 for marginal gain predicting equations.
produced by these two rations is, as any cattle feeder knows, not identi-
cal in quality. Second, the time variable, which is also correlated with
quality, confounds the comparison (i.e., it takes less time for a feeder
calf to consume 100 pounds of a high concentrate ration than a bulky
forage ration). As discussed previously, estimates of beef grades were
not provided in data upon which this study is based. However, estimates
of time required by the feeder calf to consume various quantities of

feed were made. Accordingly, the "horizontal” lines in Table 13.7 in-
dicate the estimated time required by the feeder calf to consume vari-
8
ous quantities of the fixed rations. For example, 90 days is the time
predicted for consuming 1,400 pounds of ration D (68 per cent corn, 10
per cent L.O.M., and 22 per cent forage). The corresponding gain is
218 pounds. For an equal time period, 1,000 pounds of ration H (46 per
cent corn, 8 per cent L.O.M., and 46 per cent forage) produces a gain of
149 pounds. Thus, in the same time period; the predicted gain is 69
pounds greater for the heavy concentrate ration compared to the bulky
forage ration. However, the feeder animal fed the concentrate ration
(ration D) is predicted to consume an additional 400 pounds of feed.
TIME FACTOR IN BEEF GAINS
Previous sections have indicated how given beef gains can be pro-
duced with varying combinations of corn, L.O.M., and hay. However,
as mentioned above, different proportions of these three feeds in a ra-
tion affects the rate at which the beef calf gains weight. Hence, rations
that minimize costs may not necessarily coincide with the ration that
produces the most rapid gains. As cattle feeders are well aware, time
of marketing may affect profits as much or more than feed costs.
Time Function
To allow analysis of these considerations, a quadratic function was

used to express the relationship between corn (C), L.O.M. (P), forage
(F), and the number of days (T) required to consume various quantities
of these feeds. The estimated relationship is indicated by equation
13.28 where
(13.28) T = .039328C + .076652P + .134446F - .00000130C 2

2 2
+ .Q0019013P + .0QQ01578F - .00008414CP
+ .00000568CF + .00001115PF + 3.095509
T is time in days for the feeder calf to consume a given quantity of the three
feeds, starting from the beginning of the feeding period. Consecutive obser-
vations were taken at 28-day intervals on a given ration line, with each ob-
servation being the accumulated sum of previous time intervals. For exam-
ple, T = 28 days for the first observation, with the second observation being
the sum of the first time interval plus the second time interval or 56 days,
etc.
C is total intake of corn in pounds from the time grain feeding is started to the
particular weighing date (each weighing period is separated by a 28-day time
Equation 13.28 in the following section provides the basis for these time estimates.
.
interval). Hence, the first observation of corn is from the beginning of the
experiment to the first weighing date (28 days), the second corn observation
is from the beginning of the experiment to the second weighing date (56 days),
etc.
P and F are total intake of L.O.M. and forage measured in pounds. Observations
are measured as outlined for corn.
The coefficient of determination for equation 13.28 is .99, indicating

that the three feed variables explain a major proportion of variation in
the dependent variable T. Standard errors of the regression coeffi-
cients for equation 13.28 are given in Table 13.9.
Table 13.9. Standard Errors for Regression

Coefficients in Equation 13.28
for Standard Error
C .00392100
P .01790900
p .00922700
C2 .00000377
P2 .00004390
F2 .00000345
CP .00001669
CF .00000591
PF .00002372
Time Equations for Specific Rations
Equation 13.28, the over -all time function, was reduced to several
individual time equations; one for each of the eight different feed com-
binations or rations discussed in the previous section. These time
equations are shown in Table 13.10. Estimates of the total time re-
quired to consume quantities of feed for the eight different rations are
shown in Table 13.11.
Table 13.11 shows, as previous research would suggest, that the
amount of time required to consume a "bulky" roughage ration is
greater than that required to consume the same amount of a higher con-
centrate ration. More specifically, calves require 392 days to consume
4,000 pounds of ration H (1:1 corn -forage ratio). Calves fed ration D
(3:1 corn-forage ratio) require only 221 days to consume the same
quantity of feed. The quantity of L.O.M. in the ration appears to have
little effect on the time required for the calf to consume the ration.
Rations A and E both have approximately a 2:1 corn-forage ratio. How-
ever, ration A contains 14 pounds of L.O.M., as compared to 9 pounds
for ration E, per hundred pounds of total feed (see Table 13.10). The
predicted time for feeder calves to consume 4,000 pounds of ration A is
281 days, about the same number of days required for feeder calves to
consume an equal quantity of ration E
Table 13.10. Predicted Time Equations for Good-to-Choice Feeder Calves to

Consume Selected Rations Indicated in Table 13.6
2
Ration A (13.29) ta = .072138 ot
A
- .00000069 a
A + 3.095509
Ration B (13.30) tb = .074991 cu
B + .00000008 a B2 + 3.095509
2
Ration C (13.31) Tc = .079663a
c + .00000119 a c + 3.095509
Ration D (13.32) td + .063986 a
D - .00000256 a
d
2
+ 3.095509
Ration E (13.33) te = .071223 a

E - .00000080 a 2
E + 3.095509
Ration F (13.34) tf = .074261 a F + .00000012 a f2 + 3.095509
Ration G (13.35) tg = .079410 a
G + .00000121 a G2 + 3.095509
Ration H (13.36) th = ,086068a .00000279 a 2
+ 3.095509
H+ h
Least -Cost Rations
The methods used previous chapters could be used to predict

in
least -cost rations and optimum marketing weights. However, these
steps are not taken here because of space limitations and because new
3-year experiments are being conducted to allow a better basis for such
estimates.
APPENDIX: Reduced Gain Equation for Given Ration Line
The following notes explain the method of deriving the quantities in

tables 13.6 to 13.8 from equation 13.1.
With a feed combination of X per cent corn, K per cent L.O.M., and
Z per cent forage, a new variable (a) is defined with the following con-
ditions:
Corn (C) = X i
a i
Protein (L.O.M.) = Kî

Forage (F) = .
Substituting the above conditions into equation 13.1 gives the general
form of the reduced equation in terms of Y* and
(13.37) Yj = .157402(X i )a j+ .361070(K i )a + .112332(21)0!!

i
2 2 2 2 2
- .00001612(Xi)Q! i - .00090147(K i )Q! i - .00005958(Z )o!i
- .OOOOSe^XiKOoi2 - .00000637(x i z i )o! i
2
+ .000541 44(KiZi) 2 o'!

2
+ 3.58753
Table 13.11. Predicted Total Time Required for Good-to~Choice Feeder Calves,
Weighing 400 Pounds at the Outset, to Consume Various Fixed
Rations
Pounds
Total Days Required To Feed Quantities of Rations*
of
Feed (A) (B) (C) (D) (E) (F) (G) (H)
100 10.3 10.6 11.1 9.5 10.2 10.6 11.0 11.7

200 17.5 18.1 19.1 15.9 17.3 18.1 19.0 20.4
300 24.7 25.6 27.1 22.2 24.4 25.6 27.0 29.2
400 31.8 33.1 35.2 28.4 31.5 33.1 35.1 38.0
500 39.0 40.6 43.2 34.6 38.5 40.5 43.1 46.8
600 46.1 48.1 51.3 40.8 45.5 48.0 51.2 55.7
700 53.3 55.6 59.4 46.9 52.6 55.5 59.3 64.7
800 60.4 63.1 67.6 53.0 59.6 63.0 67.4 73.7
900 67.5 70.7 75.8 59.0 66.5 70.4 75.5 82.8
1,000 74.5 78.2 84.0 65.0 73.5 77.9 83.7 92.0
1,100 81.6 85.7 92.2 70.9 80.5 85.4 91.9 101.1
1,200 88.7 93.2 100.4 76.7 87.4 92.8 100.1 110.4
1,300 95.7 100.7 108.7 |
82.6 94.3 100.3 108.4 119.7
1,400 102.7 108.2 117.0 88.4 101.2 107.7 116.6 |
1
126.0 200 lb.
1,500 109.8 115.8 125.3 94.1 108.1 115.2 124.9 139.5 gain
1,600 116.8 123.3 133.6 99.7 115.0 122.6 133.2 147.9
1,700 123.7 131.7 142.0 105.4 121.9 130.0 141.6 157,5
1,800 130.7 138.3 150.4 110.9 128.7 137.5 150.0 167.1
1,900 137.7 145.9 158.8 116.4 135.5 144.9 158.3 176.7
2,000 144.6 153.4 167.2 |
121.9 142.3 152.4 166.8 186.4
|
2,100 151.5 160.9 175.7 127.4 149.1 159.8 1

175.2 196.1 300 Xb.
2,200 158.5 168.5 184.1 132.7 155.9 167.2 183.7 205.9 gain
2,300 165.4 176.0 192.6 138.0 162.7 174.6 192.1 215.8
2,400 172.3 183.5 201.2 143.3 169.4 182.1 200.6 225.7
2,500 179.1 191.1 209.7 148.6 176.2 189.5 209.2 235.7
2,600 186.0 198.6 218.3 153.7 182.9 196.9 217.7 245.7
2,700 192.8 ]
206.2 226.9 158.8 189.6 204.3 226.3 255.8
|
2,800 199.7 213.7 235.5 163.9 196.2 211.7 234.9 266.0

2,900 206.5 221.2 241.4 170.0 |
202.9 219.1 243.6 276.2
3,000 213.3 228.8 252.8 173.9 209.6 226.5 252.2 286.4 400 lb.
3,100 220.1 236.3 261.5 178.9 216.2 233.9 260.9 296.7 gain
3,200 226.9 243.9 270.2 183.8 222.8 241.3 269.6 307.1
3,300 233.6 251.4 279.0 188.6 229.4 248.7 278.3 317.5
3,400 240.4 259.0 I 287.7 193.4 236.0 256.1 287.1 328.0
3,500 247.1 266.5 296.5 198.1 242.6 263.5 295.9 338.5
3,600 253.8 274.1 305.3 202.8 249.1 270.9 304.7 349.1
3,700 260.6 281.7 314.2 207.4 255.7 278.2 313.5 359.7
[
3,800 267.3 289.2 323.0 212.0 262.2 285.6 322.3 1 370.4 500 lb.
3,900 273.9 296.8 331.9 216.5 268.7 292.9 331.2 381.2 gain
4,000 280.6 304.3 340.8 221.0 275.2 300.4 340.1 392.0
See Table 13.10 for prediction equation corresponding to each ration.

Simplifying 13.37 yields:
2
(13.38) Yj = a i (.1574G2X i + .361070Ki+ .112332Zi) + a (_.00001612Xi 2
2 2
- .00090147Ki - .00005958Z - .00003612XiKi
- .00000637X jZ i + .00054144KiZi) + 3.58753 .
For example, with 100 pounds of ration (i.e., feeds combined in fixed
proportions) composed of 57 pounds of corn, 14 pounds of L.O.M., and
28 pounds of forage, X = .57, K = .14, and Z = .28.
Substituting the above values for X, K, and Z into equation 13.38
gives:
(13.39) Ya = .174768a A - .00001040a A + 3.587530 .
In addition, marginal gain was predicted for each ration by taking

the total derivative of Y ^ (gain in pounds) with respect to a i (feed vari-
able with components held in fixed proportions). For predicted gain
equation 13.39 the marginal gain equation becomes:
dY A
(13.40) = .174768 - .00002080a A
da A .
Similarly, total and marginal gain equations were derived for seven
other rations (see tables 13.6 and 13.7) with constant proportions of
corn, protein, and forage.
Recent Data
Preliminary results of an experiment conducted with 300 steers

over the period 1956-59 are similar to those presented above. Two
major inputs, pasture forage (F) in dry weight and corn supplemented
with protein (C), were fed along six ration lines. One resulting function,
computed through the origin, is as follows:
(13.41) Y = .1246C + .0231F - .00001240 2
(.0127) (.0019) (.000005)

2
- .00000111F -00000246CF - 1.3631T
(.0000002) (.0000015) (.4033)
where Y is
gain in pounds and T is temperature above the summer
mean. R2 Isoquants again indicate that large anima ls readily
is .995.
substitute one feed for another. While details will be explained else-
where, it is interesting to note that temperature is a significant ex-
planatory variable, gains being less as temperature rises.
CHAPTER 14
Ear! O. Heady
Crop Response Surfaces
John T. Pesek
William G. Brown and Economic Optima
John P. Doll
in Fertilizer Use
CHAPTER
T HIS
initiated in
reports initial research on crop response surfaces
in 1952. Since the original experiment in 1952,
Iowa
many additional studies employing several types of designs have
been completed for different crops at various locations in the state.
Results from a few of these experiments are reported in Chapters 14
and 15 to illustrate (a) the nature of estimated fertilizer response sur-
faces, (b) the nature of isoclines and other relationships derived for
various soil and moisture conditions by different algebraic functions,
and (c) the optimum quantities and use of fertilizer as specified by par-
ticular analysis of response. Experiments which have been completed
include those with two of the nutrients N and K 2 0 and P 2 Os variable,
with all three variable, and with one or more nutrients variable while
stand, moisture, or certain other inputs also are variable. Only a few
examples from the many experiments will be discussed in these two
chapters.
Fertilizer is an important area for production function research.
Greater use of fertilizer has been one of the important developments in
agriculture over the past decades. It is an input which is clearly divisi-
ble and for which profit -maximizing principles, for either limited or
unlimited capital situations, can easily be applied. Because of space
limitations, however, only a select set of the economic principles dis-
cussed in Chapter 2 have been applied to the several sets of quantitative
results from production function studies.
In conformity with accepted economic terminology, nutrient combi-
nations can be expressed in terms of their substitution or replacement
rates. In the chemical processes of the plant one element may not sub-
stitute for another; however, it is true that moderate yield increases
may be attained with several combinations of elements. A farmer may
obtain a 5 -bushel increase in corn from use of ammonium nitrate alone,
from phosphate alone, or from a mixed fertilizer such as 20-20-0 or ,
8-8-8. If all of the mixtures give the 5 -bushel increase, they can be
looked upon as substitutes for each other in attaining the given yield
even though physiological substitution does not actually take place. Ele-
ments Na and K may be real substitutes over wide ranges in the chemi-
cal processes of some plants. However, even though plant nutrients
such as N, P, or K do not directly serve as substitutes in the chemical
475
476 RESPONSE SURFACES IN FERTILIZER USE
functions of the plant, the fact that similar yield increases can be at-
tained with different combinations of nutrients causes them to serve as
substitutes in the decision-making framework of the farmer. Within
limits, he can use more of one nutrient and less of another in attaining
yield increases under many soil situations. While the terms "substitu-
tion” or "replacement rates” thus may not represent an entirely accu-
rate physiological concept, they are employed in the remainder of this
study in the absence of more appropriate terms. While substitution is
discussed in forthcoming sections, the biological exceptions mentioned
above should be kept in mind. From the standpoint of fertilizer ratios,
the problem is perhaps as much one of finding "optimum combinations
of nutrients” (least-cost combinations for a given yield) as in determin-
ing "substitution” rates.
INITIAL EXPERIMENTS
The experiments reported in this chapter were conducted in 1952
with corn on calcareous Ida silt loam in western Iowa and with alfalfa
and red clover on Webster and Nicollet loam in north central Iowa.
Major emphasis will be given to corn on Ida silt loam. Two variable
nutrients were used on each experiment. Nitrogen in the form of am-
monium nitrate and P 2 O s in the form of concentrated superphosphate
were applied to corn while K 2 0 in the form of potassium chloride and
P 2 0 5 in the form of concentrated superphosphate were applied to both
alfalfa and red clover. Observations were obtained from an incomplete
factorial experimental design of the nature indicated by Table 14.1.
The same design was used for alfalfa and red clover except that the
second variable nutrient was K 2 0. With replication, there were 114 ob-
servations for each of the three experiments. 1 Two cuttings were
Table 14.1. Design of Experiment for Corn; Each “X”

Represents an Experimental Plot
Pounds
^2 05 Pounds Nitrogen per Acre
per Acre
0 40 80 120 160 200 240 280 320

0 XX XX XX XX XX XX XX XX XX
40 XX XX XX XX XX
80 XX XX XX XX XX
120 XX XX XX XX XX
200 XX XX XX XX XX
240 XX XX XX XX XX
280 XX XX XX XX XX
1
The treatments were assigned at random (completely randomized block design).
RESPONSE SURFACES 3N FERTILIZER USE 477
obtained from both the alfalfa and red clover. Yield measurements for
hay were in terms of 12 per cent moisture. This design, with random-
ized plots, allows continuous observations at the extremes of applica-
tion rates with combinations of the various nutrients. It also provides
sufficient observations over other points of the production surface for
estimation of the two -variable nutrient function. In the experiments, all
resources or inputs except fertilizer were held constant except for the
variable quantities of labor and machine services for application and
harvesting; seeding rates were constant.
Weather in Experimental Year
The 1952 growing year was one favorable for the use of fertilizer.
The spring was fairly cool and wet. Rainfall was ample to mid-August
when a 2-month drought began. For these reasons, the experimental
data do not necessarily serve as a basis for inference to average years.
On corn the constant plant population of 18,000 plants per acre for all
treatments may have limited the response obtained from heavy fertili-
zation rates.
We again point out that fertility nutrients may not substitute in the
biological processes involved in producing a given amount of a specified
part of the plant. However, they do serve as substitute means of attain-
ing specified yield responses. With corn, for example, an average yield
of 24.8 bushels per acre was obtained on the plots receiving 120 pounds
of nitrogen and no P 2 Os
. The plots receiving 40 pounds of P 2 Os aver-
aged 28.6 bushels. With slightly fewer pounds of phosphoric acid, equal
increments in yield might have been attained with entirely different nu-
trient combinations of nitrogen and P 2 G 5 With 160 pounds of N and
.
40 pounds of P2 0 5 , the plots averaged 101.5 bushels; with 240 pounds of

P 2 05 and 80 pounds of N, the average was 102.5 bushels. Similarly, for
clover, the plots receiving 120 pounds of P 2 0 5 and 160 pounds of K 20
averaged 3.66 tons while those receiving 160 pounds of P 2 05 and 80
pounds of K 2 0 averaged 3.68 tons. Thus, while the nutrients may not
serve as substitutes in the chemical process of the plants, they do
serve as substitute means of attaining given yield increases. These are
the kind of data needed in farmer decision-making; it is the cost of pro-
ducing a given yield, rather than the chemical process itself, which di-
rectly concerns him.
While nutrients may serve as substitutes over a limited range in at-
taining given levels of crop response, the data also show how they even-
tually serve as technical complements as one is increased alone. By
technical complementarity, we refer to the situation where an increase
in one element without an increase in the other either (a) does not add
anything to total yield or (b) actually decreases total yield. On corn,
for example, any increase in N alone (a path followed horizontally from
left to right in Table 14.1 with P2 0 5 held “fixed” at any level in the
table) causes first an increase and then a decrease in total yield. The
same situation holds true for P 2 O s That is, yield increases and then
.
decreases down the column of the table with N fixed at specified levels,
and this decrease indicates that N also is a limitational nutrient with
P 2 05 That the two nutrients serve as limitational resources or tech-
.
nical complements to each other also was illustrated by the fact that
yields were taken to successively higher levels with diagonal move-
ments from northwest to southeast in the table; under this “movement”
over the cells and columns of the table, the two elements are, in effect,
increased simultaneously and in fixed proportions.
DERIVATION OF PRODUCTION OR YIELD FUNCTIONS
After collection of yield observations, the next step was that of de-
riving production functions, input-output or response coefficients.
Since, at the time of this experiment, practically no empirical analysis
had been made of fertilizer production surfaces, numerous functional
forms were fitted to the data. For surface estimates, forms employed
included the Cobb-Douglas and general polynomial forms, all with lin-
ear terms but variously using exponents representing powers of .5, 1.5,
and 2. The first will be termed a square root equation; the second, a
“1.5” or “3/2” function, and the latter, a quadratic. Some estimates
were made for individual nutrients where Spillman types of functions
also were fitted to the data. In the equations which follow, Y refers to
total bushels or tons per acre for corn or hay respectively, Y' refers to
yield above check plot, while N, P, and K refer, respectively, to pounds
per acre of nitrogen, P2 O s and K 2 Q.
,
PRODUCTION FUNCTIONS FOR CORN
The variables for corn on Ida silt loam were nitrogen and P2 0 5 .
The experimental field, operated by a farmer in the usual manner ex-

cept for fertilization, was very low in fertility. The four production
surface equations estimated for corn are:
4080 ,2S77
(14.1) Y' = .442P' N
(14.2) Y = -7.51 + .584N + .664P - .0016N 2 - .0018P 2 + 00081NP
(14.3) Y = -5.68 - .316N - .417P + 6.3512N’ 5 + 8.5155P-* + -5

.3410N P-
5
1 '5 1 -5
(14.4) Y = -13.62 + .984N + 1.129P - .0500N - .0576P + .0008NP .
The R and t values for the four functions are shown in Table 14.2.
All of these are significant at a one per cent level of probability.
RESPONSE SURFACES 3N FERTILIZER USE 479
Table 14.2. Values of R for Two-Variable Nutrients and Values of t for

Individual Regression Coefficients
Equation Value of R Value of t for Coefficients in Order Listed in Equations
14.1 .9255* 18.62* 15.23* ....
14.2 .9122* 9.21* 10.46* 5.24* 8.96* 10.19*

14.3 .9582* 7.91* 10.44* 7.32* 9.81* 8.85*
14,4 .9434* 8.21* 10.35* 6.21* 7.31* 10.12*
*0 < P< 0.01.
Analyses of regression for three functions are shown in Table 14.3.

Data are not included for the Cobb -Douglas function since it was not
fitted to all yield observations. It was fitted by subtracting the average
yield of check plots from the yield of plots with fertilizer above zero
level. Observations for yields smaller than check plot levels were thus
omitted and the R for Cobb-Douglas, .86, based on the total sums of
2
squares associated with the 105 observations to which the function was
fitted cannot be compared with those of other equations. The propor-
2
tions of variance in corn yield, R explained by variables in the other
,
three equations are .86 for equation 14.2, .91 for equation 14.3, and .89
for equation 14.4. The percentage of treatment sums of squares ex-
plained by regression equations decreases as the power of the equation
increases. The mean squares for derivations from regressions were
625, 215, and 495, respectively, for these three equations. Using these
and other statistics, it was decided that square root equation 14.3
served most efficiently in predicting the corn production surface. All
three of the last mentioned equations have negative yield intercepts,
with the largest negative value being for the equation with the 1.5
powers on N and P. Since the constants are statistically significant,
Table 14.3. Analyses of Variance of Regression for Equations Fitted
to All Yield Observations, Ida Silt Loam, 1952
Degrees of Sum of Mean

Source of Variation Freedom Squares Squares
Total 113 242,707 4,175

Treatments 56 233,811
Due to regression of
quadratic equation 14.2 ( 5 201,943 40,389
(
Deviations from regression (51 31,868 625

square root equation 14.3 m 222,828 44,566
v
t—
Deviations from regression in.

i 10,983 215
“3/2* equation 14.4 ( 5 208,582 41,716
(
Deviations from regression (51 25,229 495

Among plots treated alike 57 8,896 156
they were retained in the equations, although the functions might have
appeared more sensible if they had been forced to have zero intercepts.
However, the magnitude of the constant does not affect marginal pro-
ductivities and the recommendations which arise from curves which
might have different constants but the same regression coefficients.
The analysis relating to optima which follows is presented accordingly.
However, in a later section, we analyze regression equations based
only on non -zero fertilizer treatments.
Although the square root functions were selected as most appropri-
ate for specifying economic optima in this section, we do turn to some
comparisons of estimates made by the four equations. Specification of
economic optima then follow in a later section.
COMPARISON OF PREDICTED YIELDS

FROM COMPLETE PRODUCTION SURFACES FOR CORN
Corn yields predicted from equations 14.1 to 14.4 are included in
Table 14.4. Production surfaces representing these predicted yields
geometrically in three dimensions are presented in figures 14.1 through
14.4. Predictions of the equations differ considerably. The quadratic
and “1.5” functions are most nearly alike. The square root and Cobb-
Douglas functions differ most.
The surface for the square root function rises more rapidly than for
RESPONSE SURFACES IN FERTILIZER USE 481
Table 14.4. Bushels of Corn per Acre Predicted by Four Types of Production
Functions for, Specified Nitrogen and P2 05 Applications,
Ida Silt Loam, 1952
Pounds of Nitrogen per Acre

2 '“'5
per Acre 0 40 80 120 160 200 240 280 320
Cobb-Douglas function 14.1
0 15.4 18.5 19.2 19.7 20.0 20.4 20.6 20.9 21.1

40 21.8 51.5 59.5 65.0 69.2 72.8 75.9 78.6 81.1
80 24.0 63.4 74.0 81.2 86.9 91.6 95.7 99.4 102.6
120 25.5 72.0 84.5 93.1 99.9 105.4 110.2 114.5 118.4
160 26.8 79.1 93.1 102.8 110.3 116.6 122.0 126.9 131.2
200 27.8 85.2 100.6 111.1 119.4 126.3 132.2 137.5 142.3
240 28.8 90.6 107.2 118.5 127.5 134.9 141.3 147.0 152.1
280 29.7 95.5 113.1 125.2 134.7 142.6 149.5 155.5 161.0
320 30.5 100.0 118.6 131.4 141.4 149.8 157.0 163.4 169.2
Quadratic function 14.2
0 -7.5 13.3 29.2 39.8 45.5 46.1 41.7 32.2 17.6

40 16.2 38.3 55.4 67.4 74.4 76.3 73.1 64.9 51.6
80 34.1 57.5 75.9 89.2 97.5 100.7 98.8 91.9 79.9
120 46.3 71.0 90.7 105.3 114.9 119.4 118.8 113.2 102.5
160 53.0 78.7 99.8 115.6 126.6 132.3 133.0 128.7 119.3
200 53.4 80.7 103.0 120.2 132.3 139.4 141.5 138.4 130.4
240 48.3 76.9 100.5 119.0 132.5 140.8 144.1 142.4 135.7
280 37.5 67.4 92.3 112.0 126.8 136.5 141.1 140.7 135.2
320 20,9 52.1 78.3 99.4 115.4 126.4 132.3 133.2 129.0
Square root function 14.3
0 -5.7 21.8 25.8 25.9 24.0 20.9 16.8 12.1 6.8

40 31.5 72.6 82.3 88.7 88.5 88.6 87,4 85.3 82.5
80 37.1 83.9 95.9 102.1 105.4 106.8 106.9 105.9 104.1
120 37.5 88.7 102.4 110.1 114.5 116.9 117.9 117.8 116.8
160 35.3 90.1 105.4 114.2 119.6 122.9 124.6 125.2 124.9
200 31.6 89.3 105.9 115.7 122.0 126.1 128.5 129.7 130.0
240 26.1 87.0 104.8 115.6 122.6 127.4 130.4 132.2 133.0
280 19.9 83.6 102.5 114.1 121.9 127.2 130.8 133.2 134.5
320 13.1 79.2 99.2 111.5 120.0 126.0 130.1 132.9 134.7
“1.5* power function 14.4
0 -13.6 13.1 29.3 38.8 42.7 41.9 36.8 27.8 15.3

40 16.9 44.9 62.5 73.2 78.4 78.9 75.1 67.4 56.2
80 35.4 64.7 83.6 95.6 102.1 103.8 101.3 95.0 85.0
120 46.0 76.6 96.8 110.1 117.9 120.9 119.7 114.6 106.0
160 50.3 82.2 103.6 118.2 127.3 131.6 131.7 127.9 120.6
200 49.1 82.3 105.0 120.9 131.3 136.9 138.3 135.8 129.7
240 42.9 77.4 101.4 118.6 130.3 137.2 139.9 138.7 133.9
280 32.3 68.1 93.4 111.9 124.9 133.1 137.0 137.1 133.6
320 17.6 54.7 81.3 101.0 115.3 124.8 130.1 131.4 129.2
the other functions when nutrient applications are small. But as input
levels increase the square root surface has less slope than those of the
other functions. When either of the nutrients is at a zero rate, so that
interactions are not present, the square root function predicts yields
Figure 14.3. Production surface for corn predicted by square root function 14.3,
Ida silt loam, 1952
which are of a smaller magnitude. It has a surface which attains a

maximum at lower input levels than other functions. When interactions
are present, the square root function again predicts maximums, in
reading across a row or down a column of the tables, at lower yield
levels but at higher input levels.
The quadratic and “1.5” functions have lower yield intercept values
than the square root function. While yields from the square root in-
crease at initially high rates but decrease rapidly as the maximum
yield is approached, these two functions increase to the maximum at
fairly uniform rates. As explained in Chapter 3, the quadratic has a
linear marginal product equation, indicating a constant deceleration.
These two functions predict higher maximum yields at lower input
levels than the square root function. The quadratic function predicts a
maximum yield at input and yield levels slightly higher than the “ 1 5 ”
,
function.
Predictions from the Cobb-Douglas function are subject to constant
elasticity and thus do not attain a maximum. As mentioned previously,
this function is not entirely comparable with the other equations. When
both nutrients are present, predictions of the Cobb-Douglas function
increase relative to the other functions until they surpass all other pre-
dictions at the highest input rates. For example, at 320-pound applica-
tion rates for nitrogen and P2 O5 the yield prediction for the Cobb-
,
Douglas function is 169.2 bushels per acre, comparable figures for the
square root, quadratic, and “1.5” functions are 134.7, 129.0, and 129.2
bushels, respectively.
To show these differences more clearly, some curves of yield

re
sponses, with one input constant, are presented in
figures 14.5, 14.6,
and 14.7, remembering the Cobb-Douglas is based on Y'.
Marginal Physical Products and Yield Maxima
Functions 14.5 through 14.8 are marginal physical

product equations
for nitrogen, corresponding to equations 14.1
through 14.4. Marginal
physical product equations, derivatives of production
functions, predict
the slope or rate of change of yield predictions
of the functions for any
rate of nutrient input. Interactions between
the nutrients cause all of
the marginal product equations to include a
P2 O s term. Hence, the
magnitude of the marginal product of one nutrient is
dependent’upon the
Figure 14.6. Corn yield response to nitrogen when P2 0 5 is constant

at 320 pounds per acre predicted by four types of
production functions, Ida silt loam, 1952
1
(14.5) .2877N” Y'
(14.6) —dv =^
.5843 - .0032N + .0008P
( 14 -7) = .3161 + 3.1756N"- 5 + .1705P- 5 N~- 5
(14.8) = .9839 - .0749N’ 5 + .0008P
level at which the other is considered to be fixed. The marginal prod-

uct equation for the square root function is discontinuous at zero input
rates and forms a curve convex to the origin. When P 2 O is zero, s
Figure 14.7. Corn yield response to P2 0 5 when nitrogen is constant

marginal products predicted by the square root function are

of nitrogen
at first larger (the marginal product of 1 pound of nitrogen is 2.86
bushels) but rapidly decrease and are smaller than marginal products
of the other polynomials for input rates between 40 and 120 pounds. The
square root function does not have marginal products which "diminish
as markedly” as the other polynomials. Hence, its marginal product
decreases less rapidly as nutrient inputs increase to high levels. Its
marginal products again become larger at these high levels. As the
rate of P2 0 5 applications increase, similar relationships are present
but the size of the marginal products for the square root function in-
crease relative to those of other functions.
Marginal products predicted by the quadratic and the "1.5” function
are similar. Those for the "1.5” function are larger than those of the
quadratic function for nitrogen inputs up to 40 pounds, smaller between
80 and 200 pounds, about equal at 240 pounds, and larger for higher
nitrogen inputs. In relation to the polynomials, marginal products pre-
dicted by the Cobb-Douglas function are lower for low input rates but
increase consistently until they are well above all others at high input
rates.
Setting the marginal product equations for both nutrients equal to
zero and solving simultaneously, the rates of nutrient application which
maximize yield can be obtained for 14.2, 14.3, and 14.4. The square
root function predicts a maximum yield of 135.7 bushels of corn per
acre with nitrogen and P 2 0 5 rates of 398 and 337 pounds per acre, re-
spectively. This predicted yield is not as large as some yields actually
obtained in the experiment but both input rates are outside the range of
observations and should be interpreted accordingly. The quadratic
function predicts a maximum of 144.2 bushels per acre with 246 and
240 pounds per acre, respectively, of nitrogen and P2 0 5 The predicted .
maximum for the “1.5” function is 139.7 bushels, with 244 pounds of ni-
trogen and 235 pounds of P2 O s The highest average yield obtained on
.
the experimental plots, 141.3 bushels, was obtained with 200 pounds of
each nutrient. Input combinations derived for maximum yields are well
above this 200 pound figure but predicted maximum yields are all rea-
sonably close to the realized yield figure. Evidently the production sur-
faces predicted by different equations display similar characteristics at
similar yield levels, but at different nutrient input levels.
Yield Isoquants and Marginal Rates of Substitution
Isoquant equations 14.9 through 14.12 were derived from the pro-
duction functions in 14.1 through 14.4, respectively. In these equations,
P2 0 5 is expressed as a function of nitrogen and yield.
, __ „ _ log Y' - .2877 log N - log .4417

(14.9) 10g F ~
.4090
(14.10) P = 184.7001 + .2257N ± 278.2415(.3867

2 5
+ .0053N - .00001N - .007 2Y)'
5 5
(14.11) P = [10.2007 + .4088N' t 1.1979(16.4115N'
*5 2
- .4115N - 1.6696Y + 63.0275) ]
1-5
(14.12) .0576P - (1.1285 + .0008N)P - [-13.6238
1-5
+ .9839N - .0500N - Y] = 0
Isoquants predicted by these equations are presented in figures 14.8

through 14.11, respectively. Convex to the origin, all sets of isoquants
indicate decreasing marginal rates of nutrient substitution. They are
Figure 14.8. Isoquants and isoclines for corn predicted by Cobb-

Douglas function 14.1, Ida silt loam, 1952
successively farther apart, on any straight line through the origin, in-
dicating decreasing returns to fertilizer inputs as the yield level in-
creases. Isoquants of the square root function appear almost horizontal
for considerable distances as they approach the ridgelines but curve
sharply at their centers. The larger distance between isoquants for the
110- and 130-bushel yields indicates extremely low marginal products
at these high yield levels. Isoquants for the quadratic and “1.5” func-
tions curve more gradually than those of the square root equation.
Also, isoquants of the “1.5” function are located slightly lower in the
input plane than like isoquants for the quadratic function. Isoquants for
the two are very similar, however. Isoquants for the Cobb-Douglas
function are asymptotic to the nutrient axes.
Input combinations which produce given yield levels and marginal
rates of substitution for these combinations, tabular counterparts of the
isoquants, are presented in Table 14.5 for three yield levels. The mar-
ginal rates of substitution, ratios of marginal products, were predicted
by equations 14.13 through 14.17 for the respective production functions
in equations 14.1 through 14.4.
POUNDS OF NITROGEN PER ACRE

Figure 14.9. Isoquants and isoclines for corn predicted by
quadratic function 14.2, Ida silt loam, 1952
(Dashed lines are ridge lines.)
Table 14.5. Combinations of Nitrogen and P2 O s Required To Produce Specified Corn Yields and
Corresponding Marginal Rates of Substitution Predicted by Four Types of
Production Functions, Ida Silt Loam, 1952
Square Root Quadratic “1.5* Cobb- Douglas

Pounds Function 14.3 Function 14.2 Function 14.4 Function 14.1
of
Pounds Pounds Pounds Pounds
Nitrogen
per Acre
of
P2 05 ijj Of
Of
of
p2 o 5
mrs(-§)
of
P2 Os
MRS(-§)
of
p2 o 5 |«Q
•O
per acre per acre per acre per acre
Y = 50
20 16 -0.66 102 -1.92 75 -1.79 59 -2.06

40 9 -0.20 72 -1.18 49 -0.98 36 -0.63
60 6 -0.08 52 -0.83 33 -0.63 27 -0,32
80 5 -0.04 38 -0.65 22 -0.42 22 -0.19
Y = 70
40 34 -0.70 116 -1.98 95 -1.84 110 -1.93

60 25 -0.30 86 -1.16 68 -0.99 83 -0.97
80 21 -0.15 67 -0.79 52 -0.63 67 -0.59
100 19 -0.08 54 -0.57 42 -0.41 58 -0.41
Y = 90
60 76 -2.73 161 -3.94 131 -2.72 177 -2.08

80 59 -0.55 118 -1.40 97 -1.15 145 -1.27
100 51 -0.29 96 -0.86 79 -0.68 123 -1.07
120 46 -0.17 82 -0.58 68 -0.43 109 -0.64
c
320
280
o
o
<c
&
CL
lO
%
CL
li_
O
co 120
o
z:
3
o
°- 80
POUNDS OF NITRGGEN PER ACRE

Figure 14.10. Isoquants and isoclines for corn predicted by square
root function 14.3, Ida silt loam, 1952 (Dashed lines
are ridge lines.)
-1
.7034PN
.5843 - .003 2N + .0008P

.6638 - .0036P + .0008N
-3161 + (3,1756 + .1705P */2 )N~Y 2

.4174 + (4.2578 + .1705N V 2
)P“1/2
2
.9839 - .0749N^ + .0008P
1 2
1.1285 . .0864P / + .0008N
Yield Isoclines
Equations 14.17 through 14.20 are the isocline equations for the re-
spective production functions where k equals the price of nitrogen di-
vided by the price of P 2 0 5 .
RESPONSE SURFACES IN FERTILIZER USE
NUMBERS 50 TO
130 INDICATE
YIELD LEVELS
OF ISOQUANTS
O 40 80 120 180 200 240 280 320

Figure 14.11. Isoquants and isoclines for corn predicted by “1.5”
function 14.4, Ida silt loam, 1952 (Dashed lines are
ridge lines.)
P = 1.4212kN
« .6638k - .5843 + (.0008k + .0032)N

(.0008 + .0036k)
5
(14.19) P = [-9.3126 + (-1.2240k + .9271)N‘
± 2.9326(10.0845 + [-2.0078 + 5.5548k]N- s

2 2
+ [(.3161 + .4174k) + .1163k] N)-*]
2
(14.20) P = [-53.4932k t 6.88119(.0075k - .0032
5
+ .0036k + .0002N' + .000003kN)- 5 ]
2
The shape of the isocline families, illustrated in figures 14.8

through 14.11, is determined jointly with the curvature or slope of iso-
quants (the magnitudes of marginal rates of substitution). The marginal
rate of substitution is, in turn, a ratio of marginal products. Because
the square root function predicts initially large marginal products but
492 RESPONSE SURFACES 3N FERTILIZER USE
which decrease rapidly, isoquant slopes for this function are almost
vertical (or horizontal) along their extremities (near ridgelines) but
curve sharply in the center. Hence, isoclines for square root functions
tend to pass through the centers of isoquants and spread only for ex-
treme price ratios and medium or high input rates. They necessarily
begin at the origin, spread out as nutrient applications increase, and
converge at the point of maximum yield. (The point of convergence,
considerably outside the input observations at a yield of 141.3 bushels,
is not shown.) The quadratic function has linear marginal equations
and, therefore, linear isoclines which originate from either axis and
converge at the point of maximum yield. The “1.5” function has iso-
clines which curve sharply as they intersect the nutrient axes but ap-
proach linearity for yields approaching the maximum. The Cobb-
Douglas isoclines are linear, pass through the origin, and diverge with
yield level.
DETAILED ANALYSIS WITH SQUARE ROOT EQUATION
Previous analysis has indicated similarities and differences in pro-

duction functions estimated by the four equations. While each might be
appropriate for certain purposes, the square root function appears to
have a slight superiority in statistical efficiency in predicting the pro-
duction surface for the particular environmental conditions under which
the experiment was conducted. Hence, equation 14.2 is used in this
section for examining certain additional details in physical relation-
ships and in specifying some economic optima.
The predicted yield levels for this function, for various levels on N
and P, are given in Table 14.4. These quantities are the counterpart of
a production surface, except that they represent distinct points on it.
Since both nutrients were present in the soil in limited amounts,
yields were not high for either nutrient used alone. With no P O
2 s , 120
pounds of nitrogen gives a maximum yield of 25.9 bushels in Table 14.4;
with no nitrogen, 120 pounds of P2 Og gives a maximum of 37.5 bushels.2
However, with the addition of 40 pounds of P 2 0 5 , a large yield increase
takes place across the nitrogen columns; a similar change takes place
for P2 0 5 down the first column. In other words, the productivity of one
nutrient is highly limited by the amount of the other with which it is
combined. With both nutrients variable, the predicted maximum yield
These fertilizer quantities do not represent the exact maximum yield. The maximum
yields for nitrogen variable with P O5 fixed at zero or P 0 variable with nitrogen fixed
2 2 5 at
zero are determined by setting the derivatives for each variable nutrient equal to zero and
solving for N or P, respectively, as in (a) and (b) below. The maximum for
P 2 0 5 is with
104.3 pounds; the maximum for N is with 101.0 pounds of this nutrient. The corresponding
yields are 37.7 and 26.4 bushels, respectively.
(a) 0 = +.316 - 3.1756N-- 5 N = 101.0 lbs.

(b) 0 = +.417 - 4.2578P-- 5 p = 104.3 lbs.
is 135.8 bushels with 397.6 pounds of nitrogen and 336.6 pounds of

3
P2 0 5 .
Diminishing total yields for nitrogen, as the variable nutrient, are

indicated up to 200 pounds of P2 O s as the fixed nutrient. Similarly,
negative marginal products hold true for I?2 Os as the variable nutrient,
for up to 280 pounds of nitrogen as the fixed nutrient. Diminishing total
yields are not predicted, within the range of the observation, when both
nutrients are variable in a 1:1 ratio. Just as these two nutrients inter-
act to affect the productivity of each other, another variable resource,
such as stand, might well have caused different productivity coeffi-
cients for either nitrogen or P2 0 5 .
Figure 14.3 is the response surface showing these productivity

relationships more vividly. A vertical slice through this surface
per-
pendicular with the P2 0 5 axis is the counterpart of a single -nutrient
response curve with nitrogen as the variable nutrient; a slice perpen-
dicular to the nitrogen axis represents P2 O s as the variable resource
and nitrogen as the fixed nutrient. A vertical slice intersecting the
origin is the counterpart of a response curve with both nutrients varia-
ble in fixed proportions. Horizontal slices through the surface provide
yield isoquants showing all possible combinations of the two nutrients
which will produce a given yield; these quantities are provided in later
paragraphs.
Table 14.6 indicates the marginal products or yields corresponding
to the total yields of Table 14.4; they are the counterparts of the slopes
of vertical slices through Figure 14.3, at the yield levels of Table 14.4.
These figures again illustrate that the quantity of one nutrient affects
the productivity of the other. For example, movement down any column
of Table 14.4 represents an increase in the ratio of P to N; movement
across a row represents a decrease in the P to N ratio. Down any col-
umn, the marginal product of P2 O s decreases while the marginal prod-
uct of nitrogen increases; across rows the opposite holds true. Mar-
ginal yields per pound of nutrient are equal for the two nutrients when
the quantity of each is 120 pounds. The negative marginal products
represent diminishing total yields; the small positive marginal products
in much of the table correspond to the fact that the production surface
is quite flat over a large section.
Single Variable Input -Output Curves
Figures 14.12 and 14.13 provide total response as yield curves when
one nutrient is fixed at specified levels and the other is variable. With
3
The predicted maximum yield, an extrapolation beyond the observations of the experi-
ment, was obtained as follows. The partial derivatives (the marginal products) for each
nutrient were set at zero; the quantity of each nutrient, to give a partial derivative of zero,
was then computed. These are the quantities of nutrients which give a maximum yield.
They were substituted back into the original function and the maximum yield was predicted
accordingly.
—
Table 14.6. Marginal Product or Yield (Bushel per Pound of Fertilizer Nutrient) for
Combinations Indicated in Hows and Columns; Upper Figure for
Nitrogen; Lower Figure for P2 0 5 *
_ Pounds Nitrogen
Pounds
, .
of
of
—-
P205 0 40 80 120 160 200 240 280 320
0 — .19 .04
*
O CM -.07 -.09 -.11 -.12 -.14
1I
40 .36 .16 .07 .02 -.02 -.04 -.06 -.08

.26 .43 .49 .55 .60 .64 .67 .71 .74
80 .43 .21 .11 .06 .01 -.01 -.04 -.05

.06 .17 .23 .27 .30 .33 .35 .38 .40
120 .48 .24 .14 .08 .04 .01 -.01 -.03

-.03 .07 .11 .14 .17 .19 .21 .23 .25
160 .52 .28 .17 .11 .06 .03 .01 -.02

-.08 .01 .04 .07 .
.09 .11 .13 .14 .16
200 — .57 .31 .19 .13 .08 .04 .02 -.01

-.11 -.04 -.01 .02 .03 .05 .07 .09 .10
240 — .60 .33 .21 .14 .10 .06 .03 -.01

-.14 -.07 -.04 -.02 -.01 .02 .03 .04 .05
280 — .63 .36 .23 .16 .11 .07 .04 .02

-.16 -.10 -.07 -.05 -.03 -.02 -.01 .01 .02
320 — .67 .37 .25 .18 .12 .09 .06 .03

1 00 -.12 -.09 -.07 -.06 -.04 -.03 -.03 .01
These figures are the derivatives of yield in respect to the single -nutrient variable
while the other is fixed. They are derived from equation 14.3, with the nitrogen and
P2 0 5 quantities shown at the top of the columns and to the left of the rows. The .36
in the cell where both nutrients are 40 pounds is the derivative or marginal product
for nitrogen as the variable nutrient while P 2 O s is fixed at 40 pounds. The .43 is the
marginal product for P 2 0 5 as the variable nutrient while nitrogen is fixed at 40
pounds.
a zero nitrogen input for Figure 14.12, the P 2 0 5 curve falls low in the
plane with diminishing total yield indicated for small inputs of P 2 O s .
With nitrogen input at 160 and 320 pounds, the response curves for P2 0 5
cross each other. This is due to the fact that, with small quantities of
P2 0 5 , 320 pounds of nitrogen gives an excessive quantity of nitrogen;
with larger quantities of P 2 0 5 the two nutrients interact to give
,
slightly higher yields for 320 than for 160 pounds of nitrogen. A simi-
lar situation exists for nitrogen as the variable nutrient. With P 2 0 5
fixed at 160 and 320 pounds, the nitrogen response curves in Figure 14.13
again cross each other. An increase in P2 0 5 from 160 to 320 pounds
adds nothing to yield if nitrogen inputs are small. The fact that the
maximum yield from nitrogen, with no R
05 , is lower than the maxi-
mum of P2 0 & with no nitrogen, suggests that the soil, while deficient
,
in both nutrients, was lacking especially in P 2 0 5 .

0 80 160 240 320

POUNDS OF Pg 0 5 POUNDS OF NITROGEN
Figure 14.12. Total yield with P 2 0 5 Figure 14. 13. Total yield with nitrogen
variable and nitrogen variable and P2 05 fixed
fixed at three levels at three levels
“Scale Line 5
’
Response Curve With Both Nutrients Variable
Figures 14.14 and 14.15 show predicted input -output or response

curves when the two nutrients are increased in fixed ratios. The
amount of one element is always in a fixed ratio to the amount of the
other, as indicated on the bottom of the graphs. In Figure 14.14, for
example, the ratio line of IP = 2. ON means that 2 pounds of P 2 0 5 is
used for each pound of nitrogen; with a nitrogen input of 160 pounds, in-
put of P2 0 5 is 320 pounds; and with nitrogen at 320 pounds, input of
P 2 O s is 640 pounds. These two figures indicate that greatest yields can
be obtained from use of the two nutrients in a 1:1 ratio. For light ap-
plications of fertilizer, greater response per pound may be obtained
with nutrient ratios differing from 1:1.
Yield Isoquants
Yield isoquants derived from the same basic yield surface equation
were shown in Figure 14.10. The isoquants show that as higher and
Figure 14.14. Yield of corn with Figure 14.15. Yield of corn with
nutrients increased nutrients increased
in fixed proportions in fixed proportions
higher yields are attained, the marginal rates of substitution between

P 2 O s and nitrogen change along a scale line (a fixed nutrient ratio line).
In other words, the slopes of successively higher isoquants are differ-
ent at the points where they are intersected by an imaginary straight
line through the origin. This change in the slopes of the yield isoquants
indicates that the combinations of nutrients (the fertilizer ratio) which
gives lowest cost for one yield level is not the same mixture which
gives lowest cost for another yield level. In other words, the least -cost
combination is not the same for yields of 60 and 120 bushels. This
same point is illustrated in Table 14.5 which shows several predicted
combinations of the two nutrients which will produce the same yield and
the marginal rates of nutrient substitution for the indicated combina-
tions. Figures for isoquants indicate, on the one hand, the minimum
quantity of nitrogen and the maximum quantities of P2 0 5 which will pro-
duce the stated yield and, on the other hand, the maximum quantities of
P2 0 5 and the minimum quantities of nitrogen. More P2 O s must be used
with a stated amount of nitrogen for a higher yield as compared to a
lower yield. With 160 pounds of nitrogen, 165 pounds of P2 0 5 allows a
yield of 120 bushels; only 64 pounds of P2 0 5 is required with 160 pounds
of nitrogen to produce 100 bushels. The yield isoquants also indicate
that the range of N/P ratios, over which the two nutrients can be sub-
stituted in obtaining a given yield, narrows as higher yield levels are
attained. For higher yields, the nutrients become limitational in nature
as the “upper ends” of the isoquants take on an infinite slope and as the
"lower ends" take on zero slopes. Low yields can be attained by addi-
tion of one nutrient alone, but high yields can be attained only with some
minimum quantity of either nutrient. The maximum yield per acre, as
predicted from the equation, can be produced by only one combination of
p2 Qj and nitrogen (i.e., the isoquant for a yield of 135,8 bushels re-
duces to a single point corresponding to 397.6 pounds of nitrogen and
336.6 pounds of P2 O s ).
Economic Optima
The various quantities and relationships derived above provide a

basis for specifying selected economic optima, applying to the particu-
lar environmental conditions of the experiment. We first examine the
optimum level of applying one nutrient when the level of the other is
considered to be fixed. We suppose N to be variable but P2 Q> to be at
zero. The optimum level of nutrient application depends on the magni-
tude of marginal products and the nutrient to corn price ratio. With
corn at $1.40 per bushel and nitrogen at $.18 per pound (including ni-
18
trogen and the cost of application) the price ratio is or .129.
Hence the derivative corn yield with respect to ni-

of equation 14.3 for
trogen equation 14.21 below, with P 2 Og at zero
is set at this quantity in
level. Solving 14.21 for N, 53.3 pounds of nitrogen equates the mar-
ginal product, and therefore is the most profitable quantity of this nu-
trient when no P2 O s is used. The corresponding yield, from equation
14.3, is 24.8 bushels.
5
(14.21) = .316 + 3.1756N'- = .129
oN
18
With corn at $.80 and nitrogen at $.18, the price ratio is or .225
and 34.8 pounds of N is the level of fertilization to maximum profits.
With the price ratio at -r4“ or

i.û
.071, 67.4 pounds of N is most profitable.
^0
However, when 80 pounds of P2 0 5 are used and the price ratio is ,
136.9 pounds of nitrogen represents the optimum.

Using the same price ratios for P 2 05 , we obtain the values below.
For any one of the nutrient to crop price ratios shown, the optimum
quantity of P2 05 with a stated amount of nitrogen, is slightly greater
than the optimum quantity of nitrogen with the same amount of P2 O s .
Zero input of nitrogen
Price ratio of .129; P 05 optimum

2
is 60.8 pounds
Price ratio of .225; P2 0 optimum

5
is 44.0 pounds
Price ratio of .071; P2 05 optimum is 76.1 pounds

160 pounds of nitrogen

Price ratio of .071; P2 Gj optimum is 175.9 pounds
Minimum Costs for a Specified Yield
optimum quantity of a single nutrient is only a par-

Selection of the
tial solution of theeconomic problem of fertilizer use. Still to be
solved is (a) the optimum quantity of each nutrient or rate of applica-
tion when both nutrients are variable and (b) the best combination of
nutrients for any one yield level. The change in the slopes of the yield
isoquants along a scale line in Figure 14.10 suggests that the combina-
tion of the two nutrients which will give the lowest cost, for a stated
yield, changes with the level of yield. The nutrient combination which
is best for a 100-bushel yield is not also best for a 50-bushel yield.
This point also is illustrated with the isoquant and substitution data of
Table 14.7. The least -cost resource combination for a given yield is
attained when the marginal rate of substitution of the resources (i.e.,
the derivative of one nutrient in respect to the other, with yield at
stated levels) is equal to the inverse price ratio. Hence, we can illus-
trate that the proportion of the two nutrients, to give the least cost, dif-
fers with yield level. First, we derive the equation of marginal rates
of substitution (the first derivatives of change in one nutrient with re-
spect to the other) as in equation 14.15. Second, we set it equal to the
particular price ratio for the nutrients and solve for the nutrient com-
bination which minimizes cost for the particular yield. Or, we can set
the derivative of P with respect to N equal to the price of nitrogen di-
vided by the price of P2 O s Using the equation in 14.22, and equating it
.
to different price ratios, and solving for N and P, we obtain the results
in Table 14.7.
-3 + .1705N"- S P- 5
(14.22)
fip
-£5 =
5N
3.1756N
4.2578P
- *5
— —
+ .1705N'
5
P
_ 5
- .316
+ .417
price ratio
These are the least -cost combinations for the specif ied price ratios
and yield isoquants. With a N/P price ratio of 1.5, the combination for
a 50-bushel yield should total 36.1 pounds; 32.7 per cent of this should
be nitrogen and 67.3 per cent should be P2 0 5 For the same price .
ratio, a total of 180.9 pounds, composed of 43.8 per cent nitrogen and
56.2 per cent P 2 O 5 should be used to minimize fertilizer cost for a
,
100 -bushel yield. The mixture, to minimize cost for a given yield,
should contain relatively more nitrogen for higher yield levels. Tra-
ditionally, this distinction has not been made in fertilizer recommenda-
tions; the same fertilizer mix has, for a given soil and productivity
situation, usually been recommended for numerous yield levels.
Table 14.7. Combinations of Nitrogen and P2 0 5 Needed

to Minimize Fertilizer Costs per
Specified Yield Level for
Different Price Ratios
Yield Optimum Optimum

Level Pounds of N Pounds of P
Price of $.18 per lb. for N and $.12

N
per lb. for P (jj- ratio of 1.5)
50 bu. 11.8 24,3

100 bu. 79.3 101.6
Price of $.12 per lb. for N and $.18
per lb. for P (--ratio of .67)
50 bu. 19.8 16.3

100 bu. 99.1 82.7
Similarly, with a change in the price ratio from 1.5 to .67, the percent-
age of nitrogen, for a 50-bushel yield, should change from 32.7 per cent
to 54.8 per cent. For a 100-bushel yield, similar changes in the price
ratio should cause the nutrient combination to change from 43.8 per
cent to 54.5 per cent nitrogen.
Solution for Two -Variable Nutrients
In the analysis above, principles of profit maximization were used

to independently specify (a) the optimum quantity of one -variable nutri-
ent, with yield as a variable and the second nutrient fixed and (b) the
optimum combination for two -variable nutrients for a given or fixed
yield. However, these conditions need to be imposed simultaneously if
the economic optimum usage of fertilizer is to be determined. In other
words, we must simultaneously determine the optimum (a) combination
of nutrients and (b) level of application. It was explained in an earlier
section that the combination of nutrients which gives lowest cost for one
yield level does not similarly give the least cost for other yield levels.
This is true since the slopes of the yield isoquants, and hence the mar-
ginal rates of substitution between nutrients, change with higher yield
levels.
One approach to determining the dual solution outlined above is that
of successive approximation. One can use the principle that application
of more fertilizer is profitable (for a farmer with unlimited capital) as
long as the marginal product of a fertilizer nutrient is greater than the
nutrient to crop price ratio. Hence, with a price of $1.40 per bushel for
corn, $.18 per pound for nitrogen, and $.12 for P 2 0 5 , we can obtain so-
lutions by successive approximations using Table 14.6. The price

ratio is .085; we can move down the first column until we find a mar-
ginal product for P 2 O s which is greater than .085. The marginal prod-
uct of the 40th pound of P 2 O s is .26 -
hence, it is profitable. The 80th
pound of P2 05 is not profitable since its marginal product of .06 is less
than the price ratio of .85. Starting from zero nitrogen, we can then
move across the second row to determine the amount of nitrogen which
is profitable, with 40 pounds of P 2 0 5 already applied. Since the nitro-
gen to corn price ratio is .125, the 80th pound of nitrogen is profitable;
the 120th pound is not since the marginal product of .07 is less than the
18
price ratio of or .125.
Now, with 40 pounds of P2 Q> and 80 pounds of nitrogen, we move

down the third column. With 80 pounds of nitrogen, the 120th pound of
P2 0 5 becomes prpfitable since its marginal product of .11 is greater
than the price ratio of .085. With 120 pounds of P2 O s , the 120th pound
of nitrogen also becomes profitable. From the data in Table 14.4 and
with the prices quoted, the method of “ successive approximation” indi-
cates that 120 pounds of each nutrient is profitable. However, the suc-
cessive approximation may require added steps in arithmetic before
the final solution is attained.
The successive approximation indicates only which of the combina-
tions in the table are most profitable. It does not indicate the exact
combinations which might be more profitable. The exact fertilizer
combination can be solved by setting the marginal products or partial
derivatives for both nutrients equal to the price ratios and simultane-
ously solving for the quantity of the nutrients to apply for maximum
profits. These optima are attained when the partial derivatives (the
marginal products) for both nutrients are equal to the nutrient to corn
price ratio. Hence, with a price of $1.40 for corn, $.18 for nitrogen,
and $.12 for P2 O s , the equations become 14.23 and 14.24 below.
5C __ 3.1756 .1705 j/P _ .18

(14.23) -.316 +
6N yw yw i.4o
5C _ 4.2578 .1705 j/gT .12

(14.24) -.417 +
5P yp- yr
From simultaneous solution of these equations, we obtain the figures
for situation A in Table 14.8: 298.93 pounds of fertilizer should be
used, including 156.45 pounds of P2 0 5 and 142.48 pounds of nitrogen.
The same procedure has been used for the other price situations in
Table 14.8. With a decline in corn price by 36 per cent (from situation
A to situation B), total usage of fertilizer should decline by 30 per cent.
Input of nitrogen should decline 34 per cent and input of P2 O s should de-
cline 26 per cent if profit is to be at a maximum; inputs should not be
reduced by the same proportions. With a 43 per cent increase in corn
price (from situation A to situation C), total input of fertilizer should
increase by 25 per cent; input of nitrogen should increase by 30 per cent
Table 14.8. Optimum Quantity of Fertilizer and Optimum Combination

of Nutrients for Specific Price Relationships
Optimum
Optimum Fertilizer Use
Price Yield
Situation in Bushels Total pounds Pounds N Pounds P2 0 5
A: corn at $1.40;
Nat .18;
P at .12 117.21 298.93 142.48 156.45
B: corn at $ .90;
Nat .18;
P at .12 104.99 209.27 94.06 115.21
C: corn at $2.00;
Nat .18;
P at .12 124.22 374.84 185.04 189.80
D: corn at $1.40;
Nat .12;
P at .12 122.30 349.50 180.19 169.31
E: corn at $2.00;
Nat .12;
P at .18 124.91 384.18 208.72 175.46
and input of P 2 O s should increase by only 21 per cent. With corn at

$1.40 and a 1:1 price ratio for nutrients (situation D), input of nitrogen
should be greater than input of P2 0 5 With a N/P price ratio of 1.5
.
(situation C), input of phosphate should be about 5 pounds greater than

input of nitrogen. However, with a N/P price ratio of .667 (situation E),
input of nitrogen should exceed input of phosphate by 33 pounds.
We have illustrated that simultaneous solution of (a) the optimum
rate of fertilization and (b) the optimum combination of nutrients is
possible from appropriate experimental data. We also have illustrated
some points ordinarily overlooked in both economic and agronomic
recommendations: a reduction in product price not only may call for a
reduction in the total quantity of fertilizer used on corn, it also may
specify a change in the fertilizer grade. These and many other basic
principles can be applied when fertilizer experiments are designed to
provide relevant marginal quantities and the corresponding economic
analysis.
For high level yields and recent prices, the cost of the optimum nu-
trient combination (computed by both partial derivatives with their re-
spective price ratios) for corn is only slightly less than numerous other
nutrient combinations which will give yields in the neighborhood of 125
bushels. The reasons for this outcome are explained below.
Since an isocline connects all points of the same slope (i.e., equal
substitution rates) on successive isoquants, the isocline conforming to
a particular price ratio also is an expansion path. It traces all combi-
nations of nutrients which give least -cost yields. If an isocline con-
forming to a particular price ratio is nearly straight, an increase in
nutrients by a fixed proportion is nearly consistent with the least-cost
use of nutrients. If the isocline bends sharply, a fixed-ratio fertilizer

increase will not give the most economic nutrient combinations. While
little is yet known about them, isocline maps can take on
many distinct
forms. They can be established only by basic experiments. In a family
of isoclines, one denoting a substitution ratio of 1.0 may be bent; one
for a .5 substitution ratio may be linear. Hence, with a nutrient price
ratio of 1, least -cost fertilizer mixture proportions will vary across
the surface; with a .5 price ratio, the fertilizer mix should follow a
fixed ratio line, although no particular ratio can be specified without
knowledge of the function. One of a family of isoclines may be straight
(although it need not be one along a 1:1 ratio); none may be straight. All
may be straight but may not pass through the origin. The extent to
which profit is depressed, in increasing or decreasing fertilizer inputs
in fixed mixes as price levels change, depends on the slope and curva-
ture of the relevant isocline. If it is curved and passes through the
origin but has a slope which is nearly constant, the depression will not
be great. If it intersects the nutrient axes near the origin and is a
straight line, the same will hold true. Of course, the relative price of
nutrients also is important in determining profit sacrifices from using
a fixed nutrient mix for all yield levels.
PRODUCTION FUNCTIONS WITH OMISSION

OF ZERO NUTRIENT APPLICATIONS
While economic optima are illustrated only for the quadratic func-
tion, the same procedure can be applied to the other equations from 14.1
through 14.4. However, because of algebraic properties, they suggest
different nutrient mixes and amounts for any one price level. For ex-
ample, Table 14.9 shows some of the differences in predicted inputs (N
and P) and output (Y) for several price situations under each of the
functions.
Because of the algebraic nature and forced condition of constant
from the Cobb-Douglas function are unrea-
elasticity, the predictions
sonable. The square root function, equation 14.3, gives the most
Table 14.9. Predicted Optimum Inputs and Output for Several Price Levels Under the
Cobb-Douglas (14.1), Quadratic (14.2), Square Root (14.3), and “1.5*
(14.4) Equations
Price ($) of Inputs and Outputs Specified by
Equation 14.1 Equation 14.2 Equation 14.3 Equation 14.4

Y N P N P Y N P Y N P Y N P Y
1.60 .08 .09 12,539 15,832 2,195 225 220 143 240 212 129 216 207 138
1.00 .08 .09 2,666 3,369 7-57 213 203 141 185 168 123 196 192 136
1.60 .17 .13 1,757 3,266 665 205 208 141 162 166 120 190 192 135
1.00 .17 .13 374 644 236 180 189 135 108 121 108 161 169 129
conservative estimates of input applications. Changes in input levels

are not great, for selected changes in price ratios, because of the na-
ture of the production functions under the environmental conditions
studied.
It has been suggested that with marginal productivities for initial
nutrient inputs of the magnitude realized in this study, less difference

in the optima specified by the various equations might occur if zero
rates of application were omitted from the analysis. This procedure is
used in this section.
Regression Equations
The types of equations shown in equations 14.1 through 14.3 are re-
peated in equations 14.25 through 14.28, respectively, with observations
for zero rates omitted.
2119 1246
(14.25) Y' = 19.94N' P'
2
(14.26) Y = 38.3122 + .5093N + .2788P - .Q013N
2
- .0007P + .0004NP
(14.27) Y = -37.5002 + 15.5508N' 3 + 5.7321P 3

-5 -5
- .7077N - .3536P + .3941N P
1 -5
(14.28) Y = 25.2709 + .9279N + .5088P - .0449N
1 -5
- .0255P + .005NP
The coefficients of determination are .45, .77 .84, and .78, respec-
tively, for the four equations. The R 2, s for the polynomials are about
10 per cent below those resulting when all observations were included.
The difference in R 2, s of the quadratic and “1.5” equations are negligi-
2
ble. The square root equation again has the highest R (Yield inter- .
cept values have no meaning because there were no observations for

zero nutrient levels.) Analyses of variance of regression for the four
equations are presented in Table 14.10. The square root equation again
has the smallest figure for deviations from regression.
The number of observations was reduced from 114 to 80 by the de-
letion. Because these observations represented yields which were rel-
atively large distances from the mean yield level, their omission
greatly reduced total and treatment sums of squares. Also, the mean
corn yield was increased from 86.07 to 113.39 bushels per acre. Be-
cause treatment sums of squares were reduced relatively more than the
sums of squares among plots treated alike, R 2, s of the equations were
lowered by dropping the observations.
Table 14.10. Analyses of Variance of Regression for Equations

Fitted to Selected Corn Yield Observations,
Ida Silt Loam, 1952
Degrees of Sums of Mean

Source of Variation Freedom Squares Squares
Total 79 38,439
Treatments 39 31,001 795
Cobb-Douglas equation 14.25 2 17,217 8,609
Deviations from regression 37 13,784 373
quadratic equation 14.26 5 23,859 4,772
square root equation 14.27 5 25,939 5,188
*1.5” equation 14.28 5 24,274 4,855
Among plots treated alike 40 7,438 186
Figure 14.16. Production surface for corn predicted by Cobb-Douglas function

14.25, Ida silt loam, 1952
Figure 14.17. Production surface for corn predicted by quadratic function 14.26,
Ida silt loam, 1952
Figure 14.18. Production surface for corn predicted by square root function
14.27, Ida silt loam, 1952
Figure 14.19. Production surface for corn predicted by “3/2” function 14.28, Ida
silt loam, 1952
Predicted Yields and Production Surfaces
Production surfaces for the four equations are presented in figures

14.16 through 14.19. They are, of course, flatter than the earlier sur-
faces. Yields predicted by the four functions are now more comparable
than before the zero observations were dropped. Again, predictions of
the quadratic and "1.5” functions are most similar and, in this case,
are almost identical. The square root function, which characteristi-
cally predicts higher, lower, and, finally, higher yields than the other
polynomials as nutrient inputs increase, still follows this pattern, but
the differences are now quite small. For example, for 80 pounds of
both nutrients, the square root function predicts 8.3 bushels more than
the quadratic function; before the zero rates were dropped this differ-
ence was 20 bushels. Predictions of the Cobb-Douglas function are
more similar to those of other functions than before but, due to the con-
stant elasticity assumption, still vary widely for some nutrient levels.
The predicted yields of the functions for 40 -pound nutrient rates are all
very close to the mean yield for that treatment level, 71.4 bushels.
As Figure 14.20 indicates, the single nutrient input-output curves
are made much more similar through deletion of zero rates. However,
the Cobb-Douglas curve still is highly different from the others.
In summary, dropping the observations changed the response pat-
terns in this way: predicted responses to nitrogen have been increased
relative to P 2 0 5 responses. Marginal products of all functions were, in

Figure 14.20. Corn yield response to nitrogen when P2 0 5 is constant
general, lowered. The deletion affected the square root function more
than the other polynomials. Most significant was the change in the input
rates at which marginal products of this function were forced to zero.
Ridgelines for it became less curved and converged at nutrient input
rates similar to the other functions. Therefore, diminishing returns of
yields predicted by this function became more similar to those of other
functions and input levels at which maximum yields occur became al-
most identical for the three functions. Hence, while the pattern of yield
response differs between the two sets of production f met ions, omitting
the observations has made predictions of the polynomial equations as a
group more similar. The isoquant and isocline maps for the quadratic
equation 14.26, square root equation 14.27, and “1.5” equation 14.28
functions are shown in figures 14.21 through 14.23. The Cobb-Douglas

quadratic function 14.26, Ida silt loam, 1952
isoquants and isoclines for equation 14.25 are not presented since they
fan out, rather than converge to a yield maximum. The isoclines for
the square root function now converge to a yield maximum within the
observed range of inputs.
Isoclines predicted by equations 14.25 to 14.28 are more similar
than those obtained before the zero observations were omitted. Pre-
viously isoclines and ridgelines of the square root function tended to
diverge for the inputs of the experiment. They now converge at a max-
imum yield similar to those of other functions and ridgelines are nar-
rowed accordingly and limited the divergence of the isoclines. Iso-
clines for the *1.5” function have also been changed, and now are nearly
linear. Isoclines of all functions have shifted due to the change in the
response pattern.
POUNOS OF NITROGEN PER ACRE

square root function 14.27, Ida silt loam, 1952
Economic Optima
Combinations of nitrogen and P 2 0 5 which minimize the cost of pro-

ducing 90 and 130 bushels of corn for price ratios (Pn /Pp ) of 5/1, 2/1,
13/9, 1/2, and 1/5 are presented in Table 14.11.
Amounts of nitrogen predicted by the polynomial functions are quite
similar for all yield and price situations. For instance, for a yield of
130 bushels and a price ratio of 1/2, 207, 207, and 206 pounds of nitro-
gen are recommended by the square root, quadratic, and *1.5* function,
respectively. Recommendations for P2 0 5 are somewhat more variable;
the amounts being 195, 173, and 178 pounds per acre, respectively.
Combinations predicted by the three polynomial functions are most

“1.5” function 14.28, Ida silt loam, 1952
similar for the 130 -bushel yield level, indicating that predictions of
these functions become more comparable as yields increase. The
Cobb-Douglas function again differs considerably from the polynomials.
Combinations of nitrogen and P2 0 5 which maximize profits per acre
are presented in Table 14.12 for various corn and nutrient prices.
These combinations were derived from functions 14.25 to 14.28 in the
manner explained previously.
Predictions of the quadratic function and the “1.5* function are
closely comparable for all price situations; those of the quadratic being
slightly higher in all cases. Nutrient combinations and yields predicted
by the square root function are, in all situations, more conservative
than those of the other polynomials. Predicted profits are almost iden-
tical for all three polynomial functions except in the first price situa-
tion where the quadratic function predicts $5.69 more profit per acre
than the square root function. Differences among predictions of these
functions increase as nutrient prices rise relative to corn prices but
this increase is small. Hence, the functions are about equally sensitive
Table 14.11. Cost Minimizing Combinations (Pounds) of Nitrogen and P2 O s per Acre
for Corn Yields of 90 and 130 Bushels per Acre Predicted by Four
Types of Production Functions (Fitted to Selected Yield
Observations), Ida Silt Loam, 1952
Price
Ratio Cobb-Douglas Quadratic Square Root “1.5* Function
(— N P2 Os N p2 o5 N
V Y = 90
p2 o5 N p2 o5
5 19 173 54 146 57 88 55 130

2 83 98 74 87 65 60 71 80
13/9 94 79 87 64 70 53 81 63
1/2 138 40 —* — -- — —
1/5 — — — — — — — —
Y = 130
5 177 218 186 228 178 207

2 248 291 184 196 190 211 184 200
13/9 279 237 188 188 195 207 190 193
1/2 — -- 207 173 207 195 206 178
1/5 — — 218 168 215 193 217 175
*Input combinations are beyond limits of isoquant and isocline map.
to price changes. Predictions of the Cobb-Douglas function have been

reduced relative to those of the other functions by dropping the zero ob-
servations. However, nutrient predictions for many price situations
still fall beyond the limits of observation for the experiment.
In summary, differences in recommended optimum nutrient inputs
of the functions have not been eliminated by omitting the zero observa-
tions. These differences have been made more consistent, however.
Yield and profit predictions of the polynomials are closely comparable
after the deletions.
Table 14.12. Predicted Optimum Inputs and Outputs for Several Price Levels Under
Equations 14.25 Through 14.28
Price ($) of; Inputs and Outputs Specified by
YNPNPYNPYNPYNPY Equation 14.25 Equation 14.26 Equation 14.27 Equation 14.28
1.60 .08 .09 709 370 166 215 212 134 191 186 128 208 208 132
1.00 .08 .09 349 183 132 198 184 131 167 149 125 189 175 129
1.60 .17 .13 245 188 123 189 187 130 157 150 124 179 177 128
1.00 .17 .13 - - - 159 145 121 126 110 78 148 135 119
Profit Maximizing Combinations — Limited Capital
Quantities derived for limited capital situations, by the methods out-

lined in Chapter 3, are presented in Table 14.13 when zero application
*
Table 14.13. Corn Yields, and Nitrogen and P 2 0 5 Combinations

Predicted for Limited Capital Situations When
Nitrogen and P2 0 5 Cost $.13 and $.09 per
Pound, Respectively
Pounds per Acre

Function Capital Corn
N R0 5 Yield
Quadratic equation 14.26 20 98 80 97

40 181 182 129
Square root equation 14.27 20 97 81 105
40 179 186 128
“1.5” equation 14.28 20 97 81 99
40 181 183 129
rates are not included in the analysis. The figures show the optimum
amounts of N and P2 0 5 which should be used if capital for fertilizer is
limited to $20 and $40 per acre for three functions.
For capital levels of $20 and $40 per acre, nutrient recommenda-
tions are almost identical. The above comparison has been made only
for one set of nutrient prices. However, this set of prices typified fer-
tilizer prices current at the time of the study. With greatly different
nutrient prices and amounts of capital, the predictions by the three
equations would differ more.
RED CLOVER
The general empirical procedures for red clover were the same as
for corn except that only the three equations below were derived. The
methodological steps in fitting functions only to observations of nonzero
fertilizer inputs were not repeated for red clover and alfalfa. (The ex-
periments were designed for both alfalfa and red clover as outlined in
Table 14.1.) Here Y refers to yield in tons, Y' refers to yield above
check plot, P refers to P 2 0 5 and K refers to K 20 in pounds.
,
0384 1884
(14.29) Y' = .36551K* P*
(14.30) Y = 2.657 + .0019K + .0079P - .0000018K

2
2
- .Q000167P - .0000031KP
(14.31) Y = 2.46 - .000073K - .003952P + .028141K**

5 5 5
+ .128004P - .000980K* P'
Basic statistics of the equations are given in Table 14.14. The

clover data were relatively more variable than were the corn data. For
the preceding three algebraic functions, the largest portion of variance
explained by fertilizer nutrients was the 64 per cent for the square root
Table 14.14. Values of R and t for Individual Regression Coefficients
Equations Value of R Value of t for Coefficients in Order Listed
14.29 .7510* 15.16* 3.12*

t-
10.30 .7622* 2.17* 9.20* .76* 6.99* H-a 00

—
f-
—
1
14.31 .8016* .10* 5.52* oo 1— 8.23* 1.82*
*0 < P < .01

t.10 < P < .20
i-p > .40
functions. The t value for one regression coefficient in both the cross-
product and square root function was not significant at the 40 per cent
2
level of probability. The K term was dropped from the quadratic equa-
tion and the K term from the square root equation (see Table 14.15)
since these terms were not significant. The new regression coeffi-
cients are shown in equations 14.32 and 14.33.
2
(14.32) Y = 2.68 + .0013K + .0Q79P - .0000Q017P - .0000031KP
5 5
(14.33) Y = 2.468 - .003947P + .026834K- + .127892P*
5 *5
- ,000979K* P
As Table 14.15 indicates, dropping one of the equa- term from each
4
tions did not result in a significant decrease in the values. After R 2
comparing response curves and isoquants from the new two -variable
functions with (a) individual observations from the experiment and
(b) similar estimates from the single -variable function, it was decided
to use the latter four -term square root function for the estimates which
follow, i.e., to use equation 14.33.
Table 14.15. Sum of Squares and Value of F in Analysis of

Variance for Added Regression Term
Item* Quadratic Square Root
Deviation from regression,

four terms 15,026.07 12,751.18
Deviation from regression,
five terms 14,946.95 12,749.88
Reduction due to added term 79.12 1.30
Value of F .572 .011
^Degrees of freedom are 109 for four terms, 108 for five terms, and 1
for regression term analyzed.
4
The Rvalues are .76 and new quadratic and square root
.80, respectively, for the
functions. In order regression equations, the t values are:
of the coefficients in the
quadratic = 3.21, 9.20, 6.98, and 1.50: square root = 5.55, 3.08, 8.29, and 1.82.
Production Surface Estimates
The surface of Figure 14.24 shows the surface to be relatively flat

for large inputs of either or both nutrients. Diminishing total yields are
attained with extremely large quantities of P2 0 5 While the marginal
.
products of K 2 0 decline for large inputs, negative marginal products do

not exist, on the predicted surface, for this nutrient. The marginal
products for small nutrient inputs are largest for P2 O s Hence, it is
.
the most limiting two nutrients. However, with more than 160
of the
pounds of both nutrients, K 20 has higher marginal products than P2 0 5 .
The first 40 pounds of P 2 0 5 have a greater effect in increasing total

yields than for K 2 0 although increases from P2 O5 are smaller as K 2 0
is increased. This is because P2 Ob and K 2 0 substitute more for each
other in red clover production than did P 2 0 5 and N for corn.
KgO IN POUNDS
Figure 14.24. Predicted yield response surface

for red clover
Yield Isoquants and Substitution Ratio
Figure 14.25 includes the product isoquants predicted for yields of

2.8, 3.1, and 3.4 tons per acre.
Only slight quantities of P2 05 alone or P2 Os in combination with K* O
are predicted to produce a yield of 2.8 tons. This yield can be produced
with about 8 pounds of P2 Os alone; for any quantity of P2 05 1 pound of
,
the K 2 0 replaces less than 1 pound of P O Although the isoquant ex-

.
2 s
tends out as far as 80 pounds on the IC,0 axis, this quantity of K 0
2
Figure 14.25. Yield isoclines showing points of equal slope

and replacement rate for red clover isoquants
would never be profitable in producing a 2.8-ton yield. On this portion

of the isoquant, 1 pound of K 2 0 substitutes for only a very small frac-
tion of a pound of P 2 O s
. Actually, economic combinations of nutrients
for a 2.8-ton yield do not exist away from the P 2 0 5 axis. Only this nu-
trient should ever be used for a 2.8-ton yield; for all practical pur-
poses, the 2.8-ton isoquant does not exist. Somewhat the same situation
holds true for a yield of 3.1 tons. A 3.1-ton yield is predicted to be at-
tained with 37 pounds of P2 O s and none of K 2 0, or any of the other com-
binations indicated on the middle isoquant. The replacement rate of
K 2Q for P 2 O s becomes less than 1:1 with 36 pounds of P2 Ob and 2
pounds of K2 0. While the isoquant is predicted to extend far out toward
the K2 0 axis, the substitution rates are extremely low at these ex-
tremes.
The isoquant for a 3.4-ton yield has greater slope than the 3.1-ton
isoquant. Accordingly, K 2 0 replaces P 2 0 5 at a higher rate over a
wider range of Kg O inputs. One pound of KÔ replaces more than 1
pound of P 2 0 5 up to 14 pounds of K 2 0; the substitution rate becomes 1:1
with 14 pounds of K 2 0 and 98 pounds of P2 0 5 . But only a very slight
amount of K 2 0 will, according to the predictions of the production rela-
tionships, be profitable at the usual ratio of prices for the two nutrients.
However, larger amounts of K 2 0 are predicted to be necessary for
higher yields.
This point is illustrated by the lines labeled RR. They are isoclines
indicating the path of a given replacement rate over the map of yield
isoquants or contours. They indicate the point on any yield isoquant
where the replacement rate, for the stated yield, is that indicated by the
isocline. For yields of 3.4 tons or less, the three isoclines (which rep-
resent a range of price ratios which might be attained for the two
nutrients) are close to the phosphate axis. For higher yields, they are
predicted to veer in a direction specifying a larger proportion of 2 0. K
Yields as high or greater than 3.4 tons would never be profitable; the
nutrient to hay price ratio has never been low enough and the marginal
response for fertilizer is not high enough at this yield level. However,
the isoclines do predict the least -cost nutrient combination for each
possible yield level. With P2 O s and K z O costing the same amount per
pound, the price ratio is 1.0. Since the least-cost nutrient combination
is attained when the price ratio is equal to the replacement or substitu-
tion ratio, each isocline traces out the path of least-cost nutrient com-
binations as higher yields of red clover are attained.
ALFALFA
Two -variable functions derived from the alfalfa yield data are listed
below. Predictions from these were compared with (a) predictions
from 35 single -variable functions and (b) a scatter diagram of observa-
tions. These comparisons, along with the statistics of Table 14.16,
again suggested that the square root function provided the best esti-
mates of the production or yield surface and related quantities.
0542 * 1310
(14.34) Y' = .87935K* P
(14.35) Y = 2.2514 + .0033K + .0097P - .000007K 2

- .000020P 2 - .000001KP
(14.36) Y = 1.8737 - .0014K - .0050P + .06731K- 5
.001440K‘ P* S
S
+ .173513P' 5 -
Equation Value of R Value of t for Regression Coefficients in Order Listed
14.34 .7329* 9.01* 4.29* -- — --
14.35 .8128* 3.31* 9.74* 2.50* 7.31* .42*
14.36 .8793* 1.99* 6.81* 3.85* 10.83* 2.03*
*0 <P< .01
t.01 <P< .05
tp > .05
Production Surface Estimates
Optimizing quantities are shown in Table 14.17 for specified price

combinations. The square root function was used in deriving these
quantities and in providing the production surface of Figure 14.26.
RESPONSE SURFACES IN FERTILIZER “USE 517
Table 14.17. Optimum Rates and Combinations of Fertilizer for Specified Crop
and Nutrient Prices
Price per Unit Optimum Quantity

Price Tons
Lbs. Lbs. Total lbs. Lbs. Lbs. Yield
Situation
k2 o p2 o5 hay nutrient K2 0 P2 O5 (tons)
A .12 .09 $16 71.4 8.0 63.4 3.07
B .12 .09 10 41.0 3.9 37.1 2.84
C .12 .09 22 98.0 12.5 85.5 3.20
D .12 .09 28 120.8 17.2 103.6 3.29
E .09 .12 16 58.8 13.7 45.1 2.99
F .09 .12 10 31.8 6.9 24.9 2.75
G .09 .12 22 84.5 21.0 63.5 3.14
H .09 .12 28 107.5 28.1 79.4 3.24
I .08 .08 10 50.2 7.8 42.4 2.93
Diminishing marginal yields are indicated for either nutrient increased

alone (with the other one fixed at the specified levels of the rows or
columns) or for both nutrients increased in fixed proportion. The pre-
dicted maximum yield is forthcoming (i.e., the marginal products or
first derivatives are zero) at a 3.64-ton yield with 232.2 pounds of P2 O s
and 203.6 pounds of K 2 0.
Figure 14.26. Predicted yield response

surface for alfalfa
Yield Isoquants and Nutrient Combinations
Yield isoquants for alfalfa are shown in Figure 14.27. A 2.5- or

3. 0-ton yield can be attained with P 2 0 5 alone. However, the maximum
P 2 0& for a 3.5 -ton yield is 225 pounds or 90 per cent of the total nutri-
ent application. Hence, the range of possible substitution again declines
with increased yield. For 3.6 tons, the substitution ratio drops below
1.0 with less than 202.9 pounds of P 2 0 5 and more than 100 pounds of
K2 0. The range of replacement possibilities for the 2. 5 -ton yield
ranges only from zero to 120 pounds of K 2 0.
, The slopes of the contour or isoquant lines change at higher yields.
The least -cost fertilizer mixture thus differs somewhat for 2.5-, 3.0-,
or 3.5-ton yields. The isocline labeled RR of 1.0 indicates the path
over which 1 pound of K 2 0 replaces 1 pound of P2 0 5 For a yield of
.
3.2 tons, the intersection point indicates that a replacement ratio of 1:1
is attained with about 78 pounds of P 2 O s and 17 pounds of KÔ. For
points on the 3. 2 -ton yield isoquant above this point of intersection, KgO
substitutes for P 2 05 at a rate greater than 1:1; for points below the in-
tersection point, 1 pound of K 2 0 replaces less than 1 pound of P 2 0 5
.
Similarly, for a 3.6 -ton yield, the replacement rate between nutrients is
1:1 with about 198 pounds of P 2 0 5 and 107 pounds of K 2 0 (i.e., at the
point where the yield isoquant and the isocline of 1.0 intersect). Pro-
portionately more K z O is required at higher yield levels if a 1:1 sub-
stitution ratio is maintained.
The isocline RR =1.5 indicates the nutrient combination for each
successive yield level where 1 pound of K 2 0 replaces 1.5 pounds of
P2 05 Isocline RR = .8 indicates nutrient combinations where 1 pound
.
of K 20 substitutes for .8 pound of .P2 0 5

Again, these isoclines indicate
the most economical combination the least-cost combination) of
(i.e.,
nutrients for any specified yield level. With K 2 0 at 12 cents and P 0
2 5
Figure 14.27. Yield isoquants and

isoclines for alfalfa
at 8 cents per pound, the price ratio Is — or 1.5. The least-cost nu-
trient combination includes 88 pounds of P 2 0 5 and 10 pounds of K 2 0 for
a 3.2-ton yield; it includes 213 pounds of P 2 0 5 and 94 pounds of K 2 0 for
a 3.6-ton yield (although this yield level may not itself be profitable, the
nutrient combination is the one allowing the lowest fertilizer cost for
the particular yield). Since the isoclines “bend” towards the K 2 0 axis,
proportionately more K 2 0 must be used for higher yields if the least-
cost nutrient combination is to be attained. (A single nutrient combina-
tion would provide the least -cost combination for all yield levels only if
were straight lines passing through the origin.) The iso-
the isoclines
clinesconverge to a single point denoting the maximum possible yield;
replacement of one nutrient by the other is not possible for 3.64 tons of
alfalfa.
Optimum Rates for Alfalfa
Optimum rates of fertilizer application for alfalfa are predicted in

Table 14.17 from equation 14.36. These estimates were obtained by
setting the partial derivatives of hay yield in respect to K and P to equal
the appropriate price ratios and solving for the magnitudes of the vari-
ables. Only this type of optimum is considered for alfalfa; although the
alternatives considered for corn are equally appropriate.
Under price situation A, the nutrient input includes 87.4 per cent
P 2 0 5 and 12.6 per cent K 2 0. With a fall in hay price from $16 to $10 (B)
and nutrient costs remaining the same, only 41 pounds of total nutrients
should be used. However, the total nutrient input now should be com-
posed of 90.5 per cent P 2 05 and only 9.5 per cent K 2 0. An increase in
hay price to $28 (D) requires an increase to 120.8 total pounds of nutri-
ents for economic optimum. The 120.8 pounds is composed of 85.8 per
cent P2 Qj and 14.2 pounds of K 2 0.
RESIDUAL RESPONSE FUNCTIONS FOR CORN
Residual responses also are important in determining the economic

optimum use of fertilizer. Preceding discussions related to responses
in the year following fertilizer application. For the corn experiment on
Ida silt loam it was possible to obtain second-year yields; the land was
planted back to corn and no fertilizer was applied in the second year.
Hence, the second-year response functions for 1953 reported below are
due alone to the “carry-over” effect of fertilizer applied in 1952.
Response Functions and Related Data
Two methods were used in analyzing the 1953 corn response data:
(a) total response surfaces for the 2 years were computed by adding the
1953 yields to the 1952 yields and these combined

fitting functions to
data, including zero application rates and functions were fitted to (b)
the 1953 “carry-over* yields alone. The predicted production functions

are given below where all observations, including zero rates, are used;
Two-year total (1952 and 1953 data pooled and functions fitted to
total yield of 2 years):
3
(14.37) Y = - .0965 + .6464N + .8140P - .00176N
2
- .00231P + .00149NP
(14.38) Y = 12.636 - .2213N - .4614P + 4.2464N* 5

5
+ 8.7506P' + .5603N' 5 P’ 5
Second-year residual (1953 yields only):
(14.39) Y = 7.4177 + .0621N + .1502P - .000180N 2

- .000511P 2 + .000683PN
(14.40) Y = 18.317 + .0948N - .0440P - 2.1047N’

5
*5
+ .2352P + .2193N ^P*5
The R 2 values
are: quadratic 2-year total (14.37), 88 per cent;
square root 2-year total (14.38), 92 per cent; quadratic residual
(14.39), 81 per cent; and square root residual (14.40), 77 per cent.
Equation Value of R Value of t for Coefficients in Order Listed in Equations
14.37 .940 8.82* ii.il* 8.65* 11.33* 8.37

14.38 .961 4.11* 8.58* 3.63* 7.49* 10.80
14.39 .900 T
2.00 4.84* 2.08* 5.93* !
9,03
14.40 .878 3.09* 1.43*
*0 <P< .01
_ 3.16*
~ —.35 § 7.4T
t.01 <P< .05
J.10 <P< .20

§.50 <P
Although the square root function did not fit the residual
data quite
so well as the quadratic, it did fit the 2 -year total yields
better than the
crossproduct. Since the 2-year total yields combine both the
first- and
second-year response, the square root function was chosen for
use in
the following economic analysis.
If the coefficients of the second -year
residual are added to the cor-
responding coefficients of the first-year function, the result
is equal to
the 2 -year total functions given above. For example,
the coefficients of
Table 14.19. Analysis of Variance for Total or 2- Year and for

Second Year Residual Response Functions
Degrees of Sum of Mean

Source of Variation Freedom Squares Square
Quadratic 2- year total
Total 113 466,207.68
Treatments 56 450,794.90 8,049.91

Due to regression 5 411,938.30 82,387.66
Deviations from regression 51 38,856.60 761.89
Among plots treated alike 57 15,412.78 270.40
F * 305
Square root 2- year total
Total 113 466,207.68
Treatments 56 450,794.90 8,049.91

F = 318
Quadratic second year residual
Total 113 51,216.90
Treatments 56 45,742.37 816.83

F = 86.4
Square root second year residual

Total 113 51,216.90
Treatments 56 45,742.37 816.83

F = 82.2
equation 14.40 plus those of equation 14.3 equal those of equation 14.38.
This is to be expected; the production surface for 2 years is the sum of
the surfaces for the first and second years.
Yield Isoquants
Yield isoquants were derived from the basic production function in

the manner outlined in previous sections. The isoquants for the 2 -year
production surface denote diminishing productivity over the entire
522 RESPONSE SURFACES EST FERTILIZER USE
LBS. OF NITROGEN
4
Figure 14.28. Corn yield isoquants for square root residual
function 14.40, for second year alone
surface. The segments on scale or fixed ratio lines which are inter-
sected by yield isoquants (representing equal increments in yield) be-
come greater for higher yield levels. In the case of the residual or
second-year response surface, however, the scale lines show slightly
increasing returns for small applications of fertilizer (see Figure 14.28
and Table 14 20 ). However, decreasing returns in the second year
.
might have occurred if heavier fertilization had been used in the first
year. Also, the second-year response may have, first, a stage of in-
creasing returns and, second, a stage of decreasing returns.
For economic decisions, residual response must be considered in
conjunction with the first-year response. Therefore, isoquants for the
Table 14.20. Isoquant Combinations for Producing Specified Yields;

Residual Function 14.40
40-Bushel Yield 50-Bushel Yield 60-Bushel Yield 70-Bushel Yield

Lbs. Lbs. Lbs. Lbs. Lbs. Lbs. Lbs. Lbs.
p2 o5 N P2 o5 N P2 Os N p2 o 5 N
20 489.0 50 465.9 50 586.3 50 706.0
50 342.5 80 378.5 80 487.9 80 596.9
80 268.4 100 338.6 100 441.6 100 544.9
100 236.0 150 272.2 150 362.2 150 453.7
150 184.6 200 231.8 200 311.8 200 394.1
200 155.4 250 205.1 250 277.2 250 352.2
250 137.2 300 186.5 300 252.2 300 321.2
300 125.2 350 173.0 350 233.6 350 297.6

400 111.1 400 163.0 400 219.4 400 279.2
-
Figure 14.29. Corn yield isoquants and isoclines for

2-year total square root function 14.38
2-year total response surface are given in Figure 14.29 and Table
14.21. Five isoclines are presented along with four isoquants in Figure
14.29. The center isocline, RR = 1.5, is appropriate for an N to P 2 0 5
price ratio of 1.5, approximately the present price relationship. If the
price of N were twice that of P2 0 5 the isocline labeled RR = 2.0 would
,
be the appropriate one. If some new process should make nitrogen one-
half the price of available P 2 0 5 then the isocline with RR = .5 would be
the one to be followed to maximize profits (considering both the first
and second -year response). Of course, these isoclines will depend upon
or differ with the soil type and fertility level. For other soil types and
fertility conditions, a different production surface would be expected. It
is obvious for the isoclines presented that any price ratio which would
require expansion of fertilizer use along the isocline RR = .5 would give
Table 14.21. Isoquant Combinations for Producing Specified Yields; 2-Year Total
Function 14.38
50-Bushel Yield 100-Bushel Yield 150-Bushel Yield 200-Bushel Yield
Lbs. Lbs. Lbs. Lbs. Lbs. Lbs. Lbs. Lbs.

P2 O s N Pa 0 6 N p 2o 5 N P2 O s N
3 37.2 30 117.4 80 361.0 180 515.7
5 19.8 50 54.6 100 207.8 200 412.5
10 6.9 80 33.2 120 165.0 250 325.1
15 3.0 120 25.6 150 135.8 300 294.4
20 1.3 150 24.3 200 117.4 350 283.0
30 .2 180 24.7 250 113.1 400 281.3

about the same economic results as increasing nutrients in a P/N ratio

of 1.7. This isocline deviates only slightly from a straight line through
the origin. Also, some of the other isoclines have only slight curva-
ture, denoting only slight profit depression if a fixed nutrient combina-
tion is used for increasing yield. Rates of fertilization were not great
enough to define the point of maximum yield and convergence of iso-
clines for the 2 -year total yield within the range of the experiment.
Economic Optima
Using the 2-year total function (equation 14.38), the optimum inputs
of nitrogen and phosphorous can be found in the way shown in the pre-
ceding sections. Optimum inputs are determined by equating marginal
physical products with their corresponding factor -product price ratios.
This optimum solution is valid in the 2 -year case only when the ex-
pected price of corn is the same for both years and the farmer does not
discount the expected value of the second crop more than the first crop.
However, farmers generally discount the value of distant crops
more than current crops. In this case, the problem can still be solved
by adding the discounted residual production surface to the first-year
surface. As an example, assume the farmer expects the price of corn
to be $1.25 per bushel for both the first and second year. He discounts
the value of the second crop by 20 per cent due to uncertainty or other
reasons. This makes the present value of the second -year corn worth
$1 per bushel. The response function is now the first-year response
coefficients plus .80 times the second -year coefficients.
Optimum inputs of N and P for various prices of corn are in Table
14.22. The corn prices in this table are assumed to be the present dis-
counted values of the crops at the time fertilizer is applied.
A more important problem is to determine the optimum amount of
Table 14.22. Optimum First- Year Applications of Fertilizer,
Considering the Response From Both the
First and Second Years for Specific
Price Relationships*
Optimum Fertilizer
Price Situation Application
First- Second-
Year Year
Corn Corn ^2 O5 N Pounds Pounds
per Bushel per Bushel per Pound per Pound N p2 05
$1.50 .75 .10 .15 290.1 259.6
1.25 .625 .10 .15 241.5 226.6
.625 .3125 .10 .15 101.5 120.8
.50 .25 .10 .15 72.6 95.1
Prices of corn are assumed to be the present discounted value to the

farmer at the time fertilizer is applied.
fertilizer to be applied in the second year. This can not be answered

from the present data. One hypothesis is this: the optimum application
for the second year will drop back to considerably less than that for a
single year. More research regarding the relation of soil fertility to
fertilizer response appears necessary before some of these problems
5
can be adequately answered .
In this section and in previous ones, the appropriate economic prin-

ciple has been applied in specifying optima. It is recognized, of course,
that uncertainty and other factors do not allow farmers to be so precise
in their decision making. The purpose of this study, however, has been
to apply appropriate methods. Mechanical or “rule of thumb” proce-
dures can be developed for applying these basic principles with only
slight depression of profit, once additional research provides added in-
formation on basic response functions.
LIMITATIONS AND EXPERIMENTAL NEEDS
The concepts and analytical procedures employed in this study are

basic for determining economic optima in the use of fertilizer. Also,
they provide the basic physical or structural relationships of crop re-
sponses in relation to fertilizer application. The predictions apply to
particular soils in a particular year; production surfaces obtained
under other rainfall and soil conditions can be expected to differ from
those obtained in the experiments reported. These limitations are not,
however, unique to the type of experiment and empirical procedure re-
ported here. Traditional experimental procedures (wherein a few rates
of one or more nutrients are applied) also refer to the rainfall, cli-
matic, insect, and crop conditions of the particular year.
5
See Jensen, D. and Pesek, J. Generalization of yield equations in two or more variables.
Agronomy Jour., 51: 255-63. 1959.
CHAPTER 1 5
John T. Pesek Surfaces, Isoquants, and Iscolines

Earl O. Heady
From Fertilization
T CHAPTER summarizes several fertilizer production function

HIS
studies in Iowa. The several sets of predictions are provided to
illustrate the types of physical relationships which arise from
various functional forms and magnitudes of regression coefficients
which appear most efficient in response predictions under various en-
vironmental conditions.
The analysis also illustrates the types of empirical studies which
often can be made from limited resources for experimentation or from
data which already exist. Both situations held true for some of the ex-
periments reported in this chapter. However, experiments of recent
years emphasize the composite and rotatable designs discussed in
earlier chapters.
Not all experiments reported provide such “neat* functions as the
corn experiment reported in Chapter 14. However, this is the situation
which faces the farmer: response may be large under favorable rain-
fall but small or nonexistent in dry years. Hence, it is as important to
predict response under conditions where an experiment may provide
few or no significant regression coefficients as it is to conduct experi-
ments only for years where response is great. A variety of conditions
in respect to nutrient combinations, stand levels, and harvesting are re-
ported in this chapter.
CARRINGTON EXPERIMENT
This section reports a corn experiment on Carrington silt loam in

1953. The factorial experiment providing the data consisted of two
randomized blocks, each block having five levels of N, four levels of
P 2 0 5 and three levels of K 2 0. Yields were high in this experiment;
plots without fertilizer averaged almost 98 bushels per acre. Large
yield responses for fertilizer were not expected since the soil was at a
relatively high fertility level (i.e., high yields were obtained on the
check plots). However, an average increase of 9.8 bushels per acre
*This chapter summarizes a series of studies made by the authors in co-operation with
William G. Brown, John P. Doll, Owen W. McCarthy, R. P. Nicholson, and Joseph A. Stritzel.
SURFACES FROM FERTILIZATION 527
Table 15.1. Analysis of Variance of Corn Yields on Carrington

Soil, Randomized Block Design
Source of Degrees of Sum of Mean

Variation Freedom Squares Square F
Total 119 18,570.37

Blocks 1 6,336.53 6,336.53 106.39*
Treatments 59 8,719.703 147.80 2.48*
N 4 514.042 128.51 2.16
P 3 79.285 26.43 .44
K 2 4,198.086 2,099.04 35.24*
Nx P 12 523.262 43.61 .73
Nx K 8 630.862 78.86 1.32
Px K 6 870.996 145.17 2.44*
Nx Px K 24 1,903.170 79.30 1.33
Error 59 3,514.137 59.56
*P < .01
was obtained from 40 pounds of K 2 0. Application of 80 pounds of K 2 0

resulted in an average increase of 14.2 bushels over the plots with no
potash. The significant potassium effect (Table 15.1) might have been
anticipated since the experimental plot was, according to soil tests, low
in & 2 0.
Values of F in Table 15.1 provide information for variables to be
used in the estimating equation or production functions which follow.
Potash should be included since K 2 G gives a consistent and statistically
significant increase in yield. Phosphorus can be dropped from consid-
eration because, even if all of the mean square due to P were explained
by one regression term, its F value would not be significant. Nitrogen
is an intermediate case; there is some logical justification for including
it even though it is not significant at the .05 level of probability. Phos-
phorus x potash interaction is significant at the .05 level. However, it

was not included in the regression because no term was found which
would significantly account for the variance in yield due to this term.
An analysis of covariance indicated that stand had a highly significant
effect on yield. Similar results were obtained when stand was included
as a variable in the multiple regression.
Regression Analysis for Carrington Soil
The basic purpose of this study is to estimate crop yield production

functions for fertilizer. Accordingly, information on the derivation of
the regression equations is included in this section. In the preliminary
analysis for each experiment, two general types of equations were used:
a quadratic equation and a square root equation.
The highly significant difference between the yields of the two ran-
domized blocks (Table 15.1) raised a question as to whether the re-
sponse surface differed significantly between the two blocks. To test
528 SURFACES FROM FERTILIZATION
Table 15.2. Values for t for Coefficients of Individual Block Regressions and Test of Difference
Between Corresponding Coefficients of the Two Blocks
Values of
t for Values of t
Values Values
Difference for Pooled
of t for Significance of t for Significance Significance
Coefficient Between Regression Significance
Equation Level* Equation Level* Level* Level*
Equations Equation
15.1 15.2
15.1 and 15.3
15.2
K 3.515 .001 2.006 .06 .739 .47 4.118 .0001
K2 2.059 .05 .810 .43 .834 .41 1.965 .06
N .737 .48 .996 .33 .211 .84 .316 .20
N2 .719 .48 .620 .55 .045 .92 1.030 .31
S 3.757 .001 2.554 .02 .731 .46 4.556 .00002
B — — — - -- — 9.570 .00001
Probability of obtaining as large or larger value of t by chance, given the null hypothesis.
tThese t’s have been computed by subtracting each particular regression coefficient m equation 15.2
from the corresponding regression coefficient in equation 15.1 and dividing by the weighted standard
error.
whether the response differed between blocks, regressions were calcu-

lated for each block separately as indicated in equations 15.1 and 15.2.
5
(15.1) Block I: Y = 57.97 + .3800 K - .002711 K2 + .4365 N’ - .02638 N + .002552 S
5
(15.2) Block II: Y = 51.64 + .2702 K - .001162 K2 + .6414 N* - .02490 N + .002081 S
In the above equations, Y refers to predicted total yield in bushels

per acre, K refers to pounds of K 2 O per acre, N to pounds of elemental
nitrogen per acre, and S refers to stalks per acre. The t values of the
regression coefficients are given in the left half of Table 15.2. To
help determine whether the two blocks should be pooled, t tests of the
differences between corresponding regression coefficients were made
(Table 15. 2). 1 The t values for the difference between corresponding
regression coefficients of the two blocks are small. A value of t as
large or larger than the t value of difference for K2 , t = .834, could oc-
cur by chance 40 per cent of the time even though the population of K2
coefficients was the same. The other coefficients had even smaller t
values of difference. Since there was no evidence that the blocks had
different response surfaces (different regression coefficients), a re-
gression for the pooled data of the two blocks was computed as indi-
cated in equation 15.3.
2
(15.3) Y a 77.866 + .3162 K - .001813 K + .9190 N*
5
- .04453 N + .002241 S - 13.497 B
In the pooled regression above, B represents the particular block;

B is 1.0 for Block I and 2.0 for Block 31. Stand and block are used here
An analysis of variance was also computed to test for homogeneity of regression; the
results were similar to those obtained from the t tests.
as a method of adjustment similar to covariance. The experiment was

not designed to
include stand as a variable, but variation in stand did
occur. This experiment cannot be used to determine optimum stand.
However, precision of estimates is considerably improved by including
stand in the regression (as shown by its t value of 4.56 in Table 15.2).
Blocks were included in the regression to allow an estimating equa-
tion for either block and to increase precision of estimation. Including
blocks in the regression is justified since it takes out the variability
due only to the difference in stand and yield level of the two blocks.
Predicting the actual yield is secondary to predicting the response of
corn yield to fertilizer inputs. That is, more interest is in the slopes
of the production surface rather than the absolute level of yield. The
values of the N and K coefficients are important in determining the
most profitable amount of nitrogen and potash to apply. Stand and
blocks were introduced only to increase the precision of estimate of the
N and K coefficients.
For an average stand and for Block I, the intercept (N = 0; K = 0) of
equation 15.4 becomes 105.971. For an average stand and Block n, the
intercept is 92.474. Equation 15.4 is the average of the two blocks with
an average stand of around 18,000 stalks per acre and will be used in
the later economic analysis.
(15.4) Y = 99.223 + .3162 K - .001813 K 2

+ .9190 N* 5 - .04453 N
The value of t (4.118) for the linear response of yield to potash in
Table 15.2 is highly significant.Accordingly, greater reliability can be
placed in the potash response than in the nitrogen response.
The analysis of variance for the regression estimates is presented
in Table 15.3. The F value of 33 indicates that the proportion of vari-
ance explained by the regression equation 15.3 is highly significant.

However, only 65 per cent of the total sum of squares is accounted for
by equation 15.3.
Table 15.3. Analysis of Variance for Regression of Corn

Yield, Carrington Soil

Total 119 18,570.37
Due to regression
equation 6 12,013.09 2,002.18 33.04*
Deviation from
regression 113 6,557.29 58.03
Other treatment
effects 55 3,043.15 55.33
Error 58 3,514.14 60.59
*P 1.01
.
Production Surface for Carrington Soil
Equation 15.4 is used for the economic analysis of the experiment

on Carrington soil. Since the soil was fertile, yields are predicted to
start at 99 bushels per acre with no fertilizer. A yield of almost 118
bushels per acre is predicted at 80 pounds each of N and K2 0. The fig-
ures indicate ranges of both increasing and decreasing total yields (i.e.,
positive and negative marginal products)
A geometric view of the predicted production surface is provided in
Figure 15.1. The slope of the surface indicates the response to both N
and K2 0. The slope is greater along the K2 Q axis than along the N axis.
However, the surface in general is quite flat.
A slice through the surface parallel to the potash axis in Figure 15.1
would represent response of corn to K a O at a fixed level of N. Individ-
ual yield response curves to potash for given levels of nitrogen remain
the same distance from each other. This lack of “interaction” between
N and K 2 0 was probably a characteristic of the experimental site. Pre-
vious experiments have provided production surfaces with important
interactions between fertilizer nutrients. However, N and K may inter-
act less with each other than N does with P or P does with K.
Figure 15.2 shows the 95 per cent confidence limits of the yield es-
timates for K 2 0. The spread at the ends of the curve is due to the in-
creased distance from the mean, as the response is extrapolated beyond
the 80 pound limit of the K2 0 application in the experiment. Confidence
intervals for the N response are also relatively narrow, indicating some
degree of precision in estimation.
Marginal physical products of N remain the same at all levels of
K2 0 because there is no interaction between N and K 2 0 in equation 15.4.
Conversely, the marginal physical product of K 2 0 is not affected by the
level of N. This is illustrated in the marginal product equations for N
and K 2 0:
Figure 15.1. Perspective view of predicted yield surface for corn

on Carrington soil predicted from equation 15.4
Figure 15.2. Confidence limits for corn response to K 2 0

at 104 pounds of N, Carrington soil (Dashed
vertical line is limit of K 2 0 in experiment.)
6Y
(15.5) .3162 - .003626 K
5K
6Y 5
(15.6) -.04453 + .4595 N"' .
6N
The marginal yields from K correspond to the slope or incline of

the “roof* in Figure 15.1 parallel to the N axis. At 0, 20, 40, 60, 80,
K
and 100 pounds of 2 0, the marginal yields are .32, .24, .17, .10, .03,
and -.05 bushel, respectively. Similarly, marginal yields for N, com-
puted from equation 15.6 for 1, 20, 40, 60, 80, and 120 pounds of N, are
and -.003 bushel, respectively. It can be
.41, .06, .03, .01, .007, .001,
seen from the N marginal yields that while N returns a fairly large in-
crease in yield at small inputs the response soon levels out. Negative
marginal products for either N or K indicate that further inputs at the
particular levels cause a decline in total per -acre yield.
Yield Isoquants and Isoclines for Carrington Soil
Yield isoquants in Figure 15.3 are derived from equation 15.7.
5
.000323 N + .8194 .00725 Y
-V'.00666N-
^ - -
(15.7) K = 87.23 ±
.003626
Figure 15.3. Isoquants and isoclines with dashed

ridgelines, Carrington soil
The corresponding isocline equation k is the K 2 0/N price ratio
(15.8) K = 87.20 + 12.28 k - 126.72 N~’ 5 k.
Isoclines are presented as the positively sloped curves in Figure

15.3. These are bent close to the K s O axis, denoting a relatively
greater deficiency of this nutrient in obtaining yield response. However,
for the highest yields, both nutrients become relatively deficient.
The dashed lines in Figure 15.3 represent ridgelines which denote
the economic limits of the isoclines. The ridgelines define the portion
of the production surface included between the extremes of zero (or in-
finite) substitutionrates for nutrients. (In other words, they are iso-
clines with zero substitution ratios, indicating the extreme limits of
nutrient substitution in obtaining specified yields.) The ridgelines (iso-
clines of zero substitution rates) indicate the boundaries of the surface
with positive slopes along both input axes; beyond the ridges, one or
both slopes are negative. If nitrogen were free in price but 2 0 were K
not, would pay to expand production along the top ridgeline, always
it
applying 106 pounds of N and purchasing K 20 according to its cost and

return. On the other hand, if potash were free and nitrogen were not,
production should be expanded along the right hand vertical ridgeline.
Since N and K 2 0 were independent in basic surface equation 15.4, the
ridgelines are straight and meet at a right angle.
All the isoclines (including ridgelines) converge and intersect at the

point of maximum physical product. If both N and K 2 0 were free and
cost nothing to apply, inputs should be extended to 87.2 pounds of 20 K
and 106.5 pounds of N, the point of isocline convergence. A maximum
physical yield of 117.76 bushels is predicted from these inputs of N and
k 2 o.
Carrington Soil Presentation for Practical Use
While the main purpose of this study is that of dealing with certain
basic or methodological aspects of fertilizer response and economics,
it is useful to indicate how the results can be presented for farmers or
extension personnel. Since N and K s O effects were independent in the

production function, equation 15.4, the optimum rate for N can be se-
lected without regard to the level of K 2 0 and vice versa. To find the
optimum input of either nutrient, divide the price (per pound) of the nu-
trient by the price of corn. Selection of the corresponding ratio from
one of the charts in Figure 15.4 then provides the optimum input.
Assuming N to be $ .15 and K 2 0 to be $ .08 per pound with corn at
$1.00 per bushel, the appropriate N/C price ratio is .15 and the appro-
priate K/C price ratio is .08. Use of these ratios in Figure 15.4 indi-
cates an optimum input of about 65 pounds of K 2 0 and 6 pounds of N.
The gain in yield from these inputs also can be estimated. A gain of
about 13 bushels per acre from use of K 2 0 and about 1.5 bushels from
the N application Is predicted. Of course, such a chart should be used
for a Carrington soil with fertility similar to the experimental field.
Rainfall and biological conditions also would need to be as favorable as
for the experimental field in 1953.
Figure 15.4. Added bushels from Kz O and N and optimum

rates for specified price ratios of fertilizer
nutrients and corn, Carrington soil
MOODY EXPERIMENT
The experiment reported in this section was made on Moody soil
in 1953. Cropping history and soil tests indicated low availability of ni-
trogen and phosphorus, and high availability of potassium. Large re-
sponses in corn yields were obtained by adding nitrogen; in fact, yield
was more than doubled by applying 40 pounds of N. Further increases
in yield were given by 80 and 160 pounds of N. However, with 240
pounds of N, a slight decline resulted. Potassium had little effect on
yield. Phosphorus also seemed to have only a small effect since yield
was increased by less than 8 bushels in rates ranging from 0 to 120
pounds. However, examination of the average response to P2 Ofe over
all levels of N and K 2 0 hides part of the actual effect. Actually, as
careful examination of the yields indicates, there was a strong inter-
action between phosphorus and nitrogen. At zero level of N, P 2 0 5 had
a depressing effect on yield, but at 160 and 240 pounds of N it increased
yield.
Table 15.4. Analysis of Variance of Corn Yields on Moody Soil,

Randomized Block Design

Total 117 53,164.96

Blocks 1 432.82 432.82 4.57*
Treatments 59 47,330.08 802.20 8.46t
N 4 41,000.49 10,250.12 108. 15t
P 3 1,051.22 350.41 3.70*
K 2 123.05 61.53 .65
Nx P 12 3,333.88 277.82 2.93t
Nx K 8 455.83 56.98 .60
PxK 6 99.55 16.59 .18
NxPx;K 24 1,266.06 52.75 .56
Error 57 5,402.06 94.77
*P <.05
'P <.01
The analysis of variance in Table 15.4 confirms the highly signifi-

cant effect of nitrogen. The effect of P 2 0 5 was significant at the .05
probability level. Interaction between N and P 2 0 5 was highly signifi-
cant. There also was a significant difference between the yield levels
of thetwo randomized blocks. Potash and the remaining interactions
accounted for no significant portion of yield variance. The lack of pot-
ash response was expected since the experimental plots tested high
in K. Since the soil test for P was low, negative phosphorus response
at low levels of N (see above) was unexpected. However, other evidence
suggests that this can happen as a result of aggravating the nitrogen
deficiency.
Regression Analysis for Moody Soil
Several algebraic forms of the yield predicting equation were tried

2
before equation 15.9 was selected. Equation 15.9 had an R of .827 and
was logically more acceptable than certain other forms.
2 2
(15.9) Y = 13.543 + .5340N - .001743 N - .0003549 P + .001069 NP + .000873 S.
In this equation, Y refers to total yield in bushels per acre, N re-

fers to pounds of nitrogen per acre, P refers to pounds of P 2 0 5 per
acre, and S refers to stalks per acre. Values of t for the coefficients
in the order that they appear in the equation are 12.56, 14.47, 1.68,
2
5.44, and 1.50. The preceding t values for N, N and NP are signifi- ,
2
cant at the .00001 level of probability. The terms for P and S are sig-
nificant at .10 and ,14 probability levels, respectively, and are retained
2
for logical reasons. The negative P term is important because it
forces diminishing returns to phosphorus inputs. Some of the functions
fitted to the data, or the particular function without this term, did not
have this characteristic. For example, the full five -term square root
or regular quadratic functions gave increasing returns for part of the
production surface; increasing returns make it difficult to secure de-
terminate economic solutions. The five -term quadratic equation was
as follows:
2 2
(15.10) Y * 30.277 + .533 N - .00175 N - .623 P + .000066 P + .00116 NP.
Equation 15.10 was rejected in favor of equation 15.9 since the

latter gave diminishing returns and a determinate predicted maximum
yield. The stand variable was included in equation 15.9 to increase the
precision of fit of the nutrient response curves; equation 15.11 has been
adjusted to an average stand and was used for the subsequent economic
analysis. Equation 15.11 is the same as equation 15.9 except that aver-
age plot stand is fixed at 18,000 stalks per acre. With the coefficient
for S significant at the .14 probability level, the writers adopted this
procedure as being more efficient than the conventional procedure of
adjusting individual plot yields for stand.
2
(15.11) Y = 29.248 + .5340 N - .0001743 N - .0003549 P2 + .001069 NP.
The analysis of variance of the basic regression, equation 15.9, is

given in Table 15.5. The F value of 91.20 for the over -all regression
is highly significant. The mean square for deviations from regression
is smaller than the within-plot estimate of experimental error.
Production Surface for Moody Soil
Estimated yields were predicted from equation 15.9 to provide the

corn production surface for the Moody soil.

Yield on Moody Soil

Total 117 53,164.96

Blocks 1 432.82 432.82 4.49*
Due to regression,
43,991.51 8,798.30 91.20t
equation 15.9 5
Deviations from
regression 111 9,173.45 82.64
Other treatment
effects 55 3,771.39 68.57
Error 56 5,402.06 96.47
*p <: .05
tp^.01
The interaction of nitrogen and phosphorus can best be seen from

the surface in Figure 15.5. Yields increase sharply as nitrogen is ap-
plied at the zero level of phosphorus. However, even higher yields are
obtained from as P2 05 is increased. Yield at zero level of P2 O s for
N
different rates ofN is represented by the edge of the surface directly
above the nitrogen axis. A second line over the surface parallels the
first and shows yield response to N at 40 pounds of P 2 O s Thus, the .
strong positive interaction or complementarity of N and P can be seen

from the high center ridge of the surface at large inputs of N and P.
Figure 15.5. Perspective view of predicted yield

surface for corn on Moody soil
Yield Isoquants and Isoclines for Moody Soil
In the yield isoquant equation 15.12, derived from the basic regres-
sion equation 15.11, P2 0 5 is expressed as a function of N. The yield
isoquants in Figure 15.6 are based on equation 15.12.
Figure 15.6. Yield isoquants and isoclines with

dashed ridgelines, Moody soil
2
„ + V.0007581 N - .000001332 N + .0415 - .00142 Y
(15.12) P = 1.506 N -
;
0007098
For yields as low as 50 or 60 bushels in Figure 15.6, isoquants are

nearly vertical. These steep slopes for lower yields mean that many
pounds of P 2 0 5 are necessary to replace 1 pound of nitrogen in attaining
the constant yield (or, practically, that added P2 0 5 does not substitute
for N in attaining these yields when N input is low). As yield is in-
creased to 70 bushels per acre and more N is used, the isoquant be-
comes more curved as it approaches the N axis. The 70-bushel iso-
quant intersects the N axis; 70 bushels per acre are predicted from the
equation with all nitrogen and zero of V2 0 5 However, the 80-bushel
.
isoquant requires P 2 0 5 in addition to N; a yield this high requires the

complementary effect of P2 0 5 with N.
Since the slopes of the isoquants show the change in amount of P 2 0 5
required to maintain a given yield when another unit of N is added, the
curvatures of the. isoquants indicate the change in the rate of substitu-
tion of N for P2 0 5 . Substitution or replacement rates predicted from
equation 15.12 for a 70 -bushel yield show that at 100 pounds of N, an
additional unit of N replaces about 3.5 units of P2 0 5 However, at 140
.
pounds of N, an additional pound of N replaces only about one -third of a

pound of P2 0 5 , if an 80 -bushel yield is to be retained. Since the slopes
of the isoquants change along a scale line (fixed nutrient combination)

the combination of nutrients or fertilizer ratio which gives lowest cost
for one yield level is not the same fertilizer ratio which gives lowest
cost for another yield level.
Each isocline in Figure 15.6 intersects every isoquant at a point of
specified slope on the isoquant. For example, the isocline labeled P n
= 3.0 P , meaning a price of nitrogen equal to three times the price of
phosphate, goes through each of the 70-, 80-, and 90-bushel isoquants at
points where the slope (i.e., the marginal rate of substitution) is 3:1.
On the isocline labeled P n = .33 P p each isoquant is intersected where
,
the slope is 1:3. On this isocline, each pound of P2 0 5 would replace

3 pounds of N. Therefore, if the price of N were one -third the price of
P2 0 5 per pound, production should be expanded along the isocline la-
beled P n = .33 P p if the path of fertilizer ratios for least -cost yields
,
is to be traced out. Relevant isoclines in Figure 15.6 all intersect the

nitrogen axis. This suggests that for economical yield increases, ni-
trogen is the most limiting factor at the outset. However, for higher
yields which might be profitable, P 2 0 5 is relatively more limiting,
since the isoclines have slopes greater than 45 degrees.
HAYNIE EXPERIMENT
An experiment on Haynie soil in 1953 is reported in this section.

Rates of application were 0, 40, and 80 pounds of each nutrient Im-
proved estimates of the N and 05 response would have resulted if N
and P 2 0 5 levels had extended higher. Average yields of the nitrogen
responses are 8 bushels for the last 40 pounds of N applied. Similarly,
80 pounds of P2 0 5 gave almost 5 bushels more corn per acre than did
40 pounds. To estimate the N x P interaction, nitrogen and phosphorus
inputs should go far enough to cause a decline, or at least a leveling
out, of total yield. Additional increases in yield might have been ob-
tained at heavier N and P 2 0 5 combinations.
Of several algebraic forms examined, equation 15.13 was selected
as most efficient for prediction.
(15.13) Y = -.9751 + .7126 N - .004352 N2 + .5255 P - .003103 P2

+ .2546 K - .001624 K2 - .002255 PK + .003863 S
Rather than adjusting yield for stand, plant population was included as a
variable because analysis of covariance indicated a highly significant
stand effect. Ideally, stand should have been introduced as an experi-
mental variable. However, this was not the case and stand levels in
different plots varied only through weather and chance. Experiments
are reported later where stand was included as a variable. If inputs of
N and P2 05 had been extended to higher levels, the square root function
might have given better results. Higher levels of N and P2 O s might
have also revealed a significant N x P interaction.
Table 15.6. Values of t for Individual Regression Coefficients

of Equation 15.13
Variable N N2 P P2 K K2 PK S
t value 5.67 2.89 3.72 1.96 1.90 1.06 2.10 2.99

P level .00001 .005 .0004 .05 .07 .30 .04 .004
The term S again refers to stalks per acre while N, P, and K refer
to pounds of N, P2 O s , and K 2 0 per acre. The values of t for the re-
gression coefficients are given in Table 15.6. They show that the coef-
ficients for the N variables are significant in explaining yield variance.
2
The value of t for K is only 1.06. A value this large could occur by
2
chance in about one -third of the time where K had no real effect.
Nevertheless, the K term is retained for logical reasons. Without the
2
negative K term, an unlimited linear response to K would be implied.

2
The analysis of variance of equation 15.13 regression in Table 15.7

shows the over-all regression to be highly significant. The deviations
from regression mean square are about the same as the estimate of
experimental error from within plots.

Yield on Haynie Soil

Total 53 13,876.57
Due to 8- term
regression
equation 15.13 8 10,794.66 1,349.33 18.92*
Deviations from
regression 45 3,081.91 68.49
Other treatment
effects 20 1,299.43 64.97
Error 25 1,782.48 71.30
*P< .01
In equation 15.14 stand is fixed at 9,220, the average stalk count for
all experimental plots. If stand were included as a controlled variable
in the original experiment, the optimum level of stand could be deter-
mined by economic analysis. However, none of the experiments ana-
lyzed in this study was so designed, and stand is used only to improve
the precision of estimate of the fertilizer response.
(15.14) Y = 35.0587 + .7126 N - .004352 N2 + .5255 P - .003103 P 2
+ .2546 K - .001624 K2 - .002255 PK

Production Surfaces
Equation 15.14 allows production of an infinite number of three di-

mensional production surfaces when one nutrient is held fixed and the
other two are variable. Since the effects of N were predicted to be in-
dependent of P and K, the phosphorus -potassium surface retains the
same shape or curvature at different levels of N. In Figure 15.7, the
surface shows a greater rise in yield from inputs of P 2 O s than from
K 20. Also, the surface is relatively flat over the top, indicating that
yields do not change greatly for many combinations of P 2 O 5 and K 2 0.
For a line stretched diagonally over the surface from the zero corner
to the opposite corner, a sharp increase in yield is followed by a de-
crease. The dropping off at the opposite corner for high levels of both
P 2 O s and K 2 0 is due to the negative PxK
interaction. The decline at
high levels of P 2 0 5 and K 2 G is in contrast to the high ridge at high
levels of N and P 2 0 5 in Figure 15.5, caused by positive N x P interac-
tion.
Yields are predicted to increase by about 21 bushels as nitrogen in-
puts are increased to 40 pounds. For 80 pounds of N, predicted yields
are increased 29 bushels over corresponding P -K treatments receiving
no nitrogen. These relationships between nitrogen and P-K responses
for a particular angle of the P-K yield surface are shown in A, B, and C
of Figure 15.7. In A, the P-K yield surface is shown with a zero level
of N; B gives P-K yields with 40 pounds of N; C shows P-K yields with
80 pounds of N. The surface of B is 21 bushels higher than for A be-
cause of the response of the 40 -pound application of N. The shapes of
the three surfaces are exactly the same, but one may see more of the
underside of the declining surface in the higher structures. If the P-K
yield surface for 120 pounds of N were shown it would be of the same
height as in B, since predicted yields start to decline around 82 pounds
of N.
Figure 15.7. (A) Predicted yield surface for corn on Haynie soil with no N
application, ((B)! Predicted yield surface at 40 pounds of N, (C)
Predicted yield surface at 80 pounds of N
Isoquants and Isoclines
Three sets of isoquant and three sets of isocline equations can be

derived from equation 15.14 and have been used in constructing figures
15.8, 15.9, and 15.10. The symbols Pj^, Pp , and P^ indicate prices of
the three nutrients and the lines so labeled designate the isoclines which
serve as expansion paths for various price ratios for the nutrients.
Both figures 15.8 and 15.9 have ridgelines which form 90 degree angles
because interaction between nitrogen and the other two nutrients is
lacking, where negative interaction between nutrients exists, as for
p2 0 5 and K aO in Figure 15.10, ridgelines meet at an angle greater than
90 degrees: Negative P x K interaction gives the production surface a
comparatively flat top; economic limits of nutrient combination are
wide. For close complementarity and positive interaction (i.e., between
N and P 2 0 5 for corn in Figure 15.6), the ridgelines are close together
and meet at an angle of less than 90 degrees. With positive N x P inter-
action, as in Figure 15.6 where ridgelines are close together, a non-
optimum nutrient combination could be very costly. However, to deviate
slightly from the optimum fertilizer ratio line (isocline) in Figure 15.10,
if the price ratio differs only slightly from the substitution ratio indi-
cated by the isocline, may depress profits only slightly since rates of
substitution change slowly along the isoquants.
Under prices where the price of K 2 0 is 80 per cent of the P2 O s
Figure 15.8. Yield isoquants and isoclines

with dashed ridgelines at zero
level of K 2 0, Haynie soil
Figure 15.9, Yield isoquants and isoclines

level of PgQjj, Haynie soil
price, over 60 pounds of P 2 05 should be applied before any K 2 0 is used

(Figure 15.10). However, with K 2 0 at one-third the price of P2 Os it
,
would pay to apply over 30 pounds of K 2 0 before any P2 Os is used.
STAND AND NITROGEN
Several studies with stand or plant population have been completed

in Iowa. This section reports an experiment on Marshall silt loam
where stand (S) in hundreds of plants per acre and nitrogen (N) in
pounds per acre were variables. The crop is corn measured as yield
(Y) per acre. Three nitrogen and five stand levels were used on each
of the two varieties, A.E.S. 801 hybrid and Iowa 4397 hybrid. Plants
were thinned to five stand levels. A split -split -plot design was em-
ployed with nitrogen rate, stand, and variety as the whole plots, sub-
plots, and sub-subplots, respectively. With the temperature being
equally favorable (or unfavorable) the potential production of A.E.S. 801
would be expected to be relatively higher in this experiment.
Yield observations suggested decreasing total returns for higher
levels of stand, particularly when level of nitrogen is low. Response to
nitrogen was quite large, and a strong interaction exists between nitro-
gen and stand. These data also suggested that yields for variety 801
were somewhat higher than those obtained from variety 4397.
Figure 15.10. Yield isoquants and isoclines

level of N, Haynie soil
Regression Analysis
Several regression equations were considered as a basis for esti-

mating the production surface. Those appearing most appropriate are
presented below. The first two functions are quadratic forms and are
for varieties 4397 and 801, respectively.
(15.15) Y = 18.9996 •+• .2253 N + .4200 S - .000720 N2 - .001415 S

2
+ .000361 NS
(15.16) Y = 29.3249 + .1990 N + .389216 S - .000565 N2 - .001503 S

2
+ .000800 NS
2 2
Equation 15.15 has an R of .6259 while equation 15.16 has an R of
.7222. The values of t and probability levels for the regression coeffi-
cients for each equation are included in Table 15.8. The level of sig-
nificance of all coefficients appears high enough to warrant retaining
them in the equation.
Parallel coefficients of the equations for the two different hybrids
do not appear, by inspection, to be greatly different. Differences in
yielding ability of the two varieties are important to research workers
interested per se in yields from different hybrids. However, this study
is concerned with the shape of the function, in terms of optimum stand
and fertilizer levels. While the surfaces may have different vertical
Table 15.8. Values of t for the Coefficients of the Individual
Quadratic Regression Equations Relating
Corn Yields to Inputs of Stand and
Nitrogen on Marshall Silt Loam
Variety Iowa 4397 Variety A.E.S. 801

Regression
Coefficient t values for Probability t values for Probability
for equation 15.15 level* equation 15.16 level*
N 3.01 .005 2.42 .025

S 2.21 .050 2.12 .050
N2 2.63 .025 1.81 .100
2
S 2.21 .050 2.45 .025
NS 1.12 .300 2.26 .050
* Probability of drawing t as large or larger by chance, given the null

hypothesis.
locations above the resource plane, the functions will still lead to simi-
lar conclusions on optimum resource inputs, if the shape or curvature
of the surfaces are the same. Hence, a t test has been used to test the
significance of the difference between regression coefficients in equa-
tions 15.15 and 15.16. The results of these tests (shown in Table 15.9)
suggest that differences between parallel regression coefficients, as
large as those obtained, might well have occurred by chance. Hence,
the observations were pooled.
Table 15.9. Tests of Significance of Differences

Between Parallel Coefficients in
Regression Value of t for

Coefficient Probability
Difference Between
Level*
for the Two Coefficients
N .34 .80
S .17 .90
2
N .53 .60
S* .14 .90
NS 1.29 .20
* Probability of drawing a t value as large or larger by chance,

given the null hypothesis.
A
square root function was fitted to the pooled data. However, it
was logically unsatisfactory because of the signs of the coefficients.
This quadratic equation was then fitted to the pooled data:
U^“17) Y = 27.5963 + .2051 N + .3553 S - .000641 N2 - .001298 S

a
+ .000625 NS.
The R2 for equation 15.17 is .6125, and t values of the coefficients are
given in Table 15.10. All of the terms are significant at a probability
level at least as high as .05.
The significance levels of the coefficients of the pooled regression
Table 15.10. Values of t and Significance Levels

of Coefficients of Quadratic
Regression Equation 15.17
Regression
t Probability
Coefficient
Value Level*
for
N 3.14 .005
S 2.32 .025
2
N 2.63 .010
2
S 2.52 .025
NS 2.22 .050
Probability of drawing t as large or larger by

chance, given the null hypothesis.
in equation 15.17 are, for all terms, at least as high as for the indi-
vidual regressions. Hence, through equation 15.17, the pooled regres-
sion on varieties 801 and 4397, is used in the analysis which follows.
Predicted Production Surface
Decreasing total returns to stand is predicted by equation 15.17

throughout the entire range of nitrogen applications. While there are
decreasing total returns for nitrogen at the low levels of stand, this is
not true for higher stand levels. Nitrogen and stand are predicted to
have sufficient interaction to cause yields to increase continuously
along the main diagonal of the input plane. The resulting surface is
presented in Figure 15.11. The decreasing total returns to stand and
nitrogen, for the latter when stand is held at low levels, are indicated
by the negative slopes over some portion of the surface. The surface
also slopes upward from the left-hand corner above the nutrient plane
because of the stand by nitrogen interaction.
Marginal Physical Products
Equations for the marginal physical products of stand and nitrogen

are, respectively:
5Y
(15.18) = .3553 - .0026 S + .0006 N.
OS
(15.19)
~6Y •
= .2051 - .0013 N + .0006 S .
The interaction term in the original production function causes the

marginal equation for one input to contain a term due to the other input:
when one input is varied, its marginal product is affected by the level
at which the other input is fixed.
Figure 15.11. Production surface for stand and nitrogen on Marshall

siltloam, equation 15.17
By setting the marginal equations equal to zero, and solving for the
values of the inputs, the quantity of input at which the marginal product
becomes zero is obtained. When no nitrogen is applied and stand is uti-
lized to the point where its marginal product is zero (13,688 plants per
acre), the predicted corn yield is 51.92 bushels per acre. On the other
hand, when stand is utilized at a minimum rate for the experiment (i.e.,
7,000 plants per acre since a zero rate of stand would result in no
yield) and enough nitrogen is applied to drive the marginal product to
zero, yield is predicted to be 70.26 bushels per acre. The input level at
which yield is at a maximum (i.e., both partial derivatives are zero) is
determined by equating both marginal product equations to zero and
solving simultaneously for values of stand and nitrogen. Using equa-
tions 15.18 and 15.19 for this purpose, yield is predicted to be a maxi-
mum with 256.88 pounds of nitrogen and 19,873 plants per acre. The
predicted yield is 89.25 bushels, an extrapolation along the nitrogen axis
since the nitrogen input is above the limit of experimental observations.
Yield Isoquants and Isoclines
The curves of negative slope in Figure 15.12 are isoquants, depict-

ing the horizontal sections through the surface in Figure 15.11. They
Figure 15.12. Isoquants and isoclines for stand and nitrogen on Marshall
soil based on equations 15.20 and 15.21
indicate the various combinations of stand and nitrogen which will pro-
duce the specified or given yields and are based on the isoquant equa-
tion:
X
(15.20) S = 136.8798 + .2408 N t 385.2080 (.2695 + .001509 N - .000003 N2 - .005192 Y) 2
Since the isoquants are convex to the origin, stand and nitrogen are
predicted to substitute at decreasing rates. Also, since isoquants rep-
resenting equal increments in yield are successively farther apart (on
any straight line through the origin or perpendicular to either axis),
diminishing returns to the inputs are indicated.
The isocline equation derived from equation 15.20 is:
, <B 01 , 0 k (.355340) + k (.Q00625)N - .205114 + .001282 N

(15.^1; fa - .000625 + k (.002596)
where k is the Pg /%
price ratio and the other terms are as defined
previously. Isoclines (positively sloped lines) derived from this equa-
tion are presented in Figure 15.12. Again, these isoclines can be
looked upon as representing either (a) marginal rates of substitution of
given magnitudes or (b) price ratios representing particular expansion
paths. Isoclines labeled 6.0 to 10.0 approximate the range of price
ratios realized in the past. Thus, if a given yield is produced at mini-
mum cost, stand should be increased until its marginal product is al-
most zero before applying more nitrogen. In other words, the least
cost combination of stand and nitrogen approaches the upper ridgeline
(the dashed where 5N/6S is zero).
line
The dashed
lines in Figure 15.12 are the ridgelines (i.e., isoclines
representing zero substitution rates) beyond which the inputs will not
substitute for each other except at greater total physical inputs of both
variables. Hence the ridgelines represent working limits beyond which
only less favorable economic solutions can ever be found. Levels of in-
put where the marginal physical products are zero fall exactly on one
or the other of the ridgelines. This is true because the marginal rate
of substitution between the two inputs is the ratio of their marginal
physical products and the ridgelines represent all points at which the
marginal product of either stand or nitrogen is zero. Convergence of
isoclines is at a point in the input plane where the partial derivatives of
both inputs are zero. In Figure 15.12, convergence of isoclines falls at
a yield of 89.25 bushels and inputs of 256.88 pounds of nitrogen and
19,873 plants per acre. At this particular point, stand and nitrogen are
technical complements or limitational resources; neither can be sub-
stituted for the other in producing the specified yield.
Economic Optima
Optimal conditions under certainty and unlimited capital concern

selection of (a) the combination of stand and nitrogen levels which
min-
imize the cost of a given output and (b) the levels of stand and
nitrogen
which will maximize the per-acre profits. These conditions are at-
tained simultaneously when the partial derivatives of the two
input cat-
egories are equated to the ratio formed by dividing their price
(cost) by
the price of corn per bushel.
Isoclines are expansion paths, indicating the combinations
over
which nitrogen and stand should be increased in attaining
successively
higher yield levels, if each yield is to be attained at a
minimum cost.
The isoclines representing least -cost expansion paths for
reasonable
prices of nitrogen and stand fall towards the upper ridgeline
of Figure
15.12. For a price ratio of 9, the cost per pound of
fertilizer divided
by the cost of the amount of seed to provide 100 plants,
the cost of stand
is so low that economic optima seem almost consistent with the use of
planting rates which drive the marginal physical product of stand to zero.
Optimum nitrogen levels vary widely as both corn and nitrogen
prices vary. When nitrogen costs $ .08 per pound, nitrogen use ranges
from 181 to 211 pounds per acre depending on corn and stand prices.
When nitrogen costs $ .15 per pound, the predicted optimum quantities
of nitrogen are lower, ranging from 120 to 172 pounds per acre. Varia-
tions in nitrogen prices, within their relevant range, cause larger fluc-
tuations in resulting profit than variations of stand prices within their
relevant range.
Profit maximization, even with ability to predict yields and prices,
as outlined above, is possible only for the farmer with unlimited capi-
tal. Farmers with limited funds may wish to apply some fertilizer, but
divert the remainder of their capital to other uses. Also, given weather
and price uncertainty, farmers not limited on capital also may use re-
sources short of equating marginal value products with the cost of re-
sources. Hence, we suppose a farmer has capital to use only one -half
the profit maximizing nitrogen quantities for a given set of prices. We
then ask: “What stand level would be optimum?” Computations show
that stand costs are so small that most farmers could use levels which
drive stand marginal productivity to zero regardless of availability of
capital or price of seed and yield.
MULTIPLE HAY SURFACES
This section reports a fertilization and production function study for

hay. This crop, with several products forthcoming during the year,
presents analytical and decision problems which are not encountered
with commodities harvested at a single time during the year. However,
emphasis in this section is on basic production relationships rather
than on the decision criteria which might be used for hay production
when the crop has different values or prices throughout the season, the
quality differs with the time of harvest, and other such problems.
The experiment from which the data were obtained was one of a se-
ries of pasture fertilization trials. The experiment analyzed was con-
ducted on Weller silt loam in southeast Iowa in 1952. The design was a
3x3 factorial, replicated twice, giving a total of 18 observations. Two
of these were checks. P 2 0 5 and K 2 0 fertilizer were both applied at
three levels. The alfalfa was cut three times during the course of the
growing season. Three cuttings are considered normal in this area.
While improved designs are now in use, the data which follows does pro-
vide some interesting insights into the nature of hay production surfaces.
Predicted Surfaces
Experimental yield observations were used to predict five produc-

tion surface equations.First, an equation was fitted to each separate
cutting. Next, yields for the first and second cuttings were added to-
gether and an additional function was fitted to this total. Likewise a
regression equation was computed for the sum of the yields of the three
cuttings. Hay yield (Y) is measured in tons per acre while K 2 0 (K) and
P 2 O s (P) inputs are measured in pounds per acre. 2 The five equations,
selected from among several algebraic forms tried, presented in the
order mentioned above are
(15.22) Y = .82235 4. .007042 P + .006181 K - .000028 P

2
- .000027 K2 - .000010 PK
2
(15.23) Y = .811947 + .001278 P + .001403 K - .000004 P - .000006 K 2
+ .000005 PK
2
(15.24) Y = .490277 + .001431 P + .002181 K - .000005 P - .000012 K 2
- .000002 PK
2 2
(15.25) Y = 1.634182 + .008320 P + .007584 K - .000032 P - .000033K - .000005 PK
2 2
(15.26) Y s 2.124459 + .009751 P + .009765 K - .000037 P - .000045 K - .000007 PK
where the first three are for the first, second, and third cuttings, re-
spectively, while the fourth function is the sum of the first two cuttings,
and the fifth is for the sum of all cuttings. Equation 15.25 could have
been obtained by adding equations 15.22 and 15.23 and equation 15.26 by
adding equations 15.22 through 15.24.
The over-all significance of the regressions was tested by means
of the F ratio. For the first cut, the first plus second cut, and the first
plus second plus third cut, the F’s are all significant at probability
levels of less than 5 per cent. For the second and third cuts taken by
themselves, the F values fall just outside the 5 per cent level of proba-
bility. After examining the standard errors and related statistics, it
was decided to retain all of the variables indicated in equations 15.22
through 15.26. The R2 values for these five equations were respectively
.96, .86, .91, .98, and .98.
Figure 15.13 provides three of the production surfaces which can be
predicted from the above equations. All three surfaces are of similar
configuration and denote negative marginal products, even within the
range of experimental observations, for both nutrients on each surface.
The differences in heights of the three surfaces represent the addition
to total yield due to additional cuttings. The second cutting was greater
than the third, as suggested in Figure 15.13 and by equations 15.23 and
15.24.
Hay Isoquants and Isoclines
Aset of three isoquants has been derived for each production func-
tion in equations 15.22, 15.25, and 15.26. These are presented in
Subsequent interpretation of the data in terms of hay rather than

oven dry material
means that the regression coefficients would be altered slightly.
Figure 15.13. Production surfaces for 1st, lst+2nd, and lst+2nd+3rd cuttings of
alfalfa
Figure 15.14. The isoquants have relatively slight curvature. The

change in slope is gradual, indicating that the nutrients are good substi-
tutes, within the range of the experiment. This is true for each set of
isoquants.
Yield isoclines (least cost expansion paths) were derived by equat-
ing the marginal products of each production function to the nutrient
price ratio and solving for one nutrient. Isoclines were worked out for
the first, first plus second, and first plus second plus third cuttings,
corresponding to production functions in equations 15.22, 15.25, and
15.26. The respective isocline equations corresponding to these func-
tions are presented below where the nutrient price ratio is represented
by k.
(15.27)
.007887 - .000062 P - .000011 K
.006923 - .000060 K - .000011 P
.009318 - .000072 P - .000006 K

(15.28)
.008494 - .000074 K - .000006 P
.010921 - .000082 P - .000008 K

(15.29)
.010937 - .000100 K - .000008 P
An isocline family has been drawn for each equation in Figure 15.15.
Each isocline represents the least-cost P2 O* and K 2 0 combination for
the nutrient price ratio shown. The slopes of the sets of isoclines do
k 2 o, pounds per acre

Figure 15.14. Isoquants for 1st, lst+2nd, and lst+2nd+3rd cuttings of alfalfa hay
not differ greatly. Since an isocline connects points of equal slope on a

family of isoquants, the corresponding sets of isoquants are also simi-
lar. The dotted lines again are the ridgelines, beyond which the inputs
do not substitute for each other in maintaining a given yield.
The family of isoclines converges to a point of maximum yield
FOR 1st CUT FOR 1st r 2nd CUT FOR 1st + 2nd t- 3rd CUT
K2 Q
Figure 15.15. Isoclines for 1st, lst+2nd, and lst+2nd+3rd cuttings of alfalfa hay
where the partial derivatives of both factors are zero. For one cutting
alone, the predicted amounts of P2 Q* and K 2 0 required are smaller
than for two or three cuttings. Maximum yield for three cuttings is
predicted to require more P 2 0 5 but less K 2 0 than for two cuttings, a
condition which might appear unrealistic. However, this possibility
could easily exist when moisture conditions differ between the growing
periods for the three cuttings.
CHAPTER 16
Functions for Fixed Plants
and Other Farm Situations
HIS CHAPTER summarizes selected farm production functions
T derived from cross-sectional samples for farms of given size in

acreage. It also reports other selected farm and enterprise pro-
duction function studies made in Iowa. The production function esti-
mates have the limitations mentioned in earlier chapters for the par-
ticular aggregation and specification procedures used. However, the
results were useful for diagnostic purposes in analyzing farm resource
returns and capital productivity. All studies employ Cobb -Douglas
types of equations. The aggregation and sampling methods used may be
of particular interest and the studies are reported for these reasons.
DIAGNOSTIC PURPOSES
In contrast to biological functions such as those discussed in Chap-

ters 8-15, firm and industry production functions provide little oppor-
tunity for decisions on and recommendations for highly specific types of
resources. In preceding chapters, for example, it was possible to de-
rive coefficients which provide the basis for recommendations on use
of such specific resources as nitrogen or soybean oilmeal. However,
aggregation over resource categories must be much broader for firm
production functions. The coefficients which result can have application
only to similarly broad categories of inputs. Hence, firm production
functions can never be used for recommending quantities of corn feed,
tractor fuel, and similar input quantities which are optimum for an in-
dividual farmer. Ordinarily, inferences must be restricted to aggre-
gate resource categories such as labor, land, crop capital, etc.
Hence, it is not expected that data from firm production functions
will be used to make specific recommendations such as those arising
from biological production functions or budgeting and linear program-
ming analyses. Largely, firm functions must be used for more general
diagnostic purposes. These uses would include suggestions to farmers
whether they are using too much or too little capital, or whether a re-
allocation of capital from crops to livestock would be profitable. One
could never use this empirical tool to tell a farmer whether he should
invest in Jersey rather than Ayrshire cows, or that he should invest
more in a particular insecticide or treatment.
554
FUNCTIONS FOR FIXED PLANTS 555
firm production functions also may be

In this “diagnostic sense,”
used to indicate the “degree equilibrium” in farming. Coefficients
of
derived from them may give broad indication of whether returns to
capital and labor in agriculture are higher than their market prices, or
whether great differences exist between different regions in respect to
capital productivity.
FIXED PLANTS ON MARSHALL SOIL
The first study involving fixed plant size is for Marshall silt loam. 1
It was completed for the year 1945 and provides estimates of capital
and labor productivity when land per farm is fixed.
The sample of 70 farms was drawn in an area of Marshall silt loam
and associated soils in southwestern Iowa, and was restricted to 160-
acre, owner -operated farms. Farms were included in the sample only
if they met certain tests of homogeneity in respect to soils. No farm in
the area includes Marshall silt loam alone. Because of the rolling to-
pography of the area and the related factors associated with soil for-
mation, Marshall silt loam is typically found on farms in association
with some amount of either bottomland soils in the valleys, Minden silt
loam on the level ridges, or Shelby silt loam on the steeper slopes.
Several of these soil types can often be found even within a 40 -acre
field. Although Marshall silt loam is predominant, many farms scat-
tered throughout the entire geographic area have a large acreage of as-
sociated soils. In order to obtain a sample of farms reasonably homo-
geneous in respect to basic soil resources while still retaining a
somewhat typical association of soils, upper limits were placed on the
acreage of soils other than Marshall silt loam. Soil maps were made of
each individual farm indicating the soil types, degree of slope, and de-
gree of erosion for each individual farm as a means of delimiting the
farms to be retained in the sample. Farms retained in the original
sample contained no more than 25 acres of Shelby soils and between 40
and 75 acres in combination of bottomland soils or Marshall silt loam
with a slope of 4 per cent or less. The remainder of the area was thus
composed of Marshall silt loam with a slope of more than 4 per cent.
The 70 farms also were selected to fall within certain limits in re-
spect to livestock, (a) No farm was included which had more than five
milk cows. (Actually, all of the farms included in the analysis had no
less than three and no more than five milk cows.) (b) No farm was in-
cluded which had a poultry enterprise deviating greatly from a typical
farm flock, (c) All farms were excluded which had beef cows or a
sheep enterprise of any kind, (d) Farms included in the sample had
only the livestock mentioned above plus hogs and feeder cattle. In other
words, the sample farms were those with poultry and dairy enterprises
1
Heady, Earl O. Productivity and income of labor and capital on Marshall silt loam
farms in relation to conservation farming. Iowa Agr. Exp. Sta. Bui. 401. Ames. 1953.
-
556 FUNCTIONS FOR FIXED PLANTS
of the nature outlined above plus no other livestock, hogs but no feeder
cattle, or hogs and feeder cattle.
In the sample and functions derived for this 1945 study, the data
allow a transition from an intensive grain farming system on farms
where crop inputs are low to intensive forage farming where crop in-
puts are high; they allow a transition from a small dairy -poultry combi-
nation to hogs, then to hogs and feeder cattle, and finally to more feeder
cattle as livestock inputs on farms. While the derived productivity es-
timates do not trace out the function for a single type of product, they
represent a typical transition in investment as crop and livestock en-
terprises are intensified on 160 -acre farms in the area, particularly
when the transition is in the direction of more forage and a greater de-
gree of conservation farming. Farmers with little capital generally
raise only crops, while those with a little additional capital have small
milk cow and poultry enterprises. Access to more funds allows a hog
enterprise in conjunction with grain production and sales; still greater
funds generally bring about a greater forage acreage and some feeder
cattle to go with it. Ample funds usually allow larger hog and cattle
feeding enterprises, either from home grown or purchased feeds. The
sample upon which this study is based is a random one; the inferences,
however, apply only to 160-acre hog-cattle feeding farms, the size and
organization of farms most numerous in the area. Because of the strat-
ification procedure used, inputs and output are somewhat greater than
for all owner-operated farms in the area. The inferences are repre-
sentative, however, for farms using different quantities of resources
and following the production pattern outlined previously.
Production functions were derived separately for crops and live-

stock. The variables included in these functions were:
Yc is the value of crops produced and represents the dependent or output variable
in crop production.
Yx is the value of livestock products produced, including inventory changes and
represents the dependent or output variable in livestock production.
L is input of labor, measured in dollars, used on crops or livestock according to
the function indicated'.
C is input of crop services,measured in dollars, and includes seeds, insecticides,
and other supplies attached directly to crop production for the crop function.
M is input of machine services and includes fuel, grease, repairs, depreciation,
and a small charge for housing for the crop function.
F is input of fertilizer, lime, and manure and a nominal machine cost, all meas-
ured in dollars. Manure has been converted to fertilizer m
terms of N, P,
and K and priced according to commercial fertilizer. Since fertilizer and
manure applications and soil management practices were enumerated for a
period of 3 years, the average of the last 2 years was used to obtain residual
effects. A small machine cost, calculated to include costs of fertilizer appli-
cation, was included as part of the value of fertilizer inputs (as a technical
FUNCTIONS FOE FIXED PLANTS 557
complement of fertilizer itself) and was subtracted from other machine serv-
ices.
G is input of feed for livestock including grain, protein supplements, hay, pasture,
and miscellaneous minerals in the livestock function.
S is the input of livestock capital services measured in dollars in the livestock
function. It includes annual inputs for livestock (in contrast to the capital
stock itself) as follows: For chickens, fattening pigs, and feeder cattle, the
beginning value is included as an input. For milk cows, brood sows, and lay-
ing hens, depreciation is computed and used as the input. Closing inventory
values of growing and fattening stock are used as an output, while beginning
inventory figures are used as an input. This input also includes the value of
all grain, hay, pasture, and supplemental feeds and building, equipment, and
machine services used on livestock. The latter items include depreciation
and repairs and other annual inputs (in contrast to the value of the asset it-
self). No building inputs have been used in the crop function. This procedure
has been employed under the assumption that crop storage belongs to a dif-
ferent production process than crop production. (Buildings, aside from facil-
ities for drying corn, mainly contribute storage services for gaining higher
seasonal market prices or for later livestock production.)
Thus, in all the derived functions, the inputs are measured in dol-
lars and refer to the flow of services (or expenses) for the year; they
are not capital values. Calculation of profitability can be made directly;
if the marginal return is greater than $1, the particular dollar of input
or expense more than paid for itself. If the input itself costs $1 and 5
per cent interest must be paid on it, a marginal return of $1.20 denotes
the addition of $.15 to net returns. This procedure has been followed to
ease the conversion of the 1945 prices to other levels.
In examination of the livestock function, it was found that inputs of
livestock, feed, and building services were highly correlated. They
tend to be technical complements and their quantities are increased to-
gether. Most farmers make decisions in this manner. For this reason,
the three inputs were aggregated into a single category for the purpose
of calculations. Labor was retained as a separate type of input since it
was not so highly correlated with other categories of resources.
The regression or elasticity coefficients, derived separately for
crop and livestock production, are included in Table 16.1. By coinci-
dence, the sum of elasticities has a value of .935 for both functions.
The elasticity coefficients, as averages based on the functional form,
Table 16.1. Regression Coefficients or Elasticities for Crop and

Livestock Functions in 1945 Marshall Study
Crop Function Livestock Function

Resource Elasticity Resource Elasticity
Labor (L c ) .333 Labor (Li) .251

Crop services (C) .384 Livestock services (S) .684
Machine services (M) .149
Fertilizer (F) .069
Sum of elasticities .935 Sum of elasticities .935
are less than 1 for each individual resource indicating diminishing mar-
ginal productivity.
The sum of exponents, .935 for both crops and livestock, means that
if, as an average, all resources used for each product are increased by
1 per cent, the value of output is increased by only .935 per cent and di-
minishing returns hold true; each 1-unit increase in resources, on the
average, will add a smaller amount to output or return than the previ-
ous unit. Neither of these sums of elasticities differed significantly
from 1.0. For a 160 -acre farm with all acres fully in cultivation, it
would be expected that diminishing productivity in crop production re-
sources might be expected as greater inputs are applied to a fixed land
area. While increasing returns might hold true where the quantity of
resources applied is very small, farms in the sample generally used
resources beyond this range, and the entire group had all acres of their
units in operation. A small range of increasing returns might also be
expected in livestock production. It is likely, however, that most
farmers in the sample had pushed investment in livestock resources be-
yond this point. It also is entirely reasonable that over the range of
livestock investment found on these farms, constant returns to scale
might well exist even if increasing returns were not realized. Very
great livestock inputs might, of course, result in decreasing returns.
Decreasing returns in livestock may also result from the fact that the
product measured is not homogeneous in terms of enterprise; it in-
cludes the “shading” from a little poultry and dairy products to hogs
and then to feeder cattle in the manner outlined previously. This change
in product, while deviating from a homogeneous enterprise, is realistic
in terms of the pattern followed as aggregate livestock output is in-
creased.
The values of t and the correlation coefficients between pairs of re-
sources and products are shown in Table 16.2. All of the regression
coefficients, except that for machine services on crops, were significant
at a 5 per cent, or higher, level of probability; the regression coeffi-
cient for machine services is significant at the 10 per cent level of
probability.
FIXED PLANTS ON CLARION -WEBSTER SOIL
This section reports a 1951 study for farms of fixed size in north
central Iowa. 2 The farm sample was selected in the Clarion-Webster
soil area, to give opportunity of homogeneity in the basic production
functions. However, homogeneity of soils does not guarantee that hybrid
or mongrel production functions will not result from inter-farm sam-
ples where operators use different techniques.
In the analysis which follows, two production functions, one for
2
Heady, Earl O. Resource productivity and returns on 160-acre farms in north central
Iowa. Iowa Agr. Exp. Sta. Bui. 412. Ames. 1954.
Table 16.2. Correlation Coefficients and Values of t
Crop Function
Crop Machine
Labor Fertilizer
services services
Labor (r) 1.00 .49 .16 .62

Crop services (r) 1.00 .32 .12
Machine services (r) 1.00 .10
Fertilizer (r) 1.00
tvalues 3.14 8.93 1.84 3.06
Multiple correlation coefficient (R) .86
Livestock Function
Livestock
Labor
services
Labor (r) 1.00 .64

Livestock services (r) 1.00
t values 4.45 11.90
Multiple correlation coefficient (R) .96
crops and one for livestock, have been derived. The crop function is of
the form Y c = a c L^
1
M^
C)r where Y c refers to the value of crops pro-
duced, L c refers to the quantity of labor measured in months, c re- M
fers to value of all machinery expenses or inputs of the particular year
(including repairs, depreciation, fuel, oil, etc.), and C c refers to all an-
nual expenses on crops (including seed, fertilizer, lime, seed treat-
ment, etc.); labor inputs are measured in months and machinery, and
crop inputs are measured in dollars. The livestock function is of the
form Yx = a x Li where Yx refers to the value of livestock output dur-
ing the year (sales + home used + appreciation of young and fattening
stock) while L x is labor in months used on livestock, and C x is all an-
nual capital inputs, measured in dollars, used on livestock.
Sample
As a further step in introducing homogeneity into the sample, pro-

duction functions were derived only for 160 -acre farms (farms falling
in the range of 140 to 169 acres). The random sample included 108
farms. Inferences in this study, therefore, are restricted to 160-acre
units; land input is considered as a fixed resource in the production
functions presented later.
Elasticity Coefficients and Scale Returns for 160-Acre Farms
The value of the regression coefficients or elasticities for the two

production functions are presented in Table 16.3 along with related
1
Table 16.3. Elasticities and Related Statistics for Basic

Functions on Clarion-Webster Soils
Crop Livestock
Item
Function Function
Value of constant a 3.0618 .3580
Value of elasticities*
1 .0729 .2127
c .0549 .8735
m .1798
Sum of elasticities .3076 1.0862

Value of t for elasticities*
1 1.98 3.65'!'
c 1.8 2 § 20.35T
m 3.89t
*The figure for c refers to crop capital for crops and to all capital for
livestock; m refers to machine capital for crops but this category of
inputs has been included with all capital for livestock.
Probability level for t's:
1 > P > 0;
1
*5 > P > 1;
Ho > P > 5.
statistics. All but one of the elasticities are significant at the 5 or 1
per cent level of probability; the coefficient for crop service capital
was significant at a probability level approaching 5 per cent and is, with
recognition of a slightly wider probability range for inferences, ac-
cepted for the analysis which follows. Each of the elasticities for crops
or livestock is less than 1.0 — indicating that diminishing returns hold
true for the particular resource. A 1 per cent increase in input or use
of the particular resource results in an increase in value of production
by less than 1 per cent; the return per unit of the particular resource
will decline to lower magnitudes as more of the resource is used.
For labor used on crops, for example, a 1 per cent increase in la-
bor is predicted to result as an average over the sample, in an increase
of only ,07 per cent in the value of crop production. For crops, the sum
of the elasticity coefficients differs significantly from 1.0 — indicating
diminishing returns as more and more of the various resources are
used for crops on a given land area. Since all of the crop acres were
in cultivation on all farms in the sample but different quantities of capi-
tal and labor resources were used on a given acreage, diminishing re-
turns are indicated for each resource and all resources in combination.
On the other hand, the sum of the elasticities for livestock does not
differ significantly from 1.0. As an average for farms in the sample,
more resources can be put into livestock with returns as great as the
mean quantity now being realized. While diminishing returns might be
expected on the 160 -acre farms if they carried an extreme amount of
livestock, few farms approached programs of this intensity.
Resources Used and Factor Productivity
The average value of production, the mean quantity of resources

used and the marginal productivity of crop resources at the mean are
shown in Table 16.4. The marginal product figures have this meaning:
"one more unit” of the particular resource, with its input and that of
other resources at the mean of the quantities shown in the top of the
table, will add the indicated quantity to total value of production. An in-
crease in crop services will add to total value of production at the rate
of $1.08 for each added $1 input of capital services; a $1 input in crop
capital services returns itself plus $.08 in profit. The marginal prod-
uct per month of labor is $77. In other words, the use of "one more
month of labor” beyond the mean quantity per farm in the sample, would
add this amount to total value of product. Actually, there would be little
productive use for labor in crops, beyond the mean quantity used per
farm. The computed marginal product of machine capital services is
$.93, a marginal return of this amount for each $1 of annual input or
expense.
While these return figures may appear low, except for crop serv-
ices, two things should be remembered. First, the marginal returns
are measured with resource inputs at the mean. Because of the
diminishing -return nature of the productivity coefficients, marginal
returns will be greater than the indicated amounts for labor inputs
smaller than 6.2 months, for crop service inputs smaller than $334 and
for machine service inputs smaller than $1,260; and, second, the esti-
mates are made from a set of data involving sample variance, and,
while probability tests suggest that the elasticity coefficients differ
significantly from zero, we may inquire whether they differ signifi-
cantly from a level necessary to give marginal productivities equal to
Table 16.4. Mean Quantity of Resources Used on Crops,

Marginal and Gross Average Productivity
of Specified Resources
Item Amount
Average value of production per farm $6,452.00
Average input per farm:

Crop services $ 334.00
Machine services $1,260.00
Labor 6.2 months
Marginal products at resource means:

Crop services ($/$) $ 1.08
Machine services ($/$) $ .93
Labor ($/month) $ 77.87
Gross average products at resource means:

Crop services ($/$) $ 19.32
Machine services ($/$) $ 4.82
Labor ($/month) $1,041.00
Land ($/acre) $ 42.00
Table 16.5. Test of Departure Between Marginal Resource

Returns and Market Price of Resource
Value of Elasticity to Give

Resource Marginal Product Equal to Value of t
Market Price of Resource
Crop services .0548 .002*

Machine services .2070 .587*
Labor .0729 1.029t
Probability levels:
*,P > 50;
tp >30.
the market price or cost of each of the resources. In other words, does
the marginal return of $77.87 per month of labor differ significantly
from $132.44, the market wage rate (without board) for labor in north-
ern Iowa; do the capital returns of $1.08 and $.93 differ significantly
from the $1.06 cost ($1 principal plus 6 per cent interest) for crop
service and machine service capital respectively? As a test of these
possibilities, the elasticities of production necessary to give marginal
products equal to the market cost of the resources have been computed.
Probability tests have been made for the differences between the elas-
ticities derived in the study and those necessary to give marginal prod-
ucts equal to the market price of the resources. The statistics for this
test are given in Table 16.5.
Since none of the t values are significant at the 30 per cent level of
probability, we cannot say that the marginal returns per month of labor,
as an average for farms in the sample, differed significantly from the
market wage rate of $132.44 per month for labor or of $1.06 per $1 in-
put for crop or machine capital. While the condition need not hold true
for all farms in the sample, inferences based on the above statistics
are the following:
Farmers were, on the average, maximizing returns under the par-
ticular prices and yields of the year; efficiency in production had been
attained in the sense that the cost of resources approached the added
return for more of these resources used beyond the per farm mean.
These results seem reasonable in view of the fact that farmers in north
central Iowa have a capital position about as favorable as those of any
other area in the United States. Also, neither of the years included
were outstanding in crop yield. Had crop yields been equal to the 10-
year average, marginal returns on resources would likely have been
significantly greater, given the particular price of the year, than the
cost or market price of the resources to the farmers. (Productivity or
return of resources would have exceeded the cost of the resources.)
Rates of Substitution Between Machine Services

and Labor Used on Crops
The Cobb-Douglas function is not always appropriate for estimating

rates of factor substitution because of the algebraic properties outlined
in Chapter 3. However, it is used in this section to estimate the mean
machine service -labor isoquant for the 160 -acre Clarion -Webster sam-
ple and for estimating marginal rates of substitution between these two
resource categories. It should be remembered that these predictions
are simply those which result from the functional form used and the
numerical observations of the sample. The limitations discussed ear-
lier still apply, but the estimates may provide some “diagnostic in-
sights” in regard to machine -labor substitution rates.
Using the original crop production function as a basis for estimat-
ing, the data of Table 16.6 have been derived. These data indicate
(a) the combinations of machinery, measured as annual expenses of
power and machinery, and labor which are predicted to produce the
mean crop product of $6,545 and (b) marginal rates of replacement be-
tween these factors. Because of the asymptotic nature of the isoquants,
they undoubtedly overestimate substitution rates for factor combina-
tions extending away from the mean combination. The first two col-
umns suggest that the mean crop product of 1951 could be attained with
10.96 months of labor and a $1,000 annual machinery expense, 6.99
months of labor and $1,200, 4.78 months and $1,400, or any other of the
combinations shown. (The upper and lower figures of the two columns
are extremes, and, while in the range of observations, they still extend
too far for the particular algebraic function, they are included for illus-
trative purposes.) Column 3 shows the marginal rate at which machine
services substitute for labor, predicted by the Cobb-Douglas function
used. The figures are computed on the basis of derivatives at exactly
the combinations indicated in columns 1 and 2; they are not averages
between combinations.
Table 16.6. Mean Isoquant and Marginal Rates of Substitution for Machine Services and
Labor in Crop Production, 160- Acre Farms on Clarion- Webster Soils
Isoquant
(Combination Value of Labor Replaced
Marginal Rate of
Machine Service and
Substitution of
by $100 m
Annual
Labor to Produce Machine Services at
Machine Services
Average Crop
for Labor (Mos.) 50 per cent
Output of $6,545) 50 per cent
if Labor Were Replaced 1951 greater
of wage
Input of Quantity of by $100 Machine wage than wage
rate of
machine labor m Service Input rate rate of
1951
services months 1951
($) (months) (months) ($) (?) ($)
900 14.22 3.90 516 774 258

1,000 10.96 2.71 359 539 180
1,100 8.66 1.94 257 386 129
1,200 6.99 1.44 191 277 96
1,400 4.78 .84 111 167 56
1,600 3.44 .53 70 105 35
1,800 2.07 .32 41 62 21
These figures show declining rates of substitution of machine serv-

ices for labor; at larger inputs of machine services, the amount of la-
bor replaced by $100 in machine services becomes less and less. As
column 4 shows, $100 in machine services is predicted to replace a
$359 value of labor (2.71 months at $132.44 per month) when machine
service input is $1,000 and labor input is 10.96 months. Of the combi-
nations shown, it is profitable to substitute machine services through a
combination including $1,400 in machine expenses and 4.78 months of
labor; costs of the given output of $6,545 can be lessened since the
value of labor saved is more than the $100 machine services added.
With a 50 per cent increase in wage rates, machine service costs re-
maining the same, machine expenses could be extended through $1,600;
whereas they could be extended only through $1,100 with wage rates
equal to one-half those of 1951.
It is of interest that, even with relatively large changes in wage
rates (over the range $66 to $198, or from 50 per cent to 150 per cent
of $132), the most profitable machine -labor combination would, within
the combinations shown, vary only over a $500 range for machine in-
puts. In other words, a relatively small machine input is extremely
profitable, regardless of wage rates. However, once a complement of
machines for tilling, cultivating, and certain harvesting operations has
been attained, further additions to machine expenses make only limited
savings in labor and addition to income.
LONG-RUN PRODUCTION FUNCTIONS

FOR TAMA-MUSCATINE SOILS
The production function study reported in this section applies to a
sample of farms with Tama -Muscatine soils. The sample was drawn
for east central Iowa in the year 1954. 3 In contrast to the two “fixed
plant” or short -rim production function studies reported above, this
study is of long-run production functions. It is long-run in the sense
that none of the resource categories are held fixed. All farms, quali-
fying in respect to Tam a -Muscatine soils and falling in the randomly
selected sampling units, were included in deriving the production func-
tion. Functions were fitted separately for crops and livestock; specifi-
cation of resource categories were somewhat different than for the two
studies reported above.
Crop Function
The crop function, which fitted to the entire sample of 255 farms,
took the form of equation 16.1, where:
3
This section summarizes the study conducted by David W. Brown and Earl O. Heady
and reported in Brown, David W. Adjustment of value productivity estimates to changes in
price and technical relationships. Unpublished Ph.D. thesis. Iowa State University Library
Ames. 1956.
f
(16.1) Yc = 17 .9X X f° X f s X f 2 X 3
Yc denotes the dollar value of estimated gross income from crops. This gross
income includes the values of crop products on hand at the end of the year,
sold, or used on the farm. Also included are certain miscellaneous receipts:
payments for off-farm labor and machine work, government conservation
payments, and co-op dividends.
Xj, denotes the acres of cropland used in crop production during the year. Perma-
nent pasture, woodlots, and waste space are excluded.
X 2 denotes the annual labor used for crop production. These quantities are meas-
ured in terms of 10-hour days (of labor actually used and not that available).
X 3 denotes the dollar value of annual crop machine services. Included are the
values of crop machinery depreciation and repairs, machinery operating ex-
penses, and hired machine work.
X 4 denotes the dollar value of fertilizer and lime applied during the year.
X 5 denotes the dollar value of miscellaneous crop capital services. These include
expenses for seed, insecticides, seed treatment, electricity, and telephone
service.
The coefficient of multiple determination was .90, indicating that 90

per cent of the variation in crop income was associated with changes in
Input quantities. The sum of elasticities in equation 16.1 was 1.18, sug-
gesting increasing returns to scale in crop production, assuming all in-
put categories have been included in sufficient manner. All of the in-
dividual elasticities were significantly greater than zero at the 5 per
cent level of probability.
Table 16.7. Values of t and Significance Levels for

Crop Regression Coefficients
Probability
Standard
Regression Coefficient Level for
Error of
for Elasticity
Elasticity
Coefficient
Cropland .065 .001

Crop labor .065 .001
Crop machine services .048 .001
Fertilizer and lime .005 .05
Crop capital services .028 .01
The mean marginal value productivities of the crop inputs, at their

geometric means, are shown in Table 16.8. These show very high re-
turns to land, labor, fertilizer, and crop services. Like the study on
Clarion-Webster, they show a low marginal productivity of machine
capital at the geometric mean. High productivities of resources are
expected in crop production for this soil complex, one of the most pro-
ductive in the Corn Belt. Linear programming and other types of stud-
ies also show, for example, that labor used in corn production has an
extremely high return. Even at the lower fiducial limits, to express the
probability and sampling characteristics of the estimates, the produc-
tivities of labor are still high. Labor used in magnitudes to produce
Table 16.8. Marginal Productivities at Mean Input Levels,

Crop Function, 255 Tama- Muscatine Soil
Association Farms, 1954
Marginal Value
Input Geometric
Product at the
Category Mean
Geometric Mean
Cropland 139 acres $33.47 per acre

Crop labor 97 days $34.64 per day
Crop machine services $1,945 $ .73 per dollar
Fertilizer and lime $ 61 $ 1.69 per dollar
Crop capital services $ 359 $ 1.75 per dollar
crops does have high productivity when combined with the necessary
complement of other resources. Increased alone, however, against a
fixed mix of other resources, its productivity would decline rapidly for
inputs beyond those necessary for "average requirements* in planting,
tilling,and harvesting crops. This same statement would apply to
other resources used on crops.
Marginal labor productivities for the Tama -Muscatine study are
much higher than for the Clarion -Webster study. Two things probably
account for most of this difference: (a) the latter study had labor
measured in a partial availability manner, the former had labor meas-
ured as that actually used on farms; and (b) the Clarion -Webster func-
tions were for smaller farms which have relatively large labor sup-
plies. Without off-farm work, labor inputs have little cost and may be
carried to a point where their marginal productivity is quite low.
The low machinery productivity again is partly expected. A con-
siderable part of the machinery investment on these farms can be con-
sidered for consumption rather than production purposes. Power units
and equipment are found on many of these farms and serve as items of
convenience and for lessening the drudgery of hard work, an entirely
economic outlay of funds.
Livestock Function
The livestock function, also derived for the 255 farms in the sam-
ple, was estimated as equation 16.2 where
190 360 009 602
(16.2) Y
1
= 1.79Z; Z-2 Z3
-
Z4
Yjl denotes the dollar value of gross income from livestock. Included are the
values of nonbreeding stock on hand at the end of the year, sold, or used on
the farm, plus the values of all livestock products sold or used on the farm.
Zx denotes the dollar value of all feed fed during the year. Both homegrown and
purchased feeds are included.
Z2 denotes the labor used for livestock production (and not the supply available).
As in the crop function, labor is measured in terms of 10-hour days.
Z 3 denotes the number of square feet of building space used m livestock produc-
tion.
Z 4 denotes the dollar value of miscellaneous livestock services. Included are the
values of nonbreeding stock on hand at the beginning of the year or purchased
during the year; breeding stock depreciation; livestock machinery deprecia-
tion, repairs, and operating expenses; supplies and medical expenses; live-
stock commissions; and electricity and telephone charges.
The coefficient of multiple determination was .89, indicating that

this per cent of the variation in livestock income was associated with
changes in the livestock inputs in the function. The sum of the elastici-
ties was 1.16 for livestock
but did not differ significantly from 1.0. All
of the elasticities,except that for the building input category, were sig-
nificantly greater than zero as indicated in Table 16.9.
Table 16.9. Values of t and Significance Levels for

Livestock Regression Coefficients
Standard Probability Level

Error of for Elasticity
for
Elasticity Coefficient
Feed .051 .001

Livestock labor .055 .001
Building space .056 .50
Livestock capital
services .034 .001
The mean marginal value productivities of the livestock inputs,

computed at their geometric means, are shown in Table 16.10.
Again, the mean marginal productivity of labor is high. It bears a
ratio to that of labor used on crops in about the magnitude expected by
extension specialists. The marginal productivity of livestock capital is
considerably greater than the cost of capital. The marginal produc-
tivity per dollar of feed for equation 16.2 is low, amounting only to
about the average margin over feed costs expected on the basis of farm
record summaries. However, a problem of multicollinearity exists,
with a correlation coefficient of .91 between feed and livestock capital
Table 16.10. Marginal Productivities at Mean Input Levels,

Livestock Function, 255 Tama- Muscatine
Soil Association Farms, 1954
Input
Marginal Value
Geometric
Product at the
Category Mean
Geometric Mean
Feed $6,839 $ .32 per dollar

Livestock labor 199 days $21.02 per day
Building space 6,776 sq. ft. $ .015 per sq. ft.
Livestock capital services $4,893 $ 1.43 per dollar
568 FUNCTIONS FOE FIXED PLANTS
inputs. Hence, there is empirical basis for the aggregation of livestock

inputs used in the study for the Marshall and Clarion-Webster studies
described earlier in this chapter. When the feed and livestock inputs
from equation 16.2 were aggregated into a single input category with
value of livestock production being expressed as a function of labor and
this combined category, the coefficient of multiple correlation became
.89. Both elasticities were significant at a probability level of .01. The
mean marginal productivity for labor became $18.34 per day while the
mean marginal productivity for the aggregate capital input became $1.13
POULTRY ENTERPRISE FUNCTION

The study reported a laying flock enterprise. 4
in this section is for
As a poultry production function study, from analyses re-
it differs
ported in chapters 10 and 11 in this respect: it is not an experimental
short-run function derived for two feed inputs. Instead, the data came
from record-keeping farms where all inputs or their costs were re-
corded for the laying flock enterprise. In contrast to the functions re-
ported in the current chapter, observations are measured for the par-
ticular enterprise, although the farms from which the data came also
produced other products.
The records from which the observations were obtained were those
kept by poultry producers under the guidance of W. R. Whitfield, poultry
extension specialist at Iowa State University. Hence, the production
functions which follow do not serve as sample predictions for the popu-
lation of Iowa farmers. Instead, they represent an estimate of the pro-
duction function for the universe of poultry producers who kept records
summarized by Professor Whitfield.
Production functions for egg laying flocks were derived for two
breeds: hybrid hens and leghorn hens. First, production functions
were computed for each of these breeds separately. Then, the obser-
vations for both breeds were pooled and an additional production func-
tion was computed. There were 76 observations for hybrids, 64 for
leghorns, and 140 for both breeds. The functions of a power farm are
presented below where equation 16.3 is for hybrids, equation 16.4 is for
leghorns, and equation 16.5 is for the pooled data. The variables are
5
as follows:
4
The analysis which follows is based on unpublished statistical results derived by Earl
0. Heady and David W. Brown,
5
In addition to these variables, another was includedm initial calculations. It was
equipment (waterers, feeders, etc.). However, this variable was highly correlated with
hen numbers and the regression coefficient for it was nonsignificant at a 30 per cent
level of probability.
^
4507 869 ° 532

(16.3) Y = 85.76X X X^ X'5
(4.76) (5.54) (2.36) (2.55) (1.30)
(16.4) Y = 127.84X? 723 Xf Xf Xf xf 04 04 64 42
(5.02) (.48) (5.97) (3.08) (1.77)
(16.5) Y = 134.95X 1 3853 X 21132

- ,
X3216S X;13S8 X 5 670
(5.88) (3.95) (7.32) (3.93) (2.12)
Y is the number of eggs produced, X is number of hens, X 2 is square

x
feet of housing, X is hours of labor, X4 is pounds of corn equivalent of
3
grain fed, and X is pounds of 26 per cent protein concentrate equiva-
5
lent.
The R 2 value
was identical for equations 16.3, 16.4, and 16.5,
namely, .97, a quantity significant at the .01 probability level. The
values of t for the regression coefficients are included in parentheses
under the appropriate variable of each function. In equation 16.3, re-
gression coefficients or elasticities were significant at the .01 level of
probability for X x , X 2 and X 4 at the .05 level for X 3 and at the .2
, ; ;
level for X 5 In equation 16.4 they were significant at the .01 level for
.
X L , X 3 , and X 4 at the .1 per cent level for X 5 They were not signifi-
;
.
cant at the .5 level for 2 . X

In equation 16.5 all coefficients were sig-
nificant at the .01 probability level except for X 5 which was significant
at the .02 probability level. These statements suppose, in somewhat
the sense of experimental plots purposely selected by an agronomist or
lots of animals especially sorted out by an animal husbandman for their
experiments, that the data are for a sample representing a larger popu-
lation to which inferences may be made. It is true, of course, that the
data refer to a particular population of poultry flocks for which records
were kept. Interpretations of the derived statistics should be condi-
tioned accordingly.
Analysis was made, in the conventional manner of t, to test whether
significant differences exist between each parallel pair of regression
coefficients in functions indicated by equations 16.3 and 16.4. None of
these pairs of coefficients differed significantly. Hence, the pooled
function, equation 16.5, was used for estimates based on this poultry
analysis. (Functions based on the resistance formula also were com-
puted but appeared less efficient than those presented.)
The sum of elasticities in equation 16.5 is .9188. When tested
against a value of 1.0, this sum of elasticities does not differ signifi-
cantly from constant returns to scale. Hence, it appears that if hous-
ing, hens, labor, and feed are increased in equal proportions, egg out-
put also increases by the same proportion. This statement refers, of
course, to the average techniques found on farms in the study. Other
types of egg production techniques might display important cost econo-
mies, as compared to the techniques of this study.
Estimates From Functions
The pooled function, equation 16.5, was used to derive marginal re-
source productivities, given the form of function employed and obser-
vations from the particular sample explained above. For the detail in-
cluded in the analysis, predictions can be made for fairly refined input
categories such as number of hens, corn equivalent, and protein con-
centrate equivalent. The data also have been used for estimating the
effect of varying the number of hens, relative to a fixed housing ca-
pacity, on total egg production. For the 140 farms in the study the
mean square feet of housing was 3.4 per bird and mean egg production
was 212 per bird for the period of measurement. While they serve in a
somewhat different vein than the estimates of biological functions in
chapters 10 and 11, certain estimates can be made, over restricted mix
ranges, of average marginal rates of substitution between feed catego-
ries or other resource pairs.
Mean outputs and inputs are included in Table 16.11, along with
mean marginal productivities of the specific resources.
The mean marginal productivities are, of course, all lower than the
average product per unit of each resource. (For example, mean egg
production per hen was 212.) This holds true because the elasticity co-
efficient of each resource is less than 1.0 in equation 16.5 and dimin-
ishing marginal products are encountered at the outset. The average
egg productivity, at the mean of inputs, was 212 per hen, 63 per square
foot, and 219 per hour. The marginal productivities show, of course,
the predicted addition to egg production if each particular resource
were increased by one unit above its mean, the other resources being
held fixed at their mean.
The mean marginal productivity of corn feed and protein supple-
ment was, as predicted by the particular functional form in equation
16.5, nearly equal. However, this function would predict different mar-
ginal productivities for the two feeds combined in mixes other than that
represented by their means. But given such great similarity at the
means, it is suggested that, as an average, farmers in the group were
not combining feeds in profit -maximizing proportions. This is true
because the per pound price of protein concentrate is normally
Table 16.11. Mean Inputs and Mean Marginal Productivities

for Function 16.5
Geometric Mean Marginal

Output and
Mean for Productivity
Input Categories
140 Farms (Eggs)
Eggs (number) 61,170 —

Hens (number) 288.2 81.72 per hen
Housing (square feet) 979.6 7.05 per square foot
Labor (hours) 275.6 51.79 per hour
Corn (pound) 17,736.2 .47 per pound
Protein supplement (pound) 8,391.3 .49 per pound
considerably higher than for corn. Accordingly, it should be used in

proportions which cause its marginal productivity to be similarly
higher than for corn. The equation of marginal rate of substitution of
protein concentrate for corn is
6X 4 1
(16.6) -§x~= -.540X 4X r .
The numerical coefficient does not mean that 1 pound of concentrate

substitutes only for 1 pound of corn. This rate would hold true in case
hens were being fed the same proportion of the two feeds. For the
mean quantities in Table 16.11, the marginal rate of substitution of pro-
tein concentrate for corn is 1.14, the value of the partial derivative in
equation 16.6 when the X4 and X5
quantities of Table 16.12 are substi-
tuted into it. However, if corn were used in a ratio of 5 pounds to 1
pound of concentrate, the substitution rate would be predicted to be 2.70.
It is possible that “samples” of the type analyzed may provide an esti-
mate of only one slice of the production surface and substitution rates
may be estimated accordingly. As compared to all Iowa farmers, the
particular group providing the observations may feed relatively high on
protein. Hence, only the portion of the surface with low substitution
rates may be predicted.
Optimum “Crowding” of Hens
One facet of resource mixes which has concerned poultry men has
been the proportion of hens to housing. Recommendations have long
been made that a particular amount of housing space should be used per
hen; an assumption of limitational factors when recommended apart
from the price relatives for hens and housing. Analyses of the type
presented here provide a framework for such recommendations, even
though other functional forms and coefficients may eventually prove
more appropriate. Marginal rates of substitution between these factors
provide one basis for analysis of the “optimum degree of crowding.”
The “average” marginal rate of substitution of hens, derived from
equation 16.5, for housing space is
-1
(16.7) -3.407X2X! .
In comparing this rate of substitution of hens for housing space with the
price or cost per annum of hens and housing space, the optimum ratio
of birds and space might be predicted, given the particular form of the
production function used. For the mean quantities of inputs in Table
16.11, the marginal rate of substitution of hens for housing space is 8.25;
meaning that if egg production is maintained at 61,170, an added hen can
substitute for 8.25 square feet of housing. Based on equation 16.7, the
Table 16.12. Predicted Effect of Increasing Hen Numbers,

With Space Fixed at 979.6 Square Feet,
on Marginal Hen Productivity and
Total Egg Production
Increase
Marginal Total
Space in Total
Productivity Egg
Number Hens per Hen Egg
of Hens (Eggs) Production
Production
288.2 (mean) 3.40 81.7 61,170

308.2 3.18 80.7 64,568 3,398
328.2 2.98 79.7 67,920 3,352
348.2 2.81 78.8 71,239 3,319
368.2 2.66 77.4 74,053 2,814
388.2 2.52 76.7 77,300 3,247
rate of substitution would be 16.50 if the number of hens in the 979.6

square feet of housing space were 144 instead of 288. If the annual dis-
counted cost of a square foot of housing is 8.25 times greater than the
annual cost per hen, the mean quantities are, aside from time and un-
certainty of building commitments, optimum proportions of hens and
housing. If housing costs exceed hen costs by more than this propor-
tion, additional crowding would be profitable (within limitations of pre-
dictions by the particular algebraic form).
The Cobb-Douglas function overestimates for purposes of predicting
a profit -maximizing optimum. However, the function in equation 16.5
could be used for one set of such predictions. By setting the partial de-
rivative of egg production with respect to each particular resource in
equation 16.7 to equal the appropriate egg to resource price ratios, the
optimum quantities and ratios of resources would be specified. The
optimum of both quantity and ratio of resources will vary with the
prices of eggs and individual resources. It does not, even for more ap-
propriate functional forms, turn out to be a specific quantity, such as
3.5 square feet per bird, under all price conditions.
The mean marginal productivity of hen numbers increased relative
to housing space can be predicted, on the basis of the particular func-
tional form, from equation 16.5. It is, perhaps, more appropriate to
suppose that labor and feed are increased proportionately with hen
numbers relative to a given housing space. Hence, column 1 of Table
16.12 shows the number of hens increased relative to a fixed housing
space of 979.6 square feet. But at the same time, labor, corn, and pro-
tein feed are increased by the same proportion as hens. Only housing
space is held constant. Column 3 shows the marginal productivity of
hens for each of the resource mixes implied. Column 4 shows the pre-
dicted total egg production while column 5 shows incremental egg pro-
duction for the successive increases in hens, labor, and feed relative to
fixed housing space. Because of the algebraic nature of the function,
only a few predictions beyond the mean degree of crowding are used.
The optimum degree of crowding in the short run (i.e., before new
housing space can be built) thus is a function of the price of eggs and
the price of variable resources. The quantities in column 5 can be
compared with the ratio formed by dividing the price of a combined unit
of the variable resources by the price of eggs. For a quantity in col-
umn 5greater than this price ratio, the added crowding can be attained
to extend profits. Crowding should be extended up to any increment in
column 5 which is less than the price ratio.
Because (a) the function used for prediction assumes a constant
elasticity of production and (b) the sum of elasticities for variable fac-
tors considered in Table 16.12 is quite high, the marginal productivity
of hens does not decline rapidly. These data might well overestimate
the “degree of crowding” which is profitable under a given price situa-
tion. However, the estimates do indicate the principles which are ap-
plicable on recommendations to farmers relative to proportioning of
resources such as hens and housing.
Of course, a more exact method of specifying the optimum “degree
of crowding” is to equate the four partial derivatives (of eggs with re-
spect to hens, labor, corn, and protein) with the individual egg to re-
source price ratios, and solve for the magnitude of hens relative to the
fixed housing space (and for other variable resources as well). In a
long-run context, the optimum degree of crowding relative to any set of
price ratios would consider the derivative of the housing variable as
well. The five partial derivatives below would be equated to the price
ratios indicated, and magnitudes of X x through X 5 would be solved si-
multaneously .
(16.8) = • 3853YX >" hen cost -3- egg price
(16.9) - 113inrx ’" space cost -r- egg price

-§,
(16.10)
-§, 2165YX >-’ labor cost -S- egg price
(16.11)
•§. - i368yx ‘" corn price -r- egg price
(16.12) = .0670YXr protein price -3- egg price

0X 5
These quantities specify the optimum resource mix or degree of crowd-
ing in the long rim, subject to the limitations of the power function and
the particular sample.
TIME SERIES AND CROSS-SECTIONAL OBSERVATIONS
This section reports production functions derived from time series

6
and cross-sectional data in combination. The empirical method used
6
The analysis summarized here is from an unpublished study made by A1 Egbert,
Burton French, and Earl O. Heady.
is relatively simple and alternative approaches might have been used

for data of this type. The farms analyzed are those which kept records
with the Iowa Agricultural Extension Service over the 18 years,
1937-54. There were 20 such farms, providing 360 total observations.
(Observations between years are not necessarily independent and give
rise to statistical questions discussed earlier.) These farms with con-
tinuous records, but with considerable variations in inputs and outputs,
were selected in an attempt to help guarantee a study of a particular
production function. Only farms with Clarion-Webster soils in central
Iowa were used. All values were converted to a 1950-54 basis, so that
inputs and outputs might better represent physical quantities. Crop and
livestock production functions computed separately are, respectively:
514 97 ° 33 172
(16.13) Yc = 19.22X*1 Xf X-3 X'4
145 314 578
(16.14) Yl = 2.18Z; Zj Z3
The variables are the following: Yc is crop output measured in con-

stant dollars; Xjland in crop acres; X 2 is labor used on crops in hours;
X3 is annual expense of machinery used on crops in constant dollars;
X4 is other annual capital (fertilizer, seed, etc.) used on crops in con-
stant dollars; Y x is value of livestock in constant dollars; Z x is labor
used on livestock in hours; Z2 is livestock feed in constant dollars;
and Z3 is annual capital (livestock, buildings, equipment, etc.) used for
livestock in constant dollars. In derivation of these functions, each
quantity in time and space was used as a separate observation in com-
puting production functions. A time or trend variable was not signifi-
cant for the particular data. Neither could significant coefficients be
obtained for a moisture or rainfall variable.
The was .79 for the crop function and
coefficient of determination
.95 for the livestock function. All regression coefficients were signifi-
cant at a .05 level of probability, except for the power and machinery
variable for crops, even if the number of degrees of freedom is con-
sidered to be equal to the number of farms. The sum of elasticities did
not differ significantly from 1 0 ..
While the magnitudes are somewhat different, the relative relation-

ships among mean marginal productivities are similar to those re-
ported in studies summarized above. As Table 16.13 indicates, the
mean marginal productivity of labor was relatively high for both crops
and livestock. The marginal productivity of power and machinery, at
the geometric mean, was only .17 per $1 of annual expense. The rea-
sons for this small quantity are probably those explained earlier;
namely, the tendency for farmers to have overinvested for production
purposes. It is likely that this group of record-keeping farms, even
more than for farms in a random sample, has invested in machinery
for consumption purposes. The land productivity of $26.67 is somewhat
higher than the going cash rent of about $19 per acre; but fairly consist-
ent with the average share rent. It is not expected, because of risk and
Table 16.13, Inputs, Marginal Productivities, and

Elasticities for Equilibrium
Elasticity Necessary
Mean Marginal
Arithmetic to Give Marginal
Productivity in
Input Mean Productivity Equal
Dollars
to Factor Price
Crop function
Land (acres) 271 26.67 .347
Labor (hours) 1,433 2.91 .102
Power and
machinery ($) 2,794 .17 .224
Other capital
services ($) 1,870 1.29 .141
Livestock function
Labor (hours) 2,244 2.59 .056

Feed ($) 13,675 .92 .362
Capital
services ($) 19,367 1.19 .513
uncertainty, that farmers would buy or rent land until its marginal pro-
ductivity is driven to the market price or cash rental level.
Table 16.13 indicates, in column 4, the magnitude of elasticities
necessary to give marginal resource productivities, at the geometric
mean, equal to the prices of the particular resources. Using the as-
sumptions of random observations for all quantities in time and space,
a condition not entirely fulfilled, the significance of difference between
the individual elasticities in equations 16.13 and 16.14, as compared to
those of column 4 in Table 16.13, was tested. On this basis, the mar-
ginal productivities for labor on both crop and livestock, cropland, and
machinery and power for crops differed significantly from the market
prices for these factors. The marginal productivity of the crop power
and machinery input was significantly lower than the per unit price or
cost. However, the marginal productivity of crop capital services,
livestock capital services, and livestock feed did not differ significantly
from the prices of these factors. As an average among farms and over
time, it appears that these fairly progressive managers used less than
equilibrium quantities, for profit maximization, of land and labor, and
more than equilibrium quantities of machinery.
These inferences are consistent with the beliefs of extension per-
sonnel servicing these farms. Most of the managers use family labor
and may not hire an additional man for family reasons. In general
farms are of a size consistent with the total goals of the farm family,
including participation in nonfarm activities. Machinery investment by
these relatively high income families (average net income over the 18
years of $12,341 in 1950-54 dollars) has been extended into the con-
sumption category to allow greater enjoyment of work and to free time
for nonproduction activities. Yet within this family and institutional
setting, conditioning the mix of land and labor inputs, these farmers can
and do invest in livestock, feed, fertilizer, and similar resources to the
extent that marginal productivities might approach factor prices.
However, while they appear reasonable in terms of knowledge of ex-
tension personnel, the above interpretations need qualification. The
marginal productivities are means over time and between farms, com-
pared with prices converted to a 1950-54 basis and compared to aver-
age prices of this period. Since prices of individual years deviated
from this level, farms might well have equated productivities and
prices, or failed to do so, even if this condition had not been illustrated
under the empirical mechanism used above.
INTERREGIONAL PRODUCTIVITY COMPARISONS
This section reports a production function study which allows inter-

7
regional comparison of resource productivities. Random samples
were used in the four regions included in the analysis. Cobb-Douglas
equations were used and inputs and outputs were aggregated similarly
in all regions. Separate functions were computed for crops and live-
stock, with output measured in dollar values. Inputs for crops were
cropland in acres, labor in months, and all capital services measured
in dollar value of annual input. Inputs for livestock were labor in
months and all capital services measured as dollar value of annual in-
put.
Basic Statistics
As background, we present the data in tables 16.14 and 16.15.

These are based on the four random samples covering the 1950 produc-
tion year in the Alabama Piedmont, northern Iowa, southern Iowa, and a
dry -land farming area of Montana. Generally, the derived production
elasticities, which are presented in Table 16.14, can be accepted at
conventional probability levels. These statistics have been used in de-
riving the marginal quantities of Table 16.15. The different capital to
labor ratios lead to the a priori hypothesis that labor productivity
will
vary greatly in magnitude between the four areas. The ratio of capital
to labor is much higher in northern Iowa and Montana than in
southern
Iowa and Alabama; it is much higher in southern Iowa than in Alabama.
Labor productivity for crops is highest in northern Iowa and Montana
where the amount of land and capital per worker is greatest; southern
Iowa is slightly lower and Alabama is lowest. Probability tests do not
allow us to say that differences in marginal labor productivity are
7
See Heady, Earl O. and Shaw, Russell. Resource returns and
productivity coeffi-
cients in selected farming areas. Jour. Farm Econ., 36: 243-57.
1954; and Heady, Earl O.
and Baker, C. B. Resource adjustments to equate productivities
in agriculture. Southern
Econ. Jour., 21: 36-52. 1954.
Table 16.14. Regression Coefficients and Related Statistics
Item Montana Northern Iowa Southern Iowa Alabama
Crop production
Value of constant (log) .595 1.273 .718 .979
Value of elasticities
Labor .039 .076' .0881" .319*
Land .503* .912* .795* .385*
Capital services .580* .165* .393* .463*
Livestock production
Value of constant (log) .276 .359 .057 .737

Value of elasticities
Labor .084^ .077* .117* .233*
Capital services .937* . 907 * .982* .743*
Significant at:
*1 per cent probability level;
f8 per cent probability level;
J5 per cent probability level.
significantly greater in the first three areas. However, these orderings

are those expected from (a) the capital to labor ratios in the four areas,
(b) the knowledge of agricultural workers, and (c) the findings of previ-
ous scattered investigations. For capital and land resources used in
crops, differences in marginal resource productivity differ significantly
between the four areas. Important differences also exist between areas
for resources used on livestock. Returns for capital are highest in
Montana and southern Iowa where less of this resource is used than in
Table 16.15. Mean Resource Quantities, Marginal Product, and Gross Average Product of
Resource Services
Montana Northern Iowa Southern Iowa Alabama
Crop function
Arithmetic mean:
Product; actual ($) 21,419 8,551 4,777 1,322
Cropland (acre) 975.0 166.6 114.9 23.8
Labor (month) 13.7 9.5 8.7 10.4
Machine-crop capital services ($) 5,207 2,168 1,420 553
Investment per man year 67,866 62,430 32,064 3,255
Marginal product
Cropland not pasture ($/acre) 11.06 46.83 33.05 21.37
Labor ($/month) 61.80 68.04 48.05 40.57
Machine-crop capital services ($/$) 2.39 .65 1.32 1.15
Average product per month of labor 1,559 905 547 127
Livestock function
Arithmetic mean:
Product; actual ($) 12,084 13,943 9,067 1,336
Labor (month) 8.9 8.2 7.3 3.5
All capital service inputs ($) 8,896 12,543 7,614 1,017
Investment per man year 23,163 18,078 15,207 4,340
Marginal product
Labor ($/month) 113.92 130.76 144.83 89.09
Capital service inputs ($/$) 1,27 1.06 1.17 .97
Average product per month of labor 1,351 1,694 1,238 378
northern Iowa. They are lowest in Alabama where capital inputs hardly
allow livestock on a commercial scale. Since these basic figures are
analyzed in detail elsewhere, we proceed with the objectives of our
comparisons.
Equilibrium Between Products Within Regions
Having examined the mean marginal products for each resource

used in crop and livestock production in each area, we may ask: Are
these computed differences between crops and livestock significant? If
they are, our first conclusion may be that resources are not allocated
between products within a region to either (a) maximize farm profits or
(b) give a maximum national product from given resources. Since the
marginal return of crop capital in Montana is $2.39 while the return on

livestock capital is only $1.27, it appears that capital resources should
be shifted from the latter to the former product. On the basis of the
predicted marginal figures, capital should be shifted from crops to
livestock in northern Iowa and Alabama and in the opposite direction
for southern Iowa. Given the productivity figures in Table 16.15 for
labor, the other common resource for the two products, it should be al-
located from crops to livestock in four areas.
Since, however, the estimates involve sampling errors, our infer-
ences need to be subjected to a probability test. In terms of probability
and sampling error, we cannot say that the marginal productivity of
labor is higher for crops than livestock in Montana or northern Iowa
(Table 16.16). We can more nearly state that the differences expressed
for marginal labor productivity on crops and livestock are significant in
Iowa and Alabama. Greatest difficulty in a study of this kind arises in
the enumeration of labor inputs. While the farmer can quickly enumer-
ate the number of crop acres operated, the amount of fertilizer pur-
chased or expenses for tractor fuel (he usually has the latter recorded
with other expenses), it is difficult for him to specify actual labor in-
puts, except for hired help; he is prone to report the “stock” of oper-
ator and family labor available, rather than the amount actually used
in production. While detail of our study and questionnaires may have
partly eliminated this problem, it still exists. It is our hypothesis that
even more detailed labor records might lower the variance associated
with enumeration bias and actually unfold more important differences
in labor productivity. However, logic may exist for supposing that dif-
ferences in labor productivity do not exist in fact. The farmer controls
the allocation of the family labor, and does not have its use tied to cer-
tain resources or products as in the case of capital obtained through
the credit market. Accordingly, he may be more nearly able to allocate
labor to equate its marginal productivity between products than he is
for capital. Still, where capital and labor are technical complements,
the banker who specifies that the farmer can have credit for feeder
steers but not for seed or fertilizer also is specifying how labor should
.
Table 16.16. Value of t for Comparing Elasticity Derived in Crop Production with
Computed Elasticity Necessary to Give Marginal Productivity
for Crops Equal to Marginal Productivity for Livestock*
Resource Montana Northern Iowa Southern Iowa Alabama
Labor .39 § .77 § 1.43* 1.87*
Capital 2.20* 1.31* .56§ .62 §
*The simple test made here includes the following steps: First, the elasticity of pro-
duction necessary to give a marginal product in crop production equal to the marginal
product of the same resource in livestock production is computed. Our task is to
compare /3C the elasticity actually derived for crops with j3'c the elasticity neces-
, ,
sary to give equal elasticities. It is computed in the manner of (a) and (b):
?xftc
(a) (b) S' = Px
c
dR c Rc Rj; Yc Rx
The standard error, Sb p) for the pooled variances
(c) St
is then computed as in (c) and the value of t is given in (d)
0C - A (¥x R c /Yc R x )
<d)
t()<p< 5;
*10<p< 20;
§p< 40.
be used; labor for feeder steers rather than for crop services may
cause its marginal products to diverge even further, rather than to con-
verge between products.
In terms of probability, we may infer that the mean marginal pro-
ductivity of crop capital in Montana was significantly greater than for
livestock capital in 1950. Had Montana farmers known this outcome be-
forehand, they should have used more of their limited funds for crops
and less for livestock. They could not, however, predict the more fa-
vorable weather for crops in 1950 and their ex ante decisions to operate
as they did may not have been in error. Of course, many capital-short
farmers produce crops rather than livestock because they do not have
to tie up as great an investment per $1 of annual input or output as in
beef production. Following the figures of Table 16.15, less capital
should have been used for crops and more for livestock in northern
Iowa had farmers (1) been attempting to maximize profits and (2) known
the actual outcome of yields and prices. Since capital productivity is
higher for livestock than for crops in northern Iowa and since the value
of t is relatively high, we might infer that too much capital is used on
crops and too little on livestock in northern Iowa. While the level of the
computed productivity figures may not serve too accurately in predict-

ing exactly the population parameters, the ordering of productivities
and the inference outlined above seems to be in line with the structure
of production in the area. About one -half of the farms are rented;
many of these produce little or no livestock. Livestock would be profit-
able on these rented farms but because of uncertain tenure, many tenant
operators keep little livestock. On the other hand, tenants tend to in-
vest a relatively large amount of capital in machinery, one of the com-
ponents of the crop capital services outlined above. It is entirely likely
that at the margin, more investment in machinery (and hence in capital
services for crops) would give less than $1 in product for every $1 in
expense. Northern Iowa, relative to the products produced, is about as
highly mechanized for crops as any other United States region. Capital
for crops in the form of machinery may have been used, as a matter of
convenience and to ease the drudgery of farm work, beyond the input
level necessary for equating expected and discounted marginal costs
and returns.
Tests of Inter -Area Differences in Marginal Productivity
The regional differentials presented above correspond with the ex-

perience of agricultural extension workers and others closely ac-
quainted with the nation’s agriculture. It is known that farm families
in the central Corn Belt are generally more prosperous than those in
the southern Corn Belt; during years of favorable moisture, Great
Plains farmers, including those in Montana, have prospered and have
given a high out-turn of product for capital and labor; Alabama Pied-
mont farmers, as an average, are in an area where productivity is
known to be low and where the incomes of farm families are meager
relative to the national standard. These things are observed facts but
since we have samples from each area and for purposes of national
policies and guidance to individual farmers, we must ask: “Are the dif-
ferences in productivity coefficients presented significantly different be-
tween regions?” Since a productivity figure of any level involves sam-
pling errors, we must evaluate the differences in terms of the errors
attaching to each elasticity coefficient. This step has been taken by
means of simple significance tests and the relevant statistics are pre-
sented in Table 16.17. The first figure in each cell is the elasticity co-
efficient which would have been necessary, considering the mean
quan-
tity of product and resource in the area of comparison,
to give a
marginal productivity 'equal to that computed in the area of contrast.
For example, an elasticity of .0435 for labor in Montana (as compared
to the Montana sample coefficient of .039 shown in Table 16.14) would
have been necessary to give a marginal product of $68.04, the northern
Iowa average, in Montana (as compared to the sample prediction of
$61.80 for Montana in Table 16.15). In comparing the .0435 elasticity,
as a constant, against the actual elasticity of .039 in a null hypothesis
t
t
Table 16.17. Comparison of Differences in Marginal Productivities of Resources

in Different Areas; Elasticity Coefficient Necessary to Give
Marginal Resource Productivity in One State Equal to
That m
Another State, t Values (in Parentheses)
and Significance Level*
Resource and State State for Which Test is Made

Against Which Test
Montana Northern Iowa Southern Iowa Alabama
is Made
Crop production
Labor:
Montana .0684 ( .05) § .1122 ( .12)§ .4846 ( .19)5
Northern Iowa .0435 ( .05)§ .1239 ( .387)5 .5352 ( .62)5
Southern Iowa .0307 ( .12)§ .0534 ( .38) § .3780 ( .25)5
Alabama .0260 ( .19) § .0451 ( .62) § .0739 ( .25)5
Land:
Montana .2159 (8.56)* .2661 (5.63)* .2007 (1.70)*
Northern Iowa 2.1266 (8.56)* 1.1247 (3.76)* .8405 (3.85)*
Southern Iowa 1.5029 (5.63) .6448 (3.7 6)f .5994 (1.81)*
Alabama .9646 (1.70)t .4138 (3.85)t .5101 (1.81)*
Capital:
Montana .6053 (3.05)* .7085 (1.91)* .9626 (2.28)*
Northern Iowa .1579 (3.05)* .1931 (1.78)* .2619 (1.55)*
Southern Iowa .3214 (1.19)t .3352 (1.78)* .5330 (3.50)*
Alabama .2790 (2.28) .2910 (1.55)* 1.2327 (3.50)*
Labor:
Montana .0670 ( .17)5 .0917 ( .32)5 .2984 ( .32)5
Northern Iowa .0963 ( ,17)§ .1052 ( .15)5 .3426 ( .57)5
Southern Iowa .1067 ( .32)5 .0852 ( ,15)§ .3749 ( .85) 5
Alabama .0656 ( .32) § .0524 ( .57)5 .0717 ( .85)§
Capital:
Montana 1.1450 (3.94)* 1.0688 (4.38)* .9689 (3.70)*
Northern Iowa .8063 (3.94)* .8464 (2.94)* .7672 ( .47)5
Southern Iowa .7229 (4.38)* .8834 (2.94)* .7475 (2.75)*
Alabama .7186 (3.70)t .8782 ( .47)5 .8198 (2.75)*
*The marginal product in the area of comparison equal

elasticity coefficient (to give a
to the value considered as a constant) computed as the marginal product in the area
of contrast is computed this way: The marginal product, based on the total product,
and the input of the particular resource as averages for the area are used. We wish
to compare the elasticity, /3a of resource X a estimated for region a, with ,
the
elasticity necessary to give the marginal product of region b, when X a and Ya are ,
those of region a. Hence we have (a) and (b) for estimating /3 a .
Y b Xa
,
(a)
x
—dY=—a —dY—b
= = ——Y ——
/3a Ya
=
0b b
(b)
a =
/S'
Vy~ x
j -
-
dXa dx b x x a b a b
Finally, the value of t has been computed as in the footnote of Table 16.16.
Significance level in probability:
to < p < 5;
*5 < p < 8;
5p < 40.
sense, we obtain a value of only .05; a value which is not significant at

an acceptable probability level for t.
We conclude, considering the quantity of resources used, mean mar-
ginal labor productivity in Montana is not significantly lower than in
northern Iowa. Since northern Iowa uses a different quantity of re-
sources than Montana, we again compute the elasticity coefficient nec-
essary for the marginal productivity of labor in northern Iowa to equal
the marginal productivity of labor in Montana. The elasticity of .0684,
necessary to give the $61.80 marginal labor productivity of Montana, in
northern Iowa, does not differ significantly from the sample elasticity
of northern Iowa.
The coefficients of neither crops nor livestock differ significantly
among the four areas; however, a test of the average gross products for
labor shows significant differences. Thus, at the margin of the mean,
labor productivity does not differ significantly; but for parallel months
of labor and for smaller quantities, it may do so. This comparability
of marginal labor product is in contrast to certain previous inferences
on productivity based on extremely gross, residual data for parts of the
same regions. From these data, we cannot even say that at the margin
of mean labor inputs marginal productivity of labor in Alabama differs
significantly from that of other areas, although we know that it differs
in average productivity terms. It is entirely possible that the return to
“the last month of labor” used “on the average farm” in the four areas
is quite similar. However, enumeration of labor inputs is the most dif-
ficult part of this type of study. We believe that significant differences
in mean marginal productivity might be obtained in such contrasting
areas as northern Iowa and Alabama by obtaining the labor records and
by lessening variance through samples stratified more closely by soil
type and techniques of production.
As would be expected, differences in marginal productivity of land
are significant. Not only of interest in farmer investment decisions,
these data are also important for certain production policy questions.
For example, withdrawal of an acre per farm under a control program
would have greater effect in lowering total agricultural production in
northern Iowa; and the least in Montana, while the effect in southern
Iowa would be greater than for Alabama. Or, during war or other food
emergency, it is more important to keep an acre of land in northern
Iowa in production than one in Montana, southern Iowa, or Alabama.
Tests of Disequilibrium in Marginal Resource Productivity

and Factor Prices
With our four small sectors of the total economy, we can investigate
certain conditions of general equilibrium; namely, the extent
to which
the marginal productivity of resources approach factor
prices. (Also,
our examination of intra-area and inter -area differences above
involved
the general equilibrium or stability conditions relating to
the inter -firm,
Table 16.18. Comparison of Marginal Productivities From Resource

Departure of
Prices. Necessary to Give Marginal
Elasticity Coefficients
Product Equal to Market Price of Resource, Values of t
in Parentheses
Resource Montana Northern Iowa Southern Iowa Alabama
Crop production
Labor X.0234 (13.8) .1444 (1.41) .2368 (2.84) .5900 (4.09)
Land .2504 ( 2.66) .3408 (8.28) .3505 (5.60) .2179 (1.66)
Capital .2577 ( 2.64) .2688 (1.52) .3151 (1.97) .4518 ( .13)
Labor .1178 ( .62) .0765 ( .01) .1047 ( .24) .1965 ( .57)

Capital .7804 ( 3.76) .9536 (1.44) .8901 (2.66) .8222 (1.48)
inter -product, and inter -area organization of resources and production.

It is, markets are imperfect and have bar-
of course, true that factor
riers between them; in this situation, unequal marginal returns may not
mean market disequilibrium. To make the comparison, we computed
the elasticity coefficient necessary for each resource (given the mean
input of all resources and the mean product in each area) which would
have resulted in a marginal product equal to the factor price. We then
tested the significance of the difference between this elasticity coeffi-
cient and the one estimated from the sample (Table 16.18). The mar-
ginal productivity of labor for crops appears to be significantly lower
than the monthly hired wage rate in Montana and the two Iowa areas; the
inference for Alabama is that the low marginal product of labor on crops
does not differ significantly from the low wage rate for hired workers.
Differences in availability of off-farm and part-time work in Alabama
may partly explain this difference. While off-farm work during winter
months on a Montana wheat farm is not entirely unavailable, it is not
plentiful and convenient; the Iowa areas are somewhat similar. Hence,
a marginal product lower than the wage rate (weighted particularly by
summer crop work) can be consistent with rational choice in areas of
few outside opportunities.
Marginal land productivity was significantly greater than rental
rates for all areas in 1950. (The rental prices used are share rates;
cash rates would be lower and hence t values would be even greater.)
Tenant operators or purchasers of land were realizing a short-run re-
turns margin above the rent of land. Some difference is expected, of
course, because of uncertainty considerations. Contracts are made into
the future and with discounting of anticipated returns for risk, an ex
ante equilibrium may exist in a subjective sense even if the marginal
productivity and price for land services differ in ex poste measurement,
the manner of the data in this study. Also, weather and lack of pests
were favorable in 1950, except perhaps in Alabama; weather was per-
haps better than farmers’ modal expectations, which extend over sev-
eral years. Farmers’ price expectations in postwar years had centered
-
on declining prices; rental contracts, made 1-3 years ahead, had not
accurately anticipated continuance of inflationary prices.
Capital returns were significantly greater than the cost of capital
for crops in Montana and southern Iowa; it appears that mean marginal
return of capital was lower than the cost of capital in northern Iowa.
(This phenomenon may be a function either of sampling variance or the
machine investment situation mentioned earlier; machinery investment
often is driven to a point to give ease of work and lessen drudgery on
many of these high capital farms.) In terms of probability, we must in-
fer that differences did not exist between the cost and marginal product
of capital in Alabama. This statement refers to the capital forms and
techniques being used; capital in the form of other techniques and in an
amount to allow gains of scale economies might allow much higher re-
turns.
SIMULTANEOUS EQUATIONS
The functions reported in this chapter were estimated by least -
squares, single equation methods. Several attempts have been made to
use simultaneous equations models in estimating farm production func-
tions. However, none of these have met with success. This was true
for such a model applied to the data from record-keeping farms dis-
cussed in a preceding section. 8 One difficulty evidently is in the high
inter correlation between the numerous variables necessary in a logical
simultaneous equation model. Perhaps another reason is that the basic
suppositions of simultaneous models may have no special relevance in
terms of whole -farm production functions. While reasonable decision
models might be established, there appears to be no real biological or
physical basis for supposing that the magnitudes of inputs and outputs
are interdependent. Once a farmer has made a decision to use a par-
ticular input, the process is purely physical, with input magnitude
af-
fecting output magnitude but without the opposite holding true, and
should be so reflected in the model. Except for a few minor inter
year exceptions, it appears that inputs can be considered purely as
exogeneous variables in their effects on output. In these cases we are
concerned with the quantitative effect of input on output, once
decision
has been made to use a particular magnitude of input, rather
than with
the decision-making process leading to a particular input 9
level.
8
Also, as another example, see French, Burton L. Estimation by
simultaneous equa-
tions of resource productivitiesfrom time series and cross-sectional farm observations.
Unpublished Ph.D. thesis. Iowa State University Library, Ames. 1952.
S
1
®U S K e n’ L '
tL
l 226-36.
li
R A textbook of econometrics. How Peterson,’ Evanston. 1953.
pp.19o~6,
CHAPTER 17
Comparison of Production Function
Estimates From Farm Samples

Over the World
N TODAY’S world problems, both economic and social, resource

productivity and the efficiency of resource use play a vital role. In-
I ternational productivity differentials have, of course, always been of
major concern to economists; implicit assumptions about productivity
coefficients have provided the basis for most of the literature dealing
with world exchange and national welfare maximization. Since World
War n, however, resource productivity in particular regions of the
world and international productivity differentials have taken on addi-
tional importance; stirred on the one hand by ideological currents, and
on the other by recognition of the poverty in the world and the imminent
explosion of world population. Little wonder, therefore, that the need
for increasing the productivity of resources and the efficiency of re-
source use in particular areas of the globe has been deemed especially
important, as evidenced by the funds and services that the more ad-
vanced countries have invested in an endeavor to assist less developed
areas of the world.
In the main, these investments have been concerned with increasing
the productivity of resources by the introduction of new technologies.
By replacing the existing production functions by new ones that lie
higher in the output plane, it is hoped that these investments will lead
to the production of more output from the same quantity of inputs, or to
the same output from fewer inputs (thereby releasing resources for
other economic activities). Only a minor portion of these international
development funds have been concerned with increasing the efficiency
of resource use within the context of a given production function. The
latter, of course, is the more difficult task since its success depends on
the availability of empirical information for an array of situations.
Still, as is evidenced by the data presented in this chapter, there are
significant gains to be made by increasing the efficiency of resource
use within given technological environments. The collection of the nec-
essary data is likely to be well worth the cost. This will become in-
creasingly so as the technological lags throughout the world are dissi-
pated and the opportunities for shifting to new production functions
become limited. It is, therefore, worthwhile to draw together some of
the production function studies for farm -firms that have been carried
out in various countries. The primary value of comparing such
586 FUNCTIONS OVER THE WORLD
analyses lies in the direct indications they offer of economic disequi-

libria in the use of resources, both inter- and int r a -nat ionally
. Such
quantitative evidence is essential to any attempt to improve or opti-
mize the allocation of scarce resources either within or between na-
tions. Moreover, such data may serve to indicate in what directions
and to what extent it may be desirable to change the value systems of
farmers and politicians so as to free the way for greater resource use
efficiency. As well, comparison of production function estimates pro-
vides empirical evidence of the differentials that exist in resource pro-
ductivities between and within countries. Such data may enable us to
assess the relative need for the introduction of more advanced produc-
tion processes and to decide, where necessary, between developmental
programs oriented to the adoption of new production functions and those
programs whose primary aim is to increase resource use efficiency
within a given technological environment. Needless to say, when new
production processes are substituted for old, every endeavor should be
made to ensure that resources are used in optimal fashion under the
new methods.
To these ends, and also to provide a resume of research to date, an
array of cross-sectional farm -firm studies from around the world are
collated in the present chapter. Without fully searching the fugitive
literature, we have endeavored to include representative studies from
as many nations as possible. It has, of course, been impossible to in-
clude all the production function analyses that have been made; for the
United States and Japan it has only been possible to present a selection
from among the many available studies. Nor have we been able to lo-
cate studies for a number of countries — notably the Communist coun-
tries and those of South America. Still, the analyses to be compared
cover a large portion of the globe; in particular, they relate to the
United States, Canada, New Zealand, Australia, the Union of South Af-
rica, the United Kingdom, Austria, Sweden, Norway, Israel, India,
Japan, and Taiwan. Moreover, a number of functions covering differ-
ent segments of their agricultural production are presented for most of
these countries. Within each country, and internationally, the different
types of farming studied in these analyses are of varying degrees of
importance. As a result’ as the implications of each estimated function
is assessed, one must bear in mind the relative importance of the pro-
duction process to which it refers. Both inter- and intra-nat ionally, for
instance, a production function estimate for wool growing in Australia
is of far more significance than an estimate for sweet potato production
in Japan. Likewise, in evaluating the implications of the analysis for
individual resources within each of the production processes examined,
account must be taken of the relative importance of each resource. A
small measure of inefficiency per input unit in the use of a major re-
source may have severer repercussions than a large degree of ineffi-
ciency per input unit in a factor that is only used in small amounts.
Too, for an ideal evaluation, these agricultural studies should be com-
plemented by similar analyses for other regions and industries in the
FUNCTIONS OYER THE WORLD 587
various countries. However, comparable studies for nonagricultural in-

dustries are not available, although quite a number of nonagricultural
time series analyses covering some or all of the period since 1900 have
1
been carried out.
Since any over-all comparison of the various studies would be pre-
sumptuous without some understanding of the setting of each analysis,
we first present the studies on a country by country basis. Before do-
ing so, the following comments are necessary. All of the studies are
based on Cobb-Douglas type functions fitted to cross-sectional farm
data. Each function refers to a production period of 12 months within
the years 1950 to 1958. Studies referring to prior years have been ex-
cluded in the hope of maintaining a reasonable degree of homogeneity
with respect to time. Country by country, the estimated functions and
their related statistics are presented in tabular form. To facilitate
comparisons between and within countries, all land, labor, and monetary
2
statistics are listed in the following units: land in acres; labor in man-
3 4
months; and capital in 1954 United States dollars. Comparisons be-
tween and within countries will generally be made as though each of the
estimated functions related to the same calendar time period. This as-
sumption is not thought to be overly strong, since all® the functions re-
late to years between 1950 and 1958. Also, in making comparisons, the
market prices or opportunity costs prevailing at the time of each study
(but converted to 1954 United States dollar values) will be used as
benchmarks for estimating resource use efficiency. The assumption is
that these prices are not atypical to any major degree; although it is
certainly realized that agricultural prices, especially of products, are
not as stable as one would desire for such purposes. Assuming no ma-
jor technological or scale changes to have occurred so that the esti-
mated functions would still be relevant, it would be feasible to evaluate
the various functions according to current prices. The lack of the nec-
essary data has prevented the presentation of such an extensive analy-
sis here. Also, although comparisons are made without always men-
tioning that we are talking about the situation on an average farm, or
the situation that prevails on the average in the region relevant to each
fitted function, such a proviso is implicit in most of the discussion.
Likewise, in discussing resource transfers, we assume that allowance
*For a survey of many of these studies, see: Leser, C. E. V. Statistical production

functions and economic development. Scottish Journal of Political Economy, 5: 40-49. 1958;
Kuhlo, K. C. Zur Auswertung und Ausgest der Produktionsfunktion. IFO-Institut fur
Wirtschaftsforschung. Munich, 1959.
2
The conversion factors used were those listed in Webster's New International Diction-
ary. Merriam -Webster Company. Springfield, Mass. 1958. Pp. 1521-24.
3
Where necessary, conversion rates from hours and weeks to months of labor were ob-
tained by personal communication with the authors of the original production function esti-
mates, as cited later.
4
With the exception of Taiwan, the exchange rates used were those recorded by the In-
ternational Monetary Fund as listed in United Nations Monthly Bulletin of Statistics. 13:
161-67. 1959. For Taiwan, an exchange rate of 24.78 New Taiwan dollars to the United
States dollar was used. Adjustment to 1954 United States dollar values was made by way of
a composite index based on the cost of living and wholesale price indices listed in the United
Nations Statistical Yearbook. 1958. Pp. 408-28.
is madefor the natural and institutional restrictions that exist relative

to the mobility and quantity of existing resources.
While none of the functions selected for inclusion can be regarded
as perfect estimators or representations of the production processes
studied, they are all judged to be statistically reasonable. Still, it must
be remembered that they are cross-sectional studies; in consequence
they suffer in varying degree from the disadvantages peculiar to the
cross-sectional approach as outlined in Chapter 6. As well, it should
be recognized that some of the functions do not relate to a single prod-
uct; nor, in some cases, are the input categories aggregated and meas-
ured in the most desirable fashion. In this regard, there are lessons to
be drawn from the various approaches taken in these studies, bearing in
mind the recommended procedures outlined in chapters 4 to 7 Too, all .
of the analyses are based on single equation models when, in some

cases, a multiequation approach may have been more satisfactory; to
say nothing of the possibility that a single equation model based on
some algebraic model other than the Cobb-Douglas might have been
more appropriate instances. Still, within these limitations, the
in some
usefulness of the fitted functions should not be decried. For many of
the countries they constitute the only empirical estimates of resource
productivities and measure of resource use efficiency that are availa-
ble. While they may not be of great use for recommendations to indi-
vidual farmers, they are of significant importance for the determination
of over -all policies dealing with the allocation of land, labor, and capi-
tal within countries and for the rational determination of the allocation
of scarce capital funds and technical services between nations.
So far as shifts in resource use are concerned, it is true that only
capital has any great measure of international mobility. Still, labor has
some degree of mobility between countries as evidenced, for instance,
by the transfers that have occurred in recent years from Japan to South
America, from Europe to Australia, and from China to Tibet. Land, of
course, has no international mobility in the usual sense. But significant
changes may occur in its ownership — as can be seen by comparing a
1938 map of the world with one for 1958; or by enumerating the inter-
national land purchases —
Louisiana, Texas, and Alaska — by the United
States. Intr a -nationally the allocation of any economic resource is
,
subject to some measure of centralized control, as well as the ordinary

market influences. Witness the land reform and consolidation pro-
grams that have been carried out recently, or are underway, in many
countries; the direct and indirect devices that have been used to induce
domestic labor transfers and the centralized institutions that have been
developed to control, to varying degrees, the availability of credit to
5
various sectors of most countries* economies.
The more important statistics to be presented for each function are
noted below. Since detailed explanations of the derivation and
5
For a resume' of such national policies as they relate to agricultural resource alloca-
tion see: The state of food and agriculture 1958. F.A.O., Rome. 1958.
FUNCTIONS OVER THE WORLD 589
implications of these statistics have been given in chapters 2, 3, and 4,
only brief explanatory comments are made here.
Production Elasticities
For each input resource, these coefficients indicate the expected

percentage increase (or decrease) in production that would occur if the
amount of the input resource was increased (or decreased) by 1 per
cent, other input levels being held constant. Since the models used are
of Cobb -Douglas type, these estimates of the elasticities remain un-
changed over the range of input levels to which the function is fitted
and to which it might be applied. Such constant elasticities are most
likely not a true reflection of the real-world situation. They are
merely an algebraic attribute of the Cobb-Douglas model. Still, pro-
vided other restrictions are also met, the estimated elasticities do
validly represent the real-world situation at the geometric mean input
levels. Accordingly, in drawing out the implications of the elasticities,
it is always assumed that we are talking of incremental changes at the
mean input levels. For each function, those elasticities that are signif-
icantly different from zero at the 5 per cent probability level are noted.
It is recognized that such tests do not strictly apply in those instances
where the estimates are based on a nonrandom sample of observations.
Also, regardless of these significance tests but with the provisos out-
lined in Chapter 6, the estimated elasticity is still the most likely value
of the true elasticity. Indeed, there is nothing sacred about the 5 per
cent probability level. It would be far better to ascertain the proba-
bility level of error at which the relevant policy makers desire to op-
erate and to test the estimates accordingly.
Sum of Elasticities
If one is prepared to assume that no relevant input factors have
been excluded, this sum provides an indication of returns to scale. On

the other hand, if one believes constant returns to scale prevail, the
sum may be regarded as indicating the importance of omitted variables
if it is less than unity, or as roughly indicating the extent of anomalies
in the mode of aggregation if it is greater than unity. Given all the lim-
itations associated with the measurement and interpretation of the
over -all production elasticity, as noted in Chapter 7, we do not stress
this statistic in discussion. However, for completeness, we have indi-
cated whether or not each sum of elasticities is significantly different
from one at the 5 per cent probability level.
Adjusted Coefficient of Multiple Determination, R2

This statistic gives the proportion of the variation in the output ob-
servations explained by the fitted function, correction having been made
for the size of the sample studied. In every instance, the statistic is
significantly different from zero at a probability level of 5 per cent or
less.
Sample Means of Outputs and Inputs
Generally, the geometric (rather than arithmetic) mean levels are

presented for they are most relevant to the estimates of the marginal
products. Since there is probably not a great deal of difference be-
tween the geometric and arithmetic sample means, the listed means
serve satisfactorily as indicators of the scale of the farm operation to
which each function refers. As well, the relative importance of differ-
ent resources can be judged from their mean levels, thereby enabling
comparative assessment of the importance of disequilibria in the vari-
ous resources. The outputs listed are always gross output per annum,
measured in 1954 United States dollars. Land, labor, and capital are
always shown in acres per annum, man months per annum, and 1954
United States dollars per annum. The mean total quantity of inputs used
is presented in money terms. These were obtained by valuing the land,
labor, and other nonmonetary inputs at their opportunity costs, adding
this sum to the monetary inputs used. In a few of the functions, build-
ings, livestock, and machinery have been analyzed in terms of their
asset value, rather than indirectly through their annual services as
should be done. Where such was the case, these assets have been
valued at their annual cost in deriving the mean level of all inputs. In
some instances, the opportunity cost method of valuation has probably
led to overestimation of the “all inputs" mean. For instance, in many of
the situations studied, much of the labor used is family labor which was
paid less than hired labor would have cost.
Average Products
The average product of each resource is computed as the mean out-

put divided by the mean input of the resource. The resultant
average
includes the product returns of all inputs, and not simply the
product
return attributable to the single resource. Mostly, the average
prod-
ucts are presented for the geometric mean levels of output
and input.
Frequently due to a lack of marginal productivity data, average
,
prod-
ucts are used as measures of efficiency. 6 We will not
discuss them in
such fashion, for the obvious reasons that they generally
have no strict
validity as indicators of resource use efficiency and
that we have data in
marginal terms. Moreover, if analyses of average productivity
were to
be made, far better statistical sources are available
than the few sets of
data we present as byproducts of production function
research.
For example, see Strand, E. G. et al. Productivity of resources used on commercial
farms. US,DA Tech. Bui. 1128. Washington, D.C. 1955.
Marginal Products
By necessity, these estimates are given in value terms, physical

data being unavailable in most instances. Each marginal product indi-
cates the expected increase in output forthcoming from the use of an
additional unit of the relevant input, the level of other inputs remaining
unchanged. In general, the marginal productivity of any resource de-
pends on the quantity of it that is already being used and on the levels of
the other resources with which it is combined in the production process.
For this reason the estimates with widest applicability are those de-
rived at mean input levels, with the proviso that the most accurate es-
timates from Cobb-Douglas functions are obtained with all inputs held
at their geometric mean level. Accordingly, wherever possible, mar-
ginal value products are listed for the geometric mean input levels.
But in some cases, due
to lack of information, it has been necessary to
record the marginal products measured at the arithmetic mean input
levels. In none of these instances is it likely that great differences ex-
ist between the estimates of marginal products as made at the arith-
metic means and those estimates that would have been obtained using
geometric mean input levels.
Opportunity Costs
For the various resources, these are the market prices that pre-
vailed, on the average, over the twelve months to which each production
function refers. The market price of land has always been taken as the
annual cost of renting one acre since we are interested in land services
rather than land per se. On the assumption that the employment of ad-
ditional labor would imply the purchase of hired labor, the monthly wage
rates quoted refer to hired labor. In some instances this assumption
probably overvalues the opportunity cost of labor since additional family
labor might be available at a cheaper price. Where necessary, an al-
lowance for food and lodging has been included in the wage rate. The
opportunity cost of a dollar of capital has been taken as one dollar plus
the relevant annual interest charge. For capital assets, the opportunity
cost has been taken as the annual straight-line depreciation associated
with the assets plus service charges where these have not been included
in other input categories. Following the practice adopted in the original
sources, it is assumed throughout the analysis that currently used
services of land and labor are purchased out of current funds and thus
involve no interest burden. Increases in other input services are as-
sumed to require borrowed capital with its concomitant interest
charges. Realistically, each situation should be evaluated according to
how the average farmer would finance additional inputs. However, any
inconsistencies resulting from the use of the costing procedure adopted
are thought to be slight. In listing but a single opportunity cost for each
resource in each function, it is assumed that the farm population
relevant to each sample operates under free competition; the individual

farmer having no control over the prices he pays or receives. While
infringements of this assumption may occur, they must be regarded as
minor for all practical purposes.
Marginal Product to Opportunity Cost Ratios
Within the limits these ratios provide a

of statistical reliability,
measure resource use prevailing, on the average,
of the efficiency of
throughout the population of farms relevant to each of the samples
studied. Ifthe ratio is less (greater) than one, it indicates that too
much resource is being used under the ex-
(too little) of the particular
isting price conditions, given the levels at which other resources are
operating. Maximum efficiency in resource use occurs when the reve-
nue from using one additional unit of input is equal to the cost of that
additional unit, i.e., when the marginal product to opportunity cost ratio
is equal to unity. Under such conditions entrepreneurial profits will be
maximized. As well, Pareto optimality with respect to national welfare
will also tend to be obtained; nobody could be made better off without
somebody being made worse off. Strictly, however, if marginal costs
and revenues are equated in agriculture, Pareto optimality will only
be attained if marginal costs and revenues are also equated in all other
segments of the economy. In the few instances for which appropriate
data are available, whether or not the marginal product to opportunity
cost ratios are significantly different from unity at an acceptable prob-
ability level is noted in the tables. No direct indication is given of the
quantitative changes in resource use that would be required in order to
drive the marginal product to opportunity cost ratios to unity. This in-
formation is obtainable for most of the studies from the original
sources cited. In money terms, a greater change in input quantity may
be needed to induce equilibrium in the use of a major resource that is
already used fairly efficiently than would be needed to induce equilib-
rium in a factor that is used very inefficiently but in only small
amounts. In other words, while the marginal revenue to opportunity
cost ratios indicate where the next dollar might best be added to or
subtracted from the input mix, they provide only a tentative guide rela-
tive to the next thousand dollars. It should also be noted that in those
instances where the sum of the production elasticities is equal to or
greater than unity, the specification of the fitted function prevents the
attainment of a unique equilibrium position relative to all inputs. But in
the real world situation it is certain that an equilibrium position either
exists within the range of input levels examined but is masked by errors
of specification, or could eventually be attained. In the latter case,
barring specification errors, the equilibrium position could be approxi-
mated by a function fitted to data for input levels beyond those examined
in the listed studies to which this comment applies.
With the above preliminaries in mind, we may now consider the
estimated functions on a country by country basis. In large part, the

comments on each study are but a precis of those to be found in the
original references listed for each country.
UNITED STATES
Relative to other countries, there have been many cross-sectional

production function studies of United States farm -firms. We have
chosen two of these studies for present purposes, statistics for both be-
ing listed in Table 17.1. The first study was designed as a comparative
7
analysis of four areas of United States agriculture. These areas were
(a) the Piedmont region of Alabama, which is typical of a sharecropper
farm organization built around cotton and a scattering of other crops

and livestock products; (b) southern Iowa, which is somewhat typical of
the less productive portion of the United States Corn Belt where capital
input per farm and soil productivity are lower than in the central Corn
Belt; (c) northern Iowa, which is typical of the most productive parts of
the Corn Belt, where capital availability is about as great as in any
sector of United States agriculture; and (d) a dryland wheat area of
Montana, which also includes a considerable amount of cattle ranching
under grazing conditions. Between them, these four regions typify a
major part of United States agriculture. The actual data analyzed was
obtained from a random sample of commercial farms in each area.
Over these samples, capital investment per farm was almost 50 per
cent greater in Montana than in northern Iowa, which in turn was nearly
double total investment per farm in southern Iowa; the latter having
nearly eight times as much capital per farm as the Alabama group.
Since crop and livestock production are distinct activities on the farms
within each area, separate production functions were derived for the
crop and livestock activities. Hence, the four -state study involved the
estimation of eight production functions, all of which are listed in Table
17.1.
The second United States study presented relates to various farm
tenure classes in the Iowa -Illinois area of the Corn Belt. As shown in
Table 17.1, separate production functions were estimated for stratified
random samples of farm operators who were full owners, livestock-
8
share renters, and crop-share cash-renters. To reduce the confound-
ing influence of age, only owners under 45 years of age were sampled;
7
Heady, E. O. and Shaw, R. Resource returns and productivity coefficients m selected
farming areas. Journal of Farm Economics, 36: 243-57. 1954; and, Resource returns and
productivity coefficients in selected farming areas of Iowa, Montana and Alabama. Iowa
Agr. Exp. Sta. Res. Bui. 425. Ames. 1955.
Miller, W. G. Comparative efficiency of farm tenure classes m the combination of re-
8
sources. Agricultural Economics Research, 11: 6-16. 1959; Miller, W. G., et al. Relative
efficiencies of farm tenure classes in intrafirm resource allocation. Iowa Agr. Exp. Sta.
Res. Bui. 461. Ames. 1958. The reader is referred to these sources for exemplification
of the various statistical tests pertinent to any detailed analysis of resource use disequilib-
ria based on production function estimates.
they are more comparable to tenants who are generally under 45 years,
The functions fitted for these three tenure samples relate to the whole
farm operation made up in each case of crop and livestock activities.
Since the stratification of the farms in the tenure study cuts across
that of the four -state study, we will first consider the two sets of func-
tions separately.
Crop and Livestock Studies
—
Between the four regions, the crop functions as specified in Table
17.1 by their resource elasticity coefficients - are rather dissimilar.
Still, as would be expected, the least difference occurs between the
southern and northern Iowa functions. Between these areas, climatic
and organizational features of the crop enterprise are not as distinctive
as they are for the other regions. Over the four samples, the influence
of the differences in crop resource elasticities is reflected in predicta-
ble fashion in the resource marginal products. Marginal productivity of
crop land is ordered as would be expected in terms of soil type, rain-
fall, and climatic conditions. It is highest in northern Iowa ($46.70 per
acre), followed by southern Iowa ($32.32 per acre), Alabama ($19.11 per
acre), and Montana ($10.60 per acre). Marginal labor productivity on
crops displays a differential to be expected from the capital to labor
ratios and resource quantities found on the sample farms. It is also in
line with the relative differences known by agricultural workers to ex-
ist in the areas. Northern Iowa and Montana have the highest labor
marginal value products, $67.07 and $58.17 per month, respectively.
Small farms and a smaller quantity of capital per labor unit cause the
marginal productivity of labor to be lowered in Alabama and southern
Iowa, where marginal labor returns are $39.62 and $43.40 per month,
respectively. The largest marginal productivity of crop capital serv-
ices was found in Montana and southern Iowa. The relatively low re-
turns in northern Iowa are perhaps best explained by the machine com-
ponent of the capital services. These farms are as highly mechanized
as any group in the United States. On the average, added machine in-
vestment would likely add less to production on these farms than the
annual cost of the machine services.
Comparison of the livestock production functions indicates that they
are all rather similar, especially those for Montana, southern and
northern Iowa. In every case, capital services are by far the most In-
fluential input, with labor playing only a minor - but statistically sig-
nificant —role. This similarity is explained by the fact that the pro-
duction of meat from livestock is influenced far less by the climatic
environment of the production process than is crop production. How-
ever, due to the effects of price differences and variations in the scale
of production, resource marginal products in livestock production do
show some marked differences between regions. The southern and
northern Iowa samples had marginal products of livestock labor
cent.
cent.
cent,
per
per per
55 10
^< ^
of of of
level level
level
probability probability
probability
levels.
a a input a
at at at
mean
zero unity unity
from from from
geometric
different different different
the
at
Significantly
Significantly Significantly
Êstimated
t S
[
595 ]
($125.06 and $120.52 per month, respectively) that were nearly double
those found in Montana ($76.50) and Alabama ($68.75). The low mar-
ginal product for livestock labor in Alabama is to be expected because
the capital to labor ratio is low, and because of the techniques used and
of the forms of capital employed in livestock production. In contrast to
the marginal products of livestock labor, differences between the re-
gions were not so great for the marginal products of livestock capital
services. They ranged only from about $1.20 in southern Iowa and
Montana to $1.00 in northern Iowa and Alabama.
Comparing the crop and livestock functions, the marginal product
of labor was higher for livestock than for crops in all areas. In practi-
cally all cases, the difference was significant at the 5 per cent proba-
bility level. Marginal returns on livestock capital were higher than for
crop capital in northern Iowa while they were about equal in southern
Iowa and Alabama. Marginal capital productivity was higher for crops
than for livestock in Montana. Although production elasticities for both
crops and livestock in Montana are affected by the highly variable
weather of the area, crop returns were likely favored more by good
weather in 1950. Most Montana farmers have beef cow herds of rather
fixed size and can make only partial use of above average pasture yields
from favorable weather.
From the marginal return to opportunity cost ratios derived in the
four -state study, as listed in Table 17.1, the existence of severe re-
source use disequilibria is apparent. Especially is this true for land
and labor. In every region, the average farmer could have increased
profits substantially by expanding the area cropped and by reducing the
labor input on both crops and livestock. As well, it would have been
more efficient to use additional capital services for both livestock and
crops in Montana and southern Iowa, in contrast to northern Iowa where
slightly excessive capital services were being used in both enterprises.
These conclusions are supported by the fact that the resource use effi-
ciency ratios of land and labor for both crops and livestock are signifi-
cantly different from unity at the 10 per cent probability level in each
sample, the same being true for capital services in southern Iowa and
Montana. For Alabama, the use of capital services under current pro-
duction conditions appears to be satisfactory for crops and nearly so
for livestock. The general policy implications to be drawn from the
crop -livestock study are these: (a) there is too much labor in much of
United States agriculture; (b) banks and other credit agencies should
recognize that, on the average, farmers would be acting in very logical
fashion if they endeavored to increase the size of their farms; (c) the
flow of capital should be directed towards Montana and southern Iowa
rather than towards northern Iowa and Alabama, with the proviso that
capital directed to Alabama for the development of new techniques may
be very worthwhile. While Alabama is low in capital productivity it is
known, a priori, that returns on this resource could actually be quite
large if it were invested in educational resources for farmers; more
capital would also allow livestock production on a commerical basis
rather than as a semihousehold enterprise.
Supplementing these over -all implications of the study, it might be

noted that detailed analysis of the data revealed examples of resource
use inefficiencies within each sample just as large as those found be-
9
tween the various samples. Still, these intr a -regional disequilibria re-
lated to only a small proportion of the farms in each sample. They do
not vitiate the over -all interregional policy implications drawn from the
four -state data.
Farm Tenure Study
Consider now the Iowa -Illinois tenure study. The basic hypothesis
was that full owners, live stock -share renters, and crop-share cash-
renters belong to different farm -firm populations having distinctive
patterns of resource use and levels of efficiency. Over -all, the analy-
sis indicates that while full owners differ markedly from livestock-
share and crop -share renters in terms of the production elasticities of
resources and the efficiency of resource use, differences between the
two renter groups may not be too important. Moreover, in absolute
terms, all three tenure groups displayed roughly the same level of inef-
ficiency as measured by deviations between their actual resource use
pattern and the allocation that would minimize the cost of producing
their actual average level of production.
Examination of the marginal products and marginal return to oppor-
tunity cost ratios for the three tenure classes, as listed in Table 17.1,
reveals important differences between the groups. Full owners appear
to be using too few capital services and slightly excessive quantities of
labor and land. However, if the full owners used additional capital
services, the marginal products of their land and labor resources would
be increased. Capital rationing, arising from prior commitments to
land purchases, is probably the main factor limiting the use of capital
services by the full owners. For live stock -share renters, all the mar-
ginal returns are greater than resource opportunity costs. The use of
all three resources might be extended profitably. Disequilibrium is
especially evident for land, the marginal product being nearly double
resource cost and substantially above the ratios for labor (1.36) and
capital services (1.16). These ratios follow the same order for the
crop-share cash-renter group, but are slightly lower in each case. The
higher marginal product of labor for both renter classes, compared to
full owners, is most likely due to the greater quantity of fixed assets
used with their labor.
With a marginal return to opportunity cost ratio of 1.01 for capital
services, it would seem that the cash renters were using very near the
optimal quantity of capital services. This result does not coincide with
theoretical expectations: the alleged nonoptimal sharing of costs and
9
Heady, E. O. and Baker, C. B. Resource adjustments to equate productivities in agri-
culture. Southern Economic Journal, 21: 36-52. 1954.
returns under the crop -share system should cause a higher marginal
product (relative to full owners) for capital services because of these
10
resources being limitationaL Perhaps the usually noted imperfections
of crop -share cash-rent leases are balanced by such factors as the
sharing of uncertainties and the overcoming of capital shortages by the
joint contributions of landlords and tenants. It is also possible that the
consistently lower marginal return to opportunity cost ratios of the
crop -share tenants is related to superior management. The broad con-
clusion to be drawn from the tenure study is that the types of resource
adjustments needed to optimize production vary according to tenure
status. While full owners need additional capital rather than additional
land or labor, both tenant groups would contribute more to the general
welfare, and their own, if they had more land and labor.
Looking at the over -all picture presented by the two United States
studies, it is obvious that extensive disequilibria prevail in the use of
resources in United States agriculture. The data of Table 17.1 imply
that under the price conditions prevailing, United States farmers should
in general (a) increase the size of their farms, and (b) substitute capital
services for labor. At the same time, quite large differences exist both
between geographic areas and tenure groups, and even between crop and
livestock enterprises in the same area, in the extent to which these
changes should be made. Moreover, in terms of capital investment, the
average farm to which the United States studies refer is quite large
relative to most farms in other countries. In “all input* terms, only the
Alabama farms (which are of insignificant importance relative to the
whole of United States agriculture) used less than $10,000 worth of re-
sources per year. It might be expected, therefore, that the quantities of
United States resources that could be transferred fruitfully within agri-
culture, and between agriculture and industry, would be substantial
relative to the shifts that might be required to ensure equilibrium under
the smaller and less capital-intensive forms of farm organization found
in other countries. Of course, this would not be true if the patterns of
farm organization in these other countries were to be changed via the
introduction of new larger scale production methods of high capital in-
tensity. The resource transfers required in many of these countries
would then be far in excess of those needed to bring resource use in
United States agriculture to a reasonable level of efficiency.
CANADA
Table 17.2 presents production functions and their related statistics
for samples of mixed livestock-crop farms, cattle ranches, and cattle
feedlots, respectively, in the Canadian province of Alberta. 11 In terms
10
Heady, E. O. Economics of agricultural production and resource use. Prentice -Hall,
Inc.^ New
York. 1952* Pp. 587-602.
u Darcovich, W. The use of production functions in the study of resource productivity in
some beef producing areas of Alberta. Economic Annalist, 28: 85-93. 1958,
functions over the world 599
Table 17.2. Canada: Production Elasticities, Marginal Products and Opportunity

Costs of Resources, Marginal Return to Opportunity Cost Ratios,
and Related Statistics for Selected Production Function Studies
—
Production Function for Alberta
Xtoxxi
Mixed farms Cattle ranches Cattle feedlots
farms 72 36 72
Number of
1953-54 1954-55 1953

Year
Production elasticities:
0.39* 0.20*
Land
0.20 0.37*
Labor
0.34* 0.39* 0.13*
Capital services
0.84*
Livestock inventory
t 0.93 0.97 0.97

Sum of elasticities
2 0.69 0.81 0.98

R
Sample means :*
12,556.00 22.243.00 17.930.00
Output ($)
Land (acres) 522.00 5.706.00
Labor (months) 13.8 35.6
4.042.00 5.982.00 3,943.00
Capital services ($)
11.998.00
Livestock inventory ($)
7.999.00 16.496.00 15.941.00
All inputs ($)
Average products:^
24.00 3.90
Land ($ per acre)
Labor ($ per month) 909.80 624.80
3.10 3.72 4.55
Capital services ($ per $)
per 1.49
Livestock inventory ($ $)
Marginal products:^"
9.47 0.81
Land ($ per acre)
179.08 233.30
Labor ($ per month)
1.07 1.46 0.63
Capital services ($ per $)
1.27
Livestock inventory ($ per $)
Opportunity costs:
3.67 0.92
Land ($ per acre)
per month) 147.90 147.90
Labor ($
1.06 1.06 1.06
Capital ($ per $)
Marginal return to opportunity

cost ratios
2.58 § 0.88
Land
1.21 1.58 §
Labor
1.01 1.38 0.59
Capital services
1.20
Livestock inventory
of < 5 per cent.

*Significantly different from zero at a probability level
per cent.
tNot significantly different from unity at probability level of 5
Êstimated at the arithmetic mean input levels.
§ Significantly different from unity at a probability level
of ^ 10 per cent.
600 FUNCTIONS OYER THE WORLD
of scale of operation and general technological features, these Canadian

farms are broadly comparable to the United States farms considered in
Table 17.1.
While the cattle ranch and mixed farm functions refer to the over-
allfarm enterprise, the cattle feedlot analysis relates only to the feed-
lot operation on the farms studied. As noted in Table 17.2, the mar-
ginal return to opportunity cost ratios were tested statistically. These
tests showed only the ratios for land in mixed farms (2.58) and labor on
cattle ranches (1.58) to be different from unity at the 10 per cent prob-
ability level. Nonetheless, the listed values for the other inputs still
indicate their most likely level of efficiency as currently used. The
severe lack of equilibrium in land use on the mixed farms is probably
due to relative overcapitalization in machinery. Moreover, none of the
marginal return to opportunity cost ratios for the mixed farm sample
are less than one. In consequence, no scope exists for attaining equi-
librium in resource use by substitution between resources. The gen-
eral implication is thus that these mixed farms could expand their op-
erations profitably by adding land, labor, and a little capital, in that
order. The additional capital would, of course, be necessitated by the
effect of additional land and labor in raising the marginal product of
capital, hi so far as the mixed farm sample was a superior group of
farms with high livestock intensity, it is likely that the same disequi-
libria exist in more severe form on other Alberta mixed farms. Indeed,
on these more average farms it is likely that the shortage of capital
severely limits the use of all inputs.
For the cattle ranch sample, the marginal return to opportunity cost
ratios indicate too much land and too few labor and capital services
were being used. Equilibrium would be attained most quickly by sub-
stituting labor and capital for land, or if the land input is to remain
fixed, by simply using additional quantities of labor and capital services.
So far as disequilibria between the mixed farms and cattle ranches is
concerned, the only opposing tendencies appear in the land resources:
mixed farms were using too little and ranches too much. However, due
to the fixity of the land resources, land transfers between the two types
of farming are unfeasible. While both types of farm could use more
labor and capital, priority in these resources should go to the cattle
ranches.
Since the third Canadian function — for cattle feedlots — is on an en-
terprise basis, the classification of inputs as capital services and live-
stock inventory differs from the traditional tripartite one used for the
mixed farms and cattle ranches. The capital services input is total
feed input, while the livestock inventory is the input of feeder cattle at
the start of stall feeding. Unlike the United States livestock studies
presented in Table 17.1, labor input was not included in this feedlot
study Nonetheless, the fitted function implies, via its marginal return
.
to opportunity cost ratios, that the average feedlot operation was rather
inefficient under the price conditions examined; too much feed and too
few feeder cattle being used. Since much of the feed is home produced,
the feed input requires little cash outlay and is probably used extrava-
gantly. In contract, the purchase of feeder cattle requires a large cost
outlay. Farmers tend to buy them sparingly. However, given the un-
certainty that surrounds fat cattle prices under current marketing pro-
cedures, farmers may be behaving quite rationally in overfeeding a
small number of cattle. The policy implication to be drawn is that the
development of institutions to extend longer-term or more flexible
credit would probably be worthwhile. Still, on average, the production
function analysis suggests that, even under current credit arrange-
ments, increased profits would be possible if the same quantity of feed
were spread over a greater number of animals.
AUSTRALIA AND NEW ZEALAND
For convenience, the only available function for New Zealand has
been grouped in Table 17.3 with those estimates available for Australia.
Geographically and in terms of farm organization and production, the
two countries are rather close, although their climates are somewhat
different. The listed production function estimates relate to wool and
fat lamb production from sheep, and to dairying. In both New Zealand
and Australia, these lines of production are of major agricultural im-
portance. Both are conducted as grazing enterprises with a small
measure of supplementary feeding.
The New Zealand study refers to a group of properties producing fat
lambs and wool on improved pasture in the light soil area of the south
12
island’s Canterbury Plain. For Australia, only a single study is avail-
13
able of wool and fat lamb production. It is based on a random sample
of grazing properties in the southwest slope of New South Wales, an

area which has been largely converted from natural to improved pasture
in recent years. The Australian dairy production function studies listed
in Table 17.3 are from various sources. That for New South Wales was
conducted conjointly with the sheep study already mentioned. It refers
to a random sample of coastal farms producing milk for butter produc-
tion. The South Australian analysis also refers to a random sample of
14
dairy farms producing for butter production. These farms were lo-
cated on irrigated land along the Lower Murray Valley. Unlike the
other dairying studies, that for Western Australia is based on a random
15
sample of farms producing whole -milk for human consumption. The
sample was located in the Perth -Bunbury region.
12
Mason, G. Resource productivities from a sample of light plains farms, Canterbury,
N.Z. Unpublished M. Agr. Sc. thesis. Canterbury Agricultural College Library, Canterbury,
New Zealand. 1958.
ls
Marginal productivities of resources in two farming areas of N.S.W.
Dillon, J. L.
Economic Society and N.Z. Economic Monograph 188. Sydney. May, 1956.
of Australia
Jarrett, F. G. Estimation of resource productivities as illustrated by a survey of the
14
Lower Murray Valley dairying area. Australian Journal of Statistics, 1: 3-11. 1959.
Schapper, H. P. and Mauldon, R. G. A production function from farms in the whole-
15
milk region of Western Australia. Economic Record, 33: 52-59. 1957.

Table 17.3. Australia and New Zealand: Production Elasticities, Marginal Products and Opportunity
Costs of Resources, Marginal Return to Opportunity Cost Ratios, and Related
Statistics as Derived for Selected Production Function Studies
Production Function for
Sheep Dairying
Item
Canterbury Plain, New South New South Western South
New Zealand Wales Wales Australia Australia
Number of farms 50 52 99 51 48
Year 1955-56 1953-54 1952-53 1954-55 1955-56
Land 0.42* 0.10 0.28* 0.39*
Labor 0.15 0.59* 0.22* 0.23* 0.25*
Capital services 0.55* 0.42* 0.55* 0.32*
Capital assets 0.26*
Supplementary feed 0.13* 0.18*
Agistment 0.04*
Superphosphate 0.22* 0.07*
Lime 0.06
Sum of elasticities^ 1.11 1.24 0.92 0.99 1.19
R2 0.81 0.76 0.64 0.87 0.62

Sample means
Output ($) 14,428.00 13,234.00 5,152.00 9,330.00 8,029.00
Land (acres) 698.00 1,516.00 191.00 49,50
Labor (months) 18.10 23.10 21.70 20.20 22.70
Capital services ($) 3,264.00 1,572.00 1,939.00 1,291.00
Capital assets ($) 3,809.00
Supplementary feed ($) 1,739.00 1,111.00
Agistment ($) 47.80
Superphosphate ($) 838.00 406.00
Lime ($) 704.00
All mputs ($) 7,709.00 9,765.00 4,902.00 6,876.00 § 8,343.00
Average products:"^
Land ($ per acre) 20.70 8.70 27.00 162.20
Labor ($ per month) 797.10 572.90 237.40 862.90 353.70
Capital services ($ per $) 4.05 3.28 4.81 6.22
Capital assets ($ per $) 3.79
Supplementary feed ($ per $) 5.36 7.23
Agistment ($ per $) 168.00
Superphosphate ($ per $) 17.20 23.00
Lime ($ per $) 20.50
Marginal products:"^
Land ($ per acre) 8.59 0.87 6.90 79.83
Labor ($ per month) 122.52 292.70 51.07 107.97 108.66
Supplementary feed ($ per $) 0.71 1.80
Agistment ($ per $)
5.20
Superphosphate ($ per $) 4.00 1.56
Lime ($ per $) 1.16
Opportunity costs:
Land ($ per acre) 3.46 2.24 5.60 37.70
Labor ($ per month) 144.14 134.40 104.16 119.39 177.41
Capital ($ per $) 1.05 1.05 1.05 1.06 1.06
cost ratios :$
Land 2.48 0.39 1.23 2.12
Labor 0.85 2.18 0.49 0.90 0.61
Capital services 1.97 1.40 2.50 1.26
Capital assets 2.33
Supplementary feed
0.67 1.70
Agistment
4.91
Table 17.3. (Cont’d)

Sheep Dairying
Item
Canterbury Plain, New South New South Western South
New Zealand Wales Wales Australia Australia

cost ratios :$
Superphosphate 3.81 1.47 j
Lime 1.10
’’‘Significantly different from zero at a probability level of « 5 per cent.

%ot significantly different from zero at a probability level of 5 per cent with the exception of the South
Australian study for which no significance test is available.
Êstimated at the geometric mean input levels with the exception of the sample means and average
product of the South Australian sample which refer to the arithmetic means.
§ Including an allowance of $380 for land services.
As perusal of Table 17.3 indicates, considerable variation exists

between these various analyses in the way inputs have been categorized.
Labor is the only resource category common to all five studies. While
the New South Wales sheep and dairy studies use the broad “land, labor,
and capital services” approach, the other analyses generally consider
fertilizer and purchased feed inputs separately. Land was excluded
from the Western Australian study because it was found to be highly
correlated with capital services. The differences between the sample
means of output and all inputs across Table 17.3 are noteworthy. They
are explained by the fact that in both Australia and New Zealand sheep
production has been more profitable than dairying in recent years and
that, within dairying, the rationing of whole -milk production has made it
far more rewarding than dairying for butter production. Indeed, in
Australia, dairying for butter production is a rather depressed industry,
as is borne out by the marginal return to opportunity cost data of Table
17.3.
The marginal return to opportunity cost ratios for sheep show con-
siderable differences between the New Zealand and Australian samples.
While both were using insufficient capital, the New Zealand farms ap-
parently had too little land and too much labor. The reverse situation is
indicated for the New South Wales sheep properties. In consequence,
while the New Zealand sheep farms should use more capital and less
labor, those in New South Wales should use more labor as well as capi-
tal. Such adjustments would in both cases tend to remove some of the
disequilibria in land use, especially in the Australian sample, granted
that the farmers' preferences or opportunities forestalled the purchase
(in New Zealand) or sale (in Australia) of land.
The data of Table 17.3 also gives evidence of quite large differences
in resource use efficiency within Australian dairying and between dairy-
ing and sheep production. Each of the dairy samples tends to indicate
the same pattern of inefficiency: too few capital and land services and
an excessive use of labor. Particularly outstanding are the deficiencies
of land (and agistment which substitutes for land) in the South Australian
sample, and of capital services in the Western Australian whole -milk
-
study. The greatest overuse of labor appears in the New South Wales
dairy sample. This result is not surprising. The area to which the
sample refers is patently depressed, having a surfeit of farm population
with few nonfarm employment opportunities. Comparing the marginal
return to opportunity cost ratios between the Australian sheep and
dairying samples, it is apparent that under the price conditions as-
sumed in these studies, policies oriented (a) to the transfer of labor out
of dairying into sheep production and (b) to the increased use of capital
services in both types of farming would be advantageous. Too, it should
be noted that the prices paid for dairy production in Australia have been
kept artificially high by government subsidy over the period examined.
These subsidies mask the degree of inefficiency that would prevail
under free market conditions. With unsubsidized prices for dairy prod-
ucts, the marginal return to opportunity cost ratios would be reduced,
thereby further emphasizing the excessive quantity of labor used on
Australian dairy farms and, at least initially, tending to restore equi-
librium in the use of land and capital services.
UNION OF SOUTH AFRICA
The production function analyses presented in Table 17.4 for the

Union of South Africa relate to a one in four sample of cattle ranches in
the Eastern Kalahari region. 16 This area includes parts of the Kuru-
man, Vryburg, and Mafeking districts. The limited and variable rain-
fall in combination with the sandy soils of the region presents
great
hazards for cattle ranching, in terms of both grass yields and water
supplies. Moreover, grazing is the only major economic activity that
is
feasible in the region. Data for the analysis were gathered in two
sur-
veys covering the 1950-51 and 1951-52 financial years. Only ranchers
who had been established for at least two years were included in the
samples. Both survey years being rather similar to the recorded
aver-
age of production years in the region, it is thought that the derived
func-
tions may be taken as representative. Still, at the prevailing prices,
resource productivities could be considerably higher in periods
of above
average rainfall and would be much lower in droughty years.
In deriving the production functions, the data were
stratified by the
two years in which samples were taken, and also by the
two systems of
ranch organization in the region. To wit: cow-calf herds
and growing
fattening herds. As well, a function was fitted to the
pooled data for
both years and is also presented in Table 17.4. For
all five fitted func-
tions, the production elasticities are rather
similar; none of the differ-
ences for the same resource between the various functions
is statisti-
caHy significant. Hence it is to be expected that ranches
with the same
Heady, E. O. and Du Toit, S. Marginal resource

productivity for agriculture in se-
ted aS f S t Ai riCa and the United States Jou **nal of
5^ i^! Ai° ?! 1X1
l 01 ,
* Political Economy, 62; 494-
S °f Cattle ranchin S in the Eastern Kalahari,
Wh J ^ M. Sc. thesis. Iowa State University
South Africa. Unpublished
Union of
Library. Ames. 1953.
Table 17.4. Union of South Africa: Production Elasticities, Marginal Products and
Opportunity Costs of Resources, Marginal Return to Opportunity
Cost Ratios, and Related Statistics as Derived for
Selected Production Function Studies
Production Function for Eastern Kalahari

Item
Cow - Calf Fattening Pooled
Ranches
Ranches Ranches Data
Number of farms 80 82 70 92 162
Year 1950-51 1951-52 1950-52 1950-52 1950-52
Land 0.21* 0.24* 0.19* 0.28* 0.23*
Labor 0.18* 0.13 0.19* 0.13* 0.14*
Capital services 0.54* 0.59* 0.52* 0.55* 0.57*
Sum of elasticities^ 0.93 0.96 0.90 0.96 0.94

2
R 0.86 0.76 0.64 0.85 0.80
1
Sample means:
Output ($) 9,344.00 8,211.00 6,643.00 10,704.00 8,711.00
Land (acres) 6,576.00 6,790.00 5,891.00 7,358.00 6,684.00
Labor (months) 58.00 63.00 54.00 65.00 60.00
Capital services ($) 2,910.00 2,776.00 1,942.00 3,761.00 2,827.00
All inputs ($) 4,950.00 4,918.00 3,792.00 6,044.00 4,912.00
Average products:^
Land ($ per acre) 1.42 1.21 1.13 1.45 1.30
Labor ($ per month) 161.10 130.30 123.00 164.70 145.20
Capital services ($ per $) 3.21 2.96 3.42 2.85 3.08
Marginal products:^"
Land ($ per acre) 0.30 0.29 0.22 0.41 0.30
Labor ($ per month) 29.01 16.38 23.78 20.99 21.11
Capital services ($ per $) 1.72 1.76 1.77 1.57 1.76
Opportunity costs:
i
Land ($ per acre) 0.21 0.21 0.21 0.21 0.21

Labor per month)
($ 11.36 11.36 11.36 11.36 11.36
Capital ($ per $) 1.06 1.06 1.06 1.06 1.06
Marginal return to
opportunity cost ratios:
Land 1.43 1.38 1.05 1.95 1.43
Labor 2.55 1.44 2.09 1.85 1.86
Capital services 1.62 1.66 1.67 1.48 1.66
from zero at a probability level of ^ 5 per cent.

* Significantly different
tNot significantly different from unity at a probability level of 5 per cent.

lEstimatedat the geometric mean input levels.
quantity of resources would have fairly similar marginal resource pro-

ductivities under each of the fitted functions. However, since the two
types of ranch organization - cow-calf and fattening — utilize different
amounts of resources, the marginal products vary from one to the
other. Discussion is therefore mainly confined to the production func-
tions for these two types of ranch organization and to the "pooled data*
function, the latter analysis being thought to give the best guide to the
over -all situation in the region.
As evidenced by the marginal return to opportunity cost ratios, the
average sample farms exhibit extreme inefficiencies in resource use.
In the pooled data analysis, labor productivity exceeds the wage rate by
86 per cent; capital productivity exceeds its market price by 66 per
cent; and land productivity is in excess of cost by 43 per cent. The
latter effect is mainly due to the situation on fattening ranches where
land productivity is in excess of cost by 95 per cent, the excess being
only 5 per cent on cow-calf ranches. The obvious implications are that
more labor and capital services should be used on both the fattening and
cow -calf ranches. The fattening ranches should also use more land —
despite the fact that the average ranch already employs far more land
and labor, in quantitative terms although not in value terms, than is
used on the average farm in any of the other national studies that we
present. Moreover, since the resource use efficiency ratio exceeds
unity for every resource in each fitted function, no opportunities exist
for approaching optimal efficiency through the substitution of one re-
source for another. For both the breeding and the fattening farms, the
needed reorganization is one of scale. However, these changes in scale
may not need to be as large as the data of Table 17.4 imply. The rea-
son is that management was not included as a separate input in the
fitted functions due to the lack of a quantitative measure. In fact, man-
agement is quite a distinct input on these ranches, being supplied by the
rancher himself who makes no manual contribution at all. All the actual
labor involved is contributed by a hired work force. Still, the manage-
ment input is probably partly reflected in the elasticities of the other
inputs so that their true elasticities may be somewhat less than those
estimated. In consequence, the marginal return to opportunity cost ra-
tios of Table 17.4 probably overestimate the extent of resource use dis-
equilibria. Nonetheless, the exclusion of management is probably not a
severe limitation on the analysis, despite the fact that because of the
nature of the organization of each ranch its exclusion is of more impor-
tance than would otherwise be the case.
The more important factors contributing to the inefficient use of re-
sources are probably the ranchers' lack of knowledge of marginal pro-
ductivities and the influence of yield, price, and technological and soci-
ological uncertainty. The latter influences are especially important in
the area. Historically, major uncertainty has surrounded prices and
yields. Because of the low and variable rainfall with its resultant ef-
fects on pasture production and livestock water supplies, yield is highly
variable and unpredictable from year to year. Technological uncer-
tainty is also high, the techniques of beef production still being in a
stage of transition. Indeed, new techniques may have been more suc-
cessful than most ranchers anticipated, thus contributing to sub-optimal
levels of investment. Moreover, uncertainty of a sociological, psycho-
logical, and political nature also surrounds the actions of the labor
force which generally consists of a fairly large number of African
workers under occidental management. In the face of all these uncer-

tainties, ranchers are reluctant to add more resources by borrowing,
despite the comparative cheapness of land and labor. Moreover, good
and poor climatic years do not average out because livestock watering
facilities limit carrying capacity and the rancher usually keeps cattle
numbers down in anticipation of the unfavorable seasons.
UNITED KINGDOM
Three production function estimates, each relating to a sample of

United Kingdom dairy farms, are presented in Table 17.5. All are
based on material drawn from the continuing Farm Management Survey
conducted in England. The first estimate refers to England as a whole. 17
It is based on a sample of 244 farms drawn not at random but arranged
purposively to give observations exhibiting a minimum of correlation

between input categories. These inputs were labor, purchased feed, and
capital services with land services measured as rental value included in
the capital services category. As a side effect of the purposive sam-
pling procedure, the function relates only to dairy farms with inventory
values of crops, livestock, machinery, and equipment between $2,800
and $11,200. Still, this range covers the largest part of English dairy-
ing. Each of the sample farms was located in an officially classified
dairy district and all obtained at least half of their gross receipts from
milk.
The other two functions listed in Table 17.5 relate only to Bristol
Province low-lands farms of between 50 and 200 acres which received
18
at least 75 per cent of their gross income from dairying. The analysis
was essentially aimed at assessing the usefulness of production function
estimates for individual farm recommendations. And so, in recognition
of the short -rim fixity of the land resource on British farms, land was
eliminated as a variable by considering the other inputs — cows, other
large livestock, and capital services — on a per 100 acre basis. On the
assumption that resource productivities differed to an important extent
between low and high yield farms, separate functions were fitted to
these two groups in the over -all farm sample. The average yield clas-
sifications used were from 552 to 702 gallons per cow for the low pro-
duction group, and from 744 to 894 gallons per cow for the high produc-
tion group.
Both within and between the three United Kingdom samples, quite
important differences in the efficiency of resource use are evidenced by
the marginal return to opportunity cost ratios of Table 17.5. For the
all-England sample, the average picture is a fairly satisfactory one,
with some additional capital services and a modicum of labor and
17
AntiU, A. G. Towards a production function for dairy farms. The Farm Economist,
8 (1): 1-1X. 1955.
18
Wragg, S. R. and Godsell, T. E. Production functions for dairy farming and their ap-
plication. The Farm Economist, 8 (5): 1-6. 1956.
purchased feed being required for maximum efficiency. In contrast, the

Bristol Province farms appear to be organized very inefficiently. For
both the low and high yield groups, the marginal return to opportunity
cost ratios for cows of 3.71 and 3.41, respectively, are very high. Far
too few cows were being milked. At the same time, the Bristol sam-
ples were running too many large nondairy livestock (especially the low
yield farms) and using an excessive quantity of capital services (espe-
cially the high yield group). The latter result is in line with the im-
pressions of Bristol Province farm management workers that farms
with high quality dairy herds exhibit a degree of conspicuous consump-
tion in their proliferation of capital services, over and above the feed
and labor input included in these services. It is not surprising, there-
fore, that the production elasticity of capital services varies inversely
with milk yield. While a 1 per cent increase in capital services would
raise production by only 0.05 per cent on high yield farms, it would in-
duce an increase of 0.53 per cent on low yield farms. On the other
hand, the production elasticities of cows and other livestock are much
the same on both high and low yield farms.
Comparison of the Bristol Province and all -England results points
up the fact that, subset by subset, quite large differences may exist
within a parent set of farms producing the same product. Bristol Prov-
ince being part of England, it is to be expected that the disequilibria ex-
hibited in the Bristol sample may be balanced by converse inefficien-
cies in other sectors of English dairying, leading to the picture of
apparently satisfactory resource use found in the all -England sample.
This fact emphasizes the desirability of sample homogeneity, even if
the resultant estimates are to be used for the broadest of policy deci-
sions.
AUSTRIA
Only two cross-sectional production function estimates are availa-

19
ble for Austrian agriculture. Both are listed in Table 17.6. Each of
these functions is of a rather macro nature, although the data on which
they are based is built up from individual records collected for 784
farms by the Austrian Department of Agriculture. These farms were
scattered throughout Austria so that the functions refer to Austrian ag-
riculture as a whole. The first function is based on a stratification of
the 784 farms according to their location within the eight geographic
types of regions delineated for statistical purposes by the Austrian au-
thorities. The outputs and inputs of record-keeping farms within each
of these regions were aggregated and a production function fitted to the
resultant eight observations. For the other function, the 784 farms
were grouped into 37 strata on the basis of their type of production, the
19
Tintner, G. Produktionsfunktionen fur die osterreichische Landwirtschaft. Zeitschrift
fur Nationalokonomie. 17: 426-42. 1958.
Table 17.5. United Kingdom: Production Elasticities, Marginal Products

and Opportunity Costs of Resources, Marginal Return to
Opportunity Cost Ratios, and Related Statistics for
Production Function for Dairying

Bristol Province
Item
England
low yield high yield
farms farms
Number of farms 244 24 20
Year 1951-52 1953 1953
Labor 0.29*
Capital services 0.50* 0.53* 0.05
Cows 0.55* 0.54*
Other livestock 0.00 0.05
Purchased feed 0.33*
Sum of elasticities 1.12* 1.08 0.64
R2 0.84 0.64 0.67

t
Sample means:
Output ($) 7,582.00 6,815.00 9,646.00
Labor (months) 32.40
Capital services ($) 2,774.00 5,172.00 6,345.00
Cows (cow equivalents) 24.19 27.50
Other livestock (cow equivalents) 11.58 13.48
Purchased feed ($) 2,164.00
All inputs ($) 7,012.00 6,674.00 8,640.00
Average products:^
Labor ($ per month) 234.00
Capital services ($ per $) 2.73 1.32 1.52
Cows ($ per cow equivalent) 281.70 350.80
Other livestock ($ per cow equivalent) 588.50 715.60
Purchased feed ($ per $) 3.50
Marginal products:^
Capital services ($ per $) 1.37 0.71 0.08
Cows ($ per cow equivalent) 156.04 190.93
Other livestock ($ per cow equivalent) 0.48 36.74
Purchased feed ($ per $) 1.14
Opportunity costs:
Capital ($ per $) 1.05 1.05 1.05
Livestock ($ per cow equivalent) 42.00 56.00
Marginal return to opportunity cost ratios

Labor 1.06
Capital services 1.30 0.68 0.08
Cows 3.71 3.41
Other livestock 0.01 0.66
Purchased feed !
1.09
J
from zero at a probability level of « 5 per cent,

’’'Significantly different
tsignificantly different from unity at a probability level of « 5 per cent.
$ Estimated at the geometric mean input levels. For the Bristol province studies, the
output and input statistics are on a per 100 acres basis.
Table 17,6. Austria: Production Elasticities, Marginal Products and Opportunity

and Related Statistics as Derived for Selected
Production Function Studies
Item Production Functions for Austrian Agriculture
Number of observations 8 37
Year 1954-55 1954-55
Production elasticities
Land 0.14 0.13*
Labor 0.29* 0.26*
Capital services 0.59* 0.61*
Sum of elasticities ^ 1.02 1.00

2
R 0.99 0.99
Marginal products:!
Land ($ per acre) 11.87 10.78
Labor ($ per month) 22.53 19.57
Capital services ($ per $) 1.48 1.56
Opportunity costs:
Land ($ per acre) 11.67 11.67
Capital ($ per $) 1.04 1.04

cost ratios:!
Land 1.02 0.92
Labor 0.63 0.54
Capital services 1.42 1.50
Significantly different from zero at a probability level of « 5 per cent.

iNot significantly different from unity at a probability level of 5 per cent.
$Estimated at the geometric mean input levels.
geographic type of region to which they belonged, and the state in which
they were located. The outputs and inputs of farms within each of these
strata were again considered in aggregate, a production function being
fitted to the 37 observations. Because of the aggregative approach
used, the fitted functions are of interest solely in terms of their broad
policy implications. They are of no significance in relation to resource
allocation on specific farms. Indeed, the basic aim of the study was
expository with the hope of stimulating econometric research in Aus-
tria. Still, the fitted functions are of use for present purposes.
Since both of the functions are based on the same fundamental data,
it is gratifying that they exhibit very similar production elasticities. In
consequence, the marginal products of resources and the marginal re-

turn to opportunity cost ratios are also much the same for both esti-
mates. Too, at the usual probability levels there are no statistically
significant differences between corresponding statistics of the two
analyses. Before considering the implications of the two studies, it
might be noted that the sample means of inputs and output are not listed
in Table 17.6. Those statistics, given the macro nature of the Austrian
functions, would be of no value for purposes of international compari-
sons. Also, the opportunity cost figures quoted for land and labor are
strictly average figures. Without doubt, they mask quite marked differ-
ences in the price of these resources between the various regions of
Austria.
Both functions being essentially the same, their implications in
terms of resource use efficiency are similar. From the marginal re-
turn to opportunity cost ratios, it appears that Austrian farm land is
used in optimal fashion but that too much labor and too few capital serv-
ices are used. In other words, Austrian farm policy should endeavor to
persuade farmers to substitute capital services for labor if farm profits
and, within limits, national welfare are to be maximized. Concomitantly,
it may be necessary to provide off-farm employment opportunities for
the excess labor currently being used in agriculture. It must be empha-

sized that these conclusions apply in only the broadest "on average*
sense. For greater reliability and precision, an array of production
function estimates covering the types of farming in the various regions
of Austria would be necessary. The form in which the basic data were
available prevented such an approach in the present instance.
SWEDEN
Two cross-sectional production function estimates are presented

for Sweden. The first, listed with its related statistics in Table 17.7,
20
refers to milk production in southern Sweden. The investigation was
based on accounting records of the milk enterprise on a sample of 34
farms with a total of 1,722 cows. All the farms used Swedish low-land
dairy cows of fairly uniform quality. As shown in Table 17.7, the esti-
mated production elasticities for labor and capital services are nega-
tive, although not significantly different from zero. Only the feed input
categories — concentrates, pasture, and roughage — had positive elas-
ticities. In other words, under the average conditions prevailing in the
sample, the major determinant of milk output was the input of feed; no
significant role being played by the ancillary labor and capital services
It is not to be assumed that the elasticities of labor and capital would
have still been negative if they had been used in smaller quantities.
Some small amount of both is obviously essential. The difficulty, of
course, is that the Cobb-Douglas type function allows only constant
elasticities of production. The negative elasticities for labor and capi-
tal services presumably reflect the average situation prevailing in the
sample. It says nothing of the situation that may prevail at input com-
binations above or below those found in the sample.
20
Hjelm, L. Utbytesrelationer i mjolkproduktionen (with an English summary; Input-
output relationships in milk production). Institutionem for Lantbrukets Driftsekonomi.
Stockholm. 1953.
Table 17.7. Sweden: Production Elasticities, Marginal Products and Opportunity

and Related Production Function Statistics for Milk Production
Item Production Function Statistics
Number of farms 34
Year 1951-52
Labor -0.05
Concentrates 0.28*
Pasture 0.46*
Roughage 0.53*
Capital services -0.04
Sum of elasticities 1.18t

2
i 0.88
Sample means :$
Concentrates ($ per cow) 60.30
Pasture ($ per cow) 54.00
Roughage ($ per cow) 79.90
Marginal products
Labor ($ per month per cow) negative §
Concentrates ($ per $ per cow) 1.30
Pasture ($ per $ per cow) 2.59
Roughage ($ per $ per cow) 2.02
Capital services ($ per $ per cow) negative §
Opportunity costs:
Capital ($ per $) 1.05
Marginal return to opportunity cost ratios:^

Labor negative §
Concentrates 1.24
Pasture 2.47
Roughage 1.92
Capital services negative §
^Significantly different from zero at a probability level of < 5 per cent.

J Estimated at the arithmetic mean input levels.
§ Exact figures are not available.
It is apparent from the marginal return to opportunity cost ratios
that additional pasture, roughage, and perhaps concentrates could be fed

profitably, in that order. To do so, in the actual over -all farm situa-
tion, might imply that cow numbers be decreased since production of
pasture and roughage may be restricted by the land base of the firm.
Whether or not cow numbers should be decreased, however, could only
be ascertained directly if cow numbers had been included as a variable.
To this extent, the use of a per cow approach has prevented full benefit
to be derived from the original data. So far as labor and capital serv-
ices are concerned, less of these resources should be used. But so
long as the function of Table 17.7 is used as guide, labor and capital
Table 17.8a. Production Elasticities and Opportunity Costs

of Resources for a Swedish Farm -Firm
Production Function
Item Production Function Statistics
Number of farms 273

Year 1956
Crop land 0.23*
Forest land 0.12*
Labor on crops 0.12*
Labor on livestock -0.07
Labor on forest 0.00
Labor out of farm 0.00
Machine services 0.11*
Building services 0.05*
Livestock assets 0.18*
Fertilizer 0.18*
Miscellaneous expenses 0.01
Roughage -0.05*
Grain forage 0.04
Growing timber -0.06*
Management 0.10*
Sum of elasticities 0.97t

2
R 0.93
Opportunity costs:
Crop land ($ per acre) 13.80
Forest land ($ per acre) 9.58
Capital ($ per $) 1.05
Livestock assets ($ per $) 0.17
Growing timber ($ per cubic meter) 0.76
* Significantly differentfrom zero at a probability level of < 5 per cent.

fNot significantly different from unity at probability level of 5 per cent.
would always have a negative marginal product. The study, therefore,

gives no satisfactory estimate of the degree to which these inputs should
be curtailed. To obtain such an estimate it would be necessary to use a
new production function covering a sample with labor and capital serv-
ices at lower levels.
The other production function estimate for Sweden, presented in
Table 17.8a, is a preliminary result from a current cross-sectional
analysis of record data from 2400 farms. 21 As is typical of Sweden, the
majority of the farms are diversified, producing both agricultural and
forest products. To give greater reliability, the sample has been strati-
fied in terms of eight passably homogeneous geographic regions. The
is being carried out by Eje Sandqvist at the Institute of Agricultural Eco-

21
This study
nomics Royal Agricultural College of Sweden, Uppsala. We are grateful to Mr.
of the
Sandqvist and Professor Hjelm for the opportunity to present the preliminary results listed
in the text.
function presented in the table relates to only one of these regions and
refers to the over -all farm operation, the various outputs being con-
sidered as a single aggregate. Inputs, on the other hand, have been
considered in rather disaggregated fashion. In all, there are 15 input
categories, including “management.” The latter is measured tenta-
tively by the inverse of the farmer’s age, on the supposition that edu-
cation is closely related to management, and the fact that younger
farmers generally have a better educational background. The only
other exceptional “input” category in this preliminary analysis is
“labor out of the farm.” This category has a zero production elasticity,
as would be expected on the assumption that only excess family labor is
transferred out of the farm.
While labor on livestock, roughage, and growing timber have slightly
negative production elasticities, only the former two appear anomalous
since growing timber might be expected to curtail current production;
it is associated with other inputs that might otherwise be used to con-
tribute to current production. Arbitrarily selecting a production elas-

ticity of greater than 0.10 as indicative of a resource that is important,
the following resources had an influential role in terms of potential in-
creases in production under prevailing conditions: crop land, livestock
assets, fertilizer, forest land, crop labor, and machine services. Within
the over -all farm production pattern, other resources — with the possi-
ble exception of “management” — appear to be of minor influence al-
though they might be of crucial importance with respect to individual
enterprises.
For convenience in presentation, the opportunity costs of resources
are also given in Table 17.8a; the sample means, marginal products,
and resource use efficiency ratios being presented in Table 17.8b. The
latter statistics have been calculated for each group in a fivefold strati-
fication of the original 273 farms, the strata being based on total farm
output. As the sample means indicate, there is a consistent increase in
the use of all the “input” categories except labor off the farm and man-
agement as one moves from smaller to larger farms. Not unexpectedly,
labor off the farm decreases as farm size rises. However, since farm
output and the various inputs do not change in the same proportions
from one group to the next, the marginal product estimates do not ex-
hibit a uniform pattern of change as output increases: those for crop
labor, building services, livestock assets, and grain forage rise con-
sistently; for crop land, machine services, and fertilizer the
marginal
products decline steadily; those for forest land and miscellaneous ex-
penses show no consistent pattern. These various patterns of change
with farm size are reflected directly in the marginal return to opportu-
nity cost ratios since the same input prices apply to each size
group of
farms. Nothing can be said of the marginal products for labor on live-
stock, roughage, and growing timber except that they are
negative.
Also, apart from the knowledge that it is positive, no data
is available
on the marginal product of “management.”
For all strata it is evident that it would be profitable to increase the
Table 17. Bb. Sweden: Resource Input Means, Marginal Products and Marginal Return to
Opportunity Cost Ratios for Various Size Groups of Farms
Item Size Group
I II m IV V
Sample means:*
Crop land (acres) 3.44 6.72 10.36 16.19 25.86
Forest land (acres) 3.97 6.07 9.87 9.91 21.73
Labor on crops (months) 6.99 8.70 10.72 14.12 18.27
Labor on livestock (months) 6.47 9.13 10.24 10.74 4.77
Labor on forest (months) 1.36 1.56 2.61 3.05 6.26
Labor out of farm (months) 1.92 0.79 0.68 0.54 0.26
Machine services ($) 415.00 745.00 1,130.00 1,576.00 2,251.00
Building services ($) 284.00 361.00 451.00 582.00 613.00
Livestock assets ($) 1,420.00 2,121.00 2,822.00 3,636.00 4,716.00
Fertilizer ($) 177.00 321.00 504.00 780.00 1,150.00
Operating expenses ($) 114.00 255.00 281.00 462.00 1,017.00
Roughage ($) 729.00 1,075.00 1,583.00 1,778.00 2,448.00
Gram forage ($) 838.00 1,145.00 1,496.00 2,000.00 2,452.00
Growing timber (cubic meters) 990.00 1,515.00 2,635.00 2,548.00 4,779.00
Management index 0.020 0.022 0.022 0.027 0.020
All inputs ($) 5,066.00 7,621.00 10,680.00 13,041.00 18,042.00
Marginal products:*
Crop land ($ per acre) 34.88 28.21 25.52 22.92 19.77
Forest land ($ per acre) 16.02 16.55 14.18 19.85 12.49
Labor on crops ($ per month) 53.38 67.59 76.95 81.80 86.65
Labor on livestock ($ per month) .... negative! negative negative negative negative
Labor on forest ($ per month) 0.00 0.00 0.00 0.00 0.00
Labor out of farm ($ per month) .... 0.00 0.00 0.00 0.00 0.00
Machine services ($ per $) 0.80 0.70 0.65 0.65 0.63
Building services ($ per $) 0.62 0.77 0.83 0.93 1.19
Livestock assets ($ per $) 0.40 0/42 0.44 0.48 0.51
Fertilizer ($ per $) 3.23 2.83 2.50 2.26 2.12
Miscellaneous expenses ($ per $).... 0.39 0.27 0.35 0.30 0.19
Roughage ($ per $) negative negative negative negative negative
Grain forage ($ per $) 0.13 0.15 0.16 0.17 0.19
Growing timber ($ per cubic meter) . . negative negative negative negative negative
Marginal return to opportunity cost ratios:*

Cropland 2.53 2.04 1.85 1.66 1.43
Forest land 1.67 1.73 1.48 2.07 1.30
Labor on crops 0.45 0.57 0.65 0.69 0.73
Labor on livestock negative negative negative negative negative
Labor on forest 0.00 0.00 0.00 0.00 0.00
Labor out of farm 0.00 0.00 0.00 0.00 0.00
Machine services 0.76 0.67 0.62 0.62 0.60
Building services 0.59 0.73 0.79 0.89 1.13
Livestock assets 2.35 2.47 2.59 2.82 3.00
Fertilizer 3.08 2.69 2.38 2.15 2.02
Miscellaneous expenses 0.37 0.26 0.33 0.29 0.18
Roughage negative negative negative negative negative
Grain forage 0.12 0.14 0.15 0.16 0.18
Growing timber negative negative negative negative negative
Estimated at the arithmetic mean input levels,

t Exact figures unavailable.
input of crop land, forest land, livestock assets, and fertilizer, and to
decrease the input of labor, machine services, miscellaneous expenses,
and grain forage. In making such changes, priorities within each re-
source category should be as follows: crop land and fertilizer in-
creases on smaller farms, livestock asset increases on larger farms,
decreases in machine services and miscellaneous expenses on larger
farms, and decreases in crop labor and grain forage on smaller farms.
groups I
While building services should be curtailed on farms in size
should be increased in the largest
to IV, in that order, they probably
however, first priority might best be given to re-
size group. Over -all,
ductions in the use of roughage and of labor on livestock.
But for an in-
it would be
dication of just how much these inputs should be reduced,
an estimate of their production elasticities at lower
necessary to have
levels of use than those prevailing In the current sample.
Also, without
additional data on the role of growing timber in the farm organization it
would be unwise to draw any conclusions about this resource. Finally,
we emphasize again that the estimates presented in tables 17.8a and
17.8b are but the result of an exploratory analysis. The implications
drawn must be regarded as somewhat tentative.
NORWAY
Table 17.9 lists four production functions from farm samples in the
marine clay area of southeastern Norway. Each analysis relates to
1954; the main purpose of the original study being to examine the shift
from dairying to small grains that was occurring on many farms in the
22
region. The actual districts sampled lay in the provinces of Vestfold,
Rakkested, and 0stfold. Two samples, each containing only farms
within the size range 10-35 hectares (24.71-86.48 acres), were drawn.
The first consisted of 28 grain farms. These were specified by the re-
quirements that small grains be the major crop and that the farm pro-
gram should not have included commercial dairying for at least two
years, although nondairy livestock production was permitted. The sec-
ond sample consisted of 30 farms whose main income source was dairy-
ing, the farm organization being based on fodder crop production for
dairy feed with beef cattle, vegetables, and poultry as sideline enter-
prises. While the sampling procedure was on an area basis rather
than at random, comparison with census data for 10-35 hectare farms
in the region indicated that the average farm within each sample was
not atypical.
For both the dairy and grain farms samples, separate functions
were fitted for crop and livestock enterprises. While cropping and
livestock production are independent enterprises on the grain farms,
such is not the case on the dairy farms. On the latter, crop production
consists mainly of fodder, including pasture, for subsequent feeding to
livestock. Hence the allocation of resources for crop production on
these farms should depend directly on the market conditions prevailing
23
for dairy products. As well, the crop and livestock enterprises on the
dairy farms have, over various ranges, complementary, supplementary,
and competitive relationships. However, no direct allowance was made
22
Sandberg, O. R. Efficiency of resource usem farming in southeastern Norway. Un-
published M. Sc. thesis. Iowa State University Library, Ames. 1956.
23
See Heady, E. O. The economics of agricultural production and resource use. Prentice-
Hall, Inc., New York. 1952. Pp. 260-67; Nelson, M. et al The production function and linear
programming m valuation of intermediate products. Land Economics, 33; 257-61. 1957.
Table 17.9. Norway: Production Elasticities, Marginal Products and Opportunity Costs
of Resources, Marginal Return to Opportunity Cost Ratios, and Related
Statistics as Derived for Selected Production Function Studies
Production Function for Southeastern Norway
Dairy farms Grain farms

Item
crops livestock crops livestock
Number of farms 30 30 28 26
Year 1954 1954 1954 1954
Land 0.23 0.47*
Labor 0.32* 0.18 0.04 0.42*
Machine services 0.18* 0.19
Other capital services 0.39* 0.80* 0.09 0.79*
Sum of elasticities 1.12 0.98 0.79 1.21*

2
R 0.87 0.86 0.53 0.94
Sample means it
Output ($) 4,964.00 5,936.00 4,733.00 1,624.00
Land (acres) 52.60 51.30
Labor (months) 16.20 20.20 9.80 5.70
Machine services ($) 506.00 713.00
Other capital services ($) 1,132.00 599.00 1,209.00 1,858.00
All inputs ($) 2,922.00 2,080.00 2,734.00 2,276.00
Average products :$
Land ($ per acre) 94.40 92.30
Labor ($ per month) 306.40 293.90 483.00 284.90
Machine services ($ per $) 9.81 6.64
Other capital services ($ per $) 4.38 9.91 3.91 0.87
Marginal products :!
Land ($ per acre) 21.83 43.38
Labor ($ per month) 99.63 53.81 20.92 119.94
Machine services ($ per $) 1.76 1.29
Other capital services ($ per $) 1.72 0.79 0.35 0.69
Opportunity costs:
Land ($ per acre) 1.83 1.83
Labor per month)
($ 73.32 73.32 ,
73.32 73.32
Capital ($ per $) 1.04 1.04 1.04 1.04

cost ratios:!
Land 11.93 23.66
Labor 1.36 0.73 0.29 1.64
Machine services 1.69 1.24
Other capital services 1.65 0.76 0.34 0.66
^Significantly different from zero at a probability level of ^ 5 per cent.

Significantly different from unity at a probability level of « 5 per cent.
tEstimated at the geometric mean input levels.
for such relationships in the present study, the forage output on dairy
farms being considered as a final product.
With the exception of labor, the crop enterprises on both grain and
dairy farms used rather similar quantities of resources. Still, from
comparison of the resource elasticity coefficients it is sure that the
two crop enterprises were based on different production functions, re-
flecting the differences between forage and grain cropping. Likewise,
as would be expected from the distinctions between dairy and meat pro-
duction, the production elasticities and resource use patterns of the
livestock enterprises in the two samples were quite dissimilar. Too, it
might be noted that relative to the other three estimated functions, the
function for cropping on grain farms is not very reliable; it barely ex-
plains half of the variation in observed output.
According to the marginal return to opportunity cost ratios of Ta-
ble 17.9, the average dairy farmer should use more of all resources,
especially land, in his forage cropping and less of all resources in his
dairy enterprise. In other words, the livestock enterprise on dairy
farms should be curtailed and resources shifted to fodder crop produc-
tion. This result is obviously anomalous since the fodder enterprise is
but an intermediate stage in dairy production, there being no off-farm
market for fodder. If the dairy enterprise should be reduced, it seems
sure that fodder production should also be curtailed since current feed-
ing levels are judged to be satisfactory. The above inferences for dairy
farms are, however, rather untrustworthy although they strongly sup-
port the action farmers have been taking in shifting from dairy to grain
production. Through overestimation of the value of forage, the study
probably underestimates the marginal returns on labor and capital
services used in milk production. If such is the case, the marginal re-
turns of resources used in forage cropping would be overestimated.
Concomitantly, the marginal value products of inputs used in livestock
production on the dairy farms would be underestimated. Accordingly,
the average dairy farm may be operating at a position closer to the
optimum than the marginal return to opportunity cost ratios imply. The
real difficulty lies in correctly ascertaining an opportunity cost for the
forage, recognizing that no off-farm market exists for this intermediate
product. To this end linear programming procedures could be used.
For the average grain farm, the marginal return to opportunity cost
ratios suggest that additional land, especially, and perhaps machine
services, should be used for cropping with a concomitant reduction in
the quantity of labor and other capital services. As well, these farms
should use more labor for livestock production. Such labor could be
obtained by transferring some of the excess labor out of cropping.
Over -all, given the limitations noted with respect to the dairy farm
functions, the main implication of this Norwegian study appears to be
that the acre size of grain farms should be expanded as a means to the
better utilization of labor, machinery, and other capital services. At
the same time, more labor and less capital services should be used in
the livestock enterprise on grain farms. The surplus livestock capital
could be profitably used in grain production.
1
ISRAEL
The five production function estimates listed for Israel in Table

17.10 are rather unique. Each refers to the same sample of farms but
24
relates to a different year within the period 1953 to 1957. The sample
farms are family units averaging about 7.5 acres. They are partially
irrigated and produce grain, vegetables, and forage crops together with
citrus and vine fruits, having also a few cows and poultry. Such pro-
duction units, grouped in villages on a co-operative basis, are typical
of the bulk of Israel’s agriculture.
Table 17.10. Israel: Production Elasticities, Marginal Products and Opportunity

Costs of Resources, Marginal Return to Opportunity Cost
Ratios, and Related Statistics as Derived for
Item Production Function for 68 Mixed Family Farms
Year 1953 1954 1955 1956 1957
Elasticities:
Land 0.04 0.04 0.01 0.02 0.04
Labor 0.26* 0.23* 0.27* 0.29* 0.20*
Operating capital 0.62* 0.69* 0.68* 0.54* 0.71*
Building assets 0.12* 0.08* 0.06* 0.12* 0.10*
Livestock assets 0.08* 0.05* 0.09* 0.06* -0.01
Sum of elasticities 1.12 1.091 l.llt 1.03 1.04
R2 0.85 0.94 0.95 0.92 0.92
Marginal products :?
Land ($ per acre) 297.25 266.70 15.98 30.54 73.53
Labor ($ per month) 294.51 388.68 108.76 146.36 115.13
Operating capital ($ per $) 1.30 1.45 1.31 0.96 1.28
Building assets ($ per $) 0.22 0.23 0.20 0.37 0.31
Livestock assets ($ per $) 0.14 0.10 0.18 0.12 -0.01
Opportunity costs:
Labor ($ per month) 293.79 347.03 75.83 92.43 98.80
Capital ($ per $) 1.10 1.10 1.10 1.10 1.10
Building assets ($ per $) 0.10 0.10 0.10 0.10 0.10
Livestock assets ($ per $) 0.20 0.20 0.20 0.20 0.20

cost ratios:?
Labor 1.00 1.12 1.43 1.58 1.16
Operating capital 1.18 1.32 1.19 0.87 1.16
Building assets 2.20 2.30 2.00 3.70 3.10
Livestock assets 0.70 0.50 0.90 0.60 -0.05
* Significantly different from zero at a probability level of « 5 per cent,

t Significantly different from unity at a probability level of 5 per cent,
t Estimated at the geometric mean input levels.
^Mundlak, Y. Economic structure of established family farms. Falk Project for Eco-
nomic Research In Israel. Hebrew University of Jerusalem, Rehovot. The results published
here, by courtesy of Dr. Mundlak, are preliminary.
Perusal of Table 17.10 indicates that little year to year variation

occurred in the production elasticities, the estimates ranging from 0.01
to 0.04 for land; from 0.20 to 0.29 for labor; from 0.54 to 0.71 for op-
erating capital; from 0.06 to 0.12 for building assets; and from -0.01
to 0.09 for livestock assets. In no case do the estimated elasticities for
a particular resource follow a consistent pattern of change over the five
years studied. The variations that did occur may therefore be attrib-
uted mainly to statistical and climatic variability rather than to changes
in the technologies used which are known to have changed little over
the period. In other words, it may be said that the production function
pertinent to the average mixed farm was, for all practical purposes, in-
variant over the period 1953 to 1957 and that the five functions are but
alternative estimates of the same true production function.
While it has no important influence on the marginal return to oppor-
tunity cost ratios, it should be noted that the marginal products and op-
portunity costs of Table 17.10 reflect the 1955 devaluation of Israel’s
currency relative to the United States dollar. Also, no opportunity cost
or resource use efficiency ratios are given for land. The reason is that
no market for farm land exists in Israel, farms being held on long term
lease from the government under nominal rentals that are largely sym-
bolic. So far as the other input categories are concerned, the yearly
marginal return to opportunity cost ratios exhibit rather uniform pat-
terns. Certainly it would be unwarranted to say that resource use effi-
ciency has changed for the better or the worse over the five year pe-
riod analyzed. Over-all, it appears that these Israeli farms have been
using too many livestock assets and too little of the services of build-
ings, labor, and operating capital. Given the fixed land base of the
farms, resources could be used more efficiently and profits would be
raised if the input of buildings, operating capital, and labor were raised
in that order. With increases in these resources, it is likely that the
marginal product of livestock would also rise so that there might be no
need for a reduction in the quantity of livestock used. Also, insofar as
the marginal value product of land exceeds the nominal rentals paid for
land, it would be economic for the government to raise the rental price
of land; provided of course that on the basis of other welfare policies it
was not desired to subsidize the farm sector of the economy.
INDIA
Estimates from two production function studies of Indian agriculture

are given in Table 17.11. The first study, involving four functions, re-
lates to family farms on loamy soils in the Meerut and Muzaffarnagar
districts of the Indo-Gangetic Plain in the north central state of Uttar
Pradesh. Agriculturally, this area is one of the most productive and
prosperous in India. Rainfall in the region averages about 32 inches;
deficiencies in precipitation are of less concern than excesses since
nearly three-quarters of the farm land is irrigated. The listed
1
Table 17.11. India: Production Elasticities, Marginal Products and Opportunity Costs of Resources,
Marginal Return to Opportunity Cost Ratios, and Related Statistics
for Selected Production Function Studies

Uttar Pradesh Andhra Pradesh
Item |
planted irrigated irrigated dry mixed

whole farm
sugarcane wheat farms farms farms
Number of farms 49 63 63 60 18 72 106
Year 1954-55 1955-56 1955-56 1955-56 1953-54 1953-54 1953-54
Land 0.23* 0.22* 0.37* 0.50* 0.57* 0.31* 0.14
Labor 0.58* 0.29* 0.69* -0.26 0.14 0.04 0.26
Capital services 0.18* 0.25* 0.30* 0.53* -0.08 0.07 0.13*
Capital assets 0.05 0.21 -0.27 0.16
Sum of elasticities 1.04 0.97 1.09 0.93 0.63^ 0.42 0.53t
R2 0.80 0.86 0.57 0.81 0.64 0.70 0.76
Sample means:^
Output ($) 448.00 378.00 98.00 59.00 12.50 53.00 93.00
Land (acres) 9.70 11.60 1.20 1.60 1.20 9.20 10.20
Labor (months) 13.00 9.30 1.60 1.90 7.00 19.70 25.00
Capital services ($) 42.00 50.50 25.60 3.90 0.80 5.80 9.40
Capital assets ($) 50.60 17.70 2.30 3.90
All inputs ($) 439.00 367.00 74.00 63.00 82.00 213.00 562.00
Average products:^
Land ($ per acre) 46.20 32.60 81.70 36.90 10.40 5.80 9.10
Labor ($ per month) 34.50 40.60 61.20 31.00 1.80 2.70 '
3.70
Capital services ($ per $).... 10.70 7.50 3.80 *15.10 15.60 9.10 9.90
Capital assets ($ per $) 8.80 21.30 42.60 15.10
Marginal products
Land ($ per acre) 8.82 7.18 29.94 18.44 20.85 3.65 1.94
Labor ($ per month) 19.75 11.71 40.89 -8.04 0.93 0.22 1.47
Capital services ($ per $).... 1.94 1.84 1.19 8.21 -1.02 0.28 0.42
Capital assets ($ per $) 0.44 4.39 -11.78 2.43
Opportunity costs: j
Land ($ per acre) I 8.40 8.40 8.40 8.40 26.25 7.35 36.75
Labor ($ per month) 22.97 22.97 22.97 22.97 7.10 7.10 7.10
Capital ($ per $) 1.20 1.20 1.20 1.20 1.20 1.20 1.20
Capital assets ($ per $) 0.33 0.33 0.33 0.33

cost ratios :$
Land 1.05 0.86 3.60 2.22 0.79 0.50 0.05
Labor 0.86 0.51 1.78 -0.35 0.13 0.03 0.21
Capital services 1.62 1.53 0.99 6.84 -0.85 0.23 0.35
Capital assets 1.33 13.30 -35.70 7.36
* Significantly different from zero at a probability of « 5 per cent,

f Significantly different from unity at a probability level of 5 per cent.
I Estimated at the geometric mean input levels.
functions are derived from random subsamples of a much larger group

25
of farms examined in a cost -accounting farm management project .
Typically, the farms produce wheat and sugar cane. The latter crop is
generally planted biennially, two crops being harvested per planting.
The first year’s crop is known as “planted sugar cane” and the second
25
Agrawal, G. D. and Foreman, W. J. Farm resource productivity in west Uttar Pradesh.
Indian Jour. Agric. Econ., 14(4): 115-28. 1959; and Agrawal, G. D. Studies in economics of
farm management in Uttar Pradesh, 1954-55. Directorate of Economics and Statistics,
Ministry of Food and Agriculture, New Delhi. March, 1957.
year’s as “ratoon sugar cane.” Labor used on the farms is mainly

from the farm family, but often there is a more than negligible input of
hired labor. The average land base is about 10 acres with a range from
one to 25 acres.
The first two functions listed in Table 17.11 refer to the whole farm
program involving wheat, planted and ratoon sugar cane. The other two
Uttar Pradesh functions are individual enterprise studies. One refers
to planted sugar cane and the other to the wheat enterprise. The ran-
dom sample used in deriving these two functions is the same as that
used in the 1955-56 whole farm analysis. It should be noted that not all
enterprises on these farms have been examined individually, no analysis
being available for ratoon sugar cane. Also, the labor input category
refers to both human and bullock labor; the average labor input in
months over all farms consisting of approximately seven parts human
labor (costing $7.22 per man month) to three parts of bullock labor
(costing $1.05 per bullock pair day). These proportions will be assumed
to remain fixed in the following discussion of the labor statistics de-
rived from the estimated functions. Needless to say, the assumption is
one of necessity, and not of choice.
While the production elasticities of land and capital services in the
Uttar Pradesh whole farm analyses do not vary over the two years ex-
amined, there are fairly large inter -year variations in the elasticities
of labor and capital assets. These differences may be due to sampling
effects although climatic variation is probably of more importance.
While 1954-55 was a year of average rainfall, 1955-56 was above aver-
age with some flood damage occurring. The production elasticities for
planted sugar cane and wheat tend to imply that physically excessive
quantities of capital assets were being used in the planted sugar cane
enterprise while too much labor was also being used in the wheat enter-
prise. Of course, the elasticity estimates for these resources may be
much lower than their true values; despite their relatively large abso-
lute size, they are not significantly different from zero at the 5 per cent
probability level. Also, it must be remembered that 1955 -56 was a
wetter year than is usual.
Consider now the marginal return to opportunity cost ratios found in
these Uttar Pradesh studies. From the whole farm analyses, it seems
that too few capital services and assets and too much labor were used,
the input of land being very near optimal. But these general implica-
tions must be modified in terms of the individual enterprise studies.
These show quite marked differences to exist between the resource use
efficiency ratios of the individual enterprises; the differences being
“averaged out” in the whole farm analyses which lump together the pro-
duction of wheat, and planted and ratoon sugar cane. On an individual
crop basis, the statistics strongly suggest that too much labor and too
little capital were used in wheat production, while the use of more labor
and less capital — especially capital assets — would have increased the
profitability of the planted sugar crop. Still, relative to the value of
land and labor inputs, capital is only of minor importance. Hence, in
absolute terms, only small changes may be needed to induce optimality

in the use of capital assets and services assuming, of course, that their
elasticities are not negative over lower stages of the production sur-
face. As noted before, no production function estimates are available
for the ratoon sugar crop. Still, a little can be said of this crop from
the data for the other two enterprises. Thus, while land use was satis-
factory on a whole farm basis and since too little land appears to be
used for wheat and planted sugar cane, it may be concluded that too
much land was being devoted to the ratoon sugar cane crop.
Data from an area sample of villages in the Telengana district was
used to derive the Andhra Pradesh estimates; a function being fitted for
each type of farm found in the region: fully irrigated, dry, and mixed. 26
The major activity within each of these groups is cereal cropping of a
virtually subsistence character. As the sample means of output and in-
puts listed in Table 17.11 indicate, Andhra Pradesh farms are of very
small scale. Indeed, this southeast state is one of the poorest in India.
Hence it is not surprising that at market valuations the mean of all in-
puts exceeds mean output in each of the samples. However, since the
labor input is nearly all family labor paid at a nominal rate, the dis-
crepancy between output and input is probably less than is indicated by
the data of Table 17.11. In fact, the quoted opportunity cost for labor is
rather unreal. There is virtually no market for hired labor except on
large farms, and these constitute only a small proportion of the total.
Too, the prices given for land and capital are averages and mask quite
large variations from tract to tract in the region.
From the sums of the production elasticities, it is apparent that
severe decreasing returns to scale prevail with respect to land, labor,
and capital services on these Andhra Pradesh farms. Their existence
emphasizes the importance of managerial ability in farming in this
area, granted that no other inputs of importance have been omitted from
the analysis and that constant returns prevail relative to all inputs. If
these assumptions are correct, so that the difference between unity and
the sum of each set of elasticities may be taken as approximating the
production elasticity of management, then it appears that management
has a relatively large influence on production under current conditions.
Thus the percentage increase in output from a 1 per cent increase in the
input ofmanagement would be roughly 0.4, 0.6, and 0.5 per cent, respec-
tively,on irrigated, dry, and mixed farms. These implications, more-
over, are supported by the experience of agricultural technicians in the
region. Nor are they surprising on a priori grounds, given the primitive
nature of agricultural production in the area.
In terms of their marginal return to opportunity cost ratios, all
three types of farm appear to be grossly inefficient. Under the assumed
opportunity costs, each sample used too much land, labor, and capital.
26
Suryanarayana, K. S. Resource returns in Telengana farms: a production function
study. Indian Journal of Agricultural Economics, 13(2): 20-26. 1958. A noteworthy feature
of this study is that land acreage was corrected for soil heterogeneity by a soil fertility
index.
Moreover, these conclusions remain true even if the monthly wage rate
of labor is lowered from $7.10 to as little as $1.50 so as to correspond
to the minimum payments received by family labor. Indeed, were it not
for the subsistence element in this agriculture, the data of Table 17.11
would imply that farming in the area could never be profitable under
current conditions. The difficulty, of course, is that there are ex -
tremely few other means of earning a livelihood. Alternatively, farm-
ing in the area might be reorganized by the provision of nonfarm em-
ployment opportunities and the introduction of new production techniques
of a sufficiently productive nature to make commercial agriculture pos-
sible.
Comparison of the two India studies indicates that farming in the two
states has few common
features. The semicommercial farms of Uttar
Pradesh are large and profitable relative to the subsistence units of
Andhra Pradesh. While Andhra Pradesh farms exhibit severe decreas-
ing returns to land, labor, and capital, constant returns appear to pre-
vail in more prosperous Uttar Pradesh, perhaps indicating a more effi-
cient type of farm organization in the latter state. Distinct differences
are also noticeable in the patterns of resource use efficiency: Uttar
Pradesh farms used too little capital and too much labor, while Andhra
Pradesh farms had a surfeit of all resource services, with the probable
exception of management.
JAPAN
Relative to most other countries, a large number of cross-sectional
production function studies are available for Japanese agriculture.27
Results of two of these studies, involving seven functions, are presented
in Table 17.12. The first study is based on cost data from an area sam-
ple of mixed farms in the Shizuoka Prefecture of Honshu, separate
functions being estimated for each type of enterprise found in the area.28
Accordingly, the listed sample means of input and output for sweet po-
tato, tea, and paddy rice production on these Honshu farms give
only a
partial indication of the over -all scale of farm operation. The other
four functions of Table 17.12 refer to various types of rice producing
farms in Hokkaido. 29 Data for the two 1955 functions were taken from a
random sample of farms keeping records under the supervision of the
Ministry of Agriculture. The sample, consisting of 24 smaH and
27
See Kamiya, K.
On productivity of labor. Journal of Rural Economics, 17: 22-44.
941; Okawa, The theory and measurement of food economy. Hitosubashi
I.
College of
Commerce and Economics. 1945. Ch. 8; Watanabe, T. Theory of production
functions.
Journal of Rural Economics, 21. 1945; Yuwata, Y. Production
functions for rice and barley
Quarterly Journal of Agricultural Economy, 7. 1953.
Tsuchiya, K. Production functions of agriculture in Japan. Quarterly Journal
of Agri-
cultural Economy, 9. 1955.
29
Takayama, T. A study on the Cobb-Douglas production function -
with an application
to rice production in Hokkaido. Review of the Society of
Agricultural Economics of Hokkaido
University, 15: 1-24. 1959.
Table 17.12. Japan: Production Elasticities, Marginal Products and Opportunity Costs
of Resources, Marginal Return to Opportunity Cost Ratios,
and Related Statistics as Derived for Selected

Honshu Hokkaido
Item
sweet
tea rice rice
potatoes
Number of farms 11 30 69 24 36 17 42
Year 1951 1951 1951 1955 1955 1956 1958

Land 0.85* 0.29* 0.56* 0.80* 0.55* 0.87* 0.78*
Labor 0.29* 0.30* 0.29* 0.10 0.29* 0.12 0.21*
Capital assets 0.00 0.46* 0.15* 0.02 0.18* 0.00
Sum of elastic it iest 1.14 1.05 1.00 0.99 1.02 0.99 0.99
2
R 0.82 0.85 0.77 0.70 0.61 0.81 0.60
Sample means ^
Output ($) 936.00 2,039.00 1,877.00 2,036.00
Land (acres) 3.60 7.00 7.10 6.30
Labor (months) 10.90 19.70 22.20 13.30
Capital services ($) 287.20 551.20 214.50 125.30
Capital assets ($) 394.50 672.10 115.60
All inputs ($) 1,084.00 2,026.00 1,108.00 1,043.00
Average products
Land ($ per acre) 260.00 291.30 264.40 328.40
Labor ($ per month) 85.90 103.50 84.50 153.10
Capital assets ($ per $) 2.40 3.00 17.60
,
Marginal products^
Land ($ per acre) 212.53 135.03 97.46 208.70 160.21 242.87 229.66
Labor ($ per month) 25.29 21.57 8.18 10.43 7.56 15.57 28.29
Capital assets ($ per $) 0.00 0.51 0.41 0.03 0.56 0.00
Opportunity costs:
Land ($ per acre) 6.77 6.97 6.83 89.94 89.94 77.12 83.64
Labor per month)
($ 16.89 16.89 16.89 36.11 36.11 35.39 40.30
Capital ($ per $) 1.11 1.11 1.11 1.11
Capital assets ($ per $) 0.20 0.20 0.20 0.20 0.20 0.20

cost ratios:!
Land 31.39 19.37 14.27 2.32 1.78 3.15 2.75
Labor 1.50 1.28 0.48 0.29 0.21 0.44 0.70
Capital assets 0.00 2.55 2.05 0.15 2.80 0.00
^Significantly different from zero at a probability level of ^ 5 per cent.

iEstimated at the arithmetic mean input levels in the Honshu studies and at the geometric means m
the Hokkaido studies.
36 large farms, covered most of the rice producing areas of Hokkaido.

On a priori grounds, farmers in the sample were thought to be of above
average managerial capacity. Main objective of the study was to ascer-
tain if there were structural differences between large and small units
in their production process. Both the 1956 and the 1958 analyses refer
to well-established farms an alluvial soil in the
Kuriyama region of
Hokkaido. Like the 1955 study, they relate to the over -all activity of the
farm. While the 1956 function is based on an area sample, a random

sampling procedure was used in the 1958 study.
A pattern is discernible among the resource production elasticities
listed in Table 17,12. Excluding the function for tea, capital has a low
elasticity with a range from zero to 0.18. The land coefficients are
relatively large, ranging from 0.55 to 0.87. Labor appears to be of in-
termediate importance with elasticities from 0.10 to 0.30. The expla-
nation, of course, is that land is used very intensively in the production
of paddy rice and sweet potatoes. Relative to other inputs, the greatest
production response is to be expected from an expansion of the land
base with other factors held fixed; rather than from an increase in the
use of other resources with land held constant. Indeed, under the con-
ditions prevailing in the sweet potato and some of the rice samples, an
incremental increase in the use of capital assets (mainly machinery) or
services (mainly fertilizer) would have no influence on production. Tea
production does not fit such a picture because of the distinctive features
of its production process, reflecting the specialized nature of tea crop-
ping relative to other types of crop production.
The marginal return to opportunity cost ratios of Table 17.12 evi-
dence rather severe inefficiencies in Japanese resource use. Espe-
cially striking are the exceedingly high ratios for land of 31.39, 19.37,
and 14.27, respectively, in sweet potato, tea, and rice production in
Honshu. They result not so much from inefficiencies on the part of in-
dividual farmers, but from the exceedingly low land rents fixed by gov-
ernment action under the Land Reform Law. This law also limits the
possibility of renting additional land. In a normative economic sense,
the land rents are obviously too low. Still, they may be justifiable
on'
welfare grounds. We do not know.
Disequilibria in land use is not so great on the Hokkaido farms, al-
though it is apparent that substantial increases in the land input would
be profitable. With the exception of the 1955 study which implies
more
capital assets should be used on larger farms, Hokkaido farmers
ap-
pear to have used too much labor and capital. A decrease in the
input of
labor and capital, as well as tending to optimize the use of
these re-
sources, would do much towards reducing inefficiency in the use
of
land. The only common feature Honshu enterprise analy-
of the three
ses is the indication of too little
land being used. In common with the
Hokkaido rice samples, Honshu rice farmers use too much labor.
How-
ever, additional labor should be used in Honshu sweet
potato and tea
production, with additional capital being used in the tea
and rice enter-
prises rather than in sweet potato production. More of all
resources
should be used in tea production, especially land and
capital. Land and
labor could be profitably substituted for capital in
the sweet potato en-
terprise, while land and capital should be substituted
for labor in the
rice enterprise.
Recognizing the general importance of rice in Japanese
agriculture,
the above analyses tend to substantiate the opinions
of agricultural au-
thorities that these Japanese farmers generally
use more than enough
fertilizer, are overcapitalized in farm machinery, and have a surfeit of

labor. Either the land input should be increased or nonland resource
inputs curtailed, if farmers are to maximize their profits. Concomi-
tantly, it would seem necessary for the sake of national welfare to pro-
vide alternative employment opportunities for surplus farm labor. Too,
insofar as the inefficiencies shown to exist are caused by the fairly
stringent price and production controls applied in Japanese agriculture
and not by lack of knowledge on the part of farmers, strong welfare
grounds would be necessary to justify the controls.
TAIWAN
The only two production function studies available for Taiwan are
listed in Table 17.13. One relates to sugar cane and the other to a
cropping system that is competitive with sugar cane. Both types of
production are important in Taiwan's agricultural economy. The data
on which the functions are based relate to a random sample of 100
30
farms in the Tainan area of southern Taiwan. The observations ana-
lyzed, however, do not refer to the farms themselves but to a field of
sugar cane and a field of competitive crop on each farm. Hence the
sample means of output and inputs listed in Table 17.13 for each crop
give no indication, when added together, of the over -all scale of opera-
tion of the average sample farm. As well, the consideration of fields,
rather than the over -all cropping enterprise on each farm, implies that
the fitted functions may not be extrapolated to the over -all farm situa-
tion without some danger of error. Still, since the relative proportions
of the various inputs are most likely the same for a whole farm as for
a single field, the study provides an estimate of resource use disequi-
libria within and between the two cropping enterprises on a whole farm
basis. Also, given the competitive relationship between the two crop-
ping systems, the estimated functions are of direct relevance as deci-
sion guides between the two alternatives. Thus it is reassuring that the
marginal productivity estimates for the two cropping systems diverge
in the fashion expected from comparison of the labor to land and capital
to land ratios in the two systems.
While an equilibrium amount of capital services appears to have
been used for both cropping systems, proportionately too much land and
too little labor were being used. The greatest disequilibria exist in the
excessive proportion of land used in sugar cane production and the re-
strictive quantity of labor utilized relative to other inputs in the com-
petitive cropping system. To rectify these inefficiencies, the best pro-
cedure would seem to be the use of additional labor. Increasing the
labor to land ratio would increase the marginal productivity of land and
30
Wang, Y. Resource returns and productivity coefficients for selected crop systems in
Taman area. Proceedings of Agricultural Economics Seminar, Sept. 16-20, 1958. National
Taiwan University, Taipei. 1959. Pp. 90-98.
Table 17.13. Taiwan: Production Elasticities, Marginal Products and Opportunity

Costs of Resources, Marginal Return to Opportunity Cost
Ratios, and Related Statistics as Derived for
Production Function for Tainan Region

Item
Sugar cane Other crops
Number of fields 100 100
Year 1957 1957
Land 0.36* 0.44*
Labor 0.25 0.33*
Capital services 0.34 0.31*
Sum of elasticities^ 0.95 1.08
R2 0.60 0.78
Sample means :$
Output ($) 318.70 287.10
Land (acres) 1.20 0.87
Labor (months) 3.50 2.40
Capital services ($) 78.00 67.50
All inputs ($) 324.00 244.00
Average products ;!
Land ($ per acre) 265.60 330.00
Marginal products :t
Land ($ per acre) 96.44 142.41
Opportunity costs:
Land ($ per acre) 246.67 246.67
Capital ($ per $) 1.33 1.33
Marginal return to opportunity cost ratios:!
Land 0.39 0.58
Labor 1.66 2.84
Capital services 1.02 0.99
^Significantly different from zero at a probability level of « 5 per cent.

TNot significantly different from unity at a probability level of 5 per cent.
Êstimated at the arithmetic mean input levels.
decrease the marginal productivity of labor. Both of these changes

would tend to reduce the disequilibria in land and labor use. Concomi-
tantly, an increase in labor inputs would tend to raise the marginal pro-
ductivity of capital so that the marginal return to opportunity cost ratio
for capital would be greater than unity. In consequence, it would also be
desirable to use a little more capital - assuming the land input to be
held fixed — despite the fact that the resource use efficiency ratios for
capital services of Table 17.13 are very close to unity. Given the
marginal return to opportunity cost ratios prevailing in the two samples,

priority in the use of additional labor should be given to the competitive
crop enterprise.
Having considered the collected functions on a country by country
basis, we may now examine their implications within an international
framework. This will be done in two stages. First, a comparison of
the production elasticities will be made. As outlined in Chapter 2,
these elasticities of production play an important role in the determi-
nation of resource marginal productivity. International comparisons of
the efficiency of resource use in terms of profit maximization will then
be made, using the estimated marginal return to opportunity cost ratios
as a guide to the efficiency of resource use.
INTERNATIONAL COMPARISON
OF RESOURCE PRODUCTION ELASTICITIES
For the purpose of making international comparisons of input elas-

ticities, resources will be considered under three categories: land,
labor, and other services. The other services category consists of
many particular types of inputs such as livestock, miscellaneous capi-
tal services, fertilizer, purchased feed, building and machinery serv-
ices, etc. It is a catch-all category, necessitated by the fact that land
and labor are the only inputs considered individually in a majority of
the estimated functions presented above. Certainly it would have been
desirable to consider capital services as a separate category. But to
do so is unfeasible. There is insufficient correspondence between the
inputs designated as capital services in the various national studies.
For instance, while the United States, Canadian, South African, Aus-
trian, and Taiwan studies use the traditional threefold resource classi-
fication of land, labor, and capital services, the Australian, New Zea-
land, United Kingdom, Swedish, Norwegian, Indian, and Japanese studies
generally include a number of capital input categories. Accordingly, in
these latter studies we have added the nonland and nonlabor resource
elasticities together to give a single elasticity coefficient for these
other resource services. This aggregate figure may be taken as cor-
responding —admittedly roughly, but closely enough for current pur-
poses — to the capital input elasticity derived under the tripartite clas-
sification using land, labor, and capital. Still, the elasticity of "other
services” needs to be interpreted with caution. Strictly, it is the per-
centage increase in output that is expected to result from a 1 per cent
increase in the input of each of the resources included in the other
services category, the input of land and labor being held constant.
The estimated elasticities and their sums are presented in Table
17.14, the location and specification of each parent function noted on the
left. These functions are the various ones already discussed on a coun-
try by country basis, with the exception that only one estimate is in-
cluded in those cases where a number of studies are available for a
Table 17.14. Production Elasticities of Land, Labor, and Other Resource Services as Estimated in
Forty-one Cross-Sectional Cobb-Douglas Type Production Function Studies
Elasticity of Production
Sum of
Location of Sample Function for Land Labor Other
Elasticities
services services services
United States, northern Iowa corn 0.91 0.08 (27) 0.16 (25) 1.15
Japan, Honshu sweet potatoes 0.85 0.29 (9) 0.00 (31) 1.14
United States, southern Iowa corn 0.79 0.09 (26) 0.39 (18) 1.27
Japan, Hokkaido* rice 0.75 0.18 (20) 0.07 (28) 1.00
India, Andhra Pradesh irrigated farms 0.57 0.14 (24) -0.08 (32) 0.63
Japan, Honshu rice 0.56 0.29 (10) 0.15 (26) 1.00
United States, Montana wheat 0.50 0.04 (29) 0.58 (5) 1.12
India, Uttar Pradesh wheat 0.50 -0.26 (32) 0.69 (3) 0.93
Norway, southeast cereals 0.47 0.04 (30) 0.28 (24) 0.79
Taiwan, Tainan cereals 0.44 0.33 (5) 0.31 (23) 1.08
New Zealand, Canterbury sheep 0.42 C. 15 (23) 0.54 (10) 1.11
United States, Alabama crops 0.39 0.32 (6) 0.46 (15) 1.17
South Australia dairy 0.39 0.25 (13) 0.54 (11) 1.19
Canada, Alberta wheat, beef 0.39 0.20 (18) 0.34 (21) 0.93
India, Uttar Pradesh sugar cane 0.37 0.69 (1) 0.03 (30) 1.09
Taiwan, Tainan sugar cane 0.36 0.25 (14) 0.34 (22) 0.95
|
Sweden mixed farms 0.35 0.05 (28) 0.57 (6) 0.97

India,Andhra Pradesh dry farms 0.31 0.04 (31) 0.07 (29) 0.42
Japan, Honshu tea 0.29 0.30 (8) 0.46 (16) 1.05
United States, Iowa-Hlinois crop- share 0.29 0.25 (15) 0.48 (14) 1.02
Australia, New South Wales dairy 0.28 0.22 (17) 0.42 (17) 0.92
South Africa, Kalihari cattle fattening 0.28 0.13 (25) 0.55 (8) 0.96
India, Uttar Pradesh t wheat, sugar cane 0.23 0.43 (3) 0.35 (20) 1.01
Norway, southeast fodder 0.23 0.32 (7) 0.57 (7) 1.12
United States, Iowa-Hlinois livestock- share 0.23 0.18 (21) 0.53 (12) 0.95
Canada, Alberta cattleranches 0.20 0.37 (4) 0.39 (19) 0.97
South Airica, Kalihari cow-calf ranches 0.19 0.19 (19) 0.52 (13) 0.90
India, Andhra Pradesh mixed farms 0.14 0.26 (11) 0.13 (27) 0.53
Austria t mixed farming 0.13 0.26 (12) 0.61 (4) 1.00
Australia, New
South Wales sheep 0.10 0.59 (2) 0.55 (9) 1.24
United States, Iowa-Hlinois owners 0.09 0.17 (22) 0.73 (2) 0.99
Israel § mixed farms 0.03 0.25 (16) 0.80 (1) 1.08
Norway, southeast beef cattle 0.42 0.79 1.21

United Kingdom, England dairy 0.29 0.83il 1.12
Western Australia dairy 0.23 0.76 0.99
United States, Alabama livestock 0.23 0.74 0.97
Norway, southeast dairy 0.18 0.80 0.98
United States, southern Iowa hogs, cattle 0.12 0.98 1.10
United States, Montana cattle 0.08 0.94 1.02
United States, northern Iowa hogs, cattle 0.08 j
0.91 0.99
Sweden dairy -0.05 s
1.23 1.18
Average of coefficients in 1955, 1956, and 1958 analyses.

^ Average of 1954-55 and 1955-56 analyses.
* Study of 37 strata.
§ Average of analyses for 1953, 1954, 1955, 1956, and 1957.
Încludes land services.
given area and type of product. At the top of the table are listed those
analyses in which land was a relevant input. These analyses relate to
crops and grazing livestock. Listed at the bottom of the table are the
analyses in which, with the exception of the United Kingdom study, land
was not a relevant input variable. The latter studies relate to livestock
production, generally under nongrazing conditions. The exceptions are
the United Kingdom analysis wherein land services are included in
other services, and the Swedish milk production study in which pasture
intake substitutes for land services and is included in the other services
category. The only national analyses not included in Table 17.14 are
the United Kingdom Bristol Province dairy study and the Canadian beef
feedlot analysis. Both of these include only inputs in the other services
group. The following discussion will be confined to those functions
which include land, labor, and other services. In all, there are 32 such
functions. For present purposes, they are implicitly assumed to con-
stitute a random sample from the super -population of all agricultural
production function studies that might have been carried out over the
globe during the period 1950 to 1958. We will not discuss the set of
livestock functions which do not include land. This sample is too small
to warrant detailed examination.
Those analyses which include land, labor, and other services are
ordered in Table 17.14 in terms of the size of the production elasticity
of land. Such a basis of ordering is used because land is the most fixed
of the three resource categories; labor and other services being more
mobile, in that order. For convenience, the ordering of the labor and
other services coefficients is also noted in the table within brackets be-
side the estimated coefficients. The order, from larger to smaller,
ranges from (1) to (32).
Relative to the size of the land coefficient, it is obvious that no pat-
tern of ordering prevails between countries. But a strongly identifiable
sequence does exist in terms of the type of production. For cropping,
the production elasticity of land is generally higher than it is for graz-
ing livestock or mixed crop -livestock production. Thus while all of the
livestock studies have production elasticities for land of less than 0.42,
a majority of the crop functions have an elasticity for land between 0.44
and 0.91. This result is not unexpected on theoretical grounds. Land is
used far more intensively in crop production than in livestock produc-
tion. Consequently one would expect an expansion of the land input to
have a greater influence on crop output than on livestock output. Nor is
it surprising that corn production in Iowa and Japanese rice production
have the largest production elasticities for land. In both these regions,
the intensity of land use exceeds that in any of the other regions ex-
amined.
Although the pattern is not so pronounced as for the land coefficient,
the data of Table 17.14 indicates that the elasticity of other services
tends to be lower for crops than for livestock and mixed live stock -crop
activities. The justification, of course, is the converse of that which
explains the high elasticity of land for cropping relative to livestock
production. The more intensively the land is used, the less the re-
sponse to additional inputs of other services is likely to be. Within the
crop enterprise, it also appears that the elasticity of other services
tends to be higher on average for the non -Asian samples than for the
Asian. Concomitantly, the coefficient for crop labor is lower on aver-
age among the non-Asian samples than in the Asian. This result is
surprising since Asian countries are very plentifully supplied with
labor. They might be expected to be using labor at such intensive
levels that increased inputs of this resource would yield smaller
additions to output than would occur in non-Asian crop production. Still,

it is explainable in that other services tend to be of greater importance
in the non-Asian crop production samples due to the more mechanized

nature of their production processes. Unlike the elasticities for land
and other services, those for labor exhibit no pronounced pattern rela-
tive to cropping as against livestock production.
Looking at the over -all picture presented by the elasticities of Ta-
ble 17.14, the most striking feature is the variation present within each
input category. Obviously there can be little logic associated with the
selection of any particular triplet of figures as the elasticities of land,
labor, and capital in macro studies of an international or national type
involving specification of a production function for the agricultural sec-
tor of the economy. This point is again obvious from the frequency dis-
tributions of the elasticities, and the standard deviations of their means,
presented in Table 17.15. With a mean value of 0.39 over the sample of
32 studies, the other services category tends to be the most important
in terms of output response, closely followed by land with an average
elasticity of 0.38. As would be expected, given the tendency to a surfeit
of labor that generally prevails in agriculture, the average labor elas-
ticity of 0.21 is the smallest of the three elasticities. In terms of rela-
tive variability, the labor coefficient is the most variable, although its
variance is somewhat smaller than those of land and other services
which have approximately equal standard deviations.
Without implying that the mean elasticities of Table 17.15 have any
broad usefulness of an empirical nature, it is interesting to note that
they differ quite markedly from the elasticities hypothesized for agri-
cultural production in the literature on international trade and develop-
ment. For instance, Tinbergen and Polak used elasticities of 0.1, 0.7,
Table 17.15. Frequencies, Means, Standard Deviations, and

Coefficients of Variation of the Production
Elasticities of Land, Labor, and Other
Services in a Sample of Thirty-two
Cross-sectional Cobb- Douglas
Type Production Function
Studies
Statistic Land Services Labor Services Other Services

Frequencies:
< 0 1
1 1
0.00-0.19 7 13 7
0.20-0.39 14 14 7
0.40-0.59 7 3 13
0.60-0.79 2 1 4
0.80-0.99 2
Mean 0.38 0.21 0.39

Standard deviation 0.22 0.17 0.23
Coefficient of
variation 0.58 0.80 0.59
31
and 0.2 for land, labor, and capital, respectively; Palvia has hypothe-
sized an elasticity of 0.5 for labor and a figure of 0.5 for land and capi-
32
tal combined; and Belshaw very tentatively hypothesizes values of
33
0.75 for labor and 0.25 for capital, including land, In contrast, the
mean values listed in Table 17.15 are not too dissimilar from the elas-
ticities of 0.39, 0.28, and 0.33 for land, labor, and other services esti-
mated in a cross-sectional Cobb-Douglas study of world agriculture
34
based on national statistics for the year 1949. Still, the variation
shown among the elasticities of Table 17.14 bears witness to the dan-
gers associated with the use of any such global production function.
INTERNATIONAL COMPARISONS
OF RESOURCE USE EFFICIENCIES
For production under a given technological environment with both

output and input variable, the ideal measure of the efficiency of re-
source use is provided by the marginal return to opportunity cost ra-
tios of the various resources. Concomitantly, these ratios indicate the
direction of the changes that should be made in resource allocation if
profits are to be maximized. Both within and between nations such data
is essential to any thorough analysis of resource allocation and trade,
and to any endeavor to improve the pattern of resource use.
Accordingly, having already considered the intra -national implica-
tions of the available agricultural production function estimates from
around the globe, we now proceed to examine their implications for re-
source use within an international framework. Of course, such an
analysis cannot be complete. We only have estimates from a few scat-
tered locations. In contrast, even assuming the problems of statistical
and economic specification to be negligible, a full analysis would re-
quire estimates covering all the various regions and types of production
found in the world. Only then would it be possible to specify a model
from which the optimal allocation of world resources under the given
distribution of technologies might be deduced for various price regimes.
With insertion of the appropriate restrictions, such a model might take
into account the size of national factor endowments as well as the natu-
ral and institutional fixity of some resources, and the cost of resource
transfers. Even so, such an extensive model would be far from ideal.
By necessity, it could only be based on the production functions cur-
rently being used throughout the world. As we all know, the greatest
gains in efficiency and welfare in many regions of the globe are to be
31
Tinbergen, J. and Polak, J. J. The dynamics of business cycles; a study in economic
fluctuations. Routledge, London. 1950. P. 122.
32 Palvia, C. M. An econometric model for development planning. Institute of Social
Studies, The Hague. 1953.

33
Belshaw, H. Population growth and levels of consumption with special reference to
countries in Asia. Allen and Unwin, Ltd., London. 1956.
34
Bhattacharjee, J. P. Resource use and productivity in world agriculture. Journal of
Farm Economics. 37: 57-71. 1955. See also Econ. Bui. Asia and Far East, 9(1): 17-31. 1958.
made through the introduction of more efficient production functions.

In short, the optimal use of world resources does not depend solely on
shifts in resource use within the context of a given technology. Changes
in technology are just as important, if not more so. Moreover, since
technological shifts are inherently uncertain, their specification in an
empirical model would be no easy task. Too, any analysis of world re-
source allocation that omits to take account of peoples' aspirations,
needs, potentialities, cultural level, and national sentiments must be
less than complete. Hence, granted the magnitude and difficulty of any
full appraisal of world resource allocation, the humble nature of the
material we present in readily apparent. To know that it encompasses
nearly all the available data is but small consolation. Still, even these
few comparisons of resource -use efficiency are of interest. They illu-
minate some important features of current resource allocation in world
agriculture. As well, they introduce a quantitative element into an area
of discussion that is prone to be shrouded in platitudes and ideological
cliches.
Throughout, we assume both output and input are variable. In con-
sequence, provided decreasing economic returns to scale eventually
prevail (as we believe they must), the optimal pattern of resource use
occurs when all the marginal return to opportunity cost ratios are equal
to unity. Should output be regarded as fixed, the best pattern of re-
source use is that at which all the marginal return to opportunity cost
ratios are equal, but not necessarily equal to unity. Such a resource
distribution minimizes the cost of producing the given output and is
easily calculated from the estimated production function. Moreover,
should output be fixed, the discrepancy between the current pattern of
resource allocation and the least -cost pattern provides the appropriate
measure of resource use efficiency. 35 We do not present such data; the
context of our discussion precludes the assumption of fixed output.
The marginal return to opportunity cost ratios estimated for the
services of land, labor, and capital in the various studies are listed in
tables 17.16, 17.17, and 17.18, respectively. In drawing up the tables,
only an average estimate has been included for those countries where a
number of studies are available for a given region, type of farm organ-
ization, and type of product. The capital services category corresponds
to the other services input of the previous section; it is only the same
as the capital services item of the original functions in those cases
where the classical resource categories of land, labor, and capital were
used. Moreover, in deriving the marginal return to opportunity cost
ratios for capital services, the ratios for other services in the original
functions were weighted according to the proportion each of these serv-
ices was of the total input of nonland and nonlabor services. Thus the
capital services ratios indicate the relative discrepancy between the
revenue and cost associated with an input of one more unit of capital
35
See, for example, Miller, W. G. etal
.
m
Relative efficiencies of farm tenure classes
intrafirm resource allocation. Iowa Agr. Exp. Sta. Bui. 461. Ames. 1958. Also Tramel, T.
Using production functions for making recommendations. Jour. Farm Econ., 39; 790-93. 1957.
services composed of the various nonland and nonlabor services in the

proportions currently used.
In assessing the marginal return to opportunity cost ratios, it must
be remembered they are based on the prices prevailing, on the average,
over the twelve months to which each of the production functions refers.
We assume these prices are comparable over the eight years spanned
by the estimates. As well, account needs to be taken of the relative
importance of the production process to which each estimate relates.
In total welfare terms, a small degree of inefficiency for each of the
members of a large population of entrepreneurs may be of more import
than the existence of severe disequilibrium in a small population. An-
other restrictive feature of the data is that it only relates to resource
changes of an incremental nature. Although the marginal return to op-
portunity cost ratios tell us the relative effect of using one more unit of
each resource, they provide only a tentative indication of the possible
influence of large scale changes in resource use. However, such infor-
mation may be derived from the estimated production functions and is
already available in some of the original sources cited for the various
studies. We do not present such data here. Given the small sample of
estimates with which we are concerned, the additional information
would be of little worth beyond that given by the resource use efficiency
ratios at the current margin of resource use, especially since what is
needed in many of the less developed areas of the globe is not the at-
tainment of optimality under current methods of production but the in-
troduction of new technologies. In many instances current methods,
even if used in profit maximizing fashion, are quite inadequate to sus-
tain a reasonable level of welfare.
The marginal return to opportunity cost ratios for land services are
presented in Table 17.16. Of the 33 estimates, 10 imply the use of too
much land while 23 indicate additional land could have been used profit-
ably. Over -all, the ratios range from 31.39 for sweet potato cropping
in Japan to 0.05 for mixed crop farms in Andhra Pradesh, India. The
general tendency appears to be for farmers to use too little land. In
fact, half of the estimates are greater than 2.00, indicating quite severe
deficiencies in land input. Moreover, a chi square test indicates there
is less than one chance in twenty that the distribution of the ratios
between those above and below unity is due to chance. Such a result is
as would be expected, given the relative immobility of land and the
pressure upon its use as populations increase. Apart from the ex-
tremely severe disequilibria in Honshu cropping, caused by the Japa-
nese Land Reform Law, and the prevalence of too many land services
in the Austrian, Taiwan, Australian sheep, and Indian Andhra Pradesh
studies, the data exhibits no pronounced pattern. Differences within
countries in the efficiency of land use appear to be just as great as
those between nations. Still, it might be noted that all of the estimates
for Japanese cropping, Australian dairying, Norway, New Zealand,
Sweden, and South Africa imply the use of too few land services. As
well, only one of the seven United States studies indicates an excessive
Table 17.16. Marginal Return to Opportunity Cost Ratios of Land

Services as Estimated From Selected Cross-
sectional Production Function Studies
of Farm-firms
Marginal Return
Function to Opportunity
Location of Sample for Cost Ratio of
Land Services
Japan, Honshu sweet potatoes 31.39

Norway, southeast cereals 23.66
Japan, Honshu tea 19.37
Japan, Honshu rice 14.27
Norway, southeast fodder 11.93
United States, Alabama crops 4.01
India, Uttar Pradesh planted sugar cane 3.60
United States, southern Iowa corn 3.27
United States, northern Iowa corn 3.08
Japan, Hokkaido rice 2.95
United States, Montana wheat 2.81
Canada, Alberta wheat, beef 2.58
New Zealand, Canterbury Plain sheep 2.48
Japan, Hokkaido, small farms rice 2.32
India, Uttar Pradesh irrigated wheat 2.22
South Australia dairy 2.12
South Africa, Kalihari cattle fattening 1.95
United States, Iowa- Illinois livestock- share 1.93
United States, Iowa- Illinois crop- share 1.78
Sweden mixed farms 1.78
Japan, Hokkaido, large farms rice 1.78
Australia, New South Wales dairy 1.23
South Africa, Kalihari cow- calf ranches 1.05
United States, Iowa- Illinois full owners 0.98
India, Uttar Pradesh wheat, sugar cane 0.95
Austria mixed farming 0.92
Canada, Alberta cattle ranches 0.88
India, Andhra Pradesh irrigated farms 0.79
Taiwan, Tainan cereals 0.58
India, Andhra Pradesh dry farms 0.50
Australia, New South Wales sheep 0.39
Taiwan, Tainan sugar cane 0.39
India, Andhra Pradesh mixed farms 0.05
input of land, and it has a ratio of 0.98 implying near optimality.

The
other six United States estimates range from 1.78 to 4.01 so that the
use of too few land services per farm also appears to be a common fea-
ture of United States agriculture.
In terms of type of farming, 19 of the 33 ratios for land refer
to
cropping alone and 14 to production involving livestock, generally
under
grazing conditions. Nine of the 14 livestock estimates indicate the
use
of too little land. Of the 19 crop estimates, 14 imply
an insufficient in-
put of land and there is less than one chance in 20 that this
division is
due to sampling variation. Indeed, 11 of these 14 ratios are
larger than
Table 17.17. Marginal Return to Opportunity Cost Ratios of Labor

Services as Estimated From Selected Cross-
sectional Production Function Studies
of Farm-firms
Marginal Return
Location of Sample for Cost Ratio of
Labor Services

India, Uttar Pradesh planted sugar cane 1.78
Norway, southeast beef cattle 1.64
Israel mixed farms 1.26
United Kingdom, England dairy 1.06
Western Australia dairy 0.90

Norway, southeast dairy 0.73
United States, Alabama livestock 0.67
United States, southern Iowa hogs, cattle 0.62
United States, northern Iowa hogs, cattle 0.62
Australia, New South Wales dairy 0.49
United States, Montana cattle 0.32
Japan, Hokkaido, large farms rice 0.21
India, Andhra Pradesh irrigated farms 0.13
India, Uttar Pradesh irrigated wheat -0.35
Sweden dairy <0.00
the highest ratio for land in livestock or mixed crop -livestock produc-
tion. Thus the data suggest crop production is more prone to disequi-
librium in the use of land than livestock production, and that cropping
is most frequently carried out with suboptimal inputs of land services.
Justification for these tendencies is found in the inelastic nature of the
supply of land for cropping relative to grazing and the greater intensity
of land use in crop production. As the intensity of land use rises, small
variations in input have relatively greater effects on output so that the
attainment of optimality is a more difficult task in crop production than
it is in more extensive activities involving grazing livestock. More-
over, these influences are accentuated by the quantitative and locational

fixity of the land resource.
Except in terms of ownership changes, land services are immobile.
However, changes in ownership between nations are unlikely under nor-
mal circumstances. Accordingly, we will not discuss the data of Table
17.16 in terms of land transfers. To do so would be unreal.
For the labor resource, Table 17.17 lists marginal return to oppor-
tunity cost ratios as estimated from 43 analyses. Of these, 27 (63 per
cent) imply the use of too much labor. Although this distribution is not
significant at usual probability levels, it corresponds with the theoreti-
cal expectation of surplus labor in agriculture arising from frictions in
the adjustment to economic progress. 36 But compared to the efficiency
ratios for land, which range from 0.05 to 31.39 with a mean of 4.5, the
labor ratios display a much smaller degree of disequilibria, ranging
from -0.35 to 2.84 with a mean of approximately 0.9. 37 The reason for
greater efficiency in the use of labor is probably the relative mobility
and freedom of supply of labor within nations.
As for land services, no pronounced ordering is discernible between
countries in terms of their efficiency of labor use, except for the
grouping of the Indian cereal studies at the bottom of the table. Still,
the excessive use of labor on these Indian peasant farms is only slightly
worse than prevails in commercial corn and wheat production in the
United States samples for Montana and southern Iowa. More detailed
perusal of the data shows too much labor was invariably used in the
samples for Australian dairying, the Japanese island of Hokkaido, the
Indian state of Andhra Pradesh, New Zealand, Austria, Sweden, and the
United States states of Iowa, Montana, and Alabama. In contrast, all the
analyses for Taiwan, South Africa, Israel, Canada, and the United King-
dom imply additional labor inputs would be profitable. Provided the
costs of transfer did not exceed the benefits, transfer of labor from the
former to the latter group of countries would be worthwhile. Neverthe-
less, it is probable that more feasible opportunities for beneficial
36
See Ojala, E. M. Agriculture and economic progress. Oxford University Press, New
York. 1952; also Fisher, A. G. B. Economic progress and social security. Mac millan,
New York. 1946.
37
No great significance should be attached to the average values presented. If the rele-
vant information were available it would be preferable to consider the means of the esti-
mates weighted according to the size of the farm population to which they refer.
Table 17.18. Marginal Return Opportunity Cost Ratios of

to
Capital Services as Estimated From Selected
Cross-sectional Production Function
Studies of Farm-firms
Marginal Return
for Cost Ratio of
Location of Sample
Capital Services
India, Uttar Pradesh irrigated wheat 6.97

Western Australia dairy 1.61
Japan, Hokkaido, Large farms rice 1.54
Sweden dairy 1.41
Australia, NewSouth Wales dairy 1.40
United Kingdom, England dairy 1.21
United States, Montana cattle 1.14
United Kingdom, Bristol
Province low-yield dairy 1.09
Israel mixed farms 1.09
Canada, Alberta cattle feedlots 1.05

United States, Alabama livestock 0.94
Norway, southeast dairy 0.76
United Kingdom, Bristol
Province high-yield dairy 0.73
Norway, southeast beef cattle 0.66
India, Uttar Pradesh planted sugar cane -0.11
India, Andhra Pradesh irrigated farms -0.85
transfers of labor exist within these various countries. Obviously,

such domestic adjustments should normally precede international
shifts.
Nineteen of the 43 labor -efficiency estimates of Table 17.17 refer
to cropping alone and 24 to mixed crop -live stock production. Of the 19
crop estimates, slightly more than two-thirds are greater than unity,
indicating the use of too much labor. The corresponding proportion is
not quite so great for production involving livestock. If anything, there
may be a greater tendency for a surfeit of labor in crop than in live -
stock production. Certainly such a tendency would be expected for the
world as a whole, given the prevalence of subsistence agriculture which
is necessarily based on cropping.
Table 17.18 lists the resource use efficiency ratios for capital
services as estimated from 46 studies. The general tendency is
towards the use of insufficient capital. Thirty (65 per cent) of the esti-
mates are greater than unity while their average value is 1.2. More-
over, assuming the 46 studies to be a random sample from the super
population of all possible studies, there is only 5 per cent probability
that this distribution is due to chance. The range of capital service
ratios, from -0.85 to 6.97, is slightly larger than that for the labor es-
timates, but still far less than the range for land services. Relative to
land, the mobility of capital probably explains its smaller degree of in-
efficiency. In terms of type of production, 19 of the estimated ratios
refer to cropping and 27 to production involving livestock. Among the
latter, 71 per cent imply the use of too few capital services. In con-
trast, the 19 crop estimates divide rather evenly between excessive and
deficient capital services. The data strongly suggest, therefore, that
the over -all tendency to the use of too few capital services is mainly
due to the lack of sufficient capital in production involving livestock.
While all the New Zealand, Australian, South African, Austrian, Ca-
nadian, and Israeli studies indicate that more capital services should be
used, for no country do the various estimates invariably imply that less
capital should be used. Hence, even from the few scattered estimates
we have been able to assemble, it is apparent that intra -national
chances for capital transfers abound and that international shifts in
capital can only be made in the face of favorable domestic opportuni-
ties. K the unreal assumption is made that prevailing technologies are
immutable, it is obvious that the data presents no strong argument for
allocating additional capital to the underdeveloped countries rather
than to the more advanced nations. In order of priority relative to the
benefits to be derived from the use of additional capital, the various
studies for each country are dispersed throughout the listing of Table
17.18.
So far we have been considering land, labor, and capital services
separately. Such an approach is not entirely adequate. Opportunities
exist for substitution between land, labor, and capital; changes in the
input of one influence the efficiency of use of the other resources. Ac-
cordingly, we need to consider the over -all picture of resource use
Table 17.19. Frequency Distributions, Means, Standard Deviations

and Coefficients of Variation of the Marginal
Return to Opportunity Cost Ratios for Land,
Labor, and Capital Services as Estimated
From Selected Cross-sectional
of Farm-firms
Statistic Land Services Labor Services Capital Services
Frequencies:
<0 2 4
0.01-0.20 1 2 1
0.21-0.40 2 10 2
0.41-0.60 2 4 1
0.61-0.80 1 6 5
0.81-1.00 4 3 3
1.01-1.20 1 1 10
1.21-1.40 1 6 5
1.41-1.60 2 5
1.61-1.80 3 3 3
1.81-2.00 2 1 2
2.01-3.00 7 3 4
>3.01 9 1
Mean 4.5 0.9 1.2
Standard deviation 7.3 0.7 0.7
Coefficient of
variation 1.62 0.78 0.58
efficiency. The broad picture is given by Table 17.19 which lists the
frequency distributions, means, standard deviations, and coefficients of
variation of the marginal return to opportunity cost ratios for the vari-
ous resources. The wide dispersal of the ratios is readily apparent
from these statistics. As expected from the pervasive influence of
frictions in the adjustment to economic progress and population growth,
the general tendency is for farms to have too few land and capital serv-
ices, and a surfeit of labor services. Differences in the coefficients of
variation of the land, labor, and capital ratios are noteworthy: the var-
iability for capital is slightly less than for labor, while the ratios for
land are more than twice as variable as those for capital and labor.
Also, the relatively flat distribution for land contrasts with the peaked
distributions for labor and capital services. Too, while nearly half of
the land estimates are greater than 2.00, half of the labor and capital
ratios lie, respectively, in the ranges 0.21 to 0.80 and 1.01 to 1.60. As
already noted, the tendency for more efficient use of labor and capital
services is probably due to their relative mobility compared to land.
From the statistics of Table 17.19, the broadest recommendation
that might be made for world agriculture is that farmers should substi-
tute land and capital services for labor. Such a proposal, of course,
implies a net inflow of capital into agriculture (perhaps obtainable via a
reduction in the nonproductive portion of the flow of capital out of
agriculture), as well as a reduction in the number of farms since the

supply of land is highly inelastic. Concomitantly, nonfarm employment
opportunities would need to be provided for labor transferred out of ag-
riculture. Still, little weight can be attached to such a broad recom-
mendation which, we might note, could have been made on a priori
grounds anyway. Indeed, the whole rationale of the data we have pre-
sented is that, if available in sufficient quantity, it would enable far
more specific recommendations to be made.
Nevertheless, it is apparent that no simple ordering of priorities
could be made on a country by country basis. The marginal revenue to
opportunity cost ratios of the various studies within each country are
too dissimilar. For instance, for all resources examined, the United
States studies have resource use efficiency ratios that are scattered
widely over the range of the marginal return to opportunity cost ratios.
The same statement applies to the various studies from the less ad-
vanced countries. Further evidence of this intr a -national diversity is
provided by a categorization of the studies among the eight possible
combinations of too many or too few inputs of land, labor, and capital
services. These eight combinations may be paired in terms of con-
verses. For example, using capital letters to denote too few inputs and
small letters excessive inputs of the relevant resources, the combina-
tion (LAND, LABOR, capital) is the converse of (land, labor, CAPITAL).
In such form, the studies involving all three inputs are distributed as
follows:
LAND, LABOR, CAPITAL land, labor, capital
Canada, mixed farms India, Andhra Pradesh, dry farms

Israel India, Andhra Pradesh, irrigated farms
Japan, Honshu, tea India, Andhra Pradesh, mixed farms
Norway, fodder
South Africa, cattle fattening
South Africa, cow-calf
United States, livestock-share
United States, crop -share
LAND, LABOR, capital land, labor, CAPITAL

India, Uttar Pradesh, Austria
sugar cane India, Uttar Pradesh, wheat, sugar cane
United States, full owners
LAND, labor, capital land, LABOR, CAPITAL

Japan, Hokkaido, rice Australia, sheep
Norway, cereals Canada, cattle ranches
Sweden, mixed farms Taiwan, sugar cane
United States, northern
Iowa, corn
LAND, labor, CAPITAL land, LABOR, capital
Australia, New South Taiwan, cereals

Wales, dairy
Australia, South
Australia, dairy
India, Uttar Pradesh, wheat
Japan, Hokkaido, large farms
Japan, Honshu, rice
New Zealand, sheep
United States, Alabama, crops
United States, Montana, wheat
United States, southern
Iowa, corn
Such an arrangement of the studies brings out the predominance of

deficiencies in the input of land and capital services. As well, it again
emphasizes the existence of similar patterns of inefficiency in both the
advanced and less advanced nations. Hence, if the technological en-
vironment is assumed fixed, there is no economic justification for a
blanket policy of capital transfers from the more advanced to the less
advanced countries. These transfers could only be justified on humani-
tarian grounds. On the other hand, such transfers are justifiable in
terms of positive economics when technological change is admitted.
However, the estimated marginal return to opportunity cost ratios tell
us nothing of the situation that would exist over the globe if the most
advanced methods of production were used in a satisfactory manner.
Still, provided that necessary ancillary adjustments were made to these
newer technologies, they might amend much of the poverty in the less
developed areas of the world. Accordingly, accepting current scale of
enterprise, net income, and return on investment as guides to the need
for new technologies and using the more advanced countries as bench-
marks, priorities could be attached to the need for using more advanced
technologies relative to the gains that might be made from improving
the efficiency of resource use within the context of the current techno-
logical pattern. Certainly, the best approach to improving the efficiency
of world agriculture must involve both methods. Too, investments in
these activities would probably be most beneficial if carried out simul-
taneously in a number of nations.
Finally, the data of tables 17.16 and 17.17 strongly evidence dis-
equilibria in the use of land and labor that are just as great as the inef-
ficiencies found for capital services. Moreover, land and labor are
virtually immobile in an international context. It can be concluded
intr a -national resource transfers must play the dominant role in any
initial shift towards greater efficiency in world agriculture. Still, capi-
tal service transfers between countries should not be judged unimpor-
tant, especially in relation to the implementation of more efficient pro-
duction processes.
Bibliography
1. Agrawal, G. D. Studies in economics of farm management in

Uttar Pradesh, 1954-55. Directorate of Economics and Statistics,
Ministry of Food and Agriculture, New Delhi, India. March, 1957.
2. and Foreman, W. J. Farm resource productivity

in West U. P. Indian Jour. Agr. Econ. 14:115-33. 1959.
3. Allen, Carl Wendell. Substitution relationships between forage

and grain in milk production. Unpublished Ph.D. thesis. Iowa
State University Library, Ames, Iowa. 1954.
4. Allen, R. G. D. Mathematical analysis for economists. Macmillan
and Co., Ltd., London. 1938.
5. Anderson, R. L. A comparison of discrete and continuous models

in agricultural production analysis. In Baum, E. L., Heady, Earl
O., and Blackmore, John, eds. Methodological procedures in the
economic analysis of fertilizer use data. pp. 39-60. Iowa State
University Press, Ames, Iowa. 1956.
6. . The problem of autocorrelation in regression

analysis. Jour. Amer. Stat. Assn. 49:113-29. 1954.
7. . Proper planning reduces research errors. Jour.
Farm Econ. 35:572-81. 1953.
8. Some statistical problems in the analysis of

.
fertilizerresponse data. In Baum, E. L., Heady, Earl O., Pesek,

John T., and Hildreth, Clifford G., eds. Economic and technical
analysis of fertilizer innovations and resource use. pp. 187-206.
Iowa State University Press, Ames, Iowa. 1957.
9. Antill, A. G. Towards a production function for dairy farms. The
Farm Economist 8:1-11. 1955.
10. and Clark, C. The overfeeding of dairy cows.

Westminster Bank Rev. pp. 13-15. February, 1958.
11. Aukrust, O. Investment and economic growth. Productivity Meas.
Rev. No. 16, pp. 35-53. February, 1959.
12. Baird, Bruce L., and Fitts, J. W. An agronomic procedure involv-
ing the use of a central composite design for determining fertilizer
response surfaces. In Baum, E. L., Heady, Earl O., Pesek, John
T., and Hildreth, Clifford G., eds. Economic and technical analysis
of fertilizer innovations and resource use. pp. 135-43. Iowa State
645
646 BIBLIOGRAPHY
13. Balmukand, B. H. Studies in crop variation. V. The relation

between yield and soil nutrients. Indian Jour. Agr. Sci. 18:602-27.
1928.
14. Baule, B. Zu Mitscherlich’s Gesets der physiologischen Beziechun-

gen. Landw. Jahrb. 51:363-85. 1937.
15. Baum, E. L., Heady, Earl O., and Blackmore, John, eds. Methodo-
logical procedures in the economic analysis of fertilizer use data.
16. and Walkup, H. G. Some economic implications
of input-output relationships in fryer production. Jour. Farm
Econ. 35:223-36. 1953.
17. Beach, C. L. The facility of digestion of foods as a factor in

feeding. Conn. Agr. Exp. Sta. Bui. 43. 1906.
18. Bell, Robert Daniel. Methodological problems and possibilities

in farm business analysis using inter-farm production functions.
Unpublished Ph.D. thesis. Cornell Univ., Ithaca, N. Y. 1959.
19. Bhatia, R. J. The production function for Indian manufactures,

1948. Journal of Bombay University, India. Jan. 1954.
20. Bhattacharjee, Jyoti P. Resource use and productivity in world

agriculture. Jour. Farm Econ. 37:57-72. 1955.
21. Black, J. A formal proof of the concavity of the production pos-

sibility function. Econ. Jour. 67:133-35. 1957.
22. Black, John D. Dr. Schultz on farm management research. Jour.

Farm Econ. 22:570-80. 1940.
23. Blaser, R. E., Bryant, H. T., Ward, C. Y., Hammes, R. C., Jr.,
Carter, R. C., and MacLeod, N. H. Animal performance and
yields with methods of utilizing pasturage. Agron. Jour. 51:238-
42. 1959.
24. Blaxter, K. L. Energy feeding standards for dairy cattle. Nutr.

Abst. and Rev. 20:1-18. 1950.
25. . The nutritive values of feeds as sources of

energy: a review. Jour. Dairy Sci. 39:1396. 1956.
26. and Wainman, F. W. Some observations on the

digestibility of food by sheep and on related problems. British
Jour. Nutr. 10:69-91. 1956.
27. Bloom, Solomon. Effects of various dietary hay- concentrate

ratios on nutrient utilization and production response of dairy
cows. Unpublished Ph.D. thesis. Iowa State University Library,
Ames, Iowa. 1955.
28. Boldyreff, John W. The law of diminishing fertility of the soil,
BIBLIOGRAPHY 647
from the point of view of some of the Russian economists of today.

Jour. Farm Econ. 13:470-85. 1931.
29. Bondorff, K. A. Det kvantitative Forhold mellem planternes

Earnaering og Stofproduktion. Kgl. Veterinaer-og Landbohojskole
Aarsskrift. Kandrys and Wunsch. Copenhagen, Denmark.
pp. 293-336. 1924.
30. Boresch, K. Uber Ertragsgesetze bei Pflanzen. Ergebnisse der

Biologie 4:130-204. 1928.
31. Borts, George H. Production relations in the railway industry.

Econometrica 20:71-80. 1952.
32. Bose, S. R. Measurement of productivity in Indian industries.
Indian Jour. Econ. 33:271-82. 1953.
33. Boule, F. Liber die Weiternicklung. Zeitsch. Acker Pflanzenban

94:173-86. 1947.
34. Box, G. E. P. The exploration and exploitation of response sur-

faces: some general considerations and examples. Biometrics
10:16-60. 1954.
35. and Hunter, J. S. Experimental designs for the

exploration and exploitation of response surfaces. In Chew, V.,
ed. Experimental designs in industry, pp. 138-90. John Wiley
and Sons, Inc., New York. 1958.
36. and Multi-factor experimental

.
designs for exploring response surfaces. Ann. Math. Stat.

28:195-241. 1957.
37. and Wilson, K. B. On the experimental attainment

of optimum conditions. Jour. Roy. Stat. Soc. Series B 13:1-45.
1951.
38. Briggs, G. E. Plant yield and the intensity of external factors:

Mitscherlich’s “wirkungsgesetz.” Ann. Bot. 39:475-502. 1925.
39. Briggs, P. K., et al The performance of adult merino wethers

.
fed weekly on all-grain rations. Australian Vet. Jour. 32:299-304.

1956.
40. Bronfenbrenner, M. Production functions: Cobb-Douglas, inter-

firm, intrafirm. Econometrica 12:35-44. 1944.
41. Brown, David W. Adjustment of value productivity estimates to

changes in price and technical relationships. Unpublished Ph.D.
thesis. Iowa State University Library, Ames, Iowa. 1956.
42. Brown, E. H. Phelps. The meaning of the fitted Cobb-Douglas

function.Quar. Jour. Econ. 71:546-60. 1957.
43. Brown, William G. Free choice versus least-cost mixed rations

for hogs. Jour. Farm Econ. 38:863-68. 1956.
648 BIBLIOGRAPHY
44. Brown, William G., and Arscott, George H. A method for dealing
with time in determining optimum factor inputs. Jour. Farm Econ.
40:666-73. 1958.
45. and Oveson, Merrill M. Production functions

from data over a series of years. Jour. Farm Econ. 40:451-57.
1958.
46. Cannell, C. F. and Kahn, R. L. Collection of data by interviewing.

In Festinger, L. and Katz, D., eds. Research methods in the
social sciences, pp. 327-80. Dryden Press, New York. 1953.
47. Carlson, Sune. A study on the pure theory of production. P. S.
King and Son, London. 1939.
48. Carter, H. O. and Hartley, H. O. A variance formula for marginal
productivity estimates using the Cobb-Douglas function. Economet-
rica 26:306-13. 1958.
49. Case, H. C. M. and Williams, D. B. Fifty years of farm manage-

ment. University of Illinois Press, Urbana, Illinois. 1957.
50. Catron, D. V., et at. Re-evaluation of protein requirements.

Anim. Sci. 2:221-32. 1952.
51. Chenery, Hollis B. Engineering production functions. Quart. Jour.

Econ. 63:507-37. 1949.
52. Clark, J. and Bessell, J. E. Profits from dairy farming. Imperial
Chemical Industries, Ltd., Central Agr. Control, London. Bui. 7.
1956.
53. Clarke, J. W. An analysis of the application of the production

function to a sample of farms in southern Saskatchewan. Unpub-
lished M.S. thesis. University of Saskatchewan, Canada. 1950.
54. Cobb, Charles W. Production in Massachusetts manufacturing,

1890-1928. Jour. Pol. Econ. 38:705-7. 1930.
55. and Douglas, Paul H. A theory of production.

Amer. Econ. Rev. 18:139-56, Supplement. 1928.
56. Cooper, Gershon. The role of econometric models in economic

research. Jour. Farm Econ. 30:101-17. 1948.
57. Crowther, E. W., and Yates, F. Fertilizer policy in wartime:

requirements of arable crops. Empire Jour. Exp. Agr.
fertilizer
9:77-97. 1941.
58. Darcovich, W. The use of production functions in the study of

resource productivity in some beef producing areas of Alberta.
Econ. Annalist 28:85-93. 1958.
59. Marginal productivities of resources in two farm
Dillon, J. L.
areas N.S.W. Economic Society of Australia and New Zealand.
of
Economic Monograph No. 188. 1956.
BIBLIOGRAPHY 649
60. and Heady, E. O. Decision criteria for innovation.

Australian Jour. Agr. Econ. 2:113-20. 1958.
61. Doll, John Philip. Evaluation of alternative algebraic forms for

production functions. Unpublished Ph.D. thesis. Iowa State Uni-
versity Library, Ames, Iowa. 1958.
62. Douglas, Paul H. The theory of wages. The Macmillan Co., New
York. 1934.
63. and Bronfenbrenner, M. Cross-section studies

in the Cobb-Douglas function. Jour. Pol. Econ. 47:761-85. 1939.
64. and Daly, Patricia. The production function for
Canadian manufactures. Jour. Amer. Stat. Assn. 39:178-86. 1943.
65. and Olson, Ernest. The production function for

manufacturing in the United States, 1904. Jour. Pol. Econ. 51:61-
65. 1943.
66. and Gunn, Grace T. Further measurement of

marginal productivity. Quart. Jour. Econ. 54:399-428. 1940.
67. and The production function for

.
American manufacturing for 1914. Jour. Pol. Econ. 50:595-602.

1942.
68 . and _. The production function for

American manufacturing in 1919. Amer. Econ. Rev. 31:67-80.
1941.
69. and The production function for

.
Australian manufacturing. Quart. Jour. Econ. 56:108-29. 1941.
70. and Handsaker, Marjorie L. The theory of

marginal productivity tested by data for manufacturing in Victoria.
Quart. Jour. Econ. 52:214-54. 1937.
71. Durand, David. Some thoughts on marginal productivity with spe-

cial reference to Professor Douglas’ analysis. Jour. Pol. Econ.
45:740-58. 1937.
72. Du Toit, Schalk J. v. N. Analysis of cattle ranching in the eastern

Kalahari, Union of South Africa. Unpublished M.S. thesis. Iowa
State University Library, Ames, Iowa. 1953.
73. Dutt, M. M. The production function for Indian manufactures.

Sankhya. Vol. 15, part 4. 1955.
74. Edelberg, Victor. An econometric model of production and dis-

tribution. Econometrica 4:210-25. 1936.
75. Fellows, Irving F. Developing and applying production functions

in farm management. Jour. Farm Econ. 31:1058-64. 1949.
76. .
Production functions in farm management. In
650 BIBLIOGRAPHY
Haler ow, Harold G., ed. Contemporary readings in agricultural

economics, pp. 74-81. Prentice-Hall, Inc., New York. 1955.
77. Felsenthal, Leonard. Studies in the Cobb-Douglas production
function for mining and manufacturing in Germany, 1925-36. Un-
published M.A. thesis. University of Chicago, Chicago, Illinois.
1940.
78. Finney, D. J. Response curve and the planning of experiments.

Indian Jour. Agr. Sci. 23:167-86. 1953.
79. Fisher, A. G. B. Economic progress and social security.

Macmillan and Co., New York. 1946.
80. Fitts, J. W., Mason, D. D., and Cooper, Dale. Determining yield
response surfaces and economically optimum fertilizer rates for
corn under various soil and climatic conditions in North Carolina.
Conf. for Coop’s, in the TVA. Agr. Econ. Res. Act’s. March
24-26, 1959.
81. French, Burton Leroy. Estimation by simultaneous equations of

resource productivities from time series and cross sectional
farm observations. Unpublished Ph.D. thesis. Iowa State Univer-
sity Library, Ames, Iowa. 1952.
82. . Functional relationships for irrigated corn re-
sponse to nitrogen. Jour. Farm Econ. 38:736-47. 1956.
83. Simultaneous economic relationships and deriva-
.
tion of the production function.In Heady, Earl O., Johnson, Glenn

L., and Hardin, Lowell S., eds. Resource productivity, returns to
scale, and farm size. pp. 97-105. Iowa State University Press,
Ames, Iowa. 1956.
84. Gilson, J. C. and Bjarnarson, V. W. Effects of fertilizer use on

barley in northern Manitoba. Jour. Farm Econ. 40:932-41. 1958.
85. Graves, R. R., Dawson, J. R., Kopland, D. V., Watt, A. L., and Van
Horn, A. G. Feeding dairy cows on alfalfa hay alone. U. S. Dept.
Agr. Tech. Bui. 610. 1938.
86. Griliches, Zvi. Specification bias in estimates of production

functions. Jour. Farm Econ. 39:8-20. 1957.
87. Hader, R. J., Harward, M. E., Mason, D.

D., and Moore, D. P. An
some of the relationships between copper, iron
investigation of
and molybdenum in the growth and nutrition of lettuce: I. Experi-
mental design and statistical methods for characterizing the re-
sponse surface. Soil Sci. Soc. Amer. Proc. 21:59-64. 1957.
88. Hansen, Peter L. Input- output relationships in egg production.
Jour. Farm Econ. 31:687-97. 1949.
89. Harris, H. The development and use of production functions for

firms in agriculture. Scientific Agr. 27:487-95. 1947.
BIBLIOGRAPHY 651
90. Hart, B. J. Significance levels for the ratio of the mean square
successive difference to the variance. Ann. Math. Stat. 13:445-47.
1942.
91. Haver, Cecil B. Economic interpretations of production function

estimates. In Heady, Earl O., Johnson, Glenn L., and Hardin,
Lowell S., eds. Resource productivity, returns to scale, and
farm size. pp. 146-50. Iowa State University Press, Ames,
Iowa. 1956.
92. Headley, F. B. Feeding experiment with dairy cows. Nev. Agr.

Exp. Sta. Bui. 119. 1930.
93. Heady, Earl O. Application of recent economic theory in agricul-

tural production economics. Jour. Farm Econ. Proc. 32:1125-39.
1950.
94. Choice of functions in estimating input-output

.
relationships. Proceedings of 51st Annual Meeting of the Agricul-

tural Economics and Rural Sociology Section of the Association of
Southern Agricultural Workers. 1954.
95. . An econometric investigation of the technology

of agricultural production functions. Econometrica 25:249-68.
1957.
96. . Economic concepts in directing and designing

research for programming use of range resources. Jour. Farm
Econ. 38:1604-16. 1956.
97. . Economics of agricultural production and resource

use. Prentice-Hall, Inc., New York. 1952. Ch. 2-6.
98. . Elementary models in farm production economics

research. Jour. Farm Econ. 30:201-26. 1948.
99. Integration of physical sciences and agricultural

.
economics. Canadian Jour. Agr. Econ. 4:1-15. 1956.
100. . Marginal productivity of resources and imputa-

tion of shares for cash and share rented farms. Iowa Agr. Exp.
Sta. Res. Bui. 433. 1955.
101. Methodological problems in fertilizer use. In

.
Baum, E. Earl O., and Blackmore, John, eds. Methodo-

L., Heady,
logical procedures in the economic analysis of fertilizer use data,
pp. 3-21. Iowa State University Press, Ames, Iowa. 1956.
102. . Organization activities and criteria in obtaining

and fitting technical production functions. Jour. Farm Econ.
39:360-69. 1957.
103. . Output in relation to input for the agricultural

industry. Jour. Farm Econ. 40:393-406. 1958.
652 BIBLIOGRAPHY
104. Heady, Earl O. Problems in designing dairy feeding experiments
for economic analysis. In Hoglund, C. R., Johnson, Glenn L.,
Lassiter, Charles A., and McGilliard, Lon D., eds. Nutrititional
and economic aspects of feed utilization by dairy cows. pp. 193-
205. Iowa State University Press, Ames, Iowa. 1959.
105. A production function and marginal rates of sub-

.
stitution in the utilization of feed resources by dairy cows. Jour.

Farm E con. 33:485-95. 1951.
106. .
Production functions from a random sample of
farms. Jour. Farm Econ. 28:989-1004. 1946.
107. Productivity and income of labor and capital on
.
Marshall silt loam farms in relation to conservation farming.

Iowa Agr. Exp. Sta. Res. Bui. 401. 1953.
108. Relationship of scale analysis to productivity
.
analysis. In Heady, Earl O., Johnson, Glenn L., and Hardin,

farm size. pp. 82-89. Iowa State University Press, Ames, Iowa.
1956.
109. _. Resource and revenue relationships in agricul-

tural production control programs. Rev. Econ. and Stat. 33:228-
40. 1951.
Resource productivity and returns on 160-acre

farms in north central Iowa. (A production function study of
marginal returns on farms with fixed plants.) Iowa Agr. Exp.
Sta. Res. Bui. 412. 1954.
Technical considerations in estimating production

•
functions. In Heady, Earl O., Johnson, Glenn L., and Hardin,

farm size. pp. 3-15. Iowa State University Press, Ames, Iowa
1956.
Use and estimation on input-output relationships

*
or productivity coefficients. Jour. Farm Econ. Proc. 34*775-86

1952.
anc* Baker, C. B. Resource adjustments to equate

productivities in agriculture. Southern Econ. Jour. 21:36-52
1954.
Balloun, Stanley and Dean, Gerald W. Least-cost

>
rations and optimum marketing weights

for turkeys. Iowa Agr.
Exp. Sta. Res. Bui. 443. 1956.
— >
and McAlexander, Robert. Least-
cost rations and optimum marketing weights for broilers. Iowa
Agr. Exp. Sta. Res. Bui. 442. 1956.
BIBLIOGRAPHY 653
116. ,
Brown, W. G., Pesek, John T., and Stritzel, Joseph.
Production functions, isoquants, isoclines and economic optima in
corn fertilization for experiments with two and three variable
nutrients. Iowa Agr. Exp. Sta. Res. Bui. 441. 1956.
117. and Candler, Wilfred V. Linear programming

methods. Iowa State University Press, Ames, Iowa. 1958.
118. Catron, Damon V., McKee, Dean E., Ashton,

,
Gordon, and Speer, Vaughn. New procedures in estimating feed

substitution rates and in determining economic efficiency in pork
production. II. Replacement rates for growing-fattening swine
on pasture. Iowa Agr. Exp. Sta. Res. Bui. 462. 1958.
119. Doll, John P., and Pesek, John T. Fertilizer

,
production functions for corn and oats; including analysis of irri-

gated and residual return. Iowa Agr. Exp. Sta. Res. Bui. 463.
1958.
120. and Du Toit, Schalk. Marginal resource produc-

areas of South Africa and the
tivity for agriculture in selected
United States. Jour. Pol. Econ. 62:494-505. 1954.
121. and Olson, R. O. Marginal rates of substitution

and uncertainty in the utilization of feed resources with particular
emphasis on forage crops. Iowa State Jour. Sci. 26:49-70. 1951.
122. and Substitution relationships,
.
resource requirements and income variability in the utilization

of forage crops. Iowa Agr. Exp. Sta. Res. Bui. 390. 1952.
123. and Pesek, John. A fertilizer production surface

with specification of economic optima for corn grown on calcareous
Ida silt loam. Jour. Farm Econ. 36:466-82. 1954.
124. and Some methodological consid-

.
erations in the Iowa-TVA research project on economics of ferti-

lizer use. In Baum, E. L., Heady, Earl O., Pesek, John T., and
Hildreth, Clifford G., eds. Economic and technical analysis of
fertilizer innovations and resource use. pp. 144-67. Iowa State
125. ,
and Brown, William G. Crop re-
sponse surfaces and economic optima in fertilizer use. Iowa Agr.
Exp. Sta. Res. Bui. 424. 1955.
126. Schnittker, John, Bloom, Solomon, and Jacobson,

,
Norman L. Isoquants, isoclines and economic predictions in dairy

production. Jour. Farm Econ. 38:763-79. 1956.
127. ,
Jacobson, N. L., and Bloom, Solo-
,
mon. Milk production functions, hay/ grain substitution rates and

economic optima in dairy cow rations Iowa Agr. Exp. Sta. Res.
.
Bui. 444. 1956.

654 BIBLIOGRAPHY
128. Heady, Earl O., and Shaw, R. Resource returns and productivity
coefficients in selected farming areas. Jour. Farm Econ. 36:243-
57. 1954.
129. and Resource returns and produc-

.
farming areas of Iowa, Montana, and

tivity coefficients in selected
Alabama. Iowa Agr. Exp. Sta. Res. Bui. 425. 1955.
130. and Shrader, W. D. The interrelationships of
agronomy and economics in research and recommendations to
farmers. Agron. Jour. 45:496-501. 1953.
131. and Swanson, Earl R. Resource productivity in
Iowa farming (with special reference to uncertainty and capital
use in southern Iowa). Iowa Agr. Exp. Sta. Res. Bui. 388. 1952.
132. Woodworth, Roger C., Catron, Damon, and Ashton,

,
Gordon C. An experiment to derive productivity and substitution

coefficients in pork output. Jour. Farm Econ. 35:341-55. 1953.
133. , ,
and .
New procedures in estimating feed substitution rates in determin-

ing economic efficiency in pork production. I. Replacement rates
of corn and soybean oilmeal in fortified rations for growing-
fattening swine. Iowa Agr. Exp. Sta. Res. Bui. 409. 1954.
134. Herrmann, Louis F. Diminishing returns in feeding commercial

dairy herds. Jour. Farm Econ. 25:397-409. 1943.
135. Hildreth, Clifford G. Discrete models with qualitative restric-

tions. In Baum, E. L., Heady, Earl O., and Blackmore, John, eds.
Methodological procedures in the economic analysis of fertilizer
use data. pp. 62-75. Iowa State University Press, Ames, Iowa.
1956.
136. Point estimates of ordinates of concave functions.

»
Jour. Amer. Stat. Assn. 49:598-619. 1954.
137. Possible models for agronomic -economic re-

.
search. In Baum, E. L., Heady, Earl O., Pesek, John T., and
138. Hildreth, R. J. Influence of rainfall on fertilizer profits. Jour.
Farm Econ. 39:522-24. 1957.
*39* .Possible models for evaluating climate as it in-
teracts with soils to affect the production function. Conf for .
Coop’s, in the TVA Agr. Econ. Res. Act’s. March 25-27, 1958.
140. Hjelm, L. Utbytesrelationer i mjolkproduktionen.

(With an Eng-
lish summary: Input-output relationships milk production.)
in
Institutionem for Lantbrukets Drifts ekonomi. Stockholm, Sweden.
1953.
.
BIBLIOGRAPHY 655
141. Hoch, Irving. Estimation of production function parameters and

testing for efficiency. Unpublished paper presented before the
Econometric Society, Montreal, Canada. September, 1954.
142. . Simultaneous equation bias in the context of the
Cobb- Douglas production function. Econometrica 26:566-79. 1958.
143. Hoglund, C. R., Johnson, Glenn L., Lassiter, Charles A., and
McGilliard, Lon D., eds. Nutritional and economic aspects of
feed utilization by dairy cows. Iowa State University Press,
Ames, Iowa. 1959.
144. Huffman, C. F. and Duncan, C. W. The nutritive value of alfalfa
hay. Jour. Dairy Sci. 32:465-71. 1959.
145. Ibach, D. B. A graphic method of interpreting response to ferti-

lizer. U. S. Dept. Agr. Handbook No. 93. 1956.
146. . Use of standard exponential yield curves. U. S.

Dept. Agr., Agr. Res. Serv. ARS 43-69. 1958.
147. and Mendum, S. W. A simultaneous solution for
the exponential yield equation. Jour. Farm Econ. 40:469-76. 1958.
148. Jacobson, N. L. Problems in designing feeding experiments from

a nutritional standpoint. In Hoglund, C. R., Johnson, Glenn L.,
Lassiter, Charles A., and McGilliard, Lon D., eds. Nutritional
and economic aspects of feed utilization by dairy cows. pp. 206-
12. Iowa State University Press, Ames, Iowa. 1959.
149. Jarrett, F. G. Estimation of resource productivities as illustrated

by a survey of the lower Murray Valley dairying area. Australian
Jour. Stat. 1:3-11. 1959.
150. .
Resource productivities and production functions
Australian Jour. Agr. Econ. 1:67-68. 1957.
151. Jawetz, M. B. Farm size, farming intensity and the input-output

relationships of some Welsh and west of England dairy farms.
University College of Wales, Aberystwyth. 1957.
152. Jensen, D. R. and Pesek, J. T. Generalization of yield equations

in two or more variables. Agron. Jour. 51:255-63. 1959.
153. Jensen, Einar. Determining input- output relationships in milk

production. Jour. Farm Econ. 22:249-58. 1940.
154. ,
Klein, John W., Rauchenstein, Emil, Woodward,
T. E., and Smith, Roy H. Input-output relations in milk production.
U. S. Dept. Agr. Tech. Bui. 815. 1942.
155. Jensen, H. R. and Sundquist, W. B. Resource productivity and

income for a sample of west Kentucky farms. Kentucky Agr.
Exp. Sta. Bui. 630. 1955.
656 BIBLIOGRAPHY
156. Johnson, Glenn L. Interdisciplinary considerations in designing

experiments to study the profitability of fertilizer use. In Baum,
E. L., Heady, Earl O., and Blackmore, John, eds. Methodological
procedures in the economic analysis of fertilizer use data.
157. Planning agronomic- economic research in view

.
of results to date. In Baum, E. L., Heady, Earl O., Pesek, John

T., and Hildreth, Clifford G., eds. Economic and technical analy-
sis of fertilizer innovations and resource use. pp. 217-25. Iowa
State University Press, Ames, Iowa. 1957.
158. Problems in studying resource productivity and

.
size of business arising from managerial processes. In Heady,

Earl O., Johnson, Glenn L., and Hardin, Lowell S., eds. Resource
productivity, returns to scale, and farm size. pp. 16-23. Iowa
State University Press, Ames, Iowa. 1956.
159. Some contributions of microanalysis to agricul-

.
tural policy. In Baum, E. L., Heady, Earl O., Pesek, John T., and
160. Johnson, Paul R. Alternative functions for analyzing a fertilizer-

yield relationship. Jour. Farm Econ. 35:519-29. 1953.
161. An economic analysis of corn fertilization in the

.
coastal plain of North Carolina. Unpublished M.S. thesis. North

Carolina State College, Raleigh, North Carolina. 1952.
162. Kamiya, K. On productivity of labor. Jour. Rural Econ. 17:89-98.
1941.
163. Kelley, Paul L., McCoy, John H., Tucker, Henry, and Altau, Virve
T. Resource returns and productivity coefficients in central and
western Kansas country elevators of modern construction. Kansas
Agr. Exp. Sta. Tech. Bui. 88. 1957.
164. Knetsch, Jack L. Moisture uncertainties and fertility response
studies. Jour Farm Econ. 41:70-76. 1959.
.
165. ,
Robertson, L. S., Jr., and Sundquist, W. B. Eco-
nomic considerations in soil fertility research. Mich. Agr. Exp.
Sta. Quart. Bui. 39:10-17. 1956.
166. Konijn, H. S.Estimation of an average production function from

surveys. Econ. Record 35:118-25. 1959.
167. Leser, C. E. V. Production functions for the British industrial

economy. Applied Stat. 3:174-83. 1954.
168. • Production functions and British coal mining.
BIBLIOGRAPHY 657
169. Statistical production functions and economic

.
development. Scottish Jour. Pol. Econ. 5:40-49. 1958.
170. Liebig, Justus von. Die Grundsatze der Agriculturchemie mit

Rucks icht aufdie in England angestellten Untersuchungen. Fried-
rich Viewig und Sohn, Braunschweig, Germany. 1855.
171. Lloyd, A. G. Agricultural experiments and their economic sig-

nificance. Australian Jour. Agr. Econ. 2:33-42. 1958.
172. Fodder conservation in the Southern Tablelands

.
wool industry. Rev. of Marketing and Agr. Econ. 27:5-50. 1959.

173. Lomax, K. S. An agricultural production function for the United
Kingdom, 1924 to 1947. Manchester School of Economic and
Social Studies. 17:146-60. 1949.
174. Coal production functions for Great Britain.

.
Jour. Roy. Stat. Soc. 113:346-51. 1950.
175. Loosli, J. K., Lucas, H. L., and Maynard, L. A. The effect of

roughage intake upon the fat content of milk. Jour. Dairy Sci.
28:147-53. 1945.
176. Lucas, H. L. Experimental designs and analyses for feeding effi-

ciency trials with dairy cattle. In Hoglund, C. R., Johnson, Glenn
L., Lassiter, Charles A., and McGilliard, Lon D., eds. Nutritional
and economic aspects of feed utilization by dairy cows. pp. 187-91.
177. Marschak, J., and Andrews, William H., Jr. Random simultaneous
equations and the theory of production. Econometrica 12:143-205.
1944.
178. Martin, T. G., Stoddard, C. E., and Allen, R. S. Effects of varied

rates of hay feeding on body weight and production of lactating
dairy cows. Jour. Dairy Sci. 37:1233-40. 1954.
179. Mason, David D. Functional models and experimental designs for

characterizing response curves and surfaces. In Baum, E. L.,
Heady, Earl O., and Blackmore, John, eds. Methodological proce-
dures in the economic analysis of fertilizer use data. pp. 76-98.
180. Mason, G. Resource productivities from a sample of light plains
farms, Canterbury, N. Z. Unpublished M. Agr. Sc. thesis. Canter-
bury Agricultural College, New Zealand. 1958.
181. Maunder, A. H. Some farm experiments in the use of resources:
an application of marginal analysis. The Farm Economist 8:17-27.
1956.
182. May, Kenneth. A note on the pure theory of production. Economet-

rica 18:56-59. 1950.
658 BIBLIOGRAPHY
183. May, Kenneth. Structure of the production function of the firm.

184. McCarthy, William Owen. Relation of fertilizer rates to pasture
yield and utilization. Unpublished Ph.D. thesis. Iowa State Uni-
versity Library, Ames, Iowa. 1959.
185. Mendum, S. W. Spillman's solution of the exponential yield curve

and fertilizer problems. Jour. Farm Econ. 15:503-9. 1933.
186. Mendershausen, H. On the significance of Professor Douglas'
production functions. Econometrica 6:143-53. 1938.
187. Miller, W. G. Comparative efficiency of farm tenure classes in

the combination of resources. Agr. Econ. Res. 11:6-16. 1959.
188. Chryst, W. E., and Ottoson, H. W. Relative effi-

,
ciencies of farm tenure classes in intrafirm resource allocation.

Iowa Agr. Exp. Sta. Res. Bui. 461. 1958.
189. Mitscherlich, E. A.Das Gesetz des Minimums und das Gesetz

des abnehmenden Bodenertrages. Landw. Jahrb. 38:537-52. 1909.
190. . Second approximation of the law of action.

Zeitsch. Pflanzen. Dung und Boden. 12A:273-82. 1928.
191. Monroe, C. F. and Krauss, W. E. Relationship between fat content
of dairy grain mixtures and milk and butterfat production. Ohio
Agr. Exp. Sta. Bui. 644. 1943.
192. Mundlak, Y. Costs and income of established family farms. Falk
Project for Economic Research in Israel. Hebrew University of
Jerusalem, Rehovot, Israel. Unpublished report.
193. Murti, V. N. and Sastry, V. K. Production functions for Indian
industry. Econometrica 25:705-21. 1957.
194. Nath, N. C. B. Some production functions for Indian manufactur-

ing, 1947. Indian Jour Econ. 33:69-72. 1952.
.
195. Nelson, M., et al Use of the production function and linear pro-
.
gramming in valuation of intermediate products. Land Econ.

33:257-61. 1957.
196. Nichols, W. H. Labor productivity functions in meat packing.

University of Chicago Press, Chicago, Illinois. 1948.
197. Ojala, E. M. Agriculture and economic progress. Oxford Uni-

versity Press, New York. 1952.
198. Okawa, I. The theory and measurement of food economy.

Hitosubashi College of Commerce and Economics, Tokyo, Japan.
1945.
199. Olson, R. O. and Heady, E. O. Economic use of forages in

BIBLIOGRAPHY 659
livestock production on Corn Belt farms. U. S. Dept. Agr. Circ.

No. 905. 1952.
200. Orazem, Frank and Smith, Floyd W. An economic approach to

the use of fertilizer including an economic interpretation of a
corn-fertilizer experiment on verdigris -like soil in 1956. Kansas
Agr. Exp. Sta. Tech. Bui. 94. 1958.
201. Orton, F. J. The economy of feed input in milk production. The

Farm Economist 9:11-25. 1958.
202. Palvia, C. M. An econometric model for development planning.

The Hague. 1953.
Institute of Social Studies,
203. Panse, V. G., Sahasrabudhe, V. B., and Mokashi, V. K. Co-ordi-

nated manurial trials on rainfed cotton in peninsular India. Indian
Jour. Agr. Sci. 21:113-35. 1951.
204. Parish, R. M. and Dillon, J. L. Recent applications of the produc-

farm management research. Rev. of Marketing
tion function in
and Agr. Econ. 23:215-36. 1955.
205. Parks, W. L. Methodological problems in agronomic research
involving fertilizer and moisture variables. In Baum, E. L.,
206. and Knetsch, J. L. Corn yields as influenced by
nitrogen level and drouth intensity. Agron. Jour. 51:363-64. 1959.
207. Paschal, J. L. and French, B. L. A method of economic analysis

applied to nitrogen fertilizer rate experiments on irrigated corn.
208. Pesek, John T. Agronomic problems in securing fertilizer re-

sponse data desirable for economic analysis. In Baum, E. L.,
209. and Heady, E. O. Derivation and application of a

method for determining minimum recommended rates of fertiliza-
tion. Soil Sci. Soc. Amer. Proc. 22:419-23. 1958.
210. ,
Doll, John P., and Nichollson, R. P.
,
Production surfaces and economic optima for corn yields with

respect to stand and nitrogen levels. Iowa Agr. Exp. Sta. Res.
Bui. 472. 1959.
211. and Webb, John R. Economic interpretation of the

importance water solubility in phosphorus fertilizers when used
of
as hill fertilizer for corn. In Baum, E. L., Heady, Earl O., Pesek,
660 BIBLIOGRAPHY

212. Plaxico, J. S. Problems of factor-product aggregation in Cobb-
Douglas value productivity analysis. Jour. Farm Econ. 37:664-75.
1955.
213. Plessing, H. C. Udbyttekur ver med saerlight henblik paa en

.
matematisk for mulering af landbrugets udbyttelov. Nordisk

Jordbrugsforsk 25:399-424. 1943.
214. Reder, M. W. An alternative interpretation of the Cobb-Douglas
function. Econometrica 11:259-64. 1943.
215. Redman, John C. Economic aspects of feeding for milk produc-
tion. Jour. Farm Econ. 34:333-46. 1952.
216. and Allen, Stephen Q. Some interrelationships of
economic and agronomic concepts. Jour. Farm Econ. 36:453-65.
1954.
217. Reed, O. E., Fitch, J. B., and Cave, H. W. The relation of feeding
and age of calving to development of dairy heifers. Kan. Agr. Exp.
Sta. Bui. 233. 1924.
218. Reid, J. T. Pasture evaluation in relation to efficiency. In Hog-

lund, C. R., Johnson, Glenn L., Lassiter, Charles A., and McGil-
liard, Lon D., eds. Nutritional and economic aspects of feed
utilization by dairy cows. pp. 135-51. Iowa State University
Press, Ames, Iowa. 1959.
219. Rippel, A. Zur experimentellen Widerlegung des Mitscherlich -

Bauleschen Wirkungsgesetzes der Wachstumsfaktoren. Zeitsch.
Pflanzen. Dung und Boden. 8A:65-76. 1927.
220. Robinson, Joan. The production function. Econ. Jour. 65:67-71.

1955.
221. Rohtlieb, C. E. Om
grans en for jordbrukets intensifiering.
Ekonomisk Tidskrift. p. 189. 1916.
222. Romana, K. Y. Productivity: its measurement in Indian indus-
tries. Indian Jour. Econ. 34:34-35. 1953.
223. Rostas, L.Comparative productivity in British and American

industry.Occasional Paper 8. Nat. Inst, of Econ. and Soc. Res.,
Cambridge University Press, Cambridge, England. 1948.
224. Salter, W. E. G. The production function and the durability of
capital. Econ. Record 34:47-66. 1959.
225. Sandberg, Ole Romer, Jr. Efficiency of resource use in farming
in southeastern Norway. Unpublished M.S. thesis. Iowa State
University Library, Ames, Iowa. 1956.
BIBLIOGRAPHY 661
226. Schapper H. P. and Mauldon, R. G. A production function from

,
farms in the wholemilk region of western Australia. Econ. Record

33:52-59. 1957.
227. Shephard, R. W. Cost and production functions. Princeton Univer-

sity Press, Princeton, N. J. 1953.
228. Simkin, C. G. F. Aggregate production functions. Econ. Record

30:50-60. 1955.
229. Smith, Victor E. Nonlinearity in the relation between input and

output: the Canadian automobile industry, 1918-1930. Economet-
rica 13:260-72. 1945.
230. . The statistical production function. Econometrica

13:543-62. 1945.
231. Solow, R. M. The production function and the theory of capital.

Rev. Econ. Studies 23:101-8. 1955/56.
232. Technical change and the aggregate production

.
function. Rev. Econ. and Stat. 39:312-20. 1957.
233. Soper, C. S.Production functions and cross-section surveys.

Econ. Record 33:111-18. 1958.
234. Spillman, W. J. Application of the law of diminishing returns to
some fertilizer and feed data. Jour. Farm Econ. 5:36-52. 1923.
235. .Law of the diminishing increment in the fattening

of steers and hogs. Jour. Farm Econ. 6:179-91. 1924.
236. . Use of the exponential yieldcurve in fertilizer

experiments. U. S. Dept. Agr. Tech. Bui. 348. 1933.
237. and Long, Emil. The law of diminishing returns.

World Book Co., New York. 1924.
238. Strand, E. G., Heady, Earl O., and Seagraves, James A. Produc-
tivity of resources used on commercial farms. U. S. Dept. Agr.
Tech. Bui. 1128. 1955.
239. Sukhatme, P. V. Economics of manuring. Indian Jour. Agr. Sci.

11:325-37. 1941.
240. Sundquist, W. B. and Robertson, L. S., Jr. An economic analysis

of some controlled fertilizer input-output experiments in Michigan.
Mich. Agr. Exp. Sta. Tech. Bui. 269. 1959.
241. Suryanarayana, K. S. Resource returns in Telengana farms -

a production function study. Indian Jour. Agr. Econ. 13:20-26.
1958.
242. Swanson, Earl R. Determining optimum size of business from

production functions. In Heady, Earl O., Johnson, Glenn L., and
Hardin, Lowell S. Resource productivity, returns to scale, and
662 BIBLIOGRAPHY
farm size. pp. 133-43. Iowa State University Press, Ames, Iowa.
1956.
243. Swift, R. W., Thacker, E. J., Black, A., Bratzler, J. W., and Jones,
W. H. Digestibility of rations for ruminants as affected by propor-
tion of nutrients. Jour. Anim. Sci. 1:447-61. 1918.
244. Takayama, T. A study on the Cobb-Douglas production function —

with an application to rice production in Hokkaido. Rev. of the
Soc. of Agr. Econ. of Hokkaido University 15:1-24. 1959.
245. Tintner, G. A note on the derivation of production functions from

farm records. Econometrica 12:26-34. 1944.
246. Produktions Funktionen fiir dsterreichische

.
Landwirtschaft. Zeitschrift fur Nationalokonomie 17:426-42. 1958.
247. and Brownlee, O. H. Production functions derived

from farm records. Jour. Farm Econ. 26:566-71. 1944.
248. Tolley, H. R., Black, J. D., and Ezekiel, M. J. B. -Input as related

to output infarm organization and cost of production studies.
249. Tramel, Thomas E. Alternative methods for using production

functions for making recommendations. Jour. Farm Econ. 39:790-
94. 1957.
250. . A
suggested procedure for agronomic- economic
fertilizer experiments. In Baum, E. L., Heady, Earl O., Pesek,
251. Trant, G. Adjusting for price levels in production function
I.
studies. Earl O., Johnson, Glenn L., and Hardin, Lowell
In Heady,
S., eds. Resource productivity, returns to scale, and farm size,
252. Tsuchiya, K. Production functions of agriculture in Japan. Quart.

Jour. Agr. Economy. Vol. 9. 1955.
253. Verhulst, Michel J. J. The pure theory of production applied to

the French gas industry. Econometrica 16:295-308. 1948.
254. Wall, F. W. Alfalfa as the sole feed for dairy cows. Jour. Dairy
Sci. 1:447-61. 1918.
255. Wang, Y. Resource returns and productivity coefficients for se-

lected crop systems in Tainan area. Proceedings of Agricultural
Economics Seminars, National Taiwan University, Taipei, Taiwan,
Formosa, September 16-20, 1958. pp. 90-98. 1959.
256. Watanabe, T. Theory of production functions. Jour. Rural Econ.
Vol. 21. 1945.
BIBLIOGRAPHY 663
257. Wicksell, Knut. ‘Den “Kritiska punkten” i lagen for jordbrukets

avtagonde avkastning. Ekonomisk Tidskrift 18:285-92. 1916.
258. Williams, E. J. Regression analysis. John Wiley and Sons, New
York. 1959.
259. Williams, E. J. Simultaneous regression equations in experimen-
tation. Biometrika 45:96-110. 1958.
260. Willoughby, W. M. Limitations to animal production imposed by

seasonal fluctuations in pasture and by management procedures.
Australian Jour. Agr. Res. 10:248-68. 1959.
261. Wise, J. The estimation of the time-response functions in com-
plete economic systems. Econometrica 25:67-71. 1957.
262. Wold, H. Demand Analysis. John Wiley and Sons, New York.
1953. Ch. 2.
263. .
Causal inference from observational data. Jour.
Royal Statist. Soc. 119A-.28-61. 1956.
264. Ends and means in econometric model building:

.
Basic considerations reviewed. In Grenander, U., ed. Probability

and statistics: The Harald Cramer volume. John Wiley and Sons,
New York. 1959.
265. A case study of interdependent versus causal
.
chain systems. Rev. Inst. Int. Statist. 26:5-25. 1958.
266. Wolf son, R. J. An econometric investigation of regional differen-

tials in American agricultural wages. Econometrica 26:225-57.
1958.
267. Wragg, S. R., and Godsell, T. E. Production functions for dairy

farming and their application. The Farm Economist 8:1-6. 1956.
268. Yates, F., Boyd, D., and Pettit, G. Influence of changes in level of
feeding on milk production. Jour. Agr. Sci. 32:428-56. 1942.
269. Yuwata, Y. Production functions for rice and barley. Quart. Jour.
Agr. Economy 7. 1953.
270. Zellner, Arnold. An interesting general form for a production
function. Econometrica 19:188-89. 1951.
Index
Adjusted coefficient of multiple Cost curves, 59-64

correlation, R 2 589-90
, Crop response surfaces, fertilizer
Aggregation use, 475-525
over inputs, 216-17 Crop studies, United States, 594-97
within inputs 215-16
, Cross-sectional observations, 573-76
Alfalfa, 516-19 Culbertson, C. C., 452-74
Algebraic form, choice, 203-9
Applications D
economic, 31-72
practical, 299-301 Data
Ashton, Gordon C., 266-301, 302-29 beef, 452-55
Australia, 601-4 choice, 149-50
Austria, 608-11 collection, 142-94
Average products, world, 590 cross-sectional, 143-44
experimental, 144-49, 187-94
B nonexperimental, 144-49, 187-94
Data analysis, 108-41
Balloun, Stanley, 330-73, 374-403 Dean, Gerald W., 374-403
Basic concepts, 151-52 Design, experimental, 150-87, 267-68
Basic experiment, milk, 411-29 Designs
Basic nature of functions, milk, 405-6 complete factorial, 164-68
Beef, 452-74 composite, 170-75
empirical results, 455-58 fractional factorial, 168-70
gain isoquants, 458-70 randomized block, 155-64
gains and time, 470-72 response surface, 152-55
objectives and data, 452-55 rotatable, 175
reduced gain equation, 472-74 Diagnostic purposes, 554-55
substitution rates, 458-70 Doll, John P., 475-525
Bloom, Solomon, 404-51
Brown, William G., 475-525 E
C Economic
applications, 31-72
Canada, 598, 601 specification, 195-217
Carrington soil experiment, 526-33 Economic optima
Carter, Harold O., 452-74 fertilizer use, 475-525
Catron, Damon V., 266-301, 302-29 milk production functions, 443-50
Change, technological, 235-40 Elasticities, sum, 589
Choice under certainty, 71-72 Empirical problems, 218-65
Clarion-Webster soil, 558-64 Equations
Clover, red, 512-16 simultaneous, 137-41, 584
Cobb-Douglas single, 109-37
functions, 228-32 single variable, 73-83
turkeys, 390-94 Estimation of production function, hogs
Corn on pasture, 303-12
production functions, 478-80 Estimation procedures, 141
residual response, 519-25 Experimental needs, fertilizer, 525
yields, 480-92 Experiments, hogs on pasture, 302-3
665
666 INDEX
F L
Farm tenure, United States, 597-98 Least-cost rations, 298-99, 317-24

Fertilization broilers, 330-73
isoclines, 526-53 hogs on pasture, 317-25
isoquants, 526-53 turkeys, 374-403
surfaces, 526-53 Least- time rations, hogs on pasture,
Fertilizer response surfaces, 476-78 327-29
Fertilizer use, 475-525 Limitations, fertilizer, 525
Fixed plants, 554-84 Livestock studies, 240-54, 594-97
on Clarion -Webs ter soil, 558-64
on Marshall soil, 555-58 M
Forage and grain substitution, 404-51
Function adequacy, 209-12 McAlexander, Robert, 330-73
Functions McKee, Dean E., 302-29
fixed plants, 554-84 Marginal product to opportunity cost
inter production, 196-97 ratios, 51, 592-93
intra production, 196-97 Marginal products, 591
Marginal rates of substitution, forage
G and gram, 404-51
Marketing weights
Gain isoquants, beef, 458-70 broilers, 330-73
Geometric form, milk production hogs in drylot, 298-99
functions, 406 turkeys, 374-403
Grain and forage, substitution, 404-51 Marshall soil, 555-58
Milk production functions, 404-51
H Moody soil experiment, 534-38
Multi- equation models, 198-202
Hay surfaces, 549-53
Haynie soil experiment, 538-42 N
Heady, EarlO., 266-301, 302-29,
330-73, 374-403, 404-51, n resources, 83-96
452-74, 475-525, 526-53 New Zealand, 601-4
Hogs Nitrogen and stand, 542-49
in drylot, 266-301 Norway, 616-18
on pasture, 302-29
O
I
Objectives
India, 620-24 beef, 452-55
Initial experiments, fertilizer milk production functions, 404-5
response surfaces, 476-78 Omission of zero nutrient applications,
Inputs, 218-26 502-12
Interregional productivity Opportunity costs, 591-92
comparisons, 576-84 Outputs, 218-28
Israel, 619-20
P
J
Pesek, John T., 475-525, 526-53
Jacobson, Norman L., 404-51 Pork production functions
Japan, 624-27 hogs in drylot, 266-301
hogs on pasture, 302-29
INDEX 667
Poultry enterprise function, 568-73 Schnittker, John A., 404-51

Predicted yields, corn, 480-92 Simultaneous equations, 137-41, 584
Product imputation, 69-71 Single equations, 109-37, 198-202
Production elasticities, world Single variable equations, 73-83
functions, 589 Specification errors, 212-17
Production function Specification of the model, 197-209
basis for selection, 102-7 Speer, Vaughn C., 302-29
concepts, 1-4 Square root equation, 492-502
economic specification, 195-217 Stand and nitrogen, 542-49
estimates, 6-9 Statistical analysis, 268-82
estimation, 108-94, 218-65, 303-12 Stomach capacity, milk production
research, 4-6 functions, 450-51
studies, 1-30 Substitution coefficients, hogs on
Production functions pasture, 302-29
broilers, 330-73 Substitution rates, corn and soybean
compared over the world, 585-643 meal, 287-90
corn, 478-80 Surface functions, 96-102
254-65
fitted, Sweden, 611-16
forms, 73-107
historic summary, 9-30 T
milk, 404-5
omission of zero nutrient Taiwan, 627-29
applications, 502-12 Tama- Muscatine soils, 564-68
quantities, 32-40 Tenant inputs, 57-59
turkeys, 376-90 Time factor, beef, 470-72
Production surface estimates, Time of gains, 290-98
hogs on pasture, 312-17 Time series
Productivity comparisons, 64-69 data, 143-44
Profit maximizing quantities, 40-52 observation, 573-76
Time variable, milk production
R functions, 432-43
Turkeys, 374-403
Rate of gain, hogs on pasture, 325-27 basic data, 375-76
Rations experiment, 375-76
comparison, 282-87
economy, 266-67 U
Reduced gain equation, beef, 472-74
Residual response functions, corn, Union of South Africa, 604-7
519-25 United Kingdom, 607-8
Resource production elasticities, United States, 593-98
international, 629-33
Resource use efficiencies, V
international, 633-43
Resources, 52-57, 83-96 Variables
Returns to scale, 232-34 choice, 202-3
omission, 214-15
S
W
Sample means of outputs and inputs,
world functions, 590 Woodworth, Roger, 266-301
Scale returns, 69-71 World production functions, 585-643

Agricultural Production Functions.

Uploaded by

Copyright:

Available Formats

Agricultural Production Functions.

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Agricultural Production Functions.

Uploaded by

Copyright:

Available Formats

AGRICULTURAL

Iowa State University Press, ^Ames, Iowa

All rights reserved.

Library of Congress Catalog Card Number: 60-11128.

LITHOPRINTED IN THE UNITED STATES OF AMERICA BY

The application of formal production function concepts in agricul-

functions are employed, or serve most efficiently under a given set of

1. Development of Production Function Studies 1

3. Forms of Production Functions 73

4. Data Analysis for Production Function Estimation • . 108

5. Data Collection for Production Function Estimation 142

6. Economic Specification of the Production Function 195

7. Miscellaneous Empirical Problems Relating to

8. Pork Production Functions for Hogs Fed in Drylot 266

9. Pork Production Functions and Substitution Coefficients

10. Production Functions, Least-cost Rations, and Optimum

11. Least-cost Rations and Optimum Marketing Weights

12. Milk Production Functions and Marginal Rates

13. Production Functions and Substitution Coefficients

14. Crop Response Surfaces and Economic Optima

17. Comparison of Production Function Estimates

HIS BOOK deals with agricultural production functions. Its pur-

T pose is to summarize certain concepts, empirical methods, and

USE OF PRODUCTION FUNCTION CONCEPTS

The production function is a concept in physical and biological sci-

prediction of the production functions involved. In the past, these pro-

NEED FOR PRODUCTION FUNCTION RESEARCH

a manner to approach formal economic optima. Not only does a greater

Co-operative work on production functions among physical scien-

spontaneous. Also, it has mostly been on an informal basis between in-

The discussion above has centered on physical production functions.

in fact, some of the more important ones which need

Physical Production Functions

We first review studies relating to physical production functions of

where y is yield response, x is the quantity of the nutrient, and a is

where y is total yield, c is yield level in absence of applications of x,

The first attempt to define the algebraic nature of the fertilizer -

(1.3) log A - log (A - y) = cx

to explain fertilizer response allowing diminishing marginal produc-

where k is a “damage factor” due to excessive magnitudes of x.

with the latter defining the magnitude of incremental yields associated

(1.7) y = A(1 - R x )(l - Rz )

where y is output from variable inputs, A is the maximum increase

where E is a maximum yield, b is the quantity of x initially in the soil,

B ' Zu Mitscherlich s gesetz der physiologischen bezichungen. Landw.

category measured as total digestible nutrients. Estimates were de-

where w is live weight in pounds after start of the feeding period, e is

Numerous early general economists, such as Smith, Ricardo, and

by fixed proportion, but diminishing productivity for one factor in-

and, as in the early attempts of Douglas, said that a and y should

Cobb and Douglas computed the marginal productivities of capital

Country and Period b k 1-k

Massachusetts (1890-1926) 1.007 .74 .26

In an article published in 1936, Edelberg concerned himself with a

(1.20) U.K. P = cs*

Country, region, Number of industries

United States (1904) 336 2.08 .65 .31

Bronfenbrenner and Douglas also computed separate cross-sectional

Industry Number of observations b k j

Clothing -textiles 16 2.043 .978 -.070