Chapter 1

Introduction

Index number is an abstract concept which is generally used to

measure the change in a set of related variables over time or
to compare general levels in these variables across countries or
regions. The basic concept of index number is used in many
disciplines, but it appears to have earned special significance in
the discipline of economics. Index numbers are useful instru-
ments in any economist's tool kit as they can be used to show
the changes/movements in any variable of interest in a simple
and easily understandable manner. The consumer price index
(CPI), which measures the changes in prices of a range of con-
sumer goods and services, is the most widely used economic in-
dicator in any country. Other important index numbers include
the financial indices such as the Dow Jones and Nikkei indices
of stock market prices; price deflators for various national in-
come aggregates; purchasing power parities of currencies; index
of joint factor productivity; and indices of import and export
prices and of competitiveness.

1
E. A. Selvanathan et al., Index Numbers
© E. A. Selvanathan and D. S. Prasada Rao 1994
2 Introduction

1.1 Historical Background

Index numbers have a long and distinguished history in eco-

nomics, with some of the most important contributions due to
Edgeworth, Laspeyres and Paasche dating back to the late nine-
teenth century. Formulae proposed by Laspeyres (1871) and
Paasche (1874) are still very commonly used by national statis-
tical offices around the world. But it is the work of Irving Fisher
and his book, The Making of Index Numbers, published in 1922,
that recognized the possibility of many statistical formulae to
derive appropriate index numbers. Fisher also introduced the
idea of test approach to index numbers which essentially in-
volves stating a number of intuitively obvious properties and
using them for purposes of selection of an appropriate formula
or to derive suitable formulae using mathematical functional
analysis which satisfy a range of tests. This approach, due to
Fisher, falls under the atomistic approach to index numbers.

Under the atomistic approach, construction of index num-

bers for different variables such as prices of consumer and pro-
ducer goods, quantities of consumer and producer goods, prices
and quantities of inputs used, assumes that all the variables are
independent entities. The atomistic approach can be considered
to be a fairly general scheme which contains the test approach
and stochastic approach as special cases. The test approach has
also given rise to the axiomatic approach, to be found in the
works of Eichorn and Voeller (1983) and Diewert (1992), where
a set of properties are stated in the form of axioms to be met
by an index number formula. These axioms are then used in
finding a suitable formula for purposes of comparison.
The Stochastic Approach 3

It has long been recognized that treating prices and quan-

tity observations as unrelated is untenable. Standard economic
theory implies the existence of functional relationship between
different economic variables. Index numbers must be consid-
ered and computed within the framework of such theoretically
justified relationships. This has led to the functional approach
described in Frisch (1936). The functional approach postulates
and utilizes the interrelationships to form the foundations for
the definition of cost-of-living index numbers and various index
numbers to measure changes in output and productivity. Again
the functional approach has had a long history dating back to
the work of Konus in 1924. Since then this approach has been
enriched by the works of, to name a few, Keynes, Samuelson,
Diewert and Theil. Diewert (1976; 1978; 1981; and 1992) for-
malizes the functional approach and establishes the economic
theoretic properties of some of the well-known index numbers.
Diewert introduces the idea of exact and superlative index num-
bers to describe the theoretical properties of various index num-
bers. Fisher's Ideal index and the Theil-Tornqvist index figure
prominently in Diewert's work.

1.2 The Stochastic Approach

Research work on index numbers to date is predominantly pre-

occupied with finding formulae, with desirable and acceptable
economic theoretic properties that can be used in obtaining nu-
merical values of the index numbers. Once the formula is chosen,
the index number produces a single numerical value, indicating
the change in the variable of interest, based on the observed
data. This is the strategy used in the computation of the cost-
4 Introduction

of-living index numbers and others as well. The question arises

as to the significance of the numerical value thus obtained. For
example, if the CPI is found to be 1.08 for the year, what does
this indicate? Obviously it implies that prices increased by 8
percent during the year under consideration relative to the base
year. Is this figure absolute? How reliable is this increase? It
will be far more convincing if one were to interpret this 8 per-
cent to be an estimate of the price increase during the year, as
it conveys to the casual observer that there is some uncertainty
or degree of reliability associated with this numerical value of
the index. None of the approaches described above address this
question. Obviously the utility of the numerical value of the in-
dex depends upon the level of confidence that can be attached
to the value of the index.

To a certain degree this question is addressed by statisti-

cians who are interested in the sampling aspects underlying the
computation of the cost-of-living index or, for that matter, any
other index. As the indices are based on price and quantity
data for a range of commodities, selection of the commodities,
the sampling scheme used in their selection, and the precision
with which the price quotations are obtained has a bearing on
the precision of the over all index. Banerjee (1975), Kott (1984)
and others have considered this issue at length and devised pro-
cedures to obtain standard errors for the indices based on the
nature of the sampling schemes used and the reliability of data.

Even in the case where prices of all the commodities of rele-

vance are measured, and measured without any errors, the ques-
tion of reliability of a given index arises. Consider two scenar-
ios. First scenario is where prices increase by 8 percent for all
the commodities, thus leading to an index value of 8 percent. In
The Stochastic Approach 5

this case the value of the index is a reflection of the price change
and there can be no doubt about its validity or reliability. Now
consider the second scenario where some commodities exhibit a
price decline, say by 5 percent, and the rest of the commodi-
ties show an increase in the price by 15 percent, but the index
formula gives a value of 8 percent. Under both scenarios, the
index values are the same. In both cases all the commodities
are priced, i.e., no sampling is involved and without any errors
in measurement. Intuitively, the index value of 8 percent is
more satisfactory in the first scenario, but far from satisfactory
in the second scenario where the ability of the index to reflect
the price changes in all the commodities is dubious. Such a
difference should be reflected in the form of some measures of
reliability associated with each index, and these should be pub-
lished along with numerical values of the index as a matter of
course.

In the light of this discussion, it is somewhat disappointing

that there have been no attempts in the past to obtain measures
such as this. None of the basic approaches listed and discussed
thus far address this question. It is in this context the material
of the present monograph becomes relevant. It is our view that
the stochastic approach, which was very briefly dealt with by
Frisch (1936), is the only approach that has the potential to
provide estimates of reliability along with index numbers. The
approach pursued here transcends the simplistic interpretation
accorded by Frisch where he considers the stochastic approach
as the measurement of a central tendency from a distribution of
price relatives.

The stochastic approach considers the index number prob-

lem as a signal extraction problem from the messages concerning
6 Introduction

price changes for different commodities. Obviously the strength

of the signal extracted depends upon the messages received and
the information content of the messages.

1.3 A Preview of the Book

This book focuses on the stochastic approach to index num-

bers. This approach has a long history going back to Edgeworth
(1925) and Frisch (1936). Later, this approach was reconsidered
by Theil (1965), Banerjee (1975), Balk (1980), Clements and
Izan (1981) and Diewert (1981). In addition, the stochastic ap-
proach has recently attracted renewed attention, see Clements
and !zan (1987); Giles and McCann (1991); Prasada Rao and
Selvanathan (1991, 1992a, 1992b, 1992c); Selvanathan (1987,
1989, 1991, 1993); and Selvanathan and Prasada Rao (1992).
The attraction of this approach is that it provides an alterna-
tive interpretation to some of the well known index numbers
as the estimators of parameters of specific regression models.
For example the Laspeyres, Paasche, Theil-Tornqvist and other
index numbers can be derived from various regression models.
Further this approach provides standard errors for these index
numbers. These standard errors reflect the reliability of the
index as a single numerical measure of change in a set of com-
modity prices as well as the precision which is influenced by
the sampling aspects and errors in measurement. The use of
standard errors has many practical implications. For example,
the employers and the employee unions can use the standard er-
rors to construct confidence interval estimates for the consumer
price index which can form the basis for wage negotiations; wel-
fare agencies and governments could use such interval estimates
A Preview 7

for indexing social benefit payments etc., international financial

institutions could use those interval estimates for indexing the
loans/aid to various countries.

The main purpose of this monograph is to describe the

stochastic approach to index number construction, and demon-
strate the versatility and usefulness of this approach in review-
ing the traditional index number formulae in a new light. In this
book, we attempt to bring together some of the recent develop-
ments, synthesize them within a single conceptual framework,
and demonstrate the scope and power of the stochastic approach
in the measurement of changes in price and quantity levels over
time and across different regions and countries. For purposes
of exposition, the monograph concentrates on the construction
of price index numbers within the tenets of the stochastic ap-
proach.

Chapter 2 considers various approaches to the construction

of index numbers in some detail. The chapter starts with an
intuitive definition of a general price index and shows how var-
ious price index number formulae such as Laspeyres, Paasche,
Edgeworth- Marshall, Drobisch and Geary-Khamis indices can
be derived by selecting an appropriate reference quantity vec-
tor. The chapter then introduces the atomistic approach and
uses that approach to define weighted and unweighted index
numbers. It demonstrates that selection of different weighting
schemes leads to Laspeyres, Paasche, Theil-Tornqvist, Theil,
Rao and Stuvel index numbers. This chapter also looks at
the major determinants of the gap between the Laspeyres and
Paasche index numbers, which in turn provides a justification
for the definition of Fisher and Stuvel index numbers while the
8 Introduction

main focus is on price index numbers. This is followed by a dis-

cussion of the functional approach to price index numbers. A
brief section is devoted to the problem of quantity index num-
bers. The test or axiomatic approach to index numbers is briefly
described, stating a number of desirable properties expected to
be satisfied by index number formulae. A numerical illustration
concludes the chapter.

Chapter 3 outlines the stochastic approach and shows how

various index number formulae can be derived under this ap-
proach. It draws on the similarity of the index number problem
and the signal extraction problem and sets up basic regression
models for observed price relatives under alternative specifica-
tions of error covariance structure to derive estimators of the pa-
rameters of the models as price index numbers. This approach
leads to the development of standard errors of the price indices.
These standard errors will be higher when there is substantial
variation in relative prices. This agrees with the intuitive notion
that the price index will be less well defined when relative prices
change disproportionately. This chapter also introduces models
with expenditure variables to derive price index numbers and
their standard errors. The basic model is extended to rectify its
weakness that the expected value of all relative price changes is
the same for all commodities and the new model is utilized to
estimate the relative price changes.

Chapters 4, 5 and 6 extend the use of stochastic approaches

to the measurement of the rate of inflation, to the estimation of
chain base index numbers and to measure changes in prices in
the context of multilateral comparisons, respectively. Chapter
4 also considers the aggregation problem of index numbers and
shows that the estimator for the measurement of price changes
A Preview 9

is invariant to the level of commodity aggregation. These results

are illustrated with the UK private consumption data.

The fixed base index numbers are useful to compare the

price changes between two years/periods which are not too dis-
tant from each other. When the current period and base period
are far apart, the chain base index numbers are strongly rec-
ommended. Chapter 5 describes the stochastic approach based
on expenditure variables to obtained fixed and chain base index
numbers. A system approach based on Seemingly Unrelated Re-
gression (SUR) is introduced to take account of any correlation
between price movements in different commodities to result in
more efficient estimators. The system approach is used to de-
rive, Laspeyres, Paasche index numbers. It is demonstrated
that the system approach is superior to the single equation ap-
proach as it facilitates the testing of various hypotheses. An
application is also presented in that chapter to illustrate the
main results. All the results show that the standard errors are
somewhat lower for the chain based index numbers than the
fixed base index numbers.

Chapter 6 shows how the stochastic approach can be used

to derive index number formulae in the context of multilateral
comparisons. This chapter introduces the well known Caves,
Christensen and Diewert (CCD) (1982a) index and derives it
and its standard error under the stochastic approach. An index
superior to CCD called the Generalized CCD ( GCCD) is also
proposed in the chapter. The estimation of these index numbers
are illustrated with the United Nations' International Compar-
ison Programme (ICP) data. This chapter also examines the
Uniformly Minimum Variance Unbiased (UMVU) estimation of
the Theil-Tornqvist and the CCD indices. The stochastic ap-
10 Introduction

proach to the estimation of the purchasing power parities under

the Geary-Khamis method is also discussed in Chapter 6, and
its application is illustrated with the ICP data. The chapter fur-
ther assesses the quality of the estimates using Effron's (1979)
distribution free bootstrap technique.

The final chapter deviates from the general scheme of the

book and it examines a number of data-related issues that arise
in practice and illustrates how these problems can be resolved
using the stochastic approach. Four issues of importance are
discussed. First, we consider the problem of seasonal commodi-
ties and describe a method suggested by Balk (1980) to tackle
this problem. A common problem of missing price informa-
tion in the context of international comparisons is next tackled
using the country-product-dummy (CPD) method of Summers
(1973). The third issue concerns the sampling aspects of price
index number construction and examines the bias associated
with some of the widely used formulae like the Laspeyres and
Paasche indices. Rao, Prasada Rao and Selvanathan (1993) ex-
amines this issue and a few tentative suggestions are listed in
this section. Finally, the problem of quality variation in price
and quantity comparisons is discussed. The hedonic approach to
this problem based on regression models as a possible solution
to the quantity problem is enunciated.