Advanced Econometrics - 1985 - 1era Edición - Amemiya

ADVANCED
ECONOMETRICS
Takeshi Amemiya
"The book provides an excellent overview of mod-

ern developments in such major subjects as ro-
bust inference, model selection methods,
feasible generalized least squares estimation,
nonlinear simultaneous systems models, dis-
crete response analysis, and limited dependent
variable models."
Charles F. Manski,
University of Wisconsin, Madison
Advanced Econometrics is both a comprehensive

text for graduate students and a reference work
for econometricians. It will also be valuable to
those doing statistical analysis in the other
social sciences. Its main features are a thorough
treatment of cross-section models, including
qualitative response models, censored and trun-
cated regression models, and Markov and dura-
tion models, as well as a rigorous presentation of
large sample theory, classical least-squares and
generalized least-squares theory, and nonlinear
simultaneous equation models.
Although the treatment is mathematically rigor-
ous, the author has employed the theorem-proof
method with simple, intuitively accessible as-
sumptions. This enables readers to understand
the basic structure of each theorem and to gen-
eralize it for themselves depending on their needs
and abilities. Many simple applications of theo-
rems are given either in the form of examples
in the text or as exercises at the end of each
chapter in order to demonstrate their essential
points.
Advanced Econometrics
Takeshi Amemiya
Harvard University Press Cambridge, Massachusetts 1985

Copyright C 1985 by Takeshi Amemiya
All rights reserved
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1
This book is printed on acid-free paper, and its binding materials

have been chosen for strength and durability.
Library of Congress Cataloging in Publication Data

Amemiya, Takeshi.
Advanced econometrics.
Bibliography: p.
Includes index.
1. Econometrics. I. Title.
HB139.A54 1985 330'.028 85-845
ISBN 0-674-00560-0 (alk. paper)
Takeshi Amemiya is Professor of Economics.
Stanford University, and coeditor of the Journal
of Econometrics.
Preface
This book is intended both as a reference book for professional econometri-

cians and as a graduate textbook. If it is used as a textbook, the material
contained in the book can be taught in a year-long course, as I have done at
Stanford for many years. The prerequisites for such a course should be one
year of calculus, one quarter or semester of matrix analysis, one year of
intermediate statistical inference (see list of textbooks in note 1 of Chapter
3), and, preferably, knowledge of introductory or intermediate econometrics
(say, at the level of Johnston, 1972). This last requirement is not necessary, but
I have found in the past that a majority of economics students who take a
graduate course in advanced econometrics do have knowledge of introduc-
tory or intermediate econometrics.
The main features of the book are the following: a thorough treatment of
classical least squares theory (Chapter 1) and generalized least squares theory
(Chapter 6); a rigorous discussion of large sample theory (Chapters 3 and 4); a
detailed analysis of qualitative response models (Chapter 9), censored or
truncated regression models (Chapter 10), and Markov chain and duration
models (Chapter 11); and a discussion of nonlinear simultaneous equations
models (Chapter 8).
The book presents only the fundamentals of time series analysis (Chapter 5
and a part of Chapter 6) because there are several excellent textbooks on the
subject (see the references cited at the beginning of Chapter 5). In contrast, the
models! discuss in the last three chapters have been used extensively in recent
econometric applications but have not received in any textbook as complete a
treatment as I give them here. Some instructors may wish to supplement my
book with a textbook in time series analysis.
My discussion of linear simultaneous equations models (Chapter 7) is also
brief. Those who wish to study the subject in greater detail should consult the
references given in Chapter 7.! chose to devote more space to the discussion of
nonlinear simultaneous equations models, which are still at an early stage of
development and consequently have received only scant coverage in most
textbooks.
vi Preface
In many parts of the book, and in all of Chapters 3 and 4, I have used the
theorem-proof format and have attempted to develop all the mathematical
results rigorously. However, it has not been my aim to present theorems in full
mathematical generality. Because I intended this as a textbook rather than as a
monograph, I chose assumptions that are relatively easy to understand and
that lead to simple proofs, even in those instances where they could be relaxed.
This will enable readers to understand the basic structure of each theorem and
to generalize it for themselves depending on their needs and abilities. Many
simple applications of theorems are given either in the form of examples in the
text or in the form of exercises at the end of each chapter to bring out the
essential points of each theorem.
Although this is a textbook in econometrics methodology, I have included
discussions of numerous empirical papers to illustrate the practical use of
theoretical results. This is especially conspicuous in the last three chapters of
the book.
Too many people have contributed to the making of this book through the
many revisions it has undergone to mention all their names. I am especially
grateful to Trevor Breusch, Hidehiko Ichimura, Tom MaCurdy, Jim Powell,
and Gene Sevin for giving me valuable comments on the entire manuscript. I
am also indebted to Carl Christ, Art Goldberger, Cheng Hsiao, Roger
Koenker, Tony Lancaster, Chuck Mansld, and Hal White for their valuable
comments on parts of the manuscript. I am grateful to Colin Cameron, Tom
Downes, Harry Paarsch, Aaron Han, and Choon Moon for proofreading and
to the first three for correcting my English. In addition, Tom Downes and
Choon Moon helped me with the preparation of the index. Dzung Pham has
typed most of the manuscript through several revisions; her unfailing patience
and good nature despite many hours of overtime work are much appreciated.
David Criswell, Cathy Shimizu, and Bach-Hong Tran have also helped with
the typing. The financial support of the National Science Foundation for the
research that produced many of the results presented in the book is gratefully
acknowledged. Finally, I am indebted to the editors of the Journal of Eco-
nomic Literature for permission to include in Chapter 9 parts of my article
entitled "Qualitative Response Models: A Survey" (Journal of Economic
Literature 19:1483 1536, 1981) and to North-Holland Publishing Company
for permission to use in Chapter 10 the revised version of my article entitled
"Tobit Models: A Survey" ( Journal of Econometrics 24:3 61, 1984).
Contents
1 Classical Least Squares Theory 1

2 Recent Developments in Regression Analysis 45
3 Large Sample Theory 81
4 Asymptotic Properties of Extremum Estimators 105
5 Time Series Analysis 159
6 Generalized Least Squares Theory 181
7 Linear Simultaneous Equations Models 228
8 Nonlinear Simultaneous Equations Models 245
9 Qualitative Response Models 267
10 Tobit Models 360
11 Markov Chain and Duration Models 412
Appendix 1 Useful Theorems in Matrix Analysis 459

Appendix 2 Distribution Theory 463
Notes 465
References 475
Name Index 505
Subject Index 511
1 Classical Least Squares Theory
In this chapter we shall consider the basic results of statistical inference in the
classical linear regression model the model in which the regressors are inde-
pendent of the error term and the error term is serially uncorrelated and has a
constant variance. This model is the starting point of the study; the models to
be examined in later chapters are modifications of this one.
1.1 Linear Regression Model
In this section let us look at the reasons for studying the linear regression
model and the method of specifying it. We shall start by defining Model 1, to
be considered throughout the chapter.
1 .1 .1 Introduction
Consider a sequence of K random variables (y x2 x31 ,. . . , xK, ),
t = 1, 2, . . . , T. Define a T-vector Y = (Y1, Y2,. . Yr)', a (K 1)-
vector = (x2 x3 . . . , x)', and a [(K 1) X 71-vector x*
(xr, xi",. . . , xV)'. Suppose for the sake of exposition that the joint density
of the variables is given byf(y, x*, 0), where 0 is a vector of unknown parame-
ters. We are concerned with inference about the parameter vector 0 on the
basis of the observed vectors y and x*.
In econometrics we are often interested in the conditional distribution of
one set of random variables given another set of random variables; for exam-
ple, the conditional distribution of consumption given income and the condi-
tional distribution of quantities demanded given prices. Suppose we want to
know the conditional distribution of y given x*. We can write the joint density
as the product of the conditional density and the marginal density as in
f(Y, x*, 0) = f (ylx* , e 111(x* , 02). (1.1.1)
Regression analysis can be defined as statistical inferences on 01. For this
purpose we can ignoref(x*, 02), provided there is no relationship between 0,
2 Advanced Econometrics
and 02. The vector y is called the vector of dependent or endogenous variables,
and the vector x* is called the vector of independent or exogenous variables.
In regression analysis we usually want to estimate only the first and second
moments of the conditional distribution, rather than the whole parameter
vector 01. (In certain special cases the first two moments characterize 0 1
complety.)Thuswandfiergolysaticnfero
the conditional mean E(y I x*) and the conditional variance-covariance matrix
V(ylx*). Generally, these moments are nonlinear functions of x*. However, in
the present chapter we shall consider the special case in which E(y,1 x*) is equal
to E(y,14 ) and is a linear function of xt, and V(y I x*) is a constant times an
identity matrix. Such a model is called the classical (or standard) linear
regression model or the homoscedastic (meaning constant variance) linear
regression model. Because this is the model to be studied in Chapter 1, let us
call it simply Model 1.
1.1.2 Modell
By writing x, = (1, x")', we can define Model 1 as follows. Assume
y, x',/1 + u t = 1, 2, . . . , T, (1.1.2)
where y, is a scalar observable random variable, fi is a K-vector of unknown
parameters, x, is a K-vector of known constants such that x, 34 is nonsin-
gular, and u, is a scalar, unobservable, random variable (called the error term
or the disturbance) such that Eu, = 0, Vu,= a 2 (another unknown parameter)
for all t, and Eut us = 0 for t s.
Note that we have assumed x* to be a vector of known constants. This is
essentially equivalent to stating that we are concerned only with estimating
the conditional distribution of y given x*. The most important assumption of
Model I is the linearity of E(y,Ixt ); we therefore shall devote the next subsec-
tion to a discussion of the implications of that assumption. We have also made
the assumption of homoscedasticity (Vu, = a 2 for all t) and the assumption of
no serial correlation (Eut us = 0 for t # s), not because we believe that they are
satisfied in most applications, but because they make a convenient starting
point. These assumptions will be removed in later chapters.
We shall sometimes impose additional assumptions on Model 1 to obtain
certain specific results. Notably, we shall occasionally make the assumption of
serial independence of (u,) or the assumption that u, is normally distributed.
in general, independence is a stronger assumption than no correlation, al-
Classical Least Squares Theory
though under normality the two concepts are equivalent. The additional
assumptions will be stated whenever they are introduced into Model.
1 .1 .3 Implications of Linearity
Suppose random variables y, and have finite second moments and their
variance-covariance matrix is denoted by
17 121
V [ x = [a
',11 L a 12 1 22
Then we can always write

Y: = /30 + vr, (1.1.3)
where , -1= 1,124:7 12, fib= EYt 0121-2-21E4 Ev, = 0, Vv =? 17, o'12, - -
and Ex? v, =0. It is important to realize that Model 1 implies certain assump-
tions that (1.1.3) does not: (1.1.3) does not generally imply linearity of
E(y:Ixt ) because E(v,Ixt ) may not generally be zero.
We call flo + Ain (1.1.3) the best linear predictor of y, given because
flo and fi can be shown to be the values of 1 0 and b 1 that minimize
E(y, b o b )2. In contrast, the conditional mean E(y,lx* ) is called the
best predictor of y, given because E[y, E(yrix:9]2 E[y, g(4)1 2 for
any function g.
The reader might ask why we work with eq. (1.1.2) rather than with (1.1.3).
The answer is that (1.1.3) is so general that it does not allow us to obtain
interesting results. For example, whereas the natural estimators of /3 0 and /31
0 andcanbedfiyrplgthemonsfy,adxthcreiz/3
/31 with their corresponding sample moments (they actually coincide with the
least squares estimator), the mean of the estimator cannot be evaluated with-
out specifying more about the relationship between xr and v, .
How restrictive is the linearity of E(y,14 )? It holds if y, and x* are jointly
normal or if y, and 4 are both scalar dichotomous (Bernoulli) variables.' But
the linearity may not hold for many interesting distributions. Nevertheless,
the linear assumption is not as restrictive as it may appear at first glance
because 4 can be variables obtained by transforming the original indepen-
dent variables in various ways. For example, if the conditional mean of y the
supply of good, is a quadratic function of the price, p we can put
(p plr , thereby making E(y,lx* ) linear.
1.1.4 Matrix Notation

To facilitate the subsequent analysis, we shall write (1.1.2) in matrix notation
as
Y= u, (1.1.4)
where Y = (Y) , Y2, YT) ' U = (U1, U7, UT) ' and X=
, x2 , . . . , xr)'. In other words, X is the T X K matrix, the tth row of
which is x' The elements of the matrix X are described as
X12 X1K1
X2K
x=
X7-2 XTK
If we want to focus on the columns of X, we can write X =

[x(1), x(2), , X], each Ni) is a T-vector. If there is no danger of
confusing ;0 with x we can drop the parentheses and write simply ;. In
matrix notation the assumptions on X and u can be stated as follows: X' X is
nonsingular, which is equivalent to stating rank (X) = K if T K; Eu = 0;
and Euu ' = a 2I T, where IT is the T X Tidentity matrix. (Whenever the size of
an identity matrix can be inferred from the context, we write it simply as I.)
In the remainder of this chapter we shall no longer use the partition
= (ho, fl); instead, the elements of /3 will be written as f? =
(p1. /12 , . . . , A)'. Similarly, we shall not necessarily assume that x a) is the
vector of ones, although in practice this is usually the case. Most of our results
will be obtained simply on the assumption that X is a matrix of constants,
without specifying specific values.
1.2 Theory of Least Squares

In this section we shall define the least squares estimator of the parameter /3in
Model 1 and shall show that it is the best linear unbiased estimator. We shall
also discuss estimation of the error variance a 2.
1.2.1 Definition of Least Squares Estimators of ft and a2

The least squares (LS) estimator /3 of the regression parameter'? in Model 1 is
defined to be the value of /3 that minimizes the sum of squared residuals2
Classical Least Squares Theory 5
S(fi) = (Y (Y (1.2.1)
= y'y 2y'Xfi + fi'X' Xi/
Putting the derivatives of S(j3) with respect to /3 equal to 0, we have
OS =
2X'y + 2X'Xfi= 0, (1.2.2)
where asiap denotes the K-vector the ith element of which is as/api, being
the ith element of ft. Solving (1.2.2) for ft gives
/j= (X'X)'X'y. (1.2.3)
Clearly, SA attains the global minimum at
Let us consider the special case K = 2 and 74 = (1, x21 ) and represent each
of the T-observations (y xv ) by a point on the plane. Then, geometrically,
the least squares estimates are the intercept and the slope of a line drawn in
such a way that the sum of squares of the deviations between the points and the
line is minimized in the direction of the y-axis. Different estimates result if the
sum of squares of deviations is minimized in any other direction.
Given the least squares estimator /3, we define
(1.2.4)
and call it the vector of the least squares residuals. Using fl, we can estimate a 2
by
8.2 = (1.2.5)
called the least squares estimator of a 2, although the use of the term least
squares here is not as compelling as in the estimation of the regression
parameters.
Using (1.2.4), we can write
Y = Xft + = PY MY, (1.2.6)
where P = X(X' 'X ' and M = I P. Because fi is orthogonal to X (that is,
' X = 0), least squares estimation can be regarded as decomposing y into two
orthogonal components: a component that can be written as a linear combi-
nation of the column vectors of X and a component that is orthogonal to X.
Alternatively, we can call Py the projection of y onto the space spanned by the
column vectors of X and My the projection of y onto the space orthogonal to
X. Theorem 14 of Appendix 1 gives the properties of a projection matrix such
as P or M. In the special case where both y and X are two-dimensional vectors
Figure 1.1 Orthogonal decomposition of y
(that is, K= 1 and T= 2), the decomposition (1.2.6) can be illustrated as in

Figure 1.1, where the vertical and horizontal axes represent the first and
second observations, respectively, and the arrows represent vectors.
From (1.2.6) we obtain
= Y'PY Y'MY. (1.2.7)
The goodness of fit of the regression of y on X can be measured by the ratio
y' Py/y' y, sometimes called R2. However, it is more common to define R 2 as
the square of the sample correlation between y and Py:
R2 (Y LPY)2 (1.2.8)
y'L y'PLy'
where L = I T 7-111 and1 denotes the T-vector of ones. If we assume one of
the columns of X is 1 (which is usually the case), we have LP = PL. Then we
can rewrite (1.2.8) as
_ y'LPLy _ y'My
(1.2.9)
y'Ly y'Ly
Thus R2 can be interpreted as a measure of the goodness of fit of the regression
of the deviations of y from its mean on the deviations of the columns of X from
their means. (Section 2.1.4 gives a modification of R2 suggested by Theil,
1961.)
1.2.2 Least Squares Estimator of a Subset of #

It is sometimes useful to have an explicit formula for a subset of the least
squares estimates ft. Suppose we partition /I' = ),where ft1 is a K1-vec-
tor and /.3. 2 is a K2 vector such that ICI K2 = K. Partition X conformably as

-
X = (X 1 , X2 ). Then we can write X' Xfi = X ' y as

XcX 1 /11 XcX2fi2 = Xy (1.2.10)
and
Xpti + Xpi2/I2 = Yqy. (1.2.11)
Solving (1.2.11) for f-t2 and inserting it into (1.2.10), we obtain
= (X11142X1 )-IVIM2y, (1.2.12)
where M2 = I - X2 (X;X2 . Similarly,
= NMIX2 rIX2M1Y, (1.2.13)

where M1 = I X I (XIX, )-13q.
In Model 1 we assume that Xis of full rank, an assumption that implies that
the matrices to be inverted in (1.2.12) and (1.2.13) are both nonsingular.
Suppose for a moment that X 1 is of full rank but that X2 is not. In this case /32
1 still can be estimated by modifying (1.2.12) as canotbesimd,u$
= (1.2.14)
where MI = I 14(X' xt) ixr, where the columns of XIconsist of a maxi-

mal number of linearly independent columns of X2, provided that VIMIXI is

nonsingular. (For the more general problem of estimating a linear combina-
tion of the elements of fi, see Section 2.2.3.)
1.2.3 The Mean and Variance of /land 62

Inserting (1.1.4) into (1.2.3), we have
ft= (X'X)IX'y (1.2.15)
= /I+ (X'X)-- IX'u.
Clearly, Eft = fl by the assumptions of Model 1. Using the second line of
(1.2.15), we can derive the variance-covariance matrix of
Vfl E(fifl)($fi)' (1.2.16)
= E(X'X)-1X'eu'X(X'X)-
From (1.2.3) and (1.2.4), we have ill = Mu, where M = I X(X' X)X
Using the properties of the projection matrix given in Theorem 14 of Appen-

dix 1, we obtain
E62 = T- iEu' Mu (1.2.17)
= T 1E tr Muu ' by Theorem 6 of Appendix 1
= T-1a2 tr M
= (T K)o -2 by Theorems 7 and 14 of Appendix 1,
which shows that 6 2 is a biased estimator of a 2. We define the unbiased
estimator of 62 by
62 = (T 11' (1.2.18)
We shall obtain the variance of 6 2 later, in Section 1.3, under the additional
assumption that u is normal.
The quantity V13 can be estimated by substituting either 6 2 or ep (defined
above) for the a 2 that appears in the right-hand side of (1.2.16).
1.2.4 Definition of Best

Before we prove that the least squares estimator is best linear unbiased, we
must define the term best. First we shall define it for scalar estimators, then for
vector estimators.
o
DEFINITION 1.2.1. Let and 0* be scalar estimators of a scalar parameter 0.
The estimator 0 is said to be at least as good as (or at least as efficient as) the
estimator 0* if E(0 0)2 E(0* 0)2 for all parameter values. The estimator
0 is said to be better (or more efficient) than the estimator 0* if 0 is at least as
good as 0* and E(0 0)2 < E(0* 0)2 for at least one parameter value. An
estimator is said to be best (or efficient) in a class if it is better than any other
estimator in the class.
The mean squared error is a reasonable criterion in many situations and is
mathematically convenient. So, following the convention of the statistical
literature, we have defined "better" to mean "having a smaller mean squared
error." However, there may be situations in which a researcher wishes to use
other criteria, such as the mean absolute error.
DEFINITION 1.2.2. Let 0 and 0* be estimators of a vector parameter 0. Let A

and B their respective mean squared error matrices; that is,
A = E(0 0)(0 Or and B = go* eye* 9)'. Then we say 0 is better (or
more efficient) than 0* if

c'(B A)c 0 for every vector c and every parameter value
(1.2.19)
and
c ' (B A)c > 0 for at least one value of c and (1.2.20)
at least one value of the parameter.
This definition of better clearly coincides with Definition 1.2.1 if 0 is a scalar.
In view of Definition 1.2.1, equivalent forms of statements (1.2.19) and
(1.2.20) are statements (1.2.21) and (1.2.22):
c'O is at least as good as c ' 0* for every vector c (1.2.21)
and
c 'a is better than c' o* for at least one value of c. (1.2.22)
Using Theorem 4 of Appendix 1, they also can be written as
B ?_ A for every parameter value
- (1.2.23)
and
B * A for at least one parameter value. (1.2.24)
(Note that B A means B A is nonnegative definite and B > A means
B A is positive definite.)
We shall now prove the equivalence of (1.2.20) and (1.2.24). Because the
phrase "for at least one parameter value" is common to both statements, we
shall ignore it in the following proof. First, suppose (1.2.24) is not true. Then
B = A. Therefore c '(B A)c = 0 for every c, a condition that implies that
(1.2.20) is not true. Second, suppose (1.2.20) is not true. Then c' (B A)c =0
for every c and every diagonal element of B A must be 0 (choose c to be the
zero vector, except for 1 in the ith position). Also, the Oth element of B A is
0 (choose c to be the zero vector, except for 1 in the ith and jth positions, and
note that B A is symmetric). Thus (1.2.24) is not true. This completes the
proof.
Note that replacing B # A in (1.2.24) with B> A or making the corre-
sponding change in (1.2.20) or (1.2.22) is unwise because we could not then
rank the estimator with the mean squared error matrix
higher than the estimator with the mean squared error matrix
[2 0
01
A problem with Definition 1.2.2 (more precisely, a problem inherent in the
comparison of vector estimates rather than in this definition) is that often it
does not allow us to say one estimator is either better or worse than the other.
For example, consider
A = [ 1 1 and B = [ 2 (1.2.25)
01 0 11
Clearly, neither A B nor B ?_ A. In such a case one might compare the
trace and conclude that 0 is better than 9* because tr A < tr B. Another
example is
A = [2 1 and B = [ 2 (1.2.26)
12 0 21
Again, neither A 2-- B nor B A. If one were using the determinant as the
criterion, one would prefer 0 over ON because det A < det B.
Note that B A implies both tr B tr A and det B det A. The first
follows from Theorem 7 and the second from Theorem 11 of Appendix 1. As
these two examples show, neither tr B tr A nor det B det A implies
BA.
Use of the trace as a criterion has an obvious intuitive appeal, inasmuch as it
is the sum of the individual variances. Justification for the use of the determi-
nant involves more complicated reasoning. Suppose 0 N(0, V), where V is
the variance-covariance matrix of 0. Then, by Theorem 1 of Appendix 2,
(0 0)'V -1(0 0) Xi, the chi-square distribution with K degrees of free-
dom, Kbeing the number of elements of 0. Therefore the (1 cx)96 confidence
ellipsoid for 0 is defined by
( 01(ii 0)' (Os 0) < xi(a)), (1.2.27)
where xi(a) is the number such that P [xi &a)] = a. Then the volume of
the ellipsoid (1.2.27) is proportional to the determinant of V, as shown by
Anderson (1958, p. 170).
A more intuitive justification for the determinant criterion is possible for
the case in which 0 is a two-dimensional vector. Let the mean squared error
matrix of an estimator 9 = (01, 02 )' be
az an"
a22
Suppose that 02 is known; define another estimator of 0, by

0, = 0 1 + a(02 02 ). Its mean squared error is a l , + &a22 + 2aa12 and at-
tains the minimum value of a l , (02 /a22 ) when a_ /a22 The larger
a12 , the more efficient can be estimation of 0 1 . Because the larger a 12 implies
the smaller determinant, the preceding reasoning provides a justification for
the determinant criterion.
Despite these justifications, both criteria inevitably suffer from a certain
degree of arbitrariness and therefore should be used with caution. Another
useful scalar criterion based on the predictive efficiency of an estimator will be
proposed in Section 1.6.
1.2.5 Least Squares as Best Linear Unbiased Estimator (BLUE)

The class of linear estimators of fi can be defined as those estimators of the
form C' y for any T X Kconstant matrix C. We can further restrict the class by
imposing the unbiasedness condition, namely,
EC' y=fi for all ft. (1.2.28)
Inserting (1.1.4) into (1.2.28), we obtain
C' X = I. (1.2.29)
Clearly, the LS estimator /I is a member of this class. The following theorem
proves that LS is best of all the linear unbiased estimators.
THEOREM 1.2.1 (Gauss-Markov). Let,* = C' y where C isa TX K matrix
of constants such that C' X = I. Then fl is better than fl * if ft # fi*.
Proof Because /3* = + C'u because of (1.2.29), we have
VIP = EC' uu' C (1.2.30)
=
= a2(X'X)- I + a2[C' (X'X)IX'][C' (X'X) - IXT.
The theorem follows immediately by noting that the second term of the last
line of (1.2.30) is a nonnegative definite matrix.
We shall now give an alternative proof, which contains an interesting point
of its own. The class of linear unbiased estimators can be defined alternatively
as the class of estimators of the form (S ' IS' y, where S is any T X K matrix
of constants such that S' X is nonsingular. When it is defined this way, we call
it the class of instrumental variable estimators (abbreviated as IV) and call the
column vectors of S instrumental variables. The variance-covariance matrix
of IV easily can be shown to be a 2(S' X)- 'S'S(X'S) l. We get LS when we put
S = X, and the optimality of LS can be proved as follows: Because
I S(S'S)- is is nonnegative definite by Theorem 14(v) of Appendix 1, we
have
X'X X'S(S'S)-1 5'X. (1.2.31)
Inverting both sides of (1.2.31) and using Theorem 17 of Appendix 1, we
obtain the desired result:
(X'X)- ' (S'X)-1 S'S(X'S)- I. (1.2.32)
In the preceding analysis we were first given the least squares estimator and
then proceeded to prove that it is best of all the linear unbiased estimators.
Suppose now that we knew nothing about the least squares estimator and had
to find the value of C that minimizes C'C in the matrix sense (that is, in terms
of a ranking of matrices based on the matrix inequality defined earlier) subject
to the condition C' X = I. Unlike the problem of scalar minimization, cal-
culus is of no direct use. In such a situation it is often useful to minimize the
variance of a linear unbiased estimator of the scalar parameter p 'A where p is
an arbitrary K-vector of known constants.
Let c'y be a linear estimator of p'fl. The unbiasedness condition implies
X' c = p. Because Vc ' y = a 2c ' c, the problem mathematically is
Minimize c ' c subject to X ' c = p. (1.2.33)
Define
S= cic 2A'(X'c p), (1.2.34)
where 22 is a K-vector of Lagrange multipliers. Setting the derivative of S with
respect to c equal to 0 yields
c (1.2.35)
Premultiplying both sides by X' and using the constraint, we obtain
(X'X)p. (1.2.36)
Inserting (1.2.36) into (1.2.35), we conclude that the best linear unbiased
estimator of p 'ft is p ' (X' X)-1X' y. We therefore have a constructive proof of

Theorem 1.2.1.
1.3 Model 1 with Normality

In this section we shall consider Model 1 with the added assumption of the
joint normality of u. Because no correlation implies independence under
normality, the (ut ) are now assumed to be serially independent. We shall show
that the maximum likelihood estimator (MLE) of the regression parameter is
identical to the least squares estimator and, using the theory of the Cramer-
Rao lower bound, shall prove that it is the best unbiased estimator. We shall
also discuss maximum likelihood estimation of the variance a 2.
1.3.1 Maximum Likelihood Estimator

Under Model 1 with normality we have y N(V,a 2I). Therefore the likeli-
hood function is given by 3
L= (2=2 )-T/2 exp [ ().5 0-2(y _ 341), (y ... VIA . (1.3.1)
Taking the logarithm of this equation and ignoring the terms that do not
depend on the unknown parameters, we have
T 1
log L = log U 2 - (y - Xfi)' (Y Xii). (1.3.2)
2
Evidently the maximization of log L is equivalent to the minimization of
(y Xii) '(y Xfl), so we conclude that the maximum likelihood estimator of
/3 is the same as the least squares estimator fl obtained in Section 1.2. Putting ft
into (1.3.2) and setting the partial derivative of log L with respect to a 2 equal to
0, we obtain the maximum likelihood estimator of a 2:
6..2 = T- in, (1.3.3)
where fi = y 34. This is identical to what in Section 1.2.1 we called the least
squares estimator of a 2.
The mean and the variance-covariance matrix off/were obtained in Section
1.2.3. Because linear combinations of normal random variables are normal,
we have under the present normality assumption
ii. N[Acr 2 (X' X) -1 ]. (1.3.4)
The mean of 62 is given in Eq. (1.2.17). We shall now derive its variance
under the normality assumption. We have
E(u' Mu)2 = Eu ' Muu' Mu (1.3.5)

= tr (ME[(u'Mu)uu' j)
= a4 tr (M[2M + (tr M)I])
a4 [2 tr M + (tr M)2 ]
a4 [2(T K) + (T K)2 ],
where we have used Ett? = 0 and Eu, = 3a4 since u, N(0, a 2 ).
The third equality in (1.3.5) can be shown as follows. If we write the t,sth
element of M as m, the Oh element of the matrix E(u 'Mu)uu' is given by
mEut us ui uj. Hence it is equal to 2a 4n1/4 if i j and to 2a 4m1, +
a4 II I rn if i = j, from which the third equality follows. Finally, from (1.2.17)
and (1.3.5),
,62 _ 2(T K)cr 4
1
(1.3.6)
T2 '
Another important result that we shall use later in Section 1.5, where we
discuss tests of linear hypotheses, is that
le MU y2 (1.3.7)
which readily follows from Theorem 2 of Appendix 2 because M is indempo-

tent with rank T K. Because the variance ofxi_ x is 2( T K) by Theorem 1
of Appendix 2, (1.3.6) can be derived alternatively from (1.3.7).
1.3.2 Cramer-Rao Lower Bound

The Cramer-Rao lower bound gives a useful lower bound (in the matrix sense)
for the variance-covariance matrix of an unbiased vector estimator. 4 In this
section we shall prove a general theorem that will be applied to Model 1 with
normality in the next two subsections.
THEOREM 1.3.1 (Cramer-Rao). Let z be an n-component vector of random
variables (not necessarily independent) the joint density of which is given by
L(z, 0), where 0 is a K-component vector of parameters in some parameter
space 0. Let 0(z) be an unbiased estimator of 0 with a finite variance-covar-
iance matrix. Furthermore, assume that L(z, 0) and 0(z) satisfy
a log L
ae 0,
(A) E
(B ) E
a2 log L E a log L a log L
awe' ae
(C) E
a log L a log L > 0
ae
(D) oL-n 'dz= I.
f
Note: loga ivae is a K-vector, so that 0 in assumption A is a vector of Kzeroes.
Assumption C means that the left-hand side is a positive definite matrix. The
integral in assumption D is an n-tuple integral over the whole domain of the
Euclidean n-space because z is an n-vector. Finally, I in assumption D is the
identity matrix of size K.
Then we have for 0 E 0
82 10 LT1
V(b) [ E (1.3.8)
Proof Define P = E(0 0)(-6 , Q = E("e exa log Lion and

R= ga log Lioexo log Lme'). Then
[ P Qi - (1.3.9)
Q'R
because the left-hand side is a variance-covariance matrix. Premultiply both
sides of (1.3.9) by [I, Q12-1 ] and postmultiply by [I, QR-1 ] '. Then we get
P QR-1Q ' 0, (1.3.10)
where R- can be defined because of assumption C. But we have
QE m a log Li (1.3.11)
=E [ea log L] by assumption A
=E[
ol aL]
L
= f[Eizi w
alL dz
=
=
I I
ow -
by assumption D.
Therefore (1.3.8) follows from (1.3.10), (1.3.11), and assumption B.

The inverse of the lower bound, namely, E 32 log Lime', is called the
information matrix. R. A. Fisher first introduced this term as a certain mea-
sure of the information contained in the sample (see Rao, 1973, p. 329).
The Cramer-Rao lower bound may not be attained by any unbiased esti-
mator. If it is, we have found the best unbiased estimator. In the next section
we shall show that the maximum likelihood estimator of the regression pa-
rameter attains the lower bound in Model 1 with normality.
Assumptions A, B, and D seem arbitrary at first glance. We shall now
inquire into the significance of these assumptions. Assumptions A and B are
equivalent to
(A') dz = 0
80
32L
(B') dz 0,
respectively. The equivalence between assumptions A and A' follows from
E
d log
ae
L E aLIJ (1.3.12)
= Iii%1-t1 ]L, dz
= aLdz.
do
f
The equivalence between assumptions B and B' follows from
82 log L _ E [ 1 aLi} (1.3.13)
awe' ae L ae'
aL aL 1 82L 1
E L- L2 ae ae' L aeae'
=E
r L
log a log L] I
ae ae' +
8 21, dz.
awe/
Furthermore, assumptions A', B', and D are equivalent to
aL dz =
(A") I a L dz
60 ae
(B")
f a2L30 , dz _ TO
a f al,
w dz
(D') 7 , LO dz,
because the right-hand sides of assumptions A", B", and D' are 0, 0, and!,
respectively. Written thus, the three assumptions are all of the form
a a
J.1-(z 0) dz = f f(z, 0) dr
ao ' (1.3.14)
in other words, the operations of differentiation and integration can be inter-

changed. In the next theorem we shall give sufficient conditions for (1.3.14) to
hold in the case where 0 is a scalar. The case for a vector 0 can be treated in
essentially the same manner, although the notation gets more complex.
THEOREM 1.3.2. If (i) Of(z, ewe is continuous in 0 E 0 and z, where 0 is an
open set, (ii) ff(z, 0) dz exists, and (iii) f I af(z, ewe I dz < M < co for all
E
0 0, then (1.3.14) holds.
Proof We have, using assumptions (i) and (ii),
1 1 [ f(z, 0 + hh) f(z, 0)
0 (z, 0)] dz 1
a(9f
, I 1 f(z, 0 + hh) f(z, 0)

,a9-1.0 (z, 0) 1 dz
= f1 iofo(z, 0*) -24 (z, 0)1 dz,
where 0* is between 0 and 0 + h. Next, we can write the last integral as

1= fA + fA , where A is a sufficiently large compact set in the domain of z and
A is its complement. But we can make IA sufficiently small because of (i) and
f; sufficiently small because of (iii).
1.3.3 Least Squares Estimator as Best Unbiased Estimator (BUE)

In this section we shall show that under Model 1 with normality, the least
squares estimator of the regression parameters fl attains the Cramer-Rao lower
bound and hence is the best unbiased estimator. Assumptions A, B, and C of
Theorem 1.3.1 are easy to verify for our model; as for assumption D, we shall
-
verify it only for the case where /3 is a scalar and a 2 is known, assuming that the
parameter space of 11 and a2 is compact and does not contain a 2 =0.
Consider log L given in (1.3.2). In applying Theorem 1.3.1 to our model, we
put 0' = (fl' , a2 ). The first- and second-order derivatives of log L are given by
a log L 1 (X'y X'X/I), (1.3.15)
a8/3 0.2
a log L T 1,
(1.3.16)
a(a2 ) 2a2 2a4 XfIr ( Y XfA
82 log L _ 1 ,
7,i x X, (1.3.17)
8/38/3'
82 log L T 1
80.2)2 - 2a4 a6 X/0'(Y (1.3.18)
82 log L 1 ,
afia(a2) a4 (X y X' X/I). (1.3.19)
From (1.3.15) and (1.3.16) we can immediately see that assumption A is

satisfied. Taking the expectation of (1.3.17), (1.3.18), and (1.3.19), we obtain
1 vv
log L a2
E (1.3.20)
awe' Tl .
0
2a4
We have
E
a log L a log L
1
EX'uu'X =X'X ,
1
(1.3.21)
al) air a4
[a log 1 2 T2 ,-r1.tku T ,
E . ur u (1.3.22)
8072) 404 40 2a6
2 a4
because E(u'u)2 = (T2 + 2T) a4, and
a log L a log L
E a(c2 ) (1.3.23)
Oft
Therefore, from (1.3.20) to (1.3.23) we can see that assumptions B and C are
both satisfied.
We shall verify assumption D only for the case where fi is a scalar and a 2 is
known so that we can use Theorem 1.3.2, which is stated for the case of a scalar
parameter. Take I3L, as the jof that theorem. We need to check only the last
condition of the theorem. Differentiating (1.3.1) with respect to /3, we have
1
afi = fi--
0.2 (x' y /3x' x)L, (1.3.24)
where we have written X as x, inasmuch as it is a vector. Therefore, by Holder's

inequality (see Royden, 1968, p. 113),
f Illafl ' (1I-2 Y

1/2
dy ::.5 (3.
- 3 L dy Li (x' y fix'x)2L dy] . (1.3.25)
The first integral on the right-hand side is finite because fl is assumed to have a
finite variance. The second integral on the right-hand side is also finite because
the moments of the normal distribution are finite. Moreover, both integrals
are uniformly bounded in the assumed parameter space. Thus the last condi-
tion of Theorem 1.3.2 is satisfied.
Finally, from (1.3.8) and (1.3.20) we have
Vfr cr 2(X' X) (1.3.26)
for any unbiased fr. The right-hand side of (1.3.26) is the variance-covariance
matrix of the least squares estimator of fl, thereby proving that the least
squares estimator is the best unbiased estimator under Model 1 with normal-
ity. Unlike the result in Section 1.2.5, the result in this section is not con-
strained by the linearity condition because the normality assumption was
added. Nevertheless, even with the normality assumption, there may be a
biased estimator that has a smaller average mean squared error than the least
squares estimator, as we shall show in Section 2.2. In nonnormal situations,
certain nonlinear estimators may be preferred to LS, as we shall see in Sec-
tion 2.3.
1.3.4 The Cramer-Rao Lower Bound for Unbiased Estimators of a 2

From (1.3.8) and (1.3.20) the Cramer-Rao lower bound for unbiased estima-
tors of a2 in Model 1 with normality is equal to 2a 4 T -1 . We shall examine
whether it is attained by the unbiased estimator 62 defined in Eq. (1.2.18).
Using (1.2.17) and (1.3.5), we have
2a-4
Vii 2 (1.3.27)
TIC
Therefore it does not attain the Cramer-Rao lower bound, although the differ-
ence is negligible when T is large.
We shall now show that there is a simple biased estimator of cr 2 that has a
smaller mean squared error than the Cramer-Rao lower bound. Define the
class of estimators
,;.2 - fi ' fi (1.3.28)
`" N- N'
where N is a positive integer. Both 62 and 62, defined in (1.2.5) and (1.2.18),
respectively, are special cases of (1.3.28). Using (1.2.17) and (1.3.5), we can
evaluate the mean squared error of 6-1,as
EA 0.2)2 _ 2(T K)+ (T K N)2 0. 4
. (1.3.29)
N2
By differentiating (1.3.29) with respect to N and equating the derivative to
zero, we can find the value of N that minimizes (1.3.29) to be
N* = T K+ 2. (1.3.30)
Inserting (1.3.30) into (1.3.29), we have
Ecca. _ 0.2)2 - 264 (1.3.31)
' T K + 2'
which is smaller than the Cramer-Rao bound if K = 1.
1.4 Model 1 with Linear Constraints

In this section we shall consider estimation of the parameters fi and 0-2 in
Model 1 when there are certain linear constraints on the elements of A We
shall assume that the constraints are of the form
Q'fl = c, (1.4.1)
where Q is a K X q matrix of known constants and e is a q-vector of known
constants. We shall also assume q < K and rank (Q) = q.
Equation (1.4.1) embodies many of the common constraints that occur in
practice. For example, if Q' = (I, 0) where I is the identity matrix of size K 1
1 X K2 matrix of zeroes such that IC 1 + IC2 = K, then the con-and0istheK
straints mean that the elements of a K 1 -component subset of/3 are specified to
be equal to certain values and the remaining K2 elements are allowed to vary
freely. As another example, the case in which Q' is a row vector of ones and
c = 1 implies the restriction that the sum of the regression parameters is unity.
The study of this subject is useful for its own sake in addition to providing
preliminary results for the next section, where we shall discuss tests of the
linear hypothesis (1.4.1). We shall define the constrained least squares estima-
tor, present an alternative derivation, show it is BLUE when (1.4.1) is true,
and finally consider the case where ( 1.4. 1) is made stochastic by the addition of
a random error term to the right-hand side.
1.4.1 Constrained Least Squares Estimator (CLS)

The constrained least squares estimator (CLS) fp, denoted by ii, is defined to
be the value of fl that minimizes the sum of squared residuals
SM = (Y - Vi)' (Y - V) (1.4.2)
under the constraints (1.4. 1). In Section 1.2.1 we showed that (1.4.2) is mini-
mized without constraint at the least squares estimator fl. Writing SA for the
sum of squares of the least squares residuals, we can rewrite (1.4.2) as
suo = sCifo + (fi fly vx(11 A. (1.4.3)
Instead of directly minimizing (1.4.2) under (1.4.1), we minimize (1.4.3)
under (1.4.1), which is mathematically simpler. ..,
Put fl - /I = 5 and Q 'fl - c = y. Then, because S(/1) does not depend on 13,
the problem is equivalent to the minimization of 5' X' XS under Q ' (5 = y.
Equating the derivatives of 5' X' Xe5 + 2A' (Q ' 5 - 7) with respect too and the
q-vector of Lagrange multipliers A to zero, we obtain the solution
6 = (X' X)--1 Q[Q ' (X ' X)-- V]'y. (1.4.4)
Transforming from S and y to the original variables, we can write the mini-
mizing value /I of S(ft) as
fl = ft - (X' X) - V[Q ' (X' X)- i()]-1 (Q 'ft - c). (1.4.5)
The corresponding estimator of a 2 can be defined as
a.2 = 7-1 (Y -Xi3-)' (Y - xii). (1.4.6)
It is easy to show that the ii and F2 are the constrained maximum likelihood
estimators if we assume normality of u in Model 1.
1.4.2 An Alternative Derivation of the Constrained Least

Squares Estimator
Define aK X (K- q) matrix R such that the matrix (Q, R) is nonsingular and
R'Q =0. Such a matrix can always be found; it is not unique, and any matrix
that satisfies these conditions will do. Then, defining A = (Q, R)', y = Aii,
and Z = XA-1 , we can rewrite the basic model (1.1.4) as
y = Xi? + u (1.4.7)
= XA-1Afl + u
= Zy + u.
If we partition y = (yl, y)', where y l = Q'fl and 72 = R'fl, we see that the
constraints (1.4.1) specify y l and leave 72 unspecified. The vector 72 has K- q
elements; thus we have reduced the problem of estimating K parameters
subject to q constraints to the problem of estimating K- q free parameters.
Using yi = c and
A.- ' = [Q(Q'Q) - ', R(R'R)-1 j, (1.4.8)
we have from (1.4.7)
y - XQ(Q'Q)- Ic = XR(R'R)iy2 + u. (1.4.9)
Let 512 be the least squares estimator of 72 in (1.4.9):
51'2 = R ' R(R' X ' XR)- 'It' X ' [y - XQ(Q ' Q)'c]. (1.4.10)
Now, transforming from y back to /I by the relationship fl = .21 - '7, we obtain
the CLS estimator of /3:
= R(R'X'XR)- IR'X'y (1.4.11)
[I - R(R ' X ' XR) - 'It' X ' X]Q(Q ' Q)'c.
Note that (1.4.11) is different from (1.4.5). Equation (1.4.5) is valid only if
X 'X is nonsingular, whereas (1.4.11) can be defined even if X 'X is singular
provided that R' X' XR is nonsingular. We can show that if X' X is nonsingu-
lar, (1.4.11) is unique and equal to (1.4.5). Denote the right-hand sides of
(1.4.5) and (1.4.11) by fl i and A, respectively. Then it is easy to show
Q '
[R'X'l
(fli P2) 0. (1.4.12)
Therefore A = A if the matrix in the square bracket above is nonsingular. But

we have
[ ,
R'x'xi R, Q] = [X 12 1 X0' R R'X'XQ
Q Q'Q I (1.4.13)
where the matrix on the right-hand side is clearly nonsingular because non-
singularity of X ' X implies nonsingularity of R'X'XR. Because the matrix

[R, Q] is nonsingular, it follows that the matrix
r t'x'x]
Q'
is nonsingular, as we desired.
1.4.3 Constrained Least Squares Estimator as Best Linear

Unbiased Estimator
A
That is the best linear unbiased estimator follows from the fact that P2 is the
best linear unbiased estimator of y 2 in (1.4.9); however, we also can prove it
directly. Inserting (1.4.9) into (1.4.11) and using (1.4.8), we obtain
ft - - II + R(R'X'XR)-1 11'X'u. (1.4.14)
Therefore, # is unbiased and its variance-covariance matrix is given by
V(17) = a 2R(R ' X ' XR)- '12'. (1.4.15)
We shall now define the class of linear estimators by r =C'y d where C' is
aKX T matrix and d is a K-vector. This class is broader than the class of linear
estimators considered in Section 1.2.5 because of the additive constants d. We
did not include d previously because in the unconstrained model the unbi-
asedness condition would ensure d = 0. Here, the unbiasedness condition
d) = ft implies C' X = I + GQ' and d = Gc for some arbitrary K X q
matrix G. We have V(r) = a2C 'C as in Eq. (1.2.30) and CLS is BLUE
because of the identity
C/C R(R'X'XR)- 'R' (1.4.16)
= [C' R(R'X'XR)-- ill'X'][C' R(R'X'XR) - IR'XT,
where we have used C' X = I + GQ' and R'Q = 0.
1.4.4 Stochastic Constraints

Suppose we add a stochastic error term on the right-hand side of (1.4.1),
namely,
Q'fl = c + v, (1.4.17)
where visa q-vector of random variables such that Ev = 0 and Evv ' =1.21. By
making the constraints stochastic we have departed from the domain of clas-
sical statistics and entered that of Bayesian statistics, which treats the un-
known parameters fi as random variables. Although we generally adopt the
classical viewpoint in this book, we shall occasionally use Bayesian analysis
whenever we believe it sheds light on the problem at hand. 5
In terms of the parameterization of (1.4.7), the constraints (1.4.17) are
equivalent to
71 =-- c + v. (1.4.18)
We shall first derive the posterior distribution of 7, using a prior distribution
over all the elements of y, because it is mathematically simpler to do so; and
then we shall treat (1.4.18) as what obtains in the limit as the variances of the
prior distribution for 72 go to infinity. We shall assume a 2 is known, for this
assumption makes the algebra considerably simpler without changing the
essentials. For a discussion of the case where a prior distribution is assumed on
a2 as well, see Zellner (1971, p. 65) or Theil (1971, p. 670).
Let the prior density of 7 be
f(y) = (2n)-K/ 2 1111 -1 /2 exp [ (1/2)(y /.t)' srl(y p)], (1.4.19)
where fl is a known variance-covariance matrix. Thus, by Bayes's rule, the
posterior density of y given y is
.RY YV(Y) (1.4.20)
f(I)
.11(y17)./(7)dv
= c1 exp ((1/2)[cr -2 (y Zy)' (y Zy)
+ (y PA),
where c l does not depend on y. Rearranging the terms inside the bracket, we
have
a-2(y ZY)'(Y ZY) + (y fr i(Y (1.4.21)
= y'(o.-2Z'Z + crl)y 2(a -2y'Z + p'11-1 )y + cr-2y'y + Aeirly
= (y P)'(c)---2Z'Z + 0-1 )(y rf, ) j)'(cr -2Z'Z + LI-1 W
y Er 14 ,
where
(0-2z,z fr1)i( 0-2v y frIp). (1.4.22)
Therefore the posterior distribution of y is
YI Y IsT(T', [a-2Z' Z + I ]), (1.4.23)
and the Bayes estimator of >' is the posterior mean i given in (1.4.22). Because
7 = A/B, the Bayes estimator of /3 is given by
fl= ii- i[a-2(A' ) IX' XA- ' + CI- ' ]' [a -2(A ')- IX' y + 0- '0]
= (o.-2 X 1 X + ATI- lA)(a-2X' y + A 'Ll- 1/). (1.4.24)
We shall now specify p and Ca so that they conform to the stochastic con-
straints (1.4.18). This can be done by putting the first q elements ofp equal toe
(leaving the remaining K q elements unspecified because their values do not
matter in the limit we shall consider later), putting
[r2I 0
fl = (1.4.25)
0 v2 I1
and then taking v2 to infinity (which expresses the assumption that nothing is a
priori known about 72 ). Hence, in the limit we have
[ -2i,
i 0
iim Sr' = (1.4.26)
v--- 0 0j
Inserting (1.4.26) into (1.4.24) and writing the first q elements of p as c, we
finally obtain
fl= (x , x + ) 2QQ , y 1 (x , y + 22Qc) , (1.4.27)
where A2 .. 0.2tr 2 .
We have obtained the estimator /1 as a special case of the Hayes estimator,
but this estimator was originally proposed by Theil and Goldberger (1961) and
was called the mixed estimator on heuristic grounds. In their heuristic ap-
proach, Eqs. (1.1.4) and (1.4.17) are combined to yield a system of equations
Y X a _i_ u (1.4.28)
[Ac] ---- [AO] P 1 [Avi
Note that the multiplication of the second part of the equations by A renders
the combined error terms homoscedastic (that is, constant variance) so that
(1.4.28) satisfies the assumptions of Model 1. Then Theil and Goldberger
proposed the application of the least squares estimator to (1.4.28), an opera-
tion that yields the same estimator as/igiven in (1.4.27). An alternative way to
interpret this estimator as a Bayes estimator is given in Theil (1971, p. 670).
There is an interesting connection between the Bayes estimator (1.4.27) and
the constrained least squares estimator (1.4.11): The latter is obtained as the
limit of the former, taking A 2 to infinity. Note that this result is consistent with
our intuition inasmuch as ,1.2 , co is equivalent to r2 * 0, an equivalency that
implies that the stochastic element disappears from the constraints (1.4.17),
thereby reducing them to the nonstochastic constraints (1.4.1). We shall dem-
onstrate this below. [Note that the limit of (1.4.27) as 2 2 , oo is not
(QQ1-1Qc because QQ' is singular; rather, the limit is a K X K matrix with
rank q<K.]
Define
B A(2-2X'X + QQ9A' (1.4.29)
p,--2Q'x'xQ+Q'QQ'Q A-2Q'x'XR1
A-2R'x'xQ A-2R'x'xitr
Then, by Theorem 13 of Appendix 1, we have
(R'X'XR)-1 R'X'XQE-1
(1.4.30)
where
E = Q'QQ/Q + 2-2Q'X'XQ (1.4.31)
.1.-2Q'X'XR(R'X'XR) -1 R'X'XQ
and
F = .1-2R'X'XR (1.4.32)
2-4R'X'XQ(2-2Q'X'XQ + Q'QQ'Q)-1Q'X'XR.
From (1.4.27) and (1.4.29) we have
A13-1A(2-2X'y + Qc). (1.4.33)
Using (1.4.30), we have
urn A'13-1 A(A-2X'y) = R(R'X'XR)-1 R'X'y (1.4.34)
and
1 l
urn
lim A'B-1AQc = lim (Q, R)B-1 (1.4.35)
= Q(Q'Q)-1c R(R ' X ' XR)-112' X' XQ(Q'Q)-1c.

Thus we have proved
lim fr =, (1.4.36)
where fi is given in (1.4.11).

1.5 Test of Linear Hypotheses

In this section we shall regard the linear constraints (1.4.1) as a testable hy-
pothesis, calling it the null hypothesis. Throughout the section we shall as-
sume Model 1 with normality because the distributions of the commonly used
test statistics are derived under the assumption of normality. We shall discuss
the t test, the F test, and a test of structural change (a special case of the Ftest).
1.5.1 The t Test

The t test is an ideal test to use when we have a single constraint, that is, q = 1.
The F test, which will be discussed in the next section, will be used if q> 1.
Because fl is normal, as shown in Eq. (1.3.4), we have
Q'fi N[c, a 2Q' (X' X) -1()] (1.5.1)
under the null hypothesis (that is, if Q = c). With q = 1, Q' is a row vector
and c is a scalar. Therefore
Qin c
[0.2Q, pc, 30 _ IQ] 1/2
N(0, 1). (1.5.2)
This is the test statistic one would use if ci were known. As we have shown in
Eq. (1.3.7), we have
6'6 y2 (1.5.3)
2 " TK
0
The random variables (1.5.2) and (1.5.3) easily can be shown to be indepen-
dent by using Theorem 6 of Appendix 2 or by noting DV' = 0, which implies
that 0 and flare independent because they are normal, which in turn implies
that u'fl and fl are independent. Hence, by Theorem 3 of Appendix 2 we have
Q c
ST-K, (1.5.4)
[ii 2Q' (x,x)-min
which is Student's t with T K degrees of freedom, where a is the square root
of the unbiased estimator of a 2 defined in Eq. (1.2.18). Note that the denomi-
nator in (1.5.4) is an estimate of the standard deviation of the numerator. Thus
the null hypothesis Q'fl = c can be tested by the statistic (1.5.4). We can use a
one-tail or two-tail test, depending on the alternative hypothesis.
In Chapter 3 we shall show that even if u is not normal, (1.5.2) holds
asymptotically (that is, approximately when T is large) under general condi-
tions. We shall also show that 3.2 converges to a 2 in probability as T goes to
infinity (the exact definition will be given there). Therefore under general
distributions of u the statistic defined in (1.5.4) is asymptotically distributed as
N(0, 1) and can be used to test the hypothesis using the standard normal
table. 6 In this case a 2 may be used in place of a2 because 461..2 also converges to (7 2
inprobalty.
1.5.2 The F Test

In this section we shall consider the test of the null hypothesis Q c against
the alternative hypothesis Q13 c when it involves more than one constraint
(that is, q> 1).
We shall derive the F test as a simple transformation of the likelihood ratio
test. Suppose that the likelihood function is in general given by L(, 0), where
is a sample and 0 is a vector of parameters. Let the null hypothesis be Ho :
6 E So where So is a subset of the parameter space 0 and let the alternative
hypothesis be HI : 0 e S, where SI is another subset of 0. Then the likelihood
ratio test is defined by the following procedure:
max L(, 0)
11E4
Reject Ho if 2 < g, (1.5.5)
max La , 0)
eEsous,
where g is chosen so as to satisfy P(A < g) = a for a given significance level a.
The likelihood ratio test may be justified on several grounds: (1) It is intui-
tively appealing because of its connection with the Neyman-Pearson lemma.
(2) It is known to lead to good tests in many cases. (3) It has asymptotically
optimal properties, as shown by Wald (1943).
Now let us obtain the likelihood ratio test of Q1 = c against Q1 # c in
Model 1 with normality. When we use the results of Section 1.4.1, the numer-
ator likelihood becomes
max L = (27r17 2 )- exp [ 0.5a-2(y XA') (y V)] (1.5.6)
QP-c
(2 2 ) T/2 e- T/2.
Because in the present case So U Si is the whole parameter space of ft. the
maximization in the denominator of (1.5.5) is carried out without constraint.
Therefore the denominator likelihood is given by
max L = (27r612 )_ Tn exp [ 0.56-2(y X./i)' (y Xi)] (1.5.7)
(2x6 2 )_ Tn e- TI2
Hence, the likelihood ratio test of Q'fi = c is

S)O 1- 772
Reject Ql= c if A. = [-- . < g, (1.5.8)
S(11)
where g is chosen as in (1.5.5).
We actually use the equivalent test
Reject Q73 = c if Pi
T K S(A). .SCA > h, (1.5.9)
q S(A
where h is appropriately chosen. Clearly, (1.5.8) and (1.5.9) can be made
equivalent by putting h = [(T K)/q]((-2IT 1). The test (1.5.9) is called the
F test because under the null hypothesis n is distributed as F(q, T K) the
F distribution with q and T K degrees of freedom as we shall show next.
From (1.4.3) and (1.4.5) we have
S(13) S(k= (Q 'fi
c)' [Q' (X ' X)-- ' Q]' (Q73 c).
(1.5.10)

But since (1.5.1) holds even if Q is a matrix, we have by Theorem 1 of

Appendix 2
S(fl) Sth y2
(1.5.11)
4"
a 1,
Because this chi-square variable is independent of the chi-square variable

given in (1.5.3) by an application of Theorem 5 of Appendix 2 (or by the
independence of ii and //), we have by Theorem 4 of Appendix 2
T K (Q11 c)' [(r(X'X)_'Qri(Q71 c) F(q, T K).
q 11'6
(1.5.12)
We shall now give an alternative motivation for the test statistic (1.5.12).
For this purpose, consider the general problem of testing the null hypothesis
110 : 0= 00 , where 0= (00 02)' is a two-dimensional vector of parameters.
Suppose we wish to construct a test statistic on the basis of an estimator 6 that
is distributed as N(00 , V) under the null hypothesis, where we assume V to be a
known diagonal matrix for simplicity. Consider the following two tests of 1/ 0 :
Reject 1/0 if (6 ti ny (6 0)> c (1.5.13)
and
Reject 1/0 if (i). 00'1" ( 0 0 ) > d, (1.5.14)
where c and d are determined so as to make the probability of Type I error

equal to the same given value for each test. That (1.5.14) is preferred to
(1.5.13) can be argued as follows: Let al and al be the variances of 0 1 and 02 ,
respectively, and suppose ai < ai . Then a deviation of 01 from 010 provides a
greater reason for rejecting 1/0 than the same size deviation of 0 2 from 020
02 rather than by falseness becausthlrodbecausyviltof
of I/0 . The test statistic (1.5.14) precisely incorporates this consideration.
The test (1.5.14) is commonly used for the general case of a q-vector Band a
general variance-covariance matrix V. Another attractive feature of the test is
that
- 130 - ( - ) (1.5.15)
Applying this test to the test of the hypothesis (1.4.1), we readily see that the
test should be based on
cY[QPCX)-1 (2]-1(0 c) 2
(1.5.16)
Xe
a2
This test can be exactly valid only when a 2 is known and hence is analogous to
the standard normal test based on (1.5.2). Thus the Ftest statistic (1.5.12) may
be regarded as a natural adaptation of (1.5.16) to the situation where a 2 must
be estimated. (For a rigorous mathematical discussion of the optimal proper-
ties of the F-test, the reader is referred to Scheffe, 1959, p. 46.)
By comparing (1.5.4) and (1.5.12) we immediately note that if q = I (and
therefore Q' is a row vector) the F statistic is the square of the t statistic. This
fact clearly indicates that if q = 1 the t test should be used rather than the Ftest
because a one-tail test is possible only with the t test.
As stated earlier, (1.5.1) holds asymptotically even if u is not normal. Be-
cause a-2 = filAT K) converges to a2 in probability, the linear hypothesis
can be tested without assuming normality of u by using the fact that or is
asymptotically distributed as 4. Some people prefer to use the F test in this
situation. The remark in note 6 applies to this practice.
The F statistic Pi given in (1.5.12) takes on a variety of forms as we insert a
variety of specific values into Q and c. As an example, consider the case where
/3 is partitioned as //' = (fl, A), where /31 is a Krvector and fl2 is a K2-vector
such that K1 + K2 = K, and the null hypothesis specifies /12 = /12 and leaves /1 1
unspecifd.ThyotsanbewrithfomQ1=cbypuing
Q' = (0, I), where 0 is the K2 X K1 matrix of zeros and I is the identity matrix
of size K2, and by putting c = /12 . Inserting these values into (1.5.12) yields
T- K (132 - A)[R 1)(rX)-1(0, In -1(fi2 - $2) (1.5.17)

K2 Wu
- F(K2 , T - K).
We can write (1.5.17) in a more suggestive form. Partition X as X = (X 1 , X2)
conformably with the partition offl and define M I = I - XI( XIX1 )-1V, . Then
by Theorem 13 of Appendix 1 we have
[A IXVX)-1 (0, I)']- I ----- 3qM1X2 . (1.5.18)
Therefore, using (1.5.18), we can rewrite (1.5.17) as
T- K (1j2 - krXiMIX242 - T32)
71 - ir F(K2 , T- K). (1.5.19)
....2 fil
Of particular interest is the further special case of (1.5.19), where K 1 = 1, so
that 131 is a scalar coefficient on the first column of X, which we assume to be
the vector of ones (denoted by 1), and where /1 2 =0. Then M i becomes L
I - T-111. Therefore, using (1.2.13), we have
$2 = ( XPLX2)-1 Xg,y, (1.5.20)
so (1.5.19) can now be written as
T - K y'LX2(XJ,X2)- iXLy
77 - K - 1 F(K - 1, T - K). (1.5.21)
il'fi
Using the definition of R 2 given in (1.2.9), we can rewrite (1.5.21) as
T - K R2
- -11-R2 F(K - 1, T - K), (1.5.22)
71K
since fi'fi = y'Ly - y'LX2(XZLX2)-1 3qLy because of Theorem 15 of Appen-
dix 1. The value of statistic (1.5.22) usually appears in computer printouts and
is commonly interpreted as the test statistic for the hypothesis that the popula-
tion counterpart of R 2 is equal to 0.
1.5.3 A Test of Structural Change when Variances Are Equal
Suppose we have two regression regimes
yi = XI fil + u 1 (1.5.23)
and
Y2 = X2P2 + U2 , (1.5.24)
where the vectors and matrices in (1.5.23) have T 1 rows and those in (1.5.24)
T2 rows, X 1 is a Ti X K* matrix and X2 is a T2 X K* matrix, and u 1 and u2 are
normally distributed with zero means and variance-covariance matrix
E[ U2
111 ](uc, u) =[(711T'
0 aiIi .
We assume that both X 1 and X2 have rank equal to /0.7 We want to test the
null hypothesis ii, = /32 assuming al = ai(=--- a2) in the present section and
ai # ai in the next section. This test is especially important in econometric
time series because the econometrician often suspects the occurrence of a
structural change from one era to another (say, from the prewar era to the
postwar era), a change that manifests itself in the regression parameters. When
al = a3, this test can be handled as a special case of the standard F test
presented in the preceding section.
To apply the F test to the problem, combine Eqs. (1.5.23) and (1.5.24) as
y = Xfl + u, (1.5.25)
where
y _ [yi ] , x _ [Xi 0 a flz and u = 1111
Y2 0 X] ' 1' Li2 i ' u2 '
Then, since oi = a( c2), (1.5.25) is the same as Model 1 with normality;
hence we can represent our hypothesis fl, =fl2 as a standard linear hypothesis
on Model 1 with normality by putting T = T1 + T2, K = 2K*, q = K*, Q' --
(I, I), and c = 0. Inserting these values into (1.5.12) yields the test statistic
(T 1 + T2 2K*) (fti ft2)1(XIX1) -1 + (XiX2) - 'NA ft2)
1 K* y'[I X( X'20-1Xly
F(K*, T 1 + T2 2K*), (1.5.26)
where A = oc,x,)--ixly, and A = (x2(2)- ixy2.
We shall now give an alternative derivation of (1.5.26). In (1.5.25) we
combined Eqs. (1.5.23) and (1.5.24) without making use of the hypothesis
fli = /32 . If we make use of it, we can combine the two equations as
y -- X/3+ u, (1.5.27)
where we have defined X = (XI, Xi)' and fl = fli = A . Let S(ft) be the sum of
squared residuals from (1.5.25), that is,
S(ft) -- ylI X(X'X) - lMy (1.5.28)
and let S(ii) be the sum of squared residuals from (1.5.27), that is,
S(fi) = y/I X(X'X) - IX/y. (1.5.29)
Then using (1.5.9) we haves
T, + T2 - 2K* SO) ,:g/1) Ric , Ti + T2 - 2K*).
?I K* S(II)
(1.5.30)
To show the equivalence of (1.5.26) and (1.5.30), we must show

S(P) SU-3') = (iii fi2)1(XIX,) 1 + (XiX2r il- 1 CA fi2).
(1.5.31)
From (1.5.29) we have
Rip= y' [I [ X
x21 ] (XIX, + XiX2)-1 (XI , Xi)] (1.5.32)
and from (1.5.28) we have
[X,(X'A lr'X', 0
S(fi) = y' [I (1.5.33)
0 X2(XX2) -1Xil] Y
= YI[I X1(XIX1)-1Xi]Yi + Yi[I X2(XiX2)-IXi]Y2-

Therefore, from (1.5.32) and (1.5.33) we get
SO-I) S(/) (1.5.34)
= (YIX,, Y2X2)
x 1(XX1 )- ' (X'A i + X2C2)-1 (xpc, + Wc2)-1 i
. (XIX, + XiX2) -1 1 (XPE2)- ' - (XIX, + XPC2)- 1
X [NY]
3qY2
= (YIX1, Y2X2)
X [ " ,] [(XIXi) + (XiX2) 1 i-I RXiX1r 1 , (X2(2)-1 ]
X XCY'l
[
,q3r 2 ..1 '
the last line of which is equal to the right-hand side of (1.5.31). The last
equality of (1.5.34) follows from Theorem 19 of Appendix 1.
The hypothesis ft1 = ft2 is merely one of many linear hypotheses we can
impose on theft of the model (1.5.25). For instance, we might want to test the
equality of a subset of /1 1 with the corresponding subset of /12 . If the subset
consists of the first Kr elements of both/31 and ft2 , we should put T T 1 + T2
and K= 2K* as before, but q= K Q' = (I, 0, I, 0), and c =0 in the for-
mula (1.5.12).
If, however, we wish to test the equality of a single element of /1 1 with the
corresponding element of/I2 , we should use the t test rather than the F test for
the reason given in Section 1.5.2. Suppose the null hypothesis is flii = /32;
where these are the ith elements of/hand/1 2 , respectively. Let the ith columns
of X1 and X2 be denoted by xi; and x2; and let X 1(j) and X2(i) consist of the
remaining K* 1 columns of X I and X2, respectively. Define M 1(1) = I
X i (0( Vic/Alcor 'rim, = M mxii , and 51, = M1(f)y1 and similarly define
M2(0/ 221y and 512 Then using Eqs. (1.2.12) and (1.2.13), we have
/4
- =hr
IJ 1 (P 1 i (1.5.35)
AL I 11 AliAli
and
2
5125'2 472
(1.5.36)
)921 = N (fl2s/ k)
221121
Therefore, under the null hypothesis,
P11 - P21
N(0, 1). (1.5.37)
ai ai ) 1/2
ilitz)
Also, by Theorem 2 of Appendix 2

31M1n YZM2Y2 2 (1.5.38)
2 ATI+ T2-210'
(7 2
,
1 a2
where M i = I XI( 'Wi l ) 'XI and M2 = I X2(XPC2) I X;. Because

(1.5.37) and (1.5.38) are independent, we have by Theorem 3 of Appendix 2
(iiii 'i;i)( T 1 + T2 2K*) I12

STI -1-T2-2K (1.5.39)
(a _ a
ai )1/2(31M1Y
1
+
YZNI2Y2)1/2
K iiiiii A/ ai a 22 i
Putting a? = al in (1.5.39) simplifies it to

- 132i (1.5.40)
1/2 ST1+ T2-2X 0 9
iiiiu
where o2 is the unbiased estimate of a 2 obtained from the model (1.5.25),
that is,
1
a 2K* y'[I X(X'X)-1 Xly. (1.5.41)
T1 + T2-
1.5.4 A Test of Structural Change when Variances Are Unequal

In this section we shall remove the assumption a?= al and shall study how to
test the equality offi l and /32 . The problem is considerably more difficult than
the case considered in the previous section; in fact, there is no definitive
solution to this problem. Difficulty arises because (1.5.25) is no longer Model
1 because of the heteroscedasticity of u. Another way to pinpoint the difficulty
is to note that cri and ai do not drop out from the formula (1.5.39) for the t
statistic.
Before proceeding to discuss tests of the equality of fl and /12 when a?
we shall first consider a test of the equality of the variances. For, if the hypoth-
esis a? = al is accepted, we can use the F test of the previous section. The null
hypothesis to be tested is a? a 2). Under the null hypothesis we have
Y1MIYI (1.5.42)
a2 4 TI-K
and
Y2M2Y2
2 X 2 K'. (1.5.43)
a
Because these two chi-square variables are independent by the assumptions of
the model, we have by Theorem 4 of Appendix 2
T2 - K * Y1M1Y 1
Fa K*9 (1.5.44)
T1 K* YiM2Y2 T2- f(*).
Unlike the F test of Section 1.5.2, a two-tailed test should be used here because
either a large or a small value of (1.5.44) is a reason for rejecting the null
hypothesis.
Regarding the test of the equality of the regression parameters, we shall

consider only the special case considered at the end of Section 1.5.3, namely,
the test of the equality of single elements, Ai where the t test is applica-
ble. The problem is essentially that of testing the equality of two normal means
when the variances are unequal; it is well known among statisticians as the
Behrens Fisher problem. Many methods have been proposed and others are
-
still being proposed in current journals; yet there is no definitive solution to

the problem. Kendall and Stuart (1979, Vol. 2, p. 159) have discussed various
methods of coping with the problem. We shall present one of the methods,
which is attributable to Welch (1938).
As we noted earlier, the difficulty lies in the fact that one cannot derive
(1.5.40) from (1.5.39) unless ai = a. We shall present a method based on the
assumption that a slight modification of (1.5.40), namely,
/311 fl21i
)1/2 ,
21,51 1i %/iv) \
where al= (T1 K*)lycM ly, and al = (T2 K*)-1 YM2Y2, is approxi-
mately distributed as Student's t with degrees of freedom to be appropriately
determined. Because the statement (1.5.37) is still valid, the assumption that
(1.5.45) is approximately Student's t is equivalent to the assumption that w
defined by
ai
, + a2
AziAzi
w 2 (1.5.46)
ai 022
%/ 12i
is approximately x?, for some v. Because Ew = v, w has the same mean as x!.
We shall determine v so as to satisfy
Vw = 2v. (1.5.47)
Solving (1.5.47) for v, we obtain
[ 0.3 ]2
v
%Au (1.5.48)
4 4
(7 2
)2 (T2 K *)(54/22a
Finally, using the standard normal variable (1.5.37) and the approximate
chi-square variable (1.5.46), we have approximately

(1.5.49)
In practice v will be estimated by inserting 61 and di
- into the right-hand side of
(1.5.48) and then choosing the integer closest to the calculated value.
Unfortunately, Welch's method does not easily generalize to a situation
where the equality of vector parameters is involved. Toyoda (1974), like
Welch, proposed that both the denominator and the numerator chi-square
variables of (1.5.26) be approximated by the moment method; but the result-
ing test statistic is independent of the unknown parameters only under unreal-
istic assumptions. Schmidt and Sickles (1977) found Toyoda's approximation
to be rather deficient.
In view of the difficulty encountered in generalizing the Welch method, it
seems that we should look for other ways to test the equality of the regression
parameters in the unequal-variances case. There are two obvious methods
that come to mind: They are (1) the asymptotic likelihood ratio test and (2) the
asymptotic F test (see Goldfeld and Quandt, 1978). 9
The likelihood function of the model defined by (1.5.23) and (1.5.24) is
L = (27 ) -( + T21/207 r,tri 2.2 (1.5.50)
X exp [ 0.50-1-2(y, X 1 /31)'(y1 300]
X exp [ 0.5ai2(y2 -2,
X B .1 (V2 X2P2)I
The value L attained when it is maximized without constraint, de-

noted by L, can be obtained by evaluating the parameters of L at fi = fl1,
fi2 = fi2, cif = "="-- TJ I(Yi Xifil)'(Yi WO, and a=
Til(y2 X 2 /32)'(y2 X2A). The value of L attained when it is maximized
subject to the constraints fi t = fl2(=-1), denoted by L, can be obtained by
evaluating the parameters of L at the constrained maximum likelihood esti-
mates: fi l =P2(), al, and . These estimates can be iteratively obtained as
follows:
Step 1. Calculate ft (ai2X2C1 ai2xpc2)`(ii 2x1y1 + iii2)qy2).
Step 2. Calculate a? --- TN) , 1 X1 /3')'(y 1 and
Ti l(y 2 X2fl)'(Y2 X211).
Step 3. Repeat Step 1, substituting ai and cri for 6? and 63.
Step 4. Repeat Step 2, substituting the estimates of fi obtained at Step 3
for ft.
Continue this process until the estimates converge. In practice, however, the
estimates obtained at the end of Step 1 and Step 2 may be used without
changing the asymptotic result (1.5.51).
Using the asymptotic theory of the likelihood ratio test, which will be
developed in Section 4.5.1, we have asymptotically (that is, approximately
when both T 1 and T2 are large)
2 log (Lit) = T1 log crivaD + T2 log (ayai) x. (1.5.51)
The null hypothesis /31 = fl2 is to be rejected when the statistic (1.5.51) is larger
than a certain value.
The asymptotic F test is derived by the following simple procedure: First,
estimate a and alby 6? and respectively, and define = I A. Second,
multiply both sides of (1.5.24) by and define the new equation
= xtP2 + u2 (1.5.52)
where = /5y2 , = nx2 , and ut = inh. Third, treat (1.5.23) and (1.5.52) as
the given equations and perform the F test (1.5.26) on them. The method
works asymptotically because the variance of u! is approximately the same as
that of u l when T1 and T2 are large, because //converges to a l /a2 in probability.
Goldfeld and Quandt (1978) conducted a Monte Carlo experiment that
showed that, when ai # ai , the asymptotic F test performs well, closely
followed by the asymptotic likelihood ratio test, whereas the F test based on
the assumption of equality of the variances could be considerably inferior.
1.6 Prediction
We shall add to Model 1 the pth period relationship (where p> T)
yp = x;/3 + u,, , (1.6.1)
where yp and up are scalars and xp are the pth period observations on the
regressors that we assume to be random variables distributed independently of
up and u.") We shall also assume that up is distributed independently of u with
Eup = 0 and Vup = a2. The problem we shall consider in this section is how to
predict yp by a function of y, X, and xp when ft and a 2 are unknown.
We shall only consider predictors of yp that can be written in the following
form:
4 = x'pr, ( 1.6.2)
where fl* is an arbitrary estimator of/3 and a function of y and X. Here, r may
be either linear or nonlinear and unbiased or biased. Although there are more
general predictors of the form f(xp , y, X), it is natural to consider (1.6.2)

because x'p fl is the best predictor of yp if fl is known.
The mean squared prediction errorof y: conditional on xp is given by
ERY7, Yp) 2 1xpl= a2 PU3* P'xp,

(1.6.3)
where the equality follows from the assumption of independence between fi*
and up . Equation (1.6.3) clearly demonstrates that as long as we consider only
predictors of the form (1.6.2) and as long as we use the mean squared predic-
tion error as a criterion for ranking predictors, the prediction ofyp is essentially
the same problem as the estimation of x;fi. Thus the better the estimator of x'p fl
is, the better the predictor of yp .
In particular, the result of Section 1.2.5 implies that x pl, where fi is the LS
estimator, is the best predictor in the class of predictors of the form x;C ' y such
that C' X = I, which we shall state as the following theorem:
THEOREM 1.6.1. Let ft be the LS estimator and C be an arbitrary TX K
matrix of constants such that C' X = I. Then
E[(4 yp)2 1xp] E[(x;C'y yp)2 I xp], (1.6.4)
where the equality holds if and only if C = X( X'X)-
Actually we can prove a slightly stronger theorem, which states that the least
squares predictor xp'fi is the best linear unbiased predictor.
THEOREM 1.6.2. Let d be a T-vector the elements of which are either con-
stants or functions of xp . Then
E[(d '3' Y)2Ixp] E[(4 Yd2 Ixp1
for any d such that E(d'y I xp) = E(yp ixp). The equality holds if and only if
d' = x'p(X'X)- IX'.
Proof The unbiasedness condition E(d'yixp) = E(yp ixp) implies
d'X = (1.6.5)
E[(d'y yp)2I xi] = E[(d'u up)2 I xp] (1.6.6)
'a2(1 +d'd).
But from (1.6.3) we obtain
E[(x;fi yp)2 Ixp] a2[ 1 + x;,(X'X) -- 'xp]. ( L6.7)
Therefore the theorem follows from

d'd x'0(X'X)ixp = [d X(X'Xrixpild X(X'X) Ixp], (1.6.8)
where we have used (1.6.5) again.
Theorem 1.6.2 implies Theorem 1.6.1 because Cx p of Theorem 1.6.1 satisfies
the condition for d given by (1.6.5)."
In Section 1.2.4 we stated that if we cannot choose between a pair of
estimators by the criterion defined by Definition 1.2.2 (that is, if the difference
of the mean squared error matrices is neither nonnegative definite nor non-
positive definite), we can use the trace or the determinant as a criterion. The
conditional mean squared prediction error defined in (1.6.3) provides an
alternative scalar criterion, which may have a more intuitive appeal than the
trace or the determinant because it is directly related to an important purpose
to which estimation is put namely, prediction. However, it has one serious
weakness: At the time when the choice of estimators is made, x p is usually not
observed.
A solution to the dilemma is to assume a certain distribution for the random
variables xp and take the further expectation of (1.6.3) with respect to that
distribution. Following Amemiya (1966), let us assume
Exp x; = X' X. (1.6.9)
Then we obtain from (1.6.3) the unconditional mean squared prediction error
E(y: yp )2 = a2 + E( ff) X ' X(/1* /7). (1.6.10)
This provides a workable and intuitively appealing criterion for choosing an
estimator. The use of this criterion in choosing models will be discussed in
Section 2.1.5. The unconditional mean squared prediction error of the least
squares predictor x'pfl is given by
Eocil _ y102 . 0.2(1 T-1 la (1.6.11)
Exercises
1. (Section 1.1.2)
Give an example of a pair of random variables that are noncorrelated but
not independent.
2. (Section 1.1.3)
Let y and x be scalar dichotomous random variables with zero means.
Define u y Cov(y, xX Vx)- x. Prove E(ulx) = 0. Are u and x inde-

pendent?
3. (Section 1.1.3)
Let y be a scalar random variable and x be a vector of random variables.
Prove E[y E(yjx)] 2 E[y g(x)J 2 for any function g.
4. (Section 1.1.3)
A fair die is rolled. Let y be the face number showing and define x by the
rule:
x = y if y is even
=0 if y is odd.
Find the best predictor and the best linear predictor of y based on x and
compare the mean squared prediction errors.
5. (Section 1.2.1)
Assume that y is 3 X 1 and X = (X I , X2 ) is 3 X 2 and draw a three-di-
mensional analog of Figure 1.1.
6. (Section 1.2.5)
Prove that the class of linear unbiased estimators is equivalent to the class
of instrumental variables estimators.
7. (Section 1.2.5)
In Model 1 find a member of the class of linear unbiased estimators for
which the trace of the mean squared error matrix is the smallest, by
minimizing tr C' C subject to the constraint C' X = I.
8. (Section 1.2.5)
Prove that fi defined by (1.2.14) is a best linear unbiased estimator of/3 1 .
9. (Section 1.2.5)
In Model 1 further assume K = 1 and X = 1, where I is the vector of ones.
Define /3+ = 1' y/(T + 1), obtain its mean squared error, and compare it
with that of the least squares estimator fl.
10. (Section 1.2.5)
In Model 1 further assume that T = 3, K = 1, X = (1, 1, 1) ', and that
(4), t = 1, 2, 3, are independent with the distribution
u, = a with probability
= a with probability 4.
Obtain the mean squared error offtR = y' y/x ' y and compare it with that
of the least squares estimator fl. (Note that PR' the reverse least squares
estimator, is obtained by minimizing the sum of squared errors in the
direction of the x-axis.)
11. (Section 1.3.2)
Assume K = 1 in Model 1 with normality and furthermore assume
/32 = a2. Obtain the maximum likelihood estimator of /3 and obtain the
Cramer-Rao lower bound.
12. (Section 1.4.2)
Suppose
Q , = [1 1 1
1 2 3]
Find a row vector R' such that (Q, R) is nonsingular and R'Q = 0.
13. (Section 1.4.2)
Somebody has run a least squares regression in the classical regression
model (Model 1) and reported
2 41 ] [ 5 311
P ' [g2 = 4] and a 2 (X' Xr 1 = [1 2 1] .
/33 2 112
On the basis of this information, how would you estimate fi if you

believed /i1 + /32 = /13 ?
14. (Section 1.5.3)

We have T observations on the dependent variable y, and the independent
variables x, and z t = 1, 2, . . . , T. We believe a structural change
occurred from the first period consisting of T i observations to the second
period consisting of T2 observations ( T1 + T2 = T) in such a way that in
the first period Ey, depends linearly on x, (with an intercept) but not on z,
whereas in the second period Ey, depends linearly on z, (with an intercept)
but not on x,. How do you test this hypothesis? You may assume that (y,)
are independent with constant variance a2 for t = 1, 2, . . . , T.
15. (Section 1.5.3)
Consider the following two regression equations, each of which satisfies
the assumptions of the classical regression model (Model 1):
(1) yi = a l 1 + a2 xi + ui
(2) y2 = /31 1 + /32 x2 + 112 ,
where a's and ll's are scalar parameters, yi, Y2, x 1 , x2, u1, and u2 are
seven-component vectors, and 1 is the seven-component vector of ones.
Assume that u l and u2 are normally distributed with the common var-
iance a 2 and that u 1 and u2 are independent of each other. Suppose that
x l = 1' x2 = 0 and l' y1 = I' y2 = 7. Suppose also that the sample mo-
ment matrix of the four observable variables is given as
Y1 Y2 )(1 3L2
yl 9.3 7 2 1.5
Y2 7 9 3 1
xl 2 3 2 1.2
X2 1.5 1 1.2 1
For example, the table shows y y i = 9.3 and yc y2 = 7. Should you reject
the joint hypothesis "a 1 = fl, and a l + 2a2 =132" at the 5% significance
level? How about at 1%?
16. (Section 1.5.3)
Consider a classical regression model
y1 = a + /31 z1 +
Y2 = cx 2 x2 /32 z2 + U2
y3 = a 3 X3 + AZ3 u3,
where a's and Fs are scalar unknown parameters, the other variables are
vectors often elements, x's and z's are vectors of known constants, and u's
are normal with mean zero and Eu iu; = a 2 I for every i and Eu iu; =0 if
i # j. Suppose the observed vector products are as follow:
YiYi = 2, YiY2 = 5 , Y;Y3 2

i = 1, yx2 = 3, y;x3 = 2y;x
yczl = 2, yz2 = 3, y;z3 = 1
x ;xi = z;zi = 4 for every i
x;zi =0 for every i.
Test the joint hypothesis (a 1 = a2 and /32= A) at the 5% significance
level.
17. (Section 1.6)

Consider Model 1 with the added prediction-period equation (1.6.1).
Suppose Z is a TXL matrix of known constants and zp is an L-vector of
known constants. Which of the following predictors of y p is better? Ex-
plain.
9p = z;(Z'Z) ' Z'y
pp =z;(Z'Z)IZ'X(X'X)IX/y.
18. (Section 1.6)
Consider the case K = 2 in Model 1, where y, =- /31 + 132x, + up For the
prediction period we have yp = fil -I- /32x,, -I- up , where up satisfies the as-
sumptions of Section 1.6. Obtain the mean squared prediction error of the
predictor Pp = T-- ' MT., y, and compare it with the mean squared predic-
tion error of the least squares predictor.
19. (Section 1.6)
Prove that any d satisfying (1.6.5) can be written as Cxp for some C such
that C' X = I.
2 Recent Developments
in Regression Analysis
In this chapter we shall present three additional topics. They can be discussed
in the framework of Model 1 but are grouped here in a separate chapter
because they involve developments more recent than the results of the pre-
vious chapter.
2.1 Selection of Regressors'

2.1.1 Introduction
Most of the discussion in Chapter 1 proceeded on the assumption that a given
model (Model 1 with or without normality) is correct. This is the ideal situa-
tion that would occur if econometricians could unambiguously answer ques-
tions such as which independent variables should be included in the right-
hand side of the regression equation; which transformation, if any, should be
applied to the independent variables; and what assumptions should be im-
posed on the distribution of the error terms on the basis of economic theory. In
practice this ideal situation seldom occurs, and some aspects of the model
specification remain doubtful in the minds of econometricians. Then they
must not only estimate the parameters of a given model but also choose a
model among many models.
We have already considered a particular type of the problem of model
selection, namely, the problem of choosing between Model 1 without con-
straint and Model 1 with the linear constraints Q'fi = c. The model selection
of this type (selection among "nested" models), where one model is a special
case of the other broader model, is the easiest to handle because the standard
technique of hypothesis testing is precisely geared for handling this problem.
We shall encounter many instances of this type of problem in later chapters. In
Section 2.1, however, we face the more unorthodox problem of choosing
between models or hypotheses, neither of which is contained in the other
("nonnested" models).
In particular, we shall study how to choose between the two competing

regression equations
y= + (2.1.1)
and
Y u2, (2.1.2)
where X1 is a T X K1 matrix of constants, X2 is a T X K2 matrix of constants,
Eul = Eu2 =0, Eu 1 t4 = aft, and Eu2u; = ail. Note that this notation differs
from that of Chapter 1 in that here there is no explicit connection between X 1
X2: X1 and X2 may contain some common column vectors or they mayand
not; one of X i and X2 may be completely contained in the other or they may be
completely different matrices. Note also that the dependent variable y is the
same for both equations. (Selection among more general models will be dis-
cussed briefly in Section 4.5.)
This problem is quite common in econometrics, for econometricians often
run several regressions, each of which purports to explain the same dependent
variable, and then they choose the equation which satisfies them most accord-
ing to some criterion. The choice is carried out through an intuitive and
unsystematic thought process, as the analysts consider diverse factors such as a
goodness of fit, reasonableness of the sign and magnitude of an estimated
coefficient, and the value of a t statistic on each regression coefficient. Among
these considerations, the degree of fit normally plays an important role, al-
though the others should certainly not be ignored. Therefore, in the present
study we shall focus our attention on the problem of finding an appropriate
measure of the degree of fit. The multiple correlation coefficient R 2, defined in
(1.2.8), has an intuitive appeal and is a useful descriptive statistic; however, it
has one obvious weakness, namely, that it attains its maximum of unity when
one uses as many independent variables as there are observations (that is,
when K = T). Much of what we do here may be regarded as a way to rectify
that weakness by modifying R 2.
2.1.2 Statistical Decision Theory

We shall briefly explain the terminology used in statistical decision theory. For
a more thorough treatment of the subject, the reader should consult Zacks
(1971). Statistical decision theory is a branch of game theory that analyzes the
game played by statisticians against nature. The goal of the game for statisti-
cians is to make a guess at the value of a parameter (chosen by nature) on the
Recent Developments in Regression Analysis 47
basis of the observed sample, and their gain from the game is a function of how
close their guess is to the true value. The major components of the game are 0,
the parameter space; Y, the sample space; and D, the decision space (the
totality of functions from Y to 0). We shall denote a single element of each
space by the lowercase letters 0, y, d. Thus, if y is a particular observed sample
(a vector of random variables), d is a function of y (called a statistic or an
estimator) used to estimate 0. We assume that the loss incurred by choosing d
when the true value of the parameter is 0 is given by the loss function L(d, 0).
We shall define a few standard terms used in statistical decision theory.
Risk. The expected loss Ey L(d, 0) for which the expectation is taken with
respect to y (which is implicitly in the argument of the function d) is called the
risk and is denoted by R(d10).
Uniformly smaller risk. The estimator2 d i has a uniformly smaller risk than
the estimator d 2 if R(d 1 10)-5 R(d210) for all 0 E 0 and R(d 1 10) < R(d210) for at
least one 0 E 0.
Admissible. An estimator is admissible if there is no d in D that has a
uniformly smaller risk. Otherwise it is called inadmissible.
Minimax. The estimator d* is called a minimax estimator if
max R(d*10) = mm max R(d10).
0E13 deD eEe
The minimax estimator protects the statistician against the worst pos-
sible situation. If max eEe R(d10) does not exist, it should be replaced with
supeEe R(d10) in the preceding definition (and min with inf).
Posterior risk. The expected loss Ef L(d, 0) for which the expectation is
taken with respect to the posterior distribution of 0 given y is called the
posterior risk and is denoted by R(d ly). It obviously depends on the particular
prior distribution used in obtaining the posterior distribution.
Bayes estimator. The Bayes estimator, given a particular prior distribution,
minimizes the posterior risk R(dly). If the loss function is quadratic, namely,
L(d, 0) = (d 0) ' W(d 0) where W is an arbitrary nonsingular matrix, the
posterior risk Ee(d 0)' W(d 0) is minimized at d = E80, the posterior
mean of 0. An example of the Bayes estimator was given in Section 1.4.4.
Regret. Let R(d10) be the risk. Then the regret W(d10) is defined by
W(d10) = R(d10) mm R(d10).
deD
Minimax regret. The minimax regret strategy minimizes max 0,43 W(d10)
with respect to d.
Some useful results can be stated informally as remarks rather than stating
them formally as theorems.
REMARK 2.1.1. A Bayes estimator is admissible.
REMARK 2.1.2. A minimax estimator is either a Bayes estimator or the limit
of a sequence of Bayes estimators. The latter is called a generalized Bayes
estimator. (In contrast, a Bayes estimator is sometimes called a proper Bayes
estimator.)
REMARK 2.1.3. A generalized Bayes estimator with a constant risk is
minimax.
REMARK 2.1.4. An admissible estimator may or may not be minimax, and a
minimax estimator may or may not be admissible.
2.1.3. Bayesian Solution

The Bayesian solution to the selection-of-regressors problem provides a peda-
gogically useful starting point although it does not necessarily lead to a useful
solution in practice. We can obtain the Bayesian solution as a special case of
the Bayes estimator (defined in Section 2.1.2) for which both 0 and D consist
of two elements. Let the losses be represented as shown in Table 2.1, where L12
is the loss incurred by choosing model 1 when model 2 is the true model
and L21 is the loss incurred by choosing model 2 when model 1 is the true
mode1.3 Then, by the result of Section 2.1.2, the Bayesian strategy is to choose
model 1 if
LI2P(21y) < 1-21111 ly), (2.1.3)
where P(ily), i = 1 and 2, is the posterior probability that the model i is true
given the sample y. The posterior probabilities are obtained by Bayes's rule as
P( 1 lY) f f fiY1 01)f(011 1 )P( 1 )4

(2.1.4)
f(Y1 01)f(0 1 I 1)P(1)chli + f f(Y1 02).f(021 2)P(2)d02
and similarly for P(2 1Y), where 0, = (fi; , 0)% f(Y1 0i) is the joint density of y
given 0i , f(Oil i) is the prior density of 0, given the model i, and P(i) is the prior
probability that the model i is true, for i = 1 and 2.
There is an alternative way to characterize the Bayesian strategy. Let S be a
subset of the space of y such that the Bayesian chooses the model 1 if y E S.
Table 2.1 The loss matrix

True
model
Decision Model 1 Model 2
Choose model I 0 L12
Choose model 2 L21 0
Then the Bayesian minimizes the posterior risk

L12P(2)P(Y E S12) L21 /1(1)P(y E (2.1.5)
with respect to S, where S is the complement of S. It is easy to show that the
posterior risk (2.1.5) is minimized when S is chosen to be the set of y that
satisfies the inequality (2.1.3).
The actual Bayesian solution is obtained by specifying f(y10i), JAI i), and
PM in (2.1.4) and the corresponding expressions for P(2Iy). This is not done
here because our main purpose is to understand the basic Bayesian thought, in
light of which we can perhaps more clearly understand some of the classical
strategies to be discussed in subsequent subsections. The interested reader
should consult Gayer and Geisel (1974) or Zellner (1971, p. 306). Gayer and
Geisel pointed out that if we use the standard specifications, that is, f(Y10i)
normal, f(eili) "diffuse" natural conjugate, P(1) = P(2), the Bayesian solu-
tion leads tci a meaningless result unless K1 = K2 .4
2.1.4 Theirs Corrected R 2

Theil (1961, p. 213) proposed a correction of R 2 aimed at eliminating the
aforementioned weakness of R 2. Theirs corrected R 2, denoted by R 2 , is de-
fined by
1R2 - (1 R2). (2.1.6)

T K
Because we have from (1.2.9)
1 R2 Y'MY (2.1.7)
y'Ly
where M = I X(X'X) -1X' and L = I T-1 11' as before, choosing the
equation with the largest R 2 is equivalent to choosing the equation with the
smallest o2 = (T Kr y ' My . Coming back to the choice between Eqs.

(2.1.1) and (2.1.2), Theil's strategy based on his R 2 amounts to choosing Eq.
(2.1.1) if
al < Fri, (2.1.8)
where Cri =-- (T KI) iy'M ly,
M1 = I X1(XPC1)iX1,
= (T K2)- 'y'M2y, and
M2 = I - x2(xpc2)-13q.
The inequality (2.1.8) can be regarded as a constraint on y and hence defines
a subset in the space of y. Call it So ; that is, So = (yliii < al). This choice of S
can be evaluated in terms of the Bayesian minimand (2.1.5). Suppose Eq.
(2.1.2) is the true model. Then we have
-2-2 Y' 14 1Y Y'A42Y (2.1.9)

at a2 T K 1 T K2
_ flpqM IX2/32 2upVIIX2/12 -1- tqM 1u2 t012u2

TK1 T K2 '
Therefore
. /3pcivi,x2 p2
Eur- i ail2)
TKI
Therefore, in view of the fact that nothing a priori is known about whether
5-; c9 is positively or negatively skewed, it seems reasonable to expect that

P(y E So l2) < I. - (2.1.11)

For a similar reason it also seems reasonable to expect that
P(Y E gol 1 ) < i. (2.1.12)
These inequalities indicate that So does offer an intuitive appeal (though a
rather mild one) to the classical statistician who, by principle, is reluctant to
specify the subjective quantities L12, L,, P(1), and P(2) in the posterior risk
(2.1.5).
As we have seen, Theil's corrected R2 has a certain intuitive appeal and has
been widely used by econometricians as a measure of the goodness of fit.
However, its theoretical justification is not strong, and the experiences of
some researchers have led them to believe that Theirs measure does not
correct sufficiently for the degrees of freedom; that is, it still tends to favor the
equation with more regressors, although not as much as the uncorrected R 2

(se,forxamplMy1975).InSectio2wshalpremu
that corrects more for the degrees of freedom than Theil's R 2 does, and in
Section 2.1.6 we shall try to justify the proposed measure from a different
angle.
2.1.5 Prediction Criterion

A major reason why we would want to consider any measure of the goodness
of fit is that the better fit the data have in the sample period, the better
prediction we can expect to get for the prediction period. Therefore it makes
sense to devise a measure of the goodness of fit that directly reflects the
efficiency of prediction. For this purpose we shall compare the mean squared
prediction errors of the predictors derived from the two competing equations
(2.1.1) and (2.1.2). To evaluate the mean squared prediction error, however,
we must know the true model; but, if we knew the true model, we would not
have the problem of choosing a model. We get around this dilemma by
evaluating the mean squared prediction error of the predictor derived from
each model, assuming in turn that each model is the true model. This may be
called the minimini principle (minimizing the minimum risk)in contrast to
the more standard minimax principle defined in Section 2.1.2 because the
performance of each predictor is evaluated under the most favorable condi-
tion for it. Although the principle is neither more nor less justifiable than the
minimax principle, it is adopted here for mathematical convenience. Using
the unconditional mean squared prediction error given in Eq. (1.6.11) after a 2
Prediction Criterion isreplacdwthunbseimaor,wdfnth
(abbreviated PC) by
PC; = + i= 1, 2, (2.1.13)
for each model, where al =
(T K iry'M ay is the number of regressors in
model i.
If we define the modified R 2, denoted by k 2, as
1_ T+K
(1 R 2), (2.1.14)
TK
choosing the equation with the smallest PC is equivalent to choosing the
equation with the largest R 2. A comparison of (2.1.6) and (2.1.14) shows that
112 imposes a higher penalty upon increasing the number of regressors than
does Theil's R 2.
Criteria proposed by Mallows (1964) and by Akaike (1973) are similar to

the PC, although they are derived from somewhat different principles. To
define Mallows' Criterion (abbreviated MC), we must first define the matrix X
as that matrix consisting of distinct column vectors contained in the union of
X, and X2 and K as the number of the column vectors of X thus defined. These
definitions of X and K can be generalized to the case where there are more than
two competing equations. Then we have
2Ki y'M y y ' Mar
MC1 = + i = 1, 2, (2.1.15)
T T K T
where M = I X(X'X) - IX'. Akaike (1973) proposed what he calls the
Akaike Information Criterion (abbreviated AIC) for the purpose of distin-
guishing between models more general than regression models (see Section
4.5.2). When the AIC is applied to regression models, it reduces to
2K.
AIC, = log ( YiMiY ) + ' i= 1, 2. (2.1.16)
T 1 T '
All three criteria give similar results in common situations (see Amemiya,
1980a). The MC has one unattractive feature: the matrix X must be specified.
2.1 .6 Optimal Significance Level

In the preceding sections we have considered the problem of choosing be-
tween equations in which there is no explicit relationship between the com-
peting regressor matrices X, and X2. In a special case where one set of regres-
sors is contained in the other set, the choice of an equation becomes equivalent
to the decision of accepting or rejecting a linear hypothesis on the parameters
of the broader model. Because the acceptance or rejection of a hypothesis
critically depends on the significance level chosen, the problem is that of
determining the optimal significance level (or, equivalently, the optimal criti-
cal value) of the F test according to some criterion. We shall present the gist of
the results obtained in a few representative papers on this topic and then shall
explain the connection between these results and the foregoing discussion on
modifications of R 2.
Let the broader of the competing models be
y Xfi + u + X2/12 + u (2.1.17)
where X 1 and X2 are T X K, and T X K2 matrices, respectively, and for which

we assume u is normal so that model (2.1.17) is the same as Model 1 with
normality. Note that the X2 here has no relationship with the X2 that appears
in Eq. (2.1.2). Suppose we suspect /32 might be 0 and test the hypothesis /12 = 0
by the F test developed in Section 1.5.2. The appropriate test statistic is
obtained by putting /32 = 0 in Eq. (1.5.19) as
. ( T KVy'M ly _
1) F(K2 , T K). (2.1.18)
'7 k K2 A y'My
The researcher first sets the critical value d and then chooses the model
(2.1.17) if ri _ d or the constrained model
y = Xfil + u (2.1.19)
if n< d.
Conventionally, the critical value d is determined rather arbitrarily in such a
way that P(t) d) evaluated under the null hypothesis equals a preassigned
significance level such as 1 or 5 percent. We shall consider a decision-theoretic
determination of d. For that we must first specify the risk function. The
decision of the researcher who chooses between models (2.1.17) and (2.1.19)
on the basis of the F statistic rj may be interpreted as a decision to estimate fl by
the estimator ft defined as
ft = ft ifla..d (2.1.20)
--= [1)1 ] if n< d,
where ji is the least squares estimator applied to (2.1.17) and ii i is that applied
to (2.1.19). Thus it seems reasonable to adopt the mean squared error matrix
SI --- E(fi fi)(fi fi)', where the expectation is taken under (2.1.17) as our
risk (or expected loss) function. However, il is not easy to work with directly
because it depends on many variables and parameters, namely, X, ft, a2, K, K 1 ,
and d, in addition to having the fundamental difficulty of being a matrix. (For
the derivation of fl, see Sawa and Hiromatsu, 1973, or Farebrother, 1975.)
Thus people have worked with simpler risk functions.
Sawa and Hiromatsu (1973) chose as their risk function the largest charac-
teristic root of
[V(x'x)- ' 4:2] - "2Q1/Q[Q'(x'x)`Q] - "2, (2.1.21)
where Q' = (0, I) where 0 is the K2 X ICI matrix of zeros and I is the identity
matrix of size K2. This transformation of LI lacks a strong theoretical justifica-
tion and is used primarily for mathematical convenience. Sawa and Hiro-
matsu applied the minimax regret strategy to the risk function (2.1.21) and
showed that in the special case K2 = 1, d 1.88 is optimal for most reasonable
values of T K. Brook (1976) applied the minimax regret strategy to a differ-
ent transformation of 0,
tr (2.1.22)
and recommended d= 2 on the basis of his results. The risk function (2.1.22)
seems more reasonable than (2.1.21) as it is more closely related to the mean
squared prediction error (see Section 1.6). At any rate, the conclusions of these
two articles are similar.
Now on the basis of these results we can evaluate the criteria discussed in the
previous subsections by asking what critical value is implied by each criterion
in a situation where a set of the regressors of one model is contained in that of
the other model. We must choose between models (2.1.17) and (2.1.19). For
each criterion, let p denote the ratio of the value of the criterion for model
(2.1.19) over that for model (2.1.17). Then, using (2.1.18), we can easily
establish a relationship between pj and p. For Theil's criterion we have from
(2.1.6)
T T K
17 p(Theil) (2.1.23)
K K
Therefore we obtain the well-known resultthat Theil's criterion selects
(2.1.19) over (2.1.17) if and only if n < 1. Thus, compared with the optimal
critical values suggested by Brook or by Sawa and Hiromatsu, Theil's criterion
imposes far less of a penalty upon the inclusion of regressors. From the
prediction criterion (2.1.13) we get
(T K 1)(T + K) TK
p(PC) (2.1.24)

(K K 1)(T + K K K 1 *
Therefore
2T
p(PC) > 1 if and only if ri > (2.1.25)
T + K1
Table 2.2 gives the values of 2 TAT + K 1) for a few selected values of K 1 /T.
These values are close to the values recommended by Brook and by Sawa and
Hiromatsu. The optimal critical value of the F test implied by the AIC can be
easily computed for various values of Ka T and KIT from (2.1.16) and
(2.1.18). The critical value for the AIC is very close to that for the PC, although
Table 2.2 Optimal critical value of the F test implied by PC

2T
T+
1/10 1.82
1/20 1.90
1/30 1.94
it is slightly smaller. Finally, for the MC we have from (2.1.15)
KT+ K p(MC)
T K + 2K,
KKi
(2.1.26)
Therefore
p(MC) > 1 if and only if n > 2. (2.1.27)
These results give some credence to the proposition that the modified R 2
2 as a measure of the goodness prosedhiftoTel'scrdR
of fit. However, the reader should take this conclusion with a grain of salt for
several reasons: (1) None of the criteria discussed in the previous subsections is
derived from completely justifiable assumptions. (2) The results in the litera-
ture of the optimal significance level are derived from the somewhat question-
able principle of minimizing the maximum regret. (3) The results in the
literature on the optimal significance level are relevant to a comparison of the
criteria considered in the earlier subsections only to the extent that one set of
regressors is contained in the other set. The reader should be reminded again
that a measure of the goodness of fit is merely one of the many things to be
considered in the whole process of choosing a regression equation.
2.2 Ridge Regression and Stein's Estimator5

2.2.1 Introduction
We proved in Section 1.2.5 that the LS estimator is best linear unbiased in

Model 1 and proved in Section 1.3.3 that it is best unbiased in Model 1 with
normality In either case a biased estimator may be better than LS (in the sense
of having a smaller mean squared error) for some parameter values. In this
section we shall consider a variety of biased estimators and compare them to
LS in Model 1 with normality.
The biased estimators we shall consider here are either the constrained least
squares estimator discussed in Section 1.4.1 or the Bayes estimator discussed
in Section 1.4.4 or their variants. If the linear constraints (1.4.1) are true, the
constrained least squares estimator is best linear unbiased. Similarly, the
Bayes estimator has optimal properties if the regression vector # is indeed
random and generated according to the prior distribution. In this section,
however, we shall investigate the properties of these estimators assuming that
the constraints do not necessarily hold. Hence, we have called them biased
estimators. Even so, it is not at all surprising that such a biased estimator can
beat the least squares estimator over some region of the parameter space. For
example, 0 can beat any estimator when the true value of the parameter in
question is indeed 0. What is surprising is that there exists a biased estimator
that dominates the least squares estimates over the whole parameter space
when the risk function is the sum of the mean squared errors, as we shall show.
Such an estimator was first discovered by Stein (see James and Stein, 1961)
and has since attracted the attention of many statisticians, some of whom have
extended Stein's results in various directions.
In this section we shall discuss simultaneously two closely related and yet
separate ideas: One is the aforementioned idea that a biased estimator can
dominate least squares, for which the main result is Stein's, and the other is the
idea of ridge regression originally developed by Hoerl and Kennard (1970a, b)
to cope with the problem of multicollinearity. Although the two ideas were
initially developed independently of each other, the resulting estimators are
close cousins; in fact, the term Stein-type estimators and the term ridge esti-
mators are synonymous and may be used to describe the same class of estima-
tors. Nevertheless, it is important to recognize them as separate ideas. We
might be tempted to combine the two ideas by asserting that a biased estimator
can be good and is especially so if there is multicollinearity. The statement can
be proved wrong simply by noting that Stein's original model assumes
X 'X = I, the opposite of multicollinearity. The correct characterization of the
two ideas is as follows: (1) Some form of constraint is useful in estimation. (2)
Some form of constraint is necessary if there is multicollinearity.
The risk function we shall use throughout this section is the scalar
gft flY(fl -
11), (2.2.1)
where is an estimator in question. This choice of the risk function is as
general as
E(T1 fiYA(ft f1), (2.2.2)
where A is an arbitrary (known) positive definite matrix, because we can

always reduce (2.2.2) to (2.2.1) by transforming Model 1 to
y= +u (2.2.3)
= xA-1/2Aiap u
and consider the transformed parameter vector AInft. Note, however, that
(2.2.1) is not as general as the mean squared error matrix E(fl fi)(p fir ,
which we used in Section 1.2.4, since (2.2.1) is the trace of the mean squared
error matrix.
2.2.2 Canonical Model

Let H be an orthogonal matrix that diagonalizes the matrix X'X, that is,
H'H = I and H'X'XH = A, where A is the diagonal matrix consisting of the
characteristic roots of X' X. Defining X* = XH and a = 1-11, we can write Eq.
(1.1.4) as
y = X*a + u. (2.2.4)
If cr is the least squares estimator of a in model (2.2.4), we have
N(a, 0.2A-1). (2.2.5)
Because the least squares estimator is a sufficient statistic for the vector of
regression coefficients, the estimation of /3 in Model 1 with normality is
equivalent to the estimation of a in model (2.2.5). We shall call (2.2.5) the
canonical model; it is simpler to analyze than the original model. Because
= I, the risk function E(fl fl)'(fl p in Model 1 is equivalent to the
risk function E(ii (ert a) in model (2.2.5).
2.2.3 Multicollinearity and Principal Components

In Model I we assumed that X is of full rank [that is, rank(X) = K T], or,
equivalently, that X' X is nonsingular. If it is not, X' X cannot be inverted and
therefore the least squares estimator cannot be uniquely defined. In other
words, there is no unique solution to the normal equation
X'Xil = X'y. (2.2.6)
Even then, however, a subset of the regression parameters still may be esti-
mated by (1.2.14).
We shall now turn to a more general question: Can F'fl be estimated by least
squares, where F is an arbitrary K X fmatrix of rankf(f K)? To make sense
out of this question, we must first define the least squares estimator of F We
say that the least squares estimator of Fills F'/3, where /lis any solution (which
may not be unique) of the normal equation (2.2.6), provided F'fl is unique. If
F'/3 is unique, we also say that F'fl is estimable. Then it is easy to prove that
F'fl is estimable if and only if we can write F = X' A for some T X fmatrix A,
or equivalently, if and only if we can write F = X' XB for some K X fmatrix B.
(See Rao, 1973, pp. 223 - 224, for the proof.) If is estimable, it can be
shown that F'fl is the best linear unbiased estimator of F'fl.
The estimability of F'fl can be reduced to the estimability of a subset of the
regression parameters in the sense of the previous paragraph by the following
observation. Let G be a K X (K f ) matrix of rank K fsuch that G'F =0.
(We defined a similar matrix in Section 1.4.2.) Then we can write Model 1 as
y = Xfi + u (2.2.7)
F'
= X[F(F G(G'Gr [ G ,] fl + u
where the identity defines Z 1 , Z2, yi , and y2 . Then the estimability of F'fl is
equivalent to the estimability of y i .
If X' X is singular, fi is not estimable in the sense defined above (that is, a
solution of Eq. 2.2.6 is not unique). This fact does not mean that we should not
attempt to estimate fl. We can still meaningfully talk about a class of estima-
tors and study the relative merits of the members of the class. One such class
may be the totality of solutions of (2.2.6)infinite in number. Another class
may be the constrained least squares estimator satisfying linear constraints
Q'P = c. From Eq. (1.4.11) it is clear that this estimator can be defined even
when X'X is singular. A third class is the class of Bayes estimators with prior
Q/3 = c + v formed by varying Q, c, and the distribution of v. We should
mention an important member of the first class that also happens to be a
member of the second class. It is called the principal components estimator.
Suppose we arrange the diagonal elements of A defined in Section 2.2.2 in
descending order A 1 A. 2 AKand let the corresponding charac-
teristic vectors be h i , h2 , . . . , hK so that H = (h 1 , h2 , . . . , hK). Then we
call XII; the ith principal component of X. If X'X is singular, some of its
characteristic roots are 0. Partition
A = [A, 0
(2.2.8)
0 A.2]
so that the diagonal elements of A 1 are positive and those of A2 are all 0, and
partition H = (H 1 , 112) conformably. Furthermore, define Xr= XH I and
X 21' = XH2 and partition a' = (a; , a) conformably. Then X 111 =0 and hence
a2 cannot be estimated. Suppose we estimate a l by
a, = (XrXr)-1Xf'y (2.2.9)
and set 112 = 0. (It is arbitrary to choose 0 here; any other constant will do.)
Transforming ee ...- (ac, ii;) into an estimator of ft, we obtain the principal
components estimator of p by the formula6
lip = Ha = H IAVWX'y. (2.2.10)
It is easy to show that fip satisfies (2.2.6); hence, it is a member of the first class.
It is also a member of the second class because it is the constrained least
squares subject to HVI =0.
It was shown by Fomby, Hill, and Johnson (1978) that the principal compo-
nents estimator (or constrained least squares subject to H;fl = 0) has a smaller
variance-covariance matrix than any constrained least squares estimator ob-
tained subject to the constraints Ql= c, where Q and c can be arbitrary
except that Q has an equal or smaller number of columns than H2.
We shall now consider a situation where X'X is nonsingular but nearly
singular. The near singularity of X' X is commonly referred to as multicollin-
earity. Another way of characterizing it is to say that the determinant ofX 'X is
close to 0 or that the smallest characteristic root of X' X is small. (The question
of how small is "small" will be better understood later.) We now ask the
question, How precisely or imprecisely can we estimate a linear combination
of the regression parameters cli by least squaresr Because the matrix H is
nonsingular, we can write c = Hd for some vector d. Then we have
V(c'M = a2d'A- id, (2.2.11)
which gives the answer to the question. In other words, the closer c is to the
direction of the first (last) principal component, the more precisely (impreci-
sely) one can estimate cl. In particular, we note from (2.2.5) that the preci-
sion of the estimator of an element of a is directly proportional to the corre-
sponding diagonal element of A.
Suppose we partition A as in (2.2.8) but this time include all the "large"
elements in A 1 and "small" elements in A2. The consideration of which roots
to include in A2 should depend on a subjective evaluation of the statistician

regarding the magnitude of the variance to be tolerated. It makes sense to use
the principal components estimator (2.2.10) also in the present situation,
because a 2 can be only imprecisely estimated. (Here, the principal compo-
nents estimator is not unique because the choice of a 2 is a matter of subjective
judgment. Therefore it is more precise to call this estimator by a name such as
the K1 principal components estimator specifying the number of elements
chosen in a l .)
2.2.4 Ridge Regression

Hoerl and Kennard ( 1970a, b) proposed the class of estimators defined by
/1(y) = (X'X + yI)-1 X'y, (2.2.12)
called the ridge estimators. Hoerl and Kennard chose these estimators be-
cause they hoped to alleviate the instability of the least squares estimator due
to the near singularity of X ' X by adding a positive scalar y to the characteristic
roots of X' X. They proved that given /3 there exists y*, which depends upon /I,
such that E[11(y*) II] pky*) if] < E(ft fir (lj /I), where /1 = /AO).
Because 7* depends on fl, fl(y) is not a practical estimator. But the existence of
/1(y*) gives rise to a hope that one can determine 7, either as a constant or as a
function of the sample, in such a way that /1(y) is better than the least squares
estimator ji with respect to the risk function (2.2.2) over a wide range of the
parameter space.
Hoerl and Kennard proposed the ridge trace method to determine the value
of y. The ridge trace is a graph of A(y), the ith element of Ay), drawn as a
function of y. They proposed that y be determined as the smallest value at
which the ridge trace stabilizes. The method suffers from two weaknesses: (1)
The point at which the ridge trace starts to stabilize cannot always be deter-
mined objectively. (2) The method lacks theoretical justification inasmuch as
its major justification is derived from certain Monte Carlo studies, which,
though favorable, are not conclusive. Although several variations of the ridge
trace method and many analogous procedures to determine y have been
proposed, we shall discuss only the empirical Bayes method, which seems to
be the only method based on theoretical grounds. We shall present a variant of
it in the next paragraph and more in the next two subsections.
Several authors interpreted the ridge estimator (more precisely, the class of
estimators) as the Bayes estimator and proposed the empirical Bayes method
of determining 7; we shall follow the discussion of Sclove (1973). 8 Suppose
that the prior distribution of /I is N(/.z, op), distributed independently of a.

Then from (1.4.22) the Bayes estimator of /3 is given by
fl* (X'X + yI) -1(X'y + yg), (2.2.13)
where y = a2/oi. Therefore, the ridge estimator (2.2.12) is obtained by putting

= 0 in (2.2.13). By the empirical Bayes method we mean the estimation of
the parameters (in our case, cr) of the prior distribution using the sample
observations. The empirical Bayes method may be regarded as a compromise
between the Bayesian and the classical analysis. From the marginal distribu-
tion (that is, not conditional on A of y, we have
Ey'y orrj tr X'X + To-2, (2.2.14)
which suggests that we can estimate aj by
cifi Y'Y ra2 (2.2.15)

tr X'X '
where 62= X(X'X) -1 Xly as usual. Finally, we can estimate y by
62/6.3
In the next two subsections we shall discuss many more varieties of ridge
estimators and what we call generalized ridge estimators, some of which
involve the empirical Bayes method of determining y. The canonical model
presented in Section 2.2.2 will be considered.
2.2.5 Stein's Estimator: Homoscedastic Case

Let us consider a special case of the canonical model (2.2.5) in which A = I.
James and Stein (1961) showed that Ell [1 oc11 2 is minimized
for all a when c= (K 2)a2 if a2 is known and K 3, where 11x11 2 denotes the
vector product x' x. If we define
(K 2)a2 ]
Stein's estimator ii* = [ 1
6e6t a'
the result of James and Stein implies in particular that Stein's estimator is
uniformly better than the maximum likelihood estimator a with respect to the
risk function E(ei a)'(& a) 3. In other words, ii is inadmissible (see
the definition of inadmissible in Section 2.1.2). The fact that a is minimax
with a constant risk (see Hodges and Lehman, 1950) implies that Of * is mini-
max. This surprising result has had a great impact on the theory of statistics.
Translated into the regression model, the result of James and Stein implies
two facts: (1) Consider the ridge estimator (X'X + yI) -1X'y with y
(K 2)c2/[ (K 2)(3 2]. If X'X = I, it is reduced precisely to Stein's
estimator of /3 since X'y N(ft, a2I). Therefore this ridge estimator is uni-
formly better than the least squares estimator if X'X = I (the opposite of
multicollinearity). (2) Assume a general X'X in Model 1. If we define A =
(X'X) 1 /2 , we have Afl N(Afl, 0.21), where fi is the least squares estimator.
Applying Stein's estimator to AA we know ENO B)A/3 A/31I 2 <
M 12 for all /3 where B = (K 2)cr 2/fi'X'Xii. Therefore, equivalently,
(1 B)/1 is uniformly better than fl with respect to the risk function
E(ft M'X'X(ft M. Note that this is essentially the risk function we pro-
posed in (1.6.10) in Section 1.6, where we discussed prediction.
So far, we have assumed a2 is known. James and Stein showed that even
when a2 is unknown, if S is distributed independent of a and as a2x!, then
Ell [ 1 cS(61'6)- 96/ all 2 attains the minimum for all a and GI at c =
(K 2)/(n + 2) if K 3. They also showed that [1 cS(a'a)-1]a is uni-
formly better than oi if 0 < c < 2(K 2)/(n + 2). In the regression model we
can put S = y' [I X(X 1 30- 'Xly because it is independent of /I and distrib-
uted as a2A-K.
Efron and Morris (1972) interpreted Stein's estimator as an empirical Bayes
estimator. Suppose 61 N(a, a2I), where 0.2 is known, and the prior distribu-
tion of a is N(0, c2 ) - ' I). Then the Bayes estimator is ii* = (1 + y)'ol =
(1 Mil where B = y/(1 + y). The marginal distribution of& is N(0, a 233- 'I).
Therefore B&' &/a 2 xi. Because E[(x 2K) -1 ] = (K 2)-1 (see Johnson and
Kotz, 1970a, vol. 1, p. 166), we have
(K 2)a 2]
E[ B.
a'a
Thus we can use the term within the square bracket as an unbiased estimator
of B, thereby leading to Stein's estimator.
It is important not to confuse Stein's result with the statement that
E(ii* a)(6c* a)' <E(& a)( a)' in the matrix sense. This inequality
does not generally hold. Note that Stein's estimator shrinks each component
of a by the same factor B. If the amount of shrinkage for a particular compo-
nent is large, the mean squared error of Stein's estimator for that component
may well exceed that of the corresponding component of ci, even though
Ella* a11 2 < Ella al1 2. In view of this possibility, Efron and Morris
(1972) proposed a compromise: Limit the amount of shrinkage to a fixed
amount for each component. In this way the maximum possible mean
squared error for the components of a can be decreased, whereas, with luck,
the sum of the mean squared errors will not be increased by very much.
Earlier we stated that the maximum likelihood estimator a is inadmissible

because it is dominated by Stein's estimator. A curious fact is that Stein's
estimator itself is dominated by
(K 2)a2 &,
Stein's positive-rule estimator [1
eg'ii +
where [x]4 denotes max[0, x]. Hence Stein's estimator is inadmissible, as
proved by Baranchik (1970). Efron and Morris (1973) showed that Stein's
positive-rule estimator is also inadmissible but cannot be greatly improved
upon.
We defined Stein's estimator as the estimator obtained by shrinking &
toward 0. Stein's estimator can easily be modified in such a way that it shrinks
et toward any other value. It is easy to show that
(K 2)a2 i
Stein's modified estimator [1
(ii cr (ci c)] (a. c) + c
is minimax for any constant vector c. If the stochastic quantity ./C- Vet is
chosen to be c, where lis the vector of ones, then the resulting estimator can be
shown to be minimax for K 4.
2.2.6 Stein's Estimator: Heteroscedastic Case

Assume model (2.2.5), where A is a general positive definite diagonal matrix.
Two estimators for this case can be defined.
Ridge estimator: ci* = (A + 'A&,
= (1 Bi)rizi where Bi Y,
y
(Note: The transformation a = H'fl translates this estimator into the ridge
estimator (2.2.12). 7 is either a constant or a function of the sample.)
Generalized ridge estimator: a* = (A + fl 'A& where r is
diagonal,
461 --=-- (1 B i)eq where

Bi = Yi
Yi'
h* = (x ,x+ HrH')-1X'y.
Other ridge and generalized ridge estimators have been proposed by various
authors. In the three following ridge estimators, y is a positive quantity that
does not depend on therefore 111 is inversely proportional to Ai . This is an
intuitively appealing property because it seems reasonable to shrink the com-
ponent with the larger variance more. In the four following generalized ridge
estimators, exactly the opposite takes place: The amount of shrinkage B. is an
increasing function of ill an undesirable property. In some of the estimators,
a2 appears in the formula, and in some, its estimate 62, which is assumed to be
independent of ci, appears. As pointed out by Efron and Morris (1976), the
fundamental properties of Stein's estimators are not changed if 0 2 is indepen-
dently estimated.
Selected Estimators and Their Properties
Ridge Estimators
Ridge 1 (Sclove, 1973)
a2 tr A
a'Act
Ridge 2 (Hoer!, Kennard, and Baldwin, 1975) and Modified Ridge 2
(Thisted, 1976)
1062
Y -
aa
This estimator is obtained by putting A = I in Sclove's estimator. Although
the authors claimed its good properties on the basis of a Monte Carlo study,
Thisted (1976) showed that it can sometimes be far inferior to the maximum
likelihood estimator oi; he proposed a modification, y = (K 2)62/&16t, and
showed that the modified estimator is minimax for some A if a2 is known.
Ridge 3 (Thisted, 1976)
a2
if all d1 <00
doi?
fr.!
=0 otherwise,
where
A 7 14-j.
di
[(E 272) 21T
This estimator is minimax for all A if (7 2 is known. If Ai are constant, this

estimator is reduced to the modified version of Ridge 2. When the Xs are too
spread out, however, it becomes indistinguishable from & (which is minimax).
Generalized Ridge Estimators

Generalized Ridge 1 (Berger, 1975)
(K 2)a 2.1.1
a'A1/4
This estimator is minimax for all A and reduces to Stein's estimator when Ai
are constant.
Generalized Ridge 2 (Berger, 1976)
f(611A2ii),721.1.
a 1A2a
Berger (1976) obtained conditions onfunder which it is minimax and admis-
sible for all A.
Generalized Ridge 3 (Bhattacharya, 1966). B i is complicated and therefore
is not reproduced here, but it is an increasing function of A i and is minimax for
all A.
Generalized Ridge 4 (Strawderman, 1978)
2
where "a2 a2x!.
+ 02 + h +
This estimator is minimax for all A if
1-2(K- 2) ]
0 g. a +2
n
2K
g
n+2
h O.
Results
All the generalized ridge estimators are minimax for all A and generalized
ridge 2 is also admissible, whereas among the ridge estimators only Thisted's
(which is strictly not ridge because of a discontinuity) is minimax for all A.
Because & is minimax with a constant risk, any other minimax estimator
dominates a. However, the mere fact that an estimator dominates ci does not
necessarily make the estimator good in its own right. If the estimator is admis-
sible as well, like Berger's generalized ridge 2, there is no other estimator that
dominates it. Even that, however, is no guarantee of excellence because there
may be an estimator (which may be neither minimax nor admissible) that has
a lower risk over a wide range of the parameter space. It is nice to prove
minimaxity and admissibility; however, we should look for other criteria
of performance as well, such as whether the amount of shrinkage is propor-
tional to the variancethe criterion in which all the generalized ridge esti-
mators fail.
The exact distributions of Stein's or ridge estimators are generally hard to
obtain. However, in many situations they may be well approximated by the
jackknife and the bootstrap methods (see Section 4.3.4).
2.2.7 Monte Carlo and Applications

Thisted (1976) compared ridge 2, modified ridge 2, ridge 3, and generalized
ridge 1 by the Monte Carlo method and found the somewhat paradoxical
result that ridge 2, which is minimax for the smallest subset ofA, performs best
in general.
Gunst and Mason (1977) compared by the Monte Carlo method the esti-
mators (1) least squares, (2) principal components, (3) Stein's, and (4) ridge 2.
Their conclusion was that although (3) and (4) are frequently better than (1)
and (2), the improvement is not large enough to offset the advantages of (1)
and (2), namely, the known distribution and the ability to select regressors.
Dempster, Schatzoff, and Wermuth (1977) compared 57 estimators, be-
longing to groups such as selection of regressors, principal components,
Stein's and ridge, in 160 normal linear models with factorial designs using
both E(fi and E(fi ff)'X'X(/3 fl) as the risk function. The
winner was their version of ridge based on the empirical Bayes estimation of y
defined by
K ,1
12 ot2K.
Y
The fact that their ridge beat Stein's estimator even with respect to the risk
function E(fi 1)'X'X(fl /I) casts some doubt on their design of Monte

Carlo experiments, as pointed out by Efron, Morris, and Thisted in the dis-
cussion following the article.
For an application of Stein's estimator (pulling toward the overall mean),

see Efron and Morris (1975), who considered two problems, one of which is
the prediction of the end-of-season batting average from the averages of the
first forty at-bats. For applications of ridge estimators to the estimation of
production functions, see Brown and Beattie (1975) and Vinod (1976). These
authors determined 7 by modifications of the Hoerl and Kennard ridge trace
analysis. A ridge estimator with a constant y is used by Brown and Payne
(1975) in the study of election night forecasts. Aigner and Judge (1977) used
generalized ridge estimators on economic data (see Section 2.2.8).
2.2.8 Stein's Estimator versus Pre Test Estimators

-
Let a N(a, 0 2I) and S ale, (independent of a). Consider the strategy:

Test the hypothesis a = 0 by the F test and estimate a by 0 if the hypothesis is

accepted (that is, if ,S-- 'ii'ci d for an appropriate d) and estimate a by ci if the
hypothesis is rejected. This procedure amounts to estimating a by the estima-
tor /dot, where /d is the indicator function such that Id= 1 ifS- Iikii > d and 0
otherwise. Such an estimator is called a preliminary-test estimator, or a pre-
test estimator for short. Sclove, Morris, and Radhakrishnan (1972) proved
that /ddi is dominated by Stein's positive-rule estimator [1 (doSI a' a)]it.' for
some do such that d < do < 2(K 2)I(n + 2).
A pre-test estimator is commonly used in the regression model. Often the
linear hypothesis Q'fl = c is tested, and the constrained least squares estima-
tor (1.4.11) is used if the F statistic (1.5.12) is smaller than a prescribed value
and the least squares estimator is used otherwise. An example of this was
considered in Eq. (2.1.20). The result of Sclove, Morris, and Radhalcrishnan
can be extended to this regression situation in the following way. 9
Let R be as defined in Section 1.4.2 and let A = (Q, R)'. Define Z = XA - ',
y* = y Xf, and y = A(fl f) where f is any vector satisfying c = Q'f. Then
we can write Model 1 as
y*--Zy+ u. (2.2.16)
Partition Z and y conformably as Z = (Z 1 , Z2), where Z 1 = XQ(Q'Q)- ' and

Z2 = XR(12/11) -1, and as y' = (7;, )9, where yi = QV 0 and 72 --=
R'(fl f). Then the hypothesis Q'il = c in the model y = Xfl + u is equiva-
lent to the hypothesis y i = 0 in the model y* = Zy + u.
Define w = (vvo,
WI, WO as vv2 = (zz2) "2 z; vvi =
(Zii1) -1/2Z; ,
where Z 1 = [I Z2(ZV.2) -1 Z]Z1 ; and Wo is a matrix satisfying WW 1 = 0,
WW2 =0, and W;Wo = I. Then we have

W' y* N(W ' Zy, a2I), (2.2.17)
and the hypothesis yi =0 in model (2.2.16) is equivalent to the hypothesis
W Zy =0 in model (2.2.17). Define Id= 1 ifS-l y*'ZI(ZIZI)-1 Z4* > dand
0 otherwise and define D = (1 [41/y*'Z I(Zci 1 )-1Z41)+ . Then the
aforementioned result of Sclove et al. implies that
0 0
Wy*] + (1 D) [ 0 dominates
D[ WiY * WiY *
[ Wlyd + (1 /d) [ 0
Vkl Y * WiY *
in the estimation of W'Zy. Therefore, premultiplying by W (which is the
inverse of W'), we see that
DZ(Z'Z)- 'Z'y* + (1 D)Z 2(Z;Z2)- IZ;y* Z9 (2.2.18)
dominates
/aZ(Z'Z)- 'Z'y* + (1 /d)Z2(Z2Z2)- 'Z;y* Z (2.2.19)
in the estimation of Zy. Finally, we conclude that Stein's positive-rule estima-
tor 9 defined by (2.2.18) dominates the pre-test estimator 9 defined by (2.2.19)
in the sense that E(S; y)'Z'Z(9 y) yYZ'Z(i y) for all y.
Although the preceding conclusion is the only known result that shows the
dominance of a Stein-type estimator over a pre-test estimator (with respect to
a particular risk function), any Stein-type or ridge-type estimator presented in
the previous subsections can be modified in such a way that "shrinking" or
"pulling" is done toward linear constraints Q'fl = c.
We can assume Q'Q = I without loss of generality because if Q'Q # I, we
can define Q* = Q(Q'Q) -1/2 and c* = (Q/Q) -1/2c so that Q*'fl = c* and
Q*'Q* = I. Denoting the least squares estimator of ft by fi, we have
Q'ft c N[Q'fi c, cr 2Q'(X'X)-1Q]. (2.2.20)
Defining the matrix G such that G'G = I and G'Q'(X'X) - 'QG =
diagonal, we have
G'(Q'ft c) /V[G 1 (Q'fl c), a2 M-1]. (2.2.21)
Therefore, if B is the diagonal matrix with the ith diagonal element B. defined
under any of the minimax estimators presented in Section 2.2.6, we have

(I B)G'(Qqj c) ..c G'(Q'ft c) for Gi(Q'fi e), (2.2.22)
where "of 4 62 for 0" means that
E(61 /Mei 0) 5 E(02 0)'(62 0)
for all 0 with strict inequality holding for at least one value of 0. Therefore we
have
GDG'(Qift c) -zir Q'ft c for Q'fl c, (2.2.23)
where D -- I B, and consequently
GDG'(Q'fri--
c) C ] [ Q for [Q1 (2.2.24)
[
Rift ]
where R is as defined in Section 1.4.2 with the added condition R'R = I. Then
[ Q't1GDG'(Q'k c) + c] .c f.i for ii,
(2.2.25)
R' R'fl
which can be simplified as
ils QGBG'((r/i c) -4C n for fl. (2.2.26)
Let us consider a concrete example using generalized ridge estimator 1.

Putting
(K 2)(72
B ,,,,, 12 a
I
in the left-hand side of (2.2.26), we obtain the estimator
'0 .., (K 2)a2

(Q' er iQ'OUX) -10:2] -2(Q1) c) Q[4:Y(rX)-1Q]-1((113 c).

(2.2.27)
Aigner and Judge (1977) applied the estimator (2.2.27) and another estima-
tor attributed to Bock (1975) to the international trade model of Baldwin
(1971) and compared these estimates to Baldwin's estimates, which may be
regarded as pre-test estimates because Baldwin utilized a certain linear restric-
tion. Aigner and Judge concluded that the conditions under which Bock's
estimator is minimax are not satisfied by the trade data and that although
Berger's estimator (2.2.27) is always minimax, it gives results very close to the
least squares in the trade model.
2.3 Robust Regression

2.3.2 Introduction
In Chapter 1 we established that the least squares estimator is best linear

unbiased under Model 1 and best unbiased under Model 1 with normality.
The theme of the previous section was essentially that a biased estimator may
be better than the least squares estimator under the normality assumption.
The theme of this section is that, in the absence of normality, a nonlinear
estimator may be better than the least squares estimator. We shall first discuss
robust estimation in the i.i.d sample case, and in the next subsection we shall
generalize the results to the regression case.
2.3.2 Independent and Identically Distributed Case
Let Y1, Y2, , Y T be independent observations from a symmetric distribu-

tion function FRy ,u)lcr] such that F(0) = I. Thus ,u is both the population
mean and the median. Here a represents a scale parameter that may not
necessarily be the standard deviation. Let the order statistics be y(1) 5
Y(2) g g Y. We define the sample median A to be v , ((T+ I )/2) if T is odd
and any arbitrarily determined point between AT12) and 3/0772)4_ 0 if Tis even. It
has long been known that fl would be a better estimator of# than the sample
mean ,a = 7 ' Ir_ 1 yr if F has heavier tails than the normal distribution.
Intuitively speaking, this is because ii is much less sensitive to the effect of a
few wild observations than fi is.
It can be shown that fi is asymptotically normally distributed with mean #
and variance [4 Tf(0)2]', where f is the density of F.1 Using this, we can
compare the asymptotic variance of A with that ofii for various choices of F.
Consider the three densities
1
Normal e 212
4-27c
I
Laplace e-lx1
2
1 1
Cauchy
r (1 + x2) .
Table 2.3 shows the asymptotic variances of 12 and fi under these three distri-
butions. Clearly, the mean is better than the median under normality but the
Table 2.3 Asymptotic variances (times sample size) of the sample mean and the
sample median under selected distributions
TV(11)
Normal 1.57
Laplace 2 1
Cauchy co 2.47
median outperforms the mean in the case of the other two long-tailed distri-
butions. (Note that the mean is the maximum likelihood estimator under
normality and that the median, because it minimizes I ly, bl, is the maxi-

mum likelihood estimator under the Laplace distribution.) In general, we call

an estimator, such as the median, that performs relatively well under distribu-
tions heavier-tailed than normal a "robust" estimator. To be precise, there-
fore, the word robustness should be used in reference to the particular class of
possible distributions imagined. A comparison of the variances of the mean
and the median when the underlying class of distributions is a mixture of two
normal distributions can be found in Bickel and Doksum (1977, p. 371),
which has an excellent elementary discussion of robust estimation.
Another robust estimator of location that has long been in use is called the
a trimmed mean, which is simply the mean of the sample after the proportion
-
a of largest and smallest observations have been removed. These and other
similar robust estimators were often used by statisticians in the nineteenth
century (see Huber, 1972; Stigler, 1973). However, the popularity of these
robust estimators declined at the turn of the century, and in the first half of the
present century the sample mean or the least squares estimator became the
dominant estimator. This change occurred probably because many sophisti-
cated testing procedures have been developed under the normality assump-
tion (mathematical convenience) and because statisticians have put an undue
confidence in the central limit theorem (rationalization). In the last twenty
years we have witnessed a resurgence of interest in robust estimation among
statisticians who have recognized that the distributions of real data are often
significantly different from normal and have heavier tails than the normal in
most cases. Tukey and his associates in Princeton have been the leading
proponents of this movement. We should also mention Mandelbrot (1963),
who has gone so far as to maintain that many economic variables have infinite
variance. However, it should be noted that the usefulness of robust estimation
is by no means dependent on the unboundedness of the variance; the occur-
rence of heavier tails than the normal is sufficient to ensure its efficacy. For a
survey of recent developments, the reader is referred to Andrews et al. (1972),
who reported on Monte Carlo studies of 68 robust estimators, and to Huber
(1972, 1977, 1981), Hogg (1974), and Koenker (1981a).
Robust estimators of location can be classified into four groups: M, L p , L,
and R estimators. M, L, and R estimators are the terms used by Huber (1972).
Lp estimators constitute a subset of the class M, but we have singled them out
because of their particular importance. We shall briefly explain these classes of
estimators and then generalize them to the regression case.
M Estimator. The M estimator (stands for "maximum-likelihood-type"
estimator) is defined as the value of b that minimizes Mr_ ,p R y, b)/s] where s
-
is an estimate of the scale parameter a and pis a chosen function. If p is twice

differentiable and its second derivative is piecewise continuous with
Ep'[(y, u)/ s 0] = 0 where so is the probability limit of s," we can use the
results of Chapter 4 to show that the M estimator is asymptotically normal
with mean it and variance
T- 1 s2 gll[s(71(Y1 11)] 2) (2.3.1)

(EP" [cl(Yr 10 2.
Note that when X)) = .12, this formula reduces to the familiar formula for
variance of the sample mean.
Consider an M estimator proposed by Huber (1964). It is defined by
p(z) = iz2 if Izi < c (2.3.2)

---- clzi ic2 if Izi c,
where z = (y p)Is and cis to be chosen by the researcher. (The Monte Carlo
studies of Andrews et al. (1972) considered several values of c between 0.7and
2.) Huber (1964) arrived at the p function in (2.3.2) as the minimax choice
(doing the best against the least favorable distribution) when F(z)==
(1 e)(1)(z)+ ell(z), where z = (y /1)1a, H varies among all the symmetric
distributions, 4) is the standard normal distribution function, and e is a given
constant between 0 and 1. The value of c depends on e in a certain way. As for
s, one may choose any robust estimate of the scale parameter. Huber (1964)
proposed the simultaneous solution of
(2.3.3)
and
T (y, b)2 T , 222
p, p (z) e z dz (2.3.4)
r- s lin --
in terms of b and s. Huber's estimate ofy converges to the sample mean or the
sample median as c tends to 00 or 0, respectively.
Another M estimator shown to be robust in the study of Andrews et al.
(1972) is the following one proposed by Andrews (1974). Its p function is
defined by
p(z) = 1 + cos z if 121 g it (2.3.5)
=0 if Izl > 7r,
where z = (y b)ls as before. Andrews' choice ofs is (2.1) Median (Iy,
Estimator. The LI, estimator is defined as the value of b that minimizes

Mr_ ly, bIP. Values of p between 1 and 2 are the ones usually considered.
Clearly, p = 2 yields the sample mean and p = 1 the sample median. For any
p 1, the asymptotic variance of the estimator is given by (2.3.1). The ap-
proximate variance for the case p = 1 (the median) was given earlier. Note that
an estimate of a scale parameter need not be used in defining the Li, estimator.
L Estimator. The L estimator is defined as a linear combination of order
statistics Y(j) y(2) .1 ym. The sample median and the a-trimmed
mean discussed at the beginning of this subsection are members of this class.
(As we have seen, the median is a member of L I, and hence of M as well.)
Another member of this class is the Winsorized mean, which is similar to a
trimmed mean. Whereas a trimmed mean discards largest and smallest obser-
vations, a Winsorized mean "accumulates" them at each truncation point.
More precisely, it is defined as
T- I [(g + 1)y(g+ + As+2) + + yq_ g_ (g + 1)3v_.4] 2.3.6)
(
for some integer g.

A generalization of the median is the 0 th sample quantile, 0 < 6 < 1; this is
defined as yoo , where k is the smallest integer satisfying k> TO if TO is not an
integer, and an arbitrarily determined point between yu and ya.01) if TO is an
-9)
integer. Thus 0= corresponds to the median. It can be shown that the 0th
sample quantile, denoted by ii(0), minimizes
I Oly, bl+ yeCb (1

y,gb
19)1Y( bl. (2.3.7)
Previously we have written ,a(i) simply as /2. Gastwirth (1966) proposed a

linear combination of quantiles 0.3/1(i) + 0.4/7(i) + 0.317(i) as a robust esti-
mator of location. Its asymptotic distribution can be obtained from the follow-
ing general result attributed to Mosteller (1946): Suppose 0 < 0 1 <
02 < < 0 < 1. Then ii(01 ), /7(02), . . . ii(0) are asymptotically
jointly normal with the means equal to their respective population quantiles,
4u(01), /OA , it(0) (that is, 0, = FIP(0 0, and variances and covar-
iances given by
Of)
Coy 1/ 1(01), 1-1(91' (nu Tf[1t(0W[10.0 i j. (2.3.8)
R Estimator. The rank is a mapping from n real numbers to the integers 1

through n in such a way that the smallest number is given rank 1 and the next
smallest rank 2 and so on. The rank estimator ofy, denoted byy*, is defined as
follows: Construct a sequence of n = 2T observations x1 , x2 , . . . , x by
defining xi = b, i = 1, 2, . . . , T, and xn.. i = b yi, = 1, 2, . . . , T,
and let their ranks be R I , R2, . . . R. Then g * is the value of b that satisfies
Ri
0,
(2n + 1
where J is a function with the property f W(A) dA = 0. Hodges and Lehmann

(1963) proposed setting J(A) = A I. For this choice of Jo.), it* can be shown
to be equal to Median ((yi + 02), 1 j T. It is asymptotically normal
with mean it and variance
1
(2.3.9)
12 T [ f f 2(y) dy]2.
Remarks
We have covered most of the major robust estimators of location that have
been proposed. Of course, we can make numerous variations on these estima-
tors. Note that in some of the estimation methods discussed earlier there are
parameters that are left to the discretion of the researcher to determine. One
systematic way to determine them is the adaptive procedure, in which the
values of these parameters are determined on the basis of the information
contained in the sample. Hogg (1974) surveyed many such procedures. For
example, the a of the a-trimmed mean may be chosen so as to minimize an
estimate of the variance of the estimator. Similarly, the weights to be used in

the linear combination of order statistics may be determined using the
asymptotic variances and covariances of order statistics given in (2.3.8).
Most of the estimators discussed in this subsection were included among the
nearly seventy robust estimators considered in the Monte Carlo studies of
Andrews et al. (1972). Their studies showed that the performance of the
sample mean is clearly inferior. This finding is, however, contested by Stigler
(1977), who used real data (eighteenth and nineteenth century observations
on physical constants such as the speed of light and the mean density of the
earth for which we now know the true values fairly accurately). He found that
with his data a slightly trimmed mean did best and the more "drastic" robust
estimators did poorly. He believes that the conclusions of Andrews et al. are
biased in favor of the drastic robust estimators because they used distributions
with significantly heavy tails as the underlying distributions. Andrews et al.
did not offer definite advice regarding which robust estimator should be used.
This is inevitable because the performance of an estimator depends on the
assumed distributions. These observations indicate that it is advisable to per-
form a preliminary study to narrow the range of distributions that given data
are supposed to follow and decide on which robust estimator to use, if any.
Adaptive procedures mentioned earlier will give the researcher an added
flexibility.
The exact distributions of these robust estimators are generally hard to
obtain. However, in many situations they may be well approximated by
methods such as the jackknife and the bootstrap (see Section 4.3.4).
2.3.3 Regression Case
Let us generalize some of the estimation methods discussed earlier to the

regression situation.
M Estimator. The M estimator is easily generalized to the regres-
sion model: It minimizes .p [(y1 x;b)/s] with respect to the vector
b. Its asymptotic variance-covariance matrix is given by
si(X'AX) - I X 'BX(X'AX) - where A and B are diagonal matrices with the t th
diagonal elements equal to Ep"[(ye x31)Is 0] and E (p'[(y, x;13)1s 0]2), re-
spectively.
Hill and Holland (1977) did a Monte Carlo study on the regression general-
ization of Andrews' M estimator described in (2.3.5). They used s = (2.1)
Median (largest T K + 1 of 1y1 x;ftl ) as the scale factor in the p function,
where 11 is the value of b that minimizes ET.. 1 1y, 34bl. Actually, their estima-
tor, which they called one-step sine estimator, is defined as
fts = (X'DX) -1 X'Dy, (2.3.10)
where D is the diagonal matrix whose tth diagonal element d is defined by
d, [Yr 1-1 sin [ Y` if jy, 34%31 ..5 sn (2.3.11)
=0 if ly, x;i3l>
This is approximately the first step of the so-called Newton-Raphson iteration
designed to minimize (2.3.5), as we shall show next.
Put g(fl)= 7.. 1 p(z,) where z,--= xAls and expand g(//) around /3 = fi
in a Taylor series as
g(fo ag + 82
g (p fl), (2.3.12)
P) P P) (Via" 11 Ph
where the derivatives are evaluated atfl. Let fts be the value of/3 that minimizes
the right-hand side of (2.3.12). Thus
a 2g 1 2(_g_
(2.3.13)
[ afia" air
This is the first step in the Newton-Raphson iteration (see Section 4.4.1).
Inserting
ag
(2.3.14)
t-i
and
82g
(2.3.15)
aim"
where p' and p" are evaluated at (y, into (2.3.13), we obtain
T ) --1 T
= r2p"xix; (s- tp' xi r 2p"x,x',",4). (2.3.16)

1 1 P-1
Finally, inserting a Taylor approximation
p' p'(0)+ x'Ap" (2.3.17)
into the right-hand side of (2.3.16), we obtain the estimator (2.3.10).
In their Monte Carlo study, Hill and Holland assumed that the error term is
N(0, 1) with probability 1 a and N(0, c 2) with probability a, with various
values for c and a. The regressor matrix was artificially generated with various
degrees of outlying observations and with the number of regressors ranging
from 1 to 6. The sample size was chosen to be 20. The reader should look at the
table in their article (Hill and Holland, 1977) to see the striking improvement
of /3 or fls over the least squares estimator and the minor improvement of fis
over1.
Lp Estimator. This class also can easily be generalized to the regression
model. It is the value of b that minimizes Iy x; ND. A special case in
which p = 1, which was already defined as iO in the discussion about the M
estimator, will be more fully discussed later as an L estimator. For p 1, the
asymptotic variance of the Lp estimator can be obtained by the same formula
as that used for the asymptotic variance of the M estimator. Forsythe (1972)
conducted a Monte Carlo study of L estimators for p = 1.25, 1.5, 1.75, and 2
(least squares) in a regression model with one fixed regressor and an intercept
where the error term is distributed as GN(0, 1) + (1 G)N(S, R) for several
values of G, S, and R. His conclusion: The more "contamination," the smaller
p should be.
L Estimator. The 9 th sample quantile can be generalized to the regression
situation by simply replacing b by x; b in the minimand (2.3.7), as noted by
Koenker and Bassett (1978). We shall call the minimizing value the 9 th
sample regression quantile and shall denote it by AO). They investigated
the conditions for the unique solution of the minimization problem and ex-
tended Mosteller's _result to the regression case. They established that
A02), Aen) are asymptotically normal with the means equal to
+ ,u(0 I ), fi + p(0 2), . . . , 13 + /(O ), where p(k) = [,u(0 i), 0, 0, . . . , O]',
and the variance-covariance matrix is given by
Coy [i)(6i), ;4(03)] = cou(X 1 X) ', i j, (2.3.18)
where coy is given in (2.3.8). A proof for the special case ft() = ft is also given in
Bassett and Koenker (1978) (see Section 4.6.2).
Blattberg and Sargent (1971) conducted a Monte Carlo study and corn-
pared fl, the least squares estimator, and one other estimator in the model with
the regressor and no intercept, assuming the error term has the characteristic
function exp(I ala) for a = 1.1, 1.3, 1.5, 1.7, 1.9, and 2.0. Note that a = 2
gives the normal distribution and a = 1 the Cauchy distribution. They found
that ft did best in general.
Schlossmacher (1973) proposed the following iterative scheme to obtain

= (X 'DX)- I X' Dy, (2.3.19)
where D is the diagonal matrix whose tth diagonal element is given by
d, = ly 410-01 -1 if y x(41-01> E (2.3.20)
=0 otherwise,
where & is the estimate obtained at the ith iteration and E is some predefined
small number (say, E = 10-7). Schlossmacher offered no theory of conver-
gence but found a good convergence in a particular example.
Fair (1974) proposed the following alternative to (2.3.20):
cl;" = x'Ai-01 `, -1 ). (2.3.21)
Thus Fair bounded the weights from above, whereas Schlossmacher threw out
observations that should be getting the greatest weight possible. It is clear that
Fair's method is preferable because his weights are continuous and nonin-
creasing functions of the residuals, whereas Schlossmacher's are not.
A generalization of the trimmed mean to the regression model has been
proposed by several authors. We shall consider two such methods. The first,
which we shall call 11(a) for 0 <a < I, requires a preliminary estimate, which
we shall denote by A. The estimation process involves calculating the resid-
uals from A, throwing away those observations corresponding to the [Ta]
smallest and [Ta] largest residuals, and then calculating /3(a) by least squares
applied to the remaining observations.
The second method uses the 0 th regression quantile ft(0) of Koenker and
Bassett (1978) mentioned earlier. This method involves removing from the
sample any observations that have a residual from fl(a) that is negative or a
residual from /3(1 a) that is positive and then calculating the LS estimator
using the remaining observations. This estimator is denoted by fl s (a).
Ruppert and Carroll (1980) derived the asymptotic distribution of the two
estimators and showed that the properties of/3(a) are sensitive to the choice of
the initial estimate /30 and can be inefficient relative to ft* (a). However, if
= I [fl(a) + a)] and the distribution of the error term is symmetric,
fl(a) is asymptotically equivalent to II* (a).
R Estimator. This type of regression estimator was proposed by Jaeckel
(1972). He wrote the regression model as y = /301 + Xfl + u, where X is now
the usual regressor matrix except I, the column of ones. Jaeckel's estimator
minimizes D(y Xb) = [R, (T+ 1)/2](y, x;b), where R, =

rank(y, b). Note that R, is also a function of b. Jaeckel proved that D is a
nonnegative, continuous, and convex function of b and that his estimator is
asymptotically normal with mean II and variance-covariance matrix
T2 txt [1 _
11 ] xy. (2.3.22)
where T2 = 12- i[ff 2(u) du] -2. The ratio a 2/T2 is known as the Pitman effi-
ciency of the Wlcoxon rank test and is equal to 0.955 if f is normal and
greater than 0.864 for any symmetric distribution, whereas its upper bound is
infinity. Because the derivative of D exists almost everywhere, any iterative
scheme of minimization that uses only the first derivatives can be used.
(Second derivatives are identically zero.) The intercept /i 0 may be estimated by
the Hodges-Lehmann estimator, Median((a, + 12)/2), 1 jgT, where
is the vector of the least squares residuals. See the articles by McKean and
Hettmansperger (1976) and Hettmansperger and McKean (1977) for tests of
linear hypotheses using Jaeckel's estimator.
Exercises
1. (Section 2.1.3)
Show that the Bayesian minimand (2.1.5) is minimized when Sis chosen to
be the set of y that satisfies the inequality (2.1.3).
2. (Section 2.1.5)
A weakness of PC is that it does not choose the right model with probability
1 when Tgoes to infinity. (The weakness, however, is not serious.) Suppose
we must choose between regressor matrix X, and X such that X, c X. Show
that
lim P [PC chooses X, IX, is true] = P[XX-K, < 2K] < 1.
3. (Section 2.1.5)
Schwartz's (1978) criterion minimizes T log y'Miy K, log T. Show that
this criterion chooses the correct model with probability 1 as T goes to co.
4. (Section 2.2.3)
If F',5' is estimable, show that F'fl is the BLUE of Pfl, where fi is the LS
estimator.
5. (Section 2.2.3)
Show that fl , defined in (2.2.10) is a solution of (2.2.6).
6. (Section 2.2.4)
Show that for any square matrix A there exists a positive constant y o such
that for all y> yo , A + yl is nonsingular.
3 Large Sample Theory
Large sample theory plays a major role in the theory of econometrics because
econometricians must frequently deal with more complicated models than
the classical linear regression model of Chapter 1. Few finite sample results are
known for these models, and, therefore, statistical inference must be based on
the large sample properties of the estimators. In this chapter we shall present a
brief review of random variables and the distribution function, discuss various
convergence theorems including laws of large numbers and central limit
theorems, and then use these theorems to prove the consistency and the
asymptotic normality of the least squares estimator. Additional examples of
the application of the convergence theorems will be given.
3.1 A Review of Random Variables and the Distribution Function
This section is not meant to be a complete discussion of the subject; the reader
is assumed to know the fundamentals of the theory of probability, random
variables, and distribution functions at the level of an intermediate textbook
in mathematical statistics.' Here we shall introduce a few concepts that are not
usually dealt with in intermediate textbooks but are required in the subse-
quent analysis, in particular, the rigorous definition of a random variable and
the definition of the Stieltjes integra1. 2
3.1 .1 Random Variables

At the level of an intermediate textbook, a random variable is defined as a
real-valued function over a sample space. But a sample space is not defined
precisely, and once a random variable is defined the definition is quickly
forgotten and a random variable becomes identified with its probability dis-
tribution. This treatment is perfectly satisfactory for most practical applica-
tions, but certain advanced theorems can be proved more easily by using the
fact that a random variable is indeed a function.
We shall first define a sample space and a probability space. In concrete
terms, a sample space may be regarded as the set of all the possible outcomes of
an experiment. Thus, in the experiment of throwing a die, the six faces of the
die constitute the sample space; and in the experiment of measuring the height
of a randomly chosen student, the set of positive real numbers can be chosen
as the sample space. As in the first example, a sample space may be a set of
objects other than numbers. A subset of a sample space may be called an event.
Thus we speak of the event of an ace turning up or the event of an even number
showing in the throw of a die. With each event we associate a real number
between 0 and 1 called the probability of the event. When we think of a sample
space, we often think of the other two concepts as well: the collection of its
subsets (events) and the probabilities attached to the events. The term proba-
bility space refers to all three concepts collectively. We shall develop an ab-
stract definition of a probability space in that collective sense.
Given an abstract sample space LI, we want to define the collection A of
subsets of CI that possess certain desired properties.
DEFINITION 3. 1. 1. The collection A of subsets of S2 is called a a-algebra if it

satisfies the properties:
(i) E A.
(ii) EEAEEA. (E refers to the complement of E with respect
to a)
(iii) Ej E A, j = 1, 2, . . . U, E; E A.
Given a a-algebra, we shall define over it a real-valued set function satisfy-
ing certain properties.
DEFINITION 3.1.2. A probability measure, denoted by P( ), is a real-valued

set function that is defined over a a-algebra A and satisfies the properties:
(i) E E .A P(E) O.
(ii) P(S2) = 1.
(iii) If (Es ) is a countable collection of disjoint sets in A, then
P(0 E1 )--=
A probability space and a random variable are defined as follows:
DEFINITION 3.1.3. Given a sample space 0, a a-algebra A associated with

S2, and a probability measure P() defined over A, we call the triplet
(Si, A, P) a probability space.3
Large Sample Theory 83
DEFINITION 3.1.4. A random variable on (0, A, P) is a real-valued func-

tion4 defined over a sample space 0, denoted by X(co) for co E 0, such that for
any real number x,
CcolX(w) < x) E A.
Let us consider two examples of probability space and random variables
defined over them.
EXAMPLE 3. 1. 1. In the sample space consisting of the six faces of a die, all the
possible subsets (including the whole space and the null set) constitute a
a-algebra. A probability measure can be defined, for example, by assiging 1/6
to each face and extending probabilities to the other subsets according to the
rules given by Definition 3.1.2. An example of a random variable defined over
this space is a mapping of the even-numbered faces to one and the odd-num-
bered faces to zero.
EXAMPLE 3.1.2. Let a sample space be the closed interval [0, 1]. Consider
the smallest a-algebra containing all the open sets in the interval. Such a
a-algebra is called the collection of Borel sets or a Borel field. This a-algebra
can be shown to contain all the countable unions and intersections of open
and closed sets. A probability measure of a Borel set can be defined, for
example, by assigning to every interval (open, closed, or half-open and half-
closed) its length and extending the probabilities to the other Borel sets ac-
cording to the rules set forth in Definition 3.1.2. Such a measure is called
Lebesgue measure. 5 In Figure 3.1 three random variables, X, Y, and Z, each of
which takes the value 1 or 0 with probability +, are depicted over this probabil-
X( w ) Y(w) Z ( al)
) I 3
0 2"
-
I 0 1 I C -4 I 7$ 1
Figure 3.1 Discrete random variables defined over [0, 1] with Lebesgue measure
ity space. Note that Z is independent of either Xor Y, whereas b and Yare not
independent (in fact XY = 0). A continuous random variable X(w) with the
standard normal distribution can be defined over the same probability space
by X = cD-1 (co), where clo is the standard normal distribution function and
" 1" denotes the inverse function.
3.1.2 Distribution Function
DEFINITION 3.1.5. The distributionfunction F(x) of a random variable X(w)

is defined by
F(x) = P(colX(co) < x).
Note that the distribution function can be defined for any random variable
because a probability is assigned to every element of A and hence to
(co IX(w) < x) for any x. We shall write P(colX(co) < x) more compactly as
P(X < x).
A distribution function has the properties:
(i) F( co) = O.
(ii) F(00) = 1.
(iii) It is nondecreasing and continuous from the left.
[Some authors define the distribution function as F(x)= P(colX(co) x).
Then it is continuous from the right.]
Using a distribution function, we can define the expected value of a random
variable whether it is discrete, continuous, or a mixture of the two. This is
done by means of the Riemann-Stieltjes integral, which is a simple generaliza-
tion of the familiar Riemann integral. Let X be a random variable with a
distribution function F and let Y = h(X), where h( ) is Borel-measurable. 6
expected value of Y, denoted by EY as follows. Divide an Wedfinth
interval [a, b] into n intervals with the end points a = xo <x1 <. . . <
x,,_ 1 <x, = b and let xr be an arbitrary point in [x,, x i4., ]. Define the partial
sum
nI
Sn = winv (xi+,) - F(x (3.1.1)
associated with this partition of the interval [a, bl. If, for any e > 0, there exists
a real number A and a partition such that for every finer partition and for any
choice ofx*, 1S A j < e, we call A the Riemann-Stieltjes integral and denote
it by f dF(x). It exists if his a continuous function except possibly for a
countable number of discontinuities, provided that, whenever its discontinu-
ity coincides with that of F, it is continuous from the right.' Finally, we define
EY = f h(x) dF(x) = lirn

aw
b,* jab h(x) (3.1.2)
provided the limit (which may be + COD or 03) exists regardless of the way
a 00 and b 00.
If dF/dx exists and is equal to f(x), F(x 141 ) F(x i ) f(4 )(x 1+1 x i ) for
some x7 E (x1+1 , xi ] by the mean value theorem. Therefore
Eh(X)= f:a h(x)f(x) dx. (3.1.3)
On the other hand, suppose X = ci with probability pi , i = 1, 2, . . . , K.

Take a <c and cK < b; then, for sufficiently large n, each interval contains at
most one of the c;s. Then, of the n terms in the summand of (3.1.1), only K
terms containing c i's are nonzero. Therefore
fab h(x) dF(x)= h(c (3.1.4)
3.2 Various Modes of Convergence

In this section, we shall define four modes of convergence for a sequence of
random variables and shall state relationships among them in the form of
several theorems.
DEFINITION 3.2.1 (convergence in probability). A sequence of random vari-
ables (X) is said to converge to a random variable X in probability if
PDX. Ai > 0 for any > 0. We write X or plim X = X.
DEFINITION 3.2.2 (convergence in mean square). A sequence (X) is said to
converge to X in mean square if E(X X) 2 = 0. We write X, It X.
DEFINITION 3.2.3 (convergence in distribution). A sequence (X.) is said to
converge to X in distribution if the distribution function F X converges to
the distribution function F of X at every continuity point of F. We write
X -4 X, and we call F the limit distribution of (X.). If ) and {Y) have the
same limit distribution, we write X L-12 Y,,.
The reason for adding the phrase "at every continuity point of F" can be
understood by considering the following example: Consider the sequence
F( ) such that
1
F(x)=O, x<a (3.2.1)
(x _ a 1) 1 1
a --
2 n
1
=1, a + < x.
n
Then lirn F. is not continuous from the left at a and therefore is not a
distribution function. However, we would like to say that the random variable
with the distribution (3.2.1) converges in distribution to a degenerate random
variable which takes the value a with probability one. The phrase "at every
continuity point of F" enables us to do so.
DEFINITION 3.2.4 (almost sure convergence). A sequence (X.) is said to
converge to X almost surely 8 if
P(collim X(co)= X(co)) = 1.
We write X 21. X.
The next four theorems establish the logical relationships among the four
modes of convergence, depicted in Figure 3.2. 9
THEOREM 3.2.1 (Chebyshev). 0 = X. * 0.

-1)
Proof We have
f x 2 dF(x).. 2 dF(x), (3.2.2)
0. S.
M
1
p
Figure 3.2 Logical relationships among four modes of convergence
where S = (x1x 2 2). But we have
dF(x)= 1-4 dFn(X) dF(x) (3.2.3)
= F(E) + [1 F.()]
= P(X,, < e) + P(X )
P[X!> 9.
Therefore, from (3.2.2) and (3.2.3), we obtain
p[Al > 2]
(3.2.4)
The theorem immediately follows from (3.2.4).

The inequality (3.2.4) is called Chebyshev's inequality. By slightly modify-
ing the proof, we can establish the following generalized form of Chebyshev's
inequality:
Eg(X)
P[g(X )> 9 -5 , (3.2.5)
2
where g( ) is any nonnegative continuous function.
Note that the statement X I'-40 X = X. X, where X may be either a con-
stant or a random variable, follows from Theorem 3.2.1 if we regard X. Xas
the X of the theorem.
We shall state the next two theorems without proof. The proof of Theorem
3.2.2 can be found in Mann and Wald (1943) or Rao (1973, p. 122). The proof
of Theorem 3.2.3 is left as an exercise.
THEOREM 3.2.2. X - LI ) X X 1> X.
THEOREM 3.2.3. X X IL) X.
The converse of Theorem 3.2.2 is not generally true, but it holds in the
special case where Xis equal to a constant a. We shall state it as a theorem, the
proof of which is simple and left as an exercise.
THEOREM 3.2.4. X`I'aX 11 a.
The converse of Theorem 3.2.3 does not hold either, as we shall show by a
well-known example. Define a probability space (0, A, P) as follows: =
[0, 1], A = Lebesgue-measurable sets in [0, 1], and P = Lebesgue measure as
in Example 3.1.2. Define a sequence of random variables X(w) as

X i (co) = 1 fo r 0 '6 co 6. 1
1
X2( co) = 1 for 0 -6 co
2
=0 elsewhere
X3(co) = 1 for
2 2 3
=0 elsewhere
1 1 1
X4(co) = 1
=0 elsewhere
X5(co) = 1 12 12 5
=0 elsewhere
In other words, the subset of over which X assumes unity has the total
length 1/n and keeps moving to the right until it reaches the right end point of
[0, 1], at which point it moves back to 0 and starts again. For any 1 > e > 0, we
clearly have
1
and therefore I,, 0. However, because Mr_ i i-1 = co, there is no element in
fl for which lim n, X(w) = 0. Therefore P(collim n X (o))= 0) = 0, im-
plying that X not converge to 0 almost surely.
The next three convergence theorems are extremely useful in obtaining the
asymptotic properties of estimators.
THEOREM 3.2.5 (Mann and Wald). Let X and X be K-vectors of random
variables and let g( ) be a function from RK to R such that the set E of
discontinuity points of g( ) is closed and P(X E E) = 0. If X -d--* X, then
g(X) g(X).
A slightly more general theorem, in which a continuous function is replaced
by a Borel measurable function, was proved by Mann and Wald (1943). The
convergence in distribution of the individual elements of the vector X to the

corresponding elements of the vector X is not sufficient for obtaining the
above results. However, if the elements of X are independent for every n, the
separate convergence is sufficient.
THEOREM 3.2.6. Let X be a vector of random variables with a fixed finite
number of elements. Let ,f be a real-valued function continuous at a constant
vector point a. Then Xn 0 a g(X) 110 g(a).
Proof Continuity at a means that for any > 0 we can find (5 such that
II X all < (5 implies Ig(X) g(a)i< E. Therefore
P[IIX all < (5] P[Ig(Xn ) g(a) I < 1. (3.2.6)
The theorem follows because the left-hand side of (3.2.6) converges to 1 by the
assumption of the theorem.
THEOREM 3.2.7 (Slutsky). If Xn X and Y ) a, then
(i) Xn + Y X + a,
(ii) X Y S-Lo aX,
(iii) (X/ Y ) X/a, provided a # 0.
The proof has been given by Rao (1973, p. 122). By repeated applications of
Theorem 3.2.7, we can prove the more general theorem that if g is a rational
function and plim Y = a, i = 1, 2, . . . , .I, and Xm X i jointly in all i =
1, 2, . . . , K, then the limit distribution of g(X I , X2, . . . ,X, , Yzn
, ) is the same as the distribution of g(X i , X2, , X,,, a1 , ,
. . . , c(1). This is perhaps the single most useful theorem in large sample
theory.
The following definition concerning the stochastic order relationship is
useful (see Mann and Wald, 1943, for more details).
DEFINITION 3.2.5. Let (X.) be a sequence of random variables and let
(a n ) be a sequence of positive constants. Then we can write Xn = o(an ) if
a,TIX = 0 and X.= 0(a.) if for any E > 0 there exists an Ai, such
that
A1E ] 1 E
for all values of n.
Sometimes these order relationships are denoted op and 0,, respectively to
distinguish them from the cases where (X.) are nonstochastic. However, we
use the same symbols for both stochastic and nonstochastic cases because
Definition 3.2.5 applies to the nonstochastic case as well in a trivial way.
3.3 Laws of Large Numbers and Central Limit Theorems

Given a sequence of random variables (X,), define in = n-1 17_, X. A law of
large numbers (LLN) specifies the conditions under which X EX,, con-
verges to 0 either almost surely or in probability. The former is called a strong
law of large numbers and the latter a weak law. In many applications the
simplest way to show X,,-EXI'-> 0 is to show X-EX L I I> 0 and then
apply Theorem 3.2.1 (Chebyshev). We shall state two strong laws of large
numbers. For proofs, the reader is referred to Rao (1973, pp. 114-115).
THEOREM 3.3.1 (Kolmogorov LLN 1). Let (X,) be independent with finite
variance VX, a?. If Ir.., ol/t 2 < cc, then X - EX 0.
THEOREM 3.3.2 (Kolmogorov LLN 2). Let (X,) be i.i.d. (independent and
identically distributed). Then a necessary and sufficient condition that X a >
,u is that EX, exists and is equal to At.
Theorem 3.3.2 and the result obtained by putting ,Y EY,, into the X. of
Theorem 3.2.1 are complementary. If we were to use Theorem 3.2.1 to prove
X - EX L> I 0, we would need the finite variance ofX which Theorem 3.3.2
does not require; but Theorem 3.3.2 requires (X,) to be i.i.d., which Theorem
3.2.1 does not. ' 121
When all we want is the proof of k EX ,, 1, 0, Theorem 3.3.1 does not
add anything to Theorem 3.2.1 because the assumptions of Theorem 3.3.1
imply X - EX,, 4 0. The proof is left as an exercise.
Now we ask the question, What is an approximate distribution ofk when n
is large? Suppose a law of large numbers holds for a sequence (X,) so that
X - EX L> 0. It follows from Theorem 3.2.2 that X-EX 4 0. How-
ever, it is an uninteresting limit distribution because it is degenerate. It is
more meaningful to inquire into the limit distribution of Z= (VX )-112
For, if the limit distribution of Z exists, it should be nondegen- (X.-E)
erate because VZ --= 1 for all n. A central limit theorem (CLT) specifies the
conditions under which Z. converges in distribution to a standard normal
random variable [we shall write Z -0 N(0, 1)].
We shall state three central limit theorems Lindeberg-Levy, Liapounov,
and Lindeberg-Feller and shall prove only the first. For proofs of the other
two, see Chung (1974) or Gnedenko and Kolmogorov (1954). Lindeberg-
Levy and Liapounov are special cases of Lindeberg-Feller, in the sense that the
assumptions of either one of the first two central limit theorems imply those of
Lindeberg-Feller. The assumptions of Lindeberg-Levy are more restrictive in
some respects and less restrictive in other respects than those of Liapounov.
Before we state the central limit theorems, however, we shall define the
characteristic function of a random variable and study its properties.
DEFINITION 3.3.1. The characteristic function of a random variable X is
defined by Eeu
Thus if the distribution function of X is F( ), the characteristic function is
dF(x). The characteristic function is generally a complex number.
However, because e 4tx = cos Ax + i sin ;Lx, the characteristic function of a
random variable with a density function symmetric around 0 is real. The
characteristic functon of N(0, 1) can be evaluated as
x2
1 exp (blx ) dx = exp ) (3.3.1)
2 2
Define go) = Log .1`1%., eaw dF(x), where Log denotes the principal loga-
rithm." Then the following Taylor expansion is valid provided EX' exists:
' (iA)i
g(A) = KJ + o(IAI'), (3.3.2)
j-0
where KJ = (aig/a,V),t_o /if. The coefficients KJ are called the cumulants of X.

The first four cumulants are given by K 1 = EX, K2 = VX, K3 = E(X EX) 3,
and K4 = E(X EX) 4 3( VX) 2.
The following theorem is essential for proving central limit theorems.
THEOREM 3.3.3. If E exp (iAJC)> E exp (i)X) for every A and if
E exp(iAX) is continuous at A = 0, then X. L> X.
The proof of Theorem 3.3.3 can be found in Rao (1973, p. 119).
We can now prove the following theorem.
THEOREM 3.3.4 (Lindeberg-Levy CLT). Let (X,) be i.i.d. with EX,= y and
VX, a2. Then Z N(0, 1).
Proof We assume it =0 without loss of generality, for we can consider
(X, mu) if1,t #0. Define g(A) = Log E exp (iAX,). Then from (3.3.2) we have
go.) = 0.22A2 o(A2).
(3.3.3)
We have, using (3.3.3),
Log E exp (iAZ)= A Log E exp 0[1: ] (3.3.4)
= ng A
ta
22
= n o[
olni.
Therefore the theorem follows from (3.3.1) and Theorem 3.3.3.

THEOREM 3.3.5 (Liapounov CLT). Let (X,) be independent with EX, = Ur,
VX, = (4, and E[IXt m31 . If
n 1-1/2 i[ n 3 1/
urn [1 GI m 3, = 0 ,
^-.) t 1 t-1
then Z. > N(0, 1).
THEOREM 3.3.6 (Lindeberg-Feller CLT). Let (X,) be independent with
distribution functions (Fe ) and EX,= y, and vx,= a. Define Ca
01)1/2. If
1 n
li m (x p,)2 dF,(x)= 0
n--00 n 1 fixit,1>eC
for every e > 0, then z 0 N(0, 1).

In the terminology of Definition 3.2.3, central limit theorems provide con-
ditions under which the limit distribution of Zn ( VX n )'n(X EX) is
N(0, 1). We now introduce the term asymptotic distribution. It simply means
the "approximate distribution when n is large." Given the mathematical
result Z N(0, 1), we shall make statements such as "the asymptotic distri-
bution of Z is N(0, 1)" (written as Z,, N(0, 1)) or "the asymptotic distribu-
tion of X is N(EX, VX)." These statements should be regarded merely as
more intuitive paraphrases of the result Z 10 N(0, 1). Note that it would be
meaningless to say "the limit distribution of X is N(EX, VX)."
When the asymptotic distribution of X is normal, we also say that X is
asymptotically normal.
The following theorem shows the accuracy of a normal approximation to
the true distribution (see Bhattacharya and Rao, 1976, p. 110).
THEOREM 3.3.7. Let (X,) be i.i.d. with EX, = e, and EIX,I 3 = m3 .
Let F be the distribution function of Z and let 41 be that of N(0, 1). Then
IF(x) 01:0(x)Ig (0.7975)

a3 ji
for all x.
It is possible to approximate the distribution function F. of Z more accu-
rately by expanding the characteristic function of Z in the powers of n -1/2. If
EX: exists, we can obtain from (3.3.2)
E exp (az,7) = e- PP[i + K + K 4(111)4 (3.3.5)

6a3 1/11 24a4n
+ ici(a)6 4. of _3/211
72a6n 'n 'j
This is called the Edgeworth expansion (see Cramer, 1946, p. 228). Because
(see Cramer, 1946, p. 106)
J ea"' 40(x) dx = ( )Y e-1212, (3.3.6)
where cf)(6 (x) is the rth derivative of the density ofN(0, 1), we can invert (3.3.5)
to obtain
K3 /IA K4
F.(x) = ('D(x) , (t-',
D (x) + , 4) (4)(x) (3.3.7)
6 e Nin 24a n
K32
+ 4:0(6)(x) + 0(n-312 ).
72a6n
We shall conclude this section by stating a multivariate central limit
theorem, the proof of which can be found in Rao (1973, p. 128).
THEOREM 3.3.8. Let (X) be a sequence of K-dimensional vectors of ran-
dom variables. If cli converges to a normal random variable for every K-di-
mensional constant vector c #0, then X converges to a multivariate normal
random variable. (Note that showing convergence of each element of Ii
separately is not sufficient.)
3.4 Relationships among Jim E, AE, and plim

Let F. be the distribution function ofX and F 0 Fatcontinuity points ofF.
We have defined plim X. in Definition 3.2.1. We define lim E and AE as
follows:
lim EX.= lirn I x dF(x) (3.4.1)

OD
and
AE X f x dF(x). (3.4.2)
In words, AE, which reads asymptotic expectation or asymptotic mean, is the

mean of the limit distribution.
These three limit operators are similar but different; we can construct
examples of sequences of random variables such that any two of the three
concepts either differ from each other or coincide with each other. We shall
state relationships among the operators in the form of examples and theorems.
But, first, note the following obvious facts:
(i) Of the three concepts, only pfim X can be a nondegenerate random
variable; therefore, if it is, it must differ from lim EX or AE X..
(ii) If plim X a, a constant, then AE X = a. This follows immediately
from Theorem 3.2.2.
EXAMPLE 3.4.1. Let X be defined by
X = Z with probability (n 1)/n
= n with probability 1/n,
where Z N(0, 1). Then plim X = Z, firn EX= 1, and AE X = EZ = 0.
EXAMPLE 3.4.2. Let X defined by
X= 0 with probability (n 1)/n
= n2 with probability 1/n.
Then plim X.= AE X= 0, and lim EX = lim n = co.
EXAMPLE 3.4.3. Let X N(a, 1) and l' N(f3, tr- '), where /3# O. Then
XI Y. is distributed as Cauchy and does not have a mean. Therefore
fim E(XIY) cannot be defined either. But, because lc, 10 fl, ARX/Yn ) = aiii
by Theorem 3.2.7 (iii).
The following theorem, proved in Rao (1973, p. 121), gives the conditions
under which fim E = AE.
THEOREM 3.4.1. If EIXIr < M for all n, then lim EX f,= AE X.; for any
s < r. In particular, if EP, <M, then lim EX = AE X. (Note that this con-
dition is violated by all three preceding examples.)
We are now in a position to define two important concepts regarding the
asymptotic properties of estimators, namely, asymptotic unbiasedness and
consistency.
DEFINITION 3.4.1. The estimator O of 0 is said to be asymptotically unbi-
ased if AE 6 = 0. We call AE 0 0 the asymptotic bias.
Note that some authors define asymptotic unbiasedness using lim E instead
of AE. Then it refers to a different concept.
DEFINITION 3.4.2. The estimator on of 0 is said to be a consistent estimator if
plim on = e.
Some authors use the term weakly consistent in the preceding definition, to
distinguish it from the term strong consistency used to describe the property
en Ls4 0. 12
In view of the preceding discussions, it is clear that a consistent estimator is
asymptotically unbiased, but not vice versa.
3.5 Consistency and Asymptotic Normality of Least Squares Estimator

The main purpose of this section is to prove the consistency and the asympto-
tic normality of the least squares estimators of /I and a2 in Model 1 (classical
linear regression model) of Chapter 1. The large sample results of the preced-
ing sections will be extensively utilized. At the end of the section we shall give
additional examples of the application of the large sample theorems.
THEOREM 3.5.1. In Model 1 the least squares estimator is a consistent esti-

mator of ft if As(X'X) 00, where A s(X'X) denotes the smallest characteristic
root of VX.i 3
Proof The following four statements are equivalent:
(i) As(X/X) >
(ii) A/[(XIX)-1 ] -) 0, where Ai refers to the largest characteristic root.
(iii) tr(VX)- 0.
(iv) Every diagonal element of (VX)-1 converges to 0.
Statement (iv) implies the consistency of fl by Theorem 3.2.1.
The reader should verify that it is not sufficient for consistency to assume
that all the diagonal elements of X'X go to infinity.
THEOREM 3.5.2. If we assume that (u s ) are i.i.d. in Model 1, the least squares
estimator er2 defined in (1.2.5) is a consistent estimator of a2.
Proof We can write

a2 = lu'u - lu'Pu, (3.5.1)
where P = X(X/X)- 'X'. By Theorem 3.3.2 (Kolmogorov LLN 2),
T'- 'u'u a2. (3.5.2)
Because u'Pu is nonnegative, we can use the generalized Chebyshev inequality
(3.2.5) to obtain
P( lu'Pu > e) 'u'Pu = cr2E-1 1-1 K. (3.5.3)
Therefore
T-- 'u'Pu 12-0 O. (3.5.4)
Therefore the theorem follows from (3.5.1), (3.5.2), and (3.5.4) by using
Theorem 3.2.6.
We shall prove the asymptotic normality of the least square estimator fin
two steps: first, for the case of one regressor, and, second, for the case of many
regressors.
THEOREM 3.5.3. Assume that (u,) are i.i.d. in Model 1. Assume that K = 1,
and because X is a vector in this case, write it as x. If
firn (x'x) max x? = 0, (3.5.5)
IStST
then a-1 (x/x) 1/2(fr - fl) -0 N(0, 1).
Proof We prove the theorem using the Lindeberg-Feller CLT (Theorem

3.3.6). Take x,u, as the X, of Theorem 3.3.6. Then it, =0 and cd = ext. Let
Fx and Fz be the distribution functions of X and Z, respectively, where
X2 = Z. Then
x 2 dFx(x)= f z dFz (z)

1,2>c2 2>c2
for any c. (This can be verified from the definition of the Riemann-Stieltjes
integral given in Section 3.1.2.) Therefore we need to prove
1 r
Jim

y I 1 >e2a2x'x a dFt(a)= 0
0.1
(3.5.6)
for any e, where F, is the distribution function of (z,u, ) 2. Let G be the distribu-
tion function of 4. Then, because Ft(a)= P(.44 < a) --= P(14 < alx)=
G(a/x?), we have
a dFt(a)=.- a dG(a14)
Le2a2x'z L >e2a2x'x
x?). dG(A).
1A>2a2xT 2lex
Therefore (3.5.6) follows from the inequality
A dG(A) g A dG(A) (3.5.7)

fA>E2a2sr 2x'x A>e 2a2(nutxx- Ilex
r
and the assumption (3.5.5).
THEOREM 3.5.4. Assume that (u,) are i.i.d. in Model 1. Assume also
lim (x;x,)- ' max x =0 (3.5.8)
IStST
for every i = 1, 2, . . . , K. Define Z= XS- I, where S is the K X K diago-

nal matrix the ith diagonal element of which is (xx,) 1/2, and assume that
limr, Z'Z R exists and is nonsingular. Then
S(/) 13) N(0, a212-1 ).
Proof We have Sal fl) = (Z'Z)'Z'u. The limit distribution of

ci(EZ)- 'Z'u for any constant vector c is the same as that of y'Z'u where
y'= c'R-1 . But, because of Theorem 3.5.3, the asymptotic normality of y'Z'u
holds if
K 2
lim (y'Z'Zy)- max (
1
= 0, (3.5.9)
where yk is the kth element of y and z,k is the t,kth element of Z. But, because
y'Z'Zy A.,(Z'Z)y'y by Theorem 10 of Appendix 1, we have
(7 77Y) i InaX (E YkZtk) 2

k-,
(7777)-17'Y max E zi (3.5.10)
max
k-,-
, .,
A.,(Z'Z)
1 K ,2
max
where the first inequality above follows from the Cauchy-Schwarz inequality.
Therefore (3.5.9) follows from (3.5.10) because the last term of (3.5.10) con-
verges to 0 by our assumptions. Therefore, by Theorem 3.5.3,
y'Z'u
N(0, 1)
a(y'Z)1/2
for any constant vector y 0. Since y'Z'Zy c'Et lc, we have
y'Z'u N(0, a2clit lc).
Thus the theorem follows from Theorem 3.3.8.
At this point it seems appropriate to consider the significance of assumption
(3.5.5). Note that (3.5.5) implies x'x co. It would be instructive to try to
construct a sequence for which x'x 000 and yet (3.5.5) does not hold. The
following theorem shows, among other things, that (3.5.5) is less restrictive
than the commonly used assumption that lim T - Ix'x exists and is a nonzero
constant. It follows that if lim T - IX'X exists and is nonsingular, the condition
of Theorem 3.5.4. is satisfied.
THEOREM 3.5.5. Given a sequence of constants (x1 ), consider the state-
ments:
(i) T- Icr = a, where a # 0, a < CO, and c.= x.
CT =
(iii) o.
(iv) max is,57.4 = 0.
Then, (i) [(ii) and (iii)] (iv).
Proof (i) (ii) is obvious. We have
_ _ c T_ I 0.
(3.5.11)
T(T 1) T T 1
Therefore lim T, ( T - 1)-14= 0, which implies (i) (iii). We shall now

prove [(ii) and (iii)] (iv). Given e > 0, there exists 1Tsuch that for any
T .-
-' Ti
cilxi < e (3.5.12)
because of (iii). Given this T 1 , there exists T2 > T1 such that for any T> T2
c -7,' max 4 < E (3.5.13)
i at< T1
because of (ii). But (3.5.12) implies that for any T> T2
C -T-1 4 < for t = T1 , T1 + 1, . . . , T. (3.5.14)
Finally, (3.5.13) and (3.5.14) imply (iv).
It should be easy to construct a sequence for which (3.5.5) is satisfied but (i)
is not.
Next we shall prove the asymptotic normality of the least square estimator
of the variance a2.
THEOREM 3.5.6. Assume that (u,) are i.i.d. with a finite fourth moment
Eu: E--- n1 4 in Model 1. Then VT(6 2 - cr2) N(0, ma oh).
-0 -
Proof We can write

sff via _ 0.2) _ u'u - Ta2 1 ,_,
, W YU (3.5.15)
..ri- vT .
The second term of the right-hand side of (3.5.15) converges to 0 in probability
by the same reasoning as in the proof of Theorem 3.5.2, and the first term can
be dealt with by application of the Lindeberg-Levy CLT (Theorem 3.3.4).
Therefore the theorem follows by Theorem 3.2.7(i).
Let us look at a few more examples of the application of convergence
theorems.
EXAMPLE 3.5.1. Consider Model 1 where K= 1:

y = fix + u, (3.5.16)
where we assume lim 7- 'x'x = c* 0 and (c t ) are i.i.d. Obtain the probability
limit of /3A = yiy/x'y. (Note that this estimator is obtained by minimizing the
sum of squares in the direction of the x-axis.)
We can write
132 2fix'u_ u'u

_
- x'x x'x
IL - x'u
(3.5.17)
x'x
We have E(x'u/x'x) 2 = a2(x'x)' * 0 as T 0 D . Therefore, by Theorem
3.2.1,
x'u
plim =0. (3.5.18)
x'x
Also we have
,. u'u ,. u'u / x'x] 0 2
pu -- m m =p =
n (3.5.19)
x'x T T c
because of Theorem 3.2.6 and Theorem 3.3.2 (Kolmogorov LLN 2). There-
fore, from (3.5.17), (3.5.18), and (3.5.19) and by using Theorem 3.2.6 again,
we obtain
0.2
plim PR= /1 + ft -. (3.5.20)
c
Note that c may be allowed to be co, in which case At
' becomes a consistent
estimator of fl.
EXAMPLE 3.5.2. Consider the same model as in Example 3.5.1 except
that we now assume s lim 7-2x'x = co. Also assume limp__ (x'x)'
maxi stsr x = 0 so that 13 is asymptotically normal. (Give an example of a
sequence satisfying these two conditions.) Show that /I = x'y/x'x and /3,R =
y'y/x'y have the same asymptotic distribution.
Clearly, plim /3 = plim PR = II. Therefore, by Theorem 3.2.2, both estima-
tors have the same degenerate limit distribution. But the question concerns
the asymptotic distribution; therefore we must obtain the limit distribution of
each estimator after a suitable normalization. We can write
fl(x'x)Ii2(i3 fl) + "

(x'x)'/2
(xix)1/2(fiR fl) . (3.5.21)
x'u
il +
lex
But by our assumptions plim u'u( x'x )-1/2 = 0 and plim (x'u/x'x) = 0. There-
fore (x'x)'/2(i3it - ii) and (x'x)I/2(ft - /3) have the same limit distribution by
repeated applications of Theorem 3.2.7.
EXAMPLE 3.5.3. Consider Model 1 with K = 2:
Y xi + fl2x2 (3.5.22)
where we assume that (ur ) are i.i.d. Assume also lim T 1X'X = A, where A is a
2 X 2 nonsingular matrix. Obtain the asymptotic distribution of/31 / 2 , where
fl i and fl2 are the least squares estimators of /3 1 and /32 , assuming )62 # 0.
We can write
- fli) - fli(ii2 - A)] (3.5.23)

ii2/12
Because Om /32 = /32, the right-hand side of (3.5.23) has the same limit
distribution as
/3-2-14T fil) flifli-2 `/T (112 )62).
But, because our assumptions imply (3.5.8) by Theorem 3.5.5, [ -
(ft2 A)] converges to a bivariate normal variable by Theorem 3.5.4.
Therefore, by Theorem 3.2.5, we have
T -I ) -4 N(0, aly'A- iy),

A; - 7-]
[
P2 P 2
where y' = (fly', -
Exercises
1. (Section 3.1.2)
Prove that the distribution function is continuous from the left.
2. (Section 3.2)
Prove Theorem 3.2.3. HINT: Definition 3.1.2 (iii) implies that if il c SI,
for n < m and lim n.... fin = A, then lim_, P(0) = P(A).
3. (Section 3.2)
Prove Theorem 3.2.4.
4. (Section 3.3)
Let (X,) be as defined in Theorem 3.3.1. Prove
E(TY,, - EX )2 = 0.
5. (Section 3.3)
Let (a,), t = 1, 2, . . . , be a nonnegative sequence such that
(MIL, adIT < M for some M and every T. Prove lim,M?:.,(a,It 2)< co.
6. (Section 3.3)
Prove that the conditions of Theorem 3.3.6 (Lindeberg-Feller CLT) fol-
low from the conditions of Theorem 3.3.4 (Lindeberg-Levy CLT) or of
Theorem 3.3.5 (Liapounov CLT).
7. (Section 3.3)
Let (X,) be i.i.d. with EX, p. Then I it. This is a corollary of
Theorem 3.3.2 (Kolmogorov LLN 2) and is called Khinchine's WLLN
(weak law of large numbers). Prove this theorem using characteristic
functions.
8. (Section 3.5)
Show that .1. 3(X1X) > co implies x;x i > co for every i, where x; is the ith
column vector of X. Show also that the converse does not hold.
9. (Section 3.5)
Assume K = 1 in Model 1 and write X as x. Assume that (u,) are indepen-
dent. If there exist L and M such that 0 < L < x'x/ T < M for all T, show
/pH.. fl.
10. (Section 3.5)

Suppose y = y* + u and x = x* + y, where each variable is a vector of T
components. Assume y* and x* are nonstochastic and (u v1 ) is a bivar-
iate i.i.d. random variable with mean 0 and constant variances al
a, respectively, and covariance a. Assume y* = ft*, but y* and x* are
not observable so that we must estimate fi on the basis of y and x. Ob-
tain the probability limit of ft = x'y/x'x on the assumption that
T-1x*'x* = M.
11. (Section 3.5)
Consider the regression equation y = X 1fl1 + X2fl2 + u. Assume all the
assumptions of Model 1 except that X (X 1 , X2 ) may not be full rank.
Let Z be the matrix consisting of a maximal linearly independent subset
of the columns of X2 and assume that the smallest characteristic root of
the matrix (X 1 , Z)'(X I , Z) goes to infinity as T goes to infinity. Derive a
consistent estimator of fly. Prove that it is consistent.
12. (Section 3.5)

Change the assumptions of Theorem 3.5.3 as follows: (u,) are indepen-
dent with Eu, = 0, Vu, = a2, and E1 U11 3 = m3 . Prove a- i(x'x)I/2
(13 fl) 0 N(0, 1) using Theorem 3.3.5 (Liapounov CLT).
13. (Section 3.5)
Construct a sequence (x,) such that Ir_ i 4 > co but the condition (3.5.5)
is not satisfied.
14. (Section 3.5)
Construct a sequence (x,) such that the condition (3.5.5) holds but the
condition (i) of Theorem 3.5.5. does not.
15. (Section 3.5)
Let I be the vector of ones. Assuming lim T - Il'x* = N # 0 in the model of
Exercise 10, prove the consistency offl = l'y/I'x and obtain its asymptotic
distribution.
16. (Section 3.5)
Assume that (u,) are i.i.d. in Model 1. Assume K = 1 and write
X as x. Obtain the asymptotic distribution of )6 = l'y/l'x assuming
lim2, T- I(I'x)2 = 00 where I is the vector of ones.
17. (Section 3.5)
Consider the classical regression model y = ax + flz + u, where a and 11
are scalar unknown parameters, x and z are T-component vectors of
known constants, and u is a T-component vector of unobservable i.i.d.
random variables with zero mean and unit variance. Suppose we are given
an estimator 13 that is independent of u and the limit distribution of
TI/2(ii fl) is N(0, 1). Define the estimator & by
.. x'(y
a
f.tz)
x'x
Assuming Jim T- Ix'x = c and lim T- Ix'z = d, obtain the asymptotic
distribution of 6-1. Assume c #0 and d # 0.
18. (Section 3.5)
Consider the regression model y = fi(x + al) + u, where y, x, I, and u are
T-vectors and a and flare scalar unknown parameters. Assume that 1 is a
T-vector of ones, lim r_ x'l = 0, and lim r_ T- Ix'x = c, where c is a
nonzero constant. Also assume that the elements of u are i.i.d. with zero
mean and constant variance a 2. Supposing we have an estimate of a

denoted by 6 such that it is distributed independently of u and
VT(6/ a) , N(0, A2 ), obtain the asymptotic distribution of fi defined
by
a (x + cit)'y
a ( x + eil)'(x + cil)
4 Asymptotic Properties
of Extremum Estimators
By extremum estimators we mean estimators obtained by either maximizing

or minimizing a certain function defined over the parameter space. First, we
shall establish conditions for the consistency and the asymptotic normality of
extremum estimators (Section 4.1), and second, we shall apply the results to
important special cases, namely, the maximum likelihood estimator (Section
4.2) and the nonlinear least squares estimator (Section 4.3).
What we call extremum estimators Huber called M estimators, meaning
maximum-likelihood-like estimators. He developed the asymptotic proper-
ties in a series of articles (summarized in Huber, 1981). The emphasis here,
however, will be different from his. The treatment in this chapter is more
general in the sense that we require neither independent nor identically dis-
tributed random variables. Also, the intention here is not to strive for the least
stringent set of assumptions but to help the reader understand the fundamen-
tal facts by providing an easily comprehensible set of sufficient conditions.
In Sections 4.4 and 4.5 we shall discuss iterative methods for maximization
or minimization, the asymptotic properties of the likelihood ratio and asymp-
totically equivalent tests, and related topics. In Section 4.6 we shall discuss the
least absolute deviations estimator, for which the general results of Section 4.1
are only partially applicable.
4.1 General Results
4.1.1 Consistency
Because there is no essential difference between maximization and minimiza-
tion, we shall consider an estimator that maximizes a certain function of the
parameters. Let us denote the function by Q Y, 0), where y =
(Y1, Y2 YT) ' is a T-vector of random variables and 0 is a K-vector of
parameters. [We shall sometimes write it more compactly as Q r(0).] The
vector 0 should be understood to be the set of parameters that characterize the
distribution of y. Let us denote the domain of 0, or the parameter space, bye

and the "true value" of 0 by 00 . The parameter space is the set of all the
possible values that the true value 00 can take. When we take various opera-
tions on a function of y, such as expectation or probability limit, we shall use
the value 00 .
As a preliminary to the consistency theorems, we shall define three modes of
uniform convergence of a sequence of random variables.
DEFINITION 4. 1. 1 . Let gT(0) be a nonnegative sequence of random variables

depending on a parameter vector 0. Consider the three modes of uniform
convergence of gr(0) to 0:
(i) sup0E0 gT(0) = 0] = 1,
(ii) P[suPeee MO) < el = 1 for any e > 0,
(iii) lim.infeEe P[gT (0)< e] = 1 for any E > 0.
i ) holds, we say gr(0) converges to 0 almost surely uniformly in 0 E 0. If (ii)
If(i)
holds, we say gT(0) converges to 0 in probability uniformly in 0 E 0. If (iii)
holds, we say gr(0) converges to 0 in probability semiuniformly in 0 E 0.
It is easy to show that (i) implies (ii) and (ii) implies Consider an
example of a sequence for which (iii) holds but (ii) does not. Let the parameter
space 0 be [0, 1] and the sample space S/ also be [0, 1] with the probability
measure equal to Lebesgue measure. For 0 E 0 and co E SI, define gT(co, 0) by
i+ 1
gr(a), 0) = 1 if 0 = and gcog
T T
0, 1, . . . , T 1,
= 0 otherwise.
Then, for 0 < E < 1,
infeEe, P[gT(co, 0) < e] = (T 1)IT and
P[supee Ma), 0) < e] =0 for all T.
Now we shall prove the consistency of extremum estimators. Because we
need to distinguish between the global maximum and a local maximum, we
shall present two theorems to handle the two cases.
THEOREM 4.1.1. Make the assumptions:

(A) The parameter space 0 is a compact subset of the Euclidean K-space
(Ric). (Note that 00 is in 0.)
Asymptotic Properties of Extremum Estimators 107
(B) Q T(y, 0)is continuous in 0 E 0 for ally and is a measurable function of

y for all 0 E 0.
(C) T- 'Q(0) converges to a nonstochastic function Q(0) in probability
uniformly in 0 E 0 as T goes to oo, and Q(0) attains a unique global maximum
at 00 . (The continuity of Q(0) follows from our assumptions.)
Define OT as a value that satisfies
QT(OT) max QT(0). (4.1.1)
eEe
[It is understood that if o r is not unique, we appropriately choose one such
value in such a way that 0 T(y)is a measurable function of y. This is possible by
a theorem ofJennrich (1969, p. 637).] Then O T converges to 00 in probability.'
Proof Let N be an open neighborhood in RK containing 00 . Then KT n 0,
where N is the complement of N in R", is compact. Therefore max eEKtne Q(0)
exists. Denote
e = Q(090) rilax Q(e). (4.1.2)
BENne
Let AT be the event "I T-1 QT(0) Q(0)I < e/2 for all 0." Then
AT 12(67')> T-1 QT(aT) e12 (4.1.3)
and
A = Q( 00) 12. (4.1.4)
But, because QT(COT) Q T (00) by the definition of aT , we have from Exp.

(4.1.3)
AT Q(e r) > T -1 Q 2'(00) e / 2. (4.1.5)
Therefore, adding both sides of the inequalities in (4.1.4) and (4.1.5), we
obtain
A = Q( 00) . (4.1.6)
Therefore, from (4.1.2) and (4.1.6) we can conclude AT Or E N, which

implies P(A T ) RO T E N). But, since lim y, P(AT) --- 1 by assumption C,
0T converges to 00 in probability.
Note that ar is defined as the value of 0 that maximizes Q 2'(0) within the
parameter space 0. This is a weakness of the theorem in so far as the extremum
estimators commonly used in practice are obtained by unconstrained maxi-
mization or minimization. This practice prevails because of its relative com-

putational ease, even though constrained maximization or minimization
would be more desirable if a researcher believed that the true value lay in a
proper subset of RK. The consistency of the unconstrained maximum Or
defined by
QT(OT) = sup QT(0) (4.1.7)
eemc
will follow from the additional assumption

(D) lim T., P[Q T(00) > supe,je Q T(0)] = 1
because of the inequality
P[Q T(00) > sup (2,(0), POT E 0). (4.1.8)
As we shall see in later applications, QT is frequently the sum of indepen-

dent random variables or, at least, of random variables with a limited degree of
dependence. Therefore we can usually expect T-1 Q T(0) to converge to a
constant in probability by using some form of a law of large numbers. How-
ever, the theorem can be made more general by replacing T-1 Q T(0) with
h(T)'Q T(0), where h(T) is an increasing function of T.
The three major assumptions of Theorem 4.1.1 are (1) the compactness of
0, (2) the continuity of QT(0), and (3) the uniform convergence of T -1 Q T(0)
to Q(0). To illustrate the importance of these assumptions, we shall give
examples that show what things go wrong when one or more of the three
assumptions are removed. In all the examples, QT is assumed to be nonsto-
chastic for simplicity.
EXAMPLE 4.1.1. 0 = [I, 1], 00 = 1/2, 7-1 Q T not continuous and not
uniformly convergent.
T-1 Q T(0) = 1 + 0, 10 eo
=e, eo <e-5 0
DT
v
0< <1
1OT
= 0, 0 =1.
Here the extremum estimator does not exist, although lim T -1 Q T attains its
unique maximum at 00.
EXAMPLE 4.1.2. 0 == [ 1, co], 00 = 1/2, 7-1 QT continuous but not uni-

formly convergent.
Q(0) = 1 + 0, 100
= -O, 0 < 0.5_ 0
= 0, elsewhere.
T- V T(0) = Q(0) + hT(0),
where
hT(0)= 8 T, T 0 -15 T + 1
=T+2-0, T+1<05_T+2
= 0, elsewhere.
Here we have plim T = plim (T+ 1) = CD, although lim 7 ' Q T = Q attains
its unique maximum at 0.
EXAMPLE 4.1.3. 0 = [0, 2], 0 0 = 1.5, T-1 QT continuous but not uniformly
convergent.
1
T-1 QT (0) = TO, 0505
2T
1 1
= 1 TO,
27 < "- T-
-
1
=0,
_ T
1 <0_00
T + 1 (0 1),
T
00 <052.
Here we have plim O. = plim (2 T)1.-- 0, although firn 7-1 QT attains its
unique maximum at 00 .
EXAMPLE 4.1.4. 0 = [-2, 1], 00 =-1, T- 1Q T not continuous but uni-

formly convergent.
T-1 QT(0) = ( 1 2 1-T )0 + 2 22- T, 20 -1

= 0 21-TA 1 < eo
2T+1 2 0 < 0:5 1 2-(r+t)
2T+1 1 '
1 2-() 0< 1
= 0, 0 =-- 1.
Here we have plim aT = plim [1 2-( T+ 1)] = 1, although firn T-1 Q T attains
its unique maximum at 00 . (If we change this example so that 0 = [- 2, 1),
T-1 QT becomes continuous and only the compactness assumption is vio-
lated.)
The estimator oT of Theorem 4.1.1 maximizes the function Q T (0) globally.
However, in practice it is often difficult to locate a global maximum of Q(0),
for it means that we must look through the whole parameter space except in
the fortunate situation where we can prove that Q T(0) is globally concave.
Another weakness of the theorem is that it is often difficult to prove that
Q(0) attains its unique global maximum at 00. Therefore we would also like to
have a theorem regarding the consistency of a local maximum.
Still another reason for having such a theorem is that we can generally prove
asymptotic normality only for the local maximum, as we shall show in Section
4.1.2. Theorem 4.1.2 is such a theorem.
THEOREM 4.1.2. Make the assumptions:
(A) Let 0 be an open subset of the Euclidean K-space. (Thus the true value
00 is an interior point of 0.)
(B) Q T(y, 0) is a measurable function of y for all 0 E 0, and aQriao exists
and is continuous in an open neighborhood N1(00) of O. (Note that this
implies QT is continuous for 0 E
(C) There exists an open neighborhood N2(00) of 00 such that T-1 Q T(0)
converges to a nonstochastic function Q(0) in probability uniformly in 0 in
N2(00), and Q(0) attains a strict local maximum at 00 .
Let OT be the set of roots of the equation
OQT _ 0 (4.1.9)
ae
zorresponding to the local maxima. If that set is empty, set OT equal to (0).
Then, for any e > 0,

lirn P[ inf (0 00)'(0 00)> E] = 0.
T--**, 0E0r
Proof Choose a compact set S c NI n N2. Then the value of 0, say 0, that
globally maximizes QT(0) in S is consistent by Theorem 4.1.1. But because the
probability that 7-1 Q T (0) attains a local maximum at Ot. approaches 1 as T
goes to 00, limr_. P(011. E OT) = 1.
We sometimes state the conclusion of Theorem 4.1.2 simply as "there is a
consistent root of the Eq. (4.1.9)."
The usefulness of Theorem 4.1.2 is limited by the fact that it merely states
that one of the local maxima is consistent and does not give any guide as to
how to choose a consistent maximum. There are two ways we can gain some
degree of confidence that a local maximum is a consistent root: (1) if the
solution gives a reasonable value from an economic-theoretic viewpoint and
(2) if the iteration by which the local maximum was obtained started from a
consistent estimator. We shall discuss the second point more fully in Section
4.4.2.
4.1.2 Asymptotic Normality

In this subsection we shall show that under certain conditions a consistent root
of Eq. (4.1.9) is asymptotically normal. The precise meaning of this statement
will be made clear in Theorem 4.1.3.
THEOREM 4.1.3. Make the following assumptions in addition to the as-

sumptions of Theorem 4.1.2:
(A) 82Q T /8080' exists and is continuous in an open, convex neighborhood
of 00.
(B) T-1 (82Q T/8080')r, converges to a finite nonsingular matrix A(0 0) =
lim ET-, 02QT iaeafneo in probability for any sequence on such that
plim In' = 0c, -
(C) 7 v2042430k 0 N[0, WU, where B(00) = lim ET-1 (aQT1a0)00 x
(8.2T1p0le0 .
Let (0T ) be a sequence obtained by choosing one element from O T defined in

Theorem 4.1.2 such that plim OT = 00 . (We call O T a consistent root.) Then
VT(67 00) > N[0, A(00) -1 B(00)A(00)- 1i.
-.
Proof By a Taylor expansion we have

8QT1 = aQTI 82,27.1 (o o ) (4.1.10)
ae j , ae leo (wei r T
where 0* lies between and 00 Noting that the left-hand side of (4.1.10) is 0
.2
by the definition of 0T, we obtain
T 00) = [ 7,
1 a 2 ji
11 71
1.- , a
0102 710: (4.1.11)
where + denotes the Moore-Penrose generalized inverse (see note 6 of Chapter

2). Because plim 0* = 00 , assumption B implies
plim 1 (7227- A(00). (4.1.12)

T aom

Finally, the conclusion of the theorem follows from assumption C and Eqs.
(4.1.11) and (4.1.12) by repeated applications of Theorem 3.2.7.
As we noted earlier, QT is frequently the sum of independent random
variables (or, at least, of random variables with a limited degree of depen-
dence). Therefore it is not unreasonable to assume the conclusions of a law of
large numbers and a central limit theorem in assumptions B and C, respec-
tively. However, as we also noted earlier, the following more general normali-
zation may be necessary in certain cases: In assumption B, change
T-182QT18080' to il(T)a2 QT1a0a0'H(T), where 14(T) is a diagonal matrix
such that limr, H(T) =0; in assumption C, change T-1 /20Q T/80 to
H(T)0Q T/a0; and in the conclusion of the theorem, state the limit distribu-
tion in terms of H( T)i(O T 00).
Because assumption B is often not easily verifiable, we shall give two alter-
native assumptions, each of which implies assumption B. Let gT(0) g(y, 0)
be a function of a T-vector of random variables y = (yi , Y2,. . . , yr)' and a
continuous function of a K-vector of parameters G in 13, an open subset of the
Euclidean K-space, almost surely. We assume that g(y, 0) is a random variable
(that is, g..is a measurable function of y). We seek conditions that ensure
plim [gT (OT ) g(0)] = 0 whenever plim OT = 00. Note that Theorem 3.2.6
does not apply here because in that theorem g(. ) is a fixed function not
varying with T.
THEOREM 4.1.4. Assume that agr /o0 exists for 0 E 0, an open convex set,
and that for any e > 0 there exists M. such that
P(logT/801 < Mc) 1e
for all T, for allp E 0, and for all i, where 01 is the ith element of the vector 0.
Then plim gr(0T) = plim 87400 if 00 plim Or is in 0.
Proof The proof of Theorem 4.1.4 follows from the Taylor expansion
grOar) = MOO + (T - 0o)

V 6.
where 0* lies between Or and 00.
THEOREM 4.1.5. Suppose gT(0) converges in probability to a nonstochastic

function ,(0) uniformly in 0 in an open neighborhood N(00) of 00 . Then
plim gT (OT) = g(00 if plim or = 00 and g(0) is continuous at 00
) .
Proof Because the convergence of gr(0) is uniform in 0 and because OT E

N(00) for sufficiently large T with a probability as close to one as desired, we
can show that for any c > 0 and 6> 0, there exists T 1 such that for T> T1
<..
P[18T(OT) 8(e 7-)1-2S]< 2
Because g is continuous at 00 by our assumption, g(or ) converges to g(00) in
probability by Theorem 3.2.6. Therefore, for any e > 0 and 6> 0, there exists
T2 such that for T> T2
P[18(O r) 8(0 0)1 . <. (4.1.14)
Therefore, from the inequalities (4.1.13) and (4.1.14) we have for T>
max [T 1 , T 2]
P[IgT( 62.)
- g(00)I e] g 1 6. (4.1.15)
The inequality (4.1.13) requires that uniform convergence be defined by

either (i) or (ii) of Definition 4.1.1. Definition 4.1.1(iii) is not sufficient, as
shown in the following example attributed to A. Ronald Gallant. Let cbr(e) be
a continuous function with support on [ T-- T '] and Or(0) = 1. Define
gT(co, 0) = cbT(co TO) if 0 g co, 0 g 1, and gT(co, 0) = 0 otherwise. Assume
that the probability measure over 0 g co g 1 is Lebesgue measure and
Or = colT. Then
2
inf P[gT(co, 0) < e] g 1 1,
0.8,1
meaning that gr(co, 6) converges to 0 semiuniformly in 0. But gT(co, eT) = 1

for all T. Note that in this example

P[ sup gr(co, 0) < 1= 0.
()seal
Hence, gT(co, 0) does not converge uniformly in 0 in the sense of definition
4. 1. 1(u).
The following theorem is useful because it gives the combined conditions
for the consistency and asymptotic normality of a local extremum estimator.
In this way we can do away with the condition that Q(0) attains a strict local
maximum at 00.
THEOREM 4.1.6. In addition to assumptions A -C of Theorem 4.1.3, as-
sume
(A) T- V T (0) converges to a nonstochastic function Q(0) in probability
uniformly in 0 in an open neighborhood of 0 0 .
(B) A(00) as defined in assumption B of Theorem 4. 1.3 is a negative definite
matrix.
(C) plim 7-102QT/0080' exists and is continuous in a neighborhood of 00.
Then the conclusions of Theorems 4.1.2 and 4.1.3 follow.
Proof By a Taylor expansion we have in an open neighborhood of 00
1 1 1 a
QT(0) =. QT(00) + (0- 00) (4.1.16)
T T T Oa' I so
QT
1
ooy 1 Q
O2
--T1
where 0* lies between 0 and 00 . Taking the probability limit of both sides of
(4.1.16) and using assumptions B and C of Theorem 4.1.3 and A of this
theorem, we have
2(6 ) = 12(90) + 3(0 - 00) I A*(0 - 00),
1 (4.11 7)
where
_.1
A* - plim 7 82QT , itiw ie s.
But A* is a negative definite matrix because of assumption B of Theorem 4.1.3
and assumptions B and C of this theorem. Therefore
Q(0) < Q(00) for 0 # N. (4.1.18)
Thus all the assumptions of Theorem 4.1.2 are satisfied.
4.2 Maximum Likelihood Estimator

4.2.1 Definition
Let L T (0) = L(y, 0) be the joint density of a T-vector of random variables
y = (y,, Y2, , YT)' characterized by a K-vector of parameters 0. When we
regard it as a function of 0, we call it the likelihood function. The term
maximum likelihood estimator (MLE) is often used to mean two different
concepts: (1) the value of 0 that globally maximizes the likelihood function
L(y, 0) over the parameter space 0; or (2) any root of the likelihood equation
8L(0) _ 0
(4.2.1)
ae
that corresponds to a local maximum. We use it only in the second sense and
use the term global maximum likelihood estimator to refer to the first concept.
We sometimes use the term local maximum likelihood estimator to refer to
the second concept.
4.2.2 Consistency
The conditions for the consistency of the global MLE or the local MLE can be
immediately obtained from Theorem 4.1.1 or Theorem 4.1.2 by putting
QT(0) = log L T(0). We consider the logarithm of the likelihood function
because T-1 log L T(0) usually converges to a finite constant. Clearly, taking
the logarithm does not change the location of either global or local maxima.
So far we have not made any assumption about the distribution of y. If we
assume that (y,) are i.i.d. with common density functionf( , 0), we can write
log L(y, 0) = log f(y 0). (4.2.2)

i i
-
In this case we can replace assumption C of either Theorem 4.1.1 or Theorem

4.1.2 by the following two assumptions:
E sup llogf(y 0)1< M, for some positive constant M, (4.2.3)
8E0
and
log f(y 0) is a continuous function of 0 for each y (4.2.4)
In Theorem 4.2.1 we shall show that assumptions (4.2.3) and (4.2.4) imply
plim ,r1 i log f(y 0) = E log f(y 0) uniformly in 0 E 0.

(4.2.5)
Furthermore, we have by Jensen's inequality (see Rao, 1973, P. 58, for the
proof)
E log < log E flY I, 0) - o for 0 ' 00 , (4.2.6)
where the expectation is taken using the true value 00, and, therefore
E log f(y 0) <E log f(y 00) for 0 # 00 . (4.2.7)
As in (4.2.7), we have T- IE log L(0) < E log L(O0) for 0 # 00 and for
all T. However, when we take the limit of both sides of the inequality (4.2.7) as
T goes to infinity, we have
lim T-1E log L T (0) lim T-1 E log LT(80).
Hence, one generally needs to assume that lim T, E log L T (0) is

uniquely maximized at 0 = 00 .
That (4.2.3) and (4.2.4) imply (4.2.5) follows from the following theorem
when we put g,(0)= log f(y 0) E log f(y 0).
THEOREM 4.2.1. Let g(y, 0) be a measurable function of y in Euclidean

space for each 0 E 0, a compact subset of RK (Euclidean K-space), and a
continuous function of 0 E 0 for each y. Assume E g(y, =0. Let (y,) be a
sequence of i.i.d. random vectors such that E sup." ig(y < 00. Then
,g( y 0) converges to 0 in probability uniformly in 0 E 0.
Proof.3 Partition 0 into n nonoverlapping regions 07, 03, . . , 07, in

such a way that the distance between any two points within each 07 goes to 0 as
n goes to 00. Let 01, 2, . , 0,, be an arbitrary sequence of K-vectors such
oo
that 0. E 07, 1 = 1, 2, . . . , n. Then writing g for g(y 0), we have for any
e>0
P[sup1T-1 g,(0)I> e] (4.2.8)

BEe t
P[o isup7 do l > El]

0E 0 t-i
P [sup 1T-1 gi( 0)1> e]

eEe7 p-i
E]
P T-1
r-i t-i L I 19E07
where the first inequality follows from the fact that if A implies B then
P(A) 5 P(B) and the last inequality follows from the triangle inequality. Be-
cause gi(0) is uniformly continuous in 0 E 0, we have for every i
lim sup IVO) &AM= 0. (4.2.9)
eee;
But, because
sup 1,00) 5. 2 sup Ig,(0)1 (4.2.10)
8E07 BEO
and the right-hand side of the inequality (4.2.10) is integrable by our assump-
tions, (4.2.9) implies by the Lebesgue convergence theorem (Royden, 1968,
p. 88)
lim E sup Ig,(0) gr(0)1= 0 (4.2.11)
8E07
uniformly for i. Take n so large that the expected value in (4.2.11) is smaller
than e/2. Finally, the conclusion of the theorem follows from Theorem 3.3.2
(Kolmogorov LLN 2) by taking T to infinity.
This theorem can be generalized to the extent that 7 IM,T_ igt(k) and
,supoew Igi(0) g,(0i)I can be subjected to a law of large numbers. The

following theorem, which is a special case of a theorem attributed to Hoadley

(1971) (see also White, 1980b), can be proved by making a slight modification
of the proof of Theorem 4.2.1 and using Markov's law of large numbers
(Chapter 3, note 10).
THEOREM 4.2.2. Let gi(y, 0) be a measurable function of y in Euclidean
space for each t and for each 0 E 0, a compact subset of RK (Euclidean
K-space), and a continuous function of 0 for each y uniformly in t. Assume
E gr(y, 0) = 0. Let (y,) be a sequence of independent and not necessarily
identically distributed random vectors such that E sup eee lgXy, 0)1 1+6
M<
00 for some J> 0. Then 7 'ET_ st(y 0) converges to 0 in probability
uniformly in 0 E 0.
We will need a similar theorem (Jennrich, 1969, p. 636) for the case where Yr
is a vector of constants rather than of random variables.
THEOREM 4.2.3. Let y 1 , y2 , . . , YT be vectors of constants. We define the

empirical distribution function of (Y19 Y29 9 YT) by FT(a) = T-l Iti:1
x(y t < a), where x takes the value 1 or 0 depending on whether the event in its
argument occurs or not. Note that y, < a means every element of the vector y,
is smaller than the corresponding element of a. Assume that g(y, 0) is a
bounded and continuous function of y in Euclidean space and e in a compact
set 0. Also assume that FT converges to a distribution function F. Then
limT, 7-- lg.. ig(y 0) == fg(y, 0) dF(y) uniformly in O.
There are many results in the literature concerning the consistency of the
maximum likelihood estimator in the i.i.d. case. Rao (1973, p. 364) has
presented the consistency of the local MLE, which was originally proved by
Cramer (1946). Wald (1949) proved the consistency of the global MLE with-
out assuming compactness of the parameter space, but his conditions are
difficult to verify in practice. Many other references concerning the asymp-
totic properties of the MLE can be found in survey articles by Norden (1972,
1973).
As an example of the application of Theorem 4.1.1, we shall prove the
consistency of the maximum likelihood estimators of ft and a2 in Model 1.
Because in this case the maximum likelihood estimators can be written as
explicit functions of the sample, it is easier to prove consistency by a direct
method, as we have already done in Section 3.5. We are considering this more
complicated proof strictly for the purpose of illustration.
EXAMPLE 4.2.1. Prove the consistency of the maximum likelihood estima-
tors of the parameters of Model 1 with normality using Theorem 4.1.1, assum-
ing that lirn T- 'X'X is a finite positive definite matrix.
In Section 1.1 we used the symbols fi and a2 to denote the true values
because we did not need to distinguish them from the domain of the likelihood
function. But now we shall put the subscript 0 to denote the true value;
therefore we can write Eq. (1.1.4) as
Y "--- Xfio + u, (4.2.12)
where Vu r = ai. From (1.3.1) we have
T T 1
(y Xity(y X/l) (4.2.13)
log LT= - - log 27r log a2
2 2 2a2
T T
= -- log 27r log a2
2 2
1

2a2 [X(A fl) + unX(fi0 ii) + a],
where the second equality is obtained by using (4.2.12). Therefore

. 1 1 1
plim log LT= log 2tt log a2 (4.2.14)
T 2 2
.1
_ XiX 1 ao2
fly lim (Po PI 0 2 -
2a 2 T 2
Define a compact parameter space 0 by
ci a2 c2 , c3 , (4.2.15)
where ci is a small positive constant and c2 and c3 are large positive constants,
and assume that (1%, al) is an interior point of 0. Then, clearly, the conver-
gence in (4.2.14) is uniform in 0 and the right-hand side of (4.2.14) is uniquely
maximized at (fl; , as). Put 0 = (fl', a2)' and define OT by
log L(0) = max log L T (0). (4.2.16)
see
Then or is clearly consistent by Theorem 4.1.1. Now define gr. by

log L T( oT) = max log L T (0), (4.2.17)
where the maximization in (4.2.17) is over the whole Euclidean (K + 1)-space.

Then the consistency of OT, which we set out to prove, would follow from
inn P(O"' OT) = 1 - (4.2.18)
The proof of (4:2.18) would be simple if we used our knowledge of the explicit
formulae for OT in this example. But that would be cheating. The proof of
(4.2.18) using condition D given after the proof of Theorem 4.1.1 is left as an
exercise.
There are cases where the global maximum likelihood estimator is incon-
sistent, whereas a root of the likelihood equation (4.2.1) can be consistent, as
in the following example.
EXAMPLE 4.2.2. Let y t = 1, 2, . . . , T, be independent with the common

distribution defined by
f(y,) = Nu,, cri) with probability A (4.2.19)
= N(A2, al) with probability 1

This distribution is called a mixture of normal distributions. The likelihood
function is given by
L = ITT [ r--A exp E ()it ii i)2/(2a1)] (4.2.20)

t-i V 27rai
1 A
+ , e P x i- - - (1', 112) 21(2 a1)d
NI 2.7ra2
If we put A l = yl and let cr1 approach 0, the term of the product that corre-
sponds to t = 1 goes to infinity, and, consequently, L goes to infinity. Hence,
the global MLE canot be consistent. Note that this example violates assump-
tion C of Theorem 4. Li because Q(0) does not attain a global maximum at O.
However, the conditions of Theorem 4.1.2 are generally satisfied by this
model. An extension of this model to the regression case is called the switching
regression model (see Quandt and Ramsey, 1978).
It is hard to construct examples in which the maximum likelihood estima-
tor (assuming the likelihood function is correctly specified) is not consistent
and another estimator is. Neyman and Scott (1948) have presented an inter-
esting example of this type. In their example MLE is not consistent because
the number of incidental (or nuisance) parameters goes to infinity as the
sample size goes to infinity.

The asymptotic normality of the maximum likelihood estimator or, more
precisely, a consistent root of the likelihood equation (4.2.1), can be analyzed
by putting QT = log LT in Theorem 4.1.3. If (Yt) are independent, we can
write
T
log LT= I log fr(y 0), (4.2.21)
P 1
where f, is the marginal density of y1 . Thus, under general conditions onf,,, we

can apply a law of large numbers to (3 2 log L T/O0O0' and a central limit
theorem to a log L Tiao. Even if (ye) are not independent, a law of large
numbers and a central limit theorem may still he applicable as long as the
degree of dependence is limited in a certain manner, as we shall show in later
chapters. Thus we see that assumptions B and C of Theorem 4.1.3 are ex-
pected to hold generally in the case of the maximum likelihood estimator.
Moreover, when we use the characteristics of L as a joint density function,
we can get more specific results than Theorem 4.1.3, namely as we have shown
in Section 1.3.2, the regularity conditions on the likelihood function given in

assumptions A' and B' of Section 1.3.2 imply
A(00) = B(00). (4.2.22)
Therefore, we shall make (4.2.22) an additional assumption and state it for-
mally as a theorem.
THEOREM 4.2.4. Under the assumptions of Theorem 4.1.3 and assumption
(4.2.22), the maximum likelihood estimator T satisfies
1 02 log L T
NIT 5T N 10, [lim E (4.2.23)
T 808fY
If(y1) are i.i.d. with the common density function f( , 8), we can replace
assumptions B and C of Theorem 4.1.3 as well as the additional assumption
(4.2.22) with the following conditions on f( , 0) itself:
f t(13of dy 0, (4.2.24)
a2f d o (4.2.25)
J WiFr Y
1 T 02 logf a2 log f
plim E uniformly in 0 in an open
T am maw
neighborhood of 00. (4.2.26)
A sufficient set of conditions for (4.2.26) can be found by putting g1(0) =
82 log fdaoiae; in Theorem 4.2.1. Because log L T = MIL, log f(y 0) in this
case, (4.2.26) implies assumption B of Theorem 4.1.3 because of Theorem
4.1.5. Assumption C of Theorem 4.1.3 follows from (4.2.24) and (4.2.26) on
account of Theorem 3.3.4 (Lindeberg-Levy CLT) since (4.2.24) implies
ga log pomp. =0. Finally, it is easy to show that assumptions (4.2.24)
(4.2.26) imply (4.2.22).
We shall use the same model as that used in Example 4.2.1 and shall
illustrate how the assumptions of Theorem 4.1.3 and the additional assump-
tion (4.2.22) are satisfied. As for Example 4.2.1, the sole purpose of Example
4.2.3 is as an illustration, as the same results have already been obtained by a
direct method in Chapter 1.
EXAMPLE 4.2.3. Under the same assumptions made in Example 4.2.1,
prove asymptotic normality of the maximum likelihood estimator
0 = (0% 6.-2),.
We first obtain the first and second derivatives of log L:

a log L1
2- (X'Xfl X'y), (4.2.27)
ofl a
0 log L_ T 1,
5.4
E- XflY(Y (4.2.28)
ao-2 Tri
-2 -
82 log L 1 vv
(4.2.29)
vox a2
a2 log L _ T 1
a(a2)2 20 , Tr XflY(Y Xfl), (4.2.30)
02 log L 1
aa2afl (7 4 ("13 XY. (4.2.31)
From (4.2.29), (4.2.30), and (4.2.31) we can clearly see that assumptions A
and B of Theorem 4.1.3 are satisfied. Also from these equations we can
evaluate the elements of A(00):
1
02log L 1 X'X
plim (4.2.32)
T aflajr crg T

. 1 02 log L 1
plim (4.2.33)
T a(a2)2 2cr8'
1 82 log L
plim = 0. (4.2.34)
T acr2afileo
From (4.2.27) and (4.2.28) we obtain
1 alogLI _ 1 X'u
and 1
a log LI _ 1 u'u
op 00 al NIT 47; 430'2 leo 24
Thus, by applying either the Lindeberg-Feller or Liapounov CLT to a se-
quence of an arbitrary linear combination of a (K + 1)-v ect or
(xitu x2,141 , . . . , xKtu u? as), we can show
1 a log L N (0, urn (4.2.35)
rt op I A, ao T
and
1 alogL
N (0, -a (4.2.36),
VT at72
80
Figure 4.1 The log likelihood function in a nonregular case
with zero asymptotic covariance between (4.2.35) and (4.2.36). Thus assump-
tion C of Theorem 4.1.3 has been shown to hold. Finally, results (4.2.32)
through (4.2.36) show that assumption (4.2.22) is satisfied. We write the
conclusion (4.2.23) specifically for the present example as
,ff, [ ./.i /Jo ] _) N [ag (lim T- IX'X)- ' 204 1
(4.2.37)
{o
0.2 ag 0
There are cases where the global maximum likelihood estimator exists but
does not satisfy the likelihood equation (4.2.1). Then Theorem 4.1.3 cannot
be used to prove the asymptotic normality of MLE. The model of Aigner,
Amemiya, and Poirier (1976) is such an example. In their model, plim 7 '
log LT exists and is maximized at the true parameter value 00 so that MLE is
consistent. However, problems arise because plim T- I log L T is not smooth at
00 ; it looks like Figure 4.1. In such a case, it is generally difficult to prove
asymptotic normality.
4.2.4 Asymptotic Efficiency

The asymptotic normality (4.2.23) means that if Tis large the variance-covar-
iance matrix of a maximum likelihood estimator may be approximated by
82 log L T I 1-1
(4.2.38)
_ [E me' lki '
But (4.2.38) is precisely the Cramer-Rao lower bound of an unbiased estima-
tor derived in Section 1.3.2. At one time statisticians believed that a consistent
and asymptotically normal estimator with the asymptotic covariance matrix
(4.2.38) was asymptotically minimum variance among all consistent and

asymptotically normal estimators. But this was proved wrong by the following
counterexample, attributed to Hodges and reported in LeCam (1953).
EXAMPLE 4.2.4. Let U. be an estimator of a scalar parameter such that
plim 07. = 0 and VT(0 7- 0) ---* N[0, v(0)]. Define the estimator 0141- = wTOT,
where
WT. =0 if 074 < 7-1/4
71 1-1/4.=1if0
It can be shown (the proof is left as an exercise) that 47(01- 1 0) 0
N[0, v*(13)], where v*(0) = 0 and v*(0) = v(0) if 0 # O.
The estimator 01-1 of Example 4.2.4 is said to be superefficient. Despite the
existence of superefficient estimators, we can still say something good about
an estimator with the asymptotic variance-covariance matrix (4.2.38). We
shall state two such results without proof. One is the result of LeCam (1953),
which states that the set of 0 points on which a superefficient estimator has an
asymptotic variance smaller than the Cramer-Rao lower bound is of Lebesgue
measure zero. The other is the result of Rao (1973, p. 350) that the matrix
(4.2.38) is the lower bound for the asymptotic variance-covariance matrix of
all the consistent and asymptotically normal estimators for which the conver-
gence to a normal distribution is uniform over compact intervals of 0. These
results seem to justify our use of the term asymptotically efficient in the
following sense:
DEFINITION 4.2.1. A consistent estimator is said to be asymptotically effi-
cient if it satisfies statement (4.2.23).
Thus the maximum likelihood estimator under the appropriate assump-
tions is asymptotically efficient by definition. An asymptotically efficient
estimator is also referred to as best asymptotically normal (BAN for short).
There are BAN estimators other than MLE. Many examples of these will be
discussed in subsequent chapters, for example, the weighted least squares
estimator will be discussed Section 6.5.3, the two-stage and three-stage least
squares estimators in Sections 7.3.3 and 7.4, and the minimum chi-square
estimators in Section 9.2.5. Baranlcin and Gurland (1951) have presented a
general method of generating BAN estimators. Because their results are math-
ematically too abstract to present here, we shall state only a simple corollary of
their results: Let (y,) be an i.i.d. sequence of random vectors with E y, =
E(y, /1)( yr = M(0), and with exponential family density
.AY, = exP Ecto(0) + 13o(Y) +
and define zr = T- Then the minimization of [z T it(0)]'

loyi[z r - p(0)] yields a BAN estimator of 0 (see also Taylor, 1953; Fergu-
son, 1958).4
Different BAN estimators may have different finite sample distributions.
Recently many interesting articles in the statistical literature have compared
the approximate distribution (with a higher degree of accuracy than the
asymptotic distribution) of MLE with those of other BAN estimators. For
example, Ghosh and Subramanyam (1974) have shown that in the exponen-
tial family the mean squared error up to 0(T-2) of MLE after correcting its
bias up to 0(T- ') is smaller than that of any other BAN estimator with
similarly corrected bias. This result is referred to as the second-order efficiency
of MLE,5 and examples of it will be given in Sections 7.3.5 and 9.2.6.
4.2.5 Concentrated Likelihood Function

We often encounter in practice the situation where the parameter vector 00
can be naturally partitioned into two subvectors ao and tio as 00 = (c4 /3)'.
The regression model is one such example; in this model the parameters
consist of the regression coefficients and the error variance. First, partition the
maximum likelihood estimator as 0= (ci', fl')'. Then, the limit distribution of
ao) can easily be derived from statement (4.2.23). Partition the in-
verse of the asymptotic variance-covariance matrix of statement (4.2.23) con-
formably with the partition of the parameter vector as
E1 (32 log LTI [AB
(4.2.39)
T maw B'
Then, by Theorem 13 of Appendix 1, we have
ao) > N[0, (A BC- 11'r lb (4.2.40)
Let the likelihood function be L(a, fl). Sometimes it is easier to maximize L
in two steps (first, maximize it with respect ton, insert the maximizing value of
ft back into L; second, maximize L with respect to a) than to maximize L
simultaneously for a and fl. More precisely, define
L*(a) = L[a, 11(a)], (4.2.41)
where ii(a) is defined as a root of
a log LI 0
(4.2.42)
afl A '
and define ol as a root of
a log L*
aa a = 0. (4.2.43)
We call L*(a) the concentrated likelihood function of a. In this subsection we

shall pose and answer affirmatively the question: If we treat L*(a) as if it were a
proper likelihood function and obtain the limit distribution by (4.2.23), do we
get the same result as (4.2.40)?
From Theorem 4.1.3 and its proof we have
ao) I2 [plim
82mace
log L*1 1 8 log L*
4_i; aa L., (4.2.44)
where k-9 means that both sides of the equation have the same limit distribu-
tion. Differentiating both sides of (4.2.41) and evaluating the derivative at ao
yields
0 log L*1 a log L +
4'1 a log L (4.2.45)
Oa ao aa laoAao) aa aoi 9filao.;* ao)
log L
Oa lefo.Aao;
where the second equality follows from (4.2.42). By a Taylor expansion we
have
a log LI _ 8 log LI + a2 log LI [ ix%)
(4.2.46)
a 1.4ft) act 6lo
aa6iji
where 0+ lies between [c<1, %A. aon and 00 . But we have
[limE 32 L
-1 1 a log L (4.2.47)
.
'MA%) Po] La
3 - 1
T OP" Li 7T,
Therefore, from (4.2.45), (4.2.46), and (4.2.47), we obtain
a log Ll = (1 BC-1). a log L
(4.2.48)
ff Oa ao
Finally, using (4.2.22), we obtain from (L.48) lek.
1 a log L*L N(0, A BC-1 13'). (4.2.49)
aa
Next, differentiating both sides of the identity

0 log L*(a) _ a log L[a, ii(a)]
(4.2.50)
aa aa
with respect to a yields
82 log Ll _ 82 log LI 8fi'l 82 log L
(4.2.51)
aaaa' .0 &ea ' ..A.0) Oa .0 8/18a' 1.0,0140.
Differentiating both sides of (4.2.42) with respect to a yields
82 log LI 4.
l 82 log LI =0. (4.2.52)
acial3' ...Avo aa .0 Vali' ...ixazi)
Combining (4.2.51) and (4.2.52) yields
a2 log L*I _ 82 log L
(4.2.53)
Oaaa' la, aaaa Law
82 log LI 82 log L -1 a2 log L
aaafii lao.ka0)[ Vali' Li alioa' Lical
which implies
1 82 log L*1
plim (A BC-113'). (4.2.54)
T Oaaa' ao
Finally, we have proved that (4.2.44), (4.2.49), and (4.2.54) lead precisely to
the conclusion (4.2.40) as desired.
4.3 Nonlinear Least Squares Estimator

4.3.1 Definition
We shall first present the nonlinear regression model, which is a nonlinear
generalization of Model 1 of Chapter 1. The assumptions we shall make are
also similar to those of Model 1. As in Chapter 1 we first shall state only the
fundamental assumptions and later shall add a few more assumptions as
needed for obtaining particular results.
We assume
Yr = ft( 110) U t/ t = 1 1 25 f 71 (4.3.1)
where y, is a scalar observable random variable, /3 0 is a K-vector of unknown
parameters, and (u,) are i.i.d. unobservable random variables such that
Eli, 0 and Vu, = al (another unknown parameter) for all t.
The assumptions on the function f, will be specified later. Often in practice

we can write f,(flo) f(x, 110), where x, is a vector of exogenous variables
(known constants), which, unlike the linear regression model, may not neces-
sarily be of the same dimension as fl o
As in Chapter 1, we sometimes write (4.3.1) in vector form as
f( 10) + u, (4.3.2)
where y, f, and u are all T-vectors, for which the tth element is defined in
(4.3.1).
Nonlinearity arises in many diverse ways in econometric applications. For
example, it arises when the observed variables in a linear regression model are
transformed to take account of serial correlation of the error terms (cf. Section
6.3). Another example is the distributed-lag model (see Section 5.6), in which
the coefficients on the lagged exogenous variables are specified to decrease
with lags in a certain nonlinear fashion. In both of these examples, nonlinear-
ity exists only in parameters and not in variables.
More general nonlinear models, in which nonlinearity is present both in
parameters and variables, are used in the estimation of production functions
and demand functions. The Cobb-Douglas production function with an addi-
tive error term is given by
Qr = 131 1(92b13 + u (4.3.3)
where Q, K, and L denote output, capital input, and labor input, respectively. 6
TheCESproductinf (seArowtal.,196)myberinas
Q1= ATAKTA + (1 fl2)L -,-"31-043 + ut. (4.3.4)
See Mizon (1977) for several other nonlinear production functions. In the
estimation of demand functions, a number of highly nonlinear functions have
been proposed (some of these are also used for supply functions), for example,
translog (Christensen, Jorgenson, and Lau, 1975), generalized Leontief (Die-
wen, 1974), S-branch (Brown and Heien, 1972), and quadratic (Howe, Pol-
lak, and Wales, 1979).
As in the case of the maximum likelihood estimator, we can define the
nonlinear least squares estimator (abbreviated as NLLS) of fl o in two ways,
depending on whether we consider the global minimum or a local minimum.
In the global case we define it as the value of // that minimizes
sr = f,(13)1 2 (4.3.5)
over some parameter space B. In the local case we define it as a root of the
normal equation
8ST 0
(4.3.6)
op
We shall consider only the latter case because (4.3.6) is needed to prove
asymptotic normality, as we have seen in Section 4.1.2. Given the NLLS
estimator ft of fio , we define the NLLS estimator of al, denoted as 6 2, by
6 2 = TS-(ft). (4.3.7)
Note that and 62 defined above are also the maximum likelihood estima-
tors if the (tt,) are normally distributed.
4.3.2 Consistency'
We shall make additional assumptions in the nonlinear regression model so
that the assumptions of Theorem 4.1.2 are satisfied.
THEOREM 4.3.1. In the nonlinear regression model (4.3.1), make the addi-
tional assumptions: There exists an open neighborhood N of flo such that
(A) Of,/Oft exists and is continuous on N.
(B) fr(fl) is continuous in 5 E N uniformly in t; that is, given > 0 there
exists 8> 0 such that IAA) f,(132)j< e whenever (51 AAA 52) <8 for
all fi1,fi2 E N and for all t.8
(C) Er.. 1f,(fi1)f,(fi2) converges uniformly in , $2 E N.
(D) lim T- iIi 1 [ff(ti0)f,(fl)]2 # 0 if /I # So .
Then a root of (4.3.6) is consistent in the sense of Theorem 4.1.2.
Proof Inserting (4.3.1) into (4.3.5), we can rewrite 7 I times (4.3.5) as
T
ST 1 u
T
+ T1
f_ Efheo) .10)1 2
(4.3.8)
2
+
T
[mu -fruplut
v. A, +A2 +A3 .
The term A 1 converges to al in probability by Theorem 3.3.2 (Kolmogorov
LLN 2). The term A2 converges to a function that has a local minimum at $0
uniformly in fi because of assumptions C and D. We shall show that A3
converges to 0 in probability uniformly in fl E Nby an argument similar to the
proof of Theorem 4.2.1. First, plim T_. I MIT-1.fAi30)u1 = 0 because of as-

sumption C and Theorem 3.2.1. Next, consider suppEN 7-11Er_ ft(f)uel. Par-
tition N into n nonoverlapping regions N1, N2,. . . , N. Because of as-
sumption B, for any > 0 we can find a sufficiently large n such that for each
i = 1, 2, . . . , n
Ift(fl1) - fAii2)1 1 E 1)1/2

2(0+ for /32 E Ni and for all t.
(4.3.9)
Therefore, using the Cauchy-Schwartz inequality, we have

T T )1
/ 2
sup 1 I fi(P)u,1 = 14 2(4 1)1/2 ,
PEN' T
(4.3.10)
where fl is an arbitrary fixed point in Ni. Therefore
P [r:11%)' f > g P 14-- fimul > e]

P gi Afli)ut l >
+ j- u? > a2 + 1]. (4.3.11)
nP[ T
Finally, we obtain the desired result by taking the limit of both sides of the
inequality (4.3.1 1) as T goes to co. Thus we have shown that assumption C of
Theorem 4.1.2 is satisfied. Assumptions A and B of Theorem 4.1.2 are clearly
satisfied.
Assumption C of Theorem 4.3.1 is not easily verifiable in practice; therefore

it is desirable to find a sufficient set of conditions that imply assumption C and
are more easily verifiable. One such set is provided by Theorem 4.2.3. To
apply the theorem to the present problem, we should assume Aft) = f(x fl)
and take x, andf,(A)f,(h) as the y and g(y 0) of Theorem 4.2.3, respectively.
Alternatively, one could assume that (x i) are Lid. random variables and use
Theorem 4.2.1.
In the next example the conditions of Theorem 4.3.1 will be verified for a
simple nonlinear regression model.
EXAMPLE 4.3.1. Consider the nonlinear regression model (4.3.1) with

WO= log (flu + x t), where /30 and x, are scalars. Assume (i) the parameter
space B is a bounded open interval (c, d), (ii) x, + > > 0 for every t and for
every E B, and (iii) (x,) are i.i.d. random variables such that
Efilog (d+ x1)]2) < co. Prove that a root of (4.3.6) is consistent.
First, note that log (flu + x,) is well defined because of assumptions (i) and
(ii). Let us verify the conditions of Theorem 4.3.1. Condition A is clearly
satisfied because of assumptions (i) and (ii). Condition B is satisfied because
'log + x,) log (fl2 + + fi21
by the mean value theorem, where'll' (depending on x,) is between fl i and /12 ,
and because 1/37 + 4 -1 is uniformly bounded on account of assumptions (i)
and (ii). Condition C follows from assumption (iii) because of Theorem 4.2.1.
To verify condition D, use the mean value theorem to obtain
1 T
7, A [log (fl + x,) log (flo + ;OP
7,[i
=1 057 Xtrd( 3 110)2
t I
where fit (depending on x,) is between fi and fig . But

T 1
t-
and
prlim (d + x,) 2 = E(d + )0-2 > 0

because of assumptions (i), (ii), and (iii) and Theorem 3.3.2 (Kolmogorov
LLN 2). Therefore condition D holds.
When ft( /30) has a very simple form, consistency can be proved more simply
and with fewer assumptions by using Theorem 4.1.1 or Theorem 4.1.2 di-
rectly rather than Theorem 4.3.1, as we shall show in the following example.
EXAMPLE 4.3.2. Consider the nonlinear regression model (4.3.1) with
(flu + xt)2 and assume (i) a flo b where a and b are real numbers
such that a < b, (ii) H1%, T 1x1 = q, and (iii) lim. T-1 X1:. 1x?
p> q2. Prove that the value of fl that minimizes Mr_ dy, (fl + x,)9 2 in the
domain [a, b] is a consistent estimator of /30 .
We have
1T
t_1 Eyt (fl + )0212
plim-
T
cr 3 (/31 _ /32)2 + 4(J 0 _ /3)2p + 4(fli _ fl2)(/30 _
=
where the convergence is clearly uniform in /3 E [a, b]. But, because
Q(fi) al + (fl0 fl)2(flo fl + 2 0,
Q(I3) is uniquely minimized at /I = flu. Therefore the estimator in question is
consistent by Theorem 4.1.1.

We shall now prove the asymptotic normality of the NLLS estimator of /3 by
making assumptions on f, to satisfy the assumptions of Theorem 4.1.3.
First, consider assumption C of Theorem 4.1.3. We have
asr = 2
OP
i EY, AP)] 43 (4.3.12)
Therefore we have
asT i 2 i u 9f1
(4.3.13)
afi lp, NIT r- aflip:
The results of Section 3.5 show that if we assume in the nonlinear regression
model that
1 T aft l
inn (L.) (4.3.14)
t-1 afl p o P
is a finite nonsingular matrix, 9
then the limit distribution of (4.3.13) is N(0, 4agC).
Second, consider the assumption of Theorem 4.1.5 that implies assumption
B of Theorem 4.1.3, except for nonsingularity. From (4.3.12) we have
1 a2s, _ 2 ,T Of, af, 2 ,?; a2ft (4.3.15)
afl 117 Top op' T A u t aflalr
7 -5-
2 T a2f,
-77 XIAA) am,
A I + A2 A3.
We must make the assumption of Theorem 4.1.5 hold for each of the three
terms in (4.3.15). First, for A 1 to satisfy the assumption of Theorem 4.1.5, we
must assume
T aft of,
T 73-fi a" converges to a finite matrix uniformly (4.3.16)
for all /3 in an open neighborhood of flo .

For A2 to converge to 0 in probability uniformly in a neighborhood offi c, , we
require (as we can infer from the proof of the convergence of A3 in Theorem
4.3.1) that
a2f, is continuous in 13 in an open (4.3.17)
neighborhood of it uniformly in t
and
Tce
1
T2 ,_, [afi,afi;
a2ft ]2 0 for all ft (4.3.18)
in an open neighborhood of A.
Finally, the uniform convergence of A3 requires
T a 2f,
T AAA) afgair L2 converges to a finite matrix uniformly
for all fil and /32 in an open neighborhood of A. (4.3.19)
Thus under these assumptions we have
1 a2sT r of: 1 of
plim 2 li m
T ;VI (4.3.20)
T aftap
whenever plim fi A.
These results can be summarized as a theorem.
THEOREM 4.3.2. In addition to the assumptions of the nonlinear regression
model and the assumptions of Theorem 4.3.1, assume conditions (4.3.14),
and (4.3.16)-(4.3.19). Then, if fir is a consistent root of the normal equation

(4.3.6), we have
4T(13T flo) MO, or i). (4.3.21)
As in condition C of Theorem 4.3.1, a simple way to verify conditions
(4.3.16) and (4.3.19) is to assume that x, are i.i.d. and use Theorem 4.2.1.
Condition (4.3.18) follows from the uniform convergence of T - IIT_ ;
((PA/V;(74)2, which in turn can be verified by using Theorem 4.2.1. It would
be instructive for the reader to verify these conditions for Example 4.3.1 and to
prove the asymptotic normality of the estimator in question.
When f,(/10) has a simple form, the convergence of 7 - la2sTiaflay can be
easily established and, therefore, a cumbersome verification of conditions
(4.3.16)-(4.3.19) can be avoided. To illustrate this point, we shall consider the
model of Example 4.3.2 again.
EXAMPLE 4.3.3. Consider the model of Example 4.3.2 and, in addition to
conditions (i), (ii), and (iii) given there, also assume (iv) a < /1b < b and (v)
= r. Obtain the asymptotic distribution of the value of 13
that minimizes Ir_i[Yt + 492 in the domain (-00, co).
Note that here the minimization is carried out over the whole real line
whereas in Example 4.3.2 it was done in the interval [a, b]. We denote the
unconstrained minimum by /3 and the constrained minimum by /3. In Exam-
ple 4.3.2 we proved plim /3 = 43..: But, because /30 is an interior point of [a, b]
by assumption (iv), firn P[11= fl]= 1; therefore we also have plim /3= 4) .
, a2sTI 1-, asrl (4.3.22)
IN 13' = LT ay 4T-
where )3* lies between /3 and fin. We have
1 asrl 4 i uxflo (4.3.23)
NIT afl fic, t-1
Therefore, using assumptions (ii) and (iii) and Theorems 3.5.3 and 3.5.5, we
obtain
asT -' N[0, 16ag(flg + p + 2floq)]. (4.3.24)
NIT
We also have
1 a2sT1
7- air r T s-i
- i + 49. (4.3.25)
Because the right-hand side of (4.3.25) is a continuous function of a finite fixed

number of sequences of random variables, we can use Theorem 3.2.6 to
evaluate its probability limit. Thus, because plim fl* = /30 , we obtain
. 1 82ST
phm = 8(/3,3 + p + 2130 q). (4.3.26)
T
Finally, (4.3.22), (4.3.24), and (4.3.26) imply by Theorem 3.2.7 (iii) that
2
Nrt( /30) a0
N [0, 40,1 p+ 2fl0 (4.3.27)
The distribution of the NLLS estimator may be approximated more accu-

rately by using the Edgeworth expansion (see Pfanzagl, 1973).
4.3.4 Bootstrap and Jacknife Methods

In this subsection, we shall consider briefly two methods of approximating the
distribution of the nonlinear least squares estimator; these methods are called
the bootstrap and the jackknife methods (see Efron, 1982, for the details). As
the reader will see from the following discussion, they can be applied to many
situations other than the nonlinear regression model.
The bootstrap method is carried out in the following steps:
1. Calculate a, = y, -A(fl), where fl is the NLLS estimator.
2. Calculate the empirical distribution function F of (u,).
3. Generate NT random variables (u:), i = 1, 2, . N and t =-
1, 2, . . . , T, according to F, and calculate y = fill) + u*.
4. Calculate the NLLS estimator fir that minimizes
[y: - fr(/))]2 for i = 1, 2, . . . , N.

t-
5. Approximate the distribution of/lby the empirical distribution function
of (fl?).
The jackknife method works as follows: Partition y as =
(31 , y . . . , y'N), where each y, is an m-vector such that mN = T. Let fl be
,
the NLLS estimator using all data and let ft , be the N_LLS estimator ob-
tained by omitting y,. Then "pseudovalues" /1* = - (N- i =
1., 2, . . . , N, can be treated like Nobservations (though not independent) on
ft. Thus, for example, Vfl may be estimated by (N - 1)-1
where fl* = N-1 I fit n.,(157-PM?r
It is interesting to note that /I* may be regarded as an estimator of p in its

own right and is called the jackknife estimator. Alcahira (1983) showed that in
the i.i.d. sample case the jackknife estimator is asymptotically equivalent to
the bias-corrected maximum likelihood estimator (see Section 4.2.4) to the
order r-
4.3.5 Tests of Hypotheses

In the process of proving Theorem 4.3.2, we have in effect shown that asymp-
totically
fir Pb 5E (G'G)-1 G'u, (4.3.28)
where we have put G =-- ofialinpo . Note that (4.3.28) exactly holds in the linear
case because then G = X. The practical consequence of the approximation
(4.3.28) is that all the results for the linear regression model (Model 1) are
asymptotically valid for the nonlinear regression model if we treat G as the
regressor matrix. (In practice we must use 0 where /3 is the NLLS
estimator.)
Let us generalize the t and Fstatistics of the linear model by this principle. If
the linear hypothesis Q1 = c consists of a single equation, we can use the
following generalization of (1.5.4):
Q',8c A t
(31Q ,(0,0)_ 14:11 Tk, (4.3.29)
where 11 means "asymptotically distributed as" and 6 2 = (T K)'S T (1).

Gallant (1975a) examined the accuracy of the approximation (4.3.29) by a
Monte Carlo experiment using the model
fi) = fi1xig + /32x2t + 134 exP (fi3x31). (4.3.30)
For each of the four parameters, the empirical distribution of the left-hand
side of (4.3.29) matched the distribution of tr_k reasonably well, although, as
we would suspect, the performance was the poorest for /1 3 .
If = c consists of q(> 1) equations, we obtain two different approximate
F statistics depending on whether we generalize the formula (1.5.9) or the
formula (1.5.12). Generalizing (1.5.9), we obtain
T K S(ft) ST(11) A
F(q, T k), (4.3.31)
ST(ft)
where ft is the constrained NLLS estimator obtained by minimizing S(fl)
subject to Q1 = c. Generalizing the formula (1.5.12), we obtain

T K (Q 11 ) r o1. '"'
) col
1 --- I
(0

)
A F(q, T k). (4.3.32)
q Sr(n)
These two formulae were shown to be identical in the linear model, but they
are different in the nonlinear model. A Monte Carlo study by Gallant (1975b),
using the model (4.3.30), indicated that the test based on (4.3.31) has higher
power than the test based on (4.3.32).
4.4 Methods of Iteration

Be it for the maximum likelihood or the nonlinear least squares estimator, we
cannot generally solve the equation of the form (4.1.9) explicitly for 0. Instead,
we must solve it iteratively: Start from an initial estimate of 0 (say 01 ) and
obtain a sequence of estimates (0) by iteration, which we hope will converge
to the global maximum (or minimum) of QT or at least a root of Eq. (4.1.9).
Numerous iterative methods have been proposed and used. In this section we
shall discuss several well-known methods that are especially suitable for ob-
taining the maximum likelihood and the nonlinear least squares estimator.
Many of the results of this section can be found in Goldfeld and Quandt
(1972), Draper and Smith (1981), or, more extensively, in Quandt (1983).
4.4.1 Newton-Raphson Method

The Newton-Raphson method is based on the following quadratic approxi-
mation of the maximand (or minimand, as the case may be):
Q(0)== Q(61 ) +
- + ( 0 61)'H1(0
(4.4.1)
where el is an initial estimate and
gi = aQ
ao
1 and H,
(32Q
.
The second-round estimator e2 of the Newton-Raphson iteration is obtained

by maximizing the right-hand side of the approximation (4.4.1). Therefore
62=61H7'g1. (4.4.2)
The iteration (4.4.2) is to be repeated until the sequence (o) thus obtained
converges.
Inserting iteration (4.4.2) back into approximation (4.4.1) yields
12(02) 42(61) (62 0 1)'il1( 02 (4.4.3)

Equation (4.4.3) shows a weakness of this method: Even if (4.4.3) holds
exactly, Q(02) > Q(01 ) is not guaranteed unless H 1 is a negative definite ma-
trix. Another weakness is that even if 11 1 is negative definite, 02 - 01 may be
too large or too small: If it is too large, it overshoots the target; if it is too small,
the speed of convergence is slow.
The first weakness may be alleviated if we modify (4.4.2) as
62 = 61 - (H 1 (4.4.4)
where I is the identity matrix and al is a scalar to be appropriately chosen by
the researcher subject to the condition that H 1 alI is negative definite. This
modification was proposed by Gokifeld, Quandt, and Trotter (1966) and is
called quadratic hill climbing. [Goldfeld, Quandt, and Trotter (1966) and
-
Goldfeld and Quandt (1972, Chapter 1) have discussed how to choose al and
the convergence properties of the method.]
The second weakness may be remedied by the modification
62 = - 1g, (4.4.5)
where the scalar 1, is to be appropriately determined. Fletcher and Powell
(1963) have presented a method to determined 1 1 by cubic interpolation of
Q(0) along the current search direction. [This method is called the DFP
iteration because Fletcher and Powell refined the method originally proposed
by Davidson (1959).] Also, Berndt et al. (1974) have presented another
method for choosing A I .
The Newton-Raphson method can be used to obtain either the maximum
likelihood or the nonlinear least squares estimator by choosing the appropri-
ate Q. In the case of the MLE, E(c3 2 log Lime') may be substituted for
92 log Llaeae, in defining H. If this is done, the iteration is called the method of
scoring (see Rao, 1973, p. 366, or Zacks, 1971, p. 232). In view of Eq. (4.2.22),
E(0 log Limp log Liao') may be used instead; then we need not calculate
the second derivatives of log L.
4.4.2 The Asymptotic Properties of the Second-Round Estimator in the

Newton-Raphson Method
Ordinarily, iteration (4.4.2) is to be repeated until convergence takes place.
However, if 0 1 is a consistent estimator of 00 such that Nri"(01 00) has a proper
limit distribution, the second-round estimator 6 2 has the same asymptotic

distribution as a consistent root of Eq. (4.1.9). In this case further iteration
does not bring any improvement, at least asymptotically. To show this, con-
sider
-2'21 e
(4.4.6)
gi = ao awe'
:3 s
where 0* lies between I), and 00. Inserting (4.4.6) into (4.4.2) yields
0 -1 a2c,
47052 - 00) = {I - [Tei
62t0,
-J [itorcilesll - 00) (4.4.7)
_[i 82Q , r 1 aQ
ole,
(m ,rt aolo:
But, because under the condition of Theorem 4.1.3
,. 1 82Q ,. 1 82Q ,. 1 492Q 1
pum = pum - pum (4.4.8)
T ma& T ma& 6.
we have
It(112 - 00) - [plim 1 2-9-1 ]-1-1-

'342 (4.4.9)
T 8086v aole:
which proves the desired result.
4.4.3 Gauss-Newton Method

The Gauss-Newton method was specifically designed to calculate the nonlin-
ear least square estimator. Expanding AP) of Eq. (4.3.5) in a Taylor series
around the initial estimate A, we obtain
Afi) ft(A) + (P (4.4.10)

(V' A
Substituting the right-hand side of (4.4.10) for f,(fl) in (4.3.5) yields
TI 8f
S (ft fid2 . (4.4.11)
tIL atr A
The second-round estimator 2 of the Gauss-Newton iteration is obtained by
minimizing the right-hand side of approximation (4.4.1 1) with respect to /I as
of,I _K1 1-1 as

[ air L] aL (4.4.12)
where
as
ap lit = 2 IX f' ii)] l
aafiL.
; (4.4.13)
The iteration (4.4.12) is to be repeated until convergence is obtained. This

method involves only the first derivatives off, whereas the Newton-Raphson
iteration applied to nonlinear least squares estimation involves the second
derivatives off, as well.
The Gauss-Newton iteration may be alternatively motivated as follows:
Evaluating the approximation (4.4.10) at fic, and inserting it into Eq. (4.3.1),
we obtain
fil3i) -al (4.4.14)

afl' A a PI
Then the second-round estimator 112 can be interpreted as the least squares
estimate of applied to the linear regression equation (4.4.14), treating the
whole left-hand side as the dependent variable and (af,/an as the vector of
independent variables. Equation (4.4.14) reminds us of the point raised at the
beginning of Section 4.3.5, namely, the nonlinear regression model asymptot-
ically behaves like a linear regression model if we treat 8f/813' evaluated at a
good estimate of /I as the regressor matrix.
The Gauss-Newton iteration suffers from weaknesses similar to those of the
Newton-Raphson iteration, namely, the possibility of an exact or near singu-
larity of the matrix to be inverted in (4.4.12) and the possibility of too much or
too little change from fi, to /32 .
To deal with the first weakness, Marquardt (1963) proposed a modification
[7. afl aft
A8 7 1,1wiasL, (4.4.15)
where al is a positive scalar to be appropriately chosen.

To deal with the second weakness, Hartley (1961) proposed the following
modification: First, calculate
aft I' I as
1
[
Az; (4.4.16)
and, second, choose A i to minimize

S(/11 + A I A1), OA I. (4.4.17)
Hartley proved that under general conditions his iteration converges to a root
of Eq. (4.3.6). (Gallant, 1975a, has made useful comments on Marquardt's
and Hartley's algorithms.)
As in the Newton-Raphson method, it can be shown that the second-round
estimator of the Gauss-Newton iteration is asymptotically as efficient as
NLLS if the iteration is started from an estimator fi i such that NIT(A fin)
converges to a nondegenerate random variable.
Finally, we want to mention several empirical papers in which the Gauss-
Newton iteration and related iterative methods have been used. Bodkin and
Klein (1967) estimated Cobb-Douglas and CES production functions by the
Newton-Raphson method. Charatsis (1971) estimated the CES production
function by a modification of the Gauss-Newton method similar to that of
Hartley (1961) and found that in 64 out of 74 samples it converged within six
iterations. Mizon (1977), in a paper whose major aim was to choose among
nine production functions including Cobb-Douglas and CES, used the conju-
gate gradient method of Powell (1964) (see Quandt, 1983). Mizon's article also
contained interesting econometric applications of various statistical tech-
niques we shall discuss in Section 4.5, namely, a comparison of the likelihood
ratio and related tests, Akaike information criterion, tests of separate families
of hypotheses, and the Box-Cox transformation (Section 8.1.2). Sargent
(1978) estimated a rational expectations model (which gives rise to nonlinear
constraints among parameters) by the DFP algorithm.
4.5 Asymptotic Tests and Related Topics

4.5.1 Likelihood Ratio and Related Tests
Let L(x, 0) be the joint density of a T-vector of random variables x =
(x1 , x2 ,. . . , xi.)' characterized by a K-vector of parameters 0. We assume
all the conditions used to prove the asymptotic normality (4.2.23) of the
maximum likelihood estimator 0. In this section we shall discuss the asymp-
totic tests of the hypothesis
h(0) = 0, (4.5.1)
where h is a q-vector valued differentiable function with q < K We assume
that (4.5.1) can be equivalently written as
0 --= r(a), (4.5.2)
where a is a p-vector of parameters such that p = K g. We denote the

constrained maximum likelihood estimator subject to (4.5.1) or (4.5.2) as
= r(61).
Three asymptotic tests of (4.5.1) are well known; they are the likelihood
ratio test (LRT), Wald's test (Wald, 1943), and Rao's score test (Rao, 1947).
The definitions of their respective test statistics are
max L(0)
LRT = 2 log 148)-- 2[1og LA log L(T
L(0)J, (4.5.3)
max L( 0)
8h
Wald = h(o)' (4.5.4)
0 maw
801 [1 821og i as h(e)'
(Hog Ll e[a2 log LI 8 log L (4.5.5)
Rao
80' I- awe' 9 ae
Maximization of log L subject to the constraint (4.5.1) is accomplished by
setting the derivative of log L A.'h(0) with respect to 0 and A to 0, where A is
the vector of Lagrange multipliers. Let the solutions be 0 and A. Then they
satisfy
a log _ I-
ae 80 i
Inserting this equation into the right-hand side of (4.5.5) yields Rao =
where
ah ir a2 log LI 1-1 Oh'I
B=
80' eL 8080' Id 001e .
Silvey (1959) showed that B is the asymptotic variance-covariance matrix of X
and hence called Rao's test the Lagrange multiplier test. For a more thorough
discussion of the three tests, see Engle (1984).
All three test statistics can be shown to have the same limit distribution,
x2(q), under the null hypothesis. In Wald and Rao, 82 log L/0080' can be
replaced with T plim T '02 log L/8000' without affecting the limit distribu-
tion. In each test the hypothesis (4.5.1) is to be rejected when the value of the
test statistic is large.
We shall prove LRT x2(q). By a Taylor expansion we have
a log L
log L(00) = log 40 (0 (4.5.6)
afr L es)
1 - c12 log LI
+ (00 ey woe, Ir mo ej,
where 0* lies between 00 and & Noting that the second term of the right-hand
side of (4.5.6) is 0 by the definition of 0, we have
log LA log 030 -2
1--1 iT(e. 00)' 30(6 00), (4.5.7)
where we have defined
. I 02 log L
.7 0 = E (4.5.8)
T aoao'
Treating L[r(a)] L(a) as a function of a, we similarly obtain
log i() - log L(a0 ) 1=1 T(er a 0)'. 7 Jet ao), (4.5.9)
where
1 .32 log L
,,= lim E (4.5.10)
T aaaa'
Noting L(00) = gad, we have from (4.5.3), (4.5.7), and (4.5.9)
00)'30(8 00) aoY cge ao). (4.5.11)
But from Theorem 4.1.3 and its proof we have
( 00) LI) g i i I a log Li (4.5.12)

ae 14,
and
Nrt(ii a0) 1-4.3 jv I a log LI (4.5.13)

aa ao

Since
log L 1 a log L
R' (4.5.14)
aa L ae
where R = (ar/aa'), and
1 a log L
---0 N(0, .70), (4.5.15)
ae
we have from (4.5.11)-(4.5.15)
LRT 0' (. R.7 VR')u, (4.5.16)
where u N(0, 18). Finally, defining
(4.5.17)
we obtain
LRT R
I e'(I .71/2R.7 -i, IR'.7 V2*. (4.5.18)
But, because
3 a = R'.70R, (4.5.19)
I .7 y2R.7 ;1R'.7 y2 can be easily shown to be an idempotent matrix of rank q.

Therefore, by Theorem 2 of Appendix 2, LRT --* x2(q).

The proof of Wald --ox2(q) and Rao --) x2(q) are omitted; the former is very
easy and the latter is as involved as the preceding proof.
Next we shall find explicit formulae for the three tests (4.5.3), (4.5.4), and
(4.5.5) for the nonlinear regression model (4.3.1) when the error u is normal.
Let fi be the NLLS estimator offlo and let /3 be the constrained NLLS, that is,
the value of p that minimizes (4.3.5) subject to the constraint h(fl) =0. Also,
define G = (f/afi')i and G = (of /o8'). Then the three test statistics are de-
fined as
LRT = T[log T-1 ST(h log T -1 ST(8)], (4.5.20)
Th(fi)' [19-1 (C'O) 1 '21-

1' I Ilh(ii)
Wald
ar A afi A
(4.5.21)
ST(fl)
and
Rao T[Y f(fi)]'G(Gt) IGIY f(ki . (4.5.22)

ST(i3 )
Because (4.5.20), (4.5.21), and (4.5.22) are special cases of (4.5.3), (4.5.4), and
(4.5.5), all three statistics are asymptotically distributed asx 2(q) under the null
hypothesis if u is normal. 1 Furthermore, we can show that statistics (4.5.20),
(4.5.21), and (4.5.22) are asymptotically distributed as x2(q) under the null
even if u is not normal. Thus these statistics can be used to test a nonlinear
hypothesis under a nonnormal situation.
In the linear model with linear hypothesis Q11= 0, statistics (4.5.20)
(4.5.22) are further reduced to
LRT = T log [ST( ji)/S T(/3)], (4.5.23)
Wald = T[ST(ft) ST(h/ST(11), (4.5.24)
and
Rao -- T[ST(i3 ) ST(fPlar(ft). (4.5.25)
Thus we can easily show Wald LRT Rao. The inequalities hold also in
the multiequation linear model, as shown by Berndt and Savin (1977). Al-
though the inequalities do not always hold for the nonlinear model, Mizon
(1977) found Wald LRT most of the time in his samples.
Gallant and Holly (1980) obtained the asymptotic distribution of the three
statistics under local alternative hypotheses in a nonlinear simultaneous
equations model. Translated into the nonlinear regression model, their results
can be stated as follows: If there exists a sequence of true values (N.) such that
lim fiI= /30 and 6 = lim ra(fi plim /3) is finite, statistics (4.5.20),
(4.5.21), and (4.5.22) converge to chi-square with q degrees of freedom and
noncentrality parameter A, where
_1
A= a72o , 81
Op
[ Oh
,TA(G'G)-1
it
Ohl
arlho. (4.5.26)
aft A
Note that if is distributed as a q-vector N(0, V), then (C + 'V'( is
distributed as chi-square with q degrees of freedom and noncentrality
parameter g 'V In other words, the asymptotic local power of the tests
based on the three statistics is the same.
There appear to be only a few studies of the small sample properties of the
three tests, some of which are quoted in Breusch and Pagan (1980). No
clear-cut ranking of the tests emerged from these studies.
A generalization of the Wald statistic can be used to test the hypothesis
(4.5.1), even in a situation where the likelihood function is unspecified, as long
as an asymptotically normal estimator ft of 13 is available. Suppose ft is
asymptotically distributed as N(fl, V) under the null hypothesis, with V
estimated consistently by V. Then the generalized Wald statistic is defined by
G.W. = [ ah
afi' 1-` (4
h(n) .5.27)
and is asymptotically distributed as 2(q) under the null hypothesis. Note that
(4.5.21) is a special case of (4.5.27).
Another related asymptotic test is the specification test of Hausman (1978).
It can be used to test a more general hypothesis than (4.5.1). The only
requirement of the test is that we have an estimator, usually a maximum
likelihood estimator, that is asymptotically efficient under the null hypothesis
but loses consistency under an alternative hypothesis and another estimator
that is asymptotically less efficient than the first under the null hypothesis but
remains consistent under an alternative hypothesis. If we denote the first
estimator by 0: and the second by e, the Hausman test statistic is defined by

(0 0)'V - i(0 0), where V is 1 consistent estimator of the asymptotic
variance-covariance matrix of (0 0). Under the null hypothesis it is
asymptotically distributed as chi-square with degrees of freedom equal to the
dimension of the vector 0.
If we denote the asymptotic variance-covariance matrix by V, it is well
known that V( 0 0) = V(0) V(0). This equality follows from
V( 0) = V12, where V12 is the asymptotic covariance between 0 and 0. To
verify this equality, note that if it did not hold, we could define a new estimator
0 + [V(0) V 12 ][V(0 0)1'(0 0), the asymptotic variance-covariance
matrix of which is V( 0) [V(0) Y.12l[V( 0 0)] -1 [V( 0 ) V12]', which
smaller (in the matrix sense) than V( 0). But this is a contradiction because 0 is
asymptotically efficient by assumption.
4.5.2 Akaike Information Criterion

The Akaike information criterion in the context of the linear regression model
was mentioned in Section 2.1.5. Here we shall consider it in a more general
setting. Suppose we want to test the hypothesis (4.5.2) on the basis of the
likelihood ratio test (4.5.3). It means that we choose L(0) over L(a) if
LRT > d, (4.5.28)
where d is determined so that P[LRT > dIL(a)] = c, a certain prescribed
constant such as 5%. However, if we must choose one model out of many
competing models L(a,), L(a 2 ), . . . , this classical testing procedure is at
best time consuming and at worst inconclusive. Akaike (1973) proposed a
simple procedure that enables us to solve this kind of model-selection prob-
lem. Here we shall give only a brief account; for the full details the reader is
referred to Akaike's original article or to Amemiya (1980a).
We can write a typical member of the many competing models
L(a,), L(a 2 ), . . . as L(a). Akaike proposed the loss function
2 11 L(61) ] _ 0)0 dx,
( 00 , ii) = T, log 17-kjW1.4 (4.5.29)
where ii is treated as a constant in the integration. Because W(0 0 , ii) ...

W(00 , 00 ) = 0, (4.5.29) can serve as a reasonable loss function that is to be
minimized among the competing models. However, because W depends on
the unknown parameters 00 , a predictor of W must be used instead. After
rather complicated algebraic manipulation, Akaike arrived at the following
simple predictor of W, which he called the Akaike Information Criterion

(AIC):
2 2p
AIC = -- log L(a) + (4.5.30)
T 7'
where p is the dimension of the vector a. The idea is to choose the model for
which AIC is smallest. Akaike's explanation regarding why AIC (plus a certain
omitted term that does not depend on ci or p) may be regarded as a good
predictor of W is not entirely convincing. Nevertheless, many empirical re-
searchers think that AIC serves as a satisfactory guideline for selecting a
model.
4.5.3 Tests of Separate Families of Hypotheses

So far we have considered testing or choosing hypotheses on parameters
within one family of models or likelihood functions. The procedures dis-
cussed in the preceding sections cannot be used to choose between two entirely
different likelihood functions, say L1( 0) and Lg(y). For example, this case
arises when we must decide whether a particular sample comes from a lognor-
mal or a gamma population. Such models, which do not belong to a single
parametric family of models, are called nonnested models.
Suppose we want to test the null hypothesis L1 against the alternative Lg .
Cox (1961, 1962) proposed the test statistic
Rr = log L( O) [E0 log L1(0)] log Lg(P) (4.5.31)
+ [E0 log L( 0 )J,

where ye = plime 9 (meaning the probability limit is taken assuming L1( 0) is
the true model) and indicates maximum likelihood estimates. We are to
accept Lf if Rf is larger than a critical value determined by the asymptotic
distribution of Rr. Cox proved that R1 is asymptotically normal with zero
mean and variance equal to E(v) E(v1w;)(Ew1w;)-1E(v1w1), where vf=
log L1(0) log L(y 9 ) E6 [log L1( 0) log Lg(ye )] and w1 = a log L1( 0)!
ae. Amemiya (1973b) has presented a more rigorous derivation of the asymp-
totic distribution of Rf.
A weakness of Cox's test is its inherent asymmetry; the test of L fagainst Lg
based on Rf may contradict the test of Ls against L1 based on the analogous
test statistic Rg . For example, L1 may be rejected by R1 and at the same time
Lg may be rejected by Rg.
Pesaran (1982) compared the power of the Cox test and other related tests.
For other recent references on the subject of nonnested models in general, see
White (1983).
4.6 Least Absolute Deviations Estimator
The practical importance of the least absolute deviations (LAD) estimator as a

robust estimator was noted in Section 2.3. Besides its practical importance,
the LAD estimator poses an interesting theoretical problem because the gen-
eral results of Section 4.1 can be used to prove only the consistency of the LAD
estimator but not its asymptotic normality, even though the LAD estimator is
an extremum estimator. In Section 4.6.1 we shall prove the asymptotic nor-
mality of the median, which is the LAD estimator in the i.i.d. sample case,
using a method different from the method of Section 4.1.2 and shall point out
why the latter method fails in this case. In Section 4.6.2 we shall use the general
results of Section 4.1.1 to prove the consistency of the LAD estimator in a
regression model. Finally, in Section 4.6.3 we shall indicate what lines of proof
of asymptotic normality may be used for the LAD estimator in a regression
model.
The cases where the general asymptotic results of Section 4.1 cannot be
wholly used may be referred to as nonregular cases. Besides the LAD estima-
tor, we have already noted some nonregular cases in Section 4.2.3 and will
encounter more in Sections 9.5 and 9.6. It is hoped that the methods outlined
in the present section may prove useful for dealing with unsolved problems in
other nonregular cases.
4.6.1 Asymptotic Normality of the Median

Let (Y,), t = 1,2, . . . , T, be a sequence of i.i.d. random variables with
common distribution function F and density function f The population me-
dian M is defined by
1
F(M) = (4.6.1)
V
We assume F to be such that M is uniquely determined by (4.6.1), which
follows from assuming f(y)> 0 in the neighborhood of y = M. We also as-
sume that f'(y) exists for y > M in a neighborhood of M. Define the binary
random variable W,(a) by
Ina) =-- 1 if l't a (4.6.2)

=0 if Y, < a
for every real number a. Using (4.6.2), we define the sample median m r by
m T =-- inf ta I I
T
t-i
W(a) A.
T
(4.6.3)
The median as defined above is clearly unique. 11

The asymptotic normality of rn T can be proved in the following manner.
Using (4.6.3), we have for any y
T
T
P(m r < M + T -112y)=-- P[E WAM + T - lay) -2]. (4.6.4)
Define
Pt = 1 - P(Y, <M + 7 -112y).
Then, because by a Taylor expansion
1
Ps = -i - T- '12f(M)y - (4.6.5)
we have
T T
P X HI g - ) -- P[Zr + 0(T-1/2 ) f(M)y], (4.6.6)
( t-1 2
where W:' = W,(M + 7- u2y) and Z. = T -11211:1(W;' - Ps ). We now derive
the limit distribution of ZT using the characteristic function (Definition 3.3.1).
We have
T
E exp WO = r 1 E exp [iAT -112(WI I - PO]
t-1
(4.6.7)
fl
= T (P exp [iAT -1 /2(1 - Ps )]
'
+ (1 - Ps ) exp (-0.7-1/2P0)
A2
= [1 -- o(T-1)]T
8T+
-0 exp (-). 2/8),
where the third equality above is based on (4.6.5) and the expansion of the
exponent: ex = 1 + x + 2- 1X2 Therefore ZT N(0, 4-1 ), which
. . . .
implies by Theorem 3.2.7

ZT QT -112 ) -) N(0, 4-1 ). (4.6.8)
Finally, from (4.6.4), (4.6.6), and (4.6.8), we have proved
N/T(m T M) N[0, 4-1f wy2i. (4.6.9)
The consistency of m r follows from statement (4.6.9). However, it also can
be proved by a direct application of Theorem 4.1.1. Let TnT be the set of the 0
points that minimizet 2
ST = Yr el lYt
(4.6.10)
Then, clearly,
clearly, m r E Tri r We have
rM
plim T- IST = 0 + 2 I .11(A) &I, 20 f(A) chi., (4.6.11)
Jo Jo
where the convergence can be shown to be uniform in 0. The derivation of
(4.6.1 1) and the uniform convergence will be shown for a regression model,
for which the present i.i.d. model is a special case, in the next subsection.
Because
aQ
+ 2F(0) (4.6.12)
and
a2Q
= 2f(0), (4.6.13)
802
we conclude that Q is uniquely minimized at 0 = M and hence m r is consist-
ent by Theorem 4.1.1.
Next, we shall consider two complications that prevent us from proving the
asymptotic normality of the median by using Theorem 4.1.3: One is that
asTme= 0 may have no roots and the other is that 82s,"802= 0 except for a
finite number of points. These statements are illustrated in Figure 4.2, which
depicts two typical shapes of ST.
Despite these complications, assumption C of Theorem 4.1.3 is still valid if
we interpret the derivative to mean the left derivative. That is to say, define for
A>0
OST = 0 has roots

( i )T--
8
aST
( ii)
de
= 0 has no roots
Figure 4.2 Complications proving the asymptotic normality of the median
OST = lint ST(0) ST(0 A)

(4.6.14)
ae A-0:1 A
Then, from (4.6.10), we obtain
aSZI = i [1 2 W,(M)]. (4.6.15)
Because ( WAM)) are i.i.d. with mean i and variance +, we have by Theorem
3.3.4 (Lindeberg-Livy CLT)
-47 A , N(0, 1).

1 as
(4.6.16)
Assumption B of Theorem 4.1.3 does not hold because a2sTiae2 = 0 for

almost every O. But, if we substitute Ea2w802b, for plim 7 t[a2sT mo]M
in
assumption B of Theorem 4.1.3, the conclusion of the theorem yields exactly
the right result (4.6.9) because of (4.6.13) and (4.6.16).
4.6.2 Consistency of Least Absolute Deviations Estimator

Consider a classical regression model
= (4.6.17)
where X is a TX K matrix of bounded constants such that lim T-I XA is a

finite positive-definite matrix and u is a T-vector of i.i.d. random variables
with continuous density function f( ) such that .1-gf(A) = I and f(x) > 0
for all x in a neighborhood of 0. It is assumed that the parameter space B is
compact. We also assume that the empirical distribution function of the rows
of X, (34), converges to a distribution function. The LAD estimator /1 is
defined to be a value of fl that minimizes
Sr = lyi 01 1u11. (4.6.18)

t 1
This is a generalization of (4.6.10). Like the median, the LAD estimator may
not be unique.
We shall now prove the consistency of the LAD estimator using Theorem
4.1.1. From (4.6.18) we have
ST = E Jo:145), (4.6.19)
t-i
where = fi fic, and h(zia) is defined as
If a 0, h(zia) = a if z 0 (4.6.20)
=a 2z if 0 <z<a
= a if z a.
If a <0, h(zia)= a if z a
=a-F2z if a<z<0
= a if z 0.
Then h(z1x;(5) is a continuous function of uniformly in t and is uniformly

bounded in t and J by our assumptions. Therefore h(u 11x;5) Eh(u,lx:(5)
satisfies the conditions for gz(y, 0) in Theorem 4.2.2. Moreover, lim T_,
1 Eh(u,lx;(5) can be shown to exist by Theorem 4.2.3. Therefore
Q plim T-1ST (4.6.21)
= 2 lim f f.) d)
T-0., 1 4,1
1T 1 r
2 lim E
Toco
f(A) (1). Ked+ lim
Toco
4,
where the convergence of each term is uniform in 6 by Theorem 4.2.3.

Thus it only remains to show that Q attains a global minimum at 5= 0.
Differentiating (4.6.21) with respect to S yields
aQ = 2 lim Z.
2;
1
[I"'
da
f(A) (LI. x,]+ lim
T
T
x
"
(4.6.22)
which is equal to 0 at (5 = 0 because PS./(2) dA =f by our assumption. More-

over, because
a2Q 1 T
ame 2 lim f(x;o)v4 (4.6.23)
is positive definite at S =0 because of our assumptions, Q attains a local

minimum at 6 = 0. Next we shall show that this local minimum is indeed the
global minimum by showing that VIM # 0 if S # 0. Suppose awao = 0 at
61 # 0. Then, evaluating (4.6.22) at S I and premultiplying it by S', we obtain
lim I i f' f(A) dAix

5 1 = 0. (4.6.24)
T ,_, 2 erth
To show (4.6.24) is a contradiction, let a 1 , a2, and Mbe positive real numbers
such that I x;(51 I < M for all t and f(A): a 1 whenever PI *5- a2. Such numbers
exist because of our assumptions. Then we have
[1-1 4a,
f()) ddx;61= 1' f(A) dA114,1 (4.6.25)
a l a2
m(5' X'XJ, .
Therefore (4.6.24) is a contradiction because of our assumption that

lim 'X'X is positive definite. This completes the proof of the consistency of
the LAD estimator by means of Theorem 4.1.1.
4.6.3 Asymptotic Normality of Least Absolute Deviations Estimator

As noted in Section 4.6.1, the asymptotic normality of LAD cannot be proved
by means of Theorem 4.1.3; nor is the proof of Section 4.6.1 easily generaliz-
able to the regression case. We shall give a brief outline of a proof of asymptotic
normality. The interested reader should refer to Koenker and Bassett (1982)
and the references of that article."
The asymptotic normality of the LAD estimator/I is based on the following
three fundamental results:
1
xtiv(yi ft)
; --2
a 0 0, (4.6.26)
T t-i
where qi(x) = sgn (x),
T 1 T
7,
,1 xm(y, - x,p) - (4.6.27)
_6 A xt[Ey/(y, ;fi)]i 77
r 7'
, ; Eliot - x'th)} 0,
and
T
xi[Ewy, - 4/3)] ; XtEY0t - xA) (4.6.28)
v T t-i
+ plim 7,
1 x [f
c r Ev(y t x't fi)]x flu).
These results imply
,IT(ft -
1,
fi0)L-9 foim 7 i; [ (TI
,' Ev(y t - 4/3)]} (4.6.29)
T
;Out).
V T t-i
Noting Ety(y, = 1 2F(x;6), Ev(14 1) = 0, and Ity/(u) = 1, we obtain
rt(11 fie) ' N(0, 4 - If(0)-2[Iim (4.6.30)
The proof of (4.6.26) is straightforward and is given in Ruppert and Carroll

(1980, Lemma A2, p. 836). The proof of (4.6.27) is complicated; it follows
from Lemma 1 (p. 831) and Lemma A3 (p. 836) of Ruppert and Carroll, who
in turn used results of Bickel (1975). Equation (4.6.28) is a Taylor series
approximation.
Exercises
1. (Section 4.1.1)
Prove (i) = (ii) = (iii) in Definition 4.1.1.
2. (Section 4.1.1)
In the model of Exercise 11 in Chapter 1, obtain the probability limits of
the two roots of the likelihood equation, assuming lim T_. 7 'x'x = c,
where c is a finite, positive constant.
3. (Section 4.1.1)
In the model of Exercise 2 (this chapter), prove the existence of a consist-
ent root, using Theorem 4.1.2.
4. (Section 4.1.2)
Suppose that (XT ) are essentially bounded; that is, for any E> 0, there
exists ME such that P(IXTI < Me) 1 - E for all T. Show that if
plimr_ YT = 0, then plim r_. XTYT= 0. (This is needed in the proof of
Theorem 4.1.4.)
5. (Section 4.2.2)
Prove (4.2.18) by verifying Condition D given after (4.1.7). Assume for
simplicity that a 2 is known. (Proof for the case of unknown (1 2 is similar
but more complicated.)
6. (Section 4.2.2)
Let Xi i = 1, 2, . . . , n,t = 1, 2, . . . , T, be independent with the dis-
tribution N(p cr 2). Obtain the probability limit of the maximum likeli-
hood estimator of a 2 assuming that n is fixed and T goes to infinity (cf.
Neyman and Scott, 1948).
7. (Section 4.2.3)
Let {X,), t = 1, 2, . . . , T, be i.i.d. with the probability distribution
X, = 1 with probability p
= 0 with probability 1 p.
Prove the consistency and asymptotic normality of the maximum likeli-
hood estimator using Theorems 4.1.1 and 4.2.4. (The direct method is
much simpler but not to be used here for the sake of an exercise.)
8. (Section 4.2.3)
Prove the asymptotic normality of the consistent root in the model of
Exercise 2 (this chapter).
9. (Section 4.2.3)
Let (X,) be i.i.d. with uniform distribution over (0, 0). Show that if 0 is
defined by T-.:(T + 1) max (Xi , X2 XT.),
P [T(0 0) < x] = exp (0-1x 1) for x 0.
10. (Section 4.2.3)
Consider the model
Yi =Po + t = 1, 2, . . . , T,
where y, and u, are scalar random variables and /30 is a scalar unknown
parameter. If (u,) are i.i.d. with Eu,= 0, Et4 = fig, Eu?= 0, and
Eu: = m4 (note that we do not assume the normality of u,), which of the
following three estimators do you prefer most and why?
(1) = T-1171Y1,
(2) fl. ,2 which maximizes S = (T/2) log (1 /2p2)I7
(3) /33 defined as 0.5 times the value of /3 that minimizes
w=t rYt fl1 2.
L fl ri
11. (Section 4.2.3)
Derive the asymptotic variance of the estimator of fl obtained by mini-
mizing IT:. 1(y, /3x,)4, where y, is independent with the distribution
N(/30x ag) and limT, Ix? is a finite, positive constant. You may
assume consistency and asymptotic normality. Indicate the additional
assumptions on x, one needs. Note if Z N(0, a 2), Erk = cr2k(2k)!I
(2kk!).
12. (Section 4.2.4)
Complete the proof of Example 4.2.4 the derivation of the asymptotic
normality of the superefficient estimator.
13. (Section 4.2.5)
In the model of Example 4.2.3, obtain the asymptotic variance-covar-
iance matrix of /3 using the concentrated likelihood function in /3.
14. (Section 4.3.2)
What assumptions are needed to prove consistency in Example 4.3.2
using Theorem 4.3.1?
15. (Section 4.3.3)
Prove the asymptotic normality of the NLLS estimator in Example 4.3.1.
16. (Section 4.3.3)

Consider a nonlinear regression model
Yt = (fio + 42 + u
where we assume
(A) (u,) are i.i.d. with Eu, = 0 and Vu, a.
(B) Parameter space B = f].
(C) (x,) are i.i.d. with the uniform distribution over [1, 2], distributed
independently of (u,). [EX' = (r + 1)-1(2T+ 1 1) for every positive or
negative integer r except r = 1. EX -1 = log 2.]
Define two estimators of Po:
(1) minimizes ST(/J) = /11[3/1 + 49 2 over B.
(2) /3 minimizes W T(fl) = + x,)2 + log [(13 + x,)2]) over B.
If/I0 = 0, which of the two estimators do you prefer? Explain your prefer-
ence on the basis of asymptotic results.
17. (Section 4.3.5)
Your client wants to test the hypothesis a + fi = 1 in the nonlinear re-
gression model
Q, L7Kf + u t 1, 2, . . , T,
where L, and K, are assumed exogenous and (u,) are i.i.d. with Eu, = 0
and Vu, = a2. Write your answer in such a way that your client can
perform the test by reading your answer without reading anything else,
except that you may assume he can compute linear least squares estimates
and has access to all the statistical tables and knows how to use them.
Assume that your client understands high-school algebra but not calculus
or matrix analysis.
18. (Section 4.4.3)
Prove the asymptotic efficiency of the second-round estimator of the
Gauss-Newton iteration.
19. (Section 4.5.1)
Prove Wald -b x2(q), where Wald is defined in Eq. (4.5.4).
20. (Section 4.5.1)
Prove Rao > x2(q), where Rao is defined in Eq. (4.5.5).
21. (Section 4.5.1)
Show Wald LRT Rao, where these statistics are defined by (4.5.23),
(4.5.24), and (4.5.25).
22. (Section 4.6.1)

Consider the regression model y, = flox, + u where (x,) are known con-
stants such that limp_ 7 'E l.4 is a finite positive constant and (u,)
satisfy the conditions for ( Y,) given in Section 4.6.1. By modifying the
proof of the asymptotic normality of the median given in Section 4.6.1,
prove the asymptotic normality of the estimator of )6 obtained by mini-
mizing IT_ I ly, /3x11.
23. (Section 4.6.1)
Let (X1) be i.i.d. with a uniform density over (f, f) and let Y be the
binary variable taking values r- v2 and T-I/2 with equal probability
and distributed independently of (X,). Define W, = 1 if X, + Y 0 and
W, = 0 if X1 + Y< 0. Prove that T- on: 1 ( W, I) converges to a mix-
ture of N(1, i) and N( 1, I) with equal probability.
5 Time Series Analysis
Because there are many books concerned solely with time series analysis, this
chapter is brief; only the most essential topics are considered. The reader who
wishes to study this topic further should consult Doob (1953) for a rigorous
probabilistic foundation of time series analysis; Anderson (1971) or Fuller
(1976) for estimation and large sample theory; Nerlove, Grether, and Car-
valho (1979) and Harvey (1981a, b), for practical aspects of fitting time series
by autoregressive and moving-average models; Whittle (1983) for the theory
of prediction; Granger and Newbold (1977) for the more practical aspects of
prediction; and Brillinger (1975) for the estimation of the spectral density.
In Section 5.1 we shall define stationary time series and the autocovariance
function and spectral density of stationary time series. In Section 5.2 autore-
gressive models will be defined and their estimation problems discussed. In
Section 5.3 autoregressive models with moving-average residuals will be de-
fined. In Section 5.4 we shall discuss the asymptotic properties of the IS and
ML estimators in the autoregressive model, and in Section 5.5 we shall discuss
prediction briefly. Finally, in Section 5.6 we shall discuss distributed-lag
models.
5.1 Introduction
5.1.1 Stationary Time Series

A time series is a sequence of random variables {y 1 ), t = 0, 1, 2, . . . . We
assume Ey, = 0 for every t. (If Ey,* 0, we must subtract the mean before we
subject it to the analysis of this chapter.) We say a sequence (y 1 ) is strictly
stationary if the joint distribution of any finite subset..11, v v. 12 ) 1 YU de-
pends only on t2 t1 , t3 ti , . . , t, t, and not on t, . We say a sequence is
weakly stationary if Ey,y, depends only on It sl and not on t. If a process is
strictly stationary and if the autocovariances exist, the process is weakly sta-
tionary.
In Section 5.2 through 5.4 we shall be concerned only with strictly station-
ary time series. The distributed-lag models discussed in Section 5.6 are gener-
ally not stationary in either sense. Time series with trends are not stationary,
and economic time series often exhibit trends. However, this fact does not
diminish the usefulness of Section 5.2 through 5.4 because a time series may
be analyzed after a trend is removed. A trend may be removed either by direct
subtraction or by differencing. The latter means considering first differences
(y t y 1 ), second differences ((y1 ye-1) (y1-1 Yr--2 )), and so forth.
There are three fundamental ways to analyze a stationary time series. First,
we can specify a model for it, such as an autoregressive model (which we shall
study in Section 5.2) or a combined autoregressive, moving-average model
(which we shall study in Section 5.3). Second, we can examine autocovar-
iances Ey,y,, h , h = 0, 1, 2, . . . . Third, we can examine the Fourier trans-
form of autocovariances called spectral density. In Sections 5.1.2 and 5.1.3 we
shall study autocovariances and spectral density.
5.1.2 Autocovariances
Define yh= EYtYr+h, h= 0, 1, 2, . . . . A sequence (Yh ) contains important
information about the characteristics of a time series (ye). It is useful to
arrange (yh) as an autocovariance matrix
Yo Yi Y2 VT-1
Y1 Yo Vi YT-2
Y2 Y1
(5.1.1)
Vi
7T-1 YT-2 Y1 Yo
This matrix is symmetric, its main diagonal line consists only of yo , the next
diagonal lines have only yi , and so on. Such a matrix is called a Toeplitz form.
5.1.3 Spectral Density

Spectral density is the Fourier transform of autocovariances defined by
00
.11 (0) = Yhe ilupt (5.1.2)

h ce
provided the right-hand side converges.

Time Series Analysis 161
Substituting eu = cos ). + i sin A, we obtain
f(co) = yh [cos (hco) i sin (hco)] (5.1.3)
= I cos OM,
h-
where the second equality follows from yh = y_ h and sin A = sin ( A).
Therefore spectral density is real and symmetric around co = 0.
Inverting (5.1.2), we obtain
yh = (270- f elhaf(co) dco (5.1.4)
= f cos (hco)f(co) dco.
An interesting interpretation of (5.1.4) is possible. Suppose y, is a linear com-

bination of cosine and sine waves with random coefficients:
3): = [4 cos (wkt) Ck sin MO], (5.1.5)

k- I
where CO k = larin and (4) and (CO are independent of each other and inde-
pendent across k with E4 = ECk = 0 and Tik= VCk = al. Then we have
yh = ai cos (CO kh), (5.1.6)

k- 1
which is analogous to (5.1.4). Thus a stationary time series can be interpreted

as an infinite sum (actually an integral) of cycles with random coefficients, and
a spectral density as a decomposition of the total variance of y, into the
variances of the component cycles with various frequencies.
There is a relationship between the characteristic roots of the covariance
matrix (5.1.1) and the spectral density (5.1.2). The values of the spectral
density f(co) evaluated at T equidistant points of co in [ n, n] are approxi-
mately the characteristic roots of IT (see Grenander and Szego, 1958, p. 65,
or Amemiya and Fuller, 1967, p. 527).
5.2 Autoregressive Models

5.2.1 First-Order Autoregressive Model
Consider a sequence of random variables (ye ), t = 0, 1, 2, . . , which
follows
Y: = PYt-i es, (5.2.1)
where we assume
ASSUMPTION A. (,), t = 0, 1, 2, . . . . , are i.i.d. with EE,= 0 and
= a2 and independent of y,_, , yr-2 , . . .
ASSUMPTION B. Ipl < 1.
ASSUMPTION C. Ey,= 0 and Ey,y, +h = yh for all t. (That is, (y,) are wealdy
stationary.)
Model (5.2.1) with Assumptions A, B, and C is called a stationary first-order
autoregressive model, abbreviated as AR(1).
s-
= P% -s (5.2.2)
J 0-
But lim s-. E(P 3Ys-5) 2 =0 because of Assumptions B and C. Therefore we

have
40
Y1=1 fk ift-i, (5.2.3)

./3
which means that the partial summation of the right-hand side converges to y,
in the mean square. The model (5.2.1) with Assumptions A, B, and C is
equivalent to the model (5.2.3) with Assumptions A and B. The latter is called
the moving-average representation of the former.
A quick mechanical way to obtain the moving-average representation
(5.2.3) of (5.2.1) and vice-versa is to define the lag operator L such that
Ly, = y t_ i , L 2Yt = Yr 2, Then (5.2.1) can be written as
-
( 1 PL)Ys = es (5.2.4)
where 1 is the identity operator such that ly, = y1 . Therefore
Yr = ( 1 PL) -l er = piv) E (5.2.5)

7-6
which is (5.2.3).
An AR(1) process can be generated as follows: Define y o as a random

variable independent of e l , 2 ,. . , with Eyo = 0 and Ey1,= 20. p2).
Then define y, by (5.2.2) after putting s = t.
The autocovariances (yh ) can be expressed as functions of p and a2 as
follows: Multiplying (5.2.1) with yi_h and taking the expectation yields
yh = pyh_ l , h= 1,2 . . . . (5.2.6)
From (5.2.1), E(y, py i_ 1 )2 = Eq, so that we have
( 1 P2 )Y0 2PYI = a2. (5.2.7)
Solving (5.2.6) and (5.2.7), we obtain
a 2ph
Yh = h = 0, 1, 2, . . (5.2.8)
Note that Assumption C implies y_ h = yh .

Arranging the autocovariances in the form of a matrix as in (5.1.1), we
obtain the autocovariance matrix of AR(1),
pT-1
pT-2
a2
(5.2.9)
Xi= 1 p2
pr- 1 1
Now let us examine an alternative derivation of %, that is useful for deriving
the determinant and the inverse of /1 and is easily generalizable to higher-
order processes. Define T-vectors Y = (3' Y2,. , yTy and e to =
[( 1 P2 ) 1123/ 1 2, 3, err and a TX T matrix
(1 p2 )1/2 0 0 0
1 0
0 p 1
R1= (5.2.10)
0 p 10
0 0 p 1
Then we have
RI y = eto . (5.2.11)
But, because Eet;41') = (721, we obtain
-= (5.2.12)
which can be shown to be identical with (5.2.9). Taking the determinant of

both sides of (5.2.12) yields
0.2T
Izil (5.2.13)
Inverting both sides of (5.2.12) yields

1
= az II (5.2.14)
P 0 0 0
1 p2 0 0
1 p2
1+p 2
o p
By inserting (5.2.8) into (5.1.2), we can derive the spectral density of AR(1):
a2 e)
fl(W) 1 p2 plme- thw (5.2.15)
a2 [
= p2 1 + A(pekur + (pe - kih]
a2 [ pe lw ne-rro
_ p2 +
Pe) pe
a2
1 2p cos co + p2 '
5.2.2 Second-Order Autoregressive Model

A stationary second-order autoregressive model, abbreviated as AR(2), is
defined by
Yr= Pr Yr-1 P2Yr-2 Err t = 0, 1, 2, ., (5.2.16)
where we assume Assumptions A, C, and
ASSUMPTION B'. The roots of z 2 piz p2 = 0 lie inside the unit circle.
Using the lag operator defined in Section 5.2.1, we can write (5.2.16) as
(1 Pi P2L2 )Yi = et. (5.2.17)
Hence,
(1 ALM P2 nYt = (5.2.18)
where p, and /22 are the roots of z 2 piz p2 = 0. Premultiplying (5.2.18) by
(1 piL)-1 (1 122 L) l , we obtain
Yt = Pip IdLk ) Et (5.2.19)

j -0 k-0
=
j,kg0
Convergence in the mean-square sense of (5.2.19) is ensured by Assumption

B'. Note that even if p, and p2 are complex, the coefficients on the e, are
always real.
The values of p i and p2 for which the condition I,u, I Ip21< 1 is satisfied
correspond to the inner region of the largest triangle in Figure 5.1. In the
region above the parabola, the roots are real, whereas in the region below it,
they are complex.
The autocovariances may be obtained as follows: Multiplying (5.2.16) by
y, and taking the expectation, we obtain
Y1 = PiYo P2Y1. (5.2.20)
Squaring each side of (5.2.16) and taking the expectation, we obtain
Yo= (Pi Pi)Yo+ 2P1P2Yi a2- (5.2.21)
P2
Figure 5.1 Regions of the coefficients in AR(2)

Solving (5.2.20) and (5.2.21) for yo and y1 , we obtain

a2(P2 l) (5.2.22)
Yo (1 + P2 )[Pi ( 1 P2)2 ]
and
a 2p1
Yi ( (5.2.23)
1 P2)[Pi ( 1 P2)2 ] .
Next, multiplying (5.2.16) by 3,:_h and taking the expectation, we obtain
Y h P lY h-1 + P2 Y h-2 h 2. (5.2.24)

Note that (yh ) satisfy the same difference equation as (y,) except for the
random part. This is also true for a higher-order autoregressive process. Thus,
given the initial conditions (5.2.22) and (5.2.23), the second-order difference
equation (5.2.24) can be solved (see Goldberg, 1958) as
Pi (5.2.25)
Yh ,, 112Y0)
f41 P2
(Y1 -. AN) if #1 I.12
= I [h(Y 1 - PYO) 1-00] if 1 1 1 = 142 =

If u 1 and /12 are complex, (5.2.25) may be rewritten as
1h- 1 [y, sin h0 yo r sin (h 1)9]
Yh (5.2.26)
sin 0
where /1 1 = reie and /22 =
Arranging the autocovariances given in (5.2.25) in the form of (5.1.1) yields
the autocovariance matrix of AR(2), denoted 12 . We shall not write it explic-
itly; instead, we shall express it as a function of a transformation analogous to
(5.2.10). If we define a T-vector (av av
=-1-1, 2,1 + a3,v2, 32 -42 9 E T)
and a TX T matrix
al 0 0
a2 a3 0
- P2 - P1 1 0
R2= 0 - P2 - Pi 1 0 (5.2.27)

1 0
0 0 - P2 Pi 101
we have
R2y = ea) . (5.2.28)
Now, if we determine a 1 , a2 , and a3 by solving V(a l y,)= a2 ,

V(a2y1 a3y2 ) = a2, and L[a 1 y1 (a2 y1 a3y2 )] = 0, we have L'ea ) er2S = 0 2I.
Therefore we obtain from (5.2.28)
= cr2Ril(%) 1. -
(5.2.29)
Higher-order autoregressive processes can be similarly handled.
5.2.3 pth-Order Autogressive Model

A stationary pth-order autoregressive process, abbreviated as AR(p), is de-
fined by
Yt = PjYt-j + f t , t= 0, 2, , (5.2.30)
where we assume Assumptions A and C and
ASSUMPTION B". The roots of Dr_o = 0, Po = 1, lie inside the unit

circle.
A representation of the T X T autocovariance matrix ID of AR(p) analo-

gous to (5.2.12) or (5.2.29) is possible. The ], kth element (j, k = 0, 1, . . .
T 1) of can be shown to be the coefficient on VCk in
T-1 p p p-1
pAi pi C1- C3 I PA T-i

t-o ,fro .fro i-o
where Po = I (see Whittle, 1983, p. 73).
We shall prove a series of theorems concerning the properties of a general
autoregressive process. Each theorem except Theorem 5.2.4 is stated in such a
way that its premise is the conclusion of the previous theorem.
THEOREM 5.2.1. (y,) defined in (5.2.30) with Assumptions A, B", and C can
be written as a moving-average process of the form
J-0 J-0
< (5.2.31)
where (i ) are i.i.d. with at = 0 and E= a2.

Proof From (5.2.30) we have
[11 ( 1 Yt = Et (5.2.32)
where /L I , 1.42 , . . . , pp are the roots of 11_0 pi zP-j. Therefore we have
t
11 ( k2 kW)]
J-1 -o
[
Et . (5.2.33)
Equating the coefficients of (5.2.31) and (5.2.33), we obtain

=1 (5.2.34)
01 =PI +P2+. +111,
=
pStkial
Ilioui2 A..
Therefore
A 100( 1 +/44+4+. ) ( 1 1PJP <

(5.2.35)
where g m = max[1/111,11221, . . . , 11411.
THEOREM 5.2.2. Let (y,) be any sequence of random variables satisfying

(5.2.31). Then
(5.2.36)
cro
where yh = EYtYt+h-
Proof We have
70 = a201 + + .)
6r2(4201 + 0102 + .)
72 = a2(0043 2 043 and so on.
Therefore
ii 2,
(5.2.37)
iyhi,5 0'2 (7. )
from which the theorem follows.

THEOREM 5.2.3. Let (y,) be any stationary sequence satisfying (5.2.36).
Then the characteristic roots of the T X T autocovariance matrix X of (y,) are
bounded from above.
Proof Let x = (xo , x1 , . . . , x7 _,)' be the characteristic vector of X corre-
-
sponding to a root A. That is,

Xx = Ax. (5.2.38)
Suppose 1x11 max [1x01,1x11, . , IxT_ g. The (t + 1)st equation of (5.2.38)
can be written as
IV() + Y: -- I X' + + YoXt+ + YT-1tXT-1 = AXt (5.2.39)
Therefore, because A 0,
17,1+ lyt-il + +IY01+ IYT i ti
- -
(5.2.40)
Therefore
(5.2.41)
h=b
from which the theorem follows.
The premise of the next theorem is weaker than the conclusion of the
preceding theorem. In terms of the spectral density, the premise of Theorem
5.2.4 is equivalent to its existence, and the conclusion of Theorem 5.2.3 to its
continuity.
THEOREM 5.2.4. Let (y1 ) be any sequence of random variables satisfying
.vt= (/)./ Et-j, (5.2.42)

.FO 44<
where (,) are i.i.d with at = 0 and EE= a 2 . Then
lim yh = 0. (5.2.43)
h_...
Note that (5.2.31) implies (5.2.42).
Proof The theorem follows from the Cauchy-Schwartz inequality,

namely,
)2 (5.2.44)
(74(00(fin + icAh+1 +
5.3 Autogressive Models with Moving-Average Residuals

A stationary autoregressive, moving-average model is defined by
A .fro
Po = flo - 1 , t 0,1,+2,.
(5.3.1)
where we assume Assumptions A, B", C, and

ASSUMPTION D. The roots of El 0 fig q -j =0 lie inside the unit circle.
Such a model will be called ARMA(p, q) for short.
We can write (5.3.1) as
p(L)y, = fl(L)E, (5.3.2)
where p(L) = p Li and fl(L) =11_ 0 AD. Because of Assumptions B"

and C, we can express y, as an infinite moving average
= P-1(1-)IAL)f1 ' 0(4E,, (5.3.3)
where cl)(L) =_ 0 (kV. Similarly, because of Assumption D, we can express
y, as an infinite autoregressive process
IP(L)Y1-' 13-1(L)p(L)Yr = E (5.3.4)
where yi(L) = E;"_ 0
The spectral density of ARMA(p, q) is given by
2
f(w) 0.2 fre -it 5_ co (5.3.5)

2
Pie- Vt
.frO
where 1z1 2 = zi for a complex number z with i being its complex conjugate.
Note that (5.3.5) is reduced to (5.2.1 5) in the special case of AR(1). We also see
from (5.3.5) that the spectral density of a moving-average model is, except for
02, the inverse of the spectral density of an autoregressive model with the same
order and the same coefficients. Because the spectral density of a stationary
process approximately corresponds to the set of the characteristic roots of the
autocovariance matrix, as was noted earlier, we can show that the autocovar-
iance matrix of a moving-average model is approximately equal to the inverse
of the autocovariance matrix of the corresponding autoregressive model. We
shall demonstrate this for the case of MA(1).
Consider an MA(1) model defined by
Y: = E PE: -1, (5.3.6)
where IA < 1 and (,) are i.i.d. with EE, = 0 and Ve, = a 2. The T X Tautoco-
variance matrix is given by
+pi p 0
p 1+ p2
20) = 0
(5.3.7)
0
1 p2
1
Pl
p2
We wish to approximate the inverse of ; 1) . If we define a T-vector ti such that

its first element is e l Pei) (1 p2 )112e 1 and the other T 1 elements are
zeroes, we have
y = RI + ( 5.3.8)
where R I is given by (5.2.10). Therefore'
; 1) 0.2R1Rc. (5.3.9)
But we can directly verify

(5.3.10)
From (5.2.12), (5.3.9), and (5.3.10), we conclude

(7 21; (5.3.11)
1 v
`w1.
=
Whittle (1983, p. 75) has presented the exact inverse of 1, (, ) . The j, kth
element (j, k= 0, 1, . . , T 1) of 1-5, denoted Vet) , is given by
n-(J+1) _ pi + I )(r(T- lc) pT- Ic)

(r (5.3.12)
(1) Ar 1 (T+ I) /7+ I )
However, the exact inverse is complicated for higher-order moving-average

models, and the approximation (5.3.11) is attractive because the same inverse
relationship is also valid for higher-order processes.
Anderson (1971, p. 223) has considered the estimation of the parameters of
ARMA(p, q) without the normality assumption on (,), and Box and Jenkins
(1976) have derived the maximum likelihood estimator assuming normality
in the autoregressive integrated moving-average process. This is the model in
which the sequence obtained by repeatedly first-differencing the original
series follows ARMA(p, q). The computation of MLE can be time-consuming
because the inverse of the covariance matrix of the process, even if we use the
approximation mentioned in the preceding paragraph, is a complicated non-
linear function of the parameters (see Harvey, 1981b, for various computa-
tional methods).
5.4 Asymptotic Properties of Least Squares and Maximum Likelihood

Estimator in the Autoregressive Model
We shall first consider the least squares (LS) estimation of the parameters p's
and a2 of an AR(p) model (5.2.30) using T observations Yi, Y2, , Y T . In
vector notation we can write model (5.2.30) as
where
y = Yp + e,
Y = Yp+1 Yp+2
/32, . . , pp r , and
Y=
Yp Yp-1
Yp+ 1 Yp
YT) ',
Y1
Y2
1
e = (Ep+l ep+2) 9 er)'
(5.4.1)
Y7.- I YT. -2 YT-p
The LS estimator /3 of p is defined as = (Y'Y)' Y'y. In the special case of

p = 1, it becomes
(5.4.2)
Model (5.4.1) superficially looks like the classical regression model (1.1.4),
but it is not because the regressors Y cannot be regarded as constants. This
makes it difficult to derive the mean and variance offi for general p. However,
in the case of p = 1, the exact distribution of p can be calculated by direct
integration using the method of Imhof (1961). The distribution is negatively
skewed and downward biased. Phillips (1977) derived the Edgeworth expan-
sion [up to the order of 0(T-1 )] of the distribution of p in the case of p = 1
assuming the normality of (,) and compared it with the exact distribution. He
found that the approximation is satisfactory for p = 0.4 but not for p = 0.8.
In the general AR(p) model we must rely on the asymptotic properties of the
LS estimator. For model (5.4.1) with Assumptions A, B", and C, Anderson
(1971, p. 193) proved
It(05 N(0, a2IT, ), (5.4.3)
where Z, is the p X p autovariance matrix of AR(p). 2 We can estimate a 2

consitelyb
a2 = 7-1 (Y YR& YO). (5.4.4)

Because plim r, T-1Y'Y -= 4, the distribution of 1) may be approximated
by N[p, a 2(Y'Y)i]. Note that this conclusion is identical to the conclusion we
derived for the classical regression model in Theorem 3.5.4. Thus the asymp-
totic theory of estimation and hypothesis testing developed in Chapters 3 and
4 for the classical regression model (1.1.4) can be applied to the autoregressive
model (5.4.1). There is one difference: In testing a null hypothesis that speci-
fies the values of all the elements ofp, a 2rp- ' need not be estimated because it
depends only on p [see (5.4.16) for the case of p = 1].
We shall consider the simplest case of AR(1) and give a detailed proof of the
consistency and a sketch of the proof of the asymptotic normality. These are
based on the proof of Anderson (1971).
Yt-lEt
,5 p = t ( 5.4.5)
,2
We shall prove consistency by showing that 7 times the numerator con-

verges to 0 in probability and that 7 ' times the denominator converges to a
positive constant in probability.
The cross product terms in (11.- 2 Yt-iEt) 2 are of the form Yai+set+IEt+1-4-s
But their expectation is 0 because E l+ 1 +, is independent of ytyt+,tet+, . Thus

1 T )2 T-1 a4
E( t2 y 2
(5.4.6)
T2 1 p
Therefore, by Theorem 3.2.1 (Chebyshev), we have
T
PliM r
7.-
E
"' I t 2
0. (5.4.7)
Putting et = pyt_ i in (5.4.7), we have

T
ppm
. YI-1(Yt PYr- = O. (5.4.8)
By Theorem 3.3.2 (Kolmogorov LLN 2), we have

1 T
Vim. PYr- = a2. (5.4.9)
By adding (5.4.9) and 2p times (5.4.8), we have
i
t2
p 2 .--,
j
3 1 ) = a z. (5.4.10)
Therefore
plim -
1 62p2 + 1 p2 plim
T1 (y; (5.4.11)
T--
But the last term of (5.4.11) is 0 because of (3.2.5) generalized Chebyshev's
inequality. Therefore
1 T 2
Vi.m7 37 1 fl pv
, (5.4.12)
The consistency of "P follows from (5.4.5), (5.4.7), and (5.4.12) because of
Theorem 3.2.6.
Next consider asymptotic normality. For this purpose we need the follow-
ing definition: A sequence (vi) is said to be K-dependent if (vt, , v,. . . , vt.)
are independent of (v51 , vsa , . . . , vi.) for any set of integers satisfying
t, < t2 . . < t,, < s, < s2 < . . . < sm and t + K < s,
To apply a central limit theorem to MT., v split the T observations into S
successive groups with M + Kobservations [so that S(M + K)= 1] and then
in each group retain the first M observations and eliminate the remaining K
observations. If we choose S and M in such a way that S -> M -> 00, and
SIT-> 0 as T-0 cc, the elimination of K observations from each of the S
groups does not matter asymptotically.
We can write
1 T 1 r
1 Vm + ANT,
Trf-,IYt-IEt = IT t.2 (5.4.13)
i
where VArt -=': Et Mis c) P set 1 _s and ANT =-
NIT
(
I N+1
.--t .7 Pset-i-s). But we
have
EAT =
(2
I
s N+1
E
w+1
(5.4.14)
a4( T 1) - P23
T t-;+
Therefore ANT can be ignored for large enough N. (Anderson, 1971, Theorem
7.7.1, has given an exact statement of this.) We can show that for a fixed N2
Ellm = 0, Ed, = a > p 2i, and Evm vm. =0 for t * T.
Moreover, vm and VN.t +N+2 are independent for all t. Therefore (vm ) for each N
are (N + 1)-dependent and can be subjected to a central limit theorem (see
Anderson, 1971, Theorem 7 .7 .5). Therefore
T 4
1 p 2)
a (5.4.15)
VT t-
Combining (5.4.12) and (5.4.15), we have
1-T( - p) N(0, 1 - p 2 ). (5.4.16)
Now we consider the MLE in AR(1) under the normality of (e t). The
likelihood function is given by
L = (270-T121111 -112 exp [- (1/2)y'ZT I y I, (5.4.17)
where ai l and IT I are given in (5.2.13) and (5.2.14), respectively. Therefore
log L - - log (2n) - - log (12 (5.4.18)

2 2
1 1
+ - log (1 - p2) - Q
2 2a2
where Q = (1 + p2 ) IT-111 P2(Yi + Yr) 2P g.2 Yi-IYI Putting

a log L/8cr 2 = 0, we obtain
(5.4.19)
Inserting (5.4.19) into (5.4.18) yields the concentrated log likelihood function
(aside from terms not depending on the unknown parameters)
T
log L* = log (2tr) log Q+ I log (1 p 2 ). (5.4.20)
2 2 2
Setting a log L*Iap = 0 results in a cubic equation in p with a unique real root
in the range [ 1, 1]. Beach and Mackinnon (1978) have reported a method of
deriving this root.
However, by setting OQ/ap = 0, we obtain a much simpler estimator
T-1 (5.4.21)
/
We call it the approximate MLE. Note that it is similar to the least square
estimator A, given in (5.4.2), for which the range of the summation in the
denominator is from t = 2 to T. If we denote the true MLE by hi, we can
easily show that VT (iSA - p) and 47(L m p) have the same limit distribution
by using the result of Section 4.2.5.
5.5 Prediction
In the classical regression model with nonstochastic regressors, the problem of
predicting a future value of the dependent variable and the problem of esti-
mating the regression parameters can be solved in a unified way. For example,
the least squares predictor is the best linear unbiased predictor, as shown in
Section 1.6. But, in time series models such as AR or MA, in which the
regressors are stochastic, the problem of optimal prediction is not a simple
corollary of the problem of optimal estimation because (1.6.3) no longer
holds. In view of this difficulty, what we usually do in practice is to obtain the
optimal predictor on the assumption that the parameters are known and then
insert the estimates obtained by the best known method, such as maximum
likelihood, mechanically into the formula for the optimal predictor.
There is an extensive literature on the theory concerning the optimal pre-

diction of a stationary time series when the parameters of the process are
assumed known (see Whittle, 1983, and the references in it).
In the simplest case of AR(1), the optimal predictor of y t+,, given
y yt_ i , . . is obtained as follows: From (5.2.2) we obtain
n-1
3'14-n = PnYt + (5.5.1)
J
Taking the conditional expectation of both sides given y y, 1 ,. . . , we

obtain
.9t+n ETYt+nlYt, Yi - 1, -) Pnyt (5.5.2)
In the mixed autoregressive, moving-average model (5.3.1), the optimal
predictor of y, +n given y y 1 ,. . . is obtained as follows: From (5.3.3) we
obtain
GO
Yt+n :11
4 1cEt+nk (5.5.3)
k-0
Putting et+ = et+_ i = . . . = et+i =0 in (5.5.3), we obtain the optimal pre-

dictor
OD
91+n = f)
( j+AtI (5.5.4)
Finally, putting _.; = cb- '(L)y1_i in (5.5.4), we obtain

OD
A+n = 0' (L) (fij+nliYt

.1"0
(See Whittle, 1983, p. 32.)
5.6 Distributed-Lag Models

5.6.1 The General Lag
If we add an exogenous variable to the right-hand side of Eq. (5.3.2), we obtain
p(L)y,= ax, + 13(L)e1 . (5.6.1)
Such a model is called a distributed-lag model by econometricians because
solving (5.6.1) for yt yields
Ye = al)-1(4xt+ p 1 (L)fl(L)e, (5.6.2)
where p- '(L) describes a distribution of the effects of the lagged values of x,

(that is, x x1_ 1 , xi_2 , . .) on y,. There is a vast amount of econometric
literature on distributed lag models. The interested reader is referred to the
-
aforementioned works by Fuller (1976); Nerlove, Grether, and Carvalho

(1979); and Dhrymes (1971). A model of the form (5.6.2) also arises as a
solution of the rational expectations model (see Shiller, 1978; Wallis, 1980).
We shall discuss briefly two of the most common types of distributed-lag
models: the geometric lag and the Almon lag.
5.6.2 The Geometric Lag

The geometric lag model is defined by
y, = + v, (5.6.3)
J -o
This model was first used in econometric applications by Koyck (1954)

(hence, it is sometimes referred to as the Koyck lag) and by Nerlove (1958).
Griliches's survey article (1967) contains a discussion of this model and other
types of distributed-lag models. This model has the desirable property of
having only two parameters, but it cannot deal with more general types of lag
distribution where the effect of x on y attains its peak after a certain number of
lags and then diminishes.
By inverting the lag operator Mr_ o AILI,we can write (5.6.3) equivalently as
Y1 = + ax, + u (5.6.4)
where u, = v, If (u,) are i.i.d., the estimation of ) and a is not much
more difficult than the estimation in the purely autoregressive model dis-
cussed in Section 5.4. Under general conditions the least squares estimator
can be shown to be consistent and asymptotically normal (see Crowder, 1980).
For a discussion of the estimation of A. and a when (v,), rather than (u,), are
i.i.d., see Amemiya and Fuller (1967) and Hatanaka (1974) (also see Section
6.3.7 for further discussion of this model).
5.6.3 The Almon Lag

Almon (1965) proposed a distributed-lag model
I
J-1
fipx.H-1-,+ v (5.6.5)
in which /J, . , /3N lie on the curve of a qth-order polynomial; that is,
)3i = 4 + oij+ 4./2 + . . . + (5,7j4, j= 1, 2, . . . , N. (5.6.6)
By defining vectors II= (13 i , 132, , flu) and 3 = 050 , 01 , . . . , de r, we
can write (5.6.6) in vector notation as
fl= Jo, (5.6.7)
where
[ 11 1
1 2 22 24
J= .1
. .
N N2 Ng
The estimation of 6 can be done by the least squares method. Let X be a
T X N matrix, the t, jth element of which is .;_ 1 _1 . Then
o = (J'X'XJ) 'J'X'y and fl = J(5. 3 Note that fl is a special case of the con-
strained least squares estimator (1.4.11) where R = J and c = 0.
By choosing N and q judiciously, a researcher can hope to attain both a
reasonably flexible distribution of lags and parsimony in the number of pa-
rameters to estimate. Amemiya and Morimune (1974) showed that a small
order of polynomials (q = 2 or 3) works surprisingly well for many economic
time series.
Some researchers prefer to constrain the value of the polynomial to be 0 at
j = N + 1. This amounts to imposing another equation
(50 + (51(N + 1) + (52(N + 1)2 + . . . + (54(N + 1)4 = 0 (5.6.8)
in addition to (5.6.6). Solving (5.6.8) for 0 0 and inserting it into the right-hand
side of (5.6.7) yields the vector equation
= J*3*, (5.6.9)
where 6* = (61 , 02 , . . . , (5)' and J* should be appropriately defined.
Exercises
1. (Section 5.2.1)
Prove that model (5.2.1) with Assumptions A, B, and C is equivalent to
model (5.2.3) with Assumptions A and B.
2. (Section 5.2.1)
Show that the process defined in the paragraph following Eq. (5.2.5) is
AR(1).
3. (Section 5.3)
Find the exact inverse of the variance-covariance matrix of MA(1) using
(5.3.12) and compare it with the variance-covariance matrix of AR(1).
4. (Section 5.3)
In the MA(1) process defined by (5.3.6), define = e ;11 riet_ I , where
(en are i.i.d. with Eer = 0, Ye ;11 = p20.2. Show that the autocovariances of
Yt and y* are the same.
5. (Section 5.4)
If X, Y, and Z are jointly normal and ifX is independent of either Yor Z,
then EXYZ -= EXEYZ (Anderson, 1958, p. 22). Show by a counterexam-
ple that the equality does not in general hold without the normality
assumption.
6. (Section 5.4)
Show that INA p) and "(M the same limit distribution.
7. (Section 5.4)
In the AR(1) process defined by (5.2.1), define the first differences
y, y and derive phm T .g.3 4_ 1 4/ Zr_ 3 421 .
8. (Section 5.4)

In the AR(2) process defined by (5.2.16), derive plim,, M1 2 Yr-IYII
cT
Y?-1.
9. (Section 5.5)
Derive (5.5.2) from the general formula of (5.5.5).
10. (Section 5.5)
In the MA(1) process defined by (5.3.6), obtain the optimal predictor of
Yri-n given y1, y,-1,
11. (Section 5.6)
Show that (5.6.7) can be written in the equivalent form Q'fi =0, where Q
is an N X (N q 1) matrix such that [Q, J] is nonsingular and
Q'J 0. Find such a Q when N = 4 and q= 2.
6 Generalized Least Squares Theory
One of the important assumptions of the standard regression model (Model 1)

is the assumption that the covariance matrix of the error terms is a scalar times
the identity matrix. In this chapter we shall relax this assumption. In Section
6.1 the case in which the covariance matrix is known will be considered. It is
pedagogically useful to consider this case before we consider, in Section 6.2, a
more realistic case of an unknown covariance matrix. Then in Sections 6.3
through 6.7 various ways of specifying the covariance matrix will be con-
sidered.
6.1 The Case of a Known Covariance Matrix

6.1 .1 Model 6
The linear regression model we shall consider in this chapter, called Model 6,
is defined by
y = X/1 + u, (6.1.1)
where X is a TX K matrix of known constants with rank K(. T), ft is a
K-vector of unknown parameters, and u is a T-vector of random variables with
Eu = 0 and Euu' = 1, a known T X Tpositive-definite covariance matrix. (In
Section 6. 1.5 we shall consider briefly the case of a singular covariance ma-
trix.) We shall write the t,sth element of 1 as cr,.
6.1 .2 Generalized Least Squares Estimator

Because X is positive definite, we can define 1 - 0 as HD'/2H', where H is an
orthogonal matrix consisting of the characteristic vectors of 1, D is the diago-
nal matrix consisting of the characteristic roots oft, and Jr' ' 12 is obtained by
taking the (-1)th power of every diagonal element of D (see Theorem 3 of
Appendix 1). Premultiplying (6.1.1) by 1-1r2 yields
y* = XV + u*, (6.1.2)
where y* = 1-1 /2y, X* = ii2 X, and u* = X-112u. Note that Eu* 0 and
Eu*u* ' = I, so (6.1.2) is Model 1 except that we do not assume the elements
of u* are i.i.d. here. The generalized least squares (GLS) estimator of fl in
model (6.1.1), denoted fiG , is defined as the least squares estimator of ft in
model (6.1.2); namely,
PG = (X*'X*rX*' y* (6.1.3)
=
Using the results of Section 1.2, we obtain that Eno .= ft and
VftG = ( X'X-1X)-1 . ( 6.1.4)
Furthermore, GLS is the best linear unbiased estimator of Model 6. (Note that
in Section 1.2.5 we did not require the independence of the error terms to
prove that LS is BLUE.)
6.1.3 Efficiency of Least Squares Estimator

It is easy to show that in Model 6 the least squares estimator ft is unbiased with
its covariance matrix given by
(rx)ix'vwxyl. (6.1.5)
Because GLS is BLUE, it follows that
(X'X)- IX'XX(X'X) -1 (6.1.6)
which can also be directly proved. There are cases where the equality holds in
(6.1.6), as shown in the following theorem.
THEOREM 6.1.1. Let X'X and X both be positive definite. Then the follow-
ing statements are equivalent.
(A) (X'X)- IX'XX(X'X) -1 =
(B) X = XB for some nonsingular B.
(C) =
(D) X = HA for some nonsingular A where the columns of H are K
characteristic vectors of X.
(E) X'IZ =0 for any Z such that Z'X = 0.
(F) X = + ZOZ' + a2I for some and 0 and Z such that Z' X = 0.
Proof We show that statement A statement B = statement C:
Generalized Least Squares Theory 183
Statement A X'IX = X'X(X'I- IX)-1 X'X

XT/2[I I-112X(X'I-1 X)- IX'I-119I112X = 0
11/2x _-_.- I- InXB for some B using Theorem 14 of Appen-
dix 1
IX = XB for some B
(X'X)-1X'IX = B
B is nonsingular because X'IX is nonsingular
statement B
X'I- IX = (B')- IX'X and X'I-1 = (B') - IX'
statement C.
Statement C = statement D can be easily proved using Theorem 16 of Ap-
pendix 1. (Anderson, 1971, P. 561, has given the proof.) Statement D
statement A and statement B = statement E are straightforward. To prove
statement E = statement B, note
statement E X'1,(Z, X) = (0, X'IX)
X'I = (0, X'IX)[Z(Z/Z) - ', X(X'X) -1]'
= X'IX(X'X)-1X' using Theorem 15 of Appendix 1
statement B because X'IX(X'X) -1 is nonsingular.
For a proof of the equivalence of statement F and the other five statements, see
Rao (1965).
There are situations in which LS is equal to GLS. For example, consider
y = X(/' + v) + u, (6.1.7)
where /I is a vector of unknown parameters and u and v are random variables
with Eu =0, Ev = 0, Euu' = a% Evv' = r, and Euv' =0. In this model,
statement F of Theorem 6.1.1 is satisfied because E(Xv + uXXv + u)' =
XFX' + a2I. Therefore LS = GLS in the estimation of ft
There are situations in which the conditions of Theorem 6.1.1 are asymp-
totically satisfied so that LS and GLS have the same asymptotic distribution.
Anderson (1971, p. 581) has presented two such examples:
Yt = fli /321 + /33t2 + + fixt K-1 + ut (6.1.8)
and
Yr =/1 cos A l t +/32 cos ,1.2t + . . . + Ar cos 'IV + lir, (6.1.9)

where in each case (u t) follow a general stationary process defined by (5.2.42).
[That is, take the y, of (5.2.42) as the present u,.]
We shall verify that condition B of Theorem 6.1.1 is approximately satisfied

for the polynomial regression (6.1.8) with K = 3 when (4) follow the station-
ary first-order autoregressive model (5.2.1). Define X = (x1 , x2 , x3), where the
tth elements of the vectors x 1 , x2 , and x3 are xlt = 1, x2, = t, and x31 = t2,
respectively. Then it is easy to show IX XA, where Z is given in (5.2.9) and
-2p
1 ( 1 P)2 0
A = [ 0 (1 - p)2 0 .
a
0 0 (1 - p)2
The approximate equality is exact except for the first and the last rows.
We have seen in the preceding discussion that in Model 6 LS is generaly not
efficient. Use of LS in Model 6 has another possible drawback: The covariance
matrix given in (6.1.5) may not be estimated consistently using the usual
formula 6 2(X'X)-1 . Under appropriate assumptions we have Om o2 =
Jim T-1 tr MI, where M = I - X(X 'X'. Therefore plim 6 -2(X'X)-1 is
generally different from (6.1.5). Furthermore, we cannot unequivocally de-
termine the direction of the bias. Consequently, the standard t and F tests of
linear hypotheses developed in Chapter 1 are no longer valid.
6.1.4 Consistency of the Least Squares Estimator

We shall obtain a useful set of conditions on X and Z for the consistency of LS
in Model 6. We shall use the following lemma in matrix analysis.
LEMMA 6.1.1. Let A and B be nonnegative definite matrices of size n. Then

tr (AB) (A) tr B, where Al (A) denotes the largest characteristic root of A.
Proof Let H be a matrix such that H'AH = D, diagonal, and H'H = I.
Then, tr (AB) = tr (H 'AHH'BH) = tr DQ, where Q = H'BH. Let di be the
ith diagonal element of D and q 1 be of Q. Then
tr (DQ) = dirk max di gal = Ai (A) tr B.
Now we can prove the following theorem.
THEOREM 6.1.2. In Model 6 assume

(A) MI) bounded for all t,
(B) ao.
Then plimT.ft =
Proof We have
tr vfi= tr [(X ' X)-1X ' IX(X ' X) -1]
tr [IX(X'X) -2X1
(I) tr [X(X'X) -2X' ] by Lemma 6.1.1
= A1 (1) tr (X' X) -1
K
,(X' X)
But the last term converges to 0 because of assumptions A and B.
Note that Theorem 3.5.1 is a special case of Theorem 6.1.2. One interesting
implication of Theorems 6.1.2 and 5.2.3 is that LS is consistent if u follows a
stationary time series satisfying (5.2.36).
We have not proved the asymptotic normality of LS or GLS in this section
because the proof would require additional specific assumptions about the
generation of u.
6.1.5 A Singular Covariance Matrix

If the covariance matrix E is singular, we obviously cannot define GLS by
(6.1.3). Suppose that the rank oft is S < T. Then, by Theorem 3 of Appendix
I, we can find an orthogonal matrix H = (H 1 , H2), where H I is TX S and H2
is TX (T S), such that WW 1 = A, a diagonal matrix consisting of the S
positive characteristic roots oft, Wit112 =0, and HAH2 =0. The premulti-
plication of (6.1.1) by H' yields two vector equations:
= H;Xii + (6.1.11)
and
H y = HXfl. (6.1.12)
Note that there is no error term in (6.1.12) because EH = HAH2 = 0
and therefore 1111 is identically equal to a zero vector. Then the best linear
unbiased estimator offi is GLS applied to (6.1.11) subject to linear constraints
(6.1.12). 1 Or, equivalently, it is LS applied to
A-1/21-Icy = A-1/2H1Xfi + A-1/2H'iu (6.1.13)
subject to the same constraints. Thus it can be calculated by appropriately
redefining the symbols X, Q, and c in formula (1.4.5).
8.2 The Case of an Unknown Covariance Matrix

In the remainder of this chapter, we shall consider Model 6 assuming that I is
unknown and therefore must be estimated. Suppose we somehow obtain an
estimator I. Then we define the feasible generalized least squares (FGLS)
estimator by
fi.F. = (X'i-1X)- IX'i- 'y, (6.2.1)
assuming I is nonsingular.
For fiF to be a reasonably good estimator, we should at least require it to be
consistent. This means that the number of the free parameters that character-
ize I should be either bounded or allowed to go to infinity at a slower rate than
T. Thus one must impose particular structure on I, specifying how it depends
on a set of free parameters that are fewer than T in number. In this section we
shall consider five types of models in succession. For each we shall impose a
particular structure on I and then study the properties of LS and FGLS and
other estimators of ft. We shall also discuss the estimation of I. The five
models we shall consider are (1) serial correlation, (2) seemingly unrelated
regression models, (3) heteroscedasticity, (4) error components models, and
(5) random coefficients models.
In each of the models mentioned in the preceding paragraph, i is obtained
from the least squares residuals fi -=. y Xfi, where ft is the LS estimator.

Under general conditions we shall show that ft is consistent in these models

and hence that I is consistent. Using this result, we shall show that flF has the
same asymptotic distribution as fiG .
In some situations we may wish to use (6.2.1) as an iterative procedure; that
is, given fiF we can calculate the new residuals y XfiF, reestimate I., and
insert it into the right-hand side of (6.2.1). The asymptotic distribution is
unchanged by iterating, but in certain cases (for example, if y is normal)
iterating will allow convergence to the maximum likelihood estimator (see
Oberhofer and Kmenta, 1974).
6.3 Serial Correlation

In this section we shall consider mainly Model 6 where u follows AR(1)
defined in (5.2.1). Most of our results hold also for more general stationary
processes, as will be indicated. The covariance matrix of u, denoted I I , is as
given in (5.2.9), and its inverse is given in (5.2.14).
6.3.1 Asymptotic Normality of the Least Squares Estimator

As was noted earlier, the first step in obtaining FGLS is calculating LS. There-
fore the properties of FGLS depend on the properties of LS. The least squares
estimator is consistent in this model if assumption B of Theorem 6.1.2 is
satisfied because in this case assumption A is satisfied because of Theorem
5.2.3. In fact, Theorem 5.2.3 states that assumption A is satisfied even when u
follows a more general process than AR(1). So in this subsection we shall prove
only the asymptotic normality of LS, and we shall do so for a process of u more
general than AR(1) but only for the case of one regressor (that is, K = 1) in
Model 6. (Anderson, 1971, p. 585, has given the proof in the case of K
regressors. He assumed an even more general process for u than the one we
assume in the following proof.)
THEOREM 6.3.1. Assume K = 1 in Model 6. Because X is a T-vector in this

case, we shall denote it by x and its t th element by x,. Assume
x'x
(A) lim T-*, = # O.
T
(B) -01il < where (,) are i.i.d. with Ee, = 0 and
M74)
EE? = a2.
Then IN (1) N(0, cT 2c2), where c2 = lfflT T- 'x'Euu'x.
Proof We need only prove T- ox'u N(0, c2) because then the theorem
follows from assumption A and Theorem 3.2.7 (iii) (Slutsky). We can write
T T
;li t = xt1 xt ctojet_j + ITI t-iT x, J-+1 (15fit-i (6.3.1)
tI
+A2 .
But V(A 2) -5 T- lx'xcr 2(Mj_N+110,1)2. Therefore A2 can be ignored if one takes

N large enough. We have
T-N N 1 4.),
Ai= - Et E t:t3iX t +i+ 2, e, Oix,+i (6.3.2)
VT (-1 .fro VT 1-1N
1 T-t
Nrt t- 7 -k+ 1
A n +A l2 +A 13 .
But V(A l2) T-1 a2Nivi2 Iy "114)./1)2, which goes to 0 as T' for a fixed N.
The same is true for A13. The theorem follows by noting that1.;',1_ o ckixti.;
t in Theorem 3.5.3. satifehcond rx
6.3.2 Estimation of p
Because I defined in (5.2.9) depends on a 2 only through a scalar multiplica-
tion, /30 defined in (6.1.3) does not depend on a 2. Therefore, in obtaining
FGLS (6.2.1), we need to estimate only p. The most natural estimator of p is
(6.3.3)
where fi, = y, xji. The consistency ofii is straightforward. We shall prove its
Using ut = pug_ + et , we have
T
A + Ai
(6.3.4)
where
r T
= 5, (p P'xt_ ix;(fi p + A (fl - fi) 'xt_,u, (6.3.5)
1T P
T
(ft fir
XtUt-I 1RP
t-2 VI' P.2
2p t ( ,
it fi)
t-
and
1T 2 T
A2 = T Y.Rfi
- rzi - IlYx1 B) 'xi-tur-i.
t-12 +TA 0 ,-1 (6.3.6)
If we assume that lim T_,. T-IVX is a finite nonsingular matrix, it is easy to

show that both AI and A2 converge to Oin probability. For this, we need only
the consistency of the LS estimator fl and IT(fl p = 0(1) but not the
asymptotic normality. Therefore, by repeated applications of Theorem 3.2.7

(Slutsky), we conclude that iTo-p) has the same limit distribution as
tit-iUt
NIT[
Pl
4-1
1 2
Hence the asymptotic normality of */) follows from the result of Section 5.4.
6.3.3 Feasible Generalized Least Squares Estimator

Inserting3 defined in (6.3.3) into IV defined in (5.2.14) and then inserting the
estimate of X -1 into (6.2.1), we obtain the FGLS estimator fiF. As noted
earlier, a2 drops out of the formula (6.2.1) and therefore need not be estimated
in the calculation of fl,. (However, a2 needs to be estimated in order to
estimate the covariance matrix of /3F .) The consistency of /3F can be proved
easily under general conditions, but the proof of the asymptotic normality is
rather involved. The reader is referred to Amemiya (1973a), where it has been
proved that fl, has the same asymptotic distribution as fiG undergeneral
assumptions on X when 04) follow an autoregressive process with moving-
average errors. More specifically,
VT(IF /1) /V10, lim (6.3.7)
In a finite sample, /IF may not be as efficient as fiG.

In fact, there is no
assurance that PF is better than the LS estimator A Harvey (19811, p. 19 )
presented a summary of several Monte Carlo studies comparing /I F with A
Taylor (1981) argued on the basis of analytic approximations of the moments
of the estimators that the relative efficiency offl F vis--vis ft crucially depends
on the process of the independent variables.
6.3.4 A Useful Transformation for the Calculation of the Feasible

Generalized Least Squares Estimator
Let ft, be the matrix obtained by inserting A into the right-hand side of
(5.2.10). Then we have by (5.2.14) fiF = (X'AIRI X)- iX'R;R ly. Thus /IF can
be obtained by least squares after transforming all the variables by R,. An-
other, somewhat simpler transformation is defined by eliminating the first
row of R,. This is called the Cochrane-Orcutt transformation (Cochrane and
Orcutt, 1949). The resulting estimator is slightly different from FGLS but has
the same asymptotic distribution.
If (u,) follow AR(2) defined by (5.2.16), the relevant transformation is given
by (5.2.27), where a's are determined by solving V(a l ui) = cr 2,
V(a 2u1 + a3u2)= a2, and E[a1 u1(a2u1 + a3u2)] = 0. A generalization of the
Cochrane-Orcutt transformation is obtained by eliminating the first two rows
of k. Higher-order autoregressive processes can be similarly handled.
6.3.5 Maximum Likelihood Estimator

Let us derive the MLE offi, p, and a2 in Model 6, assuming that (u,) are normal
and follow AR(1). The log likelihood function of the model has the same
appearance as (5.4.18) and is given by
T T 1 1
log L --- - - log 271. - - log cr2 + - log (1 - p 2) - Q (6.3.8)
2 2 2 2a2 '
where Q= ( 1 + P2)/1-1 14 - P2(1.4 + ui) - 2pn.2 u,_ 1 u, where we have
written u,= y, - xdi for short. Putting a log LA3a2 -- 0, we obtain
1
(6.3.9)
Inserting (6.3.9) into (6.3.8), we obtain the concentrated log likelihood func-
tion (aside from terms not depending on the unknown parameters)
T 1
log L* = - -f
- log Q + - log (1 - p2).
2
For a given value of fi, p may be estimated either by the true MLE, which
maximizes log L* using the method of Beach and MacKinnon (1978), or by
the approximate MLE, which minimizes log Q, as was noted earlier in Section
5.4. Both have the same asymptotic distribution. The formula for the approx-
imate MLE is
T
E ( y,_ 1 - x-ifixy,- x/i)
_ t. 2

(6.3.11)
P T- 1
(Yt-i - x;-i1)2
I
Given p, the value of /I that maximizes log L* is clearly
ft' = (X'r- IX)- `X'I- ly, (6.3.12)
where X is as given in (5.2.9). The approximate MLE of /3, p, and a' are
obtained by solving (6.3.9), (6.3.11), and (6.3.12) simultaneously.
These equations defining the approximate MLE are highly nonlinear, so
that the MLE must be obtained either by a search method or by an iteration.
We have already described the common iterative procedure.sof going back and
forth between (6.3.11) and (6.3.12), starting from the LS A
For the asymptotic normality of MLE in Model 6 where {u,) follow a
stationary autoregressive, moving-average process, see Pierce (1971).
6.3.6 Durbin-Watson Test

In this subsection we shall consider the test of the hypothesis p = 0 in Model 6
with (u,) following AR(1). The Durbin-Watson test statistic, proposed by
Durbin and Watson (1950, 1951), is defined by
d' (6.3.13)
,-1
where (01) are the least squares residuals. By comparing (6.3.13) with (6.3.3),
we can show that
d = 2 219 + o(T- '). (6.3.14)
From this we know that plim d = 2 2p, and the asymptotic distribution of d
follows easily from the asymptotic distribution of P. derived in Section 6.3.2.
To derive the exact distribution of d under the null hypothesis, it is useful to
rewrite (6.3.13) as
u'MAMu
d (6.3.15)
u'Mu
where M = I X(X'X) IX' and A is a TX T symmetric matrix defined by
L1 1 0 01.
1 2 1 0 0
A= : (6.3.16)
0 0 1 2 1
0 0 1 1
Because MAM commutes with M, there exists an orthogonal matrix H that

diagonalizes both (see Theorem 8 of Appendix 1). Let v i , v2 ,. . . , vT-K be the
nonzero characteristic roots of MAM. Then we can rewrite (6.3.15) as
T-K
VI
d T-K ' (6.3.17)
0
where (Ci) are i.i.d. N(0, 1). In the form of (6.3.17), the density of d can be
represented as a definite integral (Plackett, 1960, p. 26). The significance
points of d can be computed either by the direct evaluation of the integral
(Imhof, 1961) or by approximating the density on the basis of the first few
moments. Durbin and Watson (1950, 1951) chose the latter method and
suggested two variants: the Beta approximation based on the first two mo-
ments and the Jacobi approximation based on the first four moments. For a
good discussion of many other methods for evaluating the significance points
of d, see Durbin and Watson (1971).
The problem of finding the moments of d is simplified because of the
independence of d and the denominator of d (see Theorem 7 of Appendix 2),
which implies
T-K
E va
E(ds) E
177 1 K
(
(6.3.18)
E Cl
for any positive integer s. Thus, for example,

1 T-K
Ed y=7 K v, (6.3.19)
and
2 TY (1 , Ed) 2
Vd (6.3.20)
(TKXTK+ 2)
where the term involving vi can be evaluated using
T-K
vf = tr (MAY. (6.3.21)
ri
The Beta approximation assumes that x = d/4 has a density given by

1
x'(1 0 x 1; 0 otherwise, (6.3.22)
B(p, q)
so that Ed = 4p/(p + q) and Vd = 16pql[(p + q) 2(p + q + 1)1 The Jacobi

approximation assumes that x = d/4 has a density given by
xg- i(1 x)1 -q ,
L -r a3k.r3(x) -r a4u4kxjj, (6.3.23)
B(q, p q + 1)
where G3 and G4 are the third- and fourth-order Jacobi polynomials as defined
by Abramowitz and Segun (1965, Definition 22.2.2). The parameters p and q
are determined so that the first two moments of (6.3.23) are equated to those
of d, and a3 and a4 are determined so that the third and fourth moments
match. In either approximation method the significance points can be com-
puted by a straightforward computer program for integration.
The distribution of d depends upon X. That means investigators must
calculate the significance points for each of their problems by any of the
approximation methods discussed in the previous section. Such computa-
tions are often expensive. To make such a computation unnecessary in certain
cases, Durbin and Watson (1950, 1951) obtained upper and lower bounds of d
that do not depend on X and tabulated the significance points for these. The
bounds for d are given by
T-K T-K
AtC? ill+K-IC1
I
dT i-1
T-K and du (6.3.24)
2 Cl
P-1 C I
where Ai -=. 2(1 cos inT-1) are the positive characteristic roots of A. (The
remaining characteristic root of A is 0.) These bounds were obtained
under the assumption that X contains a vector of ones and that no other
column of X is a characteristic vector of A.' Durbin and Watson (1951)
calculated the significance points (at the 1, 2.5, and 5% levels) of di. and du for
K= 1, 2, . . . , 6 and T= 15, 16, . . . , 100 by the Jacobi approximation.
Let dut and du, be the critical values of these bounds at the a% significance
level. Then test 1/13 : p = 0 against HI : p> 0 by the procedure:
Reject lic, if d (6.3.25)
Accept 1/0 if d du,.
If du. < d < du., the test is inconclusive. To test 1/ 0 : p = 0 against HI : p < 0,
Durbin and Watson suggested the following procedure:
Reject 1/0 if d 4d , (6.3.26)
Accept 1/0 if d 4 dua

If these tests are inconclusive, the significance points of (6.3.17) must be
evaluated directly.
6.3.7 Joint Presence of Lagged Endogenous Variables and

Serial Correlation
In this last subsection we shall depart from Model 6 and consider briefly a
problem that arises when X in (6.1.1) contains lagged values of y. An example
of such a model is the geometric distributed-lag model (5.6.4), where the errors
are serially correlated. This is an important problem because this situation
often occurs in practice and many of the results that have been obtained up to
this point in Section 6.3 are invalidated in the presence of lagged endogenous
variables.
Consider model (5.6.4) and suppose (4) follow AR(1), rit = pu,_ +
where (c) are i.i.d. with Ee, = 0 and Vet = a2. The LS estimators fit and a are
clearly inconsistent because plim '. 1y_ u, # 0. The GLS estimators
of /land a based on the true value ofp possess not only consistency but also all
the good asymptotic properties because the transformation R I defined in
(5.2.10) essentially reduces the model to the one with independent errors.
However, it is interesting to note that FGLS, although still consistent, does not
have the same asymptotic distribution as GLS, as we shall show.
Write (5.6.4) in vector notation as
y=4_ 1 -1-ax+u - Zy+u, (6.3.27)
where y- 1 = (Yo, Yi,. . . Yr i)'. Suppose, in defining FGLS, we use a con-

-
sistent estimator of p, denoted A, such that NIT( p) converges to a non-

degenerate normal variable. Then the FGLS estimator of 7, denoted is
given by
9F = ( 6.3.28)
where ft i is derived from R1 by replacing p with The asymptotic distribution
of PF can be derived from the following equations:
Nrta.- y) = (7-I V fri ftiZ)-1 T-1/2Z/ftlft iu (6.3.29)

IR(plim 7-1Z/R1R1Z)-1
X [7-1/2 Z/Rcit 1u + (T1/2Z/a u
- 7-1/2Z/R;R 1u)].
Let Hir be the first element of the two-dimensional vector (7 -1/2Z/ftlft iu -

7-1/2Z' Wi lliu). Then we have
1 T-2
WT = 11: ( 32 P2) 4; II YtUt+ I (6.3.30)

T t-1
E-2
Ytut + T Ytut+2)
t-o
LD a2
Ap _ 1 Nriv - p).
Thus we conclude that the asymptotic distribution of FGLS differs from that
n
of GLS and depends on the asymptotic distribution of (see Amemiya and
Fuller, 1967, for further discussion). Amemiya and Fuller showed how to
obtain an asymptotically efficient feasible estimator. Such an estimator is not
as efficient as GIS.
The theory of the test of independence discussed in the preceding subsec-
tion must also be modified under the present model. If X contained lagged
dependent variables, YI-1, Yl - 2 t 9 We Will still have Eq. (6.3.17) formally,
but (0 will no longer be independent normal because H will be a random
matrix correlated with u. Therefore the Durbin-Watson bounds will no longer
be valid.
Even the asymptotic distribution of d under the null hypothesis of indepen-
dence is different in this case from the case in which Xis purely nonstochastic.
The asymptotic distribution of d is determined by the asymptotic distribution
A
of because of (6.3.14). When X is nonstochastic, we have IN - p)--.
N(0, 1 - p 2) by the results of Section 6.3.2. But, if X contains the lagged
dependent variables, the asymptotic distribution of A will be different. This
can be seen by looking at the formula for It(16 - p) in Eq. (6.3.4). The third
term, for example, of the right-hand side of (6.3.5), which is
T -1121 72(fi - fl)'x04-1 does not converge to 0 in probability because x t and
ut_ 1 are correlated. Therefore the conclusion obtained there does not hold.
We consider the asymptotic distribution of fi in the simplest case

Yi = aYt-i + 14, tit = Put- 1 + Et, (6.3.31)
where lal, Ipl < 1, f,) are i.i.d. with EE, = 0 and Eq = a2, and both (y,) and
(u,) are stationary. Define
T
fit--1 0,
(6.3.32)
,-2
where 12, = y, y,_ 1 and Oi is the least squares estimator. Consider the limit
distribution of 17f; under the assumption p = 0. Because the denominator
times T'' converges to a2 in probability, we have asymptotically
ID 1 1T
NIT ;6=712 777Z k_la t (6.3.33)
LD 1 1 T ,
-
-- 1 1.141 _ lilt - (1 - a2)yt-1 141
Therefore the asymptotic variance (denoted AV) of ft 4 is given by
1 4-, V { i [u,_ 1 u (1 o?), ye_ i u,] 1

AV(Nrtfi) = -07 (6.3.34)
2
-11 [E (i )4" (1 - a2
Ut_ilig )2 E (i Yt-1 14)2
04 T t 2 i2
T T
2(1 &)E(E 14,_ 1 u, E y,_ , u,) ]
t-2 t-2
=a2(1 7'-1).
Hence, assuming that the asymptotic normality holds, we have
In- N(0, a2). (6.3.35)
Durbin (1970) obtained the following more general result: Even if higher-
order lagged values of y, and purely exogenous variables are contained among
the regressors, we have under the assumption p = 0
MO, 1 AV(Tiii)], (6.3.36)
where eil is the least squares estimate of the coefficient on 34_, . He proposed
that the test of the independence be based on the asymptotic normality

above.3
6.4 Seemingly Unrelated Regression Model

The seemingly unrelated regression (SUR) model proposed by Miner (1962)
consists of the following N regression equations, each of which satisfies the
assumptions of the standard regression model (Model 1):
ya = Xaila + ua , i = 1, 2, . . . , N, (6.4.1)
where ya and U. are T-vectors, X a is a TX IC, matrix, and ila is a Ka -vector. Let tia,
be the tth element of the vector u,. Then we assume that (u 1 . . . , um) is
an i.i.d. random vector with Euu = 0 and Coy (uu , up) = au . Defining y
YZ, II= (131,A, . u= . , u;X, and
X = diag (XI , X2, . . XN), we can write (6.4.1) as
y = Xfi + u. (6.4.2)
This is clearly a special case of Model 6, where the covariance matrix of u is
given by
Euu' = IT, (6.4.3)
where X = (au) and 0 denotes the Kronecker product (see Theorem 22 of
Appendix 1).
This model is useful not only for its own sake but also because it reduces to a
certain kind of heteroscedastic model if I is diagonal (a model that will be
discussed in Section 6.5) and because it can be shown to be equivalent to a
certain error components model (which will be discussed in Section 6.6).
The GLS estimator offi is defined by /1 0 = ('Q )'Q 'y. Because of
(6.4.3) we have trI = x-1 I, using Theorem 22 (i) of Appendix 1. In the
special case in which X 1 = X2= . . . XN, we can show flo = /3 as follows:
Denoting the common value of Xi by X and using Theorem 22(i) of Appendix
1 repeatedly, we have
/70 = P_M-1 Ihr (6.4.4)
= [(I 0 X')(X-1 0 IXI 0 X)] -1(I 0 X')(I-1 0 IV
0 0 X')y
= [I 0 (X'X)-1Xly
= 11)-IX/ y.
The same result can be also obtained by using Theorem 6.1.1. Statement E is
especially easy to verify for the present problem.
To define FGLS, we must first estimate X. A natural consistent estimator of
its i,jth element is provided by du - = T-Ifi;fii , where Ili y, Xifli are the
least squares residuals from the ith equation. These estimates are clearly
consistent as T goes to on while N is fixed. Because of the special form of ir
given in the preceding paragraph, it is quite straightforward to prove that
FGLS and GLS have the same asymptotic distribution (as T co) under
general assumptions on u and X. The limit distribution of 17-"(fiG - 17) or
rt(fi F 5) is N[0, aim T -1X11-1X) -1 ]. Suitable assumptions on u and X
can easily be inferred from Theorem 3.5.4.
FGLS is generally unbiased provided that it possesses a mean, as proved in a
simple, elegant theorem by Kakwani (1967). The exact covariance matrix of
FGLS in simple situations has been obtained by several authors and compared
with that of GLS or LS (surveyed by Srivastava and Dwivedi, 1979). A particu-
larly interesting result is attributable to Kariya (1981), who obtained the
following inequalities concerning the covariance matrices of GLS and FGLS
in a two-equation model with normal errors:
lifiG VfiF [ 1 + 2 VftG (6.4.5)

Tr 3
where r = rank[X, , X2].
6.5 Heteroscedasticity
A heteroscedastic regression model is Model 6 where is a diagonal matrix,
the diagonal elements of which assume at least two different values. Hetero-
scedasticity is a common occurrence in econometric applications and can
often be detected by plotting the least squares residuals against time, the
dependent variable, the regression mean, or any other linear combination of
the independent variables. For example, Prais and Houthakker (1955) found
that a variability of the residuals from a regression of food expenditure on
income increases with income. In the subsequent subsections we shall con-
sider various ways to parameterize the heteroscedasticity. We shall consider
the estimation of the heteroscedasticity parameters as well as the regression
coefficients. We shall also discuss tests for heteroscedasticity.
6.5.1 Unrestricted Heteroscedasticity

When heteroscedasticity is unrestricted, the heteroscedasticity is not parame-
terized. So we shall treat each of the T variances {o1) as an unknown parame-
ter. Clearly, we cannot consistently estimate these variances because we have

but one observation per variance. Nevertheless, it is possible to estimate the
regression coefficients p consistently and more efficiently than by LS and to
test for heteroscedasticity.
The idea of obtaining estimates of the regression coefficients that are more
efficient than LS in an unrestricted heteroscedastic regression model has been
developed independently by White (1982) and by Chamberlain (1982), fol-
lowing the idea of Eicker (1963), who suggested an estimator of V/1(6.1.5) that
does not require the consistent estimation of I. The following discussion
follows Amemiya (1983c).
Let W be a TX (T K) matrix of constants such that [X, W] is nonsingu-
lar and W'X = 0. Then, premultiplying (6.1.1) by [X, W]', GLS estimation of
/1 can be interpreted as GLS applied simultaneously to
X'y = X'Xfl X'u (6.5.1)
and
W'y = W'u. (6.5.2)
Thus, applying Theorem 13 of Appendix 1 to
r X'IX X'IW 1 -1
LW'IX W'INV J '
we obtain
PG = ( XX)-1X/IW(W1M-IW'y. (6.5.3)
Of course, it is also possible to derive (6.5.3) directly from (6.1.3) without
regard to the interpretation given above. An advantage of (6.5.3) over (6.1.3) is
that the former does not depend on 1-1. Note that one cannot estimate
T-1X'I- IX consistently unless I can be consistently estimated. To trans-
form (6.5.3) into a feasible estimator, one is tempted to replace by a
diagonal matrix D whose tth element is (y, x41) 2. Then it is easy to prove
that under general assumptions plim 'X'DW = plim T - IX'IW and
plim T-1 W'DW = plim 1 W'IW element by element. However, one dif-
ficulty remains: Because the size of these matrices increases with T, the result-
ing feasible estimator is not asymptotically equivalent to GLS.
We can solve this problem partially by replacing (6.5.2) with
Wy = Wu, (6.5.4)
where W 1 consists of N columns of W, N being a fixed number. When GIS is
applied to (6.5.1) and (6.5.4), it is called the partially generalized least squares
(PGLS) and the estimator is given by
fip = (X1 X)-1 X'XWI(WcIWI)-1 Wcy. (6.5.5)
PGLS is more efficient than LS because
111 Vfl = (X 'X) - IX '1W i(W '1 /W, )- 1W 'i 2X(X ' X) - I, (6.5.6)
which is clearly nonnegative definite, but itis less efficient than GLS. An
asymptotically equivalent feasible version offip is obtained by replacing the I
in (6.5.5) by the D defined in the preceding paragraph.
White (1980a) proposed testing the hypothesis al = a 2 for all t by compar-
ing (X'X) -1 X'DX(X'X) -1 with 6 2(X'X)- 1, where D is as defined earlier and
el 2 is the least squares estimator of 0 2 defined in (1.2.5). Equivalently, White
considered elements of X'DX 62X'X. If we stack the elements of the upper
triangular part of this matrix, we obtain a vector of (K 2 + K)/2 dimension
defined by S'(fi2 _ i) where 0 2 is a T-vector the tth element of which is
, 1 is a T-vector of ones, and S is a T X (K2 + K)/2 matrix, the columns of
which are (xiixii , xi2xj2 , . . , xirxir)' for 1 4
. j K, and i j. It is easy
to show that T- u2S'(112 621) N(0, A), where A = lim (7`'S'AS +
T-21' Al S'll'S T -2S' Al l'S T-2S1 1 l'AS) and A = E(u 2 a21)
(u2 a21)'. The test statistic proposed by White is
_ a201s(71)-1svi2 _ 6.209
where A is obtained from A by eliminating lim and replacing A with

D((u 6 2)2). This statistic is asymptotically distributed as chi-square with
(K2 + K)/2 degrees of freedom under the null hypothesis. It can be simply
computed as TR 2 from the regression of Q on the products and cross products
of xt .
6.5.2 Constant Variance in a Subset of the Sample

The heteroscedastic model we shall consider in this subsection represents the
simplest way to restrict the number of estimable parameters to a finite and
manageable size. We assume that the error vector u is partitioned into N
nonoverlapping subvectors as u = (uc, ui,. . . , ul,,)' such that Eu iu;
We can estimate each al consistently from the least squares residuals
provided that each Ti goes to infinity with T. Note that this model is a special
case of Zellner's SUR model discussed in Section 6.4, so that the asymptotic
results given there also hold here. We shall consider a test of homoscedasticity
and the exact moments of FGIS.
If we assume the normality of u, the hypothesis al = a2 for all i can be tested

by the likelihood ratio test in a straightforward manner. Partitioning y and X
into N subsets that conform to the partition of u, we can write the constrained
and unconstrained log likelihood functions respectively as
N
CLL = log al (yi xifi)'(yi Xiff) (6.5.7)
2 2o2
and
N N
ULL Ti log al _a (yi (y; Ur). (6.5.8)
Therefore 2 times the log likelihood ratio is given by
2 log LRT = Ti log (ii2/61), (6.5.9)
where 6.2 and al are the constrained and unconstrained MLE, respectively.
The statistic is asymptotically distributed as chi-square with N 1 degrees of
freedom.4
Taylor (1978) has considered a special case in which N = 2 in a model with
normal errors and has derived the formulae for the moments of FGLS. By
evaluating the covariance matrix of FGLS at various parameter values, Taylor
has shown that FGLS is usually far more efficient than LS and is only slightly
less efficient than GLS.
We shall sketch briefly the derivation of the moments of FGLS. Let C be a
KX K matrix such that
= (6.5.10)
and
C'XPL2C = crA (6.5.11)
where A is a diagonal matrix, the elements A,, A',. . . AK of which are the
roots of the equation
lai2XP2 AaT2X1X, I = 0. (6.5.12)
The existence of such a matrix is guaranteed by Theorem 16 of Appendix 1.
With S = C(I + A)-'a, transform the original equation y = Xfl + u to
y = X*y + u, (6.5.13)
where X* = XS and y = S- The FGLS estimator of y, denoted f, is given by
)7 = iii2Xorrn-105-1-23Cr yi ity2x*
2 IY2) , (6.5.14)
where ill = (7; K) -1(y1 Xiff)'(y, X1/1). 5
Using aT2X 11" X * = (I + A) -1 and cri2X = A(I + A)-1, we obtain
y = (I + A)[ar 2DXrui + cri2A-1(I D)Xru2], (6.5.15)
where D is a diagonal matrix the ith diagonal element of which is equal to
+ ,VriErf). Finally, the moments of f 7 can be obtained by mak-
ing use of the independence of D with X 1" u i and X 32" u2 and of known
formulae for the moments of D that involve a hypergeometric function.
Kariya (1981) has derived the following inequalities concerning the covar-
iance matrices of GLS and FGLS in a two-equation model:
VnG VfiF 5 [1 + 2( T
-
1
K
1
2) + 2( T2 K 2)
] V/IG . (6.5.16)
6.5.3 General Parametric Heteroscedasticity

In this subsection we shall assume cr? = ga 111 ) without specifying g, where /31
,
a is anotherisaubet(polywh)fergsionpamt/Id
vector of parameters unrelated to /1. In applications it is often assumed that
gt(a, 110= g(a, x;//). The estimation of a and /I can be done in several steps.
In the first step, we obtain the LS estimator of/J, denoted /3. In the second step,
a and can be estimated by minimizing IT., [in ge(a, fl,)12, where a, =
y, )4/I. The consistency and the asymptotic normality of the resulting esti-
mators, denoted a and , have been proved by Jobson and Fuller (1980). In
the third step we have two main options: FGLS using gAii, /31) or MLE under
normality using it and /3 as the initial estimates in some iterative algorithm. 6
CarolndRupet(1982a)rovdhunegalsmptioFGLS
has the same asymptotic distribution as GLS. Jobson and Fuller derived
simple formulae for the method of scoring and proved that the estimator of /I
obtained at the second iteration is asymptotically efficient (and is asymptoti-
cally more efficient than GLS or FGLS). Carroll and Ruppert (1982b) have
pointed out, however, that GLS or FGLS is more robust against a misspecifi-
cation in the g function. Carroll and Ruppert (1982a) have proposed a robust
version of FGLS (cf. Section 2.3) in an attempt to make it robust also against
nonnormality.
There are situations where FGLS has the same asymptotic distribution as
MLE. Amemiya (1973b), whose major goal lay in another area, compared the
asymptotic efficiency of FGLS vis--vis MLE in cases where y, has mean xdJ
and variance p/2(x:fl)2 and follows one of the three distributions: (1) normal,
(2) lognormal, and (3) gamma. This is the form of heteroscedasticity suggested
by Prais and Houthakker (1955). It was shown that FGLS is asymptotically
efficient if y, has a gamma distribution. Thus this is an example of the BAN
estimators mentioned in Section 4.2.4.
This last result is a special case of the following more general result attribut-
able to Nelder and Wedderburn (1972). Let ( y 1) be independently distributed
with the density
exP (aW[g( 01)Y, h(01)+ k(A, 301), (6.5.17)
where a and 0, q(x Aare scalars and A, x and /I are vectors. It is assumed
that h'(0,)= 0,g'(0,) for every t, which implies that Ey, = 0, and Vy, =
[a(A)g'(0,)] - '.Then, in the estimation of ft, the method of scoring is identical
to the Gauss-Newton nonlinear weighted least squares iteration (cf. Section
4.4.3). The binomial distribution is a special case of (6.5.17). To see this,
define the binary variable y, that takes unity with probability 0, and zero with
probability 1 0, and put a(A) = 1, g(0,)= log 0, log (1 0,), h(0,)
log (1 0,), and k(A, y,) = 0. The normal distribution is also a special case:
Take y, N(0,, A2), a(2) = A-2, g(0,)= 0,, h(0,) = 0/2, and k(A, y,) =
2-1A2 log (27A2) y?12. The special case of Amemiya (1973b) is obtained
by putting a(A) = A, g(0,)= h(0,)= log 0 and k(A, y,) =
log roo log A + A-1 (1 A) log y,.
There are series of articles that discuss tests of homoscedasticity against
heteroscedasticity of the form ol = g(a, x;i1), where a is a scalar such that
g(0, x;f1) does not depend on t. Thus the null hypothesis is stated as a = 0.
Anscombe (1961) was the first to propose such a test, which was later modified
by Bickel (1978). Hammerstrom (1981) proved that Bickel's test is locally
uniformly most powerful under normality.
6.5.4 Variance as a Linear Function of Regressors

In this subsection we shall consider the model al = 4a, where z, is a G-vector
of known constants and a is a G-vector of unknown parameters unrelated to
the regression coefficients it Such a parametric heteroscedastic regression
model has been frequently discussed in the literature. This is not surprising,
for this specification is as simple as a linear regression model and is more
general than it appears because the elements of; may be nonlinear transfor-
mations of independent variables. The elements of; can also be related to x,
without affecting the results obtained in the following discussion. We shall
consider the estimation of a and the test of the hypothesis al = a 2 under this
model.
Hildreth and Houck (1968) and Goldfeld and Quandt (1972) were among
the first to study the estimation of a. We shall call these estimators HH and
GQ for short. We shall follow the discussion of Amemiya (1977b), who pro-
posed an estimator of a asymptotically more efficient under normality than
the HH and GQ estimators.
Hildreth and Houck presented their model as a random coefficients model
(cf. Section 6.7) defined by
yi --= 34(fi a (6.5.18)
where (,) are K-dimensional i.i.d. vectors with 4 =0 and EZ,e, = Wad.
The last matrix denotes aKXK diagonal matrix, the tth element of which is
a,. Thus the Hildreth and Houck model is shown to be a special case of the
model where the variance is a linear function of regressors by putting z, =
(xi, xi, , . . . , xlre)'.
We shall compare the HH, GQ, and Amemiya estimators under the as-
sumption of the normality of u. All three estimators are derived from a
regression model in which 14 serves as the dependent variable. Noting t2, =
u, x(X' X) -1X'u, we can write
Ill = z:a + vi, 2v2, + v3 (6.5.19)
where v1, = ul al, v2, = ti,x;(X'X) -1X'u, and v3, = [x;(X'X) -'X'uj2. We
02 = la + v 1 2v2 + v3 . (6.5.20)
We assume that X fulfills the assumptions of Theorem 3.5.4 and that lim r_..
T-1EZ is a nonsingular finite matrix. (Amemiya, 1977b, has presented more
specific assumptions.)
Equation (6.5.20) is not strictly a regression equation because Ev 2 #0 and
Ev3 # 0. However, they can be ignored asymptotically because they are
0( T-1/2) and O( T-1 ), respectively. Therefore the asymptotic properties of LS,
GLS, and FGLS derived from (6.5.20) can be analyzed as if v, were the only
error term. (Of course, this statement must be rigorously proved, as has been
done by Amemiya, 1977b.)
The GQ estimator, denoted a, , is LS applied to (6.5.20). Thus
ex, = (Z'Z)- 'Z'fi2. (6.5.21)
It can be shown that It(& l a) has the same limit distribution as
../T(Z'Z)-1Z'v1 . Consequently, & is consistent and asymptotically normal

with the asymptotic covariance matrix given by
V1 = 2(Z'Z)-1 Z'D2Z(VZ)-1 , (6.5.22)
where D---= Euu' = D(4a).
The HH estimator 6/2 is LS applied to a regression equation that defines
:
62 147.01 + TV, (6.5.23)

where lq is obtained by squaring every element of M = I X(X'X) - 'X' and
Ew = O. Thus
62 (6.5.24)
It can be shown that
Eww' = 2Q, (6.5.25)
where Q = MDM * MDM. The symbol * denotes the Hadamard product
defined in Theorem 23 of Appendix 1. (Thus we could have written M as
M * M.) Therefore the exact covariance matrix of ex 2 is given by
V2 = 2(Z'1414Z) -I Z'li4Q14Z(Z'1414Z) -l .
'2 (6.5.26)
To compare the GQ and the HH estimators we shall consider a modifica-
tion of the HH estimator, called MHH, which is defined as
= (Z'MMZ) -1Z'141402. (6.5.27)
Because v2 and v3 in (6.5.20) are of 0(T-1/2) and 0(T-1), respectively, this
estimator is consistent and asymptotically normal with the asymptotic covar-
iance matrix given by
= 2(VMMZ)-1 Z'AMD2MMZ(Z'AMZ) -1 . (6.5.28)
Amemiya (1978a) showed that
Q 14D21iI, (6.5.29)
or that Q 14D214 is nonnegative definite, by showing A'A * A'A
(A * A)'(A * A) for any square matrix A (take A = DoM). The latter in-
equality follows from Theorem 23 (i) of Appendix 1. Therefore V2
thereby implying that MHH is asymptotically more efficient than HH.
Now, from (6.5.27) we note . that MHH is a "wrong" GLS applied to
(6.5.20), assuming Ev ivl were (MM) -1 when, in fact, it is 2132. Therefore we
cannot make a definite comparison between GQ and MHH, nor between GQ

and HH. However, this consideration does suggest an estimator that is
asymptotically more efficient than either GQ or MHH, namely, FGL,S ap-
plied to Eq. (6.5.20),
63 .= (vfic2z)-ivor2ip, (6.5.30)
where 1e), = Amemiya (1977b) showed that the estimator is con-
sistent and asymptotically normal with the asymptotic covariance matrix
given by
V3 = 2(Z' D-2Z)-1 . (6.5.31)
Amemiya (1977) has also shown that the estimator has the same asymptotic
distribution as MLE. 7
It should be noted that all four estimators defined earlier are consistent even
if u is not normal, provided that the fourth moment of u, is finite. All the
formulae for the asymptotic covariance matrices in this subsection have been
obtained under the normality assumption, and, therefore, the ranking of
estimators is not preserved in the absence of normality.
So far we have concentrated on the estimation of a. Given any one of the
estimators of a defined in this subsection, say 6/3 , we can estimate fl by FGLS
= (X' 61 1X)- 'X toi iy. (6.5.32)
We can iterate back and forth between (6.5.30) and (6.5.32). That is, in the
next round of the iteration, we can redefine II as y X/3F and obtain a new
estimate of 6/3, which in turn can be used to redefine D 3 , and so on. Buse
(1982) showed that this iteration is equivalent under normality to the method
of scoring. Goldfeld and Quandt (1972) presented a Monte Carlo study of the
performance of LS, FGLS, and MLE in the estimation (4 in the model where
Yt fit + /32x, + u, and Vu, = al + a2x, 0/3.4.
We shall conclude this subsection with a brief account of tests of homosce-
dasticity in the model where the variance is a linear function of regressors.
Assuming that the first column of Z is a vector of ones, the null hypothesis of
homoscedasticity can be set up as 143 : a = (a 2, 0, 0, . . . , O)' in this model.
Breusch and Pagan (1979) derived Rao's score test (cf. Section 4.5.1) of Ho as
1 (11
Rao = 2 ii21)'Z(Z'Z)- 'Z'(1 2 (6.5.33)
2cr4
where! is the T-vector of ones. It is asymptotically distributed asx_ i . Because
the asymptotic distribution may not be accurate in small sample, Breusch and
Pagan suggested estimating P(Rao > c) by simulation. They pointed out that
since Rao's score test depends only on 11 the simulation is relatively simple.
,
For a simple way to make (6.5.33) robust to nonnormality, see Koenker

(1981b).
Goldfeld and Quandt (1965) proposed the following nonparametric test of
homoscedasticity, which they called the peak test: First, order the residuals fi,
in the order of z;ii for a given estimator et. This defines a sequence (ko) where
j k if and only if z (Instead of 461, we can use any variable that we
suspect influences al most significantly.) Second, define that a peak occurs at
j > 1 if and only if I i1,1> IakI for all k <j. Third, if the number of peaks exceeds
the critical value, reject homoscedasticity. Gokifeld and Quandt (1965, 1972)
presented tables of critical values.
6.5.5 Variance as an Exponential Function of Regressors

As we mentioned before, the linear specification, however simple, is more
general than it appears. However, a researcher may explicitly specify the
variance to be a certain nonlinear function of the regressors. The most natural
choice is an exponential function because, unlike a linear specification, it has
the attractive feature of ensuring positive variances. Harvey (1976), who as-
sumed y, exp (ea)], proposed estimating a by LS applied to the
regression of log 12? on z, and showed that the estimator is consistent if 1.2704
is subtracted from the estimate of the constant term. Furthermore, the esti-
mator has an asymptotic covariance matrix equal to 4.9348(Z'Z) ' , more
than double the asymptotic covariance matrix of MLE, which is 2(Z'Z) - '.
6.6 Error Components Models

Error components models are frequently used by econometricians in analyz-
ing panel dataobservations from a cross-section of the population (such as
consumers, firms, and states, and hereafter referred to as individuals) at var-
ious points in time. Depending on whether the error term consists of three
components (a cross-section-specific component, a time-specific component,
and a component that depends both on individuals and on time) or two
components (excluding the time-specific component), these models are called
three error components models (3ECM) and two error components models
(2ECM), respectively. Two error components models are more frequently
used in econometrics than 3ECM because it is usually more important to
include a cross-section-specific component than a time-specific component.
For example, in studies explaining earnings, to which error components

models have often been applied, the cross-section-specific component may be
regarded as the error term arising from permanent earnings and is more
important than the time-specific component.
We shall discuss first 3ECM and then various generalizations of 2ECM.
These generalizations are (1) the Balestra-Nerlove model (2ECM with lagged
endogenous variables), (2) 2ECM with a serially correlated error, and (3)
2ECM with endogenous regressors.
6.6.1 Three Error Components Models

Three error components models are defined by
y,,= xj1 + u, = 1, 2, . . . , N, (6.6.1)
t = 1, 2, . . . , T,
and
uft + At err, (6.6.2)
where u, and A, are the cross-section-specific and time-specific components
mentioned earlier. Assume that the sequence (A), (A,), and (eft) are i.i.d.
random variables with zero mean and are mutually independent with the
variances a, of, and a!, respectively. In addition, assume that xi, is a K-vector
of known constants, the first element of which is 1 for all i and t.
We shall write (6.6.1) and (6.6.2) in matrix notation by defining several
symbols. Define y, u, e, and X as matrices of size NT X 1, NT X 1, NT X 1,
and NT X K, respectively, such that their [(i 1)T + t]th rows are y,,,
u, ei and 4, respectively. Also define A == (Ai, 112, ANY A
(A1 , A2,. . . , ATY,L = IN 1T,where 17. is a T-vector of ones, and I = IT.
Then we can write (6.6.1) and (6.6.2) as
y = Xi/ u (6.6.3)
and
u = L,u + + e. (6.6.4)
The covariance matrix LI Euu' can be written as
---- cri2A + aB + cr!INT , (6.6.5)
where A = LL' and B = H'. Its inverse is given by
yl A Y2B + Y3J), (6.6.6)

a,
where y1 = qs(oi + Ta)',
Y2 = db .! + Na,
Y3 = Y1Y2(2a! + Tcr,! + NoD(oi + To 12, + No1)-1 ,
and J is an NT X NT matrix consisting entirely of ones.
In this model the LS estimator of /3 is unbiased and generally consistent if
both N and T go to 00, but if El is known, GLS provides a more efficient
estimator. Later we shall consider FGLS using certain estimates of the var-
iances, but first we shall show that we can obtain an estimator of/3 that has the
same asymptotic distribution as GLS (as both N and T go to 00) but that does
not require the estimation of the variances.
To define this estimator, it is useful to separate the first element of /3 (the
intercept) from its remaining elements. We shall partition ft = (tic) fii)' and
X = (1, X1).8 We shall call this estimator the transformation estimator because
it is based on the following transformation, which eliminates the cross-sec-
tion- and time-specific components from the errors. Define the NT X NT
matrix
1 1 1
Q=I-- A-- B + J. (6.6.7)
T N NT
It is easy to show that Q is a projection matrix of rank NT N T + 1 that is
orthogonal to 1, L, and I. Let H be an NT X (NT N T+ 1) matrix, the
columns of which are the characteristic vectors of Q corresponding to the unit
roots. Premultiplying (6.6.3) by H' yields
H'y = + H'e, (6.6.8)
which is Model 1 because EH'ee'H = GI. The transformation estimator of
ft1 , denoted /3,41 , is defined as LS applied to (6.6.8):
(XiQX1)-1 XIQY. (6.6.9)

The transformation estimator can be interpreted as the LS estimator of
treating au , 142 , . . , pN and Ai , A2 , . . . , AT as if they were unknown re-
gression parameters. Then formula (6.6.9) is merely a special case of the
general formula for expressing a subset of the IS estimates given in (1.2.14).
This interpretation explains why the estimator is sometimes called the fixed-
effects estimator (since g's and )'s are treated as fixed effects rather than as
random variables) or the dummy-variable regression. Still another name for
the estimator is the covariance estimator.
To compare (6.6.9) with the corresponding GLS estimator, we need to
derive the corresponding subset of GLS /kJ . We have
/LI = [xgrlx, - (6.6.10)
X [Xlirly XP-1 1(11/-1 1)-1 11ray]
= [Xl(I 71A y2B + y4 l)X 1 1-1 X(I yot y2B + y4J)u,
where y4 = (NTasicri a)/NRoi + Ta,i)(a! + Nal). Note the similarity be-

tween Q and I 71A y2B + y4J. The asymptotic equivalence between
and /16, essentially follows from this similarity. If both Nand Tgo to co (it does
not matter how they go to co), it is straightforward to prove that under reason-
able assumptions on X and u, 41-VT(All /31) and fiCrT(fi Gt /11 ) converge to
MO, lim NT(XIQX 1)-1 ]. A proof of the special case of l'X i = 0 has been
given by Wallace and Hussain (1969).
The GLS estimator of /I0 is given by
7, ll'y l'
X ijci
oo NT P
Similarly, we can define the transformation estimator of fio by
- l'y
PQO (6.6.12)
NT
We have
131)
?Go (6.6.13)
NT
and similarly for 13Q0 - . Note that
T NT
ru=Ty li,+Nx),+E ei (6.6.14)
r-i r-1 r-1
where the probabilistic orders of the three terms in the right-hand side of
(6.6.14) are TIN, NIT, and Jrvi', respectively. Because the probabilistic
order of I' X l(floi fit) or A) is ArT, it does not affect the asymp-
totic distribution of flGo or tiQo . Hence, these two estimators have the same
asymptotic distribution. To derive their asymptotic distribution, we must
specify whether Nor Tgoes to 00 faster. If Ngrows faster than T, if (A 30 J30)

and IT( /IQ flo) have the same limit distribution as IT_ Aihrt , whereas if T
grows faster than N, Nrg(1300 flo) and ,rg fib) have the same limit
distribution as
Because of the form of CI -1 given in (6.6.6), FGLS can be calculated if we
can estimate the three variances al, a, and al. Several estimators of the
variances have been suggested in the literature. (A Monte Carlo comparison of
several estimators of the variances and the resulting FGLS has been done by
Baltagi, 1981.) Amemiya (1971) proved the asymptotic normality of the fol-
lowing so-called analysis-of-variance estimators of the variances:
a. (6.6.15)
(N 1)(T 1)'
a , [ T- A _ T 1 j id a
NT
14 (6.6.16)
T(N 1XT 1)
and
[N;
1 N1
u B
N NT j 4d 6 0.6.17)
N(N 1)(T 1)
where 11 = y XfiQ . Amemiya also proved that they are asymptotically more
efficient than the estimates obtained by using y X,8 for fi, where fi is the LS
estimator.
These estimates of the variances, or any other estimates with the respective
probabilistic order of (NT) 1/2, N-112, and T-1/2, can be inserted into the
right-hand side of (6.6.6) for calculating FGLS. Fuller and Battese (1974)
proved that under general conditions FGLS and GLS have the same asymp-
totic distribution.
6.6.2 Two Error Components Model

In 2ECM there is no time-specific error component. Thus the model is a
special case of 3ECM obtained by putting al =0. This model was first used in
econometric applications by Balestra and Nerlove (1966). However, because
in the Balestra-Nerlove model a lagged endogenous variable is included
among regressors, which causes certain additional statistical problems, we
shall discuss it separately in Section 6.6.3.
In matrix notation, 2ECM is defined by

y=Xii+u (6.6.18)
and
u = Liu + e. (6.6.19)
The covariance matrix fl Euu' is given by
= qdt + (6.6.20)
and its inverse by
Q-1 = +
(41 (INT yiA), (6.6.21)
where y1 = 0/(al + Tap as before.

We can define GLS and FGLS as in Section 6.6.1. We shall discuss the
estimation of yi required for FGLS later. The asymptotic equivalence of GL,S
and FGLS follows from Fuller and Battese (1974) because they allowed for the
possibility of al =0 in their proof. The transformation estimator of fl can be
defined as in (6.6.9) and (6.6.12) except that here the transformation matrix Q
is given by
Q=I T-1A. (6.6.22)
The asymptotic equivalence between GLS and the transformation estimator
can be similarly proved, except that in the 2ECM model NIN( Poo flo) and
au3,0 - flo) have the same limit distribution as I/iv-IA/a regardless of the
way N and T go to co.
Following Maddala (1971), we can give an interesting interpretation of GIS
in comparison to the transformation estimator. Let L be as defined in Section
6.6.1 and let F be an NT X N(T 1) matrix satisfying F'L = 0 and F'F = I.
Then (6.6.18) can be equivalently written as the following two sets of regres-
sion equations:
T- '12L'y = T - '12L'XI3 + (6.6.23)
and
F'y = F'Xfl + 42, (6.6.24)
where Dim; = + Ethn'2 = crAmT_ , and En1rY2 =0. Maddala
calls (6.6.23) the between-group regression and (6.6.24) the within-group
regression. The transformation estimator/1 Q, can be interpreted as LS applied
to (6.6.24). (Note that since the premultiplication by F' eliminates the vector
of ones, flo cannot be estimated from this equation.) GLS P G can be interpreted
as GLS applied to (6.6.23) and (6.6.24) simultaneously. Because these equa-
tions constitute the heteroscedastic regression model analyzed in Section
6.5.2, GLS has the following simple form:
PG = ( X'PX + cX'MX) - I(X'Py + cX'My), (6.6.25)
where c = (a! + To)/a, P = T- ILL', and M = I P. To define FGLS, c
may be estimated as follows: Estimate a! + Ta, by the LS estimator of the
variance obtained from regression (6.6.23), estimate e4 by the LS estimator of
the variance obtained from regression (6.6.24), and then take the ratio.
As we noted earlier, (6.6.23) and (6.6.24) constitute the heteroscedastic
regression model analyzed in Section 6.5.2. Therefore the finite-sample study
of Taylor (1978) applied to this model, but Taylor (1980) dealt with this model
specifically.
Next, following Balestra and Nerlove (1966), we derive MLE of the model
assuming the normality of u. For this purpose it is convenient to adopt the
following re-parameterization used by Balestra and Nerlove: Define a 2 =
oi+ qs , p= qda 2, and R = (1 p)I T + p17.1 Then we have
II = a2(IN 0 R),
= cr-2(IN 0 II-1 ),
It- ' =(1 P)(IT [PM P + PTAIT 1r),

and IRI= ( 1 P)9 1 + PM Mk
Using these results, we can write the log likelihood function as
1 1
L= log lal u'ir l u (6.6.26)
2 2
_ _ NT kr co _ N
2T log (1 p)
N log (1 + PT )
2 5
1

W(1 0 Iti)u,
2a2
where we have written u for y XIL From (6.6.26) it is clear that the MLE ofp
given p is the same as GLS and that the MLE of a 2 must satisfy
2 U i (1 0 R- ')u
a - (6.6.27)
NT
NT
L* = log I (x
2] (6.6.28)
1 p + pT
log [1 + -1 /=p] .
Putting aLs/ap = 0 yields

2 1 p 2
I Ult Utt
1 pP F pT( uft ) 1 p + pT )
(6.6.29)
from which we obtain
wit
i t t
P (6.6.30)
(T 1)
I I
Also, using (6.6.29,) we can simplify (6.6.27) as
I 14
a2 = (6.6.31)
NT
The MLE of fi, p, and a2 can be obtained by simultaneously solving the
formula for GLS, (6.6.30), and (6.6.31).
Both Balestra and Nerlove (1966) and Maddala (1971) pointed out that the
right-hand side of (6.6.30) can be negative. To ensure a positive estimate ofp,
Balestra and Nerlove suggested the following alternative formula for p:
2
I t i
P (6.6.32)
NT u2
I t
It is easy to show that the right-hand of (6.6.32) is always positive.
Maddala (1971) showed that the p given in (6.6.30) is less than 1. Berzeg
(1979) showed that if we allow for a nonzero covariance ao between u, and el
the formulae for MLE are the same as those given in (6.6.30) and (6.6.31) by
redefining a 2 = + 2cro,+ a! and p = (cr,i+ 2up.)1a 2 and in this model the
MLE of p lies between 0 and 1.
One possible way to calculate FGIS consists of the following steps: (1)
Obtain the transformation estimator & (2) Define 0, Q = y X,8Q ; (3) Insert
L, into the right-hand side of Eq. (6.6.32). (4) Use the resulting estimator of p
to construct FGLS. In the third step the numerator of (6.6.32) divided by
N 2 T2 can be interpreted as the sample variance of the LS estimator of p i + /10
1fl1 + L(p + fl01) + e. obtainedfrmhgsoy=X
6.6.3 Balestra-Nerlove Model

As we mentioned earlier, this is a generalization of 2ECM in the sense that a
lagged endogenous variable y, is included among the regressors. Balestra
and Nerlove (1966) used this model to analyze the demand for natural gas in
36 states in the period 1950- 1962.
All the asymptotic results stated earlier for 2ECM hold also for the Balestra-
Nerlove model provided that both N and T go to 3, as shown by Amemiya
(1967). However, there are certain additional statistical problems caused by
the presence of a lagged endogenous variable; we shall delineate these prob-
lems in the following discussion.
First, the LS estimator of/3 obtained from (6.6.18) is always unbiased and is
consistent if N goes to 00. However, if x0 contain yu_ 1 , LS is inconsistent even
when both N and T go to 00. To see this, consider the simplest case
Yft = /33/41
+ Pr Eft,
- 1 (6.6.33)
where we assume Ifil < 1 for stationarity and y i0 = 0 for simplicity. Solving the
difference equation (6.6.33) and omitting i, we obtain
1 fi t-1
Y,-i 1 + Et_1 +flt..2 . . . flt-2 Ei . (6.6.34)
Therefore, putting back the subscript i,

T
plim Pi

r . T r _i yi' t_ i = 1 13 (6.6.35)
Therefore
N T 0-2
Plinl A
rr. I j
t-1

1
(6.6.36)
which implies the inconsistency of LS.

Second, we show that the transformation estimator A QI is consistent if and

only if T goes to 00 9 We have
= (XiQX1) - 'XiQe, (6.6.37)
where Q is given by (6.6.22). We need to consider only the column of XI ,
which corresponds to the lagged endogenous variable, denoted y_ i , and only
the T- IA part of Q. Thus the consistency (as T co) of flu, follows from
1 1 N (T
(6.6.38)
NT2 -1A6 = NT2 1. I 1-.1 31 I t- I
Third, if yio (the value of y4 , at time t = 1) is assumed to be an unknown

parameter for each i, the MLE of fi is inconsistent unless T goes to 00 (Ander-
son and Hsiao, 1981, 1982). The problem of how to specify the initial values is
important in a panel data study, where typically N is large but T is not.
Fourth, the possibility of a negative MLE ofp is enhanced by the presence of
a lagged endogenous variable, as shown analytically by Maddala (1971) and
confirmed by a Monte Carlo study of Nerlove (1971). In his Monte Carlo
study, Nerlove compared various estimators of ft and concluded that the
FGLS described at the end of Section 6.6.2 performs best. He found that the
transformation estimator of the coefficient on y ia_ i has a tendency of down-
ward bias.
6.6.4 Two Error Components Model with a Serially Correlated Error

In the subsection we shall discuss the 2ECM defined by Eqs. (6.6.18) and
(6.6.19) in which e follows an AR(1) process, that is,
ft = Y41-1 + (6.6.39)
where (4) are i.i.d. with zero mean and variance al. As in the Balestra-Ner-
love model, the specification of EA) will be important if T is small. Lillard and
Willis (1978) used model (6.6.39) to explain the log of earnings by the inde-
pendent variables such as race, education, and labor force experience. They
assumed stationarity for (eir), which is equivalent to assuming Eeio = 0,
Veio ai/(1 y 2), and the independence of 6 63 from . . . Thus, in
the Lillard-Willis model, Euu' = (a1 7.1'7. + F) IN, where I' is like (5.2.9)
with a 2 = a and p = y. Lillard and Willis estimated ft and,/ by IS and a ji ,
and y by inserting the LS residuals into the formulae for the normal MLE.
Anderson and Hsiao (1982) have presented the properties of the full MLE in
this model.
The model of Lillard and Weiss (1979) is essentially the same as model
(6.6.39) except that in the Lillard-Weiss model there is an additional error

component, so that u = Lu + [I L(L'L) - 11(I 0 f)C + e, where f=
(1, 2, . . . , T)' and C is independent ofp and e. The authors used LS, FGLS,
and MLE to estimate the parameters of their mode1. 1 The MLE was calcu-
lated using the LISREL program developed by Joreskog and Sorbom (1976).
Hause (1980) has presented a similar model.
Finally, MaCurdy (1982) generalized the Lillard-Willis model to a more
general time series process for Eft. He eliminated pi by first differencing and
treating yt, 3,4_ 1 as the dependent variable. Then he tried to model the IS
predictor for eft 4,_, by a standard Box-Jenkins-type procedure. MaCurdy
argued that in a typical panel data model with small T and large N the
assumption of stationarity is unnecessary, and he assumed that the initial
values Eio, . . . are i.i.d. random variables across i with zero mean and
unknown variances.
6.6.5 Two Error Components Model with Endogenous Regressors

The two error components model with endogenous regressors is defined by
yi = + + Y i6 + + el (6.6.40)
y2 = X2 fl + Zy + Y2(1 +p + E2
yT = XT/I + Zy + YT45+ + eT,

where y t = 1, 2, . . . , T, is an N-vector, X, is an N X K matrix of known
constants, Z is an N X G matrix of endogenous variables, V, is an N X F
matrix of endogenous variables, and p and et are N-vectors of random vari-
ables with the same characteristics as in 2ECM. The variable Z is always
assumed to be correlated with both p and e whereas Y, is sometimes assumed
to be correlated with both p and et and sometimes only with p. The two error
components model with endogenous regressors was analyzed by Chamberlain
and Griliches (1975) and Hausman and Taylor (1981). Chamberlain and
Griliches discussed maximum likelihood estimation assuming normality,
whereas Hausman and Taylor considered the application of instrumental
variable procedures.
Amemiya and MaCurdy (1983) proposed two instrumental variables esti-
mators: one is optimal if Y, is correlated with et and the other is optimal if Y, is
uncorrelated with e, . When we write model (6.6.40) simply as y = Wa +

the first estimator is defined by
(6.6.41)
where P 1 is the projection matrix onto the space spanned by the column
vectors of the NT X KT 2 matrix ir oa r S), where S --
(X1 , X2,. . . , XT). Amemiya and MaCurdy have shown that it is asymp-
totically optimal among all the instrumental variables estimators if Y, is cor-
related with ei . The second estimator is defined by
= n-1/2P2a- 1 'W)- W 0-1/2P20- "2y, (6.6.42)
where P2 = I - T-112-VT 0 [IN - S(S t S)- 1St ]. It is asymptotically optimal
among all the instrumental variables estimators if Y, is uncorrelated with et .
The second estimator is a modification of the one proposed by Hausman and
Taylor (1981). In both of these estimators, must be estimated. If a standard
consistent estimator is used, however, the asymptotic distribution is not af-
fected.
6.7 Random Coefficients Models

Random coefficients models (RCM) are models in which the regression coef-
ficients are random variables, the means, variances, and covariances of which
are unknown parameters to estimate. The Hildreth and Houck model, which
we discussed in Section 6.5.4, is a special case of RCM. The error components
models, which we discussed in Section 6.6, are also special cases of RCM. We
shall discuss models for panel data in which the regression coefficients contain
individual-specific and time-specific components that are independent across
individuals and over time periods. We shall discuss in succession models
proposed by Kelejian and Stephan (1983), Hsiao (1974, 1975), Swamy (1970),
and Swamy and Mehta (1977). In the last subsection we shall mention several
other related models, including so-called varying parameter regression models
in which the time-specific component evolves with time according to some
dynamic process. As RCM have not been applied as extensively as ECM, we
shall devote less space to this section than to the last.
6.7.1 The Kelejian and Stephan Model

The RCM analyzed by Kelejian and Stephan (1983) is a slight generalization
of Hsiao's model, which we shall discuss in the next subsection. Their model is
defined by
Yft = x;tfi x;r(jz1 At) Eft) (6.7.1)
i= 1, 2, . . . , N and t = 1, 2, . . . , T. Note that we have separated out the
nonstochastic part if and the random partg i + A of the regression coefficients.
Using the symbols defined in Section 6.6.1 and two additional symbols, we
y = Xii + Xit + Xs). + e, (6.7.2)
where we have defined
X = diag (X,, X2,. . . , Xiy), X* =
(X r , r, . . . , Xr)', where X? diag (x'n , x . . . , x;T). It is assumed
,
that ,u, A, and e have zero means, are uncorrelated with each other, and have
covariance matrices given by Epp' = IN 0 = IT IA, and
Eee' =1,, where 1, IA, and ; are all nonsingular.
Kelejian and Stephan were concerned only with the probabilistic order of
the GLS estimator of fl an important and interesting topic previously over-
looked in the literature. For this purpose we can assume that Z, I A , and I are ,
known. We shall discuss the estimation of these parameters in Sections 6.7.2

and 6.7.3, where we shall consider models more specific than model (6.7.1). In
these models ; is specified to depend on a fixed finite number of parameters:
most typically, 1, = a 2INT .
The probabilistic order offiG can be determined by deriving the order of the
inverse of its covariance matrix, denoted simply as V. We have
V- = X' [ X (I 0 ;)X' + A] -1 X, (6.7.3)
where A = X *(IT IA)X*' + Z. Using Theorem 20 of Appendix 1, we
obtain
[X(IN 10)X' + (6.7.4)
= A-1 A-1 X [(IN Z) + X'A-1 1(] -1X'A-1 .
Therefore, noting X = X(1N 0 IK) and defining A = X'A-11c, we have
V-1 = ON 0 IK)'(A A[(IN ;: 1) + A)(IN 0 (6.7.5)
Finally, using Theorem 19 (ii) of Appendix 1, we can simplify the (6.7.5) as
1 = [(IN 0 ) A- r( I/C) (6.7.6)
Or as
V-1 = NI ;1-1 (IN IVY f(IN + Ar 1(1p, (6.7.7)
Equation (6.7.7) is identical with Eq. (11) of Kelejian and Stephan (1983,
p. 252).
Now we can determine the order of V -1 . If we write the i, jth block subma-
trix of f(IN 1,7 1) A] -1 , j = 1, 2, . . . , N, as GU, the second term of the
right-hand side of (6.7.7) can be written as I i-i-1 (IP1_ 1 17_ 1 Gu)I; 1 . Therefore
the order of this term is N2/T. Therefore, if T goes to C at a rate equal to or
faster than N, the order of V-1 is N. But, because our model is symmetric in i
and t, we can conclude that if Ngoes to co at a rate equal to or faster than T, the
order of V-1 is T. Combining the two, we can state the order of V -1 is
min (N, T) or that the probabilistic order of fi G is max (N- 1/2, T-112).
6.7.2 Hsiao's Model

Hsiao's model (1974, 1975) is obtained as a special case of the model of the
preceding subsection by assuming 10 and la are diagonal and putting /, =
a 2INT.
Hsiao (1975) proposed the following method of estimating 1 p ZA, and a2:
,
For simplicity assume that X does not contain a constant term. A simple
modification of the subsequent discussion necessary for the case in which X
contains the constant term is given in the appendix of Hsiao (1975). Consider
the time series equation for the ith individual:
yi = + pi) + X + ei . (6.7.8)
If we treat /2; as if it were a vector of unknown constants (which is permissible

so far as the estimation of 1,1 and a2 is concerned), model (6.7.8) is the
heteroscedastic regression model considered in Section 6.5.4. Hsiao suggested
estimating IA and a 2 either by the Hildreth-Houck estimator (6.5.24) or their
alternative estimator described in note 7 (this chapter). In this way we obtain
N independent estimates of 1, 1 and a 2. Hsiao suggested averaging these N
estimates. (Of course, these estimates can be more efficiently combined, but
that may not be worth the effort.) By applying one of the Hildreth-Houck
estimators to the cross-section equations for T time periods, 1.0 can be simi-
larly estimated. (In the process we get another estimate of a 2.)
Hsiao also discussed three methods of estimating fl. The first method is
FGLS using the estimates of the variances described in the preceding para-
graph. The second method is an analog of the transformation estimator de-
fined for ECM in Section 6.6.1. It is defined as the LS estimator of ft obtained
from (6.7.2) treating g and A as if they were unknown constants. The third
method is MLE derived under the normality of y, by which fi,;,;, and a2
are obtained simultaneously. Hsiao applied to his model the method of scor-
ing that Anderson (1969) derived for a very general random coefficients
model, where the covariance matrix of the error term can be expressed as a
linear combination of known matrices with unknown weights. (Note that
Anderson's model is so general that it encompasses all the models considered
in this chapter.) Hsiao essentially proved that the three estimators have the
same asymptotic distribution, although his proof is somewhat marred by his
assumption that these estimators are of the probabilistic order of (NT) -1/2.
6.7.3 Swamy's Model

Swamy's model (1970) is a special case of the Kelejian-Stephan model
obtained by putting
= 0 and I, = 0 IT,
where I = diag (a?, ai, . . . , air).It is more restrictive than Hsiao's model
in the sense that there is no time-specific component in Swamy's model, but it
is more general in the sense that Swamy assumes neither the diagonality of;
nor the homoscedasticity of e like Hsiao.
Swamy proposed estimating X i, and in the following steps:
Step 1. Estimate al by al = X1(X;X1)-1Xflyi/(T K).
Step 2. Define bi = (X; Xi)-1X;yi ....
Step 3. Estimate ; by 20 = N 1) -1
( ' D,Y_ I bi) X
(bi N- ` N- In. 1 61(NX;)-1 .
It is easy to show that a? and 1, are unbiased estimators of a? and Z,,
respectively.
Swamy proved that the FGLS estimator of /I using aland is asymptoti-
cally normal with the order AT-1/2 and asymptotically efficient under the
normality assumption. Note that GLS is of the order of N-I/2 in Swamy's
model because, using (6.7.7), we have in Swamy's model
V-1 = N1,7 1 -1; 1 [tri (1; 1 + ai2X;Xr XIV (6.7.9)
= 0(N) 0(N/T).
6.7.4 The Swamy-Mehta Model

The Swamy-Mehta model (1977) is obtained from the Kelejian-Stephan
model by putting a 2 = 0 but making the time-specific component more gen-
era! as [diag (X 11` , . . , X t)]A*, where EA* = 0 and E).*).* =

diag (IT , 12, .., IN). In their model, (6.7.7) is reduced to
V- '=NZ 1 ( + X; [MIT 0 Ii)X11 -1XJ-1 )
= 0(N) 0(N/T), (6.7 .10)

as in Swamy's model.
6.7.5 Other Models

In the preceding subsections we have discussed models in which cross-sec-
tion-specific components are independent across individuals and time-spe-
cific components are independent over time periods. We shall cite a few
references for each of the other types of random coefficients models. They are
classified into three types on the basis of the type of regression coefficients: (1)
Regression coefficients are nonstochastic, and they either continuously or
discretely change over cross-section units or time periods. (2) Regression
coefficients follow a stochastic, dynamic process over time. (3) Regression
coefficients are themselves dependent variables of separate regression models.
Note that type 1 is strictly not a RCM, but we have mentioned it here because
of its similarity to RCM. Types 1 and 2 together constitute the varying param-
eter regression models.
References for type 1 are Quandt (1958), Hinkley (1971), and Brown,
Durbin, and Evans (1975). References for type 2 are Rosenberg (1973),
Cooley and Prescott (1976), and Harvey (1978). A reference for type 3 is
Amemiya (1978b).
Exercises
1. (Section 6.1.2)
Consider a classic regression model
y = ax + /3z + u,
where a and /3 are scalar unknown parameters; x and z are T-component
vectors of known constants such that x'l = z'l = 0, wherel is a T-compo-
nent vector of ones; and u is a T-component vector of unobservable i.i.d.
random variables with zero mean and unit variance. Suppose we are given
an estimator /I such that Eft = Vfl = T - ', and Eufl = T- '12p1, where p is
a known constant with 0 lp I < 1. Write down the expression for the best
estimator of a you can think of. Justify your choice.
2. (Section 6.1.2)
In the model of Exercise 17 of Chapter 3, assume that the exact distribu-
tion of )3 is known to be N(/3, T -1).
a. Obtain the mean squared error of .
b. Find an estimator of a whose mean squared error is smaller than
that of a-t.
3. (Section 6.1.3)
Prove that statement D statement C in Theorem 6.1.1.
4. (Section 6.1.3)
If K = 1 in Model 6, the efficiency of LS relative to GLS can be defined by
Eff ("2
(x'X- Ix)(x'Ix)
Watson (1955) showed Eff 4A vtd(ill + 42, where Ai and As are the larg-
est and smallest characteristic roots of 1, respectively. Evaluate this lower
bound for the case where I is given by (5.2.9), using the approximation of
the characteristic roots by the spectral density (cf. Section 5.1.3).
5. (Section 6.1.3)
In Model 6 assume K = 1 and X = I, a vector of ones. Also assume Z is
equal to given in (5.2.9). Calculate the limit of the efficiency of LS as
T 00. (Efficiency is defined in Exercise 4.)
6. (Section 6.1.3)
Prove (6.1.6) directly, without using the fact that GLS is BLUE.
7. (Section 6.1.5)
Consider a regression model
y= Xii + u,
where Eu =0 and Euu' = P = Z(Z'Z) -1Z'. We assume that X and Z are
T X K and T X G matrices of constants, respectively, such that
rank(X) = K, rank(Z) = G < T, and PX = X. Find a linear unbiased esti-
mator of // the variance-covariance matrix of which is smaller than or
equal to that of any other linear unbiased estimator. Is such an estimator
unique?
8. (Section 6.2)
Suppose y N(Xfl, 2), where there is no restriction on I except that it is
positive definite. Can you obtain the MLE of2 by setting the derivative of
the log likelihood function with respect to 2 equal to 0?
9. (Section 6.3.2)
Show that A i and A2 given in (6.3.5) and (6.3.6) converge to 0 in probabil-
ity under the assumption of the text.
10. (Section 6.3.2)
Combining y, xd3+ u, and u, = pu,_, + e we can write
Yr= PYt-i x13 - P74-113+ Et.
Durbin (1960) proposed estimating p by the least squares coefficient on

in the regression of y, on y,_ 1 , x and ;_ 1 . In vector notation
_ yL I [I - Z(Z'Z)-1Zly
- Z(Z1 ZrZly_ i '
where y = (Yi, Y2 MI) Y-1 = ( YO, .Y11 Yr-i)', Z-
(X, X_ 1 ), X = (xl , x2 ,. . . , X_ 1 = (xo , xi , . . . , x2- 1)'. Show
that i3D has the same asymptotic distribution as . 1 u,_ 1 u,/. 1 u 1 if
lim 7- 'Z'Z is a finite nonsingular matrix.
11. (Section 6.3.3)
In Model 6 suppose K= 1 and (u,) follow AR(1), U = pu,_ e t . If we
,
thought (14) were i.i.d., we would estimate the variance of LS /3L by V =

62/x'x, where 6'2 - (x'x)'(y'x) 2]. But the true variance is
V= x'2x/(x'x)2, where 2 is as given in (5.2.9). What can you say about
the sign of V - V?
12. (Section 6.3.3)
Consider y, = fixt lit, u, = pu,_, + e where x, is nonstochastic, IPI < 1,
,,
(et) are i.i.d. with a, = 0, VE, = a2, (14t) are stationary, and/3,P, and a 2 are
unknown parameters. Given a sample (Yr, x t), t = 1, 2, . . . , T, and
given xT+1 , what do you think is the best predictor of y 2-+I ?
13. (Section 6.3.3)
Let y, = fit + u where (14,) follow AR(1), u, = eu,_, + e,. Define the
following two predictors of y+1: 9T+1= (T + 1)/3 and 92.4. 1 = ( T + 1)13,
where /3 and /3 are the LS and GLS estimators of ,6 based on yi ,
respectively. In defining GLS, p is assumed known. Assum-

Y21 , Yr,
ing p> 0, compare the two mean squared prediction errors.
14. (Section 6.3.5)
Suppose Ey = Xfi and Vy = X(/1), meaning that the covariance matrix of
y is a function of fl. The true distribution of y is unspecified. Show that
minimizing (y XPX(11) -1(y Xfi) with respect to fi yields an incon-
sistent estimator, whereas minimizing (y Xfi)'i(fi) 1(y Xfl) +
loglI(Al yields a consistent estimator. Note that the latter estimator
would be MLE if y were normal.
15. (Section 6.3.6)
Show lim T-.. (EdL 2)2 = 0, where 4 is defined in (6.3.24). Use that to
show plim T, 4 =2.
16. (Section 6.3.6)
In the table of significance points given by Durbin and Watson (1951),
du 4 gets smaller as the sample size increases. Explain this phe-
nomenon.
17. (Section 6.3.7)
Verify (6.3.30).
18. (Section 6.3.7)
Prove (6.3.33).
19. (Section 6.3.7)
In the model (6.3.31), show plim ji = ap(a + p)/(l + ap) (cf. Malinvaud,
1961).
20. (Section 6.3.7)
Consider model (6.3.27). Define the T-vector x_ 1 =
(x0 , xi , . . . , xT_ )' and the T X 2 matrix S = (x, x_ 1 ). Show that the
instrumental variables estimator 9 =- (S'Z) - IS' y is consistent and obtain
its asymptotic distribution. Show that the estimator h of p obtained by
using the residuals y Z9 is consistent and obtain its asymptotic distri-
bution.
21. (Section 6.4)
In the SUR model show that FGLS and GLS have the same asymptotic
distribution as Tgoes to infinity under appropriate assumptions on u and
X. Specify the assumptions.
22. (Section 6.5.1)

Consider a regression model
Yt = xifi + Ur, Put-1+ et, t = 1, 2, . . , T.
Assume
(A) x, is a vector of known constants such that Mr_ ixtx; is a nonsingu-
lar matrix.
(B) u0 =-- 0.
(C) (E) are independent with EE, = 0 and VE,= o.
(D) (a?) are known and bounded both from above and away from 0.
(E) ft and p are unknown parameters.
Explain how you would estimate ft. Justify your choice of the estimator.
23. (Section 6.5.3)
Show that the density of a multinomial model with more than two re-
sponses or the density of a normal model where the variance is propor-
tional to the square of the mean cannot be written in the form of (6.5.17).
24. (Section 6.5.4)
Show that NFT(Ci i a), where'd i is defined in (6.5.21), has the same limit
distribution as Nif(Z'Z) - 'Z'v l .
25. (Section 6.5.4)
Show (6.5.25).
26. (Section 6.5.4)
Consider a heteroscedastic nonlinear regression model
AA)
where (ut) are independent and distributed as N(0, 0 0), fib is a scalar
unknown parameter, and ao is a vector of unknown parameters.
Assume
(A) If ft is the NLLS estimator of /30 , VT.( /10) N(0, c), where
0 < c < co.
(B) (;) are vectors of known constants such that Iztl< h for some
finite vector h (meaning that every element of 1;1 is smaller than the
corresponding element of h), 4a0 is positive and bounded away from 0,
and is nonsingular for every T and converges to a finite
nonsingular matrix.
(C) for some finite M for every t and every ft.
Generalize the Goldfeld-Quandt estimator ofor0 to the nonlinear case and
derive its asymptotic distribution.
27. (Section 6.6.2)

Consider an error components model
Yit = Its eit, i 1, 2, . . . , N and
t 1, 2, . . . , T,
where xi, is a K-vector of constants and yi and Eft are scalar random
variables with zero means. Define p =(p a , u2,. . . , ANY , e i =
ea , , eaY, and e = eZ, , eNr. Then we assume
= = IN 1, and Elm' = 0, where I is a T-dimensional
diagonal matrix the tth diagonal element of which is ol > 0.Assume
and! are known. Obtain the formulae for the following two estimators
of fi:
1. The generalized least squares estimator TIG ;
2. The fixed-effects estimator fi (it is defined here as the generalized
least squares estimator of II treating /21 as if they were unknown
parameters rather than random variables).
28. (Section 6.6.2)
Consider the model y = Xfi + vz + u, where
y: T + 1, observable
X: TX K, known constants
z: T X 1, known constants
fl K X 1, unknown constants
v: scalar, unobservable, Ev = 0, Vv = a 2
u: T)< 1, unobservable, Eu = 0, Euu' I.
Assume v and u are independent and a 2 is a known constant. Also assume
X'z # 0 and rank[X, z] = K + 1. Rank the following three estimators in
terms of the mean squared error matrix:
LS(fiL): (X' X) 'X' y
GLS(fiG): (X'I - LX)- 'X'I- iy, where Z = E(vz + u)(vz + u)'
zz'
QLS(BQ): (X'JX) - 'X'Jy, where J =1
z'z
7 Linear Simultaneous Equations Models
In this chapter we shall give only the basic facts concerning the estimation of
the parameters in linear simultaneous equations. A major purpose of the
chapter is to provide a basis for the discussion of nonlinear simultaneous
equations to be given in the next chapter. Another purpose is to provide a
rigorous derivation of the asymptotic properties of several commonly used
estimators. For more detailed discussion of linear simultaneous equations, the
reader is referred to textbooks by Christ (1966) and Malinvaud (1980).
7.1 Model and Identification
We can write the simultaneous equations model as

yr = XB + U (7.1.1)
where Y is a TXN matrix of observable random variables (endogenous

variables), Xis a TXK matrix of known constants (exogenous variables), U is
a TX N matrix of unobservable random variables, and rand B are N X Nand
K X N matrices of unknown parameters. We denote the t, ith element of Y by
y,,, the ith column of Y by yi , and the tth row of Y by y('g), and similarly for X
and U. This notation is consistent with that of Chapter 1.
As an example of the simultaneous equations model, consider the following
demand and supply equations:
Demand: Pt hqt ut1.
Supply: q, = hp, + 42132 + un .
The demand equation specifies the price the consumer is willing to pay for
given values of the quantity and the independent variables plus the error term,
and the supply equation specifies the quantity the producer is willing to supply
for given values of the price and the independent variables plus the error term.
The observed price and quantity are assumed to be the equilibrium values that
satisfy both equations. This is the classic explanation of how a simultaneous
equations model arises. For an interesting alternative explanation in which
Linear Simultaneous Equations Models 229
the simultaneous equations model is regarded as the limit of a multivariate

time series model as the length of the time lag goes to 0, see articles by Strotz
(1960) and Fisher (1970).
We impose the following assumptions:
ASSUMPTION 7.1.1. The sequence of N-vectors (%)) is i.i.d. with zero mean
and an unknown covariance matrix I. (Thus EU = 0 and ET- lU'U =I.)
We do not assume normality of {u (,) ), although some estimators considered in
this chapter are obtained by maximizing a normal density.
ASSUMPTION 7.1.2. Rank of X is K, and lim T- ' X ' X exists and is nonsin-
gular.
ASSUMPTION 7.1.3. r is nonsingular.
Solving (7.1.1) for Y, we obtain
Y=XII+V, (7.1.2)
where
n = Br-1 (7A.3)
and V = UT-'. We define A =- We shall call (7.1.2) the reduced

form equations, in contrast to (7.1.1), which are called the structural equa-
tions.
We assume that the diagonal elements of r are ones. This is merely a
normalization and involves no loss of generality. In addition, we assume that
certain elements off and Bare zeros.' Let y, be the column vector consisting
of those elements of the ith column of r that are specified to be neither 1 nor 0,
and let ft, be the column vector consisting of those elements of the ith column
of B that are not specified to be 0. Also, let Y, and X, be the subsets of the
columns of Y and X that are postmultiplied by y, and /I,, respectively. Then
we can write the ith structural equation as
ye = Yiy, + + u, (7.1.4)
Z i a, + u,.
We denote the number of columns of Y1 and Xi by N, and K respectively.

Combining N such equations, we can write (7.1.1) alternatively as
y = Za + u, (7.1.5)
where yigY,
a = (ac, ..
u Jac, . . , u'N y ,
and Z diag(Z i , Z2, , ZN).
We define II = Euu ' = I. Note that (7.1.5) is analogous to the multi-

variate regression model (6.4.2), except that Z in (7.1.5) includes endogenous
variables.
We now ask, Is ai identified? The precise definition of identification differs
among authors and can be very complicated. In this book we shall take a
simple approach and use the word synonymously with "existence of a consist-
ent estimator." 2 Thus our question is, Is there a consistent estimator of 04?
Because there is a consistent estimator of II under our assumptions (for
example, the least squares estimator fl = (X' X) - 'X ' Y will do), our question
can be paraphrased as, Does (7.1.3) uniquely determine a, when II is deter-
mined?
To answer this question, we write that part of (7.1.3) that involves yi and
fli as
(7.1.6)
and
irm = 0. (7.1.7)
Here, (e, ;o)' is the ith column of n, and (M1 , my is the subset of the
columns of n that are postmultiplied by y,. The second subscript 0 or 1
indicates the rows corresponding to the zero or nonzero elements of the ith
column of B. Note that fl, is a K(0 X N, matrix, where K m = K K. From
(7.1.7) it is clear that y, is uniquely determined if and only if
rank(lio) = N . , (7.1.8)
This is called the rank condition of identifiability. It is clear from (7.1.6) that
once yi is uniquely determined,fli is uniquely determined. For (7.1.8) to hold,
it is necessary to assume
K(0 N (7.1.9)
which means that the number of excluded exogenous variables is greater than
or equal to the number of included endogenous variables. The condition
(7.1.9) is called the order condition of identifiability.3
Supply curve
I Demand curves with different

values of the independent
variables
Figure 7.1 Demand and supply curves
If (7.1.8) does not hold, we say al is not identified or is underidentified. If

(7.1.8) holds, and moreover, if Kw = Ni , we say a, is exactly identified or is
just identified. If (7.1.8) holds and K(0 > N we say ai is overidentified.
-
If fil #0 and ft2 = 0 in the demand and supply model given in the beginning
of this section, y2 is identified but y 1 is not. This fact is illustrated in Figure 7.1,
where the equilibrium values of the quantity and the price will be scattered
along the supply curve as the demand curve shifts with the values of the
independent variables. Under the same assumption on the rs, we have
1
11 /11 and 112 = 72 fi (7.1.10)
1 V1 Y2 1 - Y1 Y2 I
where 111 and 1-12 are the coefficients on x 1 in the reduced form equations for p
and q, respectively. From (7.1.10) it is clear that if fli consists of a single
element, y2 is exactly identified, whereas if fl i is a vector of more than one
element, y2 is overidentified.
7.2 Full Information Maximum Likelihood Estimator

In this section we shall define the maximum likelihood estimator of the
parameters of model (7.1.1) obtained by assuming the normality of U, and we
shall derive its asymptotic properties without assuming normality. We attach
the termfull information (FIML) to distinguish it from the limited information
maximum likelihood (LIML) estimator, which we shall discuss later.
The logarithmic likelihood function under the normality assumption is

given by
NT
log L = - log (27r) + T log IIFII - - log 1/1 (7.2.1)
2 2
1
- - tr I,(yr XB)' (yr XB),
2
where II rII denotes the absolute value of the determinant of f. We define the
FIML estimators as the values of x, r, and B that maximize log L subject to
the normalization on r and the zero restrictions on r and B. We shall derive its
properties without assuming normality; once we have defined FIML, we shall
treat it simply as an extremum estimator that maximizes the function given in
(7.2.1). We shall prove the consistency and asymptotic normality of FIML
under Assumptions 7.1.1, 7.1.2, and 7.1.3; in addition we shall assume the
identifiability condition (7.1.8) and T N + K.
Differentiating log L with respect to Z using the rules of matrix differentia-
tion given in Theorem 21 of Appendix 1 and equating the derivative to 0, we
obtain
= T,(Yr XB)' (yr XB). (7.2.2)
T
log Ls = - - log 1(Y - xsr --1 ) , (Y - (7.2.3)
2
where we have omitted terms that do not depend on the unknown parameters.
The condition T N + Kis needed because without it the determinant can be
made equal to 0 for some choice of B and F. Thus the FIML estimators of the
unspecified elements of B and r can be defined as those values that minimize
ST= IT- L(Y - XBr - ')' (Y - mar- 91. (7.2.4)
Inserting Y = XB0 ri + V. where B o and rc, denote the true values, into
(7.2.4) and taking the probability limit, we obtain
plim ST=143+ (Boro, Br,) , A(B0 r-0-1 Br )1, (7.2.5)
where A = firn X' X. Moreover, the convergence is clearly uniform in B

and r in the neighborhood of B0 and ro in the sense defined in Chapter 4.
Inasmuch as plim STis minimized uniquely at B = B o and r = ro because of
the identifiability condition, the FIML estimators of the unspecified elements

of f and B are consistent by Theorem 4.1.2. It follows from (7.2.2) that the
FIML estimator of 1 is also consistent.
The asymptotic normality of the FIML estimators of f and B can be proved
by using Theorem 4.1.3, which gives the conditions for the asymptotic nor-
mality of a general extremum estimator. However, here we shall prove it by
representing FIML in the form of an instrumental variables estimator. The
instrumental variables interpretation of FIML is originally attributable to
Durbin (1963) and is discussed by Hausman (1975).
Differentiating (7.2.1) with respect to the unspecified elements of B (cf.
Theorem 21, Appendix 1) and equating the derivatives to 0, we obtain
x/ (yr - XB)I - I ..41 0, (7.2.6)
where 4 means that only those elements of the left-hand side of (7.2.6) that
correspond to the unspecified elements of B are set equal to O. The ith column
of the left-hand side of (7.2.6) is X' (Yr - XB)ai, where a' is the ith column of
/ -1 . Note that this is the derivative of log L with respect to the ith column of
B. But, because only the K, elements of the ith column of B are nonzero,
(7.2.6) is equivalent to
X (YF XB)al = 0, i = 1, 2, . . . , N. (7.2.7)
We can combine the N equations of (7.2.7) as
[ Xoc 0
X
0
1 I) [ : 1=
(i __I
y1
Y2

-
Zi ai
Z2 012
0. (7.2.8)
[0 X'7,, YN - ZNaN
Differentiating (7.2.1) with respect to the unspecified elements of r and

equating the derivatives to 0, we obtain
- Y' (YF XB)I -1 0, (7.2.9)
where means that only those elements of the left-hand side of (7.2.9) that
correspond to the unspecified elements of f are set equal to 0. Solving (7.2.2)
for TI and inserting it into (7.2.9) yield
yr -
-17 ' ( 0, (7.2.10)
where V = xBr-I. In exactly the same way that we rewrote (7.2.6) as (7.2.8),
we can rewrite (7.2.10) as

[Y'l 0 0 yi Zi ai
0 Y2 - Z2 42 2
(I -1 0 I) = 0, (7.2.11)
0 Y'N YN ZNOiN
where V1 consistsof the nonstochastic part of those vectors of Y that appear in
the right-hand side of the ith structural equation.
Defining = Xi) and = diag(Z i , / iN), we can combine
(7.2.8) and (7.2.11) into a single equation
a= I)Zr l i'(I -1 I)y. (7.2.12)
The FIML estimator of a, denoted a, is a solution of (7.2.12), where is
replaced by the right-hand side of (7.2.2). Because both Z and Z in the right-
hand side of (7.2.12) depend on a, (7.2.12) defines a implicitly. Nevertheless,
this representation is useful for deriving the asymptotic distribution of FIML
as well as for comparing it with the 3SLS estimator (see Section 7.4).
Equation (7.2.12), with / replaced by the right-hand side of (7.2.2), can be
used as an iterative algorithm for calculating Evaluate Z and I in the
right-hand side of (7.2.12) using an initial estimate of a, thus obtaining a new
estimate of a by (7.2.12). Insert the new estimate into the right-hand side, and
so on. However, Chow (1968) found a similar algorithm inferior to the more
standard Newton-Raphson iteration (cf. Section 4.4.1).
The asymptotic normality of a, follows easily from (7.2.12). Let Z and / be
and / evaluated at a, respectively. Then we have from (7.2.12)
4-T(si a) = [ '(Z 0 I)Z]' ' I)u. (7.2.13)
Because a and are consistent estimators, we can prove, by a straightforward
application of Theorems 3.2.7, 3.5.4, and 3.5.5, that under Assumptions
7.1.1, 7.1.2, and 7.1.3
a)--0 MO, [lim (7.2.14)
7.3 Limited Information Model

7.3.1 Introduction
In this section we shall consider situations in which a researcher wishes to
estimate only the parameters of one structural equation. Although these pa-
rameters can, of course, be estimated simultaneously with the parameters of

the remaining equations by FIML, we shall consider simpler estimators that
do not require the estimation of the parameters of the other structural equa-
tions.
We shall assume for simplicity that a researcher wishes to estimate the
parameters of the first structural equation
y, = Yl y + Xi ft + u 1 (- Zi a + u1 ), (7.3.1)
where we have omitted the subscript 1 from y, A and a. We shall not specify
the remaining structural equations; instead, we shall merely specify the re-
duced form equations for 171 ,
Yi = XIII + VI , (7.3.2)
which is a subset of (7.1.2). The model defined by (7.3.1) and (7.3.2) can be
regarded as a simplified simultaneous equations model in which simultaneity
appears only in the first equation. We call this the limited information model.
In contrast, we call model (7.1.1) the full information model.
Assumptions 7.1.1 and 7.1.2 are still maintained. However, Assumption
7.1.3 need not be assumed if we assume (7.3.2). Assumption 7.1.1 implies that
the rows of (u 1 , VI ) are i.i.d. with zero mean. Throughout this section we shall
denote the variance-covariance matrix of each row of (u 1 , V 1 ) by I. Partition
I as
I [ ai 1 121
121 122
where a; is the variance of each element of u1 . We assume that a is identifi-
able. It means that, if we partition II I as 111 = al , rico ) , in such a way that X i
III1 , then the rank of IL) is equal to N,, the number ofispotmuledby
elements of y.
7.3.2 Limited Information Maximum Likelihood Estimator
The LIML estimator is obtained by maximizing the joint density of y 1 and Y 1

a, H, and I without anyunderthomaliysupnwthreco
constraint. Anderson and Rubin (1949) proposed this estimator (without the
particular normalization on r we have adopted here) and obtained an explicit
formula for the LIML estimator: 4
ek = [El(' - ) m) z, ]-- 1 v, a - Any' , (7.3.3)
where A is the smallest characteristic root of W 1 W - ', W =

[3'1, Yi] i M[Yi, Yi], W1 = [yl, Y 1 ]'1%41 [y1 , Yl], M =I X(X'X)-1 X',
and M 1 = I (XI )- I .
7.3.3 Two-Stage Least Squares Estimator

The two-stage least squares (2SLS) estimator of a, proposed by Theil (1953), 5
isdefnby
ohs = PZ1 )-1 Z Pyl , (7.3.4)
where P = X(Xli)' X.
7.3.4 Asymptotic Distribution of the Limited Information Maximum

Likelihood Estimator and the Two-Stage Least Squares Estimator
The LIML and 2SLS estimators of a have the same asymptotic distribution.
In this subsection we shall derive it without assuming the normality of the
observations.
We shall derive the asymptotic distribution of 2SLS. From (7.3.4) we have
V-77(a2s PZI)-1T- 12z; pui (7.3.5)
The limit distribution of .17(6i2s a) is derived by showing that plim

PZI exists and is nonsingular and that the limit distribution of
T-1 /2E1 Pui is normal.
First, consider the probability limit of T - IZ; PZ I . Substitute
(XII1 V1 . X1 ) for Z1 in TZ PZI . Then any term involving V I converges
to 0 in probability. For example,
plim ' X( X(X'X)'X'V = plim 7-1 3Cc X(T- X'X)- T- 1 3CW
= plim T- IX; X(plim T- IXA)-1 plim T- IX'Vi =0.
The second equality follows from Theorem 3.2.6, and the third equality
follows from plim = 0, which can be proved using Theorem 3.2.1.
Therefore
plim T- IZ1 PZI = [1 111 ] (plim T- 'X'X) [ 1111 I ] (7.3.6)

o n,o o
Es A.
Furthermore, A is nonsingular because rank (r110)= N 1 , which is assumed for
identifiability.
Next, consider the limit distribution of 7 -1/2ZcPul . From Theorem 3.2.7

we have
1 vpu ke 1 [nclu (7.3.7)

X
" 1 '
where kPmeans that both sides of it have the same limit distribution. But, using
Theorem 3.5.4, we can show
proci
jut_ N(O A). (7.3.8)
,FTLx
Thus, from Eqs. (7.3.5) through (7.3.8) and by using Theorem 3.2.7 again,
we conclude
NIT a) N(0, ai A-1). (7.3.9)
To prove that the LIML estimator ofa has the same asymptotic distribution
as 2SLS, we shall first prove
plim U(2 1) = 0. (7.3.10)
We note
. ( id 1
2= min (7.3.11)
a o'Wo
which follows from the identity
(ywi es - ow l w-1/21
J'W6 nin
where ti = W"25, and from Theorems 5 and 10 of Appendix 1. Because
1 (Mid ( 1 , 7'')W1( 1 , 77
a (5'WtS (1, 7')W(1, y')'
we have
A _ 1 (7.3.12)
WI Mu,
uc[X(X/X)-- 'X' X1 (X1 Xi)Anui
Mu i
Therefore the desired result (7.3.10) follows from noting, for example,
plim 7 v2u1X(X'X)-- JX'u 1 = 0. From (7.3.3) we have
NIT (64, a) = [7-1 ZcPZ I (A. 1)7 'ZIMZ 1 1- I (7.3.13)

X [T-1/2Zc Pit, (2 1)r- 1 a TiMuil.
But (7.3.10) implies that both (A 1)T - 'Z'i MZI and (2 1)TZ1Mu i
converge to 0 in probability. Therefore, by Theorem 3.2.7,
N/T (al., a) 1-2 N/T(ZcPZI) -1 ZiPui, (7.3.14)
which implies that the LIML estimator of a has the same asymptotic distribu-
tion as 2SLS.
7.3.5 Exact Distributions of the Limited Information Maximum Likelihood

Estimator and the Two-Stage Least Squares Estimator
The exact finite sample distributions of the two estimators differ. We shall
discuss briefly the main results about these distributions as well as their ap-
proximations. The discussion is very brief because there are several excellent
survey articles on this topic by Mariano (1982), Anderson (1982), Taylor
(1982), and Phillips (1983).
In the early years (say, until the 1960s) most of the results were obtained by
Monte Carlo studies; a summary of these results has been given by Johnston
(1972). The conclusions of these Monte Carlo studies concerning the choice
between LIML and 2SLS were inconclusive, although they gave a slight edge
to 2SLS in terms of the mean squared error or similar moment criteria.
However, most of these studies did not realize that the moment criteria may
not be appropriate in view of the later established fact that LIML does not
have a moment of any order and 2SLS has moments of the order up to and
including the degree of overidentifiability (Km Ni in our notation).
The exact distributions of the two estimators and their approximations
have since been obtained, mostly for a simple two-equation mode1. 6 Anderson
(1982) summarized these results and showed that 2SLS exhibits a greater
median-bias than LIML, especially when the degree of simultaneity and the
degree of overidentifiability are large, and that the convergence to normality
of 2SLS is slower than that of LIML.
Another recent result favors LIML over 2SLS; Fuller (1977) proposed the
modified LIML that is obtained by substituting 2 c/(T K2 ), where c is any
constant, for 2 in (7.3.3). He showed that it has finite moments and dominates
2SLS in terms of the mean squared error to 0(T-2 ). Fuller's estimator can be
interpreted as another example of the second-order efficiency of the bias-cor-
rected MLE (see Section 4.2.4).
However, we must treat these results with caution. First, both Anderson's
and Fuller's results were obtained under the assumption of normality,
whereas the asymptotic distribution can be obtained without normality. Sec-
ond, Anderson's results were obtained only for a simple model; and, more-
over, it is difficult to verify in practice exactly how large the degrees of simulta-
neity and overidentifiability should be for LIML to dominate 2SLS. Third, we
should also compare the performance of estimators under misspecified
models. Monte Carlo studies indicate that the simpler the estimator the more
robust it tends to be against misspecification. (See Taylor, 1983, for a critical
appraisal of the finite sample results.)
7.3.6 Interpretations of the Two-Stage Least Squares Estimator

There are many interpretations of 2SLS. They are useful when we attempt to
generalize 2SLS to situations in which some of the standard assumptions of
the model are violated. Different interpretations often lead to different gener-
alizations. Here we shall give four that we consider most useful.
Theil's interpretation is most famous, and from it the name 2SLS is derived.
In the first stage, the least squares method is applied to Eq. (7.3.2) and the least
squares predictor Y 1 = PYI is obtained. In the second stage, Y 1 is substituted
for Y1 in the right-hand side of (7.3.1) and y 1 is regressed on Y 1 and X1 . The
least squares estimates of y i and ft1 thus obtained are 2SLS.
The two-stage least squares estimator can be interpreted as the asymptoti-
cally best instrumental variables estimator. This is the way Basmann (1957)
motivated his estimator, which is equivalent to 2SLS. The instrumental vari-
ables estimator applied to (7.3.1) is commonly defined as (S' Z 1 )-1 S ' y1 for
some matrix S of the same size as Z 1 . But, here, we shall define it more
generally as
= (Z1 PsZi ) I Z; Ps yi , (7.3.15)
where Ps = S(S 'S) S'. The matrix S should have Trows, but the number of
columns need not be the same as that of Z 1 . In addition, we assume that S
satisfies the following three conditions:
(i) plim T- 'S' S exists and is nonsingular,
(ii) plim T-1 S ' a, = 0, and
(iii) plim T-1 5' V I = O.
Under these assumptions we obtain, in a manner analogous to the derivation
of (7.3.9),
- a) - N(0, oiC -1 ), (7.3.16)
where
c ErF" 1-4)] plim T- IX'P X

o s
[11 11
nio 0
I (7.3.17)
Because A C, we conclude by Theorem 17 of Appendix 1 that 2SLS is the

asymptotically best in the class of the instrumental variables estimators de-
fined above.'
For a third interpretation, we note that the projection matrix P
used to define 2SLS in (7.3.4) purports to eliminate the stochastic
part of Y i . If V1 were observable, the elimination of the stochastic part of Y 1
by another projection matrix couldbemrityacoplshed
M I V 1 (VI V 1 ) l . The asymptotic variance-covariance matrix of the
resulting estimator is (uI X121/2121 )A hence smaller than that of
'
2SLS. When we predict V 1 by MY,, where M = I X(X' X) -1 X ', and use it

in place of M vo we obtain 2SLS.
The last interpretation we shall give originates in the article by Anderson
and Sawa (1973). Let the reduced form for y, be
yi = + vi . (7.3.18)
Then (7.3.1), (7.3.2), and (7.3.18) imply
+ (7.3.19)
where J 1 = (X ' X' X i . From (7.3.19) we obtain
= 111 + J1fl + ( hi - ni) (111 ni )7, (7.3.20)

where and III are the least squares estimators of n i and 111 , respectively.
Then 2SLS can be interpreted as generalized least squares applied to (7.3.20).
To see this, merely note that (7.3.20) is obtained by premultiplying (7.3.1) by
7.3.7 Generalized Two-Stage Least Squares Estimator

Consider regression equations
y Zot + u (7.3.21)
and
Z = xn + V, (7.3.22)
where Eu = 0, EV =0, Euu' = IP, and u and V are possibly correlated. We
assume that 11 is a known nonsingular matrix. Although the limited informa-
tion model can be written in this way by assuming that certain elements of V
are identically equal to 0, the reader may regard these two equations in a more
abstract sense and not necessarily as arising from the limited information
model considered in Section 7.3.1.
Premultiplying (7.3.21) and (7.3.22) by T -1/2, we obtain
tp-ia y = 1,11-112z a W-112 u (7.3.23)
and
lip-1/2z = 9F-1/2xn T-1/2v . (7.3.24)
We define the G2SLS estimator of a as the 2SLS estimator of a applied to
(7.3.23) and (7.3.24); that is,
&G2S = [Z"11-1X(X"P-1X)-1X'T-1Z]'Z't11-1X(X"P-'X)iX'41-1y.
(7.3.25)
Given appropriate assumptions on '11- v2 X, 1/2 u, and '11-1/2 V, we can
show
- a) N[0, (lim T-1 11' X' T-1 X11)-1 ]. (7.3.26)
As in Section 7.3.6, we can show that G2SLS is asymptotically the best instru-
mental variables estimator in the model defined by (7.3.21) and (7.3.22). The
limit distribution is unchanged if a regular consistent estimator of is substi-
tuted.
The idea of G2SLS is attributable to Theil (1961), who defined it in another
asymptotically equivalent way: (Z 'P'1/ -1 Z)-1Z ' IRP- Iy. It has the same
asymptotic distribution as (Si ms .
7.4 Three-Stage Least Squares Estimator
In this section we shall again consider the full information model defined by
(7.1.1). The 3SLS estimator of a in (7.1.5) can be defined as a special case of
G2SLS applied to the same equation. The reduced form equation comparable
to (7.3.22) is provided by
z=xn+v (7.4.1)
where X = I 0 X, H = diag , J 1 ),(112 , J2 ), . . , ( nN, JN)], v =

diag [(V i , 0),(V2 , 0), . . . , (VN, 0)], and i1 = (X ' X)- I X ' Xi .
To define 3SLS (proposed by Zenner and Theil, 1962), we need a consistent
estimator of which can be obtained as follows:
Step 1. Obtain the 2SLS estimator of ai , i = 1, 2, . . . , N.

Step 2. Calculate Ili = yi Z 11 i = 1, 2, . . . , N.
Step 3. Estimate au by a, 7-1
Next, inserting Z = Z, X = X, and 'II = / 0 I into (7.3.25), we have after
some manipulation of Kronecker products (see Theorem 22 of Appendix 1).
a3S = [Z ' ( i- ' 0 P)Z]' Z' (i -1 0 P)Y (7.4.2)
= e
where 2 = diag(2 1 , '22, , 2N) and 2, = PZi . The second formula of
(7.4.2) is similar to the instrumental variables representation of FIML given in
(7.2.12). We can make use of this similarity and prove that 3SLS has the same
asymptotic distribution as FIML. For exact distributions of 3SLS and FIML,
see the survey articles mentioned in Section 7.3.5.
7.5 Further Topics

The following topics have not been discussed in this chapter but many impor-
tant results in these areas have appeared in the literature over the past several
years: (1) lagged endogenous variables included in X; (2) serially correlated
errors; (3) prediction; and (4) undersized samples (that is, X having more
columns than rows). Harvey (1981a) has given a detailed discussion of the first
two topics. Recent results concerning the third topic can be found in articles
by Sant (1978) and by Nagar and Sahay (1978). For a discussion of the fourth
topic, consult the articles by Brundy and Jorgenson (1974) and by Swamy
(1980).
Exercises
1. (Section 7.3.1)
In the limited information model defined by (7.3.1) and (7.3.2), let
X = (X1 , X2 ) where X 1 and X2 have Ki and K2 columns, respectively.
Suppose we define a class of instrumental variables estimators of a by
(S ' Z 1 )-1 S' y, where S = (X2 A, X1 ) with A being a K2 X NI (K2 -4 NI )
matrix of constants. Show that there exists some A for which the instru-
mental variables estimator is consistent if and only if the rank condition of
identifiability is satisfied for Eq. (7.3.1).
2. (Section 7.3.3)
Show that the LIML estimator of y is obtained by minimizing
(5' W 1 (5Ic5' W 6, where 5 = (1, y' )', with respect toy. Hence the estimator
is sometimes referred to as the least variance ratio estimator. Also, show

that the minimization of (5' W1 6 5' WS yields the 2SLS estimator of y.
3. (Section 7.3.3)
Consider the model
y=Yy+X1 /3-FuZa+u
Y=XII+V-- i-T-FV,
where
1
E(u V)'(u V) = 1 = [!" liz .
T ' '
Obtain the asymptotic variance-covariance matrix of ii --- 07'W' V" y
and compare it with that of the 2SLS estimator of a.
4. (Section 7.3.3)
In a two-equation model
Yt= YiY2+ III
Y2 = Y2Y1 + fi1X1 + /32x2 + U29

compute the LIML and 2SLS estimates of y i , given the following moment
matrices
X'X =
1 10 1 yy = i X 'Y =
[1 1
L 1]' L2 i2] ' i i I
where X = (x1 x2 and Y
, )
=(YI,Y2).
5. (Section 7.3.7)
Show that Theil's G2SLS defined at the end of Section 7.3.7 is asymptoti-
cally equivalent to the definition (7.3.25).
6. (Section 7.3.7)
Define oi = [Z'X(X411X)- IX'Z] 'Z'X(XVX)-1 X'y. Show that this is a
consistent estimator of a in model (7.3.21) but not as asymptotically as
efficient as eiG2s defined in (7.3.25).
7. (Section 7.4)
Suppose that a simultaneous equations model is defined by (7.1.5) and the
reduced form Z = [I 0 NII + V. Show that 6ems defined in (7.3.25) and
Theil's G2SLS defined at the end of Section 7.3.7 will lead to the same
3SLS when applied to this model.
8. (Section 7,4)
Consider the following model:
( 1 ) Y1 = YY2
(2) Y2 = flixi + fl2x2 + u2 Xfi + u2 ,
where y, /11 , and /I2 are scalar unknown parameters, x l and x2 are T-compo-
nent vectors of constants such that TX is nonsingular for every T and
also in the limit, y i and y2 are T-component vectors of observable random
variables, and u 1 and u2 are T-component vectors of unobservable random
variables that are independently and identically distributed with zero mean
and contemporaneous variance-covariance matrix
0. 121
,2
= [6 12 - 2
a. Prove that the 3SLS estimator of y is identical with the 2SLS estima-
tor of y.
b. Transform Eqs. (1) and (2) to obtain the following equivalent model:
1 1
(3) Y2 = Y1
(4) y i = Az, + yfl2x2 + yu2 + u l .
Define the reverse 2SLS estimator of y as the reciprocal of the 2SLS estima-
tor of 1 /y obtained from Eqs. (3) and (4). Prove that the reverse 2SLS
estimator of y has the same asymptotic distribution as the 2SLS estimator
of y.
c. Assume that at period p outside the sample the following relation-
ship holds:
(5) Yip = YY2p + u
(6) y2p = /31x1p + /32x2p + u2p = xp'fi + u2p ,
where u I p and u2p are independent of u 1 and u2 with zero mean and var-
iance-covariance matrix I. We want to predict y ip when .xlp and x24, are
given. Compare the mean squared prediction error of the indirect least
squares predictor, defined by
(7) 94, x'p(X'Xr 1X'y,
with the mean squared prediction error of the 2SLS predictor defined by
(8) Ai, 574,(X'Xr IX'y2 ,
where 5; is the 2SLS estimator of y. Can we say one is uniformly smaller than
the other?
8 Nonlinear Simultaneous Equations Models
In this chapter we shall develop the theory of statistical inference for nonlinear
simultaneous equations models. The main results are taken from the author's
recent contributions (especially, Amemiya, 1974c, 1975a, 1976a, 1977a).
Some additional results can be found in Amemiya (1983a). Section 8.1 deals
with the estimation of the parameters of a single equation and Section 8.2 with
that of simultaneous equations. Section 8.3 deals with tests of hypotheses,
prediction, and computation.
8.1 Estimation in a Single Equation

8.1.1 Nonlinear Two-Stage Least Squares Estimator
In this section we shall consider the nonlinear regression equation
y, = f,(Y Xl ao ) + u t= 1, 2, . . . , T, (8.1.1)
where y, is a scalar endogeneous variable, Y, is a vector of endogenous vari-

ables, XI , is a vector of exogenous variables, ao is a K-vector of unknown
parameters, and (u,) are scalar i.i.d. random variables with Eu, = 0 and
Vu, = a2. This model does not specify the distribution of Y,. Equation (8.1.1)
may be one of many structural equations that simultaneously define the
distribution of y, and Y but here we are not concerned with the other equa-
tions. Sometimes we shall write flY X, c1 0 ) simply as ft(ao ) or as f,. We
define T-vectors y, f, and u, the tth elements of which are y f and u
respectively, and matrices Y and X,, the tth rows of which are Y; and Xi,,
respectively.'
The nonlinear least squares estimator of a0 in this model is generally incon-
sistent for the same reason that the least squares estimator is inconsistent in a
linear simultaneous equations model. We can see this by considering (4.3.8)
and noting that plim A3 0 in general because f, may be correlated with u, in
the model (8.1.1) because of the possible dependence of Y, on y,. In this
section we shall consider how we can generalize the two-stage least squares
(2SLS) method to the nonlinear model (8.1.1) so that we can obtain a consist-
ent estimator.
Following the article by Amemiya (1974c), we define the class of nonlinear
two-stage least squares (NL2S) estimators of ao in the model (8.1.1) as the
value of a that minimizes
SAalW) = (3' f )'VV(WW) -1W(Y f), (8.1.2)
where W is some matrix of constants with rank at least equal to K.
In the literature prior to the article by Amemiya (1974c), a generalization of
2SLS was considered only in special cases of the fully nonlinear model (8.1.1),
namely, (1) the case of nonlinearity only in parameters and (2) the case of
nonlinearity only in variables. See, for example, the article by Zellner, Huang,
and Chau (1965) for the first case and the article by Kelejian (1971) for the
second case. 2 The definition in the preceding paragraph contains as special
cases the Zellner-Huang-Chau definition and the Kelejian definition, as well
as Theil's 2SLS. By defining the estimator as a solution of a minimization
problem, it is possible to prove its consistency and asymptotic normality by
the techniques discussed in Chapter 4.
First, we shall prove the consistency of NL2S using the general result for an
extremum estimator given in Theorem 4.1.2. The proof is analogous to but
slightly different from a proof given by Amemiya (1974c). The proof differs
from the proof of the consistency of NLLS (Theorem 4.3.1) in that the deriva-
tive off is used more extensively here.
THEOREM 8.1.1. Consider the nonlinear regression model (8.1.1) with the
additional assumptions:
(A) lim 'W'W exists and is nonsingular.
(B) oft/8a exists and is continuous in N(a0 ), an open neighborhood of ao
(C) T 'W'(af/aa') converges in probability uniformly in a E N(a0 ).
(D) plim T IW'(Of/oa') is full rank.
Then a solution of the minimization of (8.1.2) is consistent in the sense of
Theorem 4.1.2.
Proof Inserting (8.1.1) into (8.1.2), we can rewrite T' (8.1.2) as
T-1 ST = T- lu'Pwu + '(fo f )'Pw(fo f) (8.1.3)
+ 7-1 2(f0 f )'P wu
=A 1 +A2 +A3 ,
where Pw = W(W'W)- 'W', f = f(a), and fo = f(ao ). Note that (8.1.3) is simi-
Nonlinear Simultaneous Equations Models 247
tar to (4.3.8). First, A i converges to 0 in probability because of assumption A

and the assumptions on (u,). Second, consider A2. Assumption B enables us to
write
f = fo + G. (a ao ), (8.1.4)
where G. is the matrix the tth row of which is af,/aa* evaluated at a* between
a and ao . Therefore we have
A2 = T- i(ao a)'G's PwG. (ao a). (8.1.5)
Therefore, because of assumptions C and D, A2 converges in probability

uniformly in a E N(a0) to a function that is uniquely minimized at a0.
Finally, consider A3. We have by the Cauchy-Schwartz inequality
T-I I(fo f )'P wul = T -1 00 aYG',Pwul (8.1.6)
[T-1(cro arGP wG(ao a)1 1/2

X [T- 'u'Pwu]'o.
Therefore the results obtained above regarding A I and A2 imply that A3 con-
verges to 0 in probability uniformly in a E N(a0 ). Thus we have verified all
the conditions of Theorem 4.1.2.
Next, we shall prove the asymptotic normality of NL2S by verifying the
conditions of Theorem 4.1.3.
THEOREM 8.1.2. In addition to the assumptions of the nonlinear regression
model (8.1.1) and those of Theorem 8.1.1, make the following assumptions:
(A) a2f,/aaaa' exists and is continuous in a E Ma o ).
(B) 'W'(8 2flaai0a1 ) converges in probability to a finite matrix uni-
formly in a E N(a0 ), where a, is the ith element of the vector a.
Then, if we denote a consistent solution of the minimization of (8.1.2) by et,
we have
cer ao )> N(0, a2[plim T-I GOPwGo] -l ),
where Go is afiaa' evaluated at ao .
Proof First, consider condition C of Theorem 4.1.3. We have
asT 1 ___
2 (w ,w )iwi u. (8.1.7)
rt. act T
But, using Theorems 3.5.4 and 3.5.5, V -i'(W'W)- 'W'u N[0, a2

lim T(WW)-1 1 because of assumption A of Theorem 8.1.1 and the assump-

tions on (u t ). Therefore, because of assumption D of Theorem 8.1.1 and
because of Theorem 3.2.7, we have
1 as, 1
N(0, 4a2 plim W wG0 ). (8.1.8)
Second, consider assumption B of Theorem 4.1.3. We have, for any a such

that plim a = ao ,
as, (8.1.9)
T aaaa' a T
2 1[1 u'W (a ac,YG+W] [ WW r
1 w a2f I 11
'[7' Tv' &Oa' la II'
where ( ) in the second term of the right-hand side is the matrix the ith row of
a,
which is given inside ( ), 6 is aflea' evaluated at and G+ is the matrix the
tth row of which is a filOa' evaluated at c4F between a and alp . But the term
inside ( ) converges to 0 in probability because of assumptions A and C of
Theorem 8.1.1 and assumption B of this theorem. Therefore assumptions A
and C of Theorem 8.1.1 imply
. 1 a2s, . 2
plim = phm GPK,G0 (8.1.10)
T actaa' a T
Finally, because assumption A of this theorem implies assumption A of
Theorem 4.1.3, we have verified all the conditions of Theorem 4.1.3. Hence,
the conclusion of the theorem follows from (8.1.8) and (8.1.10).
Amemiya (1975a) considered, among other things, the optimal choice of
W. It is easy to show that plim T(W wG0 )- ' is minimized in the matrix sense
(that is, A> B means A B is a positive definite matrix) when we choose
W = G EGo . We call the resulting estimator the best nonlinear two stage -
least squares (BNL2S) estimator. The asymptotic covariance matrix of r7"

times the estimator is given by
VB = cr2 plim T(drd)-1 . ( 8.1.11)
However, BNL2S is not a practical estimator because (1) it is often difficult to
find an explicit expression for G, and (2) G generally depends on the unknown
parameter vector ao . The second problem is less serious because ao may be

replaced by any consistent member of the NL2S class using some W.
Given the first problem, the following procedure recommended by
Amemiya (1976a) seems the best practical way to approximate BNL2S:
Step I. Compute et, a member of the NL2S class.
Step 2. Evaluate G afiaoe at acall it G.
Step 3. Treat d as the dependent variables of regressions and search for the
optimal set of independent variables, denoted W o , that best predict G.
Step 4. Set W = Wo .
If we wanted to be more elaborate, we could search for a different set of
independent variables for each column of G (say, W, for the ith column Ad
and set W = , P wift2 , . .
Kelejian (1974) proposed another way to approximate -6. He proposed this
method for the model that is nonlinear only in variables, but it could also work
for certain fully nonlinear cases. Let the tth row of G o be Gtu , that is, --
(afi/8a2,0 . Then, because Go, is a function of Y, and ao , it is also a function of
u, and ao ; therefore write G,(u ao ). Kelejian's suggestion was to generate u,
independently n times by simulation and approximate EG 0, by n 4. 1
ex), where ix is some consistent estimator fa Kelejian also pointed
out that G,(0, i3t) is also a possible approximation for EGi ; although it is
computationally simpler, it is likely to be a worse approximation than that
given earlier.
8.1.2 Box-Cox Transformation

Box and Cox (1964) proposed the model3
sfl + u z,(A)=x' (8.1.12)
where, for y s > 0,
z,) if )#0 (8.1.13)
= log y, if A = 0.
Note that because limA_0(y,t 1)/A. = log y z,(A) is continuous at A = 0. It is
assumed that (lid are i.i.d. with Eu, = 0 and Vu, = a2.
The transformation z1(A) is attractive because it contains y, and log y, as
special cases, and therefore the choice between y, and log y, can be made
within the framework of classical statistical inference.
Box and Cox proposed estimating A, A and a2 by the method of maximum

likelihood assuming the normality of u i . However, u, cannot be normally
distributed unless 2= 0 because zr(A) is subject to the following bounds:
if 2> 0 (8.1.14)
if A<0.
Later we shall discuss their implications on the properties of the Box-

Cox MLE.
This topic is relevant here because we can apply NL2S to this model with
only a slight modification. To accommodate a model such as (8.1.12), we
generalize (8.1.1) as
AY', Y,, x11, a) = u1. (8.1.15)
Then all the results of the previous section are valid with only an apparent
modification. The minimand (8.1.2), which defines the class of NL2S estima-
tors, should now be written as
f 'W(W/W)-- 'WI, (8.1.16)
and the conclusions of Theorem 8.1.1 and 8.1.2 remain intact.

Define the NL2S estimator (actually, a class of estimators) of a = (A, /J)',
denoted ez, in the model (8.1.12) as the value of a that minimizes
[z(2) Xfi]'W(W'W) -1 Wlz( 2) XII], (8.1.17)
where z(A) is the vector the tth element of which is z,(2). The discussion about
the choice of W given in the preceding section applies here as well. One
practical choice of W would be to use the x's and their powers and cross
products. Using arguments similar to the ones given in the preceding section,
we can prove the consistency and the asymptotic normality of the estimator.
The asymptotic covariance matrix of Nri(& a) is given by
-
az' ] --1
az
V(iii) = er2 plim T[[
. W(W'W)IW'k--
aA X] . (8.1.18)
.1.'
[[_x' ]
The Box-Cox maximum ikelihood estimator is defined as the pseudo MLE
obtained under the assumption that (u,) are normally distributed. In other
words, the Box-Cox MLE 0 of 0 = (A, /3', a2)' maximizes
log L = -- log a2 (z, x1 )2 + (A 1) log y.

2 2cr2 t-i
(8.1.19)
Because u, cannot in reality be normally distributed, (8.1.19) should be re-
garded simply as a maximand that defines an extremum estimator rather than
a proper logarithmic likelihood function. This fact has two implications:
1. We cannot appeal to the theorem about the consistency of the maximum
likelihood estimator; instead, we must evaluate the probability limit as the
value of the parameter at which Jim T 1 E log L is maximized.
2. The asymptotic covariance matrix of ./7"(0 0) is not equal to the usual
expression firn T[E82 log Ll8e8el - 1 but is given by
V(6) = lim T[E

02 log
E
a log L a log L [E 02 log
8880'8e 80' mace J '
(8.1.20)
where the derivatives are evaluated at plim 0.
Thus, to derive the probability limit and the asymptotic covariance matrix
of 0, we must evaluate E log L and the expectations that appear in (8.1.20) by
using the true distribution of u,.
Various authors have studied the properties of the Box-Cox MLE under
various assumptions about the true distribution of u, . Draper and Cox (1969)
derived an asymptotic expansion for E a log LjaA, in the model without re-
gressors and showed that it is nearly equal to 0 (indicating the approximate
consistency of the Box-Cox MLE A) if the skewness of z,(A) is small. Zarembka
(1974) used a similar analysis and showed E a log Liao 0 if (u,) are hetero-
scedastic. Hinkley (1975) proposed for the model without regressors an inter-
esting simple estimator of A, which does not require the normality assump-
tion; he compared it with the Box-Cox MLE A under the assumption that y, is
truncated normal (this assumption was also suggested by Poirier, 1978).
Hinkley showed by a Monte Carlo study that the Box-Cox MLE A and II
perform generally well under this assumption. Amemiya and Powell (1981)
showed analytically that if y, is truncated normal, plim A <A if A> 0 and
plim A > A if A < O.
Amemiya and Powell (1981) also considered the case where y, follows a
two-parameter gamma distribution. This assumption actually alters the
model as it introduces a natural heteroscedasticity in (4 1 ). In this case, the
formula (8.1.18) is no longer valid and the correct formula is given by
V(ii)= plim T(Z'P wZ) IZ'P wDPwZ(Z'P wZ)-1 , (8.1.21)
where Z = (az/a, X) and D = Euu'. Amemiya and Powell considered a

simple regression model with one independent variable and showed that
NL2S dominates the Box-Cox MLE for most of the plausible parameter
values considered not only because the inconsistency of the Box-Cox MLE is
significant but also because its asymptotic variance is much larger.
There have been numerous econometric applications of the Box-Cox MLE.
In the first of such studies, Zarembka (1968) applied the transformation to the
study of demand for money. Zarembka's work was followed up by studies by
White (1972) and Spitzer (1976). For applications in other areas of economics,
see articles by Heckman and Polachek (1974) (earnings - schooling relation-
ship); by Benus, Kmenta, and Shapiro (1976) (food expenditure analysis); and
by Ehrlich (1977) (crime rate).
All of these studies use the Box-Cox MLE and none uses NL2S. In most of
these studies, the asymptotic covariance matrix for fi is obtained as if z AA) were
the dependent variable of the regression a practice that can lead to gross
errors. In many of the preceding studies, independent variables as well as the
dependent variable are transformed by the Box-Cox transformation but possi-
bly with different parameters.
8.1.3 Nonlinear Limited Information Maximum Likelihood Estimator

In the preceding section we assumed the model (8.1.1) without specifying the
model for Y, or assuming the normality of u, and derived the asymptotic
distribution of the class of NL2S estimators and the optimal member of the
class BNL2S. In this section we shall specify the model for Y, and shall
assume that all the error terms are normally distributed; under these assump-
tions we shall derive the nonlinear limited information maximum likelihood
(NLLI) estimator, which is asymptotically more efficient than BNL2S. The
NLLI estimator takes advantage of the added assumptions, and consequently
its asymptotic properties depend crucially on the validity of the assumptions.
Thus we are aiming at a higher efficiency at the possible sacrifice of robustness.
Assume, in addition to (8.1.1),
= X11 + V; , (8.1.22)
where V, is a vector of random variables, X, is a vector of known constants, and
II is a matrix of unknown parameters. We assume that (u VD are indepen-
dent drawings from a multivariate normal distribution with zero mean and
variance-covariance matrix
(8.1.23)
We define X and V as matrices the tth rows of which are x', and V;, respec-
tively. Because u and V are jointly normal, we can write
u = VIT21 (112 + e, (8.1.24)

where e is independent of V and distributed as N(0, as 2I), where a*2 --= a2 -
aft 2 1-21012.
The model defined in (8.1.24) may be regarded either as a simplified non-
linear simultaneous equations model in which both the nonlinearity and the
simultaneity appear only in the first equation or as the model that represents
the "limited information" of the investigator. In the latter interpretation, X,
are not necessarily the original exogenous variables of the system, some of
which appear in the arguments of f, but, rather, are the variables a linear
combination of which the investigator believes will explain Y, effectively.
Because the Jacobian of the transformation from (u, V)Ao (y, Y) is unity in
our model, the log likelihood function assuming normality can be written,
apart from a constant, as
T 1
L** = -- log IZI - - tr r `Q, (8.1.25)
2
where
u'u u'v
Q
= [ v'u v'vl
u and V representing y - f and Y - XII, respectively. Solving aL**/ax = 0 for
I yields
I = T-1Q. (8.1.26)
Substituting (8.1.26) into (8.1.25), we obtain a concentrated log likelihood
function
T
L* = - -1- (log u'u + log IV'MVI). (8.1.27)
Solving aL*/an= 0 for II, we obtain

n = (X'MuX)-1 X/MY, (8.1.28)
where M. = 1 (u'uruu'. Substituting (8.1.28) into (8.1.27) yields a further
concentrated log likelihood function
T
L = -. (log u'u + log IY1M,,Y - Y'M.X(X'AIX)- IX'MuY1),
(8.1.29)
which depends only on a. Interpreting our model as that which represents the
limited information of the researcher, we call the value of a that minimizes
(8.1.29) the NLLI estimator. The asymptotic covariance matrix of VT times
the estimator is given by
-
VL = plim T cellt2 G'M vG 042 1 17
12) G'MKG] , (8.1.30)
where M v = I V(V'V)- IV' and Mx = I X(X'X)-1X'.

The maximization of (8.1.29) may be done by the iterative procedures
discussed in Section 4.4. Another iterative method may be defined as follows:
Rewrite (8.1.27) equivalently as
L* = -- (log u'M vu + logiV'VI) (8.1.31)

2
and iterate back and forth between (8.1.28) and (8.1.31). That is, obtain
H = (X'X)- IX'Y and V = Y XII, maximize (8.1.31) with respect to a after
-
replacing V with V, call this estimator at and define fi = y flee), insert it into
(8.1.28) to obtain another estimator of n, and repeat the procedure until
convergence.
The estimator & defined in the preceding paragraph is interesting in its own
right. It is the value of a that minimizes
f MxY(rMxY) iVIVIxi(Y f). (8.1.32)
Amemiya (1975a) called this estimator the modified nonlinear two-stage
least squares (MNL2S) estimator. The asymptotic covariance matrix of
VTOL a) is given by
Vm = plim T(G'M vG)-1 [642G'MvG (8.1.33)
+ (a2 a*2)G'PxG](G'M yG)-1 .
Amemiya (1975a) proved V L < Vm < VB. It is interesting to note that if f is
linear in a and Y, MNL2S is reduced to the usual 2SLS (see Section 7.3.6).
In Sections 8.1.1 and 8.1.3, we discussed four estimators: (1) NL2S (as a
class); (2) BNL2S; (3) MNL2S; (4) NLLI. If we denote NL2S(W = X) by
SNL2S (the first S stands for standard), we have in the linear case
SNL2S BNL2S MNL2S NLLI, (8.1.34)
where =- means exact identity and means asymptotic equivalence. In the
nonlinear model defined by (8.1.1) and (8.1.22) with the normality assump-
tion, we can establish the following ranking in terms of the asymptotic covar-
iance matrix:
SNL2S BNL2S MNL2S 41 NLLI, (8.1.35)
where C means "is worse than." However, it is important to remember that
the first two estimators are consistent under more general assumptions than
those under which the last two estimators are consistent, as we shall show in
the following simple example.
Consider a very simple case of (8.1.1) and (8.1.22) given by
y, = az? + u, (8.1.36)
and
Z1 = 7LX, V (8.1.37)
where we assume the vector (u v,) is i.i.d. with zero mean and a finite
nonsingular covariance matrix. Inserting (8.1.37) into (8.1.36) yields
y, = aex? + au?, + (u,+ 2anx,v, + av? ca), (8.1.38)
where the composite error term (contained within the parantheses) has zero
mean. In this model, SNL2S is 2SLS with 4 regressed on x, in the first stage,
and BNL2S is 2SLS with 4 regressed on the constant term and .4 in the first
stage. Clearly, both estimators are consistent under general conditions with-
out further assumptions on u, and v On the other hand, it is not difficult to
show that the consistency of MNL2S and NLLI requires the additional as-
sumption
Ev?Ev?u,= Ev?Ev,u (8.1.39)
which is satisfied if u, and v, are jointly normal.
8.2 Estimation in a System of Equations
8.2.1 Introduction
Define a system of N nonlinear simultaneous equations by
fit( Yr xr =u i = 1, 2, . . . , N, t= 1, 2, . T,
(8.2.1)
where y, is an N-vector of endogenous variables, x, is a vector of exogenous
variables, and a, is a Krvector of unknown parameters. We assume that the
N-vector u, = (u11 , u2 . . . , is an i.i.d. vector random variable with

zero mean and variance-covariance matrix X. Not all of the elements of
vectors y, and x, may actually appear in the arguments of eachf, We assume
that each equation has its own vector of parameters a, and that there are no
constraints among ai's, but the subsequent results can easily be modified if
each a, can be parametrically expressed as a,(0), where the number of ele-
ments in 0 is less than 4, 1 14.
Strictly speaking, (8.2.1) is not a complete model by itself because there is no
guarantee that a unique solution for Yt exists for every possible value of u
unless some stringent assumptions are made on the form off, Therefore we
assume either that f,, satisfies such assumptions or that if there is more than
one solution for y there is some additional mechanism by which a unique
solution is chosen. 4
We shall not discuss the problem of identification in the model (8.2.1).
There are not many useful results in the literature beyond the basic discussion
of Fisher (1966), as summarized by Goldfeld and Quandt (1972, p. 221), and a
recent extension by Brown (1983). Nevertheless, we want to point out that
nonlinearity generally helps rather than hampers identification, so that, for
example, in a nonlinear model the number of excluded exogenous variables in
a given equation need not be greater than or equal to the number of parame-
ters of the same equation. We should also point out that we have actually given
one sufficient condition for identifiability that plim T - I(G,TwG0 ) in the
conclusion of Theorem 8.1.2 is nonsingular.
Definition of the symbols used in the following sections may facilitate the
discussion.
A =I 0 I, where 0 is the Kronecker product

=L(h, 3E/, ad
f, = an N-vector, the ith element of which is A,
= a T-vector, the tth element is f,
f= f2), . , an NT-vector
F = (f(j) , f(2) ,. . . , fon ), a TXN matrix
gi, = , a IC-vector
Oa,
G, = 9, a T X IC, matrix, the tth row is g'
G = thag(G 1 , G2,. . . GN), an NT X KO block diagonal matrix

8.2.2 Nonlinear Three-Stage Least Squares Estimator

As a natural extension of the class of the NL2S estimatorsthe version that
minimizes (8.1.22), Jorgenson and Laffont (1974) defined the class of nonlin-
ear three-stage least squares (NL3S) estimators as the value of a that mini-
mizes
fji-1 0 W(VVIV) - Wif, (8.2.2)
where i is a consistent estimate of I. For example,

... .
1= - xft(a)fi(ar, (8.2.3)
where it is the NL2S estimator obtained from each equation. This definiton of
the NL3S estimators is analogous to the definition of the linear 3SLS as a
generalization of the linear 2SLS. The consistency and the asymptotic nor-
mality of the NL3S estimators defined in (8.2.2) and (8.2.3) have been proved
by Jorgenson and Laffont (1974) and Gallant (1977).
The consistency of the NL2S and NL3S estimators of the parameters of
model (8.2.1) can be proved with minimal assumptions on u ft namely,
those stated after (8.2.1). This robustness makes the estimators attractive.
Another important strength of the estimators is that they retain their consist-
ency regardless of whether or not (8.2.1) yields a unique solution for y, and, in
the case of multiple solutions, regardless of what additional mechanism
chooses a unique solution. (MaCurdy, 1980, has discussed this point further.)
However, in predicting the future value of the dependent variable, we must
know the mechanism that yields a unique solution.
Amemiya (1977a) defined the class of the NL3S estimators more generally
as the value of a that minimizes
f 'A-- 'S(Sk 'S) IS'A.- 'f, (8.2.4)
where A is a consistent estimate of A and S is a matrix of constants with NT

rows and with the rank of at leastEPT.. 1 I Ci . This definition is reduced to the
Jorgenson-Laffont definition if S = diag(W, W,. . . , W). The asymptotic
variance-covariance matrix of NiT times the estimator is given by
V3 = plim T[G'A- I S(S'A-1 S) IS'A- 'Gr I . (8.2.5)
Its lower bound is equal to

VE3 = lim T[EG'A-1 .E]- ' (8.2.6)
which is attained when we choose S = EG. We call this estimator the BNL3S
estimator (B for "best").
We can also attain the lower bound (8.2.6) using the Jorgenson-Laffont
definition, but that is possible if and only if the space spanned by the column
vectors of W contains the union of the spaces spanned by the column vectors
of EG, for i = 1, 2, . . . , N. This necessitates including many columns in W,
which is likely to increase the finite sample variance of the estimator although
it has no effect asymptotically. This is a disadvantage of the Jorgenson-Laffont
definition compared to the Amemiya definition.
Noting that BNL3S is not practical, as was the case with BNL2S, Amemiya
(1976a) suggested the following approximation:
Step 1. Compute , an SNL2S estimator of at , i = 1, 2, . . . , N.
Step 2. Evaluate G, at &1 call it G,.
Step 3. Treat 'di as the dependent variables of a regression and search for the
optimal set of independent variables W i that best predict Gi .
Step 4. Choose S = diag(P, P22 ,. . . , PNdN), where P, --
-
Wi(W;Wi) ' W;.
Applications of NL3S can be found in articles by Jorgenson and Lau (1975),
who estimated a three-equation translog expenditure model; by Jorgenson
and Lau (1978), who estimated a two-equation translog expenditure model;
and by Haessel (1976), who estimated a system of demand equations, nonlin-
ear only in parameters, by both NL2S and NL3S estimators.
8.2.3 Nonlinear Full Information Maximum Likelihood Estimator

In this subsection we shall consider the maximum likelihood estimator of
model (8.2.1) under the normality assumption of u, . To do so we must
assume that (8.2.1) defines a one-to-one correspondence between y, and
This assumption enables us to write down the likelihood function in the usual
way as the product of the density of u, and the Jacobian. Unfortunately, this is
a rather stringent assumption, which considerably limits the usefulness of the
nonlinear full information maximum likelihood (NLFI) estimator in prac-
tice. There are two types of problems: (1) There may be no solution for y for
some values of u. (2) There may be more than one solution for y for some
values of u. In the first case the domain of u must be restricted, a condition that
implies that the normality assumption cannot hold. In the second case we
must specify a mechanism by which a unique solution is chosen, a condition
that would complicate the likelihood function. We should note that the NL2S
and NL3S estimators are free from both of these problems.
Assuming ut - N(0, I), we can write the log likelihood function of model
(8.2.1) as
1
L* = - - log III + log II afday,11- f f1 . (8.2. 7)
2 r-1
Solving ay yaz = 0 for 1, we get

1
= -E fef;. (8.2.8)
T
L=i
T 1
A
2- log 17, T logIafdy,1- (8.2.9)
The NLFI maximum likelihood estimator of a is defined as the value of a that

maximizes (8.2.9).
Amemiya (1977a) proved that if the true distribution of u t is normal, NLFI
is consistent, asymptotically normal, and in general has a smaller asymptotic
covariance matrix than BNL3S. It is well known that in the linear model the
full information maximum likelihood estimator has the same asymptotic
distribution as the three-stage least squares estimator. Amemiya showed that
the asymptotic equivalence occurs if and only iff can be written in the form
Ahrt, xtt ar) = Ai(ar)'z(Y x1) B i(ai, xt), (8.2.10)
where z is an N-vector of surrogate variables.

If the true distribution of u, is not normal, on the other hand, Amemiya
proved that NLFI is generally not consistent, whereas NL3S is known to be
consistent even then. This, again, is contrary to the linear case in which the full
information maximum likelihood estimator obtained under the assumption
of normality is consistent even if the true distribution is not normal. Note that
this result is completely separate from and in no way contradicts the quite
likely fact that the maximum likelihood estimator of a nonlinear model de-
rived under the assumption of a certain regular nonnormal distribution is
consistent if the true distribution is the same as the assumed distribution.
We shall see how the consistency of NLFI crucially depends on the normal-
ity assumption. Differentiating (8.2.9) with respect to a t , we obtain
r ag. T )-
( (8.2.11)
act; t- it t-i I
where ( )7 1 denotes the ith column of the inverse of the matrix within the
parentheses. The consistency of NLFI is equivalent to the condition
lim E
(8.2.12)
and hence to the condition

T T
firn E = hm Egliu;cri, (8.2.13)
tm it T
where cri is the ith column of Z-'. Now (8.2.13) could hold even if each term of
a summation is different from the corresponding term of the other, but that
event is extremely unlikely. Therefore we can say that the consistency of NLFI
is essentially equivalent to the condition
E ! z1
=
,
Egi u tai. (8.2.14)
It is interesting to note that the condition (8.2.14) holds if u, is normal

because of the following lemma. 5
LEMMA. Suppose u = (u l , u2 ,. . . , upi)' is distributed as N(0, X), where X
is positive definite. If ah(u)/au i is continuous, Elah/au il < co, and El <
co, then
ah
E = Ehu'al, (8.2.15)
au,
where a' is the ith column of 1-1.
In simple models the condition (8.2.14) may hold without normality. In the
model defined by (8.1.36) and (8.1.37), we have g l, = -z? and g2, = x,.
Therefore (8.2.14) clearly holds for i = 2 for any distribution of u, provided
that the mean is O. The equation for i = 1 gives (8.1.39), which is satisfied by a
class of distributions including the normal distribution. (Phillips, 1982, has
presented another simple example.) However, if g, is a more complicated ,
nonlinear function of the exogenous variables and the parameters (ai ) as well
as of u, (8.2.14) can be made to hold only when we specify a density that
depends on the exogenous variables and the parameters of the model. In such
a case, normality can be regarded, for all practical purposes, as a necessary and
sufficient condition for the consistency of NLFI.
It is interesting to compare certain iterative formulae for calculating NLFI
and BNL3S. By equating the right-hand side of (8.2.11) to 0 and rearranging

terms, we can obtain the following iteration to obtain NLFI:
elm = am cO l- io- id ,A- 1, (8.2.16)
where
(8.2.17)
T au;
and d --= diag(6 1 , d2 . . . , d' iv) and all the variables that appear in the sec-
ond term of the right-hand side of (8.2.16) are evaluated at am .
The Gauss-Newton iteration for BNL3S is defined by
el(2) = a (l) (7c 'A-1 G) -1- 1 t (8.2.18)

where -6; = EG; and G = diag(-6 1 , -d2 kr) as before.
Thus we see that the only difference between (8.2.16) and (8.2.18) is in the
respective "instrumental variables" used in the formulae. Note that defined
in (8.2.17) can work as a proper set of "instrumental variables" (that is,
variables uncorrelated with u,) only if u, satisfies the condition of the afore-
mentioned lemma, whereas G. is always a proper set of instrumental variables,
a fact that implies that BNL3S is more robust than NLFI. If u, is normal,
however, t, contains more of the part of G. uncorrelated with u, than G, does,
which implies that NLFI is more efficient than BNL3S under normality.
Note that (8.2.16) is a generalization of the formula (7.2.12) for the linear
case. Unlike the iteration of the linear case, however, the iteration defined by
(8.2.16) does not have the property that a(2) is asymptotically equivalent to
NLFI when ao) is consistent. Therefore its main value may be pedagogical,
and it may not be useful in practice.
8.3 Tests of Hypotheses, Prediction, and Computation

Suppose we want to test a hypothesis of the form h(a) = 0 in model (8.1.1),
where h is a q-vector of nonlinear functions. Because we have not specified the
distribution of Y we could not use the three tests defined in (4.5.3), (4.5.4),
and (4.5.5) even if we assumed the normality of u. But we can use two test
statistics: (1) the generalized Wald test statistic analogous to (4.5.21), and (2)
the difference between the constrained and the unconstrained sums of
a a
squared residuals (denoted SSRD). Let and be the solutions of the uncon-
strained and the constrained minimization of (8.1.2), respectively. Then the

generalized Wald statistic is given by
Wald = h(a)' [H(G'PwG)- LH']lh(a), (8.3.1)

S(a)
T
where fi = ph/ clot% and O = Pf/aala, and the SSRD statistic is given by
T
SSRD = 7-(a) S 2(a)]. (8.3.2)
ST(a) [S
Both test statistics are asymptotically distributed as chi-square with q degrees
of freedom.
Gallant and Jorgenson (1979) derived the asymptotic distribution (a non-
central chi-square) of the two test statistics under the assumption that a
deviates from the hypothesized constraints in the order of T-I/2.
As an application of the SSRD test, Gallant and Jorgenson tested the hy-
pothesis of homogeneity of degree zero of an equation for durables in the
two-equation translog expenditure model of Jorgenson and Lau (1978).
The Wald and SSRD tests can be straightforwardly extended to the system
of equations (8.2.1) by using NL3S in lieu of NL2S. As an application of the
SSRD test using NL3S, Gallant and Jorgenson tested the hypothesis of sym-
metry of the matrix of parameters in the three-equation translog expenditure
model of Jorgenson and Lau (1975).
If we assume (8.1.22) in addition to (8.1.1) and assume normality, the
model is specified (although it is a limited information model); therefore the
three tests of Section 4.5.1 can be used with NLLI. The same is true of NLFI in
model (8.1.22) under normality. The asymptotic results of Gallant and Holly
(1980) given in Section 4.5.1 are also applicable if we replace cr 2(G'G) ' in the
right-hand side of (4.5.26) by the asymptotic covariance matrix of the NLFI
estimator.
8.3.2 Prediction
Bianchi and Calzolari (1980) proposed a method by which we can calculate
the mean squared prediction error matrix of a vector predictor based on any
estimator of the nonlinear simultaneous equations model. Suppose the struc-
tural equations can be written as f(yp , xp , a) =-- up at the prediction period p
and we can solve for yp as yp = g(xp , a, up ). Define the predictor tp based on
a
the estimator by tp = g(Xp 0). (Note that yp is an N-vector.) We call this
the deterministic predictor. Then we have

E( Yp 9p)(Yp tp)' (8.3.3)
= E[g(xp , a, up ) g(xp , a, 0)] [g(xp , a, up ) g(xp , a, 0)]'
+ E[g(xp , a, 0) g(xp , et, 0)] [g(xp , a, 0) exp ,
A, + A2.
Bianchi and Calzolari suggested that A I be evaluated by simulation. As for A2,
we can easily obtain its asymptotic value from the knowledge of the asymp-
totic distribution of .
Mariano and Brown (1983) and Brown and Mariano (1982) compared the
deterministic predictor tp defined in the preceding paragraph with two other
predictors called the Monte Carlo predictor irp and the residual-based predic-
tor jrp defined as
_ i s
yp = E vi ), (8.3.4)
5-I
where {v5 ) are i.i.d. with the same distribution as up , and
1 T
yp = ex,,, (8.3.5)
where = f(y x
Because yp tp = (yp Eyp ) + (Eyp Sfp ) for any predictor tp , we should
compare predictors on the basis of how well Ey p = Eg(xp , a, up ) is estimated
by each predictor. Moreover, because a is common for the three predictors,
we can essentially consider the situation where & in the predictor is replaced by
the parameter a. Thus the authors' problem is essentially equivalent to that of
comparing the following three estimators of Eg(up ):
Deterministic: g(0)
i s
Monte Carlo: g(v)
S
T
Residual-based: g(u)
T
Clearly, the deterministic predictor is the worst, as Mariano and Brown (1983)
concluded. According to their other article (Brown and Mariano, 1982), the
choice between the Monte Carlo and residual-based predictors depends on the
consideration that the former can be more efficient if S is large and the
assumed distribution of up is true, whereas the latter is simpler to compute and
more robust in the sense that the distribution of up need not be specified.
8.3.3 Computation
The discussion of the computation of NLFI preceded the theoretical discus-
sion of the statistical properties of NLFI by more than ten years. The first
article on computation was by Eisenpress and Greenstadt (1966), who pro-
posed a modified Newton-Raphson iteration. Their modification combined
both (4.4.4) and (4.4.5). Chow (1973) differed from Eisenpress and Greenstadt
in that he obtained simpler formulae by assuming that different parameters
appear in different equations, as in (8.2.1). We have already mentioned the
algorithm considered by Amemiya (1977a), mainly for a pedagogical purpose.
Dagenais (1978) modified this algorithm to speed up the convergence and
compared it with a Newton-Raphson method proposed by Chow and Fair
(1973) and with the DFP algorithm mentioned in Section 4.4.1 in certain
examples of nonlinear models. The results are inconclusive. Belsley (1979)
compared the computational speed of the DFP algorithm in computing NLFI
and NL3S in five models of various degrees of complexity and found that
NL3S was three to ten times faster. Nevertheless, Belsley showed that the
computation of NLFI is quite feasible and can be improved by using a more
suitable algorithm and by using the approximation of the Jacobian proposed
by Fairsee Eq. (8.3.6).
Fair and Parke (1980) estimated Fair's (1976) macro model (97 equations,
29 of which are stochastic, with 182 parameters including 12 first-order auto-
regressive coefficients), which is nonlinear in variables as well as in parameters
(this latter nonfinearity is caused by the transformation to take account of the
first-order autoregression of the errors), by OLS, SNL2S, the Jorgenson-Laf-
font NL3S, and NLFI. The latter two estimators are calculated by a deriva-
tive-free algorithm proposed by Parke (1982).
Parke noted that the largest model for which NLFI and NL3S had been
calculated before Parke's study was the one Belsley calculated, a model that
contained 19 equations and 61 unknown coefficients. Parke also noted that
Newton's method is best for linear models and that the DFP method is pre-
ferred for small nonlinear models; however, Parke's method is the only feasi-
ble one for large nonlinear models.
Fair and Parke used the approximation of the Jacobian
log 1.4 = (log Phi + log 1J,2 1 +. . . + log 1.16,1), (8.3.6)

n
where J = n is a small integer, and t 1 , t2 ,. . . , t are equally spaced
between 1 and T.
The hypothesis (8.2.14) can be tested by Hausman's test (Section 4.5.1)
using either NLFI versus NL3S or NLFI versus NL2S. By this test, Fair found
little difference among the estimators. Fair also found that in terms of predic-
tive accuracy there is not much difference among the different estimators, but,
in terms of policy response, OLS is set apart from the rest.
Hatanaka (1978) considered a simultaneous equations model nonlinear
only in variables. Such a model can be written as F(Y, X)F + X111 = U. Define
k by F(t, x)r + X= 0, where r and B are the OLS estimates. Then
Hatanaka proposed using F(Y, X) as the instruments to calculate 3SLS.
He proposed the method-of-scoring iteration to calculate NLFI, where the
iteration is started at the aforementioned 3SLS.
Exercises
1. (Section 8.1.2)
Define what you consider to be the best estimators of a and /I in the model
(y, + a)2 = fix, + u t = 1, 2, . . . , T, where (x,) are known constants
and (u,) are i.i.d. with Eu, = 0 and Vu,= a2. Justify your choice of esti-
mators.
2. (Section 8.1.2)
In model (8.1.12) show that the minimization of IT_ [z,(2) xfl] 2 with
respect to A and fl yields inconsistent estimates.
3. (Section 8.1.3)
Prove the consistency of the estimator of a obtained by minimizing
(y f )'[I V( V'V)ily]i(y f) and derive its asymptotic variance-
covariance matrix. Show that the matrix is smaller (in the matrix sense)
than 'TM given in (8.1.33).
4. (Section 8.1.3)
In the model defined by (8.1.36) and (8.1.37), consider the following two-
stage estimation method: In the first stage, regress z, on x, and define
1, = iIx,, where i i is the least squares estimator; in the second stage, regress
y, on (4)2 to obtain the least squares estimator of a. Show that the resulting
estimator of a is inconsistent. (This method may be regarded as an appli-
cation of Theil's interpretation Section 7.3.6 to a nonlinear model.)
5. (Section 8.1.3)
In the model defined by (8.1.36) and (8.1.37), show that the consistency of
MNL2S and NLLI requires (8.1.39).
6. (Section 8.1.3)
In the model defined by (8.1.36) and (8.1.37), assume it = 1, x, = 1 for all t,
Vu,= Vv, = 1, and Cov(u v,) = c. Evaluate the asymptotic variances,
denoted V1 , V2, V3, and V4, of the SNL2S, BNL2S, MNL2S, and NLLI
estimators of a and show Ili V2 V3 V4 for every I cl < 1.
7. (Section 8.2.3)
Consider the following two-equation model (Goldfeld and Quandt, 1968):
log y i, = yl log y2, + /31 + )62x, + u lt ,
Y2t = 72Ylt /33-x1 U2,9
where y, > 0 and y2 < 0. Show that there are two solutions of yu and y2, for
a given value of (u 1 u2,). Show also that (Li u , u21 ) cannot be normally
distributed.
8. (Section 8.2.3)
Consider the following model (Phillips, 1982):
log yi, + a1x = Ul t ,
Y2t a2Y1t 1.12t
Derive the conditions on u 1, and u 2, that make NLFI consistent. Use

(8.2.14).
9 Qualitative Response Models
9.1 Introduction
Qualitative response models (henceforth to be abbreviated as QR models) are
regression models in which dependent (or endogenous) variables take discrete
values. These models have numerous applications in economics because
many behavioral responses are qualitative in nature: A consumer decides
whether to buy a car or not; a commuter chooses a particular mode of trans-
portation from several available ones; a worker decides whether to take a job
offer or not; and so on. A long list of empirical examples of QR models can be
found in my recent survey (Amemiya, 1981).
Qualitative response models, also known as quantal, categorical, or discrete
models, have been used in biometric applications longer than they have been
used in economics. Biometricians use the models to study, for example, the
effect of an insecticide on the survival or death of an insect, or the effect of a
drug on a patient. The kind of QR model used by biometricians is usually the
simplest kind univariate binary (or dichotomous) dependent variable (sur-
vival or death) and a single independent variable (dosage).
Economists (and sociologists to a certain extent), on the other hand, must
deal with more complex models, such as models in which a single dependent
variable takes more than two discrete values (multinomial models) or models
that involve more than one discrete dependent variable (multivariate models),
as well as considering a larger number of independent variables. The estima-
tion of the parameters of these complex models requires more elaborate
techniques, many of which have been recently developed by econometricians.
This chapter begins with a discussion of the simplest model the model for
a univariate binary dependent variable (Section 9.2), and then moves on to
multinomial and multivariate models (Sections 9.3 and 9.4). The emphasis
here is on the theory of estimation (and hypothesis testing to a lesser extent)
and therefore complementary to Amemiya's survey mentioned earlier, which
discussed many empirical examples and contained only fundamental results
on the theory of statistical inference. We shall also discuss important topics
omitted by the survey choice-based sampling (Section 9.5), distribution-

free methods (Section 9.6), and panel data QR models (Section 9.7).
9.2 Univariate Binary Models

9.2.1 Model Specification
A univariate binary QR model is defined by
P(yi = 1) = F(x'iflo ), i = 1, 2, . . . , n, (9.2.1)
where (y1 ) is a sequence of independent binary random variables taking the
value 1 or 0, x i is a K-vector of known constants, /10 is a K-vector of unknown
parameters, and F is a certain known function.
It would be more general to specify the probability as F(x it), but the
specification (9.2.1) is most common. As in the linear regression model,
specifying the argument ofFas x'iflo is more general than it seems because the
elements of x i can be transformations of the original independent variables.
To the extent we can approximate a general nonlinear function of the original
independent variables by x;fl o , the choice of F is not critical as long as it is a
distribution function. An arbitrary distribution function can be attained by
choosing an appropriate function H in the specification F[H(xi , f)]. .
The functional forms of F most frequently used in application are the
following:
Linear Probability Model
F(x) = x.
Probit Model
F(x) = (1)(x)
Logit Model
J L exp [ (t2/2)] dt.
ex
F(x)= A(x)
The linear probability model has an obvious defect in that F for this model"
is not a proper distribution function as it is not constrained to lie between 0
and 1. This defect can be corrected by defining F = 1 if F(x'iflo ) > 1 and F = 0
if F(x'iflo ) <0, but the procedure produces unrealistic kinks at the truncation
Qualitative Response Models 269
points. Nevertheless, it has frequently been used in econometric applica-

tions, especially during and before the 1960s, because of its computational
simplicity.
The probit model, like many other statistical models using the normal
distribution, may be justified by appealing to a central limit theorem. A major
justification for the logit model, although there are other justifications (nota-
bly, its connection with discriminant analysis, which we shall discuss in Sec-
tion 9.2.8), is that the logistic distribution function A is similar to a normal
distribution function but has a much simpler form. The logistic distribution
has zero mean and variance equal to n 2/3. The standardized logistic distribu-
tion er/(1 ex) with A = it! slightly heavier tails than the standard
normal distribution.
We shall consider two examples, one biometric and one econometric, of the
model (9.2.1) to gain some insight into the problem of specifying the probabil-
ity function.
EXAMPLE 9.2.1. Suppose that a dosage xi (actually the logarithm of dosage is

used as xi in most studies) of an insecticide is given to the ith insect and we
want to study how the probability of the ith insect dying is related to the dosage
xi . (In practice, individual insects are not identified, and a certain dosage x, is
given to each of n, insects in group t. However, the present analysis is easier to
understand if we proceed as if each insect could be identified). To formulate
this model, it is useful to assume that each insect possesses its own tolerance
against a particular insecticide and dies when the dosage level exceeds the
tolerance. Suppose that the tolerance y* of the ith insect is an independent
drawing from a distribution identical for all insects. Moreover, if the tolerance
is a result of many independent and individually inconsequential additive
factors, we can reasonably assume y* N(u, a2) because of the central limit
theorem. Defining yi = 1 if the ith insect dies and y = 0 otherwise, we have
P(y1 = 1) = P(VP < x 1) = p)/a], (9.2.2)
giving rise to a probit model where ill = and /32 = 1/u. If, on the other
hand, we assume that yr has a logistic distribution with mean it and variance
a2, we get a logit model
gy i = 1) = A [ )c ]. (9.2.3)
V3
EXAMPLE 9.2.2 (Domencich and McFadden, 1975). Let us consider the
decision of a person regarding whether he or she drives a car or travels by
transit to work. We assume that the utility associated with each mode of
transport is a function of the mode characteristics z (mainly the time and the
cost incurred by the use of the mode) and the individual's socioeconomic
characteristics w, plus an additive error term e. We define U il and U,D as the ith
person's indirect utilities associated with driving a car and traveling by transit,
respectively. Then, assuming a linear function, we have
UK) = ao + ex,fl + mlyo + 10 (9.2.4)
and
= ai + z;ift + w;71 + ell (9.2.5)
The basic assumption is that the ith person drives a car if Un > Um and travels
by transit if Un < U,. (There is indecision if U il = U,, but this happens with
zero probability if , and e l are continuous random variables.) Thus, defining
yi = 1 if the ith person drives a car, we have
P(h= 1) = P(Un > Um ) (9.2.6)
= PRA, ell <a1 ao + (zil z,0)1 + w;(Yi Yo)]

FRai ao) + (zit ZiO + 1%171 AM,
where F is the distribution function of e x, eil . Thus a probit (logit) model
arises from assuming the normal (logistic) distribution for e io E.
9.2.2 Consistency and Asymptotic Normality of the Maximum Likelihood

Estimator
To derive the asymptotic properties of the MLE, we make the following
assumptions in the model (9.2.1).
ASSUMPTION 9.2.1. F has derivative f and second-order derivative f', and
0 < F(x) < 1 and f(x)> 0 for every x.
ASSUMPTION 9.2.2. The parameter space B is an open bounded subset of the

Euclidean K-space.
ASSUMPTION 9.2.3. (xi ) are uniformly bounded in i and lim, 1 17_ 1
ix; is a finite nonsingular matrix. Furthermore the empirical distribution of x
(xi ) converges to a distribution function.
Both probit and logit models satisfy Assumption 9.2.1. The boundedness of
(x,) is somewhat restrictive and could be removed. However, we assume this
to make the proofs of consistency and asymptotic normality simpler. ItT this
way we hope that the reader will understand the essential's without being
hindered by too many technical details.
The logarithm of the likelihood function of the model (9.2.1) is given by
log L=Iyi log F(0) + > (1 y.) log [1 F(43)1. (9.2.7)
Therefore the MLE ft is a solution (if it exists) of

a log L yi Fi
(9.2.8)
oft 13 '
F;(1 Fd fixi =
where F F(30) and f, = f(x111).
To prove the consistency ofp, we must verify the assumptions of Theorem
4.1.2. Assumptions A and B are clearly satisfied. To verify C, we use Theorem
4.2.2. If we define gi(y, /I) = [y F(x;/30 )] log F(xX), g i(y, fl) in a compact
neighborhood of /10 satisfies all the conditions for gt(y, 0) in the theorem
because of Assumptions 9.2.1 and 9.2.3. Furthermore
log F exists because of Theorem 4.2.3. Therefore pi - lPiL l yi log F converges
to F,0 log F where Fi0 = F(xX0 ), in probability uniformly in
/3 e MA), an open neighborhood of fio . A similar result can be obtained for
the second term of the right-hand side of (9.2.7). Therefore
Q(11) plim ti-1 log L (9.2.9)
= lim Fa log F + lim n- ' (1 F03 ) log (1 F;),
where the convergence is uniform in /1 E MA). Because our assumptions

enable us to differentiate inside the limit operation in (9.2.9),' we obtain
= lim Fr,a fix/

lim 1 (9.2.10)
which vanishes at /3 = /30 . Furthermore, because
a2 Q I - -urn 113 F,0 ) xix;, (9.2.11)

allafi' Fioof
which is negat've definite by our assumptions, Q attains a strict local maxi-
mum at /3 = /30 ...Thus assumption C of Theorem 4.1.2. holds, and hence the
consistency of /3 has been proved.
A solution of (9.2.8) may not always exist (see Albert and Anderson, 1984).
For example, suppose that (xi ) are scalars such that x i < 0 for i c for some
integer c between 1 and n and x i > 0 for i > c and that yi = 0 for i c and
yi = 1 for 1 > c. Then log L does not attain a maximum for any finite value of
II. If (xi ) are K-vectors, the same situation arises if y = 0 for the values of x i
RK and y == 1 for the values of x i lyinglyingwthoesdfayprlni
within the other side. However, the possibility of no solution of (9.2.8) does
not affect the consistency proof because the probability of no solution ap-
proaches 0 as n goes to co as shown in Theorem 4.1.2.
Next we shall prove asymptotic normality using Theorem 4.2.4, which
means that assumptions A, B, and C of Theorem 4.1.3 and Eq. (4.2.22) need to
be verified. Differentiating (9.2.8) with respect to /3 yields
32 log L
op"
r yi
41 If1(1 Fi )
12 2 ,
ixixi
(9.2.12)
i1 L
Thus assumption A of Theorem 4.1.3 is clearly satisfied. We can verify that

Assumptions 9.2.1 and 9.2.3 imply assumption B of Theorem 4.1.3 by using
Theorems 4.1.5 and 4.2.2. Next, we have from (9.2.8)
1 a log LI1 n yi Fio
(9.2.13)
VT? 0/) Po - E
.0(1 F io) ffflxi.
The asymptotic normality of (9.2.13) readily follows from Theorem 3.3.5
(Liapounov CLT) because each term in the summation is independent with
bounded second and third absolute moments because of Assumptions 9.2.1,
9.2.2, and 9.3.3. Thus assumption C of Theorem 4.1.3 is satisfied and we
obtain
a log
afl L L --P. N(0, A), (9.2.14)
\fi
where
1 "
, X04. (9.2.15)
n fel Fm )
Finally, we obtain from (9.2.12)
1 82 log L
hm E A, (9.2.16)
n afiair Is,
verifying (4.2.22). Therefore

I(fl fl0) 0 N(0, ). (9.2.17)
For a proof of consistency and asymptotic normality under more general
assumptions in the logit model, see the article by Gourieroux and Monfort
(1981).
9.2.3 Global Concavity of the Likelihood Function in the Logit and Probit
Models
Global concavity means that a2 log LIallar is a negative definite matrix for
fi E B. Because we have by a Taylor expansion
L(fl)= log L(fl)+
log
- a log L
(fl ft) (9.2.18)
ay
a2 log L
+ 2 (ft P)'
aflafl' p
where ft* lies between fl and fi, global concavity implies that log L(k >
log L(fl) fort/ fliffl is a solution of (9.2.8). We shall prove global concavity
for logit and probit models.
For the logit model we have
an =A(1 A) and a2n = (1 2A) an
. (9.2.19)
ax ax2 ax
Inserting (9.2.19) into (9.2.12) with F= A yields
(32 log L _
2; Ai )xix;, (9.2.20)
where A i = A(41). Thus the global concavity follows from Assumption 9.2.3.
A proof of global concavity for the probit model is a little more complicated.
Putting F. = cbi , f; = ctli , and f; = x;130i , where 4) is the density function of
N(0, 1), into (9.2.12) yields
(32 log L " (9.2.21)
- oi07-20 - o,)-2[(y, - 2mi), + 01)0,
(yi- coop -
Thus we need to show the positivity of
gy(x) = (y 2y0:13 + 0 2) + (y 0)0(1 (D)x
for y = 1 and 0. First, consider the case y = 1. Because gl (x)=

(1 413)2 (0 + ox), we need to show (1) + Cox > 0. The inequality is clearly
satisfied if x 0, so assume x <0. But this is equivalent to showing
ct > (1 (D)x for x> 0, (9.2.22)
which follows from the identity (see Feller, 1961, p. 166)
x-1 exp ( x2/2) = (1 + y -2 exp ( y2/2) dy.

) (9.2.23)
Next, if y = 0, we have go(x)= 02[0 (1 (D)x], which is clearly positive if

x 0 and is positive if x> 0 because of (9.2.22). Thus we proved global
concavity for the probit case.
9.2.4 Iterative Methods for Obtaining the Maximum Likelihood Estimator

The iterative methods we discussed in Section 4.4 can be used to calculate a
root of Eq. (9.2.8). For the logit and probit models, iteration is simple because
of the global concavity proved in the preceding section. Here we shall only
discuss the method-of-scoring iteration and give it an interesting interpreta-
tion.
As we noted in Section 4.4, the method-of-scoring iteration is defined by
A ___ A _ I [Ea 2 log Li 1-1
J
a log L
ap (9.2.24)
affair
A
where is an initial estimator of fi0 and is the second-round estimator. The
iteration is to be repeated until a sequence of estimators thus obtained con-
verges. Using (9.2.8) and (9.2.12), we can write (9.2.24) as
;42 =
f XiX; +
LI ]-1
(9.2.25)
where we have defined f",= ) and f, = fix'A).
An interesting interpretation of the iteration (9.2.25) is possible. From
(9.2.1) we obtain
yi = F(00 + ui ,
) (9.2.26)
where Eui = 0 and Vu i = F(xyl0 )[1 F(xX o )]. This is a heteroscedastic non-
linear regression model. Expanding F(xX 0 ) in a Taylor series around flo = /31
and rearranging terms, we obtain

y, P,+ ftxcji, .400 + ui . (9.2.27)
Thus ft2 defined in (9.2.25) can be interpreted as the weighted least squares
(WLS) estimator of /30 applied to (9.2.27) with Vu i estimated by F1(1 P,).
For this reason the method-of-scoring iteration in the QR model is sometimes
referred to as the nonlinear weighted least squares (NLWLS) iteration
(Walker and Duncan, 1967).
9.2.5 Berkson's Minimum Chi-Square Method

There are many variations of the minimum chi-square (MIN ;(2) method, one
of which is Berkson's method. For example, the feasible generalized least
squares (FGLS) estimator defined in Section 6.2 is a MIN ;(2 estimator. An-
other example is the Baranlcin-Gurland estimator mentioned in Section 4.2.4.
A common feature of these estimators is that the minimand evaluated at the
estimator is asymptotically distributed as chi-square, from which the name is
derived.
The MIN x2 method in the context of the QR model was first proposed by
Berkson (1944) for the logit model but can be used for any QR model. It is
useful only when there are many observations on the dependent variable y
having the same value of the vector of independent variables x (sometimes
referred to as "many observations per cell") so that F(x'A ) ) for the specific
value of x can be accurately estimated by the relative frequency of y being
equal to 1.
To explain the method in detail, we need to define new symbols. Suppose
that the vector x i takes T distinct vector values x (1) , x(2) , . . , x(r) ; and
classify integers (1, 2, . . . , n) into Tdisjoint sets II , 12 . . I T by the rule:
i E I, if xi = xio Define P(1) = P(y, = 1) if i E I, . In addition, define n, as the
number of integers contained in I,, r, = iEJ y and P(t)= r,/n,. Note that
(Pio ) constitute the sufficient statistics of the model. In the following discus-
sion we shall write x (,) , P(0 , and P(s) as x P and P, if there is no ambiguity.
P, = F(x;110 ), t = 1, 2, . . . , T. (9.2.28)
If F is one-to-one (which is implied by Assumption 9.2.1), we can invert the
relationship in (9.2.28) to obtain
F- i(P)= xfi0 , (9.2.29)
where denotes the inverse function of F. Expanding F-1(P1 ) in a Taylor

series around P, (which is possible under Assumption 9.2.1), we obtain
aF-11 .-
F- 1(150=xxo + (P, P,) (9.2.30)
apt pr
= +
I EF- (Pr)] (P, Pt)
xg$0 + vt + w, ,
where II lies between Pt and Pr,
1
vt= (Pt PO,
.1.[P'(PIA
and
1
Wt =
G[F-1(Pn] [F-1 (P 1)]) fi Pt).
The fact that v, and w, depend on n has been suppressed.
Because
Vv, a2 (9.2.31)
nil 2 [F- qP,)]
and because w, is 0(nT I ) and hence can be ignored for large n, (as we shall
show rigorously later), (9.2.30) approximately defines a heteroscedastic linear
regression model. The MIN x2 estimator, denoted A is defined as the WLS
estimator applied to (9.2.30) ignoring w,. We can estimate a? by "a? obtained
by substituting P, for P, in (9.2.31). Thus
-1 T
11 &-2X X') I Er-t2X,F-1 (A). (9.2.32)
r- 1-1
We shall prove the consistency and asymptotic normality ofi(as n goes to co

with T fixed) under Assumptions 9.2.1, 9.2.3, and the following additional
assumption:
ASSUMPTION 9.2.4. lim b...,. (n,In)= c, # 0 for every t = 1, 2, . . . , T,
where T is a fixed integer.
We assume statement 9.2.4 to simplify the analysis. However, if c, = 0 for
some t, we can act as if the observations corresponding to that t did not exist
and impose Assumptions 9.2.3 and 9.2.4 on the remaining observations.
Inserting (9.2.30) into (9.2.32) and rearranging terms yield

1 T-1 1T A
flO) (- CrT 2Xtlet)-
r- ai 2x,(1), + we). (9.2.33)
n
Because T is fixed, we have by Theorem 3.2.6

r
plim - E i. iT2x clet = E 2x , x , (9.2.34)
n-. n t-1 t-1
where .572 = cf f 2[F-1(1)1 )]1[P1(1 - P1 )]. Also, using Theorem 3.2.7, we
obtain
1 T1 A A
E a72x(vi + w1 )= E vn
vn t-t 1-1
(9.2.35)
X(oliv,+ aTlwe )
-
= oVx,o7 1 1;
because plim n 112& = Ere- plim = 1, and plim aw = 0. But, be-

cause (v1 ) are independent, the vector (or 1V1, 11)2 . , (7-T iv T ) converges
to N(0, I T). Therefore
T T
o-T ix,o7 've N 0, Ert- 2x (9.2.36)
I-. I t I
Finally, we obtain from (9.2.33) through (9.2.36) and Theorem 3.2.7
-1
Nr-n- (PO) N[0,( ii.T2x,x;) ]. (9.2.37)
Because A defined in (9.2.15) is equal to Ir.., irT 2x1x't under Assumption

9.2.4, (9.2.17) and (9.2.37) show that the MIN x2 estimator has the same
asymptotic distribution as the MLE. 2 The MIN x2 estimator is simpler to
compute than the MLE because the former is explicitly defined as in (9.2.32),
whereas the latter requires iteration. However, the MIN x2 method requires a
large number of observations in each cell. If the model contains several inde-
pendent variables, this may be impossible to obtain. In the next subsection we
shall compare the two estimators in greater detail.
In the probit model, F -1(131 ) 01)-1 (fi1 ). Although (1)- ' does not have an
explicit form, it can be easily evaluated numerically. The function crr - '( ) is
called the probit transformation. In the logit model we can explicitly write
) = log [P,/(1 (9.2.38)

which is called the logit transformation or the logarithm of the odds ratio.
Cox (1970, P. 33) mentioned the following modification of (9.2.38):
= log {[P + (2n,)-1 ]/[1 (2n,)-1 ]). (9.2.39)
This modification has two advantages over (9.2.38): (1) The transformation
(9.2.39) can always be defined whereas (9.2.38) cannot be defined if /5,= 0 or
1. (Nevertheless, it is not advisable to use Cox's modification when this small.)
(2) It can be shown that EA; 1 (/),) ) is of the order of n7 2, whereas
EA- A-1(P1 ) is of the order of nT
In the preceding passages we have proved the consistency and the asymp-
totic normality of Berkson's MIN X2 estimator assuming x i = x, for i E
However, a situation may occur in practice where a researcher must proceed
as if x i = x, for i E I, even if x i # xt because individual observations x i are not
available and only their group mean x i n7 12,,E4xi is available. (Such an
example will be given in Example 9.3.1.) In this case the MIN X2 estimator is
generally inconsistent. McFadden and Reid (1975) addressed this issue and,
under the assumption of normality of x evaluated the asymptotic bias of the
MIN X2 estimator and proposed a modification of the MIN X2 estimator that
is consistent.
9.2.6 Comparison of the Maximum Likelihood Estimator and the Minimum

Chi-Square Estimator
In a simple model where the vector x, consists of 1 and a single independent
variable and where T is small, the exact mean and variance of MLE and the
MIN x2 estimator can be computed by a direct method. Berkson (1955, 1957)
did so for the logit and the probit model, respectively, and found the exact
mean squared error of the MIN x2 estimator to be smaller in all the examples
considered.
Amemiya (1980b) obtained the formulae for the bias to the order of n-1 and
the mean squared error to the order of n -2 of MLE and MIN X2 in a general
logit mode1. 3 The method employed in this study is as follows: Using (9.2.19)
and the sampling scheme described in Section 9.2.5, the normal equation
(9.2.8) is reduced to
I-1 n,[P, A(xX)]x,= 0. (9.2.40)

We can regard (9.2.40) as defining the MLE /I implicitly as a function of

P1, P2, , PT' say, fi = g(A , P2, y Pr). Expanding g in a Taylor
series around P i , P2,. . . , PT and noting that g(P i , P2, ) PT) = /30 we
,
obtain
nA t t 2 5 5
gutu, (9.2.41)
1
+ , E 1 E gurueusu,,
um
where u, = Pt - P, and g g, and g m denote the first-, second-, and third-
order partial derivatives of g evaluated at (P 1 , F2,. . . , PT), respectively.
The bias of the MLE to the order of n -1 is obtained by taking the expectation
of the first two terms of the right-hand side of (9.2.41). The mean squared error
of the MLE to the order of Pr-2 is obtained by calculating the mean squared
error of the right-hand side of (9.2.41), ignoring the terms of a smaller order
than n-2. We need not consider higher terms in the Taylor expansion because
E4 for k* 5 are at most of the order of nT 3. A Taylor expansion for the
MIN x2 estimator /I is obtained by expanding the right-hand side of (9.2.32)
around P,.
Using these formulae, Amemiya calculated the approximate mean squared
errors of MLE and the MIN x2 estimator in several examples, both artificial
and empirical, and found the MIN x2 estimator to have a smaller mean
squared error than MLE in all the examples considered. However, the differ-
ence between the two mean squared error matrices can be shown to be neither
positive definite nor negative definite (Ghosh and Sinha, 1981). In fact, Davis
(1984) showed examples in which the MLE has a smaller mean squared error
to the order of n offered an intuitive argument that showed that the
greater T, the more likely MLE is to have a smaller mean squared error.
Amemiya also derived the formulae for the n -2-order mean squared errors
of the bias-corrected MLE and the bias-corrected MIN x2 estimator and
showed that the former is smaller. The bias-corrected MLE is defined as
ft B(ft), where B is the bias to the order of n -1 , and similarly for MIN x2 .
This result is consistent with the second-order efficiency of MLE in the expo-
nential family proved by Ghosh and Subramanyam (1974), as mentioned in
Section 4.2.4. The actual magnitude of the difference of the n -2-order mean
squared errors of the bias-corrected MLE and MIN ein Amemiya's examples
was always found to be extremely small. Davis did not report the correspond-
ing results for her examples.
Smith, Savin, and Robertson (1984) conducted a Monte Carlo study of a

logit model with one independent variable and found that although in point
estimation MIN x2 did better than MLE, as in the studies of Berkson and
Amemiya, the convergence of the distribution of the MIN x2 estimator to a
normal distribution was sometimes disturbingly slow, being unsatisfactory in
one instance even at n = 480.
For further discussion of this topic, see the article by Berkson (1980) and the
comments following the article.

To test a hypothesis on a single parameter, we can perform a standard normal
test using the asymptotic normality of either MLE or the MIN x2 estimator. A
linear hypothesis can be tested using general methods discussed in Section
4.5.1. The problem of choosing a model among several alternatives can be
solved either by the Akaike Information Criterion (Section 4.5.2) or by Cox's
test of nonnested hypotheses (Section 4.5.3). For other criteria for choosing
models, see the article by Amemiya (1981).
Here we shall discuss only a chi-square test based on Berkson's MIN ;(2
estimator as this is not a special case of the tests discussed in Section 4.5. The
test statistic is the weighted sum of squared residuals (WSSR) from Eq.
(9.2.30) defined by
WSSR = ii72[F-1 (fi,) xj3]2. (9.2.42)

t-1
In the normal heteroscedastic regression model y N(Xfi, D) with known D,
(y '(y X/SG ) is distributed as xi_ K . From this fact we can deduce
that WSSR defined in (9.2.42) is asymptotically distributed as Xr-x
We can use this fact to choose between the unconstrained model P,---
F(x10 ) and the constrained model P, F(xcitio ), where x i, and filo are the
first K q elements of; and fin, respectively. Let WSSR. and WSSIt c be the
values of (9.2.42) derived from the unconstrained and the constrained
models, respectively. Then we should choose the unconstrained model if and
only if
WSSR. WSSR > (9.2.43)
where 4. denotes the a% critical value of4 . Li (1977) has given examples of
the use of this test with real data.
To choose among nonnested models, we can use the following variation of
the Akaike Information Criterion:

AIC = WSSR + K. (9.2.44)
This may be justified on the grounds that in the normal heteroscedastic
regression model mentioned earlier, (y XfiG )'irAy XfiG ) is equal to

2 log L aside from a constant term.

Instead of (9.2.42) we can also use
I n t[Pi(1
1-1
RxX)] 2, (9.2.45)
because this is asymptotically equivalent to (9.2.42).
9.2.8 Discriminant Analysis

The purpose of discriminant analysis is to measure the characteristics of an
individual or an object and, on the basis of the measurements, to classify the
individual or the object into one of two possible groups. For example, accept
or reject a college applicant on the basis of examination scores, or determine
whether a particular skull belongs to a man or an anthropoid on the basis of its
measurements.
We can state the problem statistically as follows: Supposing that the vector
of random variables x* is generated according to either a density g 1 or go , we
are to classify a given observation on x*, denoted x* , into the group character-
ized by either g 1 or go . It is useful to define yi = 1 if xf' is generated by g1 and
y, = 0 if it is generated by g0 . We are to predict y i on the basis of xt. The
essential information needed for the prediction is the conditional probability
P(yi = 114). We shall ignore the problem of prediction given the conditional
probability and address the question of how to specify and estimate the condi-
tional probability. 4
By Bayes's rule we have
P(yi = 1lx?)
(x11 )qi (9.2.46)
g1(x)q1 + go(x?)qo'
where q1 and qo denote the marginal probabilities P(y1 = 1) and P(y1 = 0),
respectively. We shall evaluate (9.2.46), assuming that g 1 and go are the densi-
ties of N(it, , ) and N(/40 , X0 ), respectively. We state this assumption for-
mally as
xr = 1) II) (9.2.47)
x*KY = 0)
This is the most commonly used form of discriminant analysis, sometimes

referred to as normal discriminant analysis.
Under (9.2.47), (9.2.46) reduces to the following quadratic logit model:
P(yi = 114) = A( A D + + xrAxn, (9.2.48)
where
fl(I) = + log qi log (10 (9.2.49)
1
- -2 log 'III+ -2 log 'Id,
/3(2) Ic7 1/10, (9.2.50)

and
A = (XV - (9.2.51)
In the special case I t = 10 , which is often assumed in econometric applica-

tions, we have A = 0; therefore (9.2.48) further reduces to a linear logit model:
P(yi = 114) = A(xX), (9.2.52)
where we have written flo) + = xy3 to conform with the notation of

Section 9.2.1.
Let us consider the ML estimation of the parameters,u i uo ,Z Io , ql , and
sq0 based on observations (y 1 , 4), i = 1, 2, . . . , n. The determination of q l
c, varies with authors. We shall adopt the approach of Warner (1963) and andq
treat ql and qo as unknown parameters to estimate. The likelihood function
can be written as
L = H [,g1 (4)qI ]Yi[g0(4)q0 ] 1-Yi. (9.2.53)
Equating the derivatives of log L to 0 yields the following ML estimators:
(9.2.54)
where n 1 =
_ no (9.2.55)
go -
where no = n n, ,
in
ni
I (9.2.56)
in
(1 (9.2.57)
no
i n
= yi(x* 111)(4 111)', (9.2.58)
and
-
=
I .0 in., I
( 1 - Yi) (xli - 410)(4 - 110)' - (9.2.59)
If = I) as is often assumed, (9.2.58) and (9.2.59) should be re-

placed by
n
1 = '2;2& i I Yi(X t
[ )(xi' )' (9.2.60)
( 1 i) (X 7
1
The ML estimators of/3(1) , flo) , and A are obtained by inserting these estimates
into the right-hand side of (9.2.49), (9.2.50), and (9.2.51).
Discriminant analysis is frequently used in transport modal choice analysis.
See, for example, articles by Warner (1962) and McGillivray (1972).
We call the model defined by (9.2.47) with Z = 1 0 and by (9.2.52) the
discriminant analysis (DA) model and call the estimator of /3 (flo) , g2) )'
obtained by inserting (9.2.56), (9.2.57), and (9.2.60) into (9.2.49) and (9.2.50)
with I I = Io the DA estimator, denoted /3DA . In contrast, if we assume only
(9.2: 52) and not (9.2.47), we have a logit model. We denote the logit MLE offi
by PA. In the remainder of this section we shall compare these two estimators.
The relative performance of the two estimators will critically depend on the
assumed true distribution for xr. If (9.2.47) with I I = I o is assumed in addi-
tion to (9.2.52), the DA estimator is the genuine MLE and therefore should be
asymptotically more efficient than the logit MLE. However, if (9.2.47) is not
assumed, the DA estimator loses its consistency in general, whereas the logit
MLE retains its consistency. Thus we would expect the logit MLE to be more
robust.
Efron (1975) assumed the DA model to be the correct model and studied the
loss of efficiency that results if /I is estimated by the logit MLE. He used the
asymptotic mean of the error rate as a measure of the inefficiency of an
estimator. Conditional on a given estimator f3 (be it /IDA or /jA ), the error rate is
defined by
Error Rate = P[x'ft 0 I x IA% (9.2.61)
P[x'n < Olx - N(u 1 , Mg,
g0 o[(inft)-1/2 140] q1 CD[-- (friii) 1/2 An] .
Efron derived the asymptotic mean of (9.2.61) for each of the cases fi = PDA
and ft = PA, using the asymptotic distributions of the two estimators. Defining
the relative efficiency of the logit ML estimator as the ratio of the asymptotic
mean of the error rate of the DA estimator to that of the logit ML estimator,
Efron found that the efficiency ranges between 40 and 90% for the various
experimental parameter values he chose.
Press and Wilson (1978) compared the classification derived from the two
estimators in two real data examples in which many of the independent
variables are binary and therefore clearly violate the DA assumption (9.2.47).
Their results indicated a surprisingly good performance by DA (only slightly
worse than the logit MLE) in terms of the percentage of correct classification
both for the sample observations and for the validation set.
Amemiya and Powell (1983), motivated by the studies of Efron, Press, and
Wilson, considered a simple model with characteristics similar to the two
examples of Press and Wilson and analyzed it using the asymptotic techniques
analogous to those of Efron. They compared the two estimators in a logit
model with two binary independent variables. The criteria they used were the
asymptotic mean of the probability ofcorrect classification (PCC) (that is, one
minus the error rate) and the asymptotic mean squared error. They found that
in terms of the PCC criterion, the DA estimator does very well only slightly
worse than the logit MLE, thus confirming the results of Press and Wilson. For
all the experimental parameter values they considered, the lowest efficiency of
the DA estimator in terms of the PCC criterion was 97%. The DA estimator
performed quite well in terms of the mean squared error criterion as well,
although it did not do as well as it did in terms of the PCC criterion and it did
poorly for some parameter values. Although the DA estimator is inconsistent
in the model they considered, the degree of inconsistency (the difference
between the probability limit and the true value) was surprisingly small in a
majority of the cases. Thus normal discriminant analysis seems more robust
against nonnormality than we would intuitively expect.
We should point out, however, that their study was confined to the case of
binary independent variables; the DA estimator may not be robust against a
different type of nonnormality. McFadden (1976a) illustrated a rather signifi-
cant asymptotic bias of a DA estimator in a model in which the marginal
distribution of the independent variable is normal. [Note that when we spoke
of normality we referred to each of the two conditional distributions given in
(9.2.47). The marginal distribution of x* is not normal in the DA model but,
rather, is a mixture of normals.] Lachenbruch, Sneeringer, and Revo (1973)
also reported a poor performance of the DA estimator in certain nonnormal
models.
9.2.9 Aggregate Prediction
We shall consider the problem of predicting the aggregate proportion r

yi in the QR model (9.2.1). This is often an important practical prob-
lem for a policymaker. For example, in the transport choice model of Example
9.2.2, a policymaker would like to know the proportion of people in a commu-
nity who use the transit when a new fare and the other values of the indepen-
dent variables x prevail. It is assumed that fl (suppressing the subscript 0) has
been estimated from the past sample. Moreover, to simplify the analysis, we
shall assume for the time being that the estimated value offl is equal to the true
value.
The prediction of r should be done on the basis of the conditional distribu-
tion of r given {x 1 ). When n is large, the following asymptotic distribution is
accurate:
r N[n- I F1 , n-2 F.(1 (9.2.62)
Once we calculate the asymptotic mean and variance, we have all the neces-
sary information for predicting r.
If we actually observe every x i , i = 1, 2, . . . , n, we can calculate the mean
and variance for (9.2.62) straightforwardly. However, because it is more real-
istic to assume that we cannot observe every x i , we shall consider the problem
of how to estimate the mean and variance for (9.2.62) in that situation. For
that purpose we assume that (x 1 ) are i.i.d. random variables with the common
K-variate distribution function G. Then the asymptotic mean and variance of
r can be estimated by EF(xl) and n- IEF(1 F), respectively, where E is the
expectation taken with respect to G.
Westin (1974) studied the evaluation of EF when F = A (logistic distribu-
tion) and xix, ix). He noted that the density of p = A(0) is given by
1 1 2
RP) exp { -- log ) , (9.2.63)
P( 1 p) 2a2 [ Pp
1
where it = fl and a2 = nj. Because the mean of this density does not
have a closed form expression, we must evaluate f pf(p) dp numerically for
given values of /.1 and a2.
McFadden and Reid (1975) showed that if F= (211 (standard normal distri-
bution) and x Ix ) as before, we have
E(x11) = CDR1 02)-1124 (9.2.64)
Thus the evaluation of EF is much simpler than in the logit case.
Neither Westin nor McFadden and Reid considered the evaluation of the
asymptotic variance of r, which constitutes an important piece of information
for the purpose of prediction.
Another deficiency of these studies is that the variability due to the estima-
tion of /I is totally ignored. We shall sugfest a partially Bayesian way to deal
with this problem. 5 Given an estimate /3 of fi, we treat ft as a random varia-
ble with the distribution mft,I p). An estimate of the asymptotic covariance
matrix of the estimator ft can be used for Z . We now regard (9.2.62) as the
,
asymptotic distribution of r conditionally on (x i ) and ft. The distribution of r

conditionally on {x i} but not on 11 is therefore given by
A
r N[E I Fa , Eprz-2 F,(1 Fi ) (9.2.65)
I
+ Vier' E F1 ].
t I
Finally, if the total observations on (x i ) are not available and (x i) can be

regarded as i.i.d. random variables, we can approximate (9.2.65) by
r N[ExEpF, rz- lEEpF(1 F)+ ElVpF]. (9.2.66)
9.3 Multinomial Models
9.3.1 Statistical Inference

In this section we shall define a general multinomial QR model and shall
discuss maximum likelihood and minimum chi-square estimation of this
model, along with the associated test statistics. In the subsequent sections we
shall discuss various types of the multinomial QR model and the specific
problems that arise with them.
Assuming that the dependent variable yi takes in ; + 1 values 0, 1, 2, .
mi , we define a general multinomial QR model as
i= 1, 2, . . . ,n and (9.3.1)
j = 1, 2, . . . , m,,
where x* and 0 are vectors of independent variables and parameters, respec-
tively. (Strictly speaking, we should write j as j i , but we shall suppress the
subscript i.) We sometimes write (9.3. 1) simply as Pu = Fu . We shall allow the
possibility that not all the independent variables and parameters are included
in the argument of every F. Note that P(yi = 0) ( ) need not be specified
because it must be equal to one minus the sum of the m i probabilities defined
in (9.3.1).
It is important to let m i depend on i because in many applications individ-
uals face different choice sets. For example, in transport modal choice analy-
sis, traveling by train is not included in the choice set of those who live outside
of its service area.
To define the maximum likelihood estimator of 0 in the model (9.3.1) it is
useful to define 7... 1 (m i + 1) binary variables
y1 = 1 if Yi =./ (9.3.2)
=0 if yi # j, i= 1, 2, . . . ,n and
j 0, 1, . . , m 1 .
Then we can write tl* log likelihood function as
n m,
log L = I yu log Fo , (9.3.3)
p- fi-o
a
which is a natural generalization of (9.2.7). The MLE of 0 is defined as a
solution of the normal equation 0 log Llae = 0.
Many of the results about the MLE in the binary case hold for the model
(9.3.1) as well. A reasonable set of sufficient conditions for its consistency and
asymptotic normality can be found using the relevant theorems of Chapter 4,
as we have done for the binary case. The equivalence of the method of scoring
and the NLWLS (NLGLS to be exact) iteration can also be shown. However,
we shall demonstrate these things under a slightly less general model than
(9.3.1). We assume6
P(1, r= .i) x;2132, . x;Hfix), (9.3.4)

i = 1, 2, . . . ,n and j = 1, 2, . . . , m,
where H is a fixed integer. In most specific examples of the multinomial QR
model we have H = m, but this need not be so for this analysis. Note that we
have now assumed that m does not depend on i to simplify the analysis.
We restate Assumptions 9.2.1, 9.2.2, and 9.2.3 with obvious modifications
as follows.
ASSUMPTION 9.3.1. F, has partial derivatives ft 8F1 /a(x2hfth ) and partial
second-order derivatives g afieva(x;ifid for every i, j, k, 1, and 0 < F, < 1
and pi, > 0 for every i, j, and k.
ASSUMPTION 9.3.2. The parameter space B is an open bounded subset of a
Euclidean space.
ASSUMPTION 9.3.3. {x ih ) are uniformly bounded in i for every h and
limn, n II7_,xigeo, is a finite nonsingular matrix for every h. Furthermore,
the empirical distribution function of {x il, ) converges to a distribution func-
tion.
Under these assumptions the MLE n of ft = . . . , "lir can be
shown to be consistent and asymptotically normal. We shall derive its asymp-
totic covariance matrix.
Differentiating (9.3.3) with respect go fi h, we obtain
a log L ^m (9.3.5)
YU FTl iffiX ilv
aPk P-C1
Differentiating (9.3.5) with respect to /r/ yields

n m
a2 log L_
Yif-6 2.111jfiljXikX;I (9.3.6)
fr.1
n m
+ I Iy , FIfi,ix akx;,.
1..1 J-0
Taking the expectation and noting I.71.0 ft! = 0, we obtain

(32 log L "m
E (9.3.7)
uPkuP1
Define A = (Aki). Then we can show under Assumptions 9.3.1, 9.3.2, and
9.3.3
-
J(fl fi) N(0, lim

n.ao
(9.3.8)
To show the equivalence between the method-of-scoring and the NLWLS

iteration (see Amemiya, 1976b), it is convenient to rewrite (9.3.5) and (9.3.7)
using the following vector notation:
Yi (Y11 Yi2, Ytm)' (9.3.9)
F, IL' , Fa , . . , Fo)'
f (Ail,
Ai = E(yi Fi )(yi Fi )' = D(F; ) FiF;,
where D(Fi ) is the diagonal matrix the jth diagonal element of which is F
Then, using the identities 0Fii = 1 and IAA = 0 and noting
AT' = D-1(Fi ) + (9.3.10)
where 1 is an m-vector of ones, we obtain from (9.3.5) and (9.3.7)
a log L (9.3.11)
afik Xxikv," Any Fi)
and
02 log L "
E x k fl-" A7lf (9.3.12)
aflkafl; " "
Suppose # is the initial estimate fp, let ri, it, and Xi be, respectively, F i , f,
and A; evaluated at fl and define
n _ _
Aid = Xikf li"AT (9.3.13)
and
E xikfrki + i). (9.3.14)

r-i
=
Then the second-round estimator fl obtained by the method-of-scoring itera-
tion started from # is given by
fl= (9.3.15)
where A = (A-k,) and C. = (Z't , ..,
We can interpret (9.3.15) as an NLGLS iteration. From (9.3.4) we obtain
yi = Fi + ui , i = 1, 2, . . . , n, (9.3.16)
where Eu i =0, Eu u = Ai , and Euiu; =0 for i j. Expanding F. in Taylor
,
series around xdih , h = 1, 2, . . . , H, we obtain
+ flpah -= filxj + u (9.3.17)

A-1
i = 1, 2, . . . , n.
The approximation (9.3.17) can be regarded as a linear regression model of
mn observations with a nonscalar covariance matrix. If we estimate A i by Ai
andplyFGLSto(9.317),hersulingtmaopecsly(9.315)
Next we shall derive the MIN x2 estimator of ft in model (9.3.4), following
Amemiya (1976b). As in Section 9.2.5, we assume that there are many obser-
vations with the same value of independent variables. Assuming that the
vector of independent variables takes T distinct values, we can write (9.3.4) as
x;2132, x;HitH), (9.3.18)
t = 1, 2, . . . , T and j = 1, 2, . . . , m.
To define the MIN x2 estimator offl, we must be able to invert m equations in
(9.3.18) for every t and solve them for H variables xfl1 , 342/12 2 X ;Hfikl
Thus for the moment, assume H = m. Assuming, furthermore, that the Jaco-
bian does not vanish, we obtain
X tdrflk = ill Pa) I Pun), (9.3.19)
t = 1, 2, . . . , T and k = 1, 2, . . . , m.
As in Section 9.2.5 we define ro = Z iE/y o , where I, is the set of t for which
xih = xth for all h, and Po = roin where n, is the number of integers in I.
Expanding Gh(Pil , Pa , . . . , Fn.) in Taylor series around (P, 1 , Pa,.
P,,) and using (9.3.19), we obtain
Gk(Pa, P,,. Am) xV3k+ gia(Pif - F,), (9.3.20)
t = 1, 2, . . . , T and k 1, 2, . . . , m,
where elk = aGidaPti. Equation (9.3.20) is a generalization of (9.2.30), but
here we write it as an approximate equation, ignoring an error term that
corresponds to w, in (9.2.30). Equation (9.3.20) is an approximate linear
regression equation with a nonscalar covariance matrix that depends
on A, E(f', Ft )(f), F,)' = ni- l[D(Fd F,F], where P, = (pa , Pa ,

, Pt,,,)' and Ft = (Ft! , Fa , . , Ft,)'. The MIN X' estimator /I of fi is
defined as FGLS applied to (9.3.20), using A, = riT'[D(P,) Ptil] as an
estimator of A,. An example of the MIN x2 estimator will be given in Example

9.3.1 in Section 9.3.2.
The consistency and the asymptotic normality of ncan - be proved by using a
method similar to the one employed in _Section 9.2.5. Let fl' the asymp-
totic covariance matrix offi, that is, .iTz(/3 fl) N(0, lim, nL). Then it
can be deduced from (9.3.20) that the k,/th subblock of LI is given by
= xtk(GrAtG,)'x,, (9.3.21)
where G; is an m X m matrix the kth row of which is equal to (glk , g,. . ,

gli ) and ( rk,1 denotes the k, /th element of the inverse of the matrix inside
( ). Now, we obtain from (9.3.12)
Aid= E xd,(F;A,- 1F.t ),dx;,, (9.3.22)
1
where F; is an m X m matrix the kth row of which is equal to Cnk, fik,

ptD. Thus the asymptotic equivalence of MIN x2 and MLE follows from the
identity G T' = C.
In the preceding discussion we assumed H ='m. If H < m, we can still
invert (9.3.18) and obtain
x i'kfik = Gk(Ptl 9 Pa Pun), (9.3.23)
t = 1, 2, . . . , T and k= 1, 2, . . . , H,
but the choice of function Gk is not unique. Amemiya (1976b) has shown that
the MIN x2 estimator is asymptotically efficient only when we choose the
correct function Gk from the many possible ones. This fact diminishes the
usefulness of the method. For this case Amemiya (1977c) proposed the follow-
ing method, which always leads to an asymptotically efficient estimator.
Step 1. Use some Gk in (9.3.23) and define
Ptk = Gda Pi2, 9
Step 2. In (9.3.17) replace Fi and ill by Ft and ffi evaluated at A ik and replace
x;kflh by ,k.
Step 3. Apply FGLS on (9.3.17) using At evaluated at A sk .
We shall conclude this subsection by generalizing WSSR defined by (9.2.42)
and (9.2.45) to the multinomial case.
We shall not write the analog of (9.2.42) explicitly because it would require
cumbersome notation, although the idea issimple. Instead, we shall merely
point out that it is of the form Or )1 - i (Y MG) obtained from the
regression equation (9.3.20). It is asymptotically distributed as chi-square with
degrees of freedom equal to mT minus the number of regression parameters.
Next, consider the generalization of (9.2.45). Define the vector F, = F 1(1).
Then, the analog of (9.2.45) is
WSSR = (P, Fd'AT l(P, F,) (9.3.24)
= n,(P, F,)' [D(P,) -1 + /1701 11'](15,

p-1
T m (13 pt.)2
= n, "
t-1 j-O Pti
It is asymptotically equivalent to the analog of (9.2.42).
9.3.2 Ordered Models

Multinomial QR models can be classified into ordered and unordered models.
In this subsection we shall discuss ordered models, and in the remainder of
Section 9.3 we shall discuss various types of unordered models.
A general definition of the ordered model is
DEFINITION 9.3.1. The ordered model is defined by
PO' = = P(Si)
for some probability measure p depending on x and 0 and a finite sequence of
successive intervals (Si) depending on x and 0 such that UfS., = R, the real line.
A model is unordered if it is not ordered. In other words, in the ordered
model the values that y takes correspond to a partition of the real line, whereas
in the unordered model they correspond either to a nonsuccessive partition of
the real line or to a partition of a higher-dimensional Euclidean space.
In most applications the ordered model takes the simpler form
P(y = fix, a, /I) = F(cti41 F(ai x73), (9.3.25)
j = 0, 1, . . , m, ao = co, af + 1, am+i = co,
for some distribution function F. If F = 41), (9.3.25) defines the ordered probit
model; and if F = A, it defines the ordered logit model. Pratt (1981) showed
that the log likelihood function of the model (9.3.25) based on observations
(y i , i = 1, 2, . . . , n, on (y, x) is globally concave iff, derivative of F, is
positive and log f is concave.
The model (9.3.25) is motivated by considering an unobserved continuous
random variable y* that determines the outcome of y by the rule
y = j if and only if a < y* < ai+1 , (9.3.26)
j= 0, 1, . . . , m.
If the distribution function of y* x'fl is F, (9.3.26) implies (9.3.25).
In empirical applications of the ordered model, y* corresponds to a certain
interpretative concept. For example, in the study of the effect of an insecticide
by Gurland, Lee, and Dahm (1960), y* signifies the tolerance of an insect
against the insecticide. Depending on the value of y*, y takes three discrete
values corresponding to the three states of an insectdead, moribund, and
alive. In the study by David and Legg (1975), y* is the unobserved price of a
house, and the observed values of y correspond to various ranges of the price of
a house. In the study by Silberman and Talley (1974), y* signifies the excess
demand for banking in a particular location and y the number of chartered
bank offices in the location. See also Example 9.4.1 in Section 9.4.1.
The use of the ordered model is less common in econometric applications
than in biometric applications. This must be due to the fact that economic
phenomena are complex and difficult to explain in terms of only a single
unobserved index variable. We should be cautious in using an ordered model
because if the true model is unordered, an ordered model can lead to serious
biases in the estimation of the probabilities. On the other hand, the cost of
using an unordered model when the true model is ordered is a loss of efficiency
rather than consistency.
We shall conclude this subsection by giving an econometric example of an
ordered model, which is also an interesting application of the MIN x2 method
discussed in Section 9.3.1.
EXAMPLE 9.3.1 (Deacon and Shapiro, 1975). In this article Deacon and
Shapiro analyzed the voting behavior of Californians in two recent referenda:
Rapid Transit Initiative (November 1970) and Coastal Zone Conservation
Act (November 1972). We shall take up only the former. Let AU i be the
difference between the utilities resulting from rapid transit and no rapid
transit for the ith individual. Deacon and Shapiro assumed that AU, is distrib-
uted logistically with mean that is, P(AU, < x) = A(x p, ) and that
the individual vote is determined by the rule:

Vote yes if A Ui > 1 (9.3.27)
Vote no if Ui <-61
Abstain otherwise.
Writing P ,(Y ) and P(N) for the probabilities that the ith individual votes yes
and no, respectively, we have
Pa(Y)= /Vizi 61) (9.3.28)
and
P ,(N) = A( p i 60. (9.3.29)
Deacon and Shapiro assumed p i = x;f11 and 51 = x132 , where x i is a vector of
independent variables and some elements offl i and 132 are a priori specified to
be zeros. (Note that if (5, 0, the model becomes a univariate binary logit
model.)
The model (9.3.27) could be estimated by MLE if the individual votes were
recorded and xi were observable. But, obviously, they are not: We only know
the proportion of yes votes and no votes in districts and observed average
values of x i or their proxies in each district. Thus we are forced to use a method
suitable for the case of many observations per cell.' Deacon and Shapiro used
data on 334 California cities. For this analysis it is necessary to invoke the
assumption that xi = x, for all i E I where I, is the set of individuals living in
the tth city. Then we obtain from (9.3.28) and (9.3.29)
PAY) x;(fli /32) (9.3.30)
log
1 Pi(Y)
and
log PAN )' x,(81 + p2 ). (9.3.31)

1 P,(N)
Let /5,( Y) and Pi(N) be the proportion of yes and no votes in the tth city. Then,
expanding the left-hand side of (9.3.30) and (9.3.31) by Taylor series around
P,(Y) and P,(N), respectively, we obtain the approximate regression equations
Y15X (9.3.32)
log .- Y) 34(fil fi2)
1 P,(
1
Ei+ 51( Y) PAY)]
P1(17)(1
and
log P(N)
- x'f(fli + /32) (9.3.33)
1 PO )
1
+ PANN1 P(N)] [PAN) PAN)].
Note that (9.3.32) and (9.3.33) constitute a special case of (9.3.20). The error
terms of these two equations are heteroscedastic and, moreover, correlated
with each other. The covariance between the error terms can be obtained from
the result Coy [P( Y), P(N)] = n 'P( Y)PAN). The MIN estimates of ;( 2
(/31 -12) and (/31 + /32) are obtained by applying generalized least squares to
(9.3.32) and (9.3.33), taking into account both heteroscedasticity and the
correlation.
9.3.3 Multinomial Logit Model

In this and the subsequent subsections we shall present various types of unor-
dered multinomial QR models. The multinomial logit model is defined by
Mg - 1 (9.3.34)
= exp (x;kfid exp (KA,
kO
i = 1, 2, . . . , n and -= 0, 1, . .
where we can assume x x, =0 without loss of generality. The log likelihood
function is given by
n
log L = I yi; log Pt,. (9.3.35)
EI fr=o
Following McFadden (1974), we shall show the global concavity of (9.3.35).
Differentiating (9.3.35) with respect to A we obtain
a log L _ vzz.
v ii (9.3.36)
oft Pii '
where 1i and ; denote E7.. 1 and Epo , respectively. Differentiating (9.3.36)
further yields
82 log L _ [ a2Pu i a1aP,J
(9.3.37)
8149' oflafli Py Op Oft'
Next, differentiating (9.3.34), we obtain after some manipulation

aP.
afi = Pa.1 (xi.1. (9.3.38)
where = [Ek exP (xiall)]-1 Ek exP (x'ikfl)x ik , and

a2pi .
alga" P,;(xu Fc" (9.3.39)
exp (x'a/3) exp (x'll11)xikx'ik + PAZ; .
Inserting (9.3.38) and (9.3.39) into (9.3.37) yields

a2 log L
(9.3.40)
OM'
which, interestingly, does not depend on h. Because Pu > 0 in this model, the
matrix (9.3.40) is negative definite unless (xi; ii)'a = 0 for every i and j for

some a # 0. Because such an event is extremely unlikely, we can conclude for

all practical purposes that the log likelihood function is globally concave in the
multinomial logit model.
We shall now discuss an important result of McFadden (1974), which shows
how the multinomial logit model can be derived from utility maximization.
Consider for simplicity an individual i whose utilities associated with three
alternatives are given by
Uu = /.4,) + , j = 0, 1, and 2, (9.3.41)
where pi, is a nonstochastic function of explanatory variables and unknown
parameters and eij is an unobservable random variable. (In the following
discussion, we shall write ei for Eu to simplify the notation.) Thus (9.3.41) is
analogous to (9.2.4) and (9.2.5). As in Example 9.2.2, it is assumed that the
individual chooses the alternative for which the associated utility is highest.
McFadden proved that the multinomial logit model is derived from utility
maximization if and only if () are independent and the distribution function
of Ei is given by exp [ exp ( c)]. This is called the Type I extreme-value
distribution, or log Weibull distribution, by Johnson and Kotz (1970, p. 272),
who have given many more results about the distribution than are given here.
Its density is given by exp ( ej) exp [ exp (c)], which has a unique mode at 0
and a mean of approximately 0.577. We shall give only a proof of the fpart.
Denoting the density given in the preceding paragraph byf( ), we can write
the probablity the ith person chooses alternative j as (suppressing the subscript
i from po as well as from e l)
P(yi = 2) = P(Ua > Un ,Ua > U) (9.3.42)
= 1'1E2 +/12 > E1, 62 + /12 /AO >

2+#2-po
= 1- .11E2)[1. .11,) A, f(e0) d0] d2
=I , exp ( e2) exp [ exp (2)]

X exp [ exp ( e2 P2 + Pl)]

X exp [ exp Li + rILA
(E-2 r-2 v, ade
--2
exp (p a)
exp exp (A)+ exp (Ja)
Expression (9.3.42) is equal to Pa given in (9.3.34) if we put mu m = x'a fl
and xi p = x'n ft. The expressions for Pk and Pil can be similarly derived.

EXAMPLE 9.3.2. As an application of the multinomial logit model, consider

the following hypothetical model of transport modal choice. We assume that
the utilities associated with three alternatives car, bus, and train (corre-
sponding to the subscripts 0, 1, and 2, respectively) are given by (9.3.41). As
in (9.2.4) and (9.2.5), we assume
(9.3.43)
where zu is a vector of the mode characteristics and Iv, is a vector of the ith
person's socioeconomic characteristics. It is assumed that a, fi, and y are
constant for all i and j. Then we obtain the multinomial logit model (9.3.34)
with m = 2 if we put xa = za zio and xi, = zil ;3 .
The fact that 13 is constant for all the modes makes this model useful in
predicting the demand for a certain new mode that comes into existence.
Suppose that an estimate ft of fl has been obtained in the model with three
modes (Example 9.3.2) and that the characteristics z a of a new mode (desig-
nated by subscript 3) have been ascertained from engineering calculations and
a sample survey. Then the probability that the ith person will use the new
mode (assuming that the new mode is accessible to the person) can be esti-
mated by
Pi3
exp (x;3fi) (9.344)
1 + exp (xj) + exp (x;2fl) + exp (343/)
where xi3 = z 3 zi0 .

We should point out a restrictive property of the multinomial logit model:

The assumption of independence of (i) implies that the alternatives are
dissimilar. Using McFadden's famous example, suppose that the three alter-
natives in Example 9.3.2 consist of car, red bus, and blue bus, instead of car,
bus, and train. In such a case, the independence between e l and 2 is a clearly
unreasonable assumption because a high (low) utility for red bus should
generally imply a high (low) utility for a blue bus. The probability P 0 =
P(U0 > U1 , U0 > U2) calculated under the independence assumption would
underestimate the true probability in this case because the assumption ignores
the fact that the event U 0 > U1 makes the event U0 > U2 more likely.
Alternatively, note that in the multinomial logit model the relative proba-
bilities between a pair of alternatives are specified ignoring the third alterna-
tive. For example, the relative probabilities between car and red bus are
specified the same way regardless of whether the third alternative is blue bus
or train. Mathematically, this fact is demonstrated by noting that (9.3.34)
implies
P(yi = jiyi = j or k).= [exp (4/3) + exp (34/1)1-1 exp (4/1).
(9.3.45)
McFadden has called this characteristic of the model independence from
irrelevant alternatives (IA).
The following is another example of the model.
EXAMPLE 9.3.3 (McFadden, 1976b). McFadden (1976b) used a multino-
mial logit model to analyze the selection of highway routes by the California
Division of Highways in the San Francisco and Los Angeles Districts during
the years 1958-1966. The ith project among n = 65 projects chooses one
from mi routes and the selection probability is hypothesized precisely as
(9.3.34), where x i./ is interpreted as a vector of the attributes of route j in
project i.
There is a subtle conceptual difference between this model and the model of
Example 9.3.2. In the latter model, j signifies a certain common type of
transport mode for all the individuals i. For example, j =0 means car for all i.
In the McFadden model, the jth route of the first project and the jth route of
the second project have nothing substantial in common except that both are
number j routes. However, this difference is not essential because in this type
of model each alternative is completely characterized by its characteristics
vector x, and a common name such as car is just as meaningless as a number j
in the operation of the model.
McFadden tested the IIA hypothesis by reestimating one of his models
using the choice set that consists of the chosen route and one additional route
randomly selected from mi . The idea is that if this hypothesis is true, estimates
obtained from a full set of alternatives should be close to estimates obtained by
randomly eliminating some nonchosen alternatives. For each coefficient the
difference between the two estimates was found to be less than its standard
deviation, a finding indicating that the hypothesis is likely to be accepted.
However, to be exact, we must test the equality of all the coefficients simulta-
neously.
Such a test, an application of Hausman's test (see Section 4.5.1), is devel-
oped with examples in the article by Hausman and McFadden (1984). They
tested the IIA hypothesis in a trichotomous logit model for which the three
alternatives are owning an electric dryer, a gas dryer, or no dryer. In one
experiment, data on the households without a dryer were discarded to obtain a
consistent but inefficient estimator, and in the other experiment, data on
those owning electric dryers were discarded. In both experiments Hausman's
test rejected the IIA hypothesis at less than 1% significance level. Alternative
tests of the IIA hypothesis will be discussed in Section 9.3.5.
9.3.4 Multinomial Discriminant Analysis

The DA model of Section 9.2.8 can be generalized to yield a multinomial DA
model defined by
41(3'; =./) Mis; Mi) (9.3.46)
and
P(y j) = q i (9.3.47)
for i = 1, 2, . . . , n and ] = 0, 1, . . . , m. By Bayes's rule we obtain
P(y = fix?) (9.3.48)

y gi,(41 )qk
X()
where gi is the density function of Just as we obtained (9.2.48) from

(9.2.46), we can obtain from (9.3.48)
=fix?) 0) + 13.;(2)41 + A4), (9.3.49)

= Olxi*) - A(/3
where flio) , /1;(2) , and A are similar to (9.2.49), (9.2.50), and (9.2.51) except that
the subscripts 1 and 0 should be changed to ] and 0, respectively.
As before, the term xrAx? drops out if all the E's are identical. If we write
/3./0) + /3;(2)4 = gxi , the DA model with identical variances can be written
exactly in the form of (9.3.34), except for a modification of the subscripts of )/
and x.
Examples of multinomial DA models are found in articles by Powers et al.
(1978) and Uhler (1968), both of which are summarized by Amemiya (1981).
9.3.5 Nested Logit Model

In Section 9.3.3 we defined the multinomial logit model and pointed out its
weakness when some of the alternatives are similar. In this section we shall
discuss the nested (or nonindependent) logit model that alleviates that weak-
ness to a certain extent. This model is attributed to McFadden (1977) and is
developed in greater detail in a later article by McFadden (1981). We shall
analyze a trichotomous model in detail and then generalize the results ob-
tained for this trichotomous model to a general multinomial case.
Let us consider the red bus- blue bus model once more for the purpose of
illustration. Let U + j = 0, 1, and 2, be the utilities associated with
car, red bus, and blue bus. (To avoid unnecessary complication in notation,
we have suppressed the subscript i.) We pointed out earlier that it is unreason-
able to assume independence between EI and E2, although 0 may be assumed
independent of the other two. McFadden suggested the following bivariate
distribution as a convenient way to take account of a correlation between e l
and 2 :
F(1 , 2) = exp (- [exp (-p - i,) + exp (- 11-12 )}9), (9.3.50)
0<p 1.
Johnson and Kotz (1972, p. 256) called this distribution Gumbel's Type B
bivariate extreme-value distribution. The correlation coefficient can be shown
to be 1 - p2. If p = 1 (the case of independence), F(ei , 2) becomes the prod-
uct of two Type I extreme-value distributions-- in other words, the multino-
mial logit model. As for 0 , we assume F(e0)= exp [exp (E0)] as in the
multinomial logit model.
Under these assumptions we can show
exp ( go )
P(y = 0) (9.3.51)
exp (ito) + [exp (p - + exp (p-1/22)?
and
p(), iy 0) _ exP ( 1 1)
(9.3.52)
exp (r 'i)+ :p(p- '112) .
The other probabilities can be deduced from (9.3.51) and (9.3.52). Therefore
these two equations define a nested logit model in the trichotomous case. By
dividing the numerator and the denominator of (9.3.51) by exP (Po) and those
of (9.3.52) by exp (r we note that the probabilities depend on 12
/21 muo , and p. We would normally specify "Li = x.;/), j = 0, 1, 2. The estima-
tion of /I and p will be discussed for more general nested logit models later.
The form of these two probabilities is intuitively attractive. Equation
(9.3.52) shows that the choice between the two similar alternatives is made
according to a binary logit model, whereas (9.3.51) suggests that the choice
between car and noncar is also like a logit model except that a certain kind of a
weighted average of exP (ii1) and exP (1 12) is used.
To obtain (9.3.51), note that
P(Y = 0) = P(U0 > U1, Uo > U2) (9.3.53)
=P(/10Eo>111+1,P0+0>112+2)
f
eo+" --01 [ o+Po
fe -11 2
exp ( E0 )
X exp [ exP ( E-o)]f(Ei , E2) dE2] A I } d0
= f exp ( Eo ) exp [ exp ( 0 )]
X exp ( (exP [ P-1(eo

+ exP [ 1 (Eo 122)]Y) d0
=
LI exp () exp [ exP ( E0)] de.0
-
where a = 1 + exP ( Po) [exP ) + exP (P-I/12)[.

To obtain (9.3.52), we first observe that

P(y = 1 ly # 0) = Ul > U2 (U1 >U0 or U2 > Uo) (9.3.54)
P(U1 > (12 ),
where the last equality follows from our particular distributional assumptions.
Next we obtain
P(U 1 > J2) = P(14 1+ El> 112+ 2) (9.3.55)
ei +ill
, E2 ) deddel
-OD
= exp ( E)(I + exP
X exp (exp (E l )
X fl exp _ ,112 )])P) cl1
= 1 + exP
where in the third equality we used
8F(e1, 2)
[exp ( El ) exp (p-1 e2 )1P-1 (9.3.56)
X exp ( le, )F(ei , 2 ).

The model defined by (9.3.51) and (9.3.52) is reduced to the multinomial
logit model if p = 1. Therefore the HA hypothesis, which is equivalent to the
hypothesis p = 1, can be tested against the alternative hypothesis of the nested
logit model by any one of the three asymptotic tests described in Section 4.5.1.
Hausman and McFadden (1984) performed the three tests using the data on
the households' holdings of dryers, which we discussed in Section 9.3.3. The
utilities of owning the two types of dryers were assumed to be correlated with
each other, and the utility of owning no dryer was assumed to be independent
of the other utilities. As did Hausman's test, all three tests rejected the IIA
hypothesis at less than 1% significance level. They also conducted a Monte
Carlo analysis of Hausman's test and the three asymptotic tests in a hypotheti-
cal trichotomous nested logit model. Their findings were as follows: (1) Even
with n = 1000, the observed frequency of rejecting the null hypothesis using
the Wald and Rao tests differed greatly from the nominal size, whereas it was
better for the LRT. (2) The power of the Wald test after correcting for the size
was best, with Hausman's test a clost. second.
Next, we shall generalize the trichotomous nested logit model defined by

(9.3.51) and (9.3.52) to the general case of m + 1 responses. Suppose that the
m + 1 integers 0, 1, . . . , m can be naturally partitioned into S groups so
that each group consists of similar alternatives. Write the partition as
(0, 1, 2, . . . , m) = B 1 U B2 U . . .0 B5, (9.3.57)
where U denotes the union. Then McFadden suggested the joint distribution
F(e0 , Ei, . . , E ns ) eXP I a, [ exp (

s I jEB,
(9.3.58)
Then it can be shown that
a,[ I exp (p s--

JEB, (9.3.59)
jEB,
Pi S
T-1
at LE
.EB,
exp (pnif)
s 1, 2, . . . , S,
and
exp (p;
111 E 135) s = 1, 2, . . . , S. (9.3.60)
exP (P.; I,uk
kEB,
Note that (9.3.59) and (9.3.60) are generalizations of (9.3.51) and (9.3.52),
respectively. Clearly, these probabilities define the model completely. As be-
fore, we can interpret
LB, exp ( 19;ip.dr

as a kind of weighted average of exp GO for j E B.
The nested logit model defined by (9.3.59) and (9.3.60) can be estimated by
MLE, but it also can be consistently estimated by a natural two-step method,
which is computationally simpler. Suppose we specify pi = x;fl. First, the part
of the likelihood function that is the product of the conditional probabilities of
the form (9.3.60) is maximized to yield an estimate of p; 111 Second, this
estimate is inserted into the right-hand side of (9.3.59), and the product of
(9.3.59) over s and i (which is suppressed) is maximized to yield estimates of
p's and a's (one of the a's can be arbitrarily set). The asymptotic covariance
matrix of these elements is given in McFadden (1981).
In a large-scale nested logit model, the two-step method is especially useful

because of the near infeasibility of MLE. Another merit of the two-step
method is that it yields consistent estimates that can be used to start a
Newton-Raphson iteration to compute MLE, as done by Hausman and
McFadden.
We shall present two applications of the nested logit model.
EXAMPLE 9.3.4 (Small and Brownstone, 1982). Small and Brownstone ap-
plied a nested logit model to analyze trip timing. The dependent variable takes
12 values corresponding to different arrival times, and the authors experi-
mented with various ways of "nesting" the 12 responses, for example, B 1 --
(1, 2, . . , 8) and B 2 = (9, 10, 11, 12) or B 1 = (1, 2, . . . , 8), B2 = (9), and
B3 = (10, 11, 12). All a's were assumed to be equal to 1, and various specifica-
tions of the p's were tried. Small and Brownstone found that the two-step
estimator had much larger variances than the MLE and often yielded unrea-
sonable values. Also, the computation of the asymptotic covariance matrix of
the two-step estimator took as much time as the second-round estimator
obtained in the Newton-Raphson iteration, even though the fully iterated
Newton-Raphson iteration took six times as much time.
EXAMPLE 9.3.5 (McFadden, 1978). A person chooses a community to live

in and a type of dwelling to live in. There are S communities; integers in B,
signify the types of dwellings available in community s. Set a, = a constant,
as = a constant, and /la = "'cc,/ + a'zc . Then
P
exp [P -1("xa + a'z,)1}
dne
P(community c is chosen)
fd
E Ee exp [p-1 (rxed +
(9.3.6 1 )
and
exp (p- Vfx,d)
P(dwelling d is chosen I c is chosen)
exp (pwx)
d'EB,
(9.3.62)
As in this example, a nested logit is useful when a set of utilities can be

naturally classified into independent classes while nonzero correlation is al-
lowed among utilities within each class.
9.3.6 Higher-Level Nested Logit Model

The nested logit model defined in the preceding section can be regarded as
implying two levels of nesting because the responses are classified into S
groups and each group is further classified into the individual elements. In this
section we shall consider a three-level nested logit. A generalization to a higher
level can be easily deduced from the three-level case.
Figure 9.1 shows examples of two-level and three-level nested logit models
for the case of eight responses.
Following McFadden (1981), we can generalize (9.3.58) to the three-level
case by defining
F(0, E l , . . . , E,) (9.3.63)
r.}1
= exp ( b. I a, [ exp (AVE)
S EC jEB,
Then (9.3.59) and (9.3.60) can be generalized to

exP (11.0s) (9.3.64)
=
exP (141P3) '
kEB,
as
JEB,
rns)
/ = Mu) (93.65)
J EB,
a, [exp 02,10
TE JEB,
Two-Level
1 2
2 13 4 6 7 8
1
Three-Level
2 3 1 4
3 4 516 7
1---1 1 1 1- 7
Figure 9.1 Two-level and three-level nested logit models
and
bu f
TEc,,
a, [
jEB,
exp (141p,)
rel
= hila. (9.3.66)
JE CN
/ { / a, [ exp ( j/p,) ]
EEC,, jes,
A three-step estimation method can be defined in which, first, the parame-

ters in (9.3.64) are estimated by maximizing the product of the terms of the
form (9.3.64). Then these estimates are inserted into (9.3.65), and the product
of the terms of the form (9.3.65) is maximized to yield the estimates of its
parameters. Finally, (9.3.66) is maximized after inserting the estimates ob-
tained by the first and second steps.
9.3.7 Generalized Extreme-Value Model

McFadden (1978) introduced the generalized extreme-value (GEV) distribu-
tion defined by
F (et , 2, ,EM) (9.3.67)
= exp ( G[exp (), exp ( 2), . . . , exp ( ,)]),
where G satisfies the conditions,
(i) G(14 1 , U2 , . .., 0, U 1 , U2 ,. . . , u, O.
(ii) G(au au2 , . , au,)= aQui , 112, um).
akG
(iii) 0 if k is odd
aui, auh au
0 if k is even, k =-- 1, 2, . . . , m.
If U + e; and the alternative with the highest utility is chosen as before,
(9.3.67) implies the GEV model
exp (14)Gi[exP GO, exP (p2), . , exP (PA (9.3.68)
G[exP exP ( 12), . . , exP
where Gi is the derivative of G with respect to its jth argument.
Both the nested logit model and the higher-level nested logit model dis-
cussed in the preceding sections are special cases of the GEV model. The only
known application of the GEV model that is not a nested logit model is in a
study by Small (1981).
The multinomial models presented in the subsequent subsections do not

belong to the class of GEV models.
9.3.8 Universal Logit Model

In Section 9.2.1 we stated that for a binary QR model a given probability
function G(4, 0) can be approximated by F[H(x7, 0)] by choosing an appro-
priate H(4, 0) for a given F. When F is chosen to be the logistic function A,
such a model is called a universal logit model. A similar fact holds in the
multinomial case as well. Consider the trichotomous case of Example 9.3.2.
A universal logit model is defined by
exp (ga)
Pi2 (9.3.69)
1 + exp (ga ) + exp (ga) '
exp (gi1 )
Pal 1 + exp (Ai ) exp (gt2)' (9.3.70)
and
1
PE0 (9.3.71)
1 + exp (AI ) + exp (gi2)'
where gii and ga are functions of all the explanatory variables of the model
Z 1 , za , and wi . Any arbitrary trichotomous model can be approximated by
this model by choosing the functions g a and ga appropriately. As long as g11
ga depend on all the mode characteristics, the universal logit model does and
not satisfy the assumption of the independence from irrelevant alternatives.
When the g's are linear in the explanatory variables with coefficients that
generally vary with the alternatives, the model is reduced to a multinomial
logit model sometimes used in applications (see Cox, 1966, p. 65), which
differs from the one defined in Section 9.3.3.
9.3.9 Multinomial Probit Model

Let U, j = 0, 1, 2, . . . , m, be the stochastic utility associated with the jth
alternative for a particular individual. By the multinomial probit model, we
mean the model in which Rif} are jointly normally distributed. Such a model
was first proposed by Aitchison and Bennett (1970). This model has rarely
been used in practice until recently because of its computational difficulty
(except, of course, when m = 1, in which case the model is reduced to a binary
probit model, which is discussed in Section 9.2). To illustrate the complexity

of the problem, consider the case of m = 2. Then, to evaluate P(y = 2), for
example, we must calculate the multiple integral
P(y = 2) = P( U2 > U1 , U2> Up) (9.3.72)
U2 f U 1 , U2) dUo dU dU2 ,

--
where f is a trivariate normal density. The direct computation of such an

integral is involved even for the case of m = 3. 9 Moreover, (9.3.72) must be
evaluated at each step of an iterative method to maximize the likelihood
function.
Hausman and Wise (1978) estimated a trichotomous probit model to ex-
plain the modal choice among driving own car, sharing rides, and riding a bus,
for 557 workers in Washington, D.C. They specified the utilities by
X3i
= Pa log x11; + Pa log x21 + Pa X4ii +
(9.3.73)
where x1 , x2 , x3 , and x4 represent in-vehicle time, out-of-vehicle time, in-

come, and cost, respectively. They assume that (fl ii , flu, fl e u) are indepen-
dently (with each other) normally distributed with means (Pi, / 32, fi3, 0) and
variances (a3, a3 , a3 , 1). It is assumed that e u are independent both through i
andj and fins are independent through i; therefore correlation between U and
Uik occurs because of the same ffs appearing for all the alternatives. Hausman
and Wise evaluated integrals of the form (9.3.72) using series expansion and
noted that the method is feasible for a model with up to five alternatives.
Hausman and Wise also used two other models to analyze the same data:
Multinomial Logit (derived from Eq. 9.3.73 by assuming that gs are nonsto-
chastic and eu are independently distributed as Type I extreme-value distribu-
tion) and Independent Probit (derived from the model defined in the preced-
ing paragraph by putting al = 0 22 = a3 = 0). We shall call the original model
Nonindependent Probit. The conclusions of Hausman and Wise are that (1)
Logit and Independent Probit give similar results both in estimation and in
the forecast of the probability of using a new mode; (2) Nonindependent
Probit differs significantly from the other two models both in estimation and
the forecast about the new mode; and (3) Nonindependent Probit fits best.
The likelihood ratio test rejects Independent Probit in favor of Nonindepen-
dent Probit at the 7% significance level. These results appear promising for
further development of the model.
Albright, Lerman, and MansIci (1977) (see also Lerman and Manslci, 1981)
developed a computer program for calculating the ML estimator (by a gra-
dient iterative method) of a multinomial probit model similar to, but slightly
more general than, the model of Hausman and Wise. Their specification is
= 34;fl1 (9.3.74)
where ft, ;) and ei = (em , n ,. . . , ei,) N(0, ;). As in Hausman
and Wise, fl1 and ei are assumed to be independent of each other and indepen-
dent through i. A certain normalization is employed on the parameters to
make them identifiable.
An interesting feature of their computer program is that it gives the user the
option of calculating the probability in the form (9.3.72) at each step of the
iteration either (1) by simulation or (2) by Clark's approximation (Clark,
1961), rather than by series expansion. The authors claim that their program
can handle as many as ten alternatives. The simulation works as follows:
Consider evaluating (9.3.72), for example. We artificially generate many ob-
servations on U0 , U1 , and U2 according tofevaluated at particular parameter
values and simply estimate the probability by the observed frequency. Clark's
method is based on a normal approximation of the distribution of max (X, Y)
when both X and Y are normally distributed. The exact mean and variance of
max (X, Y), which can be evaluated easily, are used in the approximation.
Albright, Lerman, and Manslci performed a Monte Carlo study, which
showed Clark's method to be quite accurate. However, several other authors
have contradicted this conclusion, saying that Clark's approximation can be
quite erratic in some cases (see, for example, Horowitz, Sparmann, and Da-
ganzo, 1982).
Albright et al. applied their probit model to the same data that Hausman
and Wise used, and they estimated the parameters of their model by Clark's
method. Their model is more general than the model of Hausman and Wise,
and their independent variables contained additional variables such as mode-
specific dummies and the number of automobiles in the household. They also
estimated an independent logit model. Their conclusions were as follows:
1. Their probit and logit estimates did not differ by much. (They compared
the raw estimates rather than comparing aP/ax for each independent
variable.)
2. They could not obtain accurate estimates of ; in their probit model.
3. Based on the Akaike Information Criterion (Section 4.5.2), an increase
in log L in their probit model as compared to their logit model was not
large enough to compensate for the loss in degrees of freedom.
4. The logit model took 4.5 seconds per iteration, whereas the probit model
took 60 75 CPU seconds per iteration on an IBM 370 Model 158.
Thus, though they demonstrated the feasibility of their probit model, the
gain of using the probit model over the independent logit did not seem to
justify the added cost for this particular data set.
The discrepancy between the conclusions of Hausman and Wise and those
of Albright et al. is probably due to the fact that Hausman and Wise imposed
certain zero specifications on the covariance matrix, an observation that
suggests that covariance specification plays a crucial role in this type of model.
9.3.10 Sequential Probit and Logit Models
When the choice decision is made sequentially, the estimation of multinomial

models can be reduced to the successive estimation of models with fewer
responses, and this results in computational economy. We shall illustrate this
with a three-response sequential probit model. Suppose an individual deter-
mines whether y = 2 or y * 2 and then, given y * 2, determines whether
y = 1 or 0. Assuming that each choice is made according to a binary probit
model, we can specify a sequential probit model by
P2 = IrD(X2P2) (9.3.75)
and
13 1 = [ 1 0(74/12)]0(xifl1). (9.3.76)
The likelihood function of this model can be maximized by maximizing the
likelihood function of two binary probit models.
A sequential logit model can be analogously defined. Kahn and Morimune
(1979) used such a model to explain the number of employment spells a
worker experienced in 1966 by independent variables such as the number of
grades completed, a health dummy, a marriage dummy, the number of chil-
dren, a part-time employment dummy, and experience. The dependent vari-
able yi is assumed to take one of the four values (0, 1, 2, and 3) corresponding
to the number of spells experienced by the ith worker, except that yi = 3
means "greater than or equal to 3 spells." Kahn and Morimune specified
probabilities sequentially as
P(yi = 0) = A(00), (9.3.77)
P(yi = fly # 0) = A(xWII ),
, (9.3.78)
and
P(Yi= 2137,* 0, yi # 1) -- A(02). (9.3.79)
Note that Kahn and Morimune could have used an ordered logit model
with their data because we can conjecture that a continuous unobservable
variable (interpreted as a measure of the tendency for unemployment)
exists that affects the discrete outcome. Specifying y7 = xU1 + e i would lead to
one of the ordered models discussed in Section 9.3.2.
9.4 Multivariate Models

9.4.1 Introduction
A multivariate QR model specifies the joint probability distribution of two or
more discrete dependent variables. For example, suppose there are two binary
dependent variables y i and Y2, each of which takes values 1 or 0. Their joint
distribution can be described by Table 9.1, where P.M = P(YI ..19 Y2 = k) . The
model is completed by sspecifying -P P101 and Poi as functions of independent
111-
variables and unknown parameters. (P oo is determined as one minus the sum
of the other three probabilities.)
A multivariate QR model is a special case of a multinomial QR model. For
example, the model represented by Table 9.1 is equivalent to a multinomial
model for a single discrete dependent variable that takes four values with
probabilities P11 , Pio F01 , and Poo . Therefore the theory of statistical infer-
ence discussed in regard to a multinomial model in Section 9.3.1 is valid for a
multivariate model without modification.
In the subsequent sections we shall present various types of multivariate QR
models and show in each model how the probabilities are specified to take
account of the multivariate features of the model.
Before going into the discussion of various ways to specify multivariate QR
models, we shall comment on an empirical article that is concerned with a
Table 9.1 Joint distribution of two binary random variables
Y2
1 0
Pit PLO
POI Poo
multivariate model but in which its specific multivariate features are (perhaps
justifiably) ignored. By doing this, we hope to shed some light on the distinc-
tion between a multivariate model and the other multinomial models.
EXAMPLE 9.4.1 (Silberman and Durden, 1976). Silberman and Durden
analyzed how representatives voted on two bills (House bill and Substitute
bill) concerning the minimum wage. The independent variables are the socio-
economic characteristics of a legislator's congressional district, namely, the
campaign contribution of labor unions, the campaign contribution of small
businesses, the percentage of persons earning less than $4000, the percentage
of persons aged 16 21, and a dummy for the South. Denoting House bill and
Substitute bill by H and S, the actual counts of votes on the bills are given in
Table 9.2.
The zero count in the last cell explains why Silberman and Durden did not
set up a multivariate QR model to analyze the data. Instead, they used an
ordered probit model by ordering the three nonzero responses in the order of a
representative's feeling in favor of the minimum wage as shown in Table 9.3.
Assuming that y* is normally distributed with mean linearly dependent on the
independent variables, Silberman and Durden specified the probabilities as
(9.4.1)
and
PA) + Pil = (1:0(x;fl + a), (9.4.2)
where a> 0.
An alternative specification that takes into consideration the multivariate
nature of the problem and at the same time recognizes the zero count of the
last cell may be developed in the form of a sequential probit model (Section
9.3.10) as follows:
P(H; = Yes) = (1)(x;fi 1 ) (9.4.3)
Table 9.2 Votes on House (H) and Substitute (S) bills

Vote for Vote for House bill
Substitute
bill Yes No
Yes 75 119
No 205 0
Table 9.3 Index of feeling toward the minimum wage

Y*
(feeling in favor of
the minimum wage)
0 Yes No Weakest
1 Yes Yes Medium
2 No Yes Strongest
and
P(Si = NolHi = Yes) = 0(02). (9.4.4)
The choice between these two models must be determined, first, from a
theoretical standpoint based on an analysis of the legislator's behavior and,
second, if the first consideration is not conclusive, from a statistical standpoint
based on some measure of goodness of fit. Because the two models involve
different numbers of parameters, adjustments for the degrees of freedom, such
as the Akaike Information Criterion (Section 4.5.2), must be employed. The
problem boils down to the question, Is the model defined by (9.4.3) and (9.4.4)
sufficiently better than the model defined by (9.4.1) and (9.4.2) to compensate
for a reduction in the degrees of freedom?
9.4.2 Multivariate Nested Logit Model
The model to be discussed in this subsection is identical to the nested logit
model discussed in Section 9.3.5. We shall merely give an example of its use in
a multivariate situation. We noted earlier that the nested logit model is useful
whenever a set of alternatives can be classified into classes each of which
contains similar alternatives. It is useful in a multivariate situation because the
alternatives can be naturally classified according to the outcome of one of the
variables. For example, in a 2 X 2 case such as in Table 9.1, the four alterna-
tives can be classified according to whether y 1 = 1 or yi = 0. Using a parame-
terization similar to Example 9.3.5, we can specialize the nested logit model to
a 2 X 2 multivariate model as follows:
P(yi = 1) = a 1 exp (zly)[exp (prix 111) + exp (pilxc oprl (9.4.5)
(a l exp (el 7)[exp (pi' xj) + exp (pilx1 0/1)P1
ao exP (zOY)[exP (PF) 141fi) + exP (1)040/1)1 P0},
exp (piix111)
P(Y2= 1 1311 = 1) exp (9.4.6)
Ixj) + exp (pr
and
exP (PV4IP)
P(Y2 = 1 lYi
exP (PV4ifi) exP (PO-lxSefl) .
The two-step estimation method discussed in Section 9.3.5 can be applied to
this model.
9.4.3 Log-Linear Model

A log-linear model refers to a particular parameterization of a multivariate
model. We shall discuss it in the context of the 2 X 2 model given in Table 9.1.
For the moment we shall assume that there are no independent variables and
that there is no constraint among the probabilities; therefore the model is
completely characterized by specifying any three of the four probabilities
appearing in Table 9.1. We shall call Table 9.1 the basic parameterization and
shall consider two alternative parameterizations.
The first alternative parameterization is a logit model and is given in Table
9.4, where d is the normalization chosen to make the sum of probabilities
equal to unity. The second alternative parameterization is called a log-linear
model and is given in Table 9.5, where d, again, is a proper normalization, not
necessarily equal to the d in Table 9.4.
The three models described in Tables 9.1, 9.4, and 9.5 are equivalent; they
differ only in parameterization. Parameterizations in Tables 9.4 and 9.5 have
the attractive feature that the conditional probabilities have a simple logistic
form. For example, in Table 9.4 we have
exp (A li)
P(Yi = 1 IY2 = 1 ) exp (ii 1 1) + exp (Poi) A(pi 1 izol). (9.4.8)
Note, however, that the parameterization in Table 9.5 has an additional

attractive feature: a12 = 0 if and only if y1 and y2 are independent. The role
Table 9.4 Bivariate logit model
Y2
Y1
1 cl - 'eui1 ri- iekho

0 el- ieft. d- ' eft.
Table 9.5 Log-linear model
Y2
Y1
I d- 'ecri-fa24-`1.2 d - Ie*,
0 d - ie2
of a 12 also can be seen by the following equation that can be derived from
Table 9.5:
P(Y1 = 1 13'2) = A(ai + a12Y2). (9.4.9)
The log-linear parameterization in Table 9.5 can also, be defined by

P(yi , Y2) m exP (a1 Y1 + ei2)2 + a12Y1Y2) , (9.4.10)
where oc reads "is proportional to." This formulation can be generalized to a
log-linear model of more than two binary random variables as follows (we
shall write only the case of three variables):
RYI, Y2, Y3) cc eXP (al Y1 4- 0 2Y2 03Y3 a12Y1 Y2 + a13Y1 Y3
+ a23 Y2 Y3 + a123Y1Y2Y3). (9.4.11)

The first three terms in the exponential function are called the main effects.
Terms involving the product of two variables are called second-order interac-
tion terms, the product of three variables third-order interaction terms, and so
on. Note that (9.4.11) involves seven parameters that can be put into a one-to-
one correspondence with the seven probabilities that completely determine
the distribution of y 1 , y2 , and y3 . Such a model, without any constraint among
the parameters, is called a saturated model. A saturated model for J binary
variables involves 2' 1 parameters. Researchers often use a constrained
log-linear model, called an unsaturated model, which is obtained by setting
some of the higher-order interaction terms to 0.
Example 9.4.2 is an illustration of a multivariate log-linear model.
EXAMPLE 9.4.2 (Goodman, 1972). Goodman sought to explain whether a
soldier prefers a Northern camp to a Southern camp (A) by the race of the
soldier (Y1), the region of his origin (Y2), and the present location of his camp
(North or South) (y3). Because each conditional probability has a logistic
form, a log-linear model is especially suitable for analyzing a model of this
sort. Generalizing (9.4.9), we can write a conditional probability as

1)(110 = 1 11'1, Y2, Y3) = Mao aoi a02 Y2 ao3 Y3 + a012 Y1 Y2
1013 Y1 Y3 a023 Y2 Y3 00123 Y1 Y2 Y3)"

(9.4.12)
Goodman looked at the asymptotic t value of the MLE of each a (the MLE
divided by its asymptotic standard deviation) and tentatively concluded
a012 = a013 = 00123 = 0, called the null hypothesis. Then he proceeded to
accept formally the null hypothesis as a result of the following testing proce-
dure. Define P t = 1, 2, . . . , 16, as the observed frequencies in the 16 cells
created by all the possible joint outcomes of the four binary variables. (They
can be interpreted as the unconstrained MLE's of the probabilities P,.) Define
F, as the constrained MLE of P, under the null hypothesis. Then we must
reject the null hypothesis if and only if
16 (fr _ i)2
n E
1-1 P,
> xi.., (9.4.13)
where n is the total number of soldiers in the sample and A ct is the a% critical
value of . [Note that the left-hand side of (9.4.13) is analogous to (9.3.24).]
Or, alternatively, we can use
P, logP, >
16
2n I (9.4.14)
(-1 F
We shall indicate how to generalize a log-linear model to the case of discrete
variables that take more than two values. This is done simply by using the
binary variables defined in (9.3.2). We shall illustrate this idea by a simple
example: Suppose there are two variables z and y3 such that z takes the three
values 0, 1, and 2 and y3 takes the two values 0 and 1. Define two binary (0, 1)
variables y1 and y2 by the rule: y, = 1 if z = 1 and y2 = 1 if z = 2. Then we can
specify P(z, y3) by specifying P(y, , y2 , y3), which we can specify by a log-linear
model as in (9.4.11). However, we should remember one small detail: Because
in the present case y i y2 = 0 by definition, the two terms involving y, y 2 in the
right-hand side of (9.4.11) drop out.
In the preceding discussion we have touched upon only a small aspect of the
log-linear model. There is a vast amount of work on this topic in the statistical
literature. The interested reader should consult articles by Haberman (1978,
1979), Bishop, Feinberg, and Holland (1975), or the many references to Leo
Goodman's articles cited therein.
Nerlove and Press (1973) proposed making the parameters of a log-linear

model dependent on independent variables. Specifically, they proposed the
main-effect parameters a, , a 2 , and a3 in (9.4.11)to be linear combina-
tions of independent variables. (However, there is no logical necessity to
restrict this formulation to the main effects.)
Because in a log-linear model each conditional probability has a logit form
as in (9.4.12), the following estimation procedure (which is simpler than
MLE) can be used: Maximize the product of the conditional probabilities with
respect to the parameters that appear therein. The remaining parameters must
be estimated by maximizing the remaining part of the likelihood function.
Amemiya (1975b) has given a sketch of a proof of the consistency of this
estimator. We would expect that the estimator is, in general, not asymptoti-
cally as efficient as the MLE. However, Monte Carlo evidence, as reported by
Guilkey and Schmidt (1979), suggests that the loss of efficiency may be minor.
9.4.4 Multivariate Probit Model

A multivariate probit model was first proposed by Ashford and Sowden
(1970) and applied to a bivariate set of data. They supposed that a coal miner
develops breathlessness (y, = 1) if his tolerance (yr) is less than 0. Assuming
that yr N( Ax, 1), where x = (1, Age)', we have
P(Y1 = 1 ) = (1)(flix). (9.4.15)
They also supposed that a coal miner develops wheeze (y 2 = 1) if his tolerance
(yr) against wheeze is less than 0 and that yr N(fix, 1). Then we have
P(Y2 = 1 ) = 0(/)ix)- (9.4.16)
Now that we have specified the marginal probabilities of y1 and y2 , the multi-
variate model is completed by specifying the joint probability
P(y, = 1, y2 = 1), which in turn is determined if the joint distribution of yr
and yr is specified. Ashford and Sowden assumed that yr and yr are jointly
normal with a correlation coefficient p. Thus
P(y, = 1, y2 = 1) = Fp(Ax, Ax), (9.4.17)
where F,, denotes the bivariate normal distribution function with zero means,
unit variances, and correlation p.
The parameters /3 1 , /32 , and p can be estimated by MLE or MIN x2 (if there
are many observations per cell for the latter method). Amemiya (1972) has
given the MIN x2 estimation of this model.
Muthen (1979) estimated the following sociological model, which is equiva-

lent to a bivariate probit model:
Yi = 1 if u < a l + (9.4.18)
Y2 = 1 if u 2 < + Q21
= x'y + v
u l , u2 N(0, 1), v N(0, a2).
Here y l and y2 represent responses of parents to questions about their attitudes

toward their children, pi is a latent (unobserved) variable signifying parents'
sociological philosophy, and x is a vector of observable characteristics of
parents. Muthen generalized this model to a multivariate model in which y, x,
u, v, and a are all vectors following
y1 if u < a + et/ (9.4.19)
Btl = Fx + v
u N(0, I), v N(0, M),

where 1 is a vector of ones, and discussed the problem of identification. In
addition, Lee (1982a) has applied a model like (9.4.19) to a study of the
relationship between health and wages.
It is instructive to note a fundamental difference between the multivariate
probit model and the multivariate logit or log-linear models discussed earlier:
in the multivariate probit model the marginal probabilities are first specified
and then a joint probability consistent with the given marginal probabilities is
found, whereas in the multivariate logit and log-linear models the joint proba-
bilities or conditional probabilities are specified at the outset. The conse-
quence of these different methods of specification is that marginal probabili-
ties have a simple form (probit) in the multivariate probit model and the
conditional probabilities have a simple form (logit) in the multivariate logit
and log-linear models.
Because of this fundamental difference between a multivariate probit
model and a multivariate logit model, it is an important practical problem for
a researcher to compare the two types of models using some criterion of
goodness of fit. Morimune (1979) compared the Ashford-Sowden bivariate
probit model with the Nerlove-Press log-linear model empirically in a model
in which the two binary dependent variables represent home ownership (y 1 )
and whether or not the house has more than five rooms (y 2 ). As criteria for
comparison, Morimune used Cox's test (Section 4.5.3) and his own modifica-
tion of it. He concluded that probit was preferred to logit by either test.
It is interesting to ask whether we could specify an Ashford-Sowden type
bivariate logit model by assuming the logistic distribution for yr and yliin the
Ashford-Sowden model. Although there is no "natural" bivariate logistic
distribution the marginal distributions of which are logistic (unlike the normal
case), Lee (1982b) found that Plackett's bivariate logistic distribution function
(Plackett, 1965) yielded results similar to a bivariate probit model when
applied to Ashford and Sowden's data and Morimune's data. Furthermore, it
was computationally simpler.
Given two marginal distribution functions F(x) and G(y), Plackett's class of
bivariate distributions H(x, y) is defined by
H(1 F G + H)
(9.4.20)
(F H)(G H)
for any fixed yi in (0, co).
Unfortunately, this method does not easily generalize to a higher-order
multivariate distribution, where because of the computational burden of the
probit model the logit analog of a multivariate probit model would be espe-
cially useful. Some progress in this direction has been made by Malik and
Abraham (1973), who generalized Gumbel's bivariate logistic distribution
(Gumbel, 1961) to a multivariate case.
9.5 Choice-Based Sampling

9.5.1 Introduction
Consider the multinominal QR model (9.3.1) or its special case (9.3.4). Up
until now we have specified only the conditional probabilities of alternatives
j =0, 1, . . . , m given a vector of exogenous or independent variables x and
have based our statistical inference on the conditional probabilities. Thus we
have been justified in treating x as a vector of known constants just as in the
classical linear regression model of Chapter 1. We shall now treat both j and x
as random variables and consider different sampling schemes that specify how
j and x are sampled.
First, we shall list a few basic symbols frequently used in the subsequent
discussion:
P(jl x, fl) or P(j) Conditional probability the jth
alternative is chosen, given the
exogenous variables x
P(jN, 110 ) or Po(j) The above evaluated at the true

value of fl
f(x) True density of 300
g(x) Density according to which
a researcher draws x
H(j) Probability according to which
a researcher draws j
Q(i) = QUIP = Pf(x) dx
G(i) = QUIfio = PUlx, flaf(x) dx
Let i = 1, 2, . . . , n be the individuals sampled according to some scheme.
Then we can denote the alternative and the vector of the exogenous variables
observed for the ith individual by ji and xi , respectively.
We consider two types of sampling schemes called exogenous sampling and
endogenous sampling (or choice-based sampling in the QR model). The first
refers to sampling on the basis of exogenous variables, and the second refers to
sampling on the basis of endogenous variables. The different sampling
schemes are characterized by their likelihood functions. The likelihood func-
tion associated with exogenous sampling is given by"
L.= P(j;N o fl)g(x i). (9.5.1)
The likelihood function associated with choice-based sampling is given by
Lc= fl iIx,, P.Rxda i) (9.5.2)
if Q(iA) is unknown and by
Lco (9.5.3)
if Q(JIfio) is known. Note that if g(x) =f(x) and H(j) = Q(jIflo), (9.5.1) and
(9.5.3) both become
L* =H P(jiix i , ft)f(x i ), (9.5.4)
which is the standard likelihood function associated with random sampling.

This is precisely the likelihood function considered in the previous sections.
Although (9.5.2) may seem unfamiliar, it can be explained easily as follows.
Consider drawing random variables j and x in the following order. We can first
draw j with probability H(j) and then, given j, we can draw x according to the
conditional density f(x I j). Thus the joint probability is f(x1j)H(j), which by
Bayes's rule is equal to P(jlx)f(x)Q(j) -111(j).
This sampling scheme is different from a scheme in which the proportion of
people choosing alternative j is a priori determined and fixed. This latter
scheme may be a more realistic one. (Hsieh, Manslci, and McFadden, 1983,
have discussed this sampling scheme.) However, we shall adopt the definition
of the preceding paragraphs (following Mansld and Lerman, 1977) because in
this way choice-based sampling contains random sampling as a special case
[Q(j)= H(j)] and because the two definitions lead to the same estimators
with the same asymptotic properties.
Choice-based sampling is useful in a situation where exogenous sampling or
random sampling would find only a very small number of people choosing a
particular alternative. For example, suppose a small proportion of residents in
a particular city use the bus for going to work. Then, to ensure that a certain
number of bus riders are included in the sample, it would be much less
expensive to interview people at a bus depot than to conduct a random survey
of homes. Thus it is expected that random sampling augmented with choice-
based sampling of rare alternatives would maximize the efficiency of estima-
tion within the budgetary constraints of a researcher. Such augmentation can
be analyzed in the framework of generalized choice-based sampling proposed
by Cosslett (1981a) (to be discussed in Section 9.5.4).
In the subsequent subsections we shall discuss four articles: Mansld and
Lerman (1977), Cosslett (1981a), Cosslett (198 1 b), and Manslci and McFad-
den (1981). These articles together cover the four different types of models,
varying according to whether f is known and whether Q is known, and cover
five estimators of/3 the exogenous sampling maximum likelihood estimator
(ESMLE), the random sampling maximum likelihood estimator (RSMLE),
the choice-based sampling maximum likelihood estimator (CBMLE), the
Manski-Lerman weighted maximum likelihood estimator (WMLE), and the
Manski-McFadden estimator (MME).
A comparison of RSMLE and CBMLE is important because within the
framework of choice-based sampling a researcher controls H(j), and the par-
ticular choice H(j)= Q0( j) yields random sampling. The choice of H(j) is an
important problem of sample design and, as we shall see later, H(j) = Q0(j) is
not necessarily an optimal choice.
Table 9.6 indicates how the definitions of RSMLE and CBMLE vary with
the four types of model; it also indicates in which article each case is discussed.
Note that RSMLE = CBMLE if Q is known. ESMLE, which is not listed in
Table 9.6 Models, estimators, and cross references

RSMLE CBMLE WMLE MME
Known Known MM MAL
Max. L* wrt /) Max. L wrt /1
subject to subject to
Qo = fPf dx Qo = fPfdx
Known Unknown MM
Max. L* wrt Max. L c wrt. fl.

Unknown Known C2 (see also MAL MM
Cosslett, 1978)
Max. L* wrt /I and Max. L,,, wrt /3 and
f subject to f subject to
Qo = 5 Ffdx. Qo= fPf dx.
Unknown Unknown Cl (also proves
asymptotic
efficiency)
Max. L* wrt Max. L, wrtflandf.
Note: RSMLE = random sampling maximum likelihood estimator, CBMLE = choice-based
sampling maximum likelihood estimator; WMLE = Manski-Lerman weighted maximum likeli-
hood estimator; MME = MansId-McFadden estimator.
MM = Manski and McFadden (1981); MAL = Mansld and Lerman (1977); C2 = Cosslett
(1981b); Cl = Cosslett (1981a).
Table 9.6, is the same as RSMLE except when./ is unknown and Q is known.
In that case ESMLE maximizes L. with respect to /I without constraints.
RSMLE and CBMLE for the case of known Q will be referred to as the
constrained RSMLE and the constrained CBMLE, respectively. For the case
of unknown Q, we shall attach unconstrained to each estimator.
9.5.2 Results of Manski and Lerman

Mansld and Lerman (1977) considered the choice-based sampling scheme
represented by the likelihood function (9.5.3) the case where Q is known
and proposed the estimator (WMLE), denoted fl,, which maximizes
S= w(j 1 ) log PU,Ix t , 10, (9.5.5)
where w( = Q0(j)/H(j). More precisely, fi, is a solution of the normal

equation
(9.5.6)

It will become apparent that the weights w(j) ensure the consistency of the
estimator. If weights were not used, (9.5.5) would be reduced to the exogenous
sampling likelihood function (9.5.1), and the resulting estimator to the usual
MLE (the ESMLE), which can be shown to be inconsistent unless H(j)
CAD.
It should be noted that because WMLE does not depend on f(x), it can be
used regardless of whether or not f is known.
We shall prove the consistency of the WMLE fl in a somewhat different
way from the authors' proof. 12 The basic theorems we shall use are Theorems
4.1.2 and 4.2.1. We need to make six assumptions.
ASSUMPTION 9.5.1. The parameter space B is an open subset of Euclidean
space.
ASSUMPTION 9.5.2. H(j) > 0 for every j = 0, 1, . . . , m.
ASSUMPTION 9.5.3. a log puix, gyap exists and is continuous in an open
neighborhood N1 (90 ) of flo for every j and x. (Note that this assumption
requires P(jlx, fi) > 0 in the neighborhood.)' 3
ASSUMPTION 9.5.4. P(jlx, /i) is a measurable function off and x for every
/I E B.
ASSUMPTION 9.5.5. (j, x) are i.i.d. random variables.

ASSUMPTION 9.5.6. If fl *Po, P [P(jI XI fl) Po)] > 0.
To prove the consistency of 1, we first note that Assumptions 9.5.1, 9.5.2,
and 9.5.3 imply conditions A and B of Theorem 4.1.2 for S defined in (9.5.5).
Next, we check the uniform convergence part of condition C by putting
g(y, 0) = log P(jlx, fl) E log P(jlx, /3) in Theorem 4.2.1. (Throughout
Section 9.5, E always denotes the expectation taken under the assumption
of choice-based sampling; that is, for any g(j, x), Eg(j, x)=
f /310 g(h x)P(jlx, fis)Qotir lIADRx) dx.) Because 0 < P(jlx, /3) < 1 for
ft E 'I', where 'I' is some compact subset of N I (/30 ), by Assumption 9.5.3 we
clearly have E supp.pllog P(j) E log P(i) < co. This fact, together with
Assumptions 9.5.4 and 9.5.5, implies that all the conditions of Theorem 4.2.1
are fulfilled. Thus, to verify the remaining part of condition C, it only remains
to show that limn n'ES attains a strict local maximum at fin .
Using Assumption 9.5.5, we have
priES = f w( j) [log P(j)]P0(/).Ax)Q0ur ili(i) dx (9.5.7)
= f [log P(j)]P0 (j)f(x) dx

= E* log P(j),
where E in the left-hand side denotes the expectation taken according
to the true choice-based sampling scheme, whereas E* after the last equality
denotes the expectation taken according to the hypothetical random
sampling scheme. (That is, for any g(j, x), E*g(j, x) ----
f .0 x)P(ilx, flo)fix) dx.) But, by Jensen's inequality (4.2.6) and As-
sumption 9.5.6, we have
E* log P(jlx, /I) < E log P(jlx, fio) for fl#fi0 . (9.5.8)
Thus the consistency of WMLE has been proved.
That the ESMLE is inconsistent under the present model can be shown as
follows. Replacing w( j) by 1 in the first equality of (9.5.7), we obtain
0 ci[log P(j)1P0(j)f(x) dx, (9.5.9)
where c; = Q0(/) -1 1/(j). Evaluating the derivative of (9.5.9) with respect to /I

at flo yields
an-alfESto
l _ 14f [it
(9.5.10)
It is clear that we must generally have c1 = 1 for every j in order for (9.5.10) to
be 0. 14
The asymptotic normality of ij can be proved with suitable additional
assumptions by using Theorem 4.1.3. We shall present merely an outline of
the derivation and shall obtain the asymptotic covariance matrix. The neces-
sary rigor can easily be supplied by the reader.
Differentiating (9.5.5) with respect to ft yields
as . Mid
w(h) (9.5.11)
aft r-i Pi)
u aft
Because (9.5.11) is a sum of i.i.d. random variables by Assumption 9.5.5, we
can expect n-112(3s,d8ph to converge to a normal random variable under

suitable assumptions. We can show
(9.5.12)
"
where
A = E[w(j)2yr], (9.5.13)
where y = [a log P(j)/8/1] PG . Differentiating (9.5.11) with respect to 13' yields
a2s ^ . 1 apcii)aptii)
I w(J1) PO .
air
(9.5.14)
1 (32P(ii)
w(ii) PC i i) *V'
Using the fact that (9.5.14) is a sum of i.i.d. random variables, we can show
1 a2s"
plim Ew(j)yy' (9.5.15)
n Po
A,
because
1 a2Pul _ f 821 ,,x, dx (9.5.16)
Ew(i) Po(f) alio" Ato aflaft po j"
0.
Po
Therefore, from (9.5.12) and (9.5.15), we conclude
flo ) )N(0, A'AA'). (9.5.17)
As we noted earlier, a researcher controls H(j) and therefore faces an
interesting problem of finding an optimal design: What is the optimal choice
of H(j)? We shall consider this problem here in the context of WMLE. Thus
the question is, What choice of H(j) will minimize A- 'A A'?
First of all, we should point out that H(j)= (20(j) is not necessarily the
optimal choice. If it were, it would mean that random sampling is always
preferred to choice-based sampling. The asymptotic covariance matrix of
V;i(13, fio) when H(j) = Q0(j) is (E*yy') -1 , where E* denotes the expecta-

tion taken with respect to the probability distribution P(jix, flo )f(x). Writing
w(j) simply as w, the question is whether

(Ewyy')'Ew 2yy'(Ewyy')i> (E*yy')'. (9.5.18)
The answer is no, even though we do have
(Ewyy')' Ew 2yy'(Ewyy') > (Eyy9 -1 , (9.5.19)
which follows from the generalized Cauchy-Schwartz inequality
Eyy' > Epe(Exx/) 1Exy'.
Let us consider an optimal choice of H(j) by a numerical analysis using a
simple logit model with a single scalar parameter. A more elaborate numerical
analysis by Cosslett (198 lb), which addressed this as well as some other ques-
tions, will be discussed in Section 9.5.5.
For this purpose it is useful to rewrite A and A as
A = E*w(j)yy (9.5.20)
and
A= E*yy' . (9.5.21)
Denoting the asymptotic variance of NG(/I fib) in the scalar case by V,v(H),
we define the relative efficiency of 13,,, based on the design H(j) by
Eff(H) c(Q) E*(72) , (9.5.22)

V (H) Elw(j)y 2]
because y is a scalar here.
We shall evaluate (9.5.22) and determine the optimal H(j) that maximizes
Eff(H) in a binary logit model:
P0( 1 ) = A(/304,
where Po and x are scalars.
In this model the determination of the optimal value of h=--- H(1) is simple
because we have
E*w(j)y 2 = + 1 b h , (9.5.23)
where
a = ExA.(130x)Ex 2 exp (fl0x)[1 + exP 0304]-3 (9.5.24)
and
b = Ex[l A(213041Ex 2 exP (2fl0x) [1 + exP (fl041 -3. (9.5.25)
Because a, b> 0, (9.5.23) approaches co as h approaches either 0 or 1 and

attains a unique minimum at
a - NO
h* - if a # b (9.5.26)
a b
-
=0.5 if a = b.
We assume x is binary with probability distribution:
x= 1 with probability p
=0 with probability 1 - p.
Then, inserting fl0 = log[(p + 2Q0 - 1)/(p - 2Q0 + 1)], where Qo = Q0( 1 ),
into the right hand side of (9.5.22), Eff(H) becomes a function ofp, Q 0 , and h
-
alone. In the last five columns of Table 9.7, the values of Eff(H) are shown for
various values ofp, Q0 , and h. For each combination ofp and Q0 , the value of
the optimal h, denoted h*, is shown in the third column. For example, if
p = 0.9 and Q0 = 0.75, the optimal value of h is equal to 0.481. When h is set
equal to this optimal value, the efficiency of WMLE is 1.387. The table shows
that the efficiency gain of using choice-based sampling can be considerable
and that h = 0.5 performs well for all the parameter values considered. It can
be shown that if Q0 = 0.5, then h* = 0.5 for all the values of p.
In the foregoing discussion of WMLE, we have assumed Q0(j) to be known.
However, it is more realistic to assume that Q0(j) needs to be estimated from a
separate sample and such an estimate is used to define w(j) in WMLE.
Manski and Lerman did not discuss this problem except to note in a footnote
Table 9.7 Efficiency of WLME for various designs in a binary logit model
P $20 h* h* 0.75 0.5 0.25 0.1
0.9 0.75 0.481 1.387 1 1.385 1.08 0.531

0.9 0.25 0.519 1.387 1.08 1.385 1 0.470
0.9 0.1 0.579 3.548 3.068 3.462 2.25 1
0.5 0.7 0.337 1.626 0.852 1.471 1.563 1
0.5 0.3 0.663 1.626 1.563 1.471 0.852 0.36
0.1 0.525 0.378 1.087 0.625 1.026 1 0.585
0.1 0.475 0.622 1.087 1 1.026 0.625 0.270
that the modified WMLE using an estimate 00(.0 is consistent as long as

plim Q0(j)= Q0(j). To verify this statement we need merely observe that
[00(../1) log P(Ax i , (9.5.27)
5 mpx1(20(/) 120(DI' i )- ' log P( jilx i , fl)1.
The right-hand side of (9.5.27) converges to 0 in probability uniformly in /I by

Assumptions 9.5.2 and 9.5.3 and by the consistency of 120(./).
To show the asymptotic equivalence of the modified WMLE and the origi-
nal WMLE, we need an assumption about the rate of convergence of (MA
By examining (9.5.12), we see that for asymptotic equivalence we need
plim n-1/2
[00(ji) Q0(i irEa log Puiyaingo = 0.
(9.5.28)
Therefore we need
00(i) Q0(j) = o(n- '12). (9.5.29)
If 00(j) is the proportion of people choosing alternative/ in a separate sample
of size n l ,
Ooti) 420(i) = 0 (ni 1/2 ). (9.5.30)
Therefore asymptotic equivalence requires that nIni should converge to O. See
Hsieh, Manski, and McFadden (1983) for the asymptotic distribution of the
WMLE with Q estimated in the case where nIni converges to a nonzero
constant.
An application of WMLE in a model explaining family participation in the
AFDC program can be found in the article by Hosek (1980).
9.5.3 Results of Manski and McFadden

Manski and McFadden (1981) presented a comprehensive summary of all the
types of models and estimators under choice-based sampling, including the
results of the other papers discussed elsewhere in Section 9.5. However, we
shall discuss only those topics that are not dealt with in the other papers,
namely, the consistency and the asymptotic normality of CBMLE in the cases
where f is known and Q is either known or unknown and MME in the case
where f is unknown and Q is known. Because the results presented here are
straightforward, we shall only sketch the proof of consistency and asymptotic
normality and shall not spell out all the conditions needed.
First, consider the case where bothfand Q are known. CBMLE maximizes
Lco given in (9.5.3) subject to the condition
Q0(i) = f P(ilx, fl)f(x) dx, j = 1, 2, . . . , m. (9.5.31)
The condition corresponding to j =0 is redundant because

Er_o fl)= 1, which will be implicitly observed throughout the following
analysis. Ignoring the known components of 4 0 , CBMLE is defined as maxi-
mizing 17_ log P(j ilxi , fl) with respect to fi subject to (9.5.31). Because the
above maximand is the essential part of log L* under random sampling, we
have CBMLE = RSMLE in this case. However, the properties of the estima-
tor must be derived under the assumption of choice-based sampling and are
therefore different from the results of Section 9.3.
To prove consistency, we need merely note that consistency of the uncon-
strained CBMLE follows from the general result of Section 4.2.2 and that the
probability limit of the constrained CBMLE should be the same as that of the
unconstrained CBMLE if the constraint is true.
To prove asymptotic normality, it is convenient to rewrite the constraint
(9.5.31) in the form of
fi = g(a), (9.5.32)
where a is a (k m)-vector, as we did in Section 4.5.1. Then, by a straightfor-

ward application of Theorem 4.2.4, we obtain
Nike/c m', ao )
N[0, (G'Eyy'G) ], (9.5.33)
where G = [afilaal, and y = [a log P(j)/811]. Therefore, using a Taylor

series approximation,
Avit. fio G(rxmi, ao), (9.5.34)
we obtain
Vii(A,AL flu ) N[0, G(G'Eyy'G)'G'1. (9.5.35)
As we would expect, jimi_ is asymptotically more efficient than WMLE ij. In
other words, we should have
G(G'Eyy'G) ' G' g (Ewyy') 'Ew 2yy'(Ewyn -1 , (9.5.36)
for any w and G. This inequality follows straightforwardly from (9.5.19).
Second, consider the case where f is known and Q is unknown. Here,

CBMLE is defined as maximizing (9.5.2) without constraint. The estimator is
consistent by the result of Section 4.2.2, and its asymptotic normally follows
from Theorem 4.2.4. Thus
( - /10 ) -0 N[0, (Eyy' - Et55') -1 1, (9.5.37)
where 6= [a log QI013110 . As we would expect, in, is not as efficient as fi
because
G(GiEriG)G' (Eyy') (Eyy' - Merl . (9.5.38)
Finally, consider the case where f is unknown and Q is known. We shall
discuss CBMLE for this model in Section 9.5.5. Here, we shall consider the
Manslci-McFadden estimator (MME), which maximizes
tY = " P(f'Ix', 11) ,23(i i)

(9.5.39)
11)120(f) i H(j)
.j()
The motivation for this estimator is the following: As we can see from (9.5.3),
the joint probability off and x under the present assumption is
h(j, x)=-- Mf(x) ,20(i)' HO. (9.5.40)
Therefore the conditional probability off given x is
h(j, x)
h(jfx) = (9.5.41)
I h(j, X)
which leads to the conditional likelihood function (9.5.39). The estimator is
computationally attractive because the right-hand side of (9.5.39) does not
depend on f(x), which is assumed unknown and requires a nonstandard
analysis of estimation, as we shall see in Sections 9.5.4 and 9.5.5.
To prove the consistency of the estimator, we observe that
plim n-1 log = E log P(i)QCirliti) (9.5.42)

P(i)(20tir'H(j)
.frO
= f [E+ log h( jlx)]C(x) dx,

where C(x) = o pouvourwif(x) and E+ is the expectation taken with
respect to the true conditional probability h o(jlx). Equation (9.5.42) is maxi-
mixed at 130 because (x) > 0 and

E+ log h(j1x) < E log ho(jlx), (9.5.43)
which, like (9.5.8), is a consequence of Jensen's inequality (4.2.6).
By a straightforward application of Theorem 4.1.3, we can show
- fl0) N[0, (Eyy' Eee') --1 ], (9.5.44)
where e = [a log /71.0 PWQ0(..1)-11/CiVafi[p o . The asymptotic covariance
matrix in (9.5.44) is neither larger nor smaller (in matrix sense) than the
asymptotic covariance matrix of WMLE given in (9.5.17). 15
9.5.4 Results of Cosslett: Part I

Cosslett (1981a) proved the consistency and the asymptotic normality of
CBMLE in the model where both fand Q are unknown and also proved that
CBMLE asymptotically attains the Cramer-Rao lower bound. These results
require much ingenuity and deep analysis because maximizing a likelihood
function with respect to a density function fas well as parameters 13 creates a
new and difficult problem that cannot be handled by the standard asymptotic
theory of MLE. As Cosslett noted, his model does not even satisfy the condi-
tions of Kiefer and Wolfowitz (1956) for consistency of MLE in the presence
of infinitely many nuisance parameters.
Cosslett's sampling scheme is a generalization of the choice-based sampling
we have hitherto considered. 16 His scheme, called generalized choice-based
sampling, is defined as follows: Assume that the total sample of size n is
divided into Ssubsamples, with ns people (n5 is a fixed known number) in the
sth subsample. A person in the sth subsample faces alternatives J a subset of
the total alternatives (0, 1, 2, . . . , m). He or she chooses alternative j with
probability Q(j)/0(s), where 0(s) = Zje Q(j). Given j, he or she chooses a
vector of exogenous variables x with a conditional densityf(xl j). 17 Therefore
the contribution of this typical person to the likelihood function is
0(s)-1 Q(j)f(xlj), which can be equivalently expressed as
0(s)-111f1x, Mx). (9.5.45)
Taking the product of (9.5.45) over all the persons in the sth subsample
(denoted by /1,) and then over all s, we obtain the likelihood function
Lfl H o(srpciiixi,mxi) (9.5.46)

s-1 /Et,
= 0(si )'Ptii I xi, filf(xi

In the model under consideration, the Q(j)'s are unknown. Therefore we

may wonder how, in subsample s, alternatives in J sampled in such a way
that the jth alternative is chosen with probability 0(s)-1 Q(j). To this question
Cosslett (1981b) gave an example of interviewing train riders at a train station,
some of whom have traveled to the station by their own cars and some of
whom have come by taxi. Thus this subsample consists of two alternatives,
each of which is sampled according to the correct probability by random
sampling conducted at the train station.
Cosslett's generalization of choice-based sampling is an attractive one, not
only because it contains the simple choice-based sampling of the preceding
sections as a special case (take J (s)), but also because it contains an inter-
esting special case of "enriched" samples." That is, a researcher conducts
random sampling, and if a particular alternative is chosen with a very low
frequency, the sample is enriched by interviewing more of those who chose
this particular alternative. Here we have, for example, J 1 = (0, 1, 2, . . . , m)
and J2(0), if the 0th alternative is the one infrequently chosen.,
Our presentation of Cosslett's results will not be rigorous, and we shall not
spell out all the conditions needed for consistency, asymptotic normality, and
asymptotic efficiency. However, we want to mention the following identifica-
tion conditions as they are especially important.
ASSUMPTION 9.5.6. 1 Js = (0, 1, 2, . . . , rn).
ASSUMPTION 9.5.7. A subset M of integers (1, 2, . . . , S) such that
(UsEm Js) n Cus.A7 -15) = where M is the complement of M, cannot be
found.
Note that if S= 2 and m = 1, for example, .11 = (0) and .12 = (1) violate
Assumption 9.5.7, whereas J 1 = (0) and .12 = (0, 1) satisfy it. Thus simple
choice-based sampling violates this assumption. Cosslett noted that these
assumptions are needed for the multinomial logit model but may not be
needed for other QR models. Cosslett also gave an intuitive reason for these
assumptions: If alternatives ] and k are both contained in some subsample,
then Q(j)/Q(k) can be estimated consistently, and under Assumption 9.5.7
we have enough of these ratios to estimate all the Q(j) separately. Thus
Assumption 9.5.7 would not be needed in the case where Q(j) are known.
Before embarking on a discussion of Cosslett's results, we must list a few
more symbols as an addition to those listed in Section 9.5.1:
P(s) = P(slx, j3) = Puix,
j".
P0(s) = 13 (SIXI 110) = P(jN, flO)

JEJ,
0(s) = Q(j)
Ns) = Q0(j)
JE.r,
H(s)=
From (9.5.46) we obtain
log L = log P(jsixi , + logf(xi) (9.5.47)
ns log J P(slx, Af(x)
Our aim is to maximize (9.5.47) with respect to a parameter vector fl and a

function f( ). By observing (9.5.47) we can immediately see that we should
putf(x) =0 for x # xi, i = 1, 2, . . . , n. Therefore our problem is reduced to
the more manageable one of maximizing
log LI = log P(jilxi, A + I log Iv, (9.5.48)

n
n5 log wiP(slx,,
s- I- I
with respect to ft and wi, i = 1, 2, . . . , n. Differentiating (9.5.48) with re-

spect to wi and setting the derivative equal to 0 yield
a log L 1 _ 1 ns P(slxi , A _ 0.
(9.5.49)
awi w 3-1 -
ws P(sIxo A
r-i
If we insert (9.5.49) into (9.5.48), we recognize that log L 1 depends on wi
S terms wiP(slxi , s = 1, 2, . . . , S. Thus, if we onlythruge
define
H(s)
As(i0 n (9.5.50)
wiP(slxi , A
I-1
we can, after some manipulation, write log L I as a function offt and (25(8)) as
log L I = log P(jilx i , fl) log As(fl)P(six 1 ,43) (9.5.51)
+ n, log As(//) I n, log n,.

5-1 5-1
But maximizing (9.5.51) with respect to fl and 4/3), s = 1 2, . . . , S, is
equivalent to maximizing
S
= I log P(jilxi , p log I ,t,p(six,, + ns log A,
3-1 s-i
(9.5.52)
with respect to /3 and A s = 1, 2, . . . , S. This can be shown as follows.
Inserting (9.5.50) into (9.5.49) yields
1
S (9.5.53)
n sts(fl)P(sixi , fl)
s-i
Multiplying both sides of (9.5.53) by P(slxi , A, summing over i, and using
(9.5.50) yield
ns Pcslx i ,
(9.5.54)
AAPP(slxi,
si
But setting the derivative of LI with respect to A, equal to 0 yields
n 13(sixo 17)
= 0 (9.5.55)
uAs As
AsP(slxi,
s
Clearly, a solution of (9.5.55) is a solution of (9.5.54).
In the maximization of (9.5.52) with respect to fl and (A5), some
normalization is necessary because i/(a.1. 1 , ceA2 , . . . , aA5, 11) =
n(Al2 A22 2 As, 11). The particular normalization adopted by Cosslett is
As= H(S). (9.5.56)
Thus we define CBMLE of this model as the value of fl that is obtained by

maximizing SI with respect to ft and (A,) subject to the normalization
(9.5.56).
Although our aim is to estimate /3, it is of some interest to ask what the MLE
of (its) is supposed to estimate. The probability limits of MLE /1' and (1.,) can be
obtained as the values that maximize plim, 'Q. We have
plim n- 11= E log fi) E log I A5nsIx1,i0 (9.5.57)

s-i
+ H(s) log
s-
= H(s) I [log Pulx, joia(s) - Igilx,110)f(x) cht
s-I je.r,
tH(s) f [log .1513(six,
X 00(s) P(six, ,80)f(x) dx
+ H(s) log As .
s.1
Differentiating (9.5.57) with respect to A i., := 1, 2, . . . , S 1, and fi yields
00(s) H(s
plim n-1 S2 = f ' 1 ) (six' 13) P(Tlx, 11)f(x) dx
As-P(six,
H(T)
+ = 1, 2, . . . , S 1, (9.5.58)
and
a .
n SI (9.5.59)
1
= H(s) I pu x, p aPap(i) 00(s)-1P(jix, ,80)f(x) dx
1 s aP(s)
H(s) I As op
5-1 AsAslx, 13) s'
s-i
X 0,0(s)'P(slx, flo)f(x) dx.
By studying (9.5.58) we see that the right-hand side vanishes if we put A s =
00(srH(s)0_0(S) and ft = /10 . Next, insert these same values into the right-
hand side of (9.5.59). Then each term becomes
S ap( i)
H(s)Qo(srf(x) dx
s-1 jEJ, uP
We conclude that
f PUYAS)420(S) -If(X) dx = 1 = 0.
plim = /30 (9.5.60)

and
plim i 1144(S) (9.5.61)

d(s)
It is interesting to note that if we replace As that appears in the right-hand

side of (9.5.52) by plim As given in (9.5.61), (2 becomes identical to the loga-
rithm of given in (9.5.39). This provides an alternative interpretation of the
Manski-McFadden estimator.
The asymptotic normality offi and (k) can be proved by a straightforward,
although cumbersome, application of Theorem 4.1.3. Although the same
value of /3 maximizes log L given in (9.5.47) and 1-2 given in (9.5.52), f2 is,
strictly speaking, not a log likelihood function. Therefore we must use
Theorem 4.1.3, which gives the asymptotic normality of a general extremum
estimator, rather than Theorem 4.2.4, which gives the asymptotic normality
of MLE. Indeed, Cosslett showed that the asymptotic covariance matrix of
617---- Ur, Ai , Az , . . . , As_ ir is given by a formula like that given in Theorem
4.1.3, which takes the form A - 'BA-1 , rather than by [ Ea2waticaal - I.
However, Cosslett showed that the asymptotic covariance matrix of Pis equal
to the first K X K block of [Ed2f2/Ootaa'r I.
Cosslett showed that the asymptotic covariance matrix oft/ attains the lower
bound of the covariance matrix of any unbiased estimator of ft. This remark-
able result is obtained by generalizing the Cramer-Rao lower bound (Theorem
1.3.1) to a likelihood function that depends on an unknown density function
as in (9.5.47).
To present the gist of Cosslett's argument, we must first define the concept
of differentiation of a functional with respect to a function. Let F(f) be a
mapping from a function to a real number (such a mapping is called a func-
tional). Then we define

OF' ____, F(f + C) F(f)
u rn
of lc el O E , (9.5.62)
where C is a function for which F(f + EC) can be defined. Let t 1 (a K-vector)
and t2 (a scalar) be unbiased estimators offi and ff(x)C(x) dx for some function
C such that f C(x) dx = 0 and f C(x) 2 dx = 1. Then Cosslett showed that the
(2K + 2)-vector
[
tc, t2 ,
a log L I I
0 log L
op" of c
has covariance matrix of the form
rc Ii
Li R()]'
where C is the covariance matrix of (t; , t 2)'. Because this covariance matrix is
positive definite, we have C> R(C) -1 , as in Theorem 1.3.1. Finally, it is shown
that the asymptotic covariance matrix offi is equal to the first K X K block of
max c R(C) - 1 .
Thus Cosslett seems justified in saying that ft is asymptotically efficient in
the sense defined in Section 4.2.4. As we remarked in that section, this does
not mean that fi has the smallest asymptotic covariance matrix among all
consistent estimators. Whether the results of LeCam and Rao mentioned in
Section 4.2.4 also apply to Cosslett's model remains to be shown.
9.5.5 Results of Cosslett: Part II

Cosslett (1981b) summarized results obtained elsewhere, especially from his
earlier papers (Cosslett, 1978, 1981a). He also included a numerical evalua-
tion of the asymptotic bias and variance of various estimators. We shall first
discuss CBMLE of the generalized choice-based sample model with unknown
f and known Q. Cosslett (1981b) merely stated the consistency, asymptotic
normality, and asymptotic efficiency of the estimator, which are proved in
Cosslett (1978). The discussion here will be brief because the results are analo-
gous to those given in the previous subsection.
The log likelihood function we shall consider in this subsection is similar to
(9.5.47) except that the last term is simplified because Q is now known. Thus
we have
log 1,2 E log P(jilxi, I))+ 5', log f(xi) (9.5.63)

1
n, log a(s),
- s
which is to be maximized with respect to fl and fsubject to the constraints
0.0(s) = f P(slx, dx. (9.5.64)
It is shown that this constrained maximization is equivalent to maximizing

with respect to Ai , j = 0, 1, 2, . . . , m,
al= log
(9.5.65)
1-1 AJ P(jjx,, p
j -0
subject to the contraint M7_ 0,17Q0(j) = 1. Consistency, asymptotic normality,

and asymptotic efficiency can be proved in much the same way as in the
preceding subsection.
Next, we shall report Cosslett's numerical analysis, which is in the same
spirit as that reported in Section 9.5.2 concerning the Mansld-Lerman esti-
mator. Cosslett compared RSMLE, CBMLE, WMLE, and MME in the sim-
ple choice-based sample model withfunknown and Q known." Three binary
QR models (logit, probit, and arctangent) were considered. In each model
there is only one independent variable, which is assumed to be normally
distributed. The asymptotic bias and the asymptotic variance of the estima-
tors are evaluated for different values offi (two coefficients), Q(1), H(1), and
the mean of the independent variable. The optimal design is also derived for
each estimator. Cosslett concluded that (1) RSMLE can have a large asymp-
totic bias; (2) CBMLE is superior to WMLE and MME; (3) a comparison
between WMLE and MME depends on parameters, especially Q(1); (4) the
choice of H(1) = 0.5 was generally quite good, leading to only a small loss of
efficiency compared to the optimal design. (This last finding is consistent with
the results of our numerical analysis reported in Section 9.5.2.)
9.6 Distribution-Free Methods

In this section we shall discuss two important articles by Manski (1975) and
Cosslett (1983). Both articles are concerned with the distribution-free estima-
tion of parameters in QR models Mansld for multinomial models and
Cosslett for binary models. However, their approaches are different: Manski
proposed a nonparametric method called maximum score estimation whereas
Cosslett proposed generalized maximum likelihood estimation in which the
likelihood function is maximized with respect to both F and /3 in the model
(9.2.1). Both authors proved only consistency of their estimator but not
9.6.1 Maximum Score Estimator A Binary Case

Manslci (1975) considered a multinomial QR model, but here we shall define
his estimator for a binary QR model and shall prove its consistency. Our proof
will be different from Manslci's proof. 2' We shall then indicate how to extend
the proof to the case of a multinomial QR model in the next subsection.
Consider a binary QR model
P(yi = 1) = F(x;130), i 1, 2, . . . , n, (9.6.1)
and define the score function
s(17)= 0) + (I Yi)X(xrfl < 0)], (9.6.2)
where
X(E) = 1 if event E occurs (9.6.3)
=0 otherwise.
Note that the score is the number of correct predictions we would make if we
predicted y, to be 1 whenever x;fl 0. Manski's maximum score estimator&
is defined by
S(fl) = sup S(fl), (9.6.4)
pEB
where the parameter space B is taken as
B= {fillrfl= 1 ). (9.6.5)
Clearly, (9.6.5) implies no loss of generality because S(c/7) = S(fl) for any
positive scalar C.
Because S(fl) is not continuous in A we cannot use Theorem 4.1.1 without
a modification. However, an appropriate modification is possible by general-
izing the concept of convergence in probability as follows:
DEFINITION 9.6.1. Let (0, A, P) be a probability space. A sequence of not
necessarily measurable functions g7(co) for co E is said to converge to 0 in
probability in the generalized sense if for any e> 0 there exists ATE A such
that
AT g IgT( 0))1
and lim P(A T) = 1.
Using this definition, we can modify Theorem 4.1.1 as follows:
THEOREM 9.6.1. Make the following assumptions:
(A) The parameter space 0 is a compact subset of the Euclidean K-space
(Rx).
(B) QT(y, 0) is a measurable function of y for all 0 E 0.
(C) 7-1 Q7 (0) converges to a nonstochastic continuous function Q(0) in
-
probability uniformly in 0 E 0 as Tgoes to co, and Q(0) attains a unique global

maximum at 00.
Define 0T as a value that satisfies
QT(OT) = sup QT(0). (9.6.6)
sEe
Then 0; converges to 00 in probability in the generalized sense.
We shall now prove the consistency of the maximum score estimator with
the convergence understood to be in the sense of Definition 9.6.1.
THEOREM 9.6.2. Assume the following:
(A) F is a distribution function such that F(x) - -- 0.5 if and only if x =0.
(B) (xi) are i.i.d. vector random variables with a joint density function g(x)
such that g(x) > 0 for all x.
Then any sequence {A) satisfying (9.6.4) converges to /1 0 in probability in
the generalized sense.
Proof Let us verify the assumptions of Theorem 9.6.1. Assumptions A and
B are clearly satisfied. The verification of assumption C will be done in five
steps.
First, define for A> 0
SAn(A IYAYA(x;rn + ( 1 y1)( - 41)l, (9.6.7)
where
WA(x)=O if x 0 (9.6.8)
=Ax if 0<x<2i
=1 if 01.-1 x.
Because each term of the summation in (9.6.7) minus its expected value
satisfies all the conditions of Theorem 4.2.1 for a fixed positive A, we can
conclude that for any e, > 0 there exists ni(A), which may depend on A, such
that for all n n i(A)
e 6
P sup In-1SA(il) QA(fi)I > < (9.6.9)
[pEB 3 2
where
= EF(x'flo)W2(x 71) + E[1 F(x x'fl). (9.6.10)
Second, we have
sup In- ' S(I1) tr'SA,,(P)1 (9.6.11)
5 sup In-1
- Di (x,13)1 + sup Ettt(x'ii)
ft
A l +A2 ,
where
IA(x) = 0 if ATI 5 Ixl (9.6.12)

= 1 +01.x if A-1 <x<0
= 1 Ax if 05x<A-1 .
Applying Theorem 4.2.1 to A I , we conclude that for any e, 6> 0 there exists
n2(A), which may depend on A, such that for all n n2(A)
(9.6.13)
We have
sup P[(x1) 2 < 2-21.
A2 .--1 (9.6.14)
s
But, because the right-hand side of (9.6.14) converges to 0 as A --. D because of
assumption B, we have for all A ..... A l
P (A2 > .,) = 0. (9.6.15)
Therefore, from (9.6.11), (9.6.13), and (9.6.15), we conclude that for all n ?.-.
n2(2) and for all A __ AI
sup in- IS(fl) n- '4,(f1)1> 1) < =6i. (9.6.16)

P s [
Third, define
Q(17) = EF(x' flo) + f [1 2F(x'fl0)]g(x) dx. (9.6.17)

xl<0
Then we have
sup 1Q(/3) QA(/1)15_ sup P[(x'fl) 2 < A-2]. (9.6.18)
fl $
Therefore, using the same argument that led to (9.6.15), we conclude that for
all A . A I
P[7 IQ(fl) Q A(PI > i] (9.6.19)
Fourth, because
sup In'S n(13) OP! suP In - i S(A n-15:1.(PI (9.6.20)
ft P
+ supin-1S),,(11) QA(11)1
ft
+ sup I GUI) QUM,
,
we conclude from (9.6.9), (9.6.16), and (9.6.19) that for any e, > Owe have
for all n _._ max[n 1 (21), n2(21)]
P[sup in - i Sn(P Q(fl)f> e] < 3. (9.6.21)
ft
Fifth and finally, it remains to show that Q(fi) defined in (9.6.17) attains a
unique global maximum at A. This is equivalent to showing
fefic<0
[1 2F(x'fl0)]g(x) dx (9.6.2)
> f
vp<0
[1 2F(x'flo)]g(x) cht if ft #
But, because 1 2F(x/i 0) > 0 in the region (xIxA <0) and 1

2F(x1/30) < 0 in the region (xixlio > 0) by assumption A, (9.6.22) follows
immediately from assumption B.
9.6.2 Maximum Score EstimatorA Multinomial Case

The multinomial QR model considered by Mansld has the following struc-
ture. The utility of the ith person when he or she chooses the jth alternative is
given by
Uu x'uttu + eu, i = 1, 2, . . . , n, j = 0, 1, . . . , m,
(9.6.23)
where we assume
ASSUMPTION 9.6.1. (eu ) are i.i.d. for both i and j.
ASSUMPTION 9.6.2. (x, x 1 ,. . . , x;,)' xi is a sequence of (m + 1)K-di-
mensional i.i.d. random vectors, distributed independently of (e u), with a
joint density g(x) such that g(x) > 0 for all x.
ASSUMPTION 9.6.3. The parameter space B is defined by B = (flifil = 1).
Each person chooses the alternative for which the utility is maximized.
Therefore, if we represent the event of the ith person choosing the jth alterna-
tive by a binary random variable h, we have
yu = 1 if Uu > Uo, for all k #j (9.6.24)
=0 otherwise.
We need not worry about the possibility of a tie because of Assumption 9.6.2.
The maximum score estimator/3 is defined as the value offithat maximizes
the scoren
n m
sn(fi)=-- E yux(xu]
P-1.frO
x;kil for all k j) (9.6.25)
subject to = 1, where x is defined in (9.6.3). We shall indicate how to

generalize the consistency theorem to the present multinomial model.
As in the proof of Theorem 9.6.2, we must verify the assumptions of
Theorem 9.6.1. Again, assumptions A and B are clearly satisfied. Assumption
C can be verified in a manner very similar to the proof of Theorem 9.6.2. We
merely note whatever changes are needed for the multinomial model.
We can generalize (9.6.7) as
n m
sAao = E yuw.d(xu- xi0)'fi, (x u -
E ,JCi )'/3, . (9.6.26)
(Xu Xi )1fl, (x1 Xij+ 1)1), . . , ( Xu
where
wA(z, , Z2 ,. . . , z) = 0 if min (zi ) 0 (9.6.27)
=1 if min (zi ) > A-1
= A min (z i ) otherwise.
Then the first four steps of the proof of Theorem 9.6.2 generalize straightfor-
wardly to the present case.
The fifth step is similar but involves a somewhat different approach. From
(9.6.25) we obtain
Q(p) = plim (13) (9.6.28)
= E E P(yi = 11x, fi0)X(x1fi xfi for all k j)

imo
= Eh(x, 13),
where 00 and (xi ) are random variables from which i.i.d. observations 0,
and (x1i) are drawn and x = (x, xi,. . . , x',)'. First, we want to show
h(x, fl) is uniquely maximized at flo . For this purpose consider maximizing
h*(x, (Ai )) = RYi = 11x, floWilf) (9.6.29)

,FO
for every x with respect to a nonoverlapping partition (Ai ) of the space of x.
This is equivalent to the following question: Given a particular x 0 , to which

region should we assign this point? Suppose
AY.to = ilxo, fio) > /lb= 1 1343, fio) for all .io (9.6.30)
Then it is clearly best to assign x 0 to the region defined by
= 11 x, fio > P(y; = 11x, flo) for all fo (9.6.31)
Thus (9.6.29) is maximized by the partition (Ay) defined by
Ay = (x1x5/30 x130 for k # j). (9.6.32)
Clearly, this is also a solution to the more restricted problem of maximizing
h(x, fl). This maximum is unique because we always have a strict inequality in
(9.6.30) because of our assumptions. Also our assumptions are such that if
h(x, fl) is uniquely maximized for every x at /30 , then D(x, II) is uniquely
maximized at/30 . This completes the proof of the consistency of the maximum
score estimator in the multonomial case. (Figure 9.2 illustrates the maximiza-
tion of Eq. 9.6.29 for the case where m = 3 and x is a scalar.)
The asymptotic distribution of the maximum score estimator has not yet
been obtained. A major difficulty lies in the fact that the score function is not
differentiable and hence Theorem 4.1.3, which is based on a Taylor expansion
of the derivative of a maximand, cannot be used. The degree of difficulty for
the maximum score estimator seems greater than that for the LAD estimator
discussed in Section 4.6 the method of proving asymptotic normality for
LAD does not work in the present case. In the binary case, maximizing (9.6.2)
P (Y0 =1 ix Po
I .
P(y r1 I x, Po )
P ( y2 = x So )
A2
Figure 9.2 An optimal partition of the space of an independent variable
is equivalent to minimizing x(x;fl 0)1. This shows both a similar-

ity of the maximum score estimator to the LAD estimator and the additional
difficulty brought about by the discontinuity of the x function.
Manski (1975) reported results of a Monte Carlo study in which he com-
pared the maximum score estimator with the logit MLE in a five-response
multinomial logit model with four unknown parameters and 400 observa-
tions. The study demonstrated a fairly good performance by the maximum
score estimator as well as its computational feasibility. The bias of the estima-
tor is very small, whereas the root mean squared error is somewhere in the
magnitude of double that of MLE. A comparison of the maximum score
estimator and MLE under a model different from the one from which MLE is
derived would be interesting.
9.6.3 Generalized Maximum Likelihood Estimator

Cosslett (1983) proposed maximizing the log likelihood function (9.2.7) of a
binary QR model with respct to /I and F, subject to the condition that F is a
distribution function. The log likelihood function, denoted here as w, is
v(fi, F)= (yi log F(x'S) + (1 y i) log [1 F(x;fi)]). (9.6.33)
The consistency proof of Kiefer and Wolfowitz ( 1956) applies to this kind of
model. Cosslett showed how to compute MLE nand Pand derived conditions
for the consistency of MLE, translating the general conditions of Kiefer and
Wolfowitz into this particular model. The conditions Cosslett found, which
are not reproduced here, are quite reasonable and likely to hold in most
practical applications.
Clearly some kind of normalization is needed on /I and F before we maxi-
mize (9.6.33). Cosslet adopted the following normalization: The constant
term is 0 and the sum of squares of the remaining parameters is equal to 1.
Note that the assumption of zero constant term is adopted in lieu of Mansld's
assumption F(0) = 0.5. We assume that the constant term has already been
eliminated from the ell that appears in (9.6.33). Thus we can proceed, assum-
ing /31 = 1.
The maximization of (9.6.33) is carried out in two stages. In the first stage
we shall fix ft and maximize ty(fi, F) with respect to F. Let the solution be
P(fl). Then in the second stage we shall maximize w[11, P(fl)] with respect to
Although the second stage presents a more difficult computational problem,
we shall describe only the first stage because it is nonstandard and concep-
tually more difficult.
The first-stage maximization consists of several steps:

Step 1. Given A rank order (x//). Suppose Kai < K2)/3 <. . . <X/1,
assuming there is no tie. Determine a sequence(09 ( v v (2) , Y(n)) accord-
9
ingly. Note that this is a sequence consisting only of ones and zeros.
Step 2. Partition this sequence into the smallest possible number of succes-
sive groups in such a way that each group consists of a nonincreasing se-
quence.
Step 3. Calculate the ratio of ones over the number of elements in each
group. Let a sequence of ratios thus obtained be (r 1 , r2 , . . . , rK), assuming
there are K groups. If this is a nondecreasing sequence, we are done. We define
F(xWI) = ri if the Wth observation is in the jth group.
Step 4. If, however, ri < for some j, combine the jth and (j 1)th group
and repeat step 3 until we obtain a nondecreasing sequence.
The preceding procedure can best be taught by example:
Example 1.
00 11 0 11
0 1
In this example, there is no need for step 4.

Example 2.
Y(0 00 11 0 10 11
P(x(0/1) 0 1
Here, the second and third group must be combined to yield

Y(0 00 1 1 0 1 0 11
'E(x O) 0 1
Note that P is not unique over some parts of the domain. For example,
between 7c4)fl and 7430 in Example 1, F may take any value between 0 and 3.
Asymptotic normality has not been proved for Cosslett's MLE if, nor for
any model to which the consistency proof of Kiefer and Wolfowitz is applica-
ble. This seems to be as difficult a problem as proving the asymptotic normal-
ity of Manslci's maximum score estimator.
Cosslett's MLE may be regarded as a generalization of Manslci's estimator
because the latter searches only among one-jump step functions to maximize
(9.6.33). However, this does not necessarily imply that Cosslett's estimator is
superior. Ranking of estimators can be done according to various criteria. If

the purpose is prediction, Manski's estimator is an attractive one because it
maximizes the number of correct predictions.
9.7 Panel Data OR Models

Panel data consist of observations made on individuals over time. We shall
consider models in which these observations are discrete. If observations are
independent both over individuals and overtime, a panel data QR model does
not pose any new problem. It is just another QR model, possibly with more
observations than usual. Special problems arise only when observations are
assumed to be correlated over time. Serial correlation, or temporal correla-
tion, is certainly more common than correlation among observations of dif-
ferent individuals or cross-section units. In this section we shall consider
various panel data QR models with various ways of specifying serial correla-
tion and shall study the problem of estimation in these models.
9.7.1 Models with Heterogeneity

We shall study the problem of specifying a panel data QR model by consider-
ing a concrete example of sequential labor force participation by married
women, following the analysis of Heckman and Willis (1977). As is customary
with a study of labor force participation, it is postulated that a person works if
the offered wage exceeds the reservation wage (shadow price of time). The
offered wage depends on the person's characteristics such as her education,
training, intelligence, and motivation, as well as on local labor market condi-
tions. The reservation wage depends on the person's assets, her husband's
wage rate, prices of goods, interest rates, expected value of future wages and
prices, and the number and age of her children. Let y ft = 1 if the ith person
works at time t and = 0 otherwise and let xi, be a vector of the observed
independent variables (a subset of the variables listed above). Then for fixed t
it seems reasonable to specify a binary QR model
P(y = 1) = F(34,13), i = 1, 2, . . . , n, (9.7.1)
where F is an appropriate distribution function.
However, with panel data, (9.7.1) is not enough because it specifies only a
marginal probability for fixed t and leaves the joint probability
P(yil , y,. . . , YET) unspecified. (We assume (y R ) are independent over i.)
The simplest way to specify a joint probablity is to assume independence and
specify P(yil , ya , . . . , yiT) gn i P(h). Then we obtain the binary QR

model studied in Section 9.2, the only difference being that we have nT
observations here.
The independence assumption would imply P(y11=11Yi,1-1=1)=
P(y u = 1). In other words, once we observe xis , whether or not a woman
worked in the previous year would offer no new information concerning her
status this year. Surely, such a hypothesis could not be supported empirically.
There are two major reasons why we would expect P(y u = 1 = 1) #
P(y it == 1):
1. Heterogeneity. There are unobservable variables that affect people dif-
ferently in regard to their tendency to work. 23
2. True State Dependence. For each person, present status influences future
status. For example, once you start doing something, it becomes habitual.
Heckman and Willis (1977) attempted to take account of only heterogeneity
in their study. We shall consider how to model true state dependence in
Section 9.7.2.
A straightforward approach to take account of heterogeneity is to specify
PO, it= llui) = Rx'itfl ui), (9.7.2)
= 1, 2, . . . , n, t = 1, 2, . . . , T,
and assume (y e ) are serially independent (that is, over t) conditional on u.
Therefore we have (suppressing subscript 0
P(yt t-i = 1 ) P(yt= 1) (9.7.3)
E[F(x;fl + u)F(x;_ I P + u)]
EF(413 + u)
EF(x;_ i + u)
Coy [F(x33 + u), F(x;_ i fi + u)]
EF(4_ 1 11 + u)
which can be shown to be nonnegative. The joint probability of (y e ), t =
1, 2, . . . , T, is given by
, Y12, )1 IT) (9.7.4)
= E,{ H F(x0 + F(Teitil + ui )P -Y"} .

t-i
The likelihood function of the model is the product of (9.7.4) over all individ-
uals i = 1, 2, . . . , n. It is assumed that (ct ) are i.i.d. over individuals.
The expectation occurring in (9.7.4) is computationally cumbersome in

practice, except for the case where F = (1), the standard normal distribution
function, and 14; is also normal. In this case, Butler and Moffitt (1982) pointed
out that using the method of Gaussian quadrature, an evaluation of (9.7.4) for
a model with, say, n = 1500 and t = 10 is within the capacity of most re-
searchers having access to a high-speed computer. We shall come back to a
generalization of model (9.7.4) in Section 9.7.2. But here we shall discuss a
computationally simpler model in which (9.7.4) is expressed as a product and
ratio of gamma functions.
Heckman and Willis proposed what they called a beta-logistic model, de-
fined as24
P(Yit = 1 I ui) = ur (9.7.5)
i = 1, 2, . . . , n, t = 1, 2, . . . , T,
where ui is distributed according to the beta density
f(a ; + b,)
fi(u;) (1 (9.7.6)
1-(a01-(bduri
0 ui 1, ai > 0, bi > 0,
where f(a) = fipca-1 e -x dx. It is assumed that (y e ) are serially independent
conditional on tt. (Independence over individuals is always assumed.)
The beta density has mean (a + bra and variance a(a + b)-2(a +
b + 1)ab and can take various shapes: a bell shape if a, b> 1; monotonically
increasing if a> 1 and b < 1; monotonically decreasing if a < 1 and b> 1;
uniform if a = b = 1; and U shaped if a, b < 1.
We have (suppressing subscript i)
P(y, = 1, y,_1= 1) Eu2
P(31 e= 1 1Y1--i= 1) Eu (9.7.7)
P(yt-i= 1 )
> Eu= P(y,= 1),
where the inequality follows from Vu> 0.
Heckman and Willis postulated a i = exp (x) and bi = exp (x;fl), where xi
is a vector of time-independent characteristics of the ith individual. Thus,
from (9.7.5), we have
P(ye = 1) = A[x;(a fl)]. (9.7.8)
Equation (9.7.8) shows that if we consider only marginal probabilities we have
a logit model; in this sense a beta-logistic model is a generalization of a logit
model. By maximizing flr_ I A[x;(a fl)] we can obtain a consistent esti-

mate of a /3. However, we can estimate each ofa andficonsistently and also
more efficiently by maximizing the full likelihood function. If the ith person
worked si periods out of T, i = 1, 2, . . . , n, the likelihood function of the
beta-logistic model is given by
L fl E[up(1 u i)T- ] (9.7.9)

e-i
n
rr F(a1 -1- bi ) Ra i + si )11-(bi + T si )
M F(ai + b i + T)
An undesirable feature of the beta-logistic model (9.7.9) is that the indepen-
dent variables x i are time independent. Thus, in their empirical study of 1583
women observed over a five-year period, Heckman and Willis were forced to
use the values of the independent variables at the first period as x i even though
some of the values changed with t.
9.7.2 Models with Heterogeneity and True State Dependence
In this subsection we shall develop a generalization of model (9.7.2) that can

incorporate both heterogeneity and true state dependence.
Following Heckman (1981a), we assume that there is an unobservable
continuous variable y: that determines the outcome of a binary variable y by
the rule
y= 1 if y: > 0 (9.7.10)
=0 otherwise.
In a very general model, y: would depend on independent variables, lagged
values of y*, lagged values of y iz , and an error term that can be variously
specified. We shall analyze a model that is simple enough to be computation-
ally feasible and yet general enough to contain most of the interesting features
of this type of model: Let
Y: = xiffi YYia- ' + vit, (9.7.11)
where for each i, vu is serially correlated in general. We define true state
dependenceto mean y 0 and heterogeneityto mean serial correlation of (yi ).
Model (9.7.2) results from (9.7.11) if we put y = 0 and vat = ui + ea , where
(Eft ) are serially independent. Thus we see that model (9.7.2) is restrictive not
only because it assumes no true state dependence but also because it assumes a
special form of heterogeneity. This form of heterogeneity belongs to a so-

called one-factor model and will be studied further in Section 9.7.3.
Here we shall assume that for each i, (v ft ) are serially correlated in a general
way and derive the likelihood function of the model. Because (ye ) are as-
sumed to be independent over i, the likelihood function is the product of
individual likelihood functions. Therefore we suppress i and consider the
likelihood function of a typical individual.
Define T-vectors y, yr, and v the tth elements of which are y tv and v
respectively, and assume v N(0, 2). Then the joint probability of y (hence, a
typical individual likelihood function) conditional on yo can be concisely
expressed as
P(y) = F[yi * (2y 1); 2 * (2y 1)(2y 1)1, (9.7.12)
where * denotes the Hadamard product (see Theorem 23 of Appendix 1), I is a
T-vector of ones, and F(x; 2) denotes the distribution function of N(0, 2)
evaluated at x. Note that in deriving (9.7.12) we assumed that the conditional
distribution of v given yo is N(0, I). This, however, may not be a reasonable
assumption.
As with continuous variable panel data discussed in Section 6.6.3, the
specification of the initial conditions y io (reintroducing subscript i) is an
important problem in model (9.7.11), where T is typically small and N is large.
The initial conditions y io can be treated in two ways. (We do assume they are
observable. Treating them as N unknown parameters is not a good idea when
N is large and T is small.)
I. Treat y,t, as known constants, 1 or 0. Then the likelihood function is as
given in (9.7.12).
2. Assume that y io is a random variable with the probability distribution
P(y io = 1) = (1)(x;oa), where a is a vector of unknown parameters.
9.7.3 One-Factor Models

The individual likelihood function (9.7.12) involves a T-tuple normal inte-
gral, and therefore its estimation is computationally infeasible for large T(say,
greater than 5). For this reason we shall consider in this subsection one-factor
models that lead to a simplification of the likelihood function.
We assume
viz = at ui + ei (9.7.13)
where (as ), t = 1, 2, . . , T, are unknown parameters, U. and (eft ) are nor-
mally distributed independent of each other, and (,j are serially indepen-
dent. We suppress subscript i as before and express (9.7.13) in obvious vector
notation as
v ---- au + e, (9.7.14)
where v, a, and e are T-vectors and u is a scalar. Then the joint probability of y
can be written as
P(y) = E ,F[y/ (2y 1); D * (2y 1)(2y 1)'], (9.7.15)
where iv now includes au and D = Eee' . Because D is a diagonal matrix, F in
(9.7.15) can be factored as the product of T normal distribution functions. The
estimation of this model, therefore, is no more difficult than the estimation of
model (9.7.4).
For the case T = 3, model (9.7.14) contains a stationary first-order autore-
gressive model (see Section 5.2.1) as a special case. To see this, put a =
(1 P2)-"2 (p) 1 ) P)' Vu = ("2, and take the diagonal elements of D as a 2, 0,

and o-2. Thus, if T = 3, the hypothesis of AR(1) can easily be tested within the
more general model (9.7.14). Heckman (1981c) accepted the AR(1) hypoth-
esis using the same data for female labor participation as used by Heckman
and Willis (1977). If T> 4, model (9.7.13) can be stationary if and only if a, is
constant for all t. A verification of this is left as an exercise.
Consider a further simplification of (9.7.13) obtained by assuming a, = 1,
(ui ) are i.i.d. over i, and (ER) are i.i.d. both over i and t. This model differs from
model (9.7.2) only in the presence ofy i.,_ , among the right-hand variables and
is analogous to the Balestra-Nerlove model (Section 6.6.3) in the continuous
variable case.
As in the Balestra-Nerlove model, (ut ) may be regarded as unknown param-
eters to estimate. If both N and T go to 00, A 7, and (ui) can be consistently
estimated. An interesting question is whether we can estimate 11 and 7 consist-
ently when only Ngoes to 00. Unlike the Balestra-Nerlove model, the answer to
this question is generally negative for the model considered in this subsection.
In a probit model, for example, the values of p and 7 that maximize
'S T
L = H H cb(wi,+ ui)mi eqvi,+ (9.7.16)
1 t-1
while treating (ui) as unknown constants, are not consistent. Heckman

(1981b), in a small-scale Monte Carlo study for a probit model with n = 100
and T = 8, compared this estimator (with y iv treated as given constants),
called the transformation estimator, to the random-effect probit MLE, in

which we maximize the expectation of (9.7.16) taken with respect to (141 ),
regarded as random variables, and specify the probability of y it, under the
specification 2 given at the end of Section 9.7.2.
Heckman concluded that (1) if 7 = 0, the transformation estimator per-
formed fairly well relative to the random-effect probit MLE; (2) if y # 0, the
random-effect probit MLE was better than the transformation estimator, the
latter exhibiting a downward bias in 7 as in the Balestra-Nerlove model (see
Section 6.6.3).
Exercises
1. (Section 9.2.1)
In a Monte Carlo study, Goldfeld and Quandt (1972, Chapter 4) gener-
ated (y e ) according to the model P(ye = 1) = 0(0.2 + 0.5x11 + 2x21 ) and,
using the generated (y e ) and given (x11 , x2e ), estimated the fi's in the linear
probability model P(yi= 1 )7 Po + flixii Ax21 Their estimates were
0.58, flt =0.1742, and /92 = 0.7451. How do you convert these esti-
mates into the estimates of the coefficients in the probit model?
2. (Section 9.2.2)
Consider a logit model P(ye =1) = A(110 + Ax e ), where x e is a binary
variable taking values 0 and I. This model can be also written as a linear
probability model P(ye = 1) = yo + vi xi .
a. Determine yo and yi as functions of flo and /31 .
b. Show that the MLE of A, and y i are equal to the least squares
estimates in the regression of ye on x; with an intercept.
3. (Section 9.2.3)
Show that global concavity is not invariant to a one-to-one transforma-
tion of the parameter space.
4. (Section 9.2.8)
In the model of Exercise 2, we are given the following data:
x 1 1 1 0 0 0 0 0 1 0 1
y 0 0 1 0 0 1 1 0 1 0 1.
Calculate the MLE and the DA estimates (with 1 0 = ) of /30 and fli .
5. (Section 9.2.8)
The following data come from a hypothetical durable goods purchase
study:
r,
Case (t) Constant n, r,
n, 4) ()
-1 n, C6 [1?-1 C)]
? n,
I
1 1 5 25 12 0.4800 0.0500 0.3984

2 1 7 26 16 0.6154 0.2930 0.3822
3 1 10 31 22 0.7097 0.5521 0.3426
4 1 15 27 21 0.7778 0.7645 0.2979
a. Compute the coefficient estimates A, and A using the following
models and estimators:
(1) Linear Probability ModelLeast Squares
(2) Linear Probability ModelWeighted Least Squares
(3) Logit Minimum x2
(4) Probit Minimum X2
(5) Discriminant Analysis Estimator'
b. For all estimators except (5), find the asymptotic variance-covar-
iance matrix (evaluated at the respective coefficient estimates).
c. For all estimators, obtain estimates of the probabilities P, corre-
sponding to each case t.
d. Rank the estimators according to each of the following criteria:
(1) E (fit 1302.
(2) nA) (Pt Pi )2.

(3) log L 1[re log .P,+ (n, r,) log (1 Pi )].
6. (Section 9.2.9)
It may be argued that in (9.2.62) the asymptotic variance of r should
be the unconditional variance of r, which is tr 2E7_ 1 F,(1 F,)+
V(priX7.. 1 F1 ), where V is taken with respect to random variables (x,).
What is the fallacy of this argument?
7. (Section 9.3.3)
In the multinomial logit model (9.3.34), assume j = 0, 1, and 2. For this
model define the NLGLS iteration.
8. (Section 9.3.5)
Suppose (y,}, i = 1, 2, . . . , n, are independent random variables taking
three values, 0, 1, and 2, according to the probability distribution defined
by (9.3.51) and (9.3.52), where we assume it0 = 0, = x;fil , and p2 =
x'ifl2 . Indicate how we can consistently estimate 1/ 1 , /12 , and p using only a
binary logit program.
9. (Section 9.3.5)
Write down (9.3.59) and (9.3.60) in the special case where S= 2, B 1 =
(1, 2), and B2 = (3, 4) and show for which values of the parameters the
model is reduced to a four-response independent logit model.
10. (Section 9.3.5)
You are to analyze the decision of high school graduates as to whether or
not they go to college and, if they go to college, which college they go to.
For simplicity assume that each student considers only two possible col-
leges to go to. Suppose that for each person 1,1= 1, 2, . . . , n, we observe
zi (family income and levels of parents' education) and x u (the quality
index and the cost of the jth school), j = 1 and 2. Also suppose that we
observe for every person in the sample whether or not he or she went to
college but observe a college choice only for those who went to college.
Under these assumptions, define your own model and show how to esti-
mate the parameters of the model (cf. Radner and Miller, 1970).
11. (Section 9.3.6)
Write down (9.3.64), (9.3.65), and (9.3.66) in the special case ofFigure 9.1
(three-level), that is, C 1 = (1, 2), C2 = (3, 4), B,= (1, 2), B2 = (3, 4),
B3 = (5, 6), and B4 = (7, 8).
12. (Section 9.3.10)
Suppose that yi takes values 0, 1, and 2 with the probability distribution
P(y,= 0) = A(x14 ) and P(y1 = lly1 0 0) = ). Assuming that we
have n i independent observations on yr with the same value x i of the
independent variables, t = 1, 2, . . . , T, indicate how to calculate the
MIN x2 estimates of /10 and /II .
13. (Section 9.4.3)

Consider two jointly distributed discrete random variables y and x such
that y takes two values, 0 and 1, and x takes three values, 0, 1, and 2. The
most general model (called the saturated model) is the model in which
there is no constraint among the five probabilities that characterize the
joint distribution. Consider a specific model (called the null hypothesis
model) in which
P(y = 11x) = [1 + exp ( a fix)]-1
and the marginal distribution ofx is unconstrained. Given n independent
observations on (y, x), show how to test the null hypothesis against the
saturated model. Write down explicitly the test statistic and the critical
region you propose.
14. (Section 9.4.3)
Suppose the joint distribution of y 11 and y2 t = 1, 2, . . . , T, is given by
the following table:
Y21
1
1 d7 exp (a'x, + fi'z g ) d7' exp (oex,)
0 d d
where d, = exp (a'; + + exp (a'x,) + 2. Given n, independent ob-
servations on (Yu, Y2t), define the minimum chi-square estimates of the
vectors a and ft. Assume that x, and z, are vectors of known constants.
15. (Section 9.4.3)
Let j = 1, 2, . . . , J, be discrete random variables taking Ni values.
Suppose the conditional probability P(yJIYI , , i, Y.,+1, , .r)
is given for every j and is positive. Then prove that there is at most one set
of joint probabilities consistent with the given conditional probabilities
(cf. Amemiya, 1975).
16. (Section 9.4.3)
Let yi and y2 be binary variables taking the value 1 or 0. Show that if
P(y = 1 13'2) = A(xli 1312 Y2 ) and gy2 = 1 1Y i) = A(x'fi2 1- At Yi ), then
P12 =Ai.
17. (Section 9.5.2)
In the simple logit model defined in the paragraph after (9.5.22), the
efficiency of the WMLE of fl using the optimal h* depends only on p and
Q0. Prove Eff [h*(P, Q0), 11, Qo] = Eff[h*(P, 1 Q0), 11, 1
18. (Section 9.5.2)

In the same model described in Exercise 17, show that if Q= 0.5,
h* = 0.5.
19. (Section 9.5.3)

In the same model described in Exercise 17, show that the asymptotic
variance of the MME of /lo is the same as that of the WMLE.
20. (Section 9.5.3)
Show that (9.5.36) follows from (9.5.19).
21. (Section 9.6.3)
We are given the following data:
i 1 2 345
yi 1 0 0 1 1
xli 1 1 0 0 1
x21 0 1 1 1 0
a. Obtain the set of /3 values that maximize
S(fl) = [YdC(13x11+ x2i 0) (I - Yi)X(I 3X11 x21 < 0)l,
where x(E) = 1 if event E occurs, or is zero otherwise.

b. Obtain the set of /3 values that maximize
'(fl, F.) = CV; log Rfixii + x21)
+ (1 yi ) log [1 Mix" + x21)]),

where F is also chosen to maximize tv among all the possible distribution
functions. (Note that I have adopted my own normalization, which may
differ from Mansld's or Cosslett's, so that the parameter space of/3 is the
whole real line.)
22. (Section 9.6.3)
Cosslett (1983, p. 780) considered two sequences of y; ordered according
to the magnitude of xdi:
Sequence A: 1 0 1 0 1 1 1 1
Sequence B: 0 1 1 1 1 1 1 0
He showed that Sequence A yields a higher value of log L whereas Se-
quence B yields a higher score. Construct two sequences with nine obser-
vations each and an equal number of ones such that one sequence yields a
higher value of log L and the other sequence yields a higher score. Your
sequence should not contain any of Cosslett's sequences as a subset.
23. (Section 9.7.2)

Assume that fi = 0 and (vi,) are i.i.d. across i in model (9.7.11) and derive
the stationary probability distribution of (y ft ) for a particular i.
24. (Section 9.7.3)
Show that if T> 4, model (9.7.13) can be stationary if and only if a, is
constant for all t. Show that if T = 4 and a, > 0, then model (9.7.13) can
be stationary if and only if a, is constant for all t.
10 Tobit Models
10.1 Introduction
Tobit models refer to censored or truncated regression models in which the
range of the dependent variable is constrained in some way. In economics,
such a model was first suggested in a pioneering work by Tobin (1958). He
analyzed household expenditure on durable goods using a regression model
that specifically took account of the fact that the expediture (the dependent
variable of his regression model) cannot be negative. Tobin called his model
the model of limited dependent variables. It and its various generalizations are
known popularly among economists as Tobit models, a phrase coined by
Goldberger (1964), because of similarities to probit models. These models are
also known as censored or truncated regression models. The model is called
truncated if the observations outside a specified range are totally lost and
censored if we can at least observe the exogenous variables. A more precise
definition will be given later.
Censored and truncated regression models have been developed in other
disciplines (notably, biometrics and engineering) more or less independently
of their development in econometrics. Biometricians use the model to analyze
the survival time of a patient. Censoring or truncation occurs when either a
patient is still alive at the last observation date or he or she cannot be located.
Similarly, engineers use the model to analyze the time to failure of material or
of a machine or of a system. These models are called survival or duration
models. Sociologists and economists have also used survival models to ana-
lyze the duration of such phenomena as unemployment, welfare receipt,
employment in a particular job, residence in a particular region, marriage, and
the period of time between births. Mathematically, survival models belong to
the same general class of models as Tobit models; survival models and Tobit
models share certain characteristics. However, because survival models pos-
sess special features, they will be discussed separately in Chapter 11.
Between 1958 when Tobin's article appearedand 1970, the Tobit
model was used infrequently in econometric applications, but since the early
1970s numerous applications ranging over a wide area of economics have
appeared and continue to appear. This phenomenon is due to a recent in-
Tobit Models 361
crease in the availability of micro sample survey data, which the Tobit model
analyzes well, and to a recent advance in computer technology that has made
estimation of large-scale Tobit models feasible. At the same time, many gener-
alizations of the Tobit model and various estimation methods for these
models have been proposed. In fact, models and estimation methods are now
so numerous and diverse that it is difficult for econometricians to keep track of
all the existing models and estimation methods and maintain a clear notion of
their relative merits. Thus it is now particularly useful to examine the current
situation and prepare a unified summary and critical assessment of existing
results.
We shall try to accomplish this objective by means of classifying the diverse
Tobit models into five basic types. (Our review of the empirical literature
suggests that roughly 95% of the econometric applications of Tobit models fall
into one of these five types.) Although there are many ways to classify Tobit
models, we have chosen to classify them according to the form of the likeli-
hood function. This way seems to be the statistically most useful classification
because a similarity in the likelihood function implies a similarity in the
appropriate estimation and computation methods. It is interesting to note
that two models that superficially seem to be very different from each other
can be shown to belong to the same type when they are classified according to
this scheme.
Sections 10.2 through 10.5 will deal with the standard Tobit model (or Type
1 Tobit), and Sections 10.6 through 10.10 will deal with the remaining four
types of models. Basic estimation methods, which with a slight modification
can be applied to any of the five types, will be discussed at great length in
Section 10.4. More specialized estimation methods will be discussed in rele-
vant passages throughout the chapter. Each model is illustrated with a few
empirical examples.
We shall not discuss clisequilibrium models except for a few basic models,
which will be examined in Section 10.10.4. Some general references on dis-
equilibrium models will be cited there. Nor shall we discuss the related topic of
switching regression models. For a discussion of these topics, the reader
should consult articles by Maddala (1980, 1983). We shall not discuss Tobit
models for panel data (individuals observed through time), except to mention
a few articles in relevant passages.
10.2 Standard Tobit Model (Type 1 Tobit Model)

Tobin (1958) noted that the observed relationship between household ex-
penditures on a durable good and household incomes looks like Figure 10.1,
Expenditure
0 Income
Figure 10.1 An example of censored data
where each dot represents an observation for a particular household. An

important characteristic of the data is that there are several observations
where the expenditure is 0. This feature destroys the linearity assumption so
that the least squares method is clearly inappropriate. Should we fit a nonlin-
ear relationship? First, we must determine a statistical model that can generate
the kind of data depicted in Figure 10.1. In doing so the first fact we should
recognize is that we cannot use any continuous density to explain the condi-
tional distribution of expenditure given income because a continuous density
is inconsistent with the fact that there are several observations at 0. We shall
develop an elementary utility maximization model to explain the phenome-
non in question.
Define the symbols needed for the utility maximization model as follows:
y a household's expenditure on a durable good
Yo the price of the cheapest available durable good
z all the other expenditures
x income
A household is assumed to maximize utility U(y, z) subject to the budget
constraint y + z x and the boundary constraint y yo or y =0. Suppose y*
is the solution of the maximization subject to y + z x but not the other
constraint, and assume y* = fl + /32x + u, where u may be interpreted as the
collection of all the unobservable variables that affect the utility function.
Then the solution to the original problem, denoted by y, can be defined by
Y=Y* if Y * > Yo (10.2.1)
=0 or yo if Y* YO
Tobit Models 363
If we assume that u is a random variable and that yo varies with households but
is assumed known, this model will generate data like Figure 10.1. We can write
the likelihood function for n independent observations from the model
(10.2.1) as
L = fi F4Y01) 1 (10.2.2)
where F and f', are the distribution and density function, respectively, of yr,
no means the product over those i for which yr -5 yoi , and n, means the
product over those i for which 31> yoi . Note that the actual value of y when
y* g yo has no effect on the likelihood function. Therefore the second line of
Eq. (10.2.1) may be changed to the statement "if y* g yo , one merely observes
that fact."
The model originally proposed by Tobin (1958) is essentially the same as
the model given in the preceding paragraph except that he specifically as-
sumed y* to be normally distributed and assumed y o to be the same for all
households. We define the standard Tobit model (or Type 1 Tobit) as follows:
= fi + ul, j == 1, 2, . . . , n, (10.2.3)
= yr if >0 (10.2.4)
=0 if yr g 0,
where {u i) are assumed to be i.i.d. drawings from N(0, 0 .2). It is assumed that
(yi) and (xi) are observed for i = 1, 2, . . . , n but (yr) are unobserved if
yr g 0. Defining X to be then X K matrix the ith row of which is x, we assume
that {xi ) are uniformly bounded and lim_ n-1 X'X is positive definite. We
also assume that the parameter space of /3 and a2 is compact. In the Tobit
model we need to distinguish the vectors and matrices of positive observations
from the vectors and matrices of all the observations; the latter appear with an
underbar.
Note that yr > 0 and y* g 0 in (10.2.4) may be changed to yr > y o and
g yo without essentially changing the model, whether yo is known or un-
known, because yo can be absorbed into the constant term of the regression. If,
however, yo; changes with i and is known for every i, the model is slightly
changed because the resulting model would be essentially equivalent to the
model defined by (10.2.3) and (10.2.4), where one of the elements of /3 other
than the constant term is known. The model where yoi changes with i and is
unknown is not generally estimable.
The likelihood function of the standard Tobit model is given by
L = fi [1 (xd 0)] fi 0-10[(y,- leifl)/0 ], (10.2.5)
where cto and 44 are the distribution and density function, respectively, of the
standard normal variable.
The Tobit model belongs to what is sometimes known as the censored
regression model. In contrast, when we observe neither yi nor xi when y* g: 0,
the model is known as a truncated regression model. The likelihood function
of the truncated version of the Tobit model can be written as
L= 11 (1)(4/3/0) - ' 0.- '0[(YI 70)/a]. (10.2.6)
Henceforth, the standard Tobit model refers to the model defined by (10.2.3)
and (10.2.4), namely, a censored regression model, and the model the likeli-
hood function of which is given by (10.2.6) will be called the truncated stan-
dard Tobit model.
10.3 Empirical Examples

Tobin (1958) obtained the maximum likelihood estimates of his model ap-
plied to data on 735 nonfarm households obtained from Surveys of Consumer
Finances. The dependent variable of his estimated model was actually the
ratio of total durable goods expenditure to disposable income and the inde-
pendent variables were the age of the head of the household and the ratio of
liquid assets to disposable income.
Since then, and especially since the early 1970s, numerous applications of
the standard Tobit model have appeared in economic journals, encompassing
a wide range of fields in economics. A brief fist of recent representative refer-
ences, with a description of the dependent variable (y) and the main indepen-
dent variables (x), is presented in Table 10.1. In all the references, except that
by Kotlikoff, who uses a two-step estimation method to be discussed later, the
method of estimation is maximum likelihood.
10.4 Properties of Estimators under Standard Assumptions

In this section we shall discuss the properties of various estimators of the Tobit
model under the assumptions of the model. The estimators we shall consider
are the probit maximum likelihood (ML), least squares (LS), Heckman's
two-step least squares, nonlinear least squares (NLLS), nonlinear weighted
least squares (NLWLS), and Tobit ML estimators.
Table 10.1 Applications of the standard Tobit model
Reference
Adams (1980) Inheritance Income, marital status,
number of children
Ashenfelter and Ham Ratio of unemployed Years of schooling,
(1979) hours to employed working experience
hours
Fair (1978) Number of extramarital Sex, age, number of years
affairs married, number of
children, education,
occupation, degree of
religiousness
Keeley et al. (1978) Hours worked after a Preprogram hours
negative income tax worked, change in the
program wage rate, family
characteristics
Kotlikoff (1979) Expected age of retirement Ratio of social security
benefits lost at time of
full-time employment
to full-time earnings
Reece (1979) Charitable contributions Price of contributions,
income
Rosenzweig (1980) Annual days worked Wages of husbands and
wives, education of
husbands and wives,
income
Stephenson and Family earnings after a Earnings before the
McDonald (1979) negative income tax program, husband's
program and wife's education,
other family
characteristics,
unemployment rate,
seasonal dummies
Wiggins (1981) Annual marketing of new Research expenditure of
chemical entities the pharmaceutical
industry, stringency of
government regulatory
standards
Witte (1980) Number of arrests (or Accumulated work release
convictions) per month funds, number of
after release from prison months after release
until first job, wage rate
after release, age, race,
drug use
10.4.1 Probit Maximum Likelihood Estimator

The Tobit likelihood function (10.2.5) can be trivially rewritten as
L =fi [ 1 - cD(x;fl/a)] 11 vxdva, (10.4.1)
H (13(y4fl/a) - Irk x;flya].
Then the first two products of the right-hand side of (10.4.1) constitute the
likelihood function of a probit model, and the last product is the likelihood
function of the truncated Tobit model as given in (10.2.6). The probit ML
estimator of a fl/ a, denoted is obtained by maximizing the logarithm of
,
the first two products.

Note that we can only estimate the ratio flia by this method and not flora
separately. Because the estimator ignores a part of the likelihood function that
involves fl and a, it is not fully efficient. This loss of efficiency is not surprising
when we realize that the estimator uses only the sign of yfi, ignoring its
numerical value even when it is observed.
From the results of Section 9.2 we see that the probit MLE is consistent and
follows
a A (X'D 120- 'X'DIPOw Ey.), (10.4.2)
where P o is the n X n diagonal matrix the ith element of which is 0(30), DI is

the n X n diagonal matrix the ith element of which is (130(x;a)-1[1
4)(x;a)] -1 4)(x;a)2, and w is the n-vector the ith element w of which is defined
by
1 if y* > 0 (10.4.3)
0 if yr 0.
Note that the ith element of Ew is equal to (1)(x;a). The symbol A means that
both sides have the same asymptotic distribution) As shown in Section 9.2.2,
a is asymptotically normal with mean a and asymptotic variance-covariance
matrix given by
Vet -= (X/X)'. (10.4.4)
10.4.2 Least Squares Estimator

From Figure 10.1 it is clear that the least squares regression of expenditure on
income using all the observations including zero expenditures yields biased
Tobit Models 367
estimates. Although it is not so clear from the figure, the least squares regres-
sion using only the positive expenditures also yields biased estimates. These
facts can be mathematically demonstrated as follows.
First, consider the regression using only positive observations of y 1 . We
obtain from (10.2.3) and (10.2.4)
E(yilyi > 0) = xfl + E(uilui > x;fl). (10.4.5)
The last term of the right-hand side of (10.4.5) is generally nonzero (even
without assuming ui is normal). This implies the biasedness of the LS estima-
tor using positive observation on yi under more general models than the
standard Tobit model. When we assume normality of u; as in the Tobit model,
(10.4.5) can be shown by straightforward integration to be
E( yilyi > 0) = x;/3 + (7.1(x;i9/0, (10.4.6)
where A(z) = 41(z)/(1)(z). 2 As shown later, this equation plays a key role in the
derivation of Heckman's two-step, NLLS, and NLWLS estimators.
Equation (10.4.6) clearly indicates that the LS estimator of ft is biased and
inconsistent, but the direction and magnitude of the bias or inconsistency
cannot be shown without further assumptions. Goldberger (1981) evaluated
the asymptotic bias (the probability limit minus the true value) assuming that
the elements of; (except the first element, which is assumed to be a constant)
are normally distributed. More specifically, Goldberger rewrote (10.2.3) as
Y =A + flt + ui (10.4.7)
and assumed N(0, I), distributed independently of ui . (Here, the as-
sumption of zero mean involves no loss of generality because a nonzero mean
can be absorbed into ft0.) Under this assumption he obtained
y
plim ... (10.4.8)
1 P2 Y
where y = ay-1.1.(fl0 /ay)[fl0 + ay (flo /ay)] and p2 = ay-2ficIth , where

a?, = a2 + /3fi, . It can be shown that 0 < y < 1 and 0 <p2 < 1; therefore
(10.4.8) shows that ft1 shrinks ft1 toward 0. It is remarkable that the degree of
shrinkage is uniform in all the elements of fi i . However, the result may not
hold if is not normal; Goldberger gave a nonnormal example where fi l =
(1, 1)' and plim fis, = (1.111, 0.887)'.
Consider the regression using all the observations of yi , both positive and 0.
To see that the least squares estimator is also biased in this case, we should look
at the unconditional mean of yi,
Ey; = (D(x;fl/a)x;fl + crO(x;Ncr). (10.4.9)

Writing (10.2.3) again as (10.4.7) and using the same assumptions as Gold-
berger, Greene (1981) showed
= evoicroP. , (10.4.10)
where ft1 is the LS estimator of fl in the regression of yi on ii using all the
observations. This result is more useful than (10.4.8) because it implies that
(07 1 )131 is a consistent estimator of fl1, where n 1 is the number of positive
observations of yi . A simple consistent estimator of flo can be similarly ob-
tained. Greene (1983) gave the asymptotic variances of these estimators.
Unfortunately, however, we cannot confidently use this estimator without
knowing its properties when the true distribution of f, is not norma1. 3
10.4.3 Heckman's Two-Step Estimator

Heckman (1976a) proposed a two-step estimator in a two-equation general-
ization of the Tobit model, which we shall call the Type 3 Tobit model. But his
estimator can also be used in the standard Tobit model, as well as in more
complex Tobit models, with only a minor adjustment. We shall discuss the
estimator in the context of the standard Tobit model because all the basic
features of the method can be revealed in this model. However, we should
keep in mind that since the method requires the computation of the probit
MLE, which itself requires an iterative method, the computational advantage
of the method over the Tobit MLE (which is more efficient) is not as great in
the standard Tobit model as it is in more complex Tobit models.
To explain this estimator, it is useful to rewrite (10.4.6) as
= x;fl + aA(x'i a) + Ei, for i such that yi > 0, (10.4.11)
where we have written a Na as before and e i = E( Ay ; >0) so that

Eft =0. The variance of Ei is given by
ile; = 0 a2x;aA(x;a) a 2A.(x;a)2. (10.4.12)
Thus (10.4.11) is a heteroscedastic nonlinear regression model with n 1 obser-

vations. The estimation method Heckman proposed consists of two steps:
Step 1. Estimate a by the probit MLE (denoted et) defined earlier.
Step 2. Regress yi on xi and .1.(46) by least squares, using only the positive
observations on yi .
To facilitate further the discussion of Heckman's estimator, we can rewrite
Tobit Models 369
(10.4.11) again as
= x; fi + a + ei + (10.4.13)
for i such that y,> 0,
where m = a[A(x'i ot) A(x;ii)]. We can write (10.4.13) in vector notation as
y=Xfi+al,++ti, (10.4.14)
where the vectors y, e, and ti have n, elements and matrix X has n, rows,
corresponding to the positive observations of yi . We can further rewrite
(10.4.14) as
y=27++1, (10.4.15)
where we have defined 2 = (X, I) and 7 = (fi' , a)'. Then Heckman's two-step
estimator of 7 is defined as
= (2,2)-I2'y. (10.4.16)
The consistency of follows easily from (10.4.15) and (10.4.16). We shall
derive its asymptotic distribution for the sake of completeness, although the
result is a special case of Heckman's result (Heckman, 1979). From (10.4.15)
and (10.4.16) we have
V(9 7) = (nr'2'2) -1 (nV/22'e + nV/22'n). (10.4.17)
Because the probit MLE cit is consistent, we have
plim = firn (10.4.18)
where Z = (X, A). Under the assumptions stated after (10.2.4), it can be shown
that
ni 1/22'e N(0, a2 lim (10.4.19)
where a21 Ea' is the n, X n, diagonal matrix the diagonal elements of

which are Tie, given in (10.4.12). We have by Taylor expansion of )(x &)
around A(x
= a (ii a) + 0(n- ). (10.4.20)
Using (10.4.20) and (10.4.2), we can prove
ni 1/22'n > N[0, a2Z'(I I)X(X'D 12)-1 X'(I 1)Z], (10.4.21)

where DI was defined after (10.4.2). Next, note that e and 7/ are uncorrelated
because n is asymptotically a linear function of w on account of (10.4.2) and
(10.4.20) and e and w are uncorrelated. Therefore, from (10.4.17), (10.4.18),
(10.4.19), and (10.4.21), we finally conclude that Y. is asymptotically normal
with mean y and asymptotic variance-covariance matrix given by
= a2(Z'Z)-1Z1 [I + (I 1,)X(Li'D 1 X)- 'X'( I 1)J7,(EZ)-1 .
(10.4.22)
Expression (10.4.22) may be consistently estimated either by replac-

ing the unknown parameters by their consistent estimates or by
(Z'Z) 'Z'AZ(Z'Z)- where A is the diagonal matrix the ith diagonal element
of which is [y; xfi ii2(x;iii)] 2, following the idea of White (1980).
Note that the second matrix within the square bracket in (10.4.22) arises
because A had to be estimated. If) were known, we could apply least squares
directly to (10.4.11) and the exact variance-covariance matrix would be
0.2(Ez) lEuwzy1 .
Heckman's two-step estimator uses the conditional mean of y i given in

(10.4.6). A similar procedure can also be applied to the unconditional mean of
yi given by (10.4.9). 4 That is to say, we can regress all the observations of y i ,
including zeros, on Cox i and ct after replacing the a that appears in the argu-
ment ofcI) and ct by the probit MLE a. In the same way as we derived (10.4.11)
and (10.4.13) from (10.4.6), we can derive the following two equations from
(10.4.9):
= (1)(70)[x;/3 + a2(30)] +6, (10.4.23)
and
= (1)(x;ei)[x;fl + ail.(xA)] + 6; + (10.4.24)
where 6; = y Ey ; and [b(70) 0:10(70)]x;/3 + cr[4)(x;a) 0(x;ci)]. A

vector equation comparable to (10.4.15) is
y=1152y+.5+, (10.4.25)
where to is the n X n diagonal matrix the ith element of which ise(x;eit). Note
that the vectors and matrices appear with underbars because they consist of n
elements or rows. The two-step estimator of y based on all the observations,
denoted 57, is defined as
)7= (2451)- 115y. (10.4.26)
Tobit Models 371
The estimator can easily be shown to be consistent. To derive its asymptotic

distribution, we obtain from (10.4.25) and (10.4.26)
y) = (n-I1522)-'01- 1122115,5 .4_ n-1/2211%). (10.4.27)
Here, unlike the previous case, an interesting fact emerges: By expanding
01(x;e1) and q5(x;i1) in Taylor series around Y4ci we can show =
Therefore
plim n-1/21'1% = 0. (10.4.28)
Corresponding to (10.4.18), we have
plim n- '21522 = rim 'Z'D2Z, (10.4.29)
where D is obtained from ti by replacing ei with a. Corresponding to (10.4.19),
we have
n- '12115(5-. N(0, a2 lim n 'Z'D 2L1Z), (10.4.30)
where en EtS5' is the n X n diagonal matrix the ith element of which is

a24)(x;a)((x/ia)2 + x;aA(x;a) + 1 cD(x;a)[x;a + gx;a)] 2). Therefore, from
(10.4.27) through (10.4.30), we conclude that )7is asymptotically normal with
mean y and asymptotic variance-covariance matrix given by'
Try' = (72(ZID2Z) IZIDVIAZID2Zy I . (10.4.31)
Which of the two estimators 9 and )7 is preferred? Unfortunately, the differ-

ence of the two matrices given by (10.4.22) and (10.4.31) is generally neither
positive definite nor negative definite. Thus an answer to the preceding ques-
tion depends on parameter values.
Both (10.4.15) and (10.4.25) represent heteroscedastic regression models.
Therefore we can obtain asymptotically more efficient estimators by using
weighted least squares (WLS) in the second step of the procedure for obtaining
9 and )7. In doing so, we must use a consistent estimate of the asymptotic
variance-covariance matrix of e + t, for the case of (10.4.15) and of + for
the case of (10.4.25). Because these matrices depend on y, an initial consistent
estimate of y (say, 9 or )7) is needed to obtain the WLS estimators. We call these
WLS estimators 9A, and )7w , respectively. It can be shown that they are consist-
ent and asymptotically normal with asymptotic variance-covariance matrices
given by
= a2(ER, + (I I)X(X13 12.0-1 V(I 1)]-1Z)-1 (10.4.32)
and
I/9w = a2(Z'D 2L1-1Z)- I. (10.4.33)
Again, we cannot make a definite comparison between the two matrices.
10.4.4 Nonlinear Least Squares and Nonlinear Weighted Least

Squares Estimators
In this subsection we shall consider four estimators: the NLLS and NLWLS
estimators applied to (10.4.11), denoted 9 N and pN, respectively, and the
NLLS and NLWLS estimators applied to (10.4.23), denoted ci N and itNIN
respectively.
All these estimators are consistent and their asymptotic distributions can be
obtained straightforwardly by noting that all the results of a linear regression
model hold asymptotically for a nonlinear regression model if we treat the
derivative of the nonlinear regression function with respect to the parameter
vector as the regression matrix. 6 In this way we can verify the interesting fact
that 9N and 9Nw have the same asymptotic distributions as 9 and 9w , respec-
tively. 7 We can also show that 9N and 9Nw are asymptotically normal with
mean y and with their respective asymptotic variance-covariance matrices
given by
9N= a2(S'S)-1 SIS(SiS) I
11 (10.4.34)
and
o.2(s ,r is) (10.4.35)
where S = DA, where D2 is the n 1 X ni diagonal matrix the ith element
of which is 1 + (30)2 + 4a01,(70). We cannot make a definite comparison
either between (10.4.22) and (10.4.34) or between (10.4.32) and (10.4.35).
In the two-step methods defining and 9 and their generalizations 9 w and
)7w , we can naturally define an iteration procedure by repeating the two steps.
For example, having obtained 9, we can obtain a new estimate of a, insert it
into the argument of A, and apply least squares again to Eq. (10.4.11). The
procedure is to be repeated until a sequence of estimates of a thus obtained
converges. In the iteration starting from 9 w , we use the mth-round estimate of
y not only to evaluate A but also to estimate the variance-covariance matrix of
the error term for the purpose of obtaining the (m + 1)st-round estimate.
Iterations starting from "'and 9w can be similarly defined but are probably not
Tobit Models 373
worthwhile because 9 and iTw are asymptotically equivalent to A,/ and Aqw , as
we indicated earlier. The estimators O N , imv , fro iNw) are clearly stationary
values of the iterations starting from (9, fw , 9, Av). However, they may not
necessarily be the converging values.
A simulation study by Wales and Woodland (1980) based on only one
replication with sample sizes of 1000 and 5000 showed that A, / is distinctly
inferior to the MLE and is rather unsatisfactory.
10.4.5 Tobit Maximum Likelihood Estimator

The Tobit MLE maximizes the likelihood function (10.2.5). Under the as-
sumptions given after (10.2.4), Amemiya (1973c) proved its consistency and
asymptotic normality. If we define e= (fl', a2)', the asymptotic variance-co-
variance matrix of the Tobit MLE 0 is given by
1 -1
n
. , [nI appe, 1 b,; 1..1
VO = i"'n n (10.4.36)
E b x; I cii
i- 1 i- 1
where
ai = (1-2{x;a0i [4,1/( 1 (1)1)] (Deb
bi = (112)o--3((x;a)24); + [(x'i a)cMI(1 COD, and
ci = (1/4)a-4{(x;a)30; + (Via)ctli [(Via)01/(1 (Di)] 21:1););
and cbi and ctoi stand for 0(x;a) and (1)(x;a), respectively.
The Tobit MLE must be computed iteratively. Olsen (1978) proved the
global concavity of log L in the Tobit model in terms of the transformed
parameters a = Pa and h = a',a result that implies that a standard iterative
method such as Newton-Raphson or the method of scoring always converges
to the global maximum of log L. 8 The log L in terms of the new parameters
can be written as
log L = log [1 4)(x;a)] + ni log h (10.4.37)
(hy i
2
from which Olsen obtained
a2 log L a log L (10.4.38)

acraa' &rah
(32 log L a2 log L
[
ahace ah2
xix; 0
[ 1 lksoi
nl
0
h2
[ x ix; x iyi
11 Yix'i E1 Y?
Because x;rx [1 (1)(x;a)] -10(x;a) < 0, the right-hand side of (10.4.38) is

the sum of two negative-definite matrices and hence is negative definite.
Even though convergence is assured by global concavity, it is a good idea to
start an iteration with a good estimator because it will improve the speed of
convergence. Tobin (1958) used a simple estimator based on a linear approxi-
mation of the reciprocal of Mills' ratio to start his iteration for obtaining the
MLE. Although Amemiya (1973c) showed that Tobin's initial estimator is
inconsistent, empirical researchers have found it to be a good starting value for
iteration.
Amemiya (1973) proposed the following simple consistent estimator. We
have
E(ylly i > 0) = (x;/))2 + crx;flA.(x;a) + a2. (10.4.39)
Combining (10.4.6) and (10.4.39) yields
E(Y?lYi > 0) = x;fiE(YilYi > 0) + a 2, (10.4.40)

which can be alternatively written as
A = yix;fl + a2 + Ca , for i such that ,y1 > 0, (10.4.41)
where E(Cilyi > 0) =- 0. Then consistent estimates off/ and a 2 are obtained by
applying an instrumental variables method to (10.4.41) using (94 1) as the
instrumental variables, where pi is the predictor of yi obtained by regressing
Tobit Models 375
positive yi on xi and, perhaps, powers of;. The asymptotic distribution of the

estimator has been given by Amemiya (1973c). A simulation study by Wales
and Woodland (1980) indicated that this estimator is rather inefficient.
10.4.6 The EM Algorithm

The EM algorithm is a general iterative method for obtaining the MLE; it was
first proposed by Hartley (1958) and was generalized by Dempster, Laird, and
Rubin (1977) to a form that is especially suited for censored regression models
such as Tobit models. We shall first present the definition and the properties of
the EM algorithm under a general setting and then apply it to the standard
Tobit model. It can also be applied to general Tobit models.
We shall explain the EM algorithm in a general model where a vector of
observable variables z is related to a vector of unobservable variables y* in
such a way that the value of y* uniquely determines the value of z but not vice
versa. In the Tobit model, (y7) defined in (10.2.3) constitute the elements of
y*, and (JO and (wi) defined in (10.2.4) and (10.4.3), respectively, constitute
the elements of z. Let the joint density or probability of y* bef(y*) and let the
joint density or probability of z be g(z). Also, define k(y*lz) = J.( y*)/g(z). Note
that f( z) = f(z 1 y*)f(y*) =f(y) because f(z 1 y*) = 1 inasmuch as y*
uniquely determines z. We implicitly assume that f g, and k depend on a
vector of parameters 0. The purpose is to maximize
L(0) =. n- ' log g(z) = log f( y*) n- ' log k(y*lz) (10.4.42)
with respect to 0. Define
Q(0101) = E[n- ' log f(y*10)1z, 0 1 ], (10.4.43)
where we are taking the expectation assuming 0 1 is the true parameter value
and doing this conditional on z. Then the EM algorithm purports to maximize
L(0) by maximizing Q(010 1 ) with respect to 0 where 01 is given at each step of
the iteration. The "E" of the name "EM" refers to the expectation taken in
(10.4.43) and the "M" refers to the maximization of (10.4.43).
Consider the convergence properties of the EM algorithm. Define
H(0101) = E[n- ' log k( y*I z, 0)1z, Of]. (10.4.44)
Then we have from (10.4.42), (10.4.43), and (10.4.44) and the fact that
L(010 1 ) E[rz-' log g(z)1z, 0 1 ] = L(0)
L(0) = Q(0191) H(010 1 ). (10.4.45)
But we have by Jensen's inequality (4.2.6)
H(0101) <H(01 181) for 0 # 01 . (10.4.46)
Now, given 0 1 , let M(0 1) maximize Q(010 1) with respect to 0. Then we have
L( = Q(m I el 1-1(M101).
)
(10.4.47)
But, because Q(M10 1 ) 12(0101) by definition and H(M10 1 ) H(01 101 ) by

(10.4.46), we have from (10.4.45) and (10.4.47)
L(M).. L(0 1). (10.4.48)
Thus we have proved the desirable property that L always increases or stays
constant at each step of the EM algorithm.
The preceding result implies that if L is bounded, then lim,...,,. L(0,) exists.
Let 0* satisfy the equality lim, L(0,)= L(0*). (0* exists if the range of L (0)
is closed.) We shall show that if 0* is a stationary point of the EM algorithm,
then 0* is a stationary point of L. For this we assume L is twice differentiable.
Differentiating (10.4.45) with respect to 0 and evaluating the derivative at
0= 0,, we obtain
aLl = aQ(010 1
a0 le, a el
)1 all(0101)
ao
(10.4.49)
But the last term of the right-hand side of (10.4.49) is 0 because of (10.4.46).
Therefore, if 0, is a stationary point of Q(91 0 i), it is a stationary point of L.
Unfortunately, a local maximum of Q(010 1 ) with respect to 0 may not be a
local maximum (let alone the global maximum) of L(0) because the negative
definiteness of [82Q(e1emea0'] 6, does not imply the negative definiteness of
[a2L1 awn, . However, this is true of any iterative method commonly used.
See Wu (1983) for further discussion of the convergence properties of the EM
algorithm.
Now consider an application of the algorithm to the Tobit mode1. 9 Define
0 = ( fic a2) as before. Then in the Tobit model we have
1 "
log f(y*10) = --2- log a 2
2a2 (y? - x;fl)2, (10.4.50)
and, for a given estimate 8 1 = (pc, of), the EM algorithm maximizes with

Tobit Models 377
respect to /3 and a2
E[logf(y*IMY,w01] (10.4.51)
n 2 1 1
= log
ER Y - 30)21 Wi = 01]
2a2 (Yi X;P)2 2(72
1 1
= log (12 (y1 /7)2
a2 - - - [E(Yr I = 0, 01) -
20.2
1
V(y1 I lwi= 0, 00,
- 2a 2
where
= 0, 0)= 4111 1 on = Y? (10.4.52)
and
= 0, ) = a? + 34/31 [ 1 ai ]2 , (10.4.53)
where Oil = fa,) and On = cri(xV3 1 /ai ).

From (10.4.51) it is clear that the second-round estimate of /3 in the EM
algorithm, denoted /12 , is obtained as follows: Assume without loss of general-
ity that the first n, observations of yi are positive and call the vector of those
observations y. Next, define an (n n1 ) vector y the elements of which are
-
the y? defined in (10.4.52). Then we have
Qcxrixtrol, (10.4.54)
where X was defined after (10.2.4). In other words, the EM algorithm amounts
to predicting all the unobservable values of yr by their conditional expecta-
tions and treating the predicted values as if they were the observed values. The
second-round estimate of a 2, denoted o-i, is given by
(10.4.55)
0
+ o, 0,)].
We can directly show that the MLE 0 is the equilibrium solution of the
iteration defined by (10.4.54) and (10.4.55). Partition X = (X', X')' so that
a
Xis multiplied by y and X by y. Then inserting into both sides of (10.4.54)
yields, after collecting terms,
acb(4/6) (10.4.56)
X/30= X'y X'
Li (D(x/j/c7)J'
where the last bracket denotes an (n n 1 )-dimensional vector the typical
element of which is given inside. Now, setting the derivative of log L with
respect to ft equal to 0 yields
a 1 cbi
(Di x + (yi x'S); = 0. (10.4.57)
But, clearly, (10.4.56) is equivalent to (10.4.57). Similarly, the normal equa-

tion for 172 can be shown to be equivalent to (10.4.55).
Schmee and Hahn (1979) performed a simulation study of the EM algo-
rithm applied to a censored regression model (a survival model) defined by
Yr = if 31 1` 5 c
=c if yr > c,
where N(a + fixi , a2 ). They generally obtained rapid convergence.
10.5 Properties of the Tobit Maximum Likelihood Estimator under

Nonstandard Assumptions
In this section we shall discuss the properties of the Tobit MLE the estima-
tor that maximizes (10.2.5) under various types of nonstandard assump-
tions: heteroscedasticity, serial correlation, and nonnormality. It will be
shown that the Tobit MLE remains consistent under serial correlation but not
under heteroscedasticity or nonnormality. The same is true of the other esti-
mators considered earlier. This result contrasts with the classical regression
model in which the least squares estimator (the MLE under the 'normality
assumption) is generally consistent under all of the three types of nonstandard
assumptions mentioned earlier.
Before proceeding with a rigorous argument, we shall give an intuitive
explanation of the aforementioned result. By considering (10.4.11) we see that
serial correlation of y, should not affect the consistency of the NLLS estimator,
whereas heteroscedasticity changes a to cri and hence invalidates the estima-
tion of the equation by least squares. If)/ is not normal, Eq. (10.4.11) itself is
generally invalid, which leads to the inconsistency of the NLLS estimator.
Tobit Models 379
Although the NLLS estimator is different from the ML estimator, we can

expect a certain correspondence between the consistency properties of the two
estimators.
10.5.1 Heteroscedasticity
Hurd (1979) evaluated the probability limit of the truncated Tobit MLE when
a certain type of heteroscedasticity is present in two simple truncated Tobit
models: (1) the i.i.d. case (that is, the case of the regressor consisting only of a
constant term) and (2) the case of a constant term plus one independent
variable. Recall that the truncated Tobit model is the one in which no infor-
mation is available for those observations for which 3,1 1 <0 and therefore the
MLE maximizes (10.2.6) rather than (10.2.5).
In the i.i.d. case Hurd created heteroscedasticity by generating rn observa-
tions from N(y, al) and (1 r)n observations from N( 1.1, al). In each case he
recorded only positive observations. Let y, i = 1, 2, . . . , n l , be the recorded
observations. (Note n 1 5 n). We can show that the truncated Tobit MLE ofp
and a2, denoted fi and 6 2, are defined by equating the first two population
moments of y, to their respective sample moments:
ni
6A(i116)= ni l I Yr (10.5.1)
and
ni
12 2 (212A (i2/6) + (2.2 = nr Y?. (10.5.2)
1
Taking the probability limit of both sides of (10.5.1) and (10.5.2) and express-
ing plim Fir 'Iyi and plim nryl as certain functions of the parameters
it, al, al, and r, we can define plim 2 and plim 62 implicitly as functions of
these parameters. Hurd evaluated the probability limits for various values ofp
and a, after having fixed r = 0.5 and a2 = 1. Hurd found large asymptotic
biases in certain cases.
In the case of one independent variable, Hurd generated observations from
N(a F flx, al) after having generated xi and log 1a11from bivariate
N(0, 0, V, V, p). For given values of a, /3, V1 , V2 and p, Hurd found the
values of a, )6, and a 2 that maximize E log L, where L is as given in (10.2.6).
Those values are the probability limits of the MLE of a, fl, and a 2 under
Hurd's model if the expectation of log L is taken using the same model. Again,
Hurd found extremely large asymptotic biases in certain cases.
Arabmazar and Schmidt (1981) showed that the asymptotic biases of the
censored Tobit MLE in the i.i.d. case are not as large as those obtained by
Hurd.
10.5.2 Serial Correlation

Robinson (1982a) proved the strong consistency and the asymptotic normal-
ity of the Tobit MLE under very general assumptions about ui (normality is
presupposed) and obtained its asymptotic variance-covariance matrix, which
is complicated and therefore not reproduced here. His assumptions are
slightly stronger than the stationarity assumption but are weaker than the
assumption that 14 possesses a continuous spectral density (see Section 5.1.3).
His results are especially useful because the full MLE that takes account of
even a simple type of serial correlation seems computationally intractable.
The autocorrelations of u need not be estimated to compute the Tobit MLE
but must be estimated to estimate its asymptotic variance-covariance matrix.
The consistent estimator proposed by Robinson (1982b) may be used for that
purpose.
10.5.3 Nonnormality
Goldberger (1983) considered an i.i.d. truncated sample model in which data
are generated by a certain nonnormal distribution with mean it and variance 1
and are recorded only when the value is smaller than a constant c. Let y
represent the recorded random variable and let 37 be the sample mean. The
researcher is to estimate p by the MLE, assuming that the data are generated
by N(p, 1). As in Hurd's i.i.d. model, the MLE it is defined by equating the
population mean of y to its sample mean:
A (c 1 2) = (10.5.3)
Taking the probability limit of both sides of (10.5.3) under the true model and
putting plim ft = it* yield
(10.5.4)
where h(c p) = E(p yly < c), the expectation being taken using the true
model. Defining m = p* p and 0 = c p, we can rewrite (10.5.4) as
m = A(0 m) h(0). (10.5.5)
Tobit Models 381
Goldberger calculated m as a function of 0 when the data are generated by

Student's t with various degrees of freedom, Laplace, and logistic distribu-
tions. The asymptotic bias was found to be especially great when the true
distribution was Laplace. Goldberger also extended the analysis to the regres-
sion model with a constant term and one discrete independent variable.
Arabmazar and Schmidt (1982) extended Goldberger's analysis to the case of
an unknown variance and found that the asymptotic bias was further accen-
tuated.
10.5.4 Tests for Normality

The fact that the Tobit MLE is generally inconsistent when the true distribu-
tion is nonnormal makes it important for a researcher to test whether the data
are generated by a normal distribution. Nelson (1981) devised tests for nor-
mality in the i.i.d. censored sample model and the Tobit model. His tests are
applications of the specification test of Hausman (1978) (see Section 4.5.1).
Nelson's i.i.d. censored model is defined by
= .11 if >0
-- 0 if y* g 0, 1 = 1, 2, . . . , n,
where ;I N(g, a 2 ) under the null hypothesis. Nelson considered the esti-
mation of P(34' > 0). Its MLE is 02)(i2ic9), where it and 6 are the MLE of the
respective parameters. A consistent estimator is provided by n, In, where, as
before, n 1 is the number of positive observations of y. Clearly, nl ln is a
consistent estimator of P(y* > 0) under any distribution, provided that it is
i.i.d. The difference between the MLE and the consistent estimator is used as a
test statistic in Hausman's test. Nelson derived the asymptotic variances of the
two estimators under normality.
If we interpret what is being estimated by the two estimators as
n- I P(yr > 0), Nelson's test can be interpreted as a test of the null
hypothesis against a more general misspecification than just nonnormality. In
fact, Nelson conducted a simulation study to evaluate the power of the test
against a heteroscedastic alternative. The performance of the test was satisfac-
tory but not especially encouraging.
In the Tobit model Nelson considered the estimation of
n- IEX ' y = Ix, [0(x;a)x;fi + a4)(x ;a)]. Its MLE is given by the right-
hand side of this equation evaluated at the Tobit MLE, and its consistent
estimator is provided by rt - X' y. Hausman's test based on these two estima-
tors will work because this consistent estimator is consistent under general
distributional assumptions on y. Nelson derived the asymptotic variance-co-

variance matrices of the two estimators.
Nelson was ingenious in that he considered certain functions of the original
parameters for which estimators that are consistent under very general as-
sumptions can easily be obtained. However, it would be better if a general
consistent estimator for the original parameters themselves could be found.
An example is Powell's least absolute deviations estimator, to be discussed in
the next subsection.
Bera, Jarque, and Lee (1982) proposed using Rao's score test in testing for
normality in the standard Tobit model where the error term follows the
two-parameter Pearson family of distributions, which contains normal as a
special case.
10.5.5 Powell's Least Absolute Deviations Estimator

Powell (1981, 1983) proposed the least absolute deviations (LAD) estimator
(see Section 4.6) for censored and truncated regression models, proved its
consistency under general distributions, and derived its asymptotic distribu-
tion. The intuitive appeal for the LAD estimator in a censored regression
model arises from the simple fact that in the i.i.d. sample case the median (of
which the LAD estimator is a generalization) is not affected by censoring
(more strictly, left censoring below the mean), whereas the mean is. In a
censored regression model the LAD estimator is defined as that which mini-
mizes 17_ 1 1 max (0, 4/01. The motivation for the LAD estimator in a
truncated regression model is less obvious. Powell defined the LAD estimator
in the truncated case as that which minimizes Mk, Iy max (27: 3/ 1 ,34/1)I. In
the censored case the limit distribution of Vii(fl /3), where ft is the LAD
estimator, is normal with zero mean and variance-covariance matrix [4f(0) 2
n, n-1 17_ Ix(x;fl > 0)x1 x;]- ', wherefis the density of the error term and xlim
is the indicator function taking on unity if xdi > 0 holds and 0 otherwise. In
the truncated case the limit distribution of ITO A is normal with zero
mean and variance-covariance matrix 2- 1A- IBA- 1, where
A = lim n-1 E x(70 > 0)[f(0) f(x;13)]F(Vr

and
B = lim x(xdi > 0)[F(4) F(0)1F(xdir I XiX;

/7-4140
1
where F is the distribution function of the error term.

Tobit Models 383
Powell's estimator is attractive because it is the only known estimator that is

consistent under general nonnormal distributions. However, its main draw-
back is its computational difficulty. Paarsch (1984) conducted a Monte Carlo
study to compare Powell's estimator, the Tobit MLE, and Heckman's two-
step estimator in the standard Tobit model with one exogenous variable under
situations where the error term is distributed as normal, exponential, and
Cauchy. Paarsch found that when the sample size is small (50) and there is
much censoring (50% of the sample), the minimum frequently occurred at the
boundary of a wide region over which a grid search was performed. In large
samples Powell's estimator appears to perform much better than Heckman's
estimator under any of the three distributional assumptions and much better
than the Tobit MLE when the errors are Cauchy.
Another problem with Powell's estimator is finding a good estimator of the
asymptotic variance-covariance matrix that does not require the knowledge
of the true distribution of the error. Powell (1983) proposed a consistent
estimator.
Powell observed that his proof of the consistency and asymptotic normality
of the LAD estimator generally holds even if the errors are heteroscedastic.
This fact makes Powell's estimator even more attractive because the usual
estimators are inconsistent under heteroscedastic errors, as noted earlier.
Another obvious way to handle nonnormality is to specify a nonnormal
distribution for the ui in (10.2.3) and use the MLE. See Amemiya and Bosldn
(1974), who used a lognormal distribution with upper truncation to analyze
the duration of welfare dependency.
10.6 Generalized Tobit Models

As stated in Section 10.1, we can classify Tobit models into five common types
according to similarities in the likelihood function. Type 1 is the standard
Tobit model, which we have discussed in the preceding sections. In the follow-
ing sections we shall define and discuss the remaining four types of Tobit
models.
It is useful to characterize the likelihood function of each type of model
schematically as in Table 10.2, where each y1, j = 1, 2, and 3, is assumed to be
distributed as N(xA, ), and P denotes a probability or a density or a
combination thereof. We are to take the product of each P over the observa-
tions that belong to a particular category determined by the sign of y1 . Thus, in
Type 1 (standard Tobit model), P(y i <0) P(y1 ) is an abbreviated notation
for II0 P(yri <0) Il lfu (y"), where fu is the density of N(xl J11 , at). This
Table 10.2 Likelihood functions of the five types of Tobit models
Type Likelihood function

1 P(y, <0) P(y,)
2 P(y, < 0) P(Y1 > 0, Y2)
3 P(yi <0) P(yi , y2 )
4 P(yi <0, Y3) P(YI, Y2)
5 P(yi <0, 3/3) P(Yi >0, Y2)
expression can be rewritten as (10.2.5) after dropping the unnecessary sub-

script 1.
Another way to characterize the five types is by the classification of the three
dependent variables that appear in Table 10.3. In each type of model, the sign
of yi determines one of the two possible categories for the observations, and a
censored variable is observed in one category and unobserved in the other.
Note that when yl is labeled C, it plays two roles: the role of the variable the
sign of which determines categories and the role of a censored variable.
We allow for the possibility that there are constraints among the parameters
of the model (,fl, ), j = 1, 2, or 3. For example, constraints will occur if the
original model is specified as a simultaneous equations model in terms of
, Y2 and y3 . Then the /I's denote the reduced-form parameters.
We shall not discuss models in which there is more than one binary variable
and, hence, models the likelihood function of which consists of more than two
components. Such models are computationally more burdensome because
they involve double or higher-order integration ofjoint normal densities. The
only exception occurs in Section 10.10.6, which includes models that are
Table 10.3 Characterization of the five types of Tobit models
Dependent variables
Type Yi Y2 Y3
1 C
2 B C
3 C C
4 C C C
5 B C C
Note: C = censored; B = binary.
Tobit Models 385
obvious generalizations of the Type 5 Tobit model. Neither shall we discuss a

simultaneous equations Tobit model of Amemiya (1974b). The simplest two-
equation case of this model is defined by y u = max (y, y2; + xci/31 + u , 0) and
y2; = max (y2 yu + ,4i/32 + 142i , 0), where (u u , u21 ) are bivariate normal and
yl y2 < I must be assumed for the model to be logically consistent? ) A sche-
matic representation of the likelihood function of this two-equation model is
RYI, Y2) RYI < /3 3) 9 )1 P(Y2 < 0 9 Y4) Rh < OP Y4 < 0 )
with y's appropriately defined.
10.7 Type 2 Tobit Model: P(Yi <0) P(Yi > 0, Y2)

10.7.1 Definition and Estimation
The Type 2 Tobit model is defined as follows:
= xirfil uir (10.7.1)
= x02 + u2i
Y2i=Y if y7>O
=0 if yri 5 0, i = 1, 2, . . . , n,
where (u 11 , u21 ) are i.i.d. drawings from a bivariate normal distribution with
zero mean, variances ai and al, and covariance a 12 . It is assumed that only
the sign of yri is observed and that Ai is observed only when yri > 0. It is
assumed that x11 are observed for all i but that x21 need not be observed for i
such that yri 0. We may also define, as in (10.4.3),
=1 if yri > 0 (10.7.2)
=0 if
Then (Mr, hi) constitute the observed sample of the model. It should be
noted that, unlike the Type 1 Tobit, y21 may take negative values." As in
(10.2.4), y21 = 0 merely signifies the event yri g 0.
The likelihood function of the model is given by
L= P(3)* 0) Hf(y211Yri > 0)P(31; > 0), (10.7.3)
where Ho and H i stand for the product over those i for which y2; =0 and
Y2, # 0, respectively, andf( lyri > 0) stands for the conditional density of yri
; > 0. Note the similarity between (10.4.1) and (10.7.3). As in Type 1 givenyr
Tobit, we can obtain a consistent estimate of $ 1 /a1 by maximizing the probit

part of (10.7.3).
Probit L = HP(yri 0) II P(yr; > 0). (10.7.4)
Also, (10.7.4) is a part of the likelihood function for every one of the five types
of models; therefore a consistent estimate of /3 1 /al can be obtained by the
probit MLE in each of these types of models.
We can rewrite (10.7.3) as
L = Hp(yiti:4 0) jj f /Uri, Y2i) dYri, (10.7.5)

0
wheref( , ) denotes the joint density of yr ; and yt. We can write the joint
density as the product of a conditional density and a marginal density, that is
1(Y ft, hi) flYril V(Y2i )9 and can determine a specific form for f(Y Y2i)
from the well-known fact that the conditional distribution of yt, given
yv is normal with mean /cc b-i B -a12aT - z 2 ( v2i - x1/i2) and variance
a? a 12 a2 2. Thus we can further rewrite (10.7.5) as
L = H [1 - (10(xc ifi1 ai')] (10.7.6)
X H (1)([xl ifl 1 cr1l + anal 'a 2 2 ( y21 x02)]
X [ 1 a2er1 2 a2 21"2)a2 '[a2 '(Y21 xii/32)1-

Note that L depends on al only through filar and cri2ai 1 ; therefore, if there is
no constraint on the parameters, we can put a l = 1 without any loss of gener-
ality. Then the remaining parameters can be identified. If, however, there is at
least one common element in fli and fl2 , al can also be identified.
We shall show how Heckman's two-step estimator can be used in this
model. To obtain an equation comparable to (10.4.11), we need to evaluate
E(yrilyt > 0). For this purpose we use
= x02 + cr12ai2(Yri xi1/31) + C211 (10.7.7)
where C2i is normally distributed independently of y'r; with zero mean and
variance ai a 2 ar2. Using (10.7.7), we can express E(yrilyri > 0) as a sim-
ple linear function of E(yrilyt > 0), which was already obtained in Section
10.4. Using (10.7.7), we can also derive V(yllyt> 0) easily. Thus we obtain
Y21 = X21/32 + Gri2ai A(xliai) + 21, (10.7.8)
for i such that y2; # 0,
Tobit Models 387
where al = E21 = 0, and

ai2ai 2[xiia1 1(xl1a1) Mxiia1)21. (10.7.9)
As in the case of the Type 1 Tobit, Heckman's two-step estimator is the LS
estimator applied to (10.7.8) after replacing al with the probit MLE. The
asymptotic distribution of the estimator can be obtained in a manner similar
to that in Section 10.4.3 by defining ,h ; in the same way as before. It was first
derived by Heckman (1979).
The standard Tobit (Type 1) is a special case of Type 2 in which yt = y't .
Therefore (10.7.8) and (10.7.9) will be reduced to (10.4.11) and (10.4.12) by
putting 4/31 = xA and al = cri = to-- 12
A generalization of the two-step method applied to (10.4.23) can easily be
defined for this model but will not be discussed.
Note that the consistency of Heckman's estimator does not require the joint
normality of)/ and y'2" provided that yr is normal and that Eq. (10.7.7) holds
with C2 independently distributed of yr but not necessarily normal (Olsen,
1980). For then (10.7.8) would still be valid. As pointed out by Lee (1982c),
the asymptotic variance-covariance matrix of Heckman's estimator can be
consistently estimated under these less restrictive assumptions by using
White's estimator analogous to the one mentioned after Eq. (10.4.22). Note
that White's estimator does not require (10.7.9) to be valid.
10.7.2 A Special Case of Independence

Dudley and Montmarquette (1976) analyzed whether or not the United States
gives foreign aid to a particular country and, if it does, how much foreign aid it
gives using a special case of the model (10.7.1), where the independence of u l;
u2i is assumed. In their model the sign of yr ; determines whether aid is and
given to the ith country, and y, determines the actual amount of aid. They
used the probit MLE to estimate /1 1 (assuming al = 1) and the least squares
regression of y2; on x2; to estimate A. The LS estimator of fi2 is consistent in
their model because of the assumed independence between u li and u21 . This
makes their model computationally advantageous. However, it seems unreal-
istic to assume that the potential amount of aid, y, is independent of the
variable that determines whether or not aid is given, yr. This model is the
opposite extreme of the Tobit model, which can be regarded as a special case of
Type 2 model where there is total dependence between y't and yr. in the whole
spectrum of models (with correlation between yr and y varying from 1 to
+ 1) contained in Type 2.
Because of the computational advantage mentioned earlier this "indepen-

dence" model and its variations were frequently used in econometric applica-
tions in the 1960s and early 1970s. In many of these studies, authors made the
additional linear probability assumption: P(yri > 0) = x 1ft1 , which enabled
them to estimate /A (as well as /3'2) consistently by the least squares method. For
examples of these studies, see the articles by Huang (1964) and Wu (1965).
10.7.3 Gronau's Model

Gronau (1973) assumed that the offered wage W is given to each housewife
independently of hours worked (H), rather than as a schedule W''(H). Given
W, a housewife maximizes her utility function U(C, X) subject to X =
WH + V and C + H = T, where C is time spent at home for childcare, X
represents all other goods, T is total available time, and V is other income.
Thus a housewife does not work if
rau (auril (10.7.10)
Lac \ax jii-O > w
and works if the inequality in (10.7.10) is reversed. If she works, the hours of
work H and the actual wage rate W must be such that
au ( aur _ w
ac\x)
Gronau called the left-hand side of (10.7.10) the housewife's value of time or,
more commonly, the reservation wage, denoted W r. 12
Assuming that both W and Wr can be written as linear combinations of
independent variables plus error terms, his model may be statistically de-
scribed as follows:
= x;.1fl2 + u2i (10.7.11)
= z;a + vi
W, = W7 if W' > W
=0 if W? 5 W, i = 1,2, . . . , n,
where (u21 , ) are i.i.d. drawings from a bivariate normal distribution with
zero mean, variances a! and 01, , and covariance cr. Thus the model can be
written in the form of (10.7.1) by putting W? Wri = yri and W? = ;it. Note
that H (hours worked) is not explained by this statistical model although it is
Tobit Models 389
determined by Gronau's theoretical model. A statistical model explaining H

as well as W was developed by Heckman (1974) and will be discussed in
Section 10.8.2.
Because the model (10.7.11) can be transformed into the form (10.7.1) in
such a way that the parameters of (10.7.11) can be determined from the
parameters of (10.7.1), all the parameters of the model are identifiable except
V( W? Wr,), which can be set equal to 1 without loss of generality. If, how-
ever, at least one element of x 2, is not included in z, , all the parameters are
identifiable.I 3 They can be estimated by the MLE or Heckman's two-step
estimator by procedures described in Section 10.7.1. We can also use the
probit MLE (the first step of Heckman's two-step) to estimate a certain subset
of the parameters. 14
10.7.4 Other Applications of the Type 2 Tobit Model

Nelson (1977) noted that a Type 2 Tobit model arises if yo in (10.2.1) is
assumed to be a random variable with its mean equal to a linear combination
of independent variables. He reestimated Gronau's model by the MLE.
In the study of Westin and Gillen (1978), yr represents the parking cost with
x2 including zonal dummies, wage rate (as a proxy for value of walking time),
and the square of wage rate. A researcher observes y = y2 if yI < C where C
represents transit cost, which itself is a function of independent variables plus
an error term.
10.8 Type 3 Tobit Model: P(y, <0) P(y., , y 2 )

Yri= 141i (10.8.1)
Yt= 1421
Yu= Yri if Yfi> 0
=0 if yr, g 0
Y2/ = ir if y r, > 0
=0 if yr, g 0, = 1, 2, . . . , n,
where (uu , u21 ) are i.i.d. drawings from a bivariate normal distribution with
zero mean, variances ai and al, and covariance cr 12 . Note that this model
differs from Type 2 only in that in this model yri is also observed when it is
positive.
Because the estimation of this model can be handled in a manner similar to
the handling of Type 2, we shall discuss it only briefly. Instead, in the following
we shall give a detailed discussion of the estimation of Heckman's model
(1974), which constitutes the structural equations version of the model
(10.8.1).
The likelihood function of the model (10.8.1) can be written as
L =fl P(y* 0) fir(yli, y21), (10.8.2)
wheref( , ) is the joint density of y, and y. Because yri is observed when it

is positive, all the parameters of the model are identifiable, including al .
Heckman's two-step estimator was originally proposed by Heckman
(1976a) for this model. Here we shall obtain two conditional expectation
equations, (10.4.11) and (10.7.8), for y i and y2 , respectively. [Add subscript 1
to all the variables and the parameters in (10.4.11) to conform to the notation
of this section.] In the first step of the method, al =ft1 01 1 is estimated by the
probit MLE . In the second step, least squares is applied separately to
(10.4.11) and (10.7.8) after replacing al by &I . The asymptotic variance-co-
variance matrix of the resulting estimates of (fl i , al ) is given in (10.4.22) and
that for (//2 , a
- 1 2a- ') can be similarly obtained. The latter is given by Heck-
man (1979). A consistent estimate of cr 2 can be obtained using the residuals of
Eq. (10.7.8). As Heckman (1976a) suggested and as was noted in Section
10.4.3, a more efficient WLS can be used for each equation in the second step
of the method. An even more efficient GLS can be applied simultaneously to
the two equations. However, even GLS is not fully efficient compared to
MLE, and the added computational burden of MLE may be sufficiently
compensated for by the gain in efficiency. A two-step method based on un-
conditional means of y1 and y2 , which is a generalization of the method
discussed in Section 10.4.3, can also be used for this model.
Wales and Woodland (1980) compared the LS estimator, Heckman's two-
step estimator, probit MLE, conditional MLE (using only those who worked),
MLE, and another inconsistent estimator in a Type 3 Tobit model in a
simulation study with one replication (sample size 1000 and 5000). The
particular model they used is the labor supply model of Heckman (1974),
which will be discussed in the next subsection." The LS estimator was found
Tobit Models 391
to be poor, and all three ML estimators were found to perform well. Heck-
man's two-step estimator was ranked somewhere between LS and MLE.
10.8.2 Heckman's Model

Beckman's model (Heckman, 1974) differs from Gronau's model (10.7.11) in
that Heckman included the determination of hours worked (H) in his
mode1. 16 Like Gronau, Heckman assumes that the offered wage W is given
independently of H; therefore Heckman's W equation is the same as
Gronau's:
W7 "="- 4112 + U2i (10.8.3)
Heckman defined Wr = oulacy(ulax) and specified"
W; = y1-1; + z:cx+ v i . (10.8.4)
It is assumed that the ith individual works if
Wri (Hi = 0) + I); < (10.8.5)
and then the wage Wi and hours worked 'hare determined by solving (10.8.3)
and (10.8.4) simultaneously after putting WT = W; = W. Thus we can de-
fine Heckman's model as
U21 (10.8.6)
and
= yH1 Z;01 vi (10.8.7)
for those i for which desired hours of work
Hit xrufh + Uli > 0, (10.8.8)
where x 1fl1 = (xA z;a) and u li = y- '(u21 vi ). Note that (10.8.5) and

(10.8.8) are equivalent because y> 0.

Call (10.8.6) and (10.8.7) the structural equations; then (10.8.6) and the
identity part of (10.8.8) constitute the reduced-form equations. The reduced-
form equations of Heckman's model can be shown to correspond to the Type
3 Tobit model (10.8.1) if we put H* = yr, H = yi , W = yt, and W= y2 .
We have already discussed the estimation of the reduced-form parameters
in the context of the model (10.8.1), but we have not discussed the estimation
of the structural parameters. Heckman (1974) estimated the structural param-
eters by MLE. In the next two subsections we shall discuss three alternative
methods of estimating the structural parameters.
10.8.3 Two-Step Estimator of Heckman

Heckman (1976a) proposed the two-step estimator of the reduced-form pa-
rameters (which we discussed in Section 10.8.1); but he also reestimated the
labor supply model of Heckman (1974) using the structural equations version.
Because (10.8.6) is a reduced-form as well as a structural equation, the estima-
tion of fl2 is done in the same way as discussed in Section 10.8.1, namely, by
applying least squares to the regression equation for E(Willf* > 0) after esti-
mating the argument of A (the hazard rate) by probit MLE. So we shall discuss
only the estimation of (10.8.7), which we can rewrite as
Hi = r I y- vi . (10.8.9)
By subtracting E(v,11-1* > 0) from vi and adding the same, we can rewrite
(10.8.9) further as
Hi = riWi Ziarl aiv ar r /al ) y- lei , (10.8.10)
where ah, = Coy (u h , vi ), cri = N o and e = vi > 0). Then the
consistent estimates of y -1 , a y-1 , and alai 1 y-1 are obtained by the least
squares regression applied to (10.8.10) after replacing /a, by its probit MLE
and Wi by W1 , the least squares predictor of Wi obtained by applying Heck-
man's two-step estimator to (10.8.6). The asymptotic variance-covariance
matrix of this estimator can be deduced from the results in the article by
Heckman (1978), who considered the estimation of a more general model
(which we shall discuss in the section on Type 5 Tobit models).
Actually, there is no apparent reason why we must first solve (10.8.7) for Hi
and proceed as indicated earlier. Heckman could jusfas easily have subtracted
and added E(vi llit > 0) to (10.8.7) itself and proceeded similarly. This
method would yield alternative consistent estimates. Inferring from the well-
known fact that the two-stage least squares estimates of the standard simulta-
neous equations model yield asymptotically equivalent estimates regardless of
which normalization is chosen, the Heckman two-step method applied to
(10.8.7) and (10.8.9) should also yield asymptotically equivalent estimates of y
and a.
Lee, Maddala, and Trost (1980) extended Heckman's simultaneous equa-
Tobit Models 393
tions two-step estimator and its WLS version (taking account of the hetero-
scedasticity) to more general simultaneous equations Tobit models and ob-
tained their asymptotic variance-covariance matrices.
10.8.4 Amemiya's Least Squares and Generalized Least

Squares Estimators
Amemiya (1978c, 1979) proposed a general method of obtaining the esti-
mates of the structural parameters from given reduced-form parameter esti-
mates in general Tobit-type models and derived the asymptotic distribution.
Suppose that a structural equation and the corresponding reduced-form
equations are given by
y = Yy + XJ) + u (10.8.11)
[y, Y] = X[n, II] + V,
where X 1 is a subset of X. Then the structural parameters y and fl are related to
the reduced-form parameters ir and n in the following way:
= fl + JP, (10.8.12)
where J is a known matrix consisting of only ones and zeros. If, for example,
X1 constitutes the first K 1 columns of X (K= K 1 + K2 ), then we have
J = (I, 0) ', where I is the identity matrix of size K 1 and 0 is the "2 X K1 matrix
of zeros. It is assumed that lc, y, and fl are vectors and n and J are matrices of
conformable sizes. Equation (10.8.12) holds for Heckman's model and more
general simultaneous equations Tobit models, as well as for the standard
simultaneous equations model (see Section 7.3.6).
Now suppose certain estimates ic and n of the reduced-form parameters are
given. Then, using them, we can rewrite (10.8.12) as
= + J + ( ir 7r) (11 11) y (10.8.13)
Amemiya proposed applying LS and GLS estimation to (10.8.13). From
Amemiya's result (Amemiya, 1978c), we can infer that Amemiya's GLS
applied to Heckman's model yields more efficient estimates than Heckman's
simultaneous equations two-step estimator discussed earlier. Amemiya
(1983b) showed the superiority of the Amemiya GLS estimator to the WLS
version of the Lee-Maddala-Trost estimator in a general simultaneous equa-
tions Tobit model.
10.8.5 Other Examples of Type 3 Tobit Models

Roberts, Maddala, and Enholm (1978) estimated two types of simultaneous
equations Tobit models to explain how utility rates are determined. One of
their models has a reduced form that is essentially Type 3 Tobit; the other is a
simple extension of Type 3.
The structural equations of their first model are
Y= !hi (10.8.14)
and
YYti xac u31, (10.8.15)
where yt is the rate requested by the ith utility firm, yti is the rate granted for
the ith firm, x2i includes the embedded cost of capital and the last rate granted
minus the current rate being earned, and x 3; includes only the last variable
mentioned. It is assumed that ylt and yt are observed only if
yri z;ct + vi > 0, (10.8.16)
where zi include the earnings characteristics of the ith firm. ( Vvi is assumed to
be unity.) The variable yt may be regarded as an index affecting a firm'deci-
sion as to whether or not it requests a rate increase. The model (10.8.14) and
(10.8.15) can be labeled as P(y, <0) P(yi >0, Y2 y3) in our shorthand
notation and therefore is a simple generalization of Type 3. The estimation
method of Roberts, Maddala, and Enholm is that of Lee, Maddala, and Trost
(1980) and can be described as follows:
Step 1. Estimate a by the probit MLE.
Step 2. Estimate /32 by Heckman's two-step method.
Step 3. Replace yt in the right-hand side of (10.8.15) by pit obtained in step
2 and estimate y and 113 by the least squares applied to (10.8.15) after adding
the hazard rate term E(u 31l > 0).
The second model of Roberts, Maddala, and Enholm is the same as the first
model except that (10.8.16) is replaced by
> Ri, (10.8.17)
where Ri refers to the current rate being earned, an independent variable. Thus
this model is essentially Type 3. (It would be exactly Type 3 if RE = 0.) The
estimation method is as follows:
Step 1. Estimate /32 by the Tobit MLE.
Step 2. Repeat step 3 described in the preceding paragraph.
Tobit Models 395
Nakamura, Nakamura, and Cullen (1979) estimated essentially the same

model as Heckman (1974) using Canadian data on married women. They
used the WLS version of Heckman's simultaneous equations two-step esti-
mators, that is, they applied WLS to (10.8.10).
Hausman and Wise (1976, 1977, 1979) used Type 3 and its generalizations
to analyze the labor supply of participants in the negative income tax (NIT)
experiments. Their models are truncated models because they used observa-
tions on only those persons who participated in the experiments. The first
model of Hausman and Wise (1977) is a minor variation of the standard Tobit
model, where earnings Y follow
Yi = Y* if Yt < Li , Y N(0, (72 ), (10.8.18)
where Li is a (known) poverty level that qualifies the ith person to participate
in the NIT program. It varies systematically with family size. The model is
estimated by LS and MLE. (The LS estimates were always found to be smaller
in absolute value, confirming Greene's result given in Section 10.4.2.) In the
second model of Hausman and Wise (1977), earnings are split into wage and
hours as Y = W H, leading to the same equations as those of Heckman (Eqs.
10.8.6 and 10.8.7) except that the conditioning event is
log W, + log Hi < log L, (10.8.19)
instead of (10.8.8). Thus this model is a simple extension of Type 3 and
belongs to the same class of models as the first model of Roberts, Maddala, and
Enholm (1978), which we discussed earlier, except for the fact that the model
of Hausman and Wise is truncated. The model of Hausman and Wise (1979)
is also of this type. The model presented in their 1976 article is an extension of
(10.8.18), where earnings observations are split into the preexperiment (sub-
script 1) and experiment (subscript 2) periods as
= Yr, and Y.21 = Y , if rt, < Li . (10.8.20)
Thus the model is essentially Type 3, except for a minor variation due to the
fact that L. with i.
10.9 Type 4 Tobit Model: P(y, <0, Y3) P(Y, , Y2)

Yr/ = x,p 1 + uli (10.9.1)
= xfl + u2i
= 'LA + 14 3,
=
=0 if yri .6 0
Y2i = 341 if >0
=0 if yri .6 0
Y3i = Yr/ 6 0
=0 if > 0, i = 1, 2, . n,
where (u 11 , uv , u31 ) are i.i.d. drawings from a trivariate normal distribution.
This model differs from Type 3 defined by (10.8.1) only by the addition of
Ai , which is observed only if A', 6 0. The estimation of this model is not
significantly different from that of Type 3. The likelihood function can be
written as
0
L=111 J3- 010 hi) dArlf2(Yii, Y21), (10.9.2)
o
where f3 ( , ) is the joint density of yr, and yt and f2 ( , ) is the joint
density of yri and 34. Heckman's two-step method for this model is similar to
the method for the preceding model. However, we must deal with three
conditional expectation equations in the present model. The equation for
will be slightly different from the other two because the variable is nonzero
when yr, is nonpositive. We obtain
gY31iYri 5- 0 ) = x;1fl3 - 07 1 ( ). (10.9.3)
We shall discuss three examples of the Type 4 Tobit model in the following
subsections: the model of Kenny et al. (1979); the model of Nelson and Olson
(1978); and the model of Tomes (1981). In the first two models the y* equa-
tions are written as simultaneous equations, like Heckman's model (1974), for
which the reduced-form equations take the form of (10.9.1). Tomes' model
has a slight twist. The estimation of the structural parameters of such models
can be handled in much the same way as the estimation of Hecicman's model
(1974), that is, by either Heckman's simultaneous equations two-step method
(and its Lee-Maddala-Trost extension) or by Amemiya's LS and GLS, both of
which were discussed in Section 10.8.
Tobit Models 397
In fact, these two estimation methods can easily accommodate the follow-
ing very general simultaneous equations Tobit model:
r'ylt = B'x; + ui , i = 1, 2, . . . , n, (10.9.4)
where the elements of the vector y 1 1 contain the following three classes of
variables: (1) always completely observable, (2) sometimes completely observ-
able and sometimes observed to lie in intervals, and (3) always observed to lie
in intervals. Note that the variable classified as C m Table 10.3 belongs to class
2 and the variable classified as B belongs to class 3. The models of Heckman
(1974), Kenny et al. (1979), and Nelson and Olson (1978), as well as a few
more models discussed under Type 5, such as that of Heckman (1978), are all
special cases of the model (10.9.4).
10.9.2 Model of Kenny, Lee, Maddala, and Trost

Kenny et al. (1979) tried to explain earnings differentials between those who
went to college and those who did not. We shall explain their model using the
variables appearing in (10.9.1). In their model, yr refers to the desired years of
college education, to the earnings of those who go to college, and yto the
earnings of those who do not go to college. A small degree of simultaneity is
introduced into the model by letting yrappear in the right-hand side of the
equation. Kenny and his coauthors used the MLE. They noted that the MLE
iterations did not converge when started from the LS estimates but did con-
verge very fast when started from Heckman's two-step estimates (simulta-
neous equations version).
10.9.3 Model of Nelson and Olson

The empirical model actually estimated by Nelson and Olson (1978) is more
general than Type 4 and is a general simultaneous equations Tobit model
(10.9.4). The Nelson-Olson empirical model involves four elements of the
vector y*:
yr Time spent on vocational school training, completely observed if
> 0, and otherwise observed to lie in the interval ( co, 0]
3, 11 Time spent on college education, observed to lie in one of the three
intervals (cc, 0], (0, 1], and (1, 00)
Wage, always completely observed
y: Hours worked, always completely observed
These variables are related to each other by simultaneous equations. How-

ever, they merely estimate each reduced-form equation seperately by various
appropriate methods and obtain the estimates of the structural parameters
from the estimates of the reduced-form parameters in an arbitrary way.
The model that Nelson and Olson analyzed theoretically in more detail is
the two-equation model:
Yu = 71 X1 Vli (10.9.5)
and
Y21 = 724i + X2112 V21 (10.9.6)
where y21 is always observed and y is observed to be yii if yri > 0. This model
may be used, for example, if we are interested in explaining only yr and yin
the Nelson-Olson empirical model. The likelihood function of this model may
be characterized by P(yi <0, Y2) P(Yi , Y2), and therefore, the model is a
special case of Type 4.
Nelson and Olson proposed estimating the structural parameters of this
model by the following sequential method:
Step 1. Estimate the parameters of the reduced-form equation for yr by the
Tobit MLE and those of the reduced-form equation for y2 by LS.
Step 2. Replace y21 in the right-hand side of (10.9.5) by its LS predictor
obtained in step 1 and estimate the parameters of (10.9.5) by the Tobit MLE.
Step 3. Replace yr; in the right-hand side of (10.9.6) by its predictor ob-
tained in step 1 and estimate the parameters of (10.9.6) by LS.
Amemiya (1979) obtained the asymptotic variance-covariance matrix of the
Nelson-Olson estimator and showed that the Amemiya GLS (see Section
10.8.4) based on the same reduced-form estimates is asymptotically more
efficient.
10.9.4 Model of Tomes

Tomes (1981) studied a simultaneous relationship between inheritance and
the recipient's income. Although it is not stated explicitly, Tomes' model can
be defined by
Y = Y1 hi xcifli + u u , (10.9.7)
Y21 = 72.V1i X21112 U219 (10.9.8)

and

Tobit Models 399
Yu yri
y11 if yfi > (10.9.9)
=0 if yri 0,
where yt is the potential inheritance, y ii is the actual inheritance, and Y21 is the
recipient's income. Note that this model differs from Nelson's model defined
by (10.9.5) and (10.9.6) only in that y ,, not yt, appears in the right-hand side
of (10.9.8). Assuming yi y2 < 1 for the logical consistency of the model (as in
Amemiya, 1974b, mentioned in Section 10.6), we can rewrite (10.9.7) as
= (1 Y1 Y2 )- I [YI (X21112 + U21) x, 1fi uu] (10.9.10)
and (10.9.8) as
Y21 = Y(2I1) ( 1 )1172)-1 [72(xiufil + 140 + xi1/32 u21] (10.9.11)
if yri > 0,
y(21) 34ifi2 + u2i if yt g 0.

Thus the likelihood function of the model is
L =f 0 -
0
fot, y(2T dyti 34?), (10.9.12)
which is the same as (10.9.2).
10.10 Type 5 Tobit Model: P(y, <0, n) P(Y, > 0, y2)

The Type 5 Tobit model is obtained from the Type 4 model (10.9.1) by
omitting the equation for yu . We merely observe the sign of yr,. Thus the
model is defined by
= xii/31 Uli (10.10.1)
= 4/12 U2i
Yx;ifi3
Y21 = if yr' > 0
=0 if y g 0
Y3i if yr, 0
=0 if yri > 0, i = 1, 2, . . . , n,
where (Liu , u21 , au ) are i.i.d. drawings from a trivariate normal distribution.
The likelihood function of the model is
L= flo j. .1-3(y*, hi) dyti IT0 o f2(Yti, ) CIA (10.10.2)
where f3 and f2 are as defined in (10.9.2). Because this model is somewhat

simpler than Type 4, the estimation methods discussed in the preceding
section apply to this model afortiori. Hence, we shall go immediately into the
discussion of applications.
10.10.2 Model of Lee

In the model of Lee (1978), y represents the logarithm of the wage rate of the
ith worker in case he or she joins the union and represents the same in case
he or she does not join the union. Whether or not the worker joins the union is
determined by the sign of the variable
= z; + vi. (10.10.3)
Because we observe only A; if the worker joins the union and y if the worker
,
does not, the logarithm of the observed wage, denoted yi , is defined by
Yi= if yr' > 0 (10.10.4)
= Yt if yr, 0.
Lee assumed that x2 and x3 (the independent variables in the As' and 3/1
equations)cldhvuarcteisofmandwrkesuch
regional location, city size, education, experience, race, sex, and health,
whereas z includes certain other individual characteristics and variables that
represent the monetary and nonmonetary costs of becoming a union member.
Because yris unobserved except for the sign, the variance of yrcan be assumed
to be unity without loss of generality.
Lee estimated his model by Heckman's two-step method applied separately
to the yt and y equations. In Lee's model simultaneity exists only in the y'r
equation and hence is ignored in the application of Heckman's two-step
method. Amemiya's LS or GLS, which accounts for the simultaneity, will, of
course, work for this model as well, and the latter will yield more efficient
estimates although, of course, not as fully efficient as the MLE.
Tobit Models 401
10.10.3 Type 5 Model of Heckman

The Type 5 model of Heckman (1978) is a simultaneous equations model
consisting of two equations
Yr( = YiY2i (51 141r (10.10.5)
and
= Y2Yri X2S2 (52W1 UN/ (10.10.6)
where we observe y2i , x11 , x21 , and w, defined by

=1 if yri > 0 (10.10.7)
=0 if yri O.
There are no empirical results in the 1978 article, but the same model was
estimated by Heckman (1976b); in this application yT i represents the average
income of black people in the ith state, yri the unobservable sentiment toward
blacks in the ith state, and w i = 1 if an antidiscrimination law is instituted in
the ith state.
When we solve (10.10.5) and (10.10.6) for yr i , the solution should not
depend upon Ivo for that would clearly lead to logical inconsistencies. There-
fore we must assume
YiO2 (51 = 0 (10.10.8)
for Heckman's model to be logically consistent. Using the constraint

(10.10.8), we can write the reduced-form equations (although strictly speaking
not reduced-form because of the presence of w i ) of the model as
yr, = x,r1 + v1 1 (10.10.9)
and
312i = (52Wl xiir2 + v21, (10.10.10)
where we can assume Vv ii = 1 without loss of generality. Thus Heckman's
model is a special case of Type 5 with just a constant shift between y'21' and yf
(that is, yt = x;n2 + v2i and y, = X;X2 U21). Moreover, if (52 = 0, it is a
special case of Type 5 where yr =-
Let us compare Heckman's reduced-form model defined by (10.10.9) and
(10.10.10) with Lee's model. Equation (10.10.9) is essentially the same
as (10.10.3) of Lee's model. Equation (10.10.4) of Lee's model can be rewrit-
ten as
yi= w1(xZifi2+ u21) + (1 wi)(x03+ u31) 00.10.10

= x03+ U31 Wi(X21fl2 + UV - X03 - U31).
By comparing (10.10.10) and (10.10.11), we readily see that Heckman's re-

duced-form model is a special case of Lee's model in which the coefficient
multiplied by wi is a constant.
Heckman proposed a sequential method of estimation for the structural
parameters, which can be regarded as an extension of Hecicman's simulta-
neous equations two-step estimation discussed in Section 10.8.3. His method
consists of the following steps:
Step 1. Estimate 7r 1 by applying the probit MLE to (10.10.9). Denote the
estimator 1 1 and define Fa = F(x;ii1 ).
Step 2. Insert (10.10.9) into (10.10.6), replace 7r 1 with 11 and wi with Pi , and
then estimate y2, A, and 62 by least squares applied to (10.10.6).
Step 3. Solve (10.10.5) for y21 , eliminate yri by (10.10.9), and then apply
least squares to the resulting equation after replacing 7r 1 by fr, and wi by Fi t
estimate n i, yr' fl and yr '6 1 .
Amemiya (1978c) derived the asymptotic variance-covariance matrix of
Heckman's estimator defined in the preceding paragraph and showed that
Amemiya's GLS (defined in Section 10.8.4) applied to the model yields an
asymptotically more efficient estimator in the special case of (5 1 = = 0. As
pointed out by Lee (1981), however, Amemiya's GLS can also be applied to
the model with nonzero o's as follows:
Step 1. Estimate ni by the probit MLE 1 applied to (10.10.9).
Step 2. Estimate (52 and n2 by applying the instrumental variables method
to (10.10.10), using Fi as the instrument for wi . Denote these estimators as 6 2
12 . and
Step 3. Derive the estimates of the structural parameters y l , fl1, 61 , y2 , /32 ,
and 62 from , 12 , and 62 , using the relationship between the reduced-form
parameters and the structural parameters as well as the constraint (10.10.8) in
the manner described in Section 10.8.4.
The resulting estimator can be shown to be asymptotically more efficient than
Heckman's estimator.
10.10.4 Disequilibrium Models

Disequilibrium models constitute an extensive area of research, about which
numerous papers have been written. Some of the early econometric models
have been surveyed by Maddala and Nelson (1974). A more extensive and
Tobit Models 403
up-to-date survey has been given by Quandt (1982). See, also, the article by
Hartley (1976a) for a connection between a disequilibrium model and the
standard Tobit model. Here we shall mention two basic models first discussed
in the pioneering work of Fair and Jaffee (1972).
The simplest disequilibrium model of Fair and Jaffee is a special case of the
Type 5 model (10.10.1), in which y , is the quantity demanded in the ith
period, yt is the quantity supplied in the ith period, and yr i = yti . Thus
the actual quantity sold, which a researcher observes, is the minimum of
supply and demand. The fact that the variance-covariance matrix of
(yf, y , yT) is only of rank 2 because of the linear relationship above does not
essentially change the nature of the model because the likelihood function
(10.10.2) involves only bivariate densities.
In another model Fair and Jaffee added the price equation to the model of
the preceding paragraphs as
Yai = Y Air), (10.10.12)
where 3,41 denotes a change in the price at the ith period. The likelihood
function of this model can be written as"
0
L =fJ f f3(yri,y3ily41)f(y41)dyri (10.10.13)
0
x f2(yri, y2ily4i).fty4i) dyri

H fo
The form of the likelihood function does not change if we add a normal error
term to the right-hand side of (10.10.12). In either case the model may be
schematically characterized by
P(Y, < Y3, Y4) P(Yi > /3 1 Y2 s Y4), (10.10.14)
which is a simple generalization of the Type 5 model.
10.10.5 Multivariate Generalizations

By a multivariate generalization of Type 5, we mean a model in which y, and
y t in (10.10.1) are vectors, whereas yr ; is a scalar variable the sign of which is
observed as before. Therefore the Fair-Jaffee model with likelihood function
characterized by (10.10.14) is an example of this type of model.
In Lee's model (1977) the yri equation is split into two equations
= XVI2 + U2 (10.10.15)
and
= 4;a2 + v2, (10.10.16)
where Cti and Tti denote the cost and the time incurred by the ith person
traveling by a private mode of transportation, and, similarly, the cost and the
time of traveling by a public mode are specified as
= xVi3 U3 (10.10.17)
and
TT, = z;ice3 + v3 . (10.10.18)
Lee assumed that Ct and 71 are observed if the ith person uses a private mode
and Ct and Tit are observed if he or she uses a public mode. A private mode is
used if yri > 0, where yri is given by
y it; = s;(51 + (52 T, A- 33 71 + 4 ( C, )+ Ei . (10.10.19)
Lee estimated his model by the following sequential procedure:
Step 1. Apply the probit MLE to (10.10.19) after replacing the starred
variables with their respective right-hand sides.
Step 2. Apply LS to each of the four equations (10.10.15) through (10.10.18)
after adding to the right-hand side of each the estimated hazard from step 1.
Step 3. Predict the dependent variables of the four equations (10.10.15)
through (10.10.18), using the estimates obtained in step 2; insert the predictors
into (10.10.19) and apply the probit MLE again.
Step 4. Calculate the MLE by iteration, starting from the estimates obtained
at the end of the step 3.
Willis and Rosen (1979) studied earnings differentials between those who
went to college and those who did not, using a more elaborate model than that
of Kenny et al. (1979), which was discussed in Section 10.9.2. In the model of
Kenny et al., y, (the desired years of college education, the sign of which
determines whether an individual attends college) is specified not to depend
directly on yt and y it (the earnings of the college-goer and the non-college-
goer, respectively). The first inclination of a researcher might be to hypothe-
size yri = A. However, this would be an oversimplification because the
decision to go to college should depend on the difference in expected lifetime
earnings rather than in current earnings.
Willis and Rosen solved this problem by developing a theory of the maxi-
mization of discounted, expected lifetime earnings, which led to the following
model:
Tobit Models 405
= x02 + 142 , (10.10.20)

Gti = z,a2 + 112, (10.10.21)
Il = X03 u3 , (10.10.22)
= 41a3 + v3 , (10.10.23)
and
12; = + ei , i = 1, 2, . . . ,n (10.10.24)
where I , and G , denote the initial earnings (in logarithm) and the growth rate
of earnings for the college-goer, It, and Gt denote the same for the non-col-
lege-goer, and 12, denotes the discount rate. It is assumed that the ith person
goes to college if yri > 0 where
(10.10.25)
and that the variables with subscript 2 are observed if y', > 0, those with
subscript 3 are observed if yri 0, and Ri is never observed. Thus the model is
formally identical to Lee's model (1977). Willis and Rosen used an estimation
method identical to that of Lee, given earlier in this subsection.
Borjas and Rosen (1980) used the same model as Willis and Rosen to study
the earnings differential between those who changed jobs and those who did
not within a certain period of observation.
10.10.6 Multinomial Generalizations

In all the models we have considered so far in Section 10.10, the sign of yr ;
detrminwobascegrfvations,uch mebrs
versus nonunion members, states with an antidiscrimination law versus those
without, or college-goers versus non-college-goers. By a multinomial general-
ization of Type 5, we mean a model in which observations are classified into
more than two categories. We shall devote most of this subsection to a discus-
sion of the article by Duncan (1980).
Duncan presented a model of joint determination of the location of a firm
and its inputoutput vectors. A firm chooses the location for which profits are
maximized, and only the input output vector for the chosen location is
observed. Let si(k) be the profit of the ith firm when it chooses location k,
i = 1, 2, . . . , n and k = 1, 2, . . . , K, and let y,(k) be the inputoutput
vector for the ith firm at the kth location. To simplify the analysis, we shall
subsequently assume yi (k) is a scalar, for a generalization to the vector case is
straightforward. It is assumed that

si (k)= 34) 1 + u,k (10.10.26)
and
y1(k) = xg) ' fl + vac, (10.10.27)
where x(i112 and 42 are vector functions of the inputoutput prices and eco-
nomic theory dictates that the same fl appears in both equations. I9 It is as-
sumed that (un , u,2 , . . , U Uc vil , v a, . . , v,K ) are i.i.d. drawings from a
,
2K-variate normal distribution. Suppose .s,(k,)> sa(j) for any j # ki . Then a

researcher observes yi (k,) but does not observe y1 (j) forj # ki .
For the following discussion, it is useful to define K binary variables for each
i by
wi (k)= 1 if ith firm chooses kth location (10.10.28)
=0 otherwise
and define the vector w, = [w,(1), wi(2), . . , w,(K)]' . Also define
Pik= P[w,(k) = 1] and the vector Pi = (Pa , Pa,. . . , Pac)'.
There are many ways to write the likelihood function of the model, but
perhaps the most illuminating way is to write it as
L I f (ki ) I wi (ki) =1]Pik (10.10.29)
where k, is the actual location the ith firm was observed to choose.
The estimation method proposed by Duncan can be outlined as follows:
Step I. Estimate the 13 that characterize fin (10.10.29) by nonlinear WLS.
Step 2. Estimate the fl that characterize Pin (10.10.29) by the multinomial
probit MLE using the nonlinear WLS iteration.
Step 3. Choose the optimum linear combination of the two estimates of /I
obtained in steps 1 and 2.
To describe step 1 explicitly, we must evaluate pi E[y,(101w,(k,)= 1]
and (4 = V[yi (k,)1w,(k,)= 1] as functions of fl and the variances and covar-
iances of the error terms of Eqs. (10.10.26) and (10.10.27). These conditional
moments can be obtained as follows. Define zi (j) = .s,(k) s i (j) and the
(K 1)-vector z i = [;(1), . . . , zi (ki 1), z i(ki + 1), . . , z,(K)]'. To
simplify the notation, write z, as z, omitting the subscript. Similarly, write
y,(k,) as y. Also, define R = E(y Ey)(z Ez)' [E(z Ez)(z Ez)V and
Q = Vy RE(z Ez)(y Ey). Then we obtain"
E(ylz> 0) = Ey + RE(z1z> 0) REz (10.10.30)
Tobit Models 407
and
= V(ylz > 0) = Riqzlz > 0)R' + Q. (10.10.31)
The conditional moments of z appearing in (10.10.30) and (10.10.31) can be
found in the articles by Amemiya (1974b, p. 1002) and Duncan (1980, p. 850).
Finally, we can describe the nonlinear WLS iteration of step 1 as follows:
Estimate al by inserting the initial estimates (for example, those obtained by
minimizing [yi (ki ) pi]2 ) of the parameters into the right-hand side of

(10.10.31)call it as? . Minimize

2 (10.10.32)
6 T2 [Y1(k1) Ai]
with respect to the parameters that appear in the right-hand side of (10.10.30).
Use these estimates to evaluate the right-hand side of (10.10.31) again to get
another estimate of al . Repeat the process to yield new estimates of ft.
Now consider step 2. Define
= E(wi Pi )(wi P,)' = Di P,P , (10.10.33)
where Di is the K X K diagonal matrix the kth diagonal element of which is
Pik. To perform the nonlinear WLS iteration, first, estimate I, by inserting the
initial estimates of the parameters into the right-hand side of (10.10.33) (de-
note the estimate thus obtained as X i ); second, minimize
(wi - (wi Pi), (10.10.34)
where the minus sign in the superscript denotes a generalized inverse, with
respect to the parameters that characterize P i , and repeat the process until the
estimates converge. A generalized inverse A- of A is any matrix that satisfies
AAA = A (Rao, 1973, p. 24). A generalized inverse IT is obtained from the
matrix DT P II', where I is a vector of ones, by replacing its kth column
and row by a zero vector. It is not unique because we may choose any k.
Finally, regarding step 3, if we denote the two estimates of /I obtained
by steps 1 and 2 by ft1 and fl2 , respectively, and their respective asymp-
totic variance-covariance matrices by V I and V2, the optimal linear com-
bination of the two estimates is given by (VI' + V') - 'VT' +
(VT' + V -2-1 ) iVV/32 . This final estimator is asymptotically not fully effi-
cient, however. To see this, suppose the regression coefficients of (10.10.26)
and (10.10.27) differ: Call them fi2 and , say. Then, by a result of Amemiya
(1976b), we know that fi2 is an asymptotically efficient estimator of fi2 . How-
ever, as we have indicated in Section 10.4.4, /1 1 is not asymptotically efficient.
So a weighted average of the two could not be asymptotically efficient.
Dubin and McFadden (1984) used a model similar to that of Duncan in

their study of the joint determination of the choice of electric appliances and
the consumption of electricity. In their model, s1 (k) may be interpreted as the
utility of the ith family when they use the kth portfolio of appliances, and y1 (k)
as the consumption of electricity for the ith person holding the kth portfolio.
The estimation method is essentially similar to Duncan's method. The main
difference is that Dubin and McFadden assumed that the error terms of
(10.10.26) and (10.10.27) are distributed with Type I extreme-value distribu-
tion and hence that the P part of (10.10.29) is multinomial logit.
Exercises
1. (Section 10.4.3)
Verify (10.4.19).
2. (Section 10.4.3)
Verify (10.4.28).
3. (Section 10.4.3)
Consider V9v, and Vy7w given in (10.4.32) and (10.4.33). As stated in
the text, the difference of the two matrices is neither positive definite
nor negative definite. Show that the first part of Viw , namely,
o-2 (Z' is smaller than VAv in the matrix sense.
4. (Section 10.4.5)
In the standard Tobit model (10.2.3), assume that a 2 = 1,/i is a scalar and
the only unknown parameter, and (x i) are i.i.d. binary random variables
taking 1 with probability p and 0 with probability 1 p. Derive the
formulae of p AV [17i(fl [3)] for /3= Probit MLE, Tobit MLE, Heck-
man's LS, and NLLS. Evaluate them for fi =0, 1, and 2.
5. (Section 10.4.6)
= 1 if yt 0
=0 if 31 <0, i= 1, 2, . . . , n,
where (y*) are independent N(x'i ft, 1). It is assumed that ( yi ) are observed
but (y*) are not. Write a step-by-step instruction of the EM algorithm to
obtain the MLE offi and show that the MLE is an equilibrium solution of
the iteration.
Tobit Models 409
6. (Section 10.6)
31i=x1fl1 +
Yir= 142i
Yli = Yri if yri > 0
0 if yt g 0
Y2, 1 if A; > 0
=0 if yt .S. 0, i = 1, 2, . . . , n,
where (u 11 , u2i ) are i.i.d. with the continuous densityf( , ). Denote the
marginal density of 14 11 by f; ( ) and that of u 2; by f2 ( ).
a. Assuming that y ii , y, xli and x2i are observed for i = 1,
2, . . . , n, express the likelihood function in terms off;f1 and f2 .
,
b. Assuming that yu , y21 , x 11 , and x21 are observed for all i, express the
likelihood function in terms off, fi and f2 .
7. (Section 10.6)
= az, +
zr = fivi +
= 1 if yr ?.- 0
=0 if y* < 0
zi = 1 if z0
=0 if < 0,
where ui and vi are jointly normal with zero means and nonzero covar-
iance. Assume that y*, z*, u, and v are unobservable and y and z are
observable. Show that the model makes sense (that is, y and z are uniquely
determined as functions of u and v) if and only if afl = 0.
8. (Section 10.6)
In the model of Exercise 7, assume that /1 = 0 and that we have n
observations on (yi , zi ), i = 1, 2, . . . , n. Write the likelihood function
of a. You may write the joint density of (u, v) as simply f(u, v) without
explicitly writing the bivariate normal density.
9. (Section 10.6)
Suppose y? and 4, i = 1, 2, . . . , n, are i.i.d. and jointly normally dis-
tributed with nonzero correlation. For each i, we observe (1) only y?, (2)
only 41 , or (3) neither, according to the following scheme:
(1) Observe y = y, and do not observe z7 if y' 4 0.
(2) Observe 4 = z, and do not observe yi if z* > 0.
(3) Do not observe either if yr <0 or 4. < 0.
Write down the likelihood function of the model. You may write the joint
normal density simply as f( , ).
10. Section (10.7.1)

Write the likelihood function of the following two models (cf. Cragg,
1971).
a. (yr, yr) Bivariate N(x;111 , xV12 , 1, ai, a12)
Y2 if yr > 0 and yr > 0
= 0 otherwise.
We observe only y2 .
b. (yr, yt) Bivariate N(x;fi l , x02 , 1, al, au ) with yfruncated so
that yr > 0
Y2 = if y>O
=0 if yf 5 0
We observe only Y2.
11. (Section 10.9.4)

In Tomes' model defined by (10.9.7) through (10.9.9), consider the fol-
lowing estimation method: Step I. Regress y2; on x 11 and x2i and obtain the
least squares predictor 92i. Step 2. Substitute hi for y2i in (10.9.7) and
apply the Tobit MLE to Eqs. (10.9.7) and (10.9.9). Will this method yield
consistent estimates of y l and /31 ?
12. (Section 10.10)

Suppose the joint distribution of a binary variable w and a continuous
variable y is determined by P(w = fly) = A(y, y) and f(ylw)=
N(y2w, a 2 ). Show that we must assume a 2 y1 = y2 for logical consistency.
Tobit Models 411
13. (Section 10.10.1)

In model (10.10.1), Type 5 Tobit, define an observed variable y i by
Yi= A I, if Yri> 0
= yt if
and assume that a researcher does not observe whether yt i > 0 or -5 0;
that is, the sample separation is unknown. Write the likelihood function
of this model.
14. (Section 10.10.4)
Let (A, yti , yt) be a three-dimensional vector of continuous random
variables that are independent across i = 1, 2, . . . , n but may be corre-
lated among themselves for each i. These random variables are unob-
served; instead, we observe zi and yi defined as follows
Zj = if yri > 0
if O.
0 with probability A
Yi= 1 if
with probability 1 A
=0 if yri O.
Write down the likelihood function. Use the following symbols:
f2i(yri , A) joint density of yt and yt
f3i(y*, A) joint density of and yt.
15. (Section 10.10.4)
Consider a regression model: A= uir and y, = 34iA + uv,
where the observable random variable y, is defined by y = yt i with proba-
bility and yi = yri with probability 1 A. This is called a switching
regression model. Write down the likelihood function of the model, as-
suming that (u11 , uv) are i.i.d. with joint density f( , ).
16. (Section 10.10.6)

Show Z1Z = 1i where I, is given in (10.10.33) and IT is given after
(10.10.34). Let w* and P7' be the vectors obtained by eliminating the kth
element from wi and Pi , where k can be arbitrary, and let be the
variance-covariance matrix of w?. Then show (w; 1T(w, Pi)
Ow7 r an-l (wr PP).
11 Markov Chain and Duration Models
We can use the term time series models in a broad sense to mean statistical
models that specify how the distribution of random variables observed over
time depends on their past observations. Thus defined, Markov chain models
and duration models, as well as the models discussed in Chapter 5, are special
cases of time series models. However, time series models in a narrow sense
refer to the models of Chapter 5, in which random variables take on continu-
ous values and are observed at discrete times. Thus we may characterize the
models of Chapter 5 as continuous-state, discrete-time models. Continuous-
state, continuous-time models also constitute an important class of models,
although we have not discussed them. In contrast, Markov chain models (or,
more simply, Markov models) may be characterized as discrete-state, dis-
crete-time models, and duration models (or, survival models) as discrete-state,
continuous-time models. In this chapter we shall take up these two models in
turn.
The reader who wishes to pursue either of these topics in greater detail than
is presented here should consult the textbooks by Bartholomew (1982) for
Markov models and by Kalbfleisch and Prentice (1980) or Miller (1981) for
duration models. For recent results on duration models with econometric
applications, see Heckman and Singer (1984b).
11.1 Markov Chain Models

11.1.1 Basic Theory
Define a sequence of binary random variables
y(t) =1 if ith person is in state j at time t (1 1 .1.1)
= 0 otherwise,
i = 1, 2, . . . , N, t = 1, 2, , . . , T,
j = 1, 2, . . . , M.
Markov Chain and Duration Models 413
Markov chain models specify the probability distribution of yi(t) as a func-

tion of yi(s), k= 1, 2, . . . , M and s = t 1, t 2, . . . , as well as (possi-
bly) of exogenous variables.
Markov chain models can be regarded as generalizations of qualitative
response models. As noted in Section 9.7, Markov chain models reduce to QR
models if yj(t) are independent over t. In fact, we have already discussed one
type of Markov models in Section 9.7.2.
Models where the distribution of y1(t) depends on yik (t 1) but not on
yik (t 2), AO 3), . . . are called first-order Markov models. We shall pri-
marily discuss such models, although higher-order Markov models will also
be discussed briefly.
First-order Markov models are completely characterized if we specify the
transition probabilities defined by
4(0 = Prob [ith person is in state k at time t given that (11.1.2)
he was in state j at time t 1]
and the distribution of yi(0), the initial conditions.
The following symbols will be needed for our discussion:
At) = M-vector the jth element of which is y(t) (11.1.3)
MO= E At)
I-1 j
n jk(t) = yj(t 1),4(t)

i-1
nc,= I )7* 1)ylk (t)

t-1
N
nik= nik(t) -= njk
i-t 1=1
Pi(t)= (IWO), an M X M matrix
14(0 = Prob[ith person is in state ) at time t]
p/(t) = M-vector the jth element of which is et).
The matrix P(t) is called a Markov matrix. It has the following properties:
(1) Every element of P(t) is nonnegative. (2) The sum of each row is unity. (In
other words, if 1 is an M-vector of ones, then Pi(t)1= 1.)
If yi(0) is a binary random variable taking the value of 1 with probability
p'(0), the likelihood function of the first-order Markov model can be written
as
L=n fl n pj,(0),J0-04(0 n fi pjoyin, (11.1.4)
ilk j If
where t ranges from 1 to T, i ranges from 1 to N, and k and j range from 1 to M

unless otherwise noted. If the initial values yi(0) are assumed to be known
constants, the last product term of the right-hand side of (11.1.4) should be
dropped.
Clearly, all the parameters Pik (t) and pi(0) cannot be estimated consist-
ently. Therefore they are generally specified as functions of a parameter vector
0, where the number of elements in 0 is either fixed or goes to co at a sufficiently
slow rate (compared to the sample size). Later we shall discuss various ways to
parameterize the transition probabilities Pik.(t).
A Markov model where PM = Pik for all t is called the stationary Markov
model. If Pik (t) = PA (t) for all i, the model is called homogeneous. (Its anto-
nym is heterogeneous.) A parameterization similar to the one used in QR
models is to specify 4(0 = Pik [30(013] for some functions Fik such that
V_ IPA = 1. Examples will be given in Sections 11.1.3 and 11.1.4. The case
where y(t) are independent over t (the case of pure QR models) is a special
case of the first-order Markov model obtained by setting Pik (t) = Pi,k (t) for all
j and j' for each i and t.
For QR models we defined the nonlinear regression models: (9.2.26) for the
binary case and (9.3.16) for the multinomial case. A similar representation for
the first-order Markov model can be written as
Llyi(t)lykt 1), ykt 2), . . = PktYykt 1) (11.1.5)
Or
ykt) = Kt)'ykt 1) + ui(t). (11.1.6)

Because these M equations are linearly dependent (their sum is 1), we elimi-
nate the Mth equation and write the remaining M 1 equations as
y Yt) =1-17"(0'ykt 1) + ikt). (11,1.7)
Conditional on ykt 1), ykt 2), . . . , we have Eil(t) = 0 and V rikt) =

D(g) pit', where it = P ktrykt 1) and D(p) is the diagonal matrix with the
elements of g in the diagonal. Strictly speaking, the analog of (9.3.16) is
(11.1.7) rather than (11.1.6) because in (9.3.16) a redundant equation has
been eliminated.
As in QR models, the NLGLS estimator of the parameters that characterize

Pi(t), derived from (11.1.7), yields asymptotically efficient estimates. The
presence of yi(t 1) in the right-hand side of the equation does not cause any
problem asymptotically. This is analogous to the fact that the properties of the
least squares estimator in the classical regression model (Model 1) hold
asymptotically for the autoregressive models discussed in Chapter 5. We shall
discuss the NLWLS estimator in greater detail in Section 11.1.3. There we
shall consider a two-state Markov model in which Pi(t) depends on exogenous
variables in a specific way.
As in QR models, minimum chi-square estimation is possible for Markov
models in certain cases. We shall discuss these cases in Section 11.1.3.
Taking the expectation of both sides of (11.1.5) yields
et) = Pi(t)' pia 1). (11.1.8)
It is instructive to rewrite the likelihood function (11.1.4) as a function of
Pik (t) and pi(t) as follows. Because
k(1) yjo_ oyvo
Plk(0 40-1)*0 Pik(0 4(t), (11.1.9)
ji Pi(0
we can write (11.1.4) alternatively as T
40_1)4,
II II Pik(0
[ --(t)
y)
H pi (o yiko (11.1.10)
t ikj 0 I j
- "=" L1 L2.
If we specify 4(0 and py0) to be functions of a parameter vector 0, then by

(11.1.8) L2 is also a function of 0. The partial likelihood function L2(0) has the
same form as the likelihood function of a QR model, and maximizing it will
yield consistent but generally asymptotically inefficient estimates of 0.
As we noted earlier, if yyt) are independent over t, the rows of the matrix
Pi(t) are identical. Then, using (11.1.8), we readily see PO) = pi(t). There-
fore L = L2, implying that the likelihood function of a Markov model is
reduced to the likelihood function of a QR model.
Because pyt) is generally a complicated function of the transition probabili-
ties, maximizing L2 cannot be recommended as a practical estimation
method. However, there is an exception (aside from the independence case
mentioned earlier), that is, the case when /4(0) are equilibrium probabilities.
The notion of equilibrium probability is an important concept for station-
ary Markov models. Consider a typical individual and therefore drop the
superscript i from Eq. (11.1.8). Under the stationarity assumption we have

p(t) P'p(t 1). (11.1.11)
By repeated substitution we obtain from (11.1.11)

P(t) = (P7P(0). (11.1.12)
if limt., (PT exists, then

P(w) = (PTP(0). (11.1.13)
We call the elements of p(00) equilibrium probabilities. They exist if every

element of P is positive.
It is easy to prove (Bellman, 1970, p. 269) that if every element of P is
positive, the largest (in absolute value or modulus) characteristic root of P is
unity and is unique. Therefore, by Theorem 1 of Appendix 1, there exist a
matrix H and a Jordan canonical form D such that P' = HMV'. Therefore
we obtain
(P') = HJH I , (11.1.14)
where J is the M X M matrix consisting of 1 in the northwestern corner and 0

elsewhere. Equilibrium probabilities, if they exist, must satisfy
P(w) = P iP(m), (11.1.15)
which implies that the first column of H is p(c0) and hence the first row of 11 -1
istheranpofM-vctruniydeo1.Thrf,m(14)
(PT = P(x)1'. (11.1.16)
Inserting (11.1.16) into (11.1.13) yields the identity p(c0) = p(c0)1'p(0) = p(op)
for any value of p(0). If p(00) exists, it can be determined by solving (11.1.15)
subject to the constraint l'p(00) = 1. Because the rank of! P' is M 1 under

the assumption, the p(co) thus determined is unique.

If pi(0) = Ow), L2 reduces to
= 11 H p.gco) Yi(0, (11.1.17)
i J
which is the likelihood function of a standard multinomial QR model. Even if

pj(0) # pl(co), maximizing Ltyields a consistent estimate as T goes to infinity.
We shall show this in a simple case in Section 11.1.3.
Now consider the simplest case of homogeneous and stationary Markov
models characterized by
Plk(0 = Pik for all i and t. (11.1.18)

The likelihood function (11.1.4) conditional on yj(0) is reduced to
L = fi fl (11.1.19)
It is to be maximized subject to M constraints Er-iPik = 1,j = 1, 2, . . . , M.

Consider the Lagrangian
S= nfic log Pik Pik 1) . (11.1.20)
Setting the derivative of S with respect to Pik equal to 0 yields
njk = AjPjk (11.1.21)

Summing both sides of (11.1.21) over k and using the constraints, we obtain
the MLE
Pjk = njk I njk (11.1.22)
See Anderson and Goodman (1957) for the asymptotic properties of the MLE
(11.1.22).
Anderson and Goodman also discussed the test of various hypotheses in the
homogeneous stationary Markov model. Suppose we want to test the null
hypothesis that Pik is equal to a certain (nonzero) specified value for
k= 1, 2, . . . , M and for a particular j. Then, using a derivation similar to
(9.3.24), we can show
IM
Si nik it (15 ;lc Plia2 A 2 (11.1.23)
M- 1
.07
.1 ) k I
where fijk is the MLE. Furthermore, if ilk is given for] = 1, 2, . . . , Mas well
as k, we can use the test statistic M .71 1 Si , which is asymptotically distributed as
chi-square with M(M 1) degrees of freedom. Next, suppose we want to test
(11.1.18) itself against a homogeneous but nonstationary model characterized
by PM = PA (t). This can be tested by the likelihood ratio test statistic with
the following distribution:
2 log H H H rfi, /15, _,k ,, nik (I) XtT- IMAM- 1) I (11.1.24)
r j k
where fijk (t) = nik (t)Inf_ i nik (t).

In the same article Anderson and Goodman also discussed a test of the
first-order assumption against a second-order Markov chain.
11.1.2 Empirical Examples of Markov Models without

Exogenous Variables
In this subsection we shall discuss several empirical articles, in which Markov
chain models without exogenous variables are estimated, for the purpose of
illustrating some of the theoretical points discussed in the preceding subsec-
tion. We shall also consider certain new theoretical problems that are likely to
be encountered in practice.
Suppose we assume a homogeneous stationary first-order Markov model
and estimate the Markov matrix (the matrix of transition probabilities) by the
MLE Pik given by (11.1.22). Let P.(02,) be a transition probability of lag two; that
is, /2 ) denotes the probability a person is in state k at time t given that he or she
was in state j at time t 2. If we define n(t) = gsr_ i yl(t ... 2)yi(t) and n5k2) =
IT 2 Pl5k2)(0, a consistent estimate of 11 ) is also given by P5122 = ni(122/Mict- in,;k2)
Now, if our assumption is correct, we should have approximately
fx2) a 152, (11.1.25)
where 2 =-- f.P Many studies of the mobility of people among classes of
income, social status, or occupation have shown the invalidity of the approxi-
mate equality (11.1.25). There is a tendency for the diagonal elements of P (2)
2 (see, for example, Bartholomew, 1982, Chapter 2). tobelarghnsfP
This phenomenon may be attributable to a number of reasons. Two empirical
articles have addressed this issue: McCall (1971) explained it by population
heterogeneity, whereas Shorrocks (1976) attributed the phenomenon in his
data to a violation of the first-order assumption itself.
McCall (1971) analyzed a Markov chain model of income mobility, where
the dependent variable is classified into three states: low income, high income,
and unknown. McCall estimated a model for each age sex race combina-
tion so that he did not need to include the exogenous variables that represent
these characteristics. Using the moverstayer model (initially proposed by
Blumen, Kogan, and McCarthy, 1955, and theoretically developed by Good-
man, 1961), McCall postulated that a proportion Si of people, j = 1, 2, and 3,
stay in state j throughout the sample period and the remaining population
follows a nonstationary first-order Markov model. Let Vik (t) be the probabil-
ity a mover is in state k at time t given that he or she was in state j at time t 1.
Then the transition probabilities of a given individual, unidentified to be
either a stayer or a mover, are given by
Pp) = + (1 Si)Vii(t) and (11.1.26)
Pik(t) S1 )Vik(t) if j # k.
McCall assumed that Vik (t) depends on t (nonstationarity) because of eco-

nomic growth, but he got around the nonstationarity problem simply by
estimating Vik(t) for each t separately. He used the simplest among several
methods studied by Goodman (1961). Stayers in state fare identified as people
who remained in state j throughout the sample periods. Once each individual
is identified to be either a stayer or a mover, S i and Vik(t) can be estimated in a
straightforward manner. This method is good only when there are many
periods. If the number of periods is small (in McCall's data T = 10), this
method will produce bias (even if n is large) because a proportion of those who
stayed in a single state throughout the sample periods may actually be movers.
After obtaining estimates of Vik(t), McCall regressed them on a variable repre-
senting economic growth to see how economic growth influences income
mobility.
For a stationary Markov model, Goodman discussed several estimates of S i
Vjk that are consistent (as n goes to infinity) even if T is small (provided and
T> 1). We shall mention only one simple consistent estimator. By defining
matrices V = Vik ), P = (Pik ), and a diagonal matrix S = D(S), we can write
(11.1.26) as
P = S + (I S)V. (11.1.27)
The matrix of transition probabilities of lag two is given by
P(2) = S + (I S)V 2. (11.1.28)
Now, P and 1*2) can be consistently estimated by the MLE mentioned earlier.
Inserting the MLE into the left-hand side of (11.1.27) and (11.1.28) gives us
2M(M 1) equations But since there are only M 2 parameters to estimate in S
and V, solving M2 equations out of the 2M(M 1) equations for S and V will
yield consistent estimates.
The empirical phenomenon mentioned earlier can be explained by the
mover stayer model as follows: From (11.1.27) and (11.1.28) we obtain after
some manipulation
1*2) - P2 = S - SW 2 [S + - SW][S + - SW]
= (I - S)(I V)S(I - V). (11.1.29)
Therefore the diagonal elements of 1* 2) P2 are positive if the diagonal

elements of (I V)S(I V) are positive. But the jth diagonal element of
(I V)S(I V) is equal to
sjo - vid2+ SkVikriki,
!' j
which is positive.
Shorrocks (1976) accounted for the invalidity of (11.1.25) in a study of
income mobility by postulating a second-order Markov model. Depending on
the initial conditions and the values of the parameters, a second-order Markov
model can lead to a situation where the diagonal elements of P 42) are larger
than the corresponding elements of P 2.
The likelihood function of a second-order Markov model conditional on
the initial values yii( 1) and y(0) is given by
L= nrilum
t ilk 1
P jki 4 0-2b0 -1 )Yi(i), (11.1.30)
where Piiki (t) is the probability the ith person is in state 1 at time t given that he
or she was in state j at time t 2 and in state k at time t 1. If homogeneity

and stationarity are assumed, then P(t) = P. Even then the model con-
tains M 2(M 1) parameters to estimate. Shorrocks grouped income into five
classes (M = 5), thus implying 100 parameters. By assuming the Champer-
nowne process (Champernowne, 1953), where income mobility at each time
change is restricted to the three possibilities staying in the same class or
moving up or down to an adjacent income class, he reduced the number of
parameters to six, which he estimated by ML. We can see this as follows: Let
= 1, 0, 1 represent the three possible movements. Then the Champer-

nowne process is characterized by P,.,, the probability a person moves by

from t 1 to t given that he or she moved by ri from t 2 to t 1. Notice that
the model is reduced to a first-order Markov chain with M = 3.
For the purpose of illustrating the usefulness of equilibrium probabilities in
an empirical problem and of introducing the problem of entry and exit, let us
consider a model presented by Adelman (1958). Adelman analyzed the size
distribution of firms in the steel industry in the United States using the data for
the periods 1929- 1939 and 1945- 1956 (excluding the war years). Firms are
grouped into six size classes according to the dollar values of the firms' total
assets. In addition, state 0 is affixed to represent the state of being out of the
industry. Adelman assumed a homogeneous stationary first-order model.
Movements into and from state 0 are called exit and entry, and they create a
special problem because the number of firms in state Oat any particular time is
not observable. In our notation it means that 4_0 0/ is not observable, and
hence, MLE Pok = nok 'Minot cannot be evaluated. Adelman circumvented the
problem by setting0 /i-o nor = 100,000 arbitrarily. We shall see that although
changing /t_onoi changes the estimates of the transition probabilities, it does
not affect the equilibrium relative size distribution of firms (except relative to
size 0).
Let p p(co) be a vector of equilibrium probabilities. Then, as was shown
earlier, p can be obtained by solving
rE1 _ roi (11.1.31)
L jP - Lij ,
where * means eliminating the first row of the matrix. Now, consider pi lpk for
j, k 0, where pi and Pk are solved from (11.1.31) by Cramer's rule. Because
I?_onoi affects only the first column of [I P']* proportionally, it does not
affect p,Ipk .
Duncan and Lin (1972) criticized Adelman's model, saying that it is unreal-
istic to suppose a homogeneous pool of firms in state 0 because a firm that
once goes out of business is not likely to come back. Duncan and Lin solved
the problem by treating entry and exit separately. Exit is assumed to be an
absorbing state. Suppose j = 1 is an absorbing state, then P 11 = 1 and P1k = 0
for k = 2, 3, . . . , M. Entry into state km k (t) firms at time t, k=
2, 3, . . . , Mis assumed to follow a Poisson distribution:
Fk(t) = k(O nIkM Annik(t)!] -1I . (11.1.32)
Then Pik and !Lk are estimated by maximizing the likelihood function
T m
L*LEIF k (t), (11.1.33)
I 2
where Lis as given in (11.1.4). This model is applied to data on five classes of
banks according to the ratio of farm loans to total loans. Maximum likelihood
estimates are obtained, and a test of stationarity is performed following the
Anderson-Goodman methodology.
11.1.3 Two State Models with Exogenous Variables

-
We shall consider a two-state Markov model with exogenous variables, which

accounts for the heterogeneity and nonstationarity of the data. This model is
closely related to the models considered in Section 9.7.2. We shall also discuss
an example of the model, attributed to Bosldn and Nold (1975), to illustrate
several important points.
This model is similar to a univariate binary QR model. To make the subse-

quent discussion comparable to the discussion of Chapter 9, assume that j = 0
or 1 rather than 1 or 2. Let yi, = 1 if the ith person is in state 1 at time t and
= 0 otherwise. (Note that h is the same as the y(t) we wrote earlier.) The
model then can be written as
P(Yit= 1 IY0-1 ) = RP'xit+ ),
(11.1.34)
where F is a certain distribution function. Note that (11.1.34) is equivalent to
P,S 1 (t)= F(fi'xit ) (11.1.35)
P 1 (t) = F [(a + fl)'xij.
Writing the model as (11.1.34) rather than as (11.1.35) makes clearer the
similarity of this model to a QR model. The model defined by (9.7.11) is a
special case of (11.1.34) if (v i,) in (9.7.11) are i.i.d.
Using (11.1.4), we can write the likelihood function of the model condi-
tional on yio as
L Hi I
Pi, (0AApl0cty,-10-yo (11.1.36)
X Ili (t)(lY1-1visp1 )0 (00 - Y4f- X1 -yo.

However, the similarity to a QR model becomes clearer if we write it alterna-
tively as
L= n (11.1.37)
I t
where Fi,= Rfi'xit+
Because we can treat (i, 0 as a single index, model (11.1.34) differs from a
binary QR model only by the presence of y1,,_ 1 in the argument of F. But, as
mentioned earlier, its presence causes no more difficulty than it does in the
continuous-variable autoregressive model; we can treat Yi.-i as if it were an
exogenous variable so far as the asymptotic results are concerned. Thus, from
(9.2.17) we can conclude that the MLE of y = ( fl', air follows
--1
NITv7 (9 7) -4 N {0, [plim
Fit (i - Fit)witwit] } I
NT i
(11.1.38)
where f , is the derivative of Fi, and
WU.
-1 xi! ].
xitY0-1
The equivalence of the NLWLS iteration to the method of scoring also holds
for this model.
Similarly, the minimum chi-square estimator can also be defined for this
model as in Section 9.2.5. It is applicable when there are many observations of
with the same value of xi,. Although a more general grouping can be
handled by the subsequent analysis, we assume x i, = ; for every i, so that yo
i = 1, 2, . . . , N, are associated with the same vector ; .
Define
Yit( 1 Y,-i)
(11.1.39)
I (1 -
and
_
I YitY4i.-1
(11.1.40)
I
Then we have
F-1 (P?) = ft'; + t = 1, 2, . . . , T (11.1.41)
and
F-1(PD (ot + fl)'; + th, t= 1, 2, . . . , T. (11.1.42)
The error terms and nt approximately have zero means, and their respective
conditional variances given y4,_, , i = 1, 2, . . . , N, are approximately
11(1 11) (11.1.43)
f2 [F-1 (in 3,41-1)
and
Pi (1 Pi) (11.1.44)
V(11) N
f 2[F- VD] y41-1
I-1
where P(,) = Pt, (t) and P) = P (t).The MIN estimator of y is the weighted
)( 2
least squares estimator applied simultaneously to the heteroscedastic regres-

sion equations (11.1.41) and (11.1.42). As in QR models, this estimator has

the same asymptotic distribution as the MLE given in (11.1.38), provided that
N goes to cc for a fixed T.
To illustrate further the specific features of the two-state Markov model, we
shall examine a model presented by Boskin and Nold (1975). In this model,
j 1 represents the state of an individual being on welfare and j = 0 the state
of the individual being off welfare. Hoskin and Nold postulated
Pio (t) = Ai (11.1.45)
/la (t) = A(fl'xi)
where A(x) = (1 + e-x), a logistic distribution. The model can be equiva-
lently defined by
P(Yit= 1 1Y i,t-i)= AUrxi (a + (11.1.46)
Thus we see that the Boskin-Nold model is a special case of (11.1.34).
Note that the exogenous variables in the Boskin-Nold model do not depend
on t, a condition indicating that their model is heterogeneous but stationary.
The exogenous variables are dummy variables characterizing the economic
and demographic characteristics of the individuals. The model is applied to
data on 440 households (all of which were on welfare initially) during a
60-month period (thus t denotes a month).
Because of stationarity the likelihood function (11.1.36) can be simplified
for this model as
L =fl 40( 1 - Ai )'4 /3:41( 1 - Bi )'eco. (11.1.47)
It is interesting to evaluate the equilibrium probability the probability

that a given individual is on welfare after a sufficiently long time has elapsed
since the initial time. By considering a particular individual and therefore
omitting the subscript i, the Markov matrix of the Boskin-Nold model can be
written as
[1 B
P' = (11.1.48)
A 1 13] .
Solving (11.1.15) together with the constraint l'p(co) = 1 yields
A
(11.1.49)
A + B' A + B]
P(ccr [B
The first component of the vector signifies the equilibrium probability that the
person is on welfare. It increases proportionally with the transition probability

Pol ('B).
It is instructive to derive the characteristic roots and vectors of P' given in
(11.1.48). Solving the determinantal equation
1 1 AA B
det
A 1B.1 =0
yields the characteristic roots A i = 1 and A2 = 1 - AB. Solving
[ A B h roi
A LO
yields (the solution is not unique) the first characteristic vector h 1 = (B, AY .
Solving
[AB AB] h2 = [0]
0
yields the second characteristic vector h 2 = ( 1, 1)'. Therefore

1 [ 1
H = [ 4 1] and H- 1
A+B A Bl
1
Using these results, we obtain
it 0 1
(PT = lim H (11.1.50)
t. [0 (1 A BY] 11-1
[1 0]
=H H-1
00
B B
A+B A+B
A Al'
A+B A+B
which, of course, could have been obtained directly from (11.1.16) and
(11.1.49).
Although the asymptotic variance-covariance matrix of the MLE and in a
the Boslcin-Nold model can be derived from (11.1.38), we can derive it directly
using the likelihood function (11.1.47). We shall do so only for et because the
derivation for /3 is similar. We shall assume T*co, for this assumption en-
ables us to obtain a simple formula.
We need to consider only the part of (11.1.47) that involves A i ; its loga-
rithm is
log L = nlo log A ; + ni l log (1 A i ). (11.1.51)
Differentiating (11.1.51) with respect to a, we obtain
a log L _ v nio aA, _ ni l aA ;

(11.1.52)
aa A, 8a 1 A i aa
and
82 log L n 10 n u 1 A, 8,41
v (11.1.53)
&tea' t[ Al (1 A. ea ea'
v nio _ 1 02A ;
[ A i 1 A i ] &ea'
To take the expectation of (11.1.53), we need to evaluate En io and Enc l . We

have
Enlo = E yi(t 1)316(t) (11.1.54)

1-1
T-1
= 1111(1)]
T-1
= 0) [ E (P7] p(0)P1 0 .
If we define D to be the 2 X 2 diagonal matrix consisting of the characteristic

roots of P', we have from (11.1.50)
T-1 T-1
E (r)= H [ Di] 11- ' (11.1.55)
r-o rzb
T [B B
A+BA
where the approximation is valid for large T. Inserting (11.1.55) into (11.1.54)
yields
(11.1.56)
Similarly,
Ai)Bi
En (11.1.57)
1 Ai + Bi
From (11.1.53), (11.1.56), and (11.1.57) we obtain
02logL T A.(1 A i )B; ,
E E xi xi (11.1.58)
aaaa' i A.+ B,
Hence, 1
7, Ai)Bi
AV(&) = [ "] . (11.1.59)
i Ai +
Similarly,
B (1 Bi )A i
,
AV(n) = [T xixi . (11.1.60)

i A i + .ui
Next, let us consider the consistency of the estimator of a and fi obtained by
maximizing LI' defined in (11.1.17). Using (11.1.50), we can write the loga-
rithm of (11.1.17) in the Boskin-Nold model as
log LI= [ r o
-
yl(t)] log [Bil(A i + Bi )] (11.1.61)
+ [ A(t)] log [A,/(A, + BM.

i 0-0
Defining y(t) = [yi (t), A(t)r , we have

T
. 1 T
ppm . y, r(t) = mr
7, (P'Yy l(0) (11.1.62)
1 [AD
A m + AD Ai '
where the subscript 0 denotes the true value. Therefore we obtain
1
(11.1.63)
NT IN LI
A,0
11111 '7;
1 l
N os IV Al LAD+BiDB,D10g (
ir,1 B.L
A i + Bi) A m + B3
log (
- \
Ai 11
A i + Bi i j'
from which we can conclude the consistency of the estimator.
We can now introduce in the context of the Boskin-Nold model an impor-

tant concept in the Markov model called duration. Let t1 be the number of
months a particular individual stays on welfare starting from the beginning of
the sample period. It is called the duration of the first spell of welfare. Its
probability is defined as the probability that the individual stays on welfare up
to time t i and then moves off welfare at time t 1 . Therefore
PO I ) = (1 A)t' -1A. (11.1.64)
The mean duration can be evaluated as
Di = E To -0-1A = A
-a (11.1.65)
T-1 ar
=A
a a r
A 1r
A 1
(1 r)2 A'
where r = 1 A. In words, this equation says that the mean duration on
welfare is the inverse of the probability of moving off welfare.
Suppose the ith person experiences H welfare spells of duration
ti , t 12 , . . . , PH and K off-welfare spells of duration si , ..., If we
generalize the Boskin-Nold model and let the ith person's transition probabil-
ities A. and B. vary with spells (but stay constant during each spell), the
likelihood function can be written as
L= 11 [ 1 -A;(h)r -- , A,00 ti [ 1 - NkAsi - lBi(k)} . (11.1.66)
Equation (11.1.66) collapses to (11.1.47) ifA i and B not depend on h and k

.
and therefore is of intermediate generality between (11.1.36) and (11.1.47).

Expressing the likelihood function in terms of duration is especially useful in
the continuous-time Markov model, which we shall study in Section 11.2.
11.1.4 Multistate Models with Exogenous Variables
Theoretically, not much need be said about this model beyond what we have
discussed in Section 11.1.1 for the general case and in Section 11.1.3 for the
two-state case. The likelihood function can be derived from (11.1.4) by speci-
fying P.iik(t) as a function of exogenous variables and parameters. The equiva-
lence of the NLWLS to the method of scoring iteration was discussed for the
general case in Section 11.1.1, and the minimum chi-square estimator defined
for the two-state case in Section 11.1.3 can be straightforwardly generalized to
the multistate case. Therefore it should be sufficient to discuss an empirical
article by Toikka (1976) as an illustration of the NLWLS (which in his case is a
linear WLS estimator because of his linear probability specification).
Toikka's model is a three-state Markov model of labor market decisions in
which the three states (corresponding to j = 1, 2, and 3) are the state of being
employed, the state of being in the labor force (actively looking for a job) and
unemployed, and the state of being out of the labor force.
The exogenous variables used by Toikka consist of average (over individ-
uals) income, average wage rate, and seasonal dummies, all of which depend
on time (months) but not on individuals. Thus Toikka's model is a homoge-
neous and nonstationary Markov model. Moreover, Toikka assumed that
transition probabilities depend linearly on the exogenous variables. 2 Thus, in
his model, Eq. (11.1.7) can be written as
= [ yi(t 1)' 0 x111 + (11.1.67)
which is a multivariate heteroscedastic linear regression equation. As we

indicated in Section 11.1.1, the generalized least squares estimator of II is
asymptotically efficient.
Let Y, be the N X M matrix the ith row of which is yi(t)' and let Y, be the
N X (M 1) matrix consisting of the first M 1 columns of Y, . Define U,
similarly. Then we can write (11.1.67) as
xi 170
Y, 3( U2
11+ (11.1.68)
i7T YT-I UT
The LS estimator fl is therefore given by
T
fl= [i 0 x,x;)] 0 x,). (11.1.69)
f I
To define the FGLS estimator of H write (11.1.68) as Y = X11+ U and

,
write the columns of Y, II and U explicitly as V [Yi Y2) YG1)

,
11= [iri 7r21 nG], and U = [u i , u2, . . , uG ], where G M 1.

Also define y = . . . , y G# )', ir = W2 , . . , )', and u
. . . , 0'. Then (11.1.68) can be written as
y -- Xar+u. (11.1.70)
The FGLS estimator of it is (Vii iX) iX f ol
y, where Cl is a consistent
estimator of Cl = Euu'. Here, Cl has the following form:
Di1 D12 DIG
D21 D22 D2G
0= (11.1.71)
DG1 DGG
I=
where each Dik is a diagonal matrix of size NT. If each Dik were a constant
times the identity matrix, (11.1.70) would be Zellner's seemingly unrelated
regression model (see Section 6.4), and therefore the LS estimator would be
asymptotically efficient. In fact, however, the diagonal elements of Dik are not
constant.
Toikka's estimator of n, denoted fl, is defined by
= [I (I x,x;)] 1 [(V IV I )iV,_ 1'17, 0 xj. (11.1.72)
Because (V 1Y,_ 1 )- IV IV, is the first M 1 columns of the unconstrained

MLE of the Markov matrix P(t), Toikka's estimator can be interpreted as the
LS estimator in the regression of P(t) on x 1 . Although this idea may seem
intuitively appealing, Toikka's estimator is asymptotically neither more nor
less efficient than H. Alternatively, Toikka's estimator can be interpreted as
premultiplying (11.1.68) by the block-diagonal matrix
- CYOY0)-IY;
(IWO-1 r
and then applying least squares. If, instead, generalized least squares were
applied in the last stage, the resulting estimator of H would be identical with
the GLS estimator of n derived from (11.1.68).
11.1.5 Estimation Using Aggregate Data

Up to now we have assumed that a complete history of each individual, yi(t)
for every i, t, and j, is observed. In this subsection we shall assume that only the
aggregate data ni (t) . 1 yj(t) are available. We shall first discuss LS and
GLS estimators and then we shall discuss MLE briefly.
Suppose the Markov matrix Pi(t) is constant across i, so that Pi(t) = P(t).
Summing both sides of (11.1.7) over i yields
'(t) =ii(t)' yi(t 1) + i i(t). (11.1.73)
The conditional covariance matrix of the error term is given by
v[ E - = E [D(4u1 ) -
u], (11.1.74)
where is; = P(t)'ykt 1). Depending on whether P(t) depends on unknown

parameters linearly or nonlinearly, (11.1.73) defines a multivariate heterosce-
dastic linear or nonlinear regression model. The parameters can be estimated
by either LS or GLS (NLLS or NLGLS in the nonlinear case), and it should be
straightforward to prove their consistency and asymptotic normality as NT
goes to co.
The simplest case occurs when Pi(t) is constant across both i and t, so that
Pi(t) = P. If, moreover, P is unconstrained with M(M 1) free parameters,
(11.1.73) becomes a multivariate linear regression model. We can apply either
a LS or a GLS method, the latter being asymptotically more efficient, as was
explained in Section 11.1.4. See the article by Telser (1963) for an application
of the LS estimator to an analysis of the market shares of three major cigarette
brands in the period 1925 1943.
If P(t) varies with t, the ensuing model is generally nonlinear in parameters,
except in Toikka's model. As we can see from (11.1.68), the estimation on the
basis of aggregate data is possible by LS or GLS in Toildca's model, using the
equation obtained by summing the rows of each Y. A discussion of models
where the elements of P(t) are nonlinear functions of exogenous variables and
parameters can be found in an article by MacRae (1977). MacRae also dis-
cussed maximum likelihood estimation and the estimation based on incom-
plete sample.
In the remainder of this subsection, we shall again consider a two-state
Markov model (see Section 11.1.3). We shall present some of the results we
have mentioned so far in more precise terms and shall derive the likelihood
function.
Using the same notation as that given in Section 11.1.3, we define r, =
jYit . The conditional mean and variance of r, given r,_ I are given by
Er = Fir = Pfr1 .. 1 + PRAT r 1 _ 1 ) (11.1.75)

and
Vrt = Fir (I Fir) --- Pg1 PR' rt-1).

(11.1.76)
The NLWLS estimation of y is defined as that which minimizes
N 2
T (T_ F 1 )
i-1
(11.1.77)
Pr
where Pr, is obtained by estimating PI and ilby F(f3'x, + aixt ) and F(Ifx t ),
respectively, where et and fl are consistent estimates obtained, for example, by
minimizing Ii... 1(r, 114. 1F02. Alternatively we can minimize
N )2
T N
Vrt
+E log 1 Ft,(1 F11 ), (11.1.78)
i-i 1-1 r=1
which will asymptotically give the same estimator. 3 Let i be the estimator
obtained by minimizing either (11.1.77) or (11.1.78). Then we have
Ar aFit N
v v aF,t -1
1 T A av A av/
rATTYY s y) . N ( 0, plim A 7 i-i ,
xi Fit (1 Fft )
-
(11.1.79)
The asymptotic variance-covariance matrix of can be shown to be larger
(in matrix sense) than that of given in (11.1.38) as follows: The inverse of the
latter can be also written as
1 TN 1 aFit
plim (11.1.80)
NT Fu ) ay ay'
Put zt, = [F,,(1 Fii)]- "28Fii/aY and au = [Fii( 1 Fri)1 112 . Then the desired
inequality follows from
aft zit E az;,

(11.1.81)
1-1
I
Finally, we can define the MLE, which maximizes the joint probability of r,
and r,_ 1 , t = 1, 2, . . . , T. The joint probability can be shown to be
T min [N [( r
(11.1.82)
t-11-maxv,-.,_,,o] r/
X (N ri-1 ) (P o) 1 _
\ 1
The maximization of (11.1.82) is probably too difficult to make this estimator
of any practical value. Thus we must use the NLLS or NLWLS estimator if
only aggregate observations are available, even though the MLE is asymptoti-
cally more efficient, as we can conclude from the study of Baranldn and
Gurland (1951).
11.2 Duration Models

11.2.1 Stationary Models Basic Theory
We shall first explain a continuous-time Markov model as the limit of a
discrete-time Markov model where the time distance between two adjacent
time periods approaches 0. Paralleling (11.1.2), we define
.14(t) At = Prob[ith person is in state k at time t + At (11.2.1)
given that he or she was in state j at time t].

In Sections 11.2.1 through 11.2.4 we shall deal only with stationary (possibly
heterogeneous) models so that we have Aik(t) = Ajk for all t.
Let us consider a particular individual and omit the subscript i to simplify
the notation. Suppose this person stayed in state j in period (0, t) and then
moved to state k in period (1, t + At). Then, assuming t/ At is an integer, the
probability of this event (called A) is
P(A) = (1 AtraA ik At, (11.2.2)
where = (EkL 1 ).ik ) determines the probability of exiting j. But using the
well-known identity (1 n )" = e', we obtain for small At
P(A) = exp ( )f t)ilik At. (11.2.3)
Because At does not depend on unknown parameters, we can drop it and

regard exp Aj tMik as the contribution of this event to the likelihood func-
tion. The complete likelihood function of the model is obtained by first taking
the product of these terms over all the recorded events of an individual and
then over all the individuals in the sample. Because of (11.2.3), a stationary
model is also referred to as an exponential model.
Suppose M = 3 and a particular individual's event history is as follows: This
person stays in state 1 in period (0, t1 ), moves to state 2 at time t / and stays
there until time t 1 + t2 , moves to state 3 at time t i + t2 and stays there until
t1 + t2 + t3 , at which point this person is observed to move back to state 1.
(The observation is terminated after we have seen him or her move to state 1.)
Then this person's likelihood function is given by
L = exp (, _t 1 , 12 ern ( t _23
2_2, exn ( /1. 3t 3,1/1.31. (11.2.4)
If we change this scenario slightly and assume that we observe this person to
leave state 3 at time t 1 + t2 + t3 but do not know where he or she went, we
should change A31 to A3 in (11.2.4). Furthermore, if we terminate our observa-
tion at time t, + t 2 + t3 without knowing whether he or she continues to stay in
state 3 or not, we should drop A31 altogether from (11.2.4). In this last case we
say "censoring (more exactly, right-censoring) occurs at time t 1 + t2 + t3 ."
Let us consider the simple case of M = 2. In this case we have A i = Al2 and
/12 = A21. (We are still considering a particular individual and therefore have
suppressed the subscript i.) To have a concrete idea, let us suppose that state 1
signifies unemployment and state 2, employment. The event history of an
individual may consist of unemployment spells and employment spells. (If
the observation is censored from the right, the last spell is incomplete.) The
individual's likelihood function can be written as the product of two terms
the probability of unemployment spells and the probability of employment
spells. We shall now concentrate on unemployment spells. Suppose our typi-
cal individual experienced r completed unemployment spells of duration
t1 , t2,. . . , tr during the observation period. Then the contribution of these r
spells to the likelihood function is given by
L= (11.2.5)
where we have defined T = Ej_ l t; and have written A for A,. The individual's
complete likelihood function is (11.2.5) times the corresponding part for the
employment spells.
We now wish to consider closely the likelihood function of one complete
spell: e/1.. At the beginning of this section, we derived it by a limit operation,
but, here, we shall give it a somewhat different (although essentially the same)
interpretation. We can interpret e-.4 as P(T> t)where Tis a random variable
that signifies the duration of an unemployment spell. Therefore the distribu-
tion function F( ) of T is given by

F(t) = 1 P(T > t)= 1 e-A (11.2.6)
Differentiating (11.2.6) we obtain the density function f( ) of T:
f(t) = (11.2.7)
Thus we have interpreted e -Aq, as the density of the observed duration of an
unemployment spell.
From (11.2.6) and (11.2.7) we have
(11.2.8)
1
1(t).
F
The meaning of 2 becomes clear when we note
f(t)Lt
A At (11.2.9)
1 F(t)
= Prob [leaves unemployment in
(t, t At)lhas not left unemployment in
(0, 0].
We call A the hazard rate. The term originates in survival analysis, where the
state in question is "life" rather than unemployment.
Still concentrating on the unemployment duration, let us suppose that we
observe one completed unemployment spell of duration ti for the ith individ-
ual, i = 1, 2, . . . , N. Then, defining f i(ti ) = 1V exp ( '1'4), we can write the

likelihood function as
L = ll fkti), (11.2.10)
i-1
which is a standard likelihood function of a model involving continuous
random variables. Suppose, however, that individuals i = 1, 2, . . . , n com-
plete their unemployment spells of duration t, but individuals i = n + 1,
. . . , N are right-censored at time C. Then the likelihood function is
given by
L = ll f '(ti) fi [1 Fi(rio)], (11.2.11)
imn+1
which is a mixture of densities and probabilities just like the likelihood func-
tion of a standard Tobit model. Thus we see that a duration model with
right-censoring is similar to a standard Tobit model.
11.2.2 Number of Completed Spells

The likelihood function (11.2.5) depends on the observed durations
t1, 12, , tr only through r and T. In other words, r and T constitute the
sufficient statistics. This is a property of a stationary model. We shall show an

alternative way of deriving the equivalent likelihood function.
We shall first derive the probability of observing two completed spells in
total unemployment time T, denoted P(2, T). The assumption that there are
two completed unemployment spells implies that the third spell is incomplete
(its duration may be exactly 0). Denoting the duration of the three spells by t 1 ,
t2, and t3 , we have
P(2, T)= P(0 1< T, 0<1 2 T- t i , t3 T t- (11.2.12)
T
= 10 f(Z1){1 f(Z2)[ f(Z3) dZ3]dZ21dZI
0
T
= Io exp (-).z 1 )
X (1. exp (- Az2 ){exP A(T z1 - z2)1) dz2) dzi
_(A.T) 2e -IT
2
It is easy to deduce from the derivation in (11.2.12) that the probability of
observing r completed spells in total time T is given by
( Ant e-AT
P(r,T) - , (11.2.13)
r!
which is a Poisson distribution. This is equivalent to (11.2.5) because Tr and ri
do not depend on the unknown parameters.
We can now put back the subscript i in the right-hand side of (11.2.5) and
take the product over i to obtain the likelihood function of all the individuals:
L exp (- Ai Ti ). (11.2.14)
Assuming that A, depends on a vector of the ith individual's characteristics x i ,

we can specify
= exp (a + /NO, (11.2.15)
where the scalar a and the vector II are unknown parameters. We shall derive
the ML estimators of these parameters. 4 In the following discussion it will
become apparent why we have separated the constant term a from the other
regression coefficients.
The log likelihood function is given by
log L = ri(a + xi) E exp + (11.2.16)
The ML estimators t and n are the solutions of the normal equations:

a log L
aa ri ea 7; exp ("xi ) =- 0 (11.2.17)
and
o log L rot, ea Ti exp (Ind; = (11.2.18)
O.
Solving (11.2.17) for ea and inserting it into (11.2.18) yield

E E Ti exp (IN);
roc, I
E Ti exp (irxi) 0. (11.2.19)
Thus ft can be obtained from (11.2.19) by the method given in the following
paragraph. Inserting fl into (11.2.17) yields an explicit solution for a.
The following method of obtaining ft was proposed by Holford (1980).
Define
[Ti exp ('
I3 xi )
(11.2.20)
-H i Ti exp(fl'x i)
Therefore, we have
log L, = r1 log + rift' xi (11.2.21)
ri) log Ti exp (fl' xi).
Setting the derivative of (11.2.21) with respect to ft equal to 0 yields

ri T; exp (fl'xi)xi
d log LI
op rix; 0, (11.2.22)
exp (13' xi)
which is identical to (11.2.19). Note that L I is the likelihood function of a

multinomial logit model (9.3.34). To see this, pretend that the i in (11.2.20)
refers to the ith alternative and ri people chose the ith alternative. This is a
model where the exogenous variables depend only on the characteristics of the
alternatives and not on those of the individuals. Thus the maximization of L 1
andhectsoluif(1.29)canbeomplishdytanru-
nomial logit routine.
We can write L 1 as a part of the likelihood function L as follows:
( I ri) !
L = L I L2
Hi Ti" (11.2.23)
where
Mi n
exp [ exp (a+fi'xi )][2 exp (a -1-fl' xi )
L2 -
( ri)!
(11.2.24)
Note that L2 is a Poisson distribution. Setting a log L2 = 0 yields (11.2.17).
We can describe the calculation of the MLE as follows: First, maximize L i
A second, insert fl into L2 and maximize it with respect to a. withrespco
11.2.3 Durations as Dependent Variables of a Regression Equation
Suppose that each individual experiences one complete spell. Then the likeli-
hood function is
L= [IA exp (-4). (11.2.25)
The case of a person having more than one complete spell can be handled by
behaving as if these spells belonged to different individuals. Assume as before
exp (fl' xi ). (11.2.26)
But, here, we have absorbed the constant term a into fi as there is no need to
separate it out.
We shall derive the asymptotic covariance matrix of the MLE n. We have
82 log L_
Etiexp (IF x i )x,4. (11.2.27)
a
But we have
1
Et, = f zA i exp ( Aiz) dz = , (11.2.28)
which is the continuous-time analog of (11.1.65). Therefore we have

a2 log L
E I (11.2.29)
apar
Therefore, by Theorem 4.2.4, fi is asymptotically normal with asymptotic
covariance matrix
-1
= xix) . (11.2.30)
Now, suppose we use log ti as the dependent variable of a linear regression

equation. For this purpose we need the mean and variance of log ti . We have5
E log ti = (log z) exp ( Az) dz (11.2.31)
= c log Ai ,
where c=-
-= 0.577 is Euler's constant, and
E(log ti )2 = A (log z)2 exp (A i z) dz (11.2.32)
2
= n6- + (c + log A, )2.
Therefore we have
7r2
V log t = (11.2.33)
i
We can write (11.2.31) and (11.2.33) as a linear regression
log ti + c = fi'x, u, (11.2.34)
where Eui = 0 and Vu i = 7r2/6. Because (u1 ) are independent, (11.2.34) de-
fines a classical regression model, which we called Model) in Chapter 1.
Therefore the exact covariance matrix of the LS estimator ikc is given by
7,2 --1
I xix,) . (11.2.35)
6
Comparing (11.2.35) with (11.2.30), we see exactly how much efficiency we

lose by this method.
Alternatively, we can define a nonlinear regression model using ti itself as
the dependent variable. From (11.2.28) we have
exp ( xi ) + vi , (11.2.36)
where Evi = 0. We have by integration by parts

2
Et1 = z2 exp ( ili z) dz = (11.2.37)
Therefore we have
Vv i = exp ( 2fl' x i ). (11.2.38)
This shows that (11.2.26) defines a heteroscedastic nonlinear regression

model. The asymptotic normality of the NLWLS estimator ftwis can be
proved under general assumptions using the results of Sections 4.3.3 and
6.5.3. Its asymptotic covariance matrix can be deduced from (4.3.21) and
(6.1.4) and is given by
v/iwis = f,
exp (2fl' xi ) -ay) x.x' . (11.2.39)
Therefore the NLWLS estimator is asymptotically efficient. In practice Vv i

will be estimated by replacing fi by some consistent estimator of/3, such as fl
defined earlier, which converges to flat the speed of .sfrt. The NLWLS estima-
tor using the estimated Vv i has the same asymptotic distribution.
The foregoing analysis was based on the assumption that ti is the duration of
a completed spell of the ith individual. If some of the spells are right-censored
and not completed, the regression method cannot easily handle them. How-
ever, maximum likelihood estimation can take account of right-censoring, as
we indicated in Section 11.2.1.
11.2.4 Discrete Observations

In the analysis presented in the preceding three subsections, we assumed that
an individual is continuously observed and his or her complete event history
during the sample period is provided. However, in many practical situations a
researcher may be able to observe the state of an individual only at discrete
times. If the observations occur at irregular time intervals, it is probably more
reasonable to assume a continuous-time Markov model rather than a dis-

crete-time model.
We shall derive the likelihood function based on discrete observations in the
case of M = 2. The underlying model is the same stationary (or exponential)
model we have so far considered. We define
PO) = Prob [ith person is in state k at time t (11.2.40)
given that he or she was in state j at time 0].
Note that this definition differs slightly from the definition (11.1.2) of the
same symbol used in the Markov chain model. We shall use the definition
(11.2.1) with the stationarity assumption Ajk (t) = Ajk for all t.
As before, we shall concentrate on a particular individual and suppress the
superscript i. When At is sufficiently small, we have approximately
P12(t + At) = P11 (0A l2 At + P12 (t)(1 .1,21 At). (11.2.41)
Dividing (11.2.41) by At and letting At go to 0 yield
dP12 p (11.2.42)
11 , 12 P12 A21
dt
Performing an analogous operation on P11, P211 and P22, we obtain the linear
vector differential equation
dt
ddP: 1
r_ A21 111 P21 (11.2.43)
dPi2 dP22 L Al2 A21]
[P12 P22
dt dt
dP
o
r dt
A solution of (11.2.43) can be shown to be 6
P' = H (11.2.44)
where the columns H are characteristic vectors of A, D is the diagonal matrix
consisting of the characteristic roots of A, and eD is the diagonal matrix
consisting of exp (di t), di, being the elements of D.
We shall derive D and H. Solving the determinantal equation
IA dIl = 0 (11.2.45)
yields the two characteristic roots d 1 = 0 and d2 = (Al2 + A21). Therefore we

have
D=
F0 0
(11.2.46)
LO (2 12 221 )] .
Let h 1 be the first column of H (the characteristic vector corresponding to the
zero root). Then it should satisfy A h i = 0, which yields a solution (which is
not unique) h 1 = (221, 212). Next, the second column h 2 of H should satisfy
[A + (2 12 + 221 )Ilh2 = 0, which yields a solution h2 = (-1, 1) '. Combining
the two vectors, we obtain
-1
(11.2.47)
1
Finally, inserting (11.2.46) and (11.2.47) into (11.2.44) yields the following
expressions for the elements of P' (putting back the superscript i):
(t) = 1 exp ( - 3; 0 (11.2.48)
P12(t) = Yi yi exP ( 6 ; 0
n(t)= 1 y, (1 yi ) exp
13 22 (t) = yi+ ( 1 yi) exP ( t),
where y i = NI2AAl2 + 11 ) and (5i= Ai2 A . Suppose we observe the ith indi-
vidual in state j1 at time 0 and in state ki at time to i = 1, 2, . . . , N. Then the
likelihood function is given by
L = 11 plk,(ti ). (11.2.49)
11 .2.5 Nonstationary Models

So far we have assumed 4(t) = Aik for all t (constant hazard rate). Now we
shall remove this assumption. Such models are called nonstationary or semi-
Markov.
Suppose a typical individual stayed in state j in period (0, 1) and then moved
to state k in period (t, t + At). We call this event A and derive its probability
P(A), generalizing (11.2.2) and (11.2.3). Defining m = At and using
log (1 e) e for small E, we obtain for sufficiently large m
P(A) = 1[11 _ 117_7 ]} /WO (11.2.50)
= exp [ Ai (z) dz]} A ik (t)
The likelihood function of this individual is the last expression in (11.2.50)

except for Um.
Let us obtain the likelihood function of the same event history we consid-
ered in the discussion preceding (11.2.4). The likelihood function now be-
comes
L = exp[ (z) dz] Al2 ) (11.2.51)

0
ti+12
X exp [ f A2 (z) dziA 23 (t, + 12 )
ti+4+13
X exp [ (z) dz1 A31 (t + t2 + t) ).
Ito-t2
As in (11.2.4), A31 should be changed to A3 if the individual is only observed to
leave state 3, and A31 should be dropped if right-censoring occurs at that time.
We shall concentrate on the transition from state 1 to state 2 and write Al2 (t)
simply as A(t), as we did earlier. The distribution function of duration under a
nonstationary model is given by
F(t) = 1 exp [ Az) dz], (11.2.52)
which is reduced to (11.2.6) if A(t) = A. The density function is given by
f(t) = A(t) exp [ A(z) dd. (11.2.53)
The likelihood function again can be written in the form of (11.2.10) or

(11.2.11), depending on whether right-censoring is absent or present.
Thus we see that there is no problem in writing down the likelihood func-
tion. The problem, of course, is how to estimate the parameters. Suppose we
specify A! (t) generally as
Aqt) = g(x i fl). (11.2.54)
To evaluate ffyli(z) dz, which appears in the likelihood function, a researcher

must specify xi, precisely as a continuous function oft not an easy thing to
do in practice. We shall discuss several empirical articles and shall see how
these articles deal with the problem of nonstationarity, as well as with other
problems that may arise in empirical work.
Model of Tuma. Tuma (1976) analyzed the duration of employment at a
particular job. The ith individual's hazard rate at time t is specified as
,1!(t)= r xi + ai t + a2 t2, (11.2.55)
where xi is a vector of socioeconomic characteristics of the ith individual: 7 The
parameter values are assumed to be such that the hazard rate is nonnegative
over a relevant range of t. Because of the simple form of the hazard rate, it can
easily be integrated to yield
l t2 +
Mz) dz = frx i t + a az t3. (11.2.56)
2 3
Some people terminate their employment during the sample period, but some
remain in their jobs at the end of the sample period (right-censoring). There-
fore Tuma's likelihood function is precisely in the form of (11.2.11).
Model of Tuma, Hannan, and Groeneveld. Tuma, Hannan, and Groene-
veld (1979) studied the duration of marriage. They handled nonstationarity
by dividing the sample period into four subperiods and assuming that the
hazard rate remains constant within each subperiod but varies across different
subperiods. More specifically, they specified
Mt) = ffp xi for t E Tp , p= 1, 2, 3, 4, (11.2.57)
where 7; is the pth subperiod. This kind of a discrete change in the hazard rate
creates no real problem. Suppose that the event history of an individual
consists of a single completed spell of duration L and that during this period a
constant hazard rate ).(1) holds from time 0 to time rand another constant rate
).(2) holds from time T to time t. Then this individual's likelihood function is
given by
L = e-uoTe -42xt-rw2) . (11.2.58)
Model of Lancaster. Lancaster (1979) was concerned with unemployment
duration and questioned the assumption of a constant hazard rate. Although a
simple search theory may indicate an increasing hazard rate, it is not clear
from economic theory alone whether we should expect a constant, decreasing,
or increasing hazard rate for unemployment spells. Lancaster used a Weibull

distribution, which leads to a nonconstant hazard rate. The Weibull distribu-
tion is defined by
F(t)= 1 exp ( )t ). (11.2.59)
If a = 1, (11.2.59) is reduced to the exponential distribution considered in the
preceding subsections. Its density is given by
f(t)= Actta -1 exp ( )r) (11.2.60)
and its hazard rate by
f(t)
A(t) Aceta - l. (11.2.61)
1 F(t)
Whether a is greater than 1 or not determines whether the hazard rate is
increasing or not, as is shown in the following correspondence:
a> 1 '- 0 (increasing hazard rate) (11.2.62)
OA
a = 1 *-0 = 0 (constant hazard rate)
at
OA
a< 1 <0 (decreasing hazard rate).
Lancaster specified the ith person's hazard rate as

Ai(t) = at' ("xi ), (11.2.63)
where xi is a vector of the ith person's characteristics. His ML estimate of a
turned out to be 0.77, a result indicating a decreasing hazard rate. However,
Lancaster reported an interesting finding: His estimate of a increases as he
included more exogenous variables in the model. This result indicates that the
decreasing hazard rate implied by his first estimate was at least partly due to
the heterogeneity caused by the initially omitted exogenous variables rather
than true duration dependence.
Because it may not be possible to include all the relevant exogenous vari-
ables, Lancaster considered an alternative specification for the hazard rate
'AO= vdf(t), (11.2.64)
where gt) is as given in (11.2.63) and v i is an unobservable random variable
independently and identically distributed as Gamma(1, a 2 ). The random
variable vi may be regarded as a proxy for all the unobservable exogenous

variables. By (11.2.52) we have (suppressing i)
F(t1 = 1 exp [ vA(t)], (11.2.65)
where A(t) = f t(z) dz. Taking the expectation of (11.2.65) yields
F*(t)=- EL,F(tiv)= 1 [1 + a2A(O] 2. (11.2.66)
Therefore, using (11.2.8), we obtain
A*(t)= A(t)[1 + cr2A(O]' (11.2.67)
= )(t)[1 F*(t)]a2 .
Because [1 F*(0]c 2 is a decreasing function of t, (11.2.67) shows that the
heterogeneity adds a tendency for a decreasing hazard rate. Under this new
model (with (7 2 as an additional unknown parameter), Lancaster finds the
MLE of a to be 0.9. Thus he argues that a decreasing hazard rate in his model is
caused more by heterogeneity than true duration dependence.
Model of Heckman and Borjas. In their article, which is also concerned
with unemployment duration, Heckman and Borjas (1980) introduced an-
other source of variability of in addition to the Weibull specification and
heterogeneity. In their model, A also varies with spells. Let I denote the /th
unemployment spell the ith individual experiences. Then Heckman and
Borjas specify the hazard rate as
Au(t ) = at' exp (irixd + vi ), (11.2.68)
where vi is unobservable and therefore should be integrated out to obtain the
marginal distribution function of duration. 8
Model of Flinn and Heckman. In a study of unemployment duration,
Flinn and Heckman (1982) generalized (11.2.68) further as
t
Al 1 tA2 1
)."(t) = exp [gx11(t) + civi + 71 + Y2 1 (11.2.69)
A2 ] .
The function (t A - 1)/ A is the Box-Cox transformation (see Section 8.1.2) and
approaches log t as A approaches 0. Therefore putting A l = 0 and y2 = 0 in
(11.2.69) reduces it to a Weibull model. Note that x u is assumed to depend on t
in (11.2.69). Flinn and Heckman assumed that changes in xu(t) occurred only
at the beginning of a month and that the levels were constant throughout the
month. The authors devised an efficient computation algorithm for handling
the heterogeneity correlated across spells and the exogenous variables varying
with time.
11 .2.6 Problem of Left-Censoring

In this subsection we shall consider the problem of left-censoring in models for
unemployment duration. When every individual is observed at the start of his
or her unemployment spell, there is no problem of left-censoring. Often,
however, an individual is observed to be in an unemployment spell at the start
of the sample period. This complicates the estimation.
Suppose that an unemployment spell of a particular individual starts at
time s, the individual is interviewed at time 0, and the spell terminates at t.
(To simplify the analysis, we assume that right-censoring at time t does not
occur.) The treatment of left-censoring varies according to the following three
cases: (1) s is observed but t is not observed, (2) both s and t are observed, and
(3) t is observed but s is not observed. For each case we shall derive the relevant
likelihood function.
The first case corresponds to the situation analyzed by Nickell (1979),
although his is a Markov chain model. Assuming that the underlying distribu-
tion of the duration is F( ) and its densityf(-), we can derive the density g(s).
Denoting the state of "being unemployed" by U, we have for sufficiently
small As
g(s) As = P[U started in (s As, s)IU at 01 (11.2.70)
_ P[U at 01 U started in (s As, s)]P[U started in (s As, s)]
I (Numerator)
_ P[U at 01 U started in (s As, s)] As

ds
J.
: (Numerator) ds
[1 F(s)] As _ [1 F(s)] As
ES '
.1.-[1 _F(s)] ds
0
where ES = f;sf(s) ds. In (11.2.70) the third equality follows from the as-
sumption that P[U started in (s As, s)] does not depend on s (the as-
sumption of constant entry rate), and the last equality follows from integra-
tion by parts. By eliminating As from both sides of (11.2.70), we obtain
1 F(s)
g(s) (11.2.71)
ES
In the second case we should derive the joint density g(s,t)= g(tls)g(s). The
density g(s) has been derived, but we still need to derive g(t1s). Let X denote
total unemployment duration. First evaluate
P(X>s+t,X>s)
P(X>s+tlX>s) (11.2.72)
P(X> s)
_ P(X>s+t)
P(X > s)

1 F(s + t)
1 F(s)
If we denote the distribution function of g(t1s) by G(t1s), (11.2.72) implies
F(s + t)
G(t1s) (11.2.73)
1 F(s) .
Therefore, differentiating (11.2.73) with respect to t, we obtain
f(s + t)
g(t1s) (11.2.74)
1 F(s) .
Finally, from (11.2.71) and (11.2.74), we obtain
f(s + t)
g(s,t) (11.2.75)
ES
This situation holds for Lancaster (1979), as he observed both s and t.
However, Lancaster used the conditional density (11.2.74) rather than the
joint density (11.2.75) because he felt uncertain about the assumption of
constant entry rate.
Finally, in the third case we need g(1). This can be obtained by integrating
g(s,t) with respect to s as follows:
1 r-
o f(s + t) ds (11.2.76)
g(t) = -is J
_1 F(t)
ES
See Flinn and Heckman (1982) for an alternative derivation of g(t).
11.2.7 Cox's Partial Maximum Likelihood Estimator

Cox (1972) considered a hazard rate of the form
Ai(t) = A(t) exp (ff (11.2.77)
which generalizes (11.2.63). This model is often referred to as a proportional

hazards model. Cox proposed a partial MLE (PMLE) obtained by maximiz-
ing a part of the likelihood function. We shall first derive the whole likelihood
function and then write it as a product of the part Cox proposed to maximize
and the remainder.
Cox allowed for right-censoring. Let ti , i 1, 2, . . . , n, be completed
durations and let ti , i = n + 1, n + 2, . . . , N, be censored durations. We
assume for simplicity (ti) are distinct.9 Then the likelihood function has the
form (11.2.11): specifically,
L =fi exp (fi'x1).1.(ti) exp [ exp (fixi )A(ti )] (11.2.78)

i-1
exp [ exp (ffxi )A(4)1,

i n+ I
where A(t) = f l(z) dz. Combining the exp functions that appear in both
terms and rewriting the combined term further, we obtain
L= exp (ifxdAti ) exp [ E exp (fi'xi ) A(z) dzi (11.2.79)
= ft exp (fi'xi)A.(ti) exp { [exp (xh )] At) dt},

helt(t)
where R(t)= (ilt i t). To understand the second equality in (11.2.79), note
that I hF_Ao) exp xh ) is a step function described in Figure 11.1.
Cox's PMLE flp maximizes
exp (fl' xi)

L1 =ft (11.2.80)
E exp (fl'xh)
hER(4)
It is a part of L because we can write L as
L = L iL2 (11.2.81)
e l311 "
ijeo'xN
t) t2 t N-1 tN
Figure 11.1 IhERio exP (rxh ) as a function of t
where
L2 = [ exp (fi'x h )A.(ti )] (11.2.82)

1 11F_R(4)
X exp { [ I exp (fi'xh )]t A( ) dt}.

J o hF_Ru)
Because L 1 does not depend on 2(1), Cox's method enables us to estimate fi
without specifying A(t). Cox (1975) suggested and Tsiatis (1981) proved that
under general conditions the PMLE is consistent and asymptotically normal
with the asymptotic covariance matrix given by
.92 log LIT (11.2.83)
This result is remarkable considering that L i is not even a conditional likeli-

hood function in the usual sense.m
However, L 1 and L2 do have intuitive meanings. We shall consider their
meanings in a simple example. Suppose N = 2 and t 1 is a completed duration
(say, the first person dies at time t 1 ) and t2 a censored duration (the second
person lives at least unti time t 2 ), with 12 > 11 . Then we have
12
I
L = 21 01 ) exp [ A' (z) dz]- exp [ f 22 (z) dz]. (11.2.84)
Jo
We can write (11.2.84) as

L = LIL21 1-221-23, (11.2.85)
where
(ti)
Li (11.2.86)
L21 = A' (t i ) + A2 (11), (11.2.87)

ti
L22 = eXP [ A' (z) dz] exp A2 (z) dz], (11.2.88)
0
and
L23 = exp [- 1:2 112 (Z) (11.2.89)

t,
These four components of L can be interpreted as follows:

L, = P(#1 dies at t i 'both #1 and #2 live until t, and either #1 or #2 dies
at t 1 )
L21 = P(either #1 or #2 dies at t, both #1 and #2 live until t, )
L22 = P(both #1 and #2 live until t i )
L23 = P(#2 lives at least until t 2 1#2 lives until t, ).
Note that L21L22L23 corresponds to L2 of (11.2.82).
Kalbfleisch and Prentice (1973) gave an alternative interpretation of Cox's
partial likelihood function. First, consider the case of no censoring. Let
1, <t2 <. . . < tN be an ordered sequence of durations. Then by successive
integrations we obtain
P(t, < t2 < . . . < tN ) (11.2.90)
Go f N
A(t i) exp (fl' x i )
4 L -1 1..1
i,
X exp [exp (fl'x i ) 0 A(z) dz dt N dtN_ i . . dt i
CO f CO E. N-2
A(t i ) exp xi )
fo
X exp [ exp (fl'xi ) A(z) dzi
Jo
X A(t N_ 1 ) exp (fi'xN_ I )
6y I
X exp { [exP (fi'xN- 1 ) + exP (IrxN )] J. 2(z) di
X dt N_ I dtN- 2 dt
co le
.1, , -, H
co N-2
. 2(0 exp (fi'xi )
10 z, fI
6
X exp [ exp (11'10 f2(z) dzi

o
exp (fi'xx-i)
X
exp (fi'xN-1) + exP (fi'xN)
exp {[exp (fi'xN_ I )+ exp (fi'xN)] f 2(z) di
X dtN_2 dtiy_3 dt,

= exp (fi'xN) exp (Pc exp
+ (exp (fi'xN)[exp (fi'xN) + exp (fl'xN_ 1 )]
[exp (ft'xN) + exp (fi'xN_,) +. . . + exp (fi'x i )]),
Next, suppose that completed durations are ordered as t, <t 2 <. . <
and in the interval [4,4+1 ), i= 1, 2, . . . , n (with the understanding
40. 1 = co), we observe censored durations t n , . . . , tic . Then we have
POI < /2 < . . . < t,ti , ta, . , tie for all i) (11.2.91)
P(t, t, 1 ,t, ..,i , for all *1 ,12 , . . tn)
X f (Of 2 (t2) IVO dt dt_i . dt1

n
n
2(4) exp (P'xi )
x exp[exp (fix,) + (exp fi'x u)] I 2(z) dz}

.1-1
X dt dt_, . .dt 1
= exp (fi'x) exp ("x--1) exP (ft'XI)

(LexP (fi'xn) + caexP + exP (rx-1) + c1 +
[exp (trx.)+ . . . + exp ("x i ) + c + . . + 1),
where ci = exp (fi'xu).
Because the parameter vector fl appears in both L i and L2, we expect Cox's
PMLE to be asymptotically less efficient than the full MLE in general. This is
indeed the case. We shall compare the asymptotic covariance matrix of the
two estimators in the special case of a stationary model where the A(t) that
appears in the right-hand side of (11.2.77) is a constant. Furthermore, we shall
suppose there is no censoring. In this special case, Cox's model is identical to
the model considered in Section 11.2.3. Therefore the asymptotic covariance
matrix of the MLE fl can be derived from (11.2.30) by noting that the 13 of the
present section can be regarded as the vector consisting of all but the last
element of the ft of Section 11.2.3. Hence, by Theorem 13 of Appendix 1, we
obtain
Vft = [ (xi )1)(xi (11.2.92)
The asymptotic covariance matrix of Cox's PMLE can be derived from

(11.2.83) as
N -2
VA, =
i E I xp
[1 e (fl'xh )] (11.2.93)
il h
X I xh x;, exp (flixh )I exp (Pi Xh)

h h
xh exp ( fifxh ) x'h exp ("Xis )
11-1,
where Ih denotes IhEito) and E denotes the expectation taken with respect to
random variables ti that appear in R(4).
This expectation is rather cumbersome to derive in general cases. There-
fore, we shall make a further simplification and assume /3 is a scalar and equal
to 0. Under this simplifying assumption, we can show that the PMLE is
asymptotically efficient. (Although this is not a very interesting case, we have
considered this case because this is the only case where the asymptotic var-
iance of the PMLE can be derived without lengthy derivations while still
enabling the reader to understand essentially what kind of operation is in-
volved in the expectation that appears in (11.2.93). For a comparison of the
two estimates in more general cases, a few relevant references will be given at
the end.) Under this simplification (11.2.93) is reduced to
(Ixh) 2
VA, 1E
}
] -1
= ;1 (11.2.94)
We shall first evaluate (11.2.94) for the case N = 3 and then for the case of
general N. If t, <12 < 13 , we have
v3 .X11 + x3i + xi
(11.2.95)
[ 4- + x13
If we change the rank order of (t i, 12, 13 ), the right-hand side of (11.2.95) will
change correspondingly. But, under our stationarity (constant A) and homoge-
neity (fl = 0) assumptions, each one of six possible rank orderings can happen
with an equal probability. Therefore we obtain
Et [ _ V] (11.2.96)
Similarly, if t i <12 < 13 , we have

3 [ Xh [
1 X + X2 + X3] 2 [X2 + X31 + [X31 2.
(11.2.97)
3 2 1
By using an argument similar to that above, we obtain
E [
/Xi, 2
1 1
= + 1) tX, (11.2.98)
031 ( 222 + 6
i 2x xi
3 / j-i+1 i
Generalizing (11.2.96) and (11.2.98) to a general N, we obtain
X h2 N
x, 1
EE [-- Ex? (11.2.99)
and
(11.2.100)
The coefficient on IliZiE./,`:_i+1 2x1x; can be derived as follows: There are N!

permutations of N integers 1, 2, . . . , N, each of which happens with an
equal probability. Let Pk be the number of permutations in which a
given pair of integers, say, 1 and 2, appear in the last k positions. Then
Pk= ()2[(N 2)!]. Therefore the desired coefficient is given by vsnyiI,P,1_2
(q)2[(N 2)!]/k2 ).
Finally, from (11.2.94), (11.2.99), and (11.2.100), we conclude
N k_ 1N -
= E (11.2.101)
[ iv k-2 i-I
Compare (11.2.101) with Vft' given in (11.2.92). Because

I log N 1 N k
v
(11.2.102)
1 Nl < N- 1 h2 k I'
we see that Cox's PMLE is asymptotically efficient in this special case.
Kalbfleisch (1974) evaluated the asymptotic relative efficiency of the
PMLE also for nonzero /3 in a model that is otherwise the same as the one we
have just considered. He used a Taylor expansion of (11.2.93) around 13 = 0.
Kay (1970) extended Kalbfleisch's results to a case where II is a two-dimen-
sional vector. Han (1983), using a convenient representation of the asymp-
totic relative efficiency of the PMLE obtained by Efron (1977), evaluated it for
Weibull as well as exponential models with any number of dimensions of the ft
vector and with or without censoring. Cox's estimator is found to have a high
asymptotic relative efficiency in most of the cases considered by these authors.
However, it would be useful to study the performance of Cox's estimator in
realistic situations, which are likely to occur in econometric applications.
Exercises
1. (Section 11.1.1)
Using (11.1.5), express the unconditional mean and variance of yl (2) as a
function of the mean and variance of y'(0).
2. (Section 11.1.1)
Write L and L2 defined in (11.1.10) explicitly in the special case where
Pi/JO = Pik, M = 2, and T = 2.
3. (Section 11.1.1)
Find an example of a matrix that is not necessarily positive but for which a
unique vector of equilibrium probabilities exists.
4. (Section 11.1.1)
Prove the statement following (11.1.15).
5. (Section 11.1.1)
Verify statement (11.1.23).
6. (Section 11.1.3)
Derive the asymptotic variance-covariance matrix of the MLE of a in the
Bosldn-Nold model using (11.1.38).
7. (Section 11.1.3)
The mean duration is derived in (11.1.65). Using a similar technique,
derive Vt,
8. (Section 11.1.5)
Let the Markov matrix of a two-state (1 or 0) stationary homogeneous
first-order Markov chain model be
[P11 P10]_[1 A Al
Pol Poo
where A is the only unknown parameter of the model. Define the follow-
ing symbols:
nik Number of people who were in state j at time 0 and are in state k
at time 1
/If . Number of people who were in state j at time 0
n. Number of people who are in state j at time 1
We are to treat n. as given constants and nik and n. as random variables.

a. Supposing n, . = 10, no . = 5, n . 1 = 8, and n. 0 = 7, compute the least
squares estimate of A based on Eq. (11.1.73). Also, compute an estimate
of its variance conditional on nJ . .
b. Supposing n 11 = 7 and no, = 1 in addition to the data given in a,

compute the MLE of A and an estimate of its variance conditional on n i .
9. (Section 11.1.5)
Prove that minimizing (11.1.78) yields the asymptotically same estimator
as minimizing (11.1.77).
10. (Section 11.1.5)
Write down the aggregate likelihood function (11.1.82) explicitly in the
following special case: T= 1, N= 5, ro = 3, and r, = 2.
11. (Section 11.2.5)

Verify (11.2.50).
12. (Section 11.2.5)
Consider a particular individual. Suppose his hazard rate of moving from
state 1 to state 2 is a i t + fli k, where t is the duration in state 1 and k
denotes the kth spell in state I. The hazard rate of moving from state 2 to 1
is a21 + /32k, where t is the duration in state 2 and k denotes the kth spell in
state 2. Suppose he was found in state 1 at time 0. This is his first spell in
state 1. (He may have stayed in state 1 prior to time 0.) Then he moved to
state 2 at time t 1 , completed his first spell in state 2, moved back to state 1
at time t2 , and stayed in state 1 at least until time t 3 . Write down his
contribution to the likelihood function. Assume a i , a2 > 0.
13. (Section 11.2.5)
Consider a homogeneous nonstationary duration model with the hazard
rate
.1.(t) = at, a> 0.
Supposing we observe n completed spells of duration t 1 , t2 ,. . . , t, and

N n censored spells of duration t-n+1 -n+2
t 9 9 tN, derive the MLE of
a and its asymptotic variance.

14. (Section 11.2.5)
Let F(t1).) = 1 e -Al where A is a random variable distributed with den-
sity g( ). Define A(t) = f(01[1 F(t)], where F(t) = EAF(tIA) and f(t) =
anat. Show mtvat < 0 (cf. Flinn and Heckman, 1982).
Appendix 1
Useful Theorems in Matrix Analysis
The theorems listed in this appendix are the ones especially useful in econo-
metrics. All matrices are assumed to be real. Proofs for many of these
theorems can be found in Bellman (1970).
1. For any square matrix A, with distinct characteristic roots, there exists a
nonsingular matrix P such that PAP-1 = A, where A is a diagonal matrix with
the characteristic roots of A in the diagonal. If the characteristic roots are not
distinct, A takes the Jordan canonical form (see Bellman, p. 198).
2. The determinant of a matrix is the product of its characteristic roots.
3. For any symmetric matrix A, there exists an orthogonal matrix H such
that H' H = I and H 'AH = A, where A is a diagonal matrix with the charac-
teristic roots (which are real) of A in the diagonal. The ith column of H is called
the characteristic vector of A corresponding to the characteristic root of A that
is the ith diagonal element of A.
4. For a symmetric matrix A, the following statements are equivalent:
(i) A is a positive definite. (Write A> 0.)
(ii) x 'Ax is positive for any nonzero vector x.
(iii) Principal minors of A are all positive.
(iv) Characteristic roots of A are all positive.
The above is true if we change the word positive to nonnegative.
5. For any matrices A and B, the nonzero characteristic roots of AB and BA
are the same, whenever both AB and BA are defined.
6. tr AB = tr BA.
7. For any square matrix A, tr A is equal to the sum of its characteristic
roots.
8. Let A, B be symmetric matrices of the same size. A necessary and
460 Appendix 1
sufficient condition that there exists an orthogonal matrix H such that both
H'AH and H'BH are diagonal is that AB = BA.
9. Any nonnegative (semipositive) definite matrix A can be written as
A = TT', where T is a lower triangular matrix with nonnegative diagonal
elements.
10. Let A i , 22, . . . , 2,, be the characteristic roots of a symmetric matrix A
in descending order (2 1 being the largest). Then
x'Ax ( x',Axi
A., = max = ,
,
x'Ax ( xAx2
.1.2 = max
e )
xt o x'x 3c x2 -'
x 'Ax
23= max-, and so on.
xxe-o x x
exv-o
11. Let A and B be symmetric matrices (n X n) with B nonnegative defi-
nite. Then it i(A + B) A i(A), i, = 1, 2, . . . , n, where ).'s and p's are the
characteristic roots in descending order. The strict inequality holds if B is
positive definite.
12. det [' -1CI if 'DI* O.

15:01 = IDI IA BD
[A 11 1 = [ E -1 E -1BD -1 ]
13.
CD D -10E -1 F -1
where E = A BD -1 C, F= D CA-1 B, E -1 = A -1 + A -111F- ICA -1 ,
and F -1 = D-1 + D -10E -1 BD-1 .
14. Let A be an n X r matrix of rank r. A matrix of the form P =
A(A'A)- 'A' is called a projection matrix and is of special importance in
statistics.
(i) P = P' = P 2 (Hence, P is symmetric and idempotent.)
(ii) rank (P) = r.
(iii) Characteristic roots of P consist of r ones and n r zeros.
(iv) If x = Ac for some vector c, then Px = x. (Hence the word pro-
jection.)
(v) M = I A(A'A)-1A' is also idempotent, with rank n r, its char-
acteristic roots consisting of n r ones and r zeros, and if x = Ac,
then Mx =0.
Appendix 1 461
(vi) P can be written as G'G, where GG' = I, or as v l v c + v2 v +

. . . + v rv ' where v i is a vector and r = rank (P).
15. Let A be an n X r matrix of rank r. Partition A as A = (A,, A2) such that
Al is n X ri and A2 is n X r2 . If we define A2 = [I - A l (Al A i ) 1A ]A2 , then
we have A(A'A) -1A' = Al(AA1 )-1A1 + A2(--kA2)-1A.
16. If A is positive definite and B is symmetric, there exists a nonsingular
matrix C such that C' AC = I, C' BC = D, where D is diagonal and the diago-
nal elements of D are the roots of 1B 1A1= 0.
17. Let A and B be symmetric nonsingular matrices of the same size. Then
A B 0 implies 13 -1 -..- A - '.
18. Let A be any nonsingular matrix. The characteristic roots of A -1 are the
reciprocals of the characteristic roots of A.
19. Let A and B be nonsingular matrices of the same size such that A + B
and A -1 + B -1 are nonsingular. Then
(i) (A + B)-1 = A -1 (A -1 + B-I)iB-1
(ii) A -1 (A + Br ' = A -1(A -1 + B -1 )-1A -1 .
20. Let X be a matrix with a full column rank. Then
(I + XX 1 )-1 = I X(I + X 1 X)-1 X'.
21. Rules for Matrix Differentiation. is an element of X. Every other
symbol is a matrix. II II denotes the absolute value of the determinant.
d dX
(i) -c tr XY = tr w Y
d dX
-
(ii) log IIXII = tr ?C1
og cg
d
(iii) log II All = (A ' )'
dA
d _ dX
(iv)
W
c X' = X-' 7-
g X-1
d
(v) T- tr X -1 Y ft (X -4
dX X-1 1')
g
d
(vi) a tr A-1 B = A -1 BA -1
462 Appendix 1
(vii) tr AB = B'
(viii) tr ABA'C = CAB + C' AB'
tr A 'BAC = B ' AC' + BAC

(ix) dA
22. Kronecker Products. Let A = {a u) be a KXL matrix and let B be a
M X N matrix. Then the Kronecker product A 0 B is a IW X LN matrix
defined by
ll B (212B a 'L B 1
a2 1 B a22B a2L B
A 0 B [a
='
. .
am B aK2 B aKL B
(i) (A 0 B)(C D) = AC BD if AC and BD can be defined.
(ii) tr (A CD B) = tr A tr B if A and B are square.
(iii) Let A be a KXK matrix with characteristic roots
, A2, , ;grand let B be a M X Mmatrix with characteristic
rootsit i , #2, , Pm. Then the KMcharacteristic roots of A 0 B
are A iy;, i= 1, 2, . .. , K, j = 1, 2, . .. , M.
(iv) Let A and B be as in (iii). Then
IA BI = 'Al m IBIK.
(v) Let A, B, and C be matrices of appropriate sizes such that ABC can
be defined. Suppose A has L columns and write the L columns as
A = (A1, A2,. . . , A L). Define vec (A) = ,A...,
Then
vec (ABC) = (C' 0 A) vec (B).
23. Hadamard Products. (Minc and Marcus, 1964, p. 120.) Let A = (a u)
and B = {b 1 ) be matrices of the same size. Then the Hadamard product A * B
is defined by A * B = (au b u ).
(i) Let A and B be both n X n. Then A * B = S' (A 0 B)S, where S is the
n 2 X n matrix the ith column of which has 1 in its [(i 1)n + i]th position
and 0 elsewhere.
Appendix 2
Distribution Theory
The theorems listed in this appendix, as well as many other results concerning
the distribution of a univariate continuous random variable, can be found in
Johnson and Kotz (1970a,b). "Rao" stands for Rao (1973) and "Plackett" for
Plackett (1960).
1. Chi-Square Distribution. (Rao, p. 166.) Let an n-component random

vector z be distributed as N(0, I). Then the distribution of z ' z is called the
chi-square distribution with n degrees of freedom. Symbolically we write
z' z A. Its density is given by
f(x) = 2-n121-(n12)e-x12x- n/2-1,
where I- ( p) = F,AP-1 e-1 chi, is called a gamma function. Its mean is n and its
variance 2n.
2. (Rao, p. 186). Let z N(0, I) and let A be a symmetric and idempotent
matrix with rank n. Then z' Az A.
3. Student's t Distribution. (Rao, p. 170.) If z N(0, 1) and w A and if z
and w are independent, n I12 ZW-112 has the Student's t distribution with n
degrees of freedom, for which the density is given by
1 ng l( x2-0+
) 0/2
f(x) =PO
T 1 -I-
where
beta function.
Symbolically we write
(n) 1/2
Sn .
464 Appendix 2
4. F Distribution. (Rao, p. 167.) If 11,1 4 and w2 4 and if w 1 and w2

n-1 1 w 1 n2 wi ' has the F distribution with n i and n2 degrees ofareindpt,
freedom, for which the density is given by
(n 1I 2)","
n 2 xh,/2-1
f(x)
9[(n 1 /2),(n2 /2)](1 + n 1 x/n2)(. -"0/2
Symbolicawert
w, In,
F(n, , n2).
W2 /n2
5. (Plackett, p. 24.) Let z N(0, I). Let A and B be symmetric matrices.
Then z' Az is independent of z'Bz if and only if AB = 0.
6. (Plackett, p. 30.) Let z N(0, I). Then c' z and z' Az are independent if
and only if Ac =0.
7. (Plackett, p. 30.) Let z N(0, I) and let A be a symmetric matrix. Then
z ' Az/z' z and z 'z are independent.
Notes
1. Classical Least Squares Theory

1. This statement is true only when x7 is a scalar. If y, x1 , and x2 are scalar dichoto-
mous random variables, we have E(ylx, , x 2) = flo + fl1 x1 + ,62x2 + 133x1 x2 , where the
ft's are appropriately defined.
2. The ft that appears in (1.2.1) denotes the domain of the function S and hence,
strictly speaking, should be distinguished from the parameter/1, which is unknown and
yet can be regarded as fixed in value. To be precise, therefore, we should use two
different symbols; however, we shall use the same symbol to avoid complicating the
notation, so the reader must judge from the context which of the two meanings the
symbol conveys.
3. Again, the warning of note 2 is in order here. The // and a2 that appear in the
likelihood function represent the domain of the function whereas those that appear in
Model 1 are unknown fixed values of the parametersso-called true values.
4. The Cramer-Rao lower bound is discussed, for example, by Cox and Hinkley
(1974), Bickel and Doksum (1977), Rao (1973), and Zacks (1971), roughly in ascend-
ing order of sophistication.
5. Good references for 13ayesian statistics are Zenner (1971) and Box and Tiao
(1973). For a quick review, Lindley (1972) is useful.
6. Even in this situation some people prefer using the t test. The procedure is also
asymptotically correct, but there is no theoretical reason for the t test to be superior to
the standard normal test in nonnormal cases. The same remark applies to the F test
discussed later. See Pearson and Please (1975) for a discussion of the properties of the t
and F tests in nonnormal cases. See White and MacDonald (1980) for tests of nonnor-
mality using the least squares residuals.
7. Chow (1960) considered the case where T2 < K and indicated how the subse-
quent analysis can be modified to incorporate this case.
8. This formula can easily be generalized to the situation where there are n regimes
(n > 2). Simply combine n equations like (1.5.25) and (1.5.27) and calculate the sum
of squared residuals from each equation. Then the numerator chi-square has
(n 1)K* degrees of freedom and the denominator chi-square has 17_, T1 nr* de-
grees of freedom.
9. These two tests can easily be generalized to the case of n regimes mentioned in
note 8.
466 Notes to Pages 38-81
10. If we have cross-section data, p refers to a cross-section unit not included in the
sample.
11. Theorem 1.6.1 is true even if we let C depend on x, . Then the two theorems are
equivalent because any d satisfying (1.6.5) can be written as Cz i, for some C such that
C'X = I.
2. Recent Developments in Regression Analysis
1. A much more detailed account of this topic can be found in Amemiya (1980a).
2. We use the term estimator here, but all the definitions and the results of this
subsection remain the same if we interpret d as any decision function mapping Y
into 0.
3. We assume that the losses do not depend on the parameters of the models.
Otherwise, the choice of models and the estimation of the parameters cannot be
sequentially analyzed, which would immensely complicate the problem. However, we
do not claim that our assumption is realistic. We adopt this simplification for the
purpose of illustrating certain basic ideas.
4. For ways to get around this problem, see, for example, Alcaike (1976) and
Schwartz (1978).
5. See Thisted (1976) for an excellent survey of this topic, as well as for some original
results. More recent surveys are given by Draper and Van Nostrand (1979) and Judge
and Bock (1983).
6. The matrix Hair 'HI is sometimes referred to as the Moore-Penrose generalized
inverse of X'X and is denoted by ( VX)+. See Rao (1973, p. 24) for more discussion of
generalized inverses.
7. This question was originally posed and solved by Silvey (1969). .
8. Although Sclove considers the random coefficients model and the prediction
(rather than estimation) of the regression vector, there is no essential difference be-
tween his approach and the Bayesian estimation.
9. What follows simplifies the derivation of Sclove et al. (1972).
10. See Section 4.6.1. In Chapter 3 we shall discuss large sample theory and make
the meaning of the term asymptotically more precise. For the time being, the reader
should simply interpret it to mean "approximately when T is large."
11. See the definition of probability limit in Chapter 3. Loosely spealdng, the state-
ment means that when T is large, s is close to so with probability close to 1.
3. Large Sample Theory

1. Representative textbooks are, in a roughly increasing order of difficulty, Hoel
(1971); Freund (1971); Mood, Graybill, and Bees (1974); Cox and Hinkley (1974), and
Bickel and Doksum (1977).
2. For a more complete study of the subject, the reader should consult Chung (1974)
or Loeve (1977), the latter being more advanced than the former. Rao (1973), which is
Notes to Pages 81 -107 467
an excellent advanced textbook in mathematical statistics, also gives a concise review

of the theory of probability and random variables.
3. If a set function satisfies only (i) and (iii) of Definition 3.1.2, it is called a
measure. In this case the triplet defined in Definition 3.1.3 is called a measure space.
Thus the theory of probability is a special case of measure theory. (A standard textbook
for measure theory is Halmos, 1950.)
4. If we have a measure space, then a function that satisfies the condition of Defini-
tion 3.1.4 is called a measurable function. Thus a random variable is a function
measurable with respect to the probability measure.
5. Lebesgue measure can be defined also for certain non-Borel sets. The Lebesgue
measure defined only for Borel sets is called Bore! measure.
6. For Yto be a random variable, h must satisfy the condition (col h[X(co)] y) E .A
for every y. Such a function is said to be Bore!-measurable. A continuous function
except for a countable number of discontinuities is Borel-measurable.
7. The following is a simple example in which the Riemann-Stieltjes integral does
not exist. Suppose F(x) = 0 for a x c and F(x) = 1 for c < x b and suppose
h(x)= F(x). Then, depending on the point x* we choose in an interval that contains c,
S is equal to either 1 or 0. This is a weakness of the Riemann-Stieltjes integral. In this
example the Lebesgue-Stieltjes integral exists and is equal to 0. However, we will
not go into this matter further. The reader may consult references cited in Note 2 to
Chapter 3.
8. Sometimes we also say X converges to X almost everywhere or with probability
one. Convergence in probability and convergence almost surely are sometimes re-
ferred to as weak convergence and strong convergence, respectively.
9. Between M and a.s., we cannot establish a definite logical relationship without
further assumptions.
10. The law of large numbers implied by Theorem 3.2.1 (Chebyshev) can be slightly
generalized so as to do away with the requirement of a finite variance. Let {X,} be
independent and suppose EIX,I ' + 6 < M for some 6> 0 and some M < OD. Then X
ET I* 0. This is called Markov's law of large numbers.
11. The principal logarithm of a complex number rem is defined as log r + i0.
12. It seems that for most practical purposes the weak consistency of an estimator is
all that a researcher would need, and it is not certain how much more practical benefit
would result from proving strong consistency in addition.
13. Lai, Robbins, and Wei (1978) proved that if (u,) are assumed to be independent
in Model 1 and if the conditions of Theorem 3.5.1 are met, the least squares estimator
is strongly consistent. Furthermore, the homoscedasticity assumption can be relaxed,
provided the variances of {u 1) are uniformly bounded from above.
4. Asymptotic Properties of Extremum Estimators
1. In the proof of Theorem 4.1.1, continuity of Q T(0) is used only to imply continu-
ity of Q(0) and to make certain the measurability of O. Therefore we can modify this
theorem in such a way that we assume continuity of Q(0) but not of Q T() and define
convergence in probability in a more general way that does not require measurability
of 'or . This is done in Theorem 9.6.1.
Also note that the proof of Theorem 4.1.1 can easily be modified to show that if the
convergence in the sense of (i) holds in assumption C, 07-converges to 00 almost surely.
2. Strictly speaking, (4.1.10) is defined only for those T's for which Otis nonempty.
However, the probability that OT is nonempty approaches 1 as T goes to infinity
because of Assumption C of Theorem 4.1.2. As an aside, Jennrich (1969) proved that
0* is a measurable function.
3. This proof is patterned after the proof of a similar theorem by Jennrich (1969).
4. Note that if z r is merely defined as a random variable with mean p(0) and
variance-covariance matrix 1(0), the minimization of the quadratic form does not
even yield a consistent estimator of 0.
5. The term second-order efficiency is sometimes used to denote a related but differ-
ent concept (see Pfanzagl, 1973).
6. If the error term is multiplicative as in Q = 13 1 KP2Lheu, the log transformation
reduces it to a linear regression model. See Bodkin and Klein (1967) for the estimation
of both models.
7. The methods of proof used in this section and Section 4.3.3 are similar to those of
Jennrich (1969).
8. Because a function continuous on a compact set is uniformly continuous, we can
assume without loss of generality that f,() is uniformly continuous in p E N.
9. The nonsingularity is not needed here but is assumed as it will be needed later.
10. Note that in the special case where the constraint h(/)) = 0 is linear and can be
written as Q1 = c, (4.5.21) is similar to (4.3.32). We cannot unequivocally determine
whether the chi-square approximation of the distribution of (4.5.21) is better or worse
than the F approximation of the distribution of (4.3.32).
11. If the sample is (1, 2, 3), 2 is the unique median. If the sample is (1, 2, 3, 4), any
point in the closed interval [2, 3] may be defined as a median. The definition (4.6.3)
picks 2 as the median. If f(x) > 0 in the neighborhood of x == M, this ambiguity
vanishes as the sample size approaches infinity.
12. The second term of the right-hand side of (4.6.10) does not affect the minimiza-
tion but is added so that plim T 'ST can be evaluated without assuming the existence
of the first absolute moment of Y,. This idea originates in Huber (1965).
13. Alternative methods of proof of asymptotic normality can be found in Bassett
and Koenker (1978) and in Amemiya (1982a).
5. Time Series Analysis

1. We are using the approximation sign to mean that most elements of the matri-
ces of both sides are equal.
2. The subscript p denotes the order of the autoregression. The size of the matrix
should be inferred from the context.
Notes to Pages 179-229 469
3. Almon (1965) suggested an alternative method of computing') based on Lagran-

gian interpolation polynomials. Cooper (1972) stated that although the method dis-
cussed in the text is easier to understand, Almon's method is computationally su-
perior.
6. Generalized Least Squares Theory
1. If rank (MX) = K, ft is uniquely determined by (6.1.2).

2. Farebrother (1980) presented the relevant tables for the case in which there is no
intercept.
3. Breusch (1978) and Godfrey (1978) showed that Durbin's test is identical to
Rao's score test (see Section 4.5.1). See also Breusch and Pagan (1980) for the Lagrange
multiplier test closely related to Durbin's test.
4. In the special case in which N = 2 and /I changes with i (so that we must estimate
both fi l and A), statistic (6.5.9) is a simple transformation of the F statistic (1.5.44). In
this case (1.5.44) is preferred because the distribution given there is exact.
5. It is not essential to use the unbiased estimators al here. If al are used, the
distribution of is only trivially modified.
6. In either FGLS or MLE we can replace A by ft1 without affecting the asymptotic
distribution.
7. Hildreth and Houck suggested another estimator (Z'MZ) -1Z'62, which is the
instrumental variables estimator applied to (6.5.23) using Z as the instrumental vari-
ables. Using inequality (6.5.29), we can show that this estimator is also asymptotically
less efficient than a3. See Hsiao (1975) for an interesting derivation of the two estima-
tors of Hildreth and Houck from the MINQUE principle of Rao (1970). Froehlich
(1973) suggested FGLS applied to (6.5.23). By the same inequality, the estimator can
be shown to be asymptotically less efficient than a 3 . Froehlich reported a Monte Carlo
study that compared the FGLS subject to the nonnegativity of the variances with
several other estimators. A further Monte Carlo study has been reported by Dent and
Hildreth (1977).
8. The symbol 1 here denotes an NT-vector of ones. The subscripts will be omitted
whenever the size of 1 is obvious from the context. The same is true for the identity
matrix.
9. For the consistency of liQo we need both N and T to go to co. For further discus-
sion, see Anderson and Hsiao (1981, 1982).
10. The presence of the vector f in a variance term would cause a problem in the
derivation of the asymptotic results if T were allowed to go to Op, but in the Lillard-
Weiss model, as in most of econometric panel-data studies, T is small and N is large.
7. Linear Simultaneous Equations Models

1. More general constraints on the elements of r and B have been discussed by
Fisher (1966) and Hsiao (1983).
2. For alternative definitions of identification, see Hsiao (1983).

3. Identification of the structural parameters may be possible because of constraints
on the variance-covariance matrix of the errors (see Fisher, 1966, or Hsiao, 1983).
4. For an alternative derivation, see Koopmans and Hood (1953).
5. Basmann (1957) independently proposed the same estimator and called it the
generalized classical linear estimator.
6. Exact distributions and densities are expressed as twofold or threefold infinite
series and therefore are difficult to work with. For this reason approximations based on
various asymptotic expansions have been proposed. They are mainly classified into
two types: one based on a Taylor expansion of the logarithm of the characteristic
function, as in Anderson and Sawa (1973), and one that is a direct expansion of the
probability, as in Anderson (1974).
7. An instrumental variables estimator satisfying conditions (i) and (ii) but not (iii)
is consistent and could be asymptotically more efficient than 2SLS.
8. Nonlinear Simultaneous Equations Models
1. A slightly more general model defined by fi(y Y Xi ao) = u, will be consid-

ered in Section 8.1.2.
2. Tsurumi (1970) estimated a CES production function by first linearizing the
function around certain initial estimates of the parameters and then proceeding as if he
had the model nonlinear only in variables. Thus his method was in effect a Gauss-
Newton iteration for obtaining NL2S. However, Tsurumi did not discuss the statistical
properties of the estimator he used. Applications of models nonlinear only in variables
can be found in Strickland and Weiss (1976) and Rice and Smith (1977).
3. The subscript 0 indicating the true value is henceforth suppressed to simplify the
notation.
4. Goldfeld and Quandt (1968) considered a two-equation model nonlinear only in
variables that yields two solutions of dependent variables for a given vector value of the
error terms. For this model they conducted a Monte Carlo study of how the perform-
ance of various estimators is affected by different mechanisms of choosing a unique
solution.
5. This lemma, stated slightly differently, is originally attributed to Stein (1973).
Amemiya (1977a) independently rediscovered the lemma. The proof can be found in
an article by Amemiya (1982b).
9. Qualitative Response Models

1. If lima_o [g(x + A) gdx)]/A = dglcbc uniformly in n, g)/cbc =
dgId.x. See, for example, Apostol (1974, p. 221) for a proof.
2. The asymptotic efficiency of Bericson's MIN x2 estimator can also be proved as a
corollary of the Baranldn and Gurland theorem quoted in Section 4.2.4. See Taylor
(1953).
Notes to Pages 278-332 471
3. There is a slight error in the formula for the mean squared error of MLE, which
has been corrected by Amemiya (1984b).
4. See Anderson (1958, Ch. 6) for a thorough discussion of discriminant analysis.
5. Certain areas of statistics are more amenable to Bayesian analysis than to classical
analysis. The simultaneous determination of prediction and estimation is one of them.
6. To simplify the notation we shall omit the subscript 0 that denotes the true value.
The reader should be able to understand from the context whether a symbol denotes
the true value of a parameter or the domain of the parameter space.
7. The possible inconsistency due to this approximation was briefly discussed at the
end of Section 9.2.5.
8. Deacon and Shapiro actually used (9.3.32) and the equation obtained by sum-
ming (9.3.32) and (9.3.33). The resulting estimates off), and fl2 are the same as those
obtained by the method described in the text.
9. The triple integral in (9.3.72) can be reduced to a double integral by a certain
transformation. In general, we must evaluate m-tuple integrals for m + 1 responses.
10. We assume that a component of the vector xis either discrete or continuous, so
that f is actually the density and the probability combined (density with respect to
some measure). Thus the integration with respect to x that will appear later should be
regarded as the integral and summation combined (the Lebesgue-Stieltjes integral with
respect to the appropriate measure).
11. Because g(xi) is known, it can be ignored in the maximization of L e . However,
we retained it to remind the reader of the sampling scheme. In (9.5.2), H(j i) is retained
for the same reason.
12. The main difference between the two proofs is that we define as a solution of
(9.5.6) whereas Mansld and Lerman define it as the value of fi that attains the global
maximum of (9.5.5) over a compact parameter space containing the true value.
13. The following analysis can be modified to allow for the possibility that for j and
x, P(jlx, fi) =0 for allfl. Such a case arises when certain alternatives are unavailable for
a certain individual. See Mansld and McFadden (1983, footnote 23, p. 13).
14. Manski and Lerman quote an interesting result attributed to McFadden, which
states that in a multinomial logit model with alternative-specific interceptsthat is,
the model in which a in (9.3.43) varies with j as in (9.2.4) and (9.2.5), the inconsistency
is confined to the parameters (a/).
15. In the simple logit model considered after (9.5.22), it can be shown that the
asymptotic variances of MME and WMLE are identical, so that Table 9.7 applies to
MME as well.
16. Cosslett's sampling scheme can be generalized further to yield a general strati-
fied sampling (see Manski and McFadden, 1983, p. 28).
17. This interpretation does not contradict the fact that in actual decisions the
determination of x precedes that of j.
18. There is a subtle difference. In the simple choice-based sampling defined earlier,
a person choosing alternative j is sampled with probability H(j), whereas in the gener-
alized choice-based sampling the number of people in the subsample s is fixed and not
random.
19. It is interesting to note that maximizing n with respect to fl and A, is equivalent
to maximizing W defined in (9.5.39) with respect to and Qo(j). Mansld and McFad-
den (1983, p. 24) suggested this version of MME without realizing its equivalence to
CBMLE.
20. Cosslett defined WMLE and MME using the actual proportions of people
choosing ] instead of H(]).
21. Manslci's proof is not complete because in the fourth line from the bottom on
page 218 of his article, double limit operations are interchanged without verifying the
necessary conditions. It seems that we would have to make more assumptions than
made by Manski in order for the necessary conditions to hold. A correct proof for the
binary case can be found in Manski (1985).
22. Manslci (1975) considered a more general score function than that defined here.
23. The heterogeneity problem is also known as the mover-stayer problem in the
literature of Markov chain models. Among the first to discuss the problem were
Blumen, Kogan, and McCarthy (1955), who found that individuals who changed
occupations most frequently in the past were more likely to change in the future.
24. The (u1) in (9.7.5) are not i.i.d., unlike the (u,) in (9.7.2).
10. Tobit Models

1. More precisely, A means in this particular case that both sides of the equation
multiplied by VT: have the same limit distribution.
2. A( ) is known as the hazard rate and its reciprocal is known as Mills's ratio. Tobin
(1958) has given a figure that shows that A(z) can be closely approximated by a linear
function of z for 1 < z < 5. Johnson and Kotz (1970a, p. 278f.) gave various expan-
sions of Mills' ratio.
3. Chung and Goldberger (1984) generalized the results of Goldberger (1981) and
Greene (1981) to the case where y* and it are not necessarily jointly normal but ail Y *)
is linear in y*.
4. This was suggested by Wales and Woodland (1980).
5. See Stapleton and Young (1984).
6. See Section 4.3.5. Hartley (1976b) proved the asymptotic normality of i N and
iNw and that they are asymptotically not as efficient as the MLE.
7. The asymptotic equivalence of iN and was proved by Stapleton and Young
(1984).
8. Amemiya (1973c) showed that the Tobit likelihood function is not globally
concave with respect to the original parameters ft and (72.
9. For an alternative account, see Hartley (1976c).
10. An inequality constraint like this is often necessary in simultaneous equations
models involving binary or truncated variables. For an interesting unified approach to
this problem, see Gourieroux, Laffont, and Monfort (1980).
Notes to Pages 385 429 473
11. See Cragg (1971) for models that ensure the nonnegativity of y 2 as well as of y1 .
12. For a more elaborate derivation of the reservation wage model based on search
theory, see Gronau (1974).
13. Gronau specified that the independent variables in the Wr equation include a
woman's age and education, family income, number of children, and her husband's
age and education, whereas the independent variables in the W equation include only
a woman's age and education. However, Gronau readily admitted to the arbitrariness
of the specification and the possibility that all the variables are included in both.
14. Gronau assumed the independence between u 2 and v and used an estimator
different from those mentioned here. Amemiya (1984a) pointed out an error in
Gronau's procedure. The independence between u 2 and v is unnecessary if we use
either the MLE or Heckman's two-step estimator.
15. Although Heckman's model (1974) is a simultaneous equations model, the
two-step estimator of Heckman studied by Wales and Woodland is essentially a re-
duced-form estimator, which we have discussed in this subsection, rather than the
structural equations version we shall discuss in the next subsection.
16. For a panel data generalization of Heckman's model, see Heckman and Ma-
Curdy (1980).
17. Actually, Heckman used log Wr and log W. The independent variables x 2
includehsba'wg,tincomeprsadviulchrtesandz
includes housewife's schooling and experience.
18. A more explicit expression for the likelihood function was obtained by Ame-
miya (1974a), who pointed out the incorrectness of the likelihood function originally
given by Fair and Jaffee.
19. Equation (10.10.26) is the maximized profit function and (10.10.27) is an input
demand or output supply function obtained by differentiating (10.10.26) with respect
to the own input or output price (Hotelling's lemma). For convenience only one input
or output has been assumed; so, strictly speaking, 4) and xa) are scalars.
20. These two equations correspond to the two equations in a proposition of Dun-
can (1980, p. 851). It seems that Duncan inadvertently omitted the last term from
(10.10.30).
11. Markov Chain and Duration Models
1. Note that this result can be obtained from (11.1.38) using the Boskin-Nold
assumptions as well as the assumption that T goes to co. The result (11.1.38) is valid
even if T does not go to co provided that NT goes to C. If we assumed p(0) = p(00), the
approximate equality in (11.1.55) would be exact and hence (11.1.59) would be valid
without assuming T 00 provided that NT I .
2. Actually, Toiklca is interested only in three out of the six transition probabilities,
and he lets those three depend linearly on exogenous variables. However, for the
simplicity of analysis, we shall proceed in our discussion as if all the six transition
probabilities depended linearly on exogenous variables.
3. Note that minimizing (11.1.78) is equivalent to maximizing the pseudo likeli-

hood function under the assumption that r, is normally distributed.
4. The likelihood function (11.2.14) pertains only to unemployment spells. When
we also observe employment spells, (11.2.14) should be multiplied by a similar expres-
sion for employment spells. When there is no common unknown parameter in both
parts of the likelihood function, each part can be separately maximized.
5. See Gradshteyn and Ryzhik (1965, p. 576).
6. See Bellman (1970, Chapter 10).
7. Here we have simplified Tuma's model slightly to bring out the essential points.
8. To integrate out v, we must specify its distribution as Lancaster did. A wrong
distribution would introduce a specification bias. Heckman and Singer (1982, 1984a)
discussed an interesting procedure of treating the distribution of v as unknown and
maximizing the likelihood function with respect to that distribution as well as the other
unknown parameters of the model.
9. See ICalbfleisch and Prentice (1980) for a procedure for modifying the following
analysis in the case of ties.
10. Let L(y, 0) be a joint density function (or a likelihood function if it is considered
as a function of 9) of a vector of random variables y. If y is transformed into (v, w) by a
transformation not dependent on 0, the conditional density L(v1w, 0) is called a condi-
tional likelihood function.
References
Abramowitz, M., and!. A. Segun. 1965. Handbook of Mathematical Functions. New

York: Dover Publishing.
Adams, J. D. 1980. "Personal Wealth Transfers." Quarterly Journal of Economics
95:159-179.
Adelman, I. G. 1958. "A Stochastic Analysis of the Size Distribution of Finns."
Journal of the American Statistical Association 53:893-904.
Aigner, D. J., T. Amemiya, and D. J. Poirier. 1976. "On the Estimation of Production
Frontiers: Maximum Likelihood Estimation of the Parameters of a Discontin-
uous Density Function." International Economic Review 17:377 396.
Aigner, D. J., and G. G. Judge. 1977. "Application of Pre-Test and Stein Estimators to
Economic Data." Econometrica 45:1279 1288.
Aitchison, J., and J. Bennett. 1970. "Polychotomous Quantal Response by Maximum
Indicant." Biometrilca 57:253-262.
Akahira, M. 1983. "Asymptotic Deficiency of the Jackknife Estimator." The Austra-
lian Journal of Statistics 25:123-129.
Akaike, H. 1973. "Information Theory and an Extension of the Maximum Likelihood
Principle," in B. N. Petrov and F. Csald, eds., Second International Symposium
on Information Theory, pp. 267-281. Budapest: Akademiai Kiado.
1976. "On Entropy Maximization Principle." Paper presented at the Sympo-
sium on Applications of Statistics, Dayton, Ohio.
Albert, A., and J. A. Anderson. 1984. "On the Existence of Maximum Likelihood
Estimates in Logistic Regression Models." Biometrilca 71:1-10.
Albright, R. L., S. R. Lerman, and C. F. Mansld. 1977. "Report on the Development of
an Estimation Program for the Multinomial Probit Model." Mimeographed
paper prepared for the Federal Highway Administration.
Almon, S. 1965. "The Distributed Lag between Capital Appropriations and Expendi-
tures." Econometrica 33:178 196.
Amemiya, T. 1966. "On the Use of Principal Components of Independent Variables
in Two-Stage Least-Squares Estimation." International Economic Review 7:283
303.
1967. "A Note on the Estimation of Balestra-Nerlove Models." Technical
Report no. 4, Institute for Mathematical Studies in the Social Sciences, Stanford
University, Calif.
476 References
1971. "The Estimation of the Variances in a Variance-Components Model."

International Economic Review 12:1 - 13.
1972. "Bivariate Probit Analysis: Minimum Chi-Square Methods." Journal of
the American Statistical Association 69:940-944.
1973a. "Generalized Least Squares with an Estimated Autocovariance Ma-
trix." Econometrica 41:723 - 732.
1973b. "Regression Analysis When the Variance of the Dependent Variable Is
Proportional to the Square of Its Expectation." Journal of the American Statistical
Association 68:928 -934.
1973c. "Regression Analysis When the Dependent Variable Is Truncated
Normal." Econometrica 41:997 - 1016.
1974a. "A Note on a Fair and Jaffee Model." Econometrica 42:759-762.
1974b. "Multivariate Regression and Simultaneous Equation Models When
the Dependent Variables Are Truncated Normal." Econometrica 42:999-1012.
1974c. "The Nonlinear Two-Stage Least-Squares Estimator." Journal ofEcon-
ometrics 2:105-110.
1975a. "The Nonlinear Limited-Information Maximum-Likelihood Estima-
tor and the Modified Nonlinear Two-Stage Least-Squares Estimator." Journal of
Econometrics 3:375-386.
1975b. "Qualitative Response Models." Annals of Economic and Social Mea-
surement 4:363-372.
1976a. "Estimation in Nonlinear Simultaneous Equation Models." Paper
presented at and published by Institut National de la Statistique et des Etudes
Economiques, Paris. Published in French: E. Malinvaud, ed., Cahiers du semin-
aire d'econometrie, no. 19, 1978.
1976b. "The Maximum Likelihood, the Minimum Chi-Square and the Non-
linear Weighted Least-Squares in the General Qualitative Response Model."
Journal of the American Statistical Association 71:347 - 351.
1977a. "The Maximum Likelihood and the Nonlinear Three-Stage Least
Squares Estimator in the General Nonlinear Simultaneous Equation Model."
Econometrica 45:955-968.
1977b. "A Note on a Heteroscedastic Model." Journal of Econometrics
6:365-370.
1977c. "The Modified Second-Round Estimator in the General Qualitative
Response Model." Journal of Econometrics. 5:295 - 299.
1978a. "Corrigenda: A Note on a Heteroscedastic Model." Journal of Econo-
metrics 8:265.
1978b. "A Note on a Random Coefficients Model." International Economic
Review 19:793-796.
1978c. "The Estimation of a Simultaneous Equation Generalized Probit
Model." Econometrica 46:1193-1205.
References 477
1979. "The Estimation of a Simultaneous-Equation Tobit Model." Interna-

tional Economic Review 20:169 - 181.
I 980a. "Selection of Regressors." International Economic Review 21:331 -
345.
1980b. "The rr 2-order Mean Squared Errors of the Maximum Likelihood and
the Minimum Logit Chi-Square Estimator." Annals of Statistics 8:488-505.
1981. "Qualitative Response Models: A Survey." Journal of Economic Litera-
ture 19:1483-1536.
1982a. "Two Stage Least Absolute Deviations Estimators." Econometrica
50:689-711.
1982b. "Correction to a Lemma." Econometrica 50:1325- 1328.
1983a. "Non-Linear Regression Models," in Z. Griliches and M. D. Intrilliga-
tor, eds., Handbook of Econometrics, 1:333-389. Amsterdam: North-Holland
Publishing.
1983b. "A Comparison of the Amemiya GLS and the Lee-Maddala-Trost
G2SLS in a Simultaneous-Equations Tobit Model." Journal of Econometrics
23:295-300.
1983c. "Partially Generalized Least Squares and Two-Stage Least Squares
Estimators." Journal of Econometrics. 23:275-283.
1984a. "Tobit Models: A Survey." Journal of Econometrics 24:3-61.
1984b. "Correction." Annals of Statistics 12:783.
Amemiya, T., and M. Boslcin. 1974. "Regression Analysis When the Dependent Vari-
able Is Truncated Lognormal, with an Application to the Determinants of the
Duration of Welfare Dependency." International Economic Review 15:485-496.
Amemiya, T., and W. A. Fuller. 1967. "A Comparative Study of Alternative Estima-
tors in a Distributed-Lag Model." Econometrica 35:509-529.
Amemiya, T., and T. E. MaCurdy. 1983. "Instrumental Variable Estimation of an
Error Components Model." Technical Report no. 414, Institute for Mathemati-
cal Studies in the Social Sciences, Stanford University, Calif. (Forthcoming in
Econometrica.)
Amemiya, T., and K. Morimune. 1974. "Selecting the Optimal Order of Polynomial
in the Almon Distributed Lag." Review of Economics and Statistics 56:378-386.
Amemiya, T., and J. L. Powell. 1981. "A Comparison of the Box-Cox Maximum
Likelihood Estimator and the Non-Linear Two-Stage Least Squares Estimator."
Journal of Econometrics 17:351 -381.
1983. "A Comparison of the Logit Model and Normal Discriminant Analysis
When the Independent Variables Are Binary," in S. Karlin, T. Amemiya, and
L. A. Goodman, eds., Studies in Econometric, Time Series, and Multivariate
Statistics, pp.3- 30. New York: Academic Press.
Anderson, T. W. 1958. Introduction to Multivariate Statistical Analysis. New York:
John Wiley & Sons.
478 References
1969. "Statistical Inference for Covariance Matrices with Linear Structure," in

P. R. Krishnaiah, ed., Proceedings of the Second International Symposium on
Multivariate Analysis, pp. 55 66. New York: Academic Press.
-
1971. The Statistical Analysis of Time Series. New York: John Wiley and
Sons.
1974."An Asymptotic Expansion of the Distribution of the Limited Informa-
tion Maximum Likelihood Estimate of a Coefficient in a Simultaneous Equation
System." Journal of the American Statistical Association 60:565 573.
-
1982. "Some Recent Developments of the Distributions of Single-Equation

Estimators," in W. Hildenbrand, ed., Advances in Econometrics pp. 109-122.
Cambridge: Cambridge Univerity Press.
Anderson, T. W., and L. A. Goodman. 1957. "Statistical Inference about Markov
Chains." Annals of Mathematical Statistics 28:89 110.
-
Anderson, T. W., and C. Hsiao. 1981. "Estimation of Dynamic Models with Error
Components." Journal of the American Statistical Association 76:598 606. -
1982. "Formulation and Estimation of Dynamic Models Using Panel Data."

Journal of Econometrics 18:47 82. -
Anderson, T. W., and H. Rubin. 1949. "Estimator of the Parameters of a Single

Equation in a Complete System of Stochastic Equations." Annals of Mathemati-
cal Statistics 20:46 63.
-
Anderson, T. W., and T. Sawa. 1973. "Distributions of Estimates of Coefficients of a

Single Equation in a Simultaneous System and Their Asymptotic Expansions."
Econometrica 41:683 714.-
Andrews, D. F. 1974. "A Robust Method for Multiple Linear Regression." Techno-
metrics 16:523-531.
Andrews, D. F., P. J. Bickel, F. R. Hampel, P. J. Huber, W. H. Rogers, and J. W.
Tukey. 1972. Robust Estimates of Location. Princeton: Princeton University
Press.
Anscombe, F. J. (1961). "Examination of Residuals," in J. Neyman, ed., Proceedings
of the Fourth Berkeley Symposium on Mathematical Statistics and Probability,
1:1 36. Berkeley: University of California Press.
-
Apostol, T. M. 1974. Mathematical Analysis, 2nd ed. Reading, Mass.: Addison-

Wesley.
Arabmazar, A., and P. Schmidt. 1981. "Further Evidence on the Robustness of the
Tobit Estimator to Heteroscedasticity." Journal of Econometrics 17:253-258.
1982. "An Investigation of the Robustness of the Tobit Estimator to Non-Nor-
mality." Econometrica 50:1055-1063.
Arrow, K. J., H. B. Chenery, B. S. Minhas, and R. M. Solow. 1961. "Capital-Labor
Substitution and Economic Efficiency." Review of Economics and Statistics
43:225-250.
Ashenfelter, 0., and J. Ham. 1979. "Education, Unemployment, and Earnings."
Journal of Political Economy 87:S99 S116.
-
References 479
Ashford, J. R., and R. R. Sowden. 1970. "Multivariate Probit Analysis." Biometrics

26:535-546.
Baldwin, R. E. 1971. "Determinants of the Commodity Structure of U.S. Trade."
American Economic Review 61:126-146.
Balestra, P., and M. Nerlove. 1966. "Pooling Cross-Section and Time Series Data in
the Estimation of a Dynamic Model: The Demand for Natural Gas." Economet-
rica 34:585-612.
Baltagi, B. H. 1981. "Pooling: An Experimental Study of Alternative Testing and
Estimation Procedures in a Two-Way Error Component Model." Journal of
Econometrics 17:21-49.
Baranchik, A. J. 1970. "A Family of Minimax Estimators of the Mean of a Multivar-
iate Normal Distribution." Annals of Mathematical Statistics 41:642-645.
Barankin, E. W., and J. Gurland. 1951. "On Asymptotically Normal Efficient Estima-
tors: I." University of California Publications in Statistics 1:86-130.
Bartholomew, D. J. 1982. Stochastic Models for Social Processes, 3rd ed. New York:
John Wiley & Sons.
Basmann, R. L. 1957. "A Generalized Classical Method of Linear Estimation of
Coefficients in a Structural Equation." Econometrica 25:77-83.
Bassett, G., Jr., and R. Koenker. 1978. "Asymptotic Theory of Least Absolute Error
Regression." Journal of the American Statistical Association 73:618-622.
Beach, C. M., and J. G. MacKinnon. 1978. "A Maximum Likelihood Procedure for
Regression with Auto-correlated Errors," Econometrica 46:51-58.
Bellman, R. 1970. Introduction to Matrix Analysis, 2d ed. New York: McGraw-Hill.
Belsley, D. A. 1979. "On the Computational Competitiveness of Full-Information
Maximum-Likelihood and Three-Stage Least-Squares in the Estimation of Non-
linear Simultaneous-Equations Models." Journal of Econometrics 9:315 -342.
Benus, J., J. Kmenta, and H. Shapiro. 1976. "The Dynamics of Household Budget
Allocation to Food Expenditures." Review of Economics and Statistics 58:129 -
138.
Bera, A. K., C. M. Jarque, and L. F. Lee. 1982. "Testing for the Normality Assumption
in Limited Dependent Variable Models." Mimeographed paper, Department of
Economics, University of Minnesota.
Berger, J. 0. 1975. "Minimax Estimation of Location Vectors for a Wide Class of
Densities." Annals of Statistics 3:1318-1328.
1976. "Admissible Minimax Estimation of a Multivariate Normal Mean with
Arbitrary Quadratic Loss." Annals of Statistics 4:223-226.
Berkson, J. 1944. "Application of the Logistic Function to Bio-Assay." Journal of the
American Statistical Association 39:357-365.
1955. "Maximum Likelihood and Minimum x2 Estimates of the Logistic
Function." Journal of the American Statistical Association 50:130 - 162.
1957. "Tables for Use in Estimating the Normal Distribution Function by
Normit Analysis." Biometrilca 44:411-435.
480 References
1980. "Minimum Chi-Square, Not Maximum Likelihood!" Annals of Statis-

tics 8:457-487.
Berndt, E. R., B. H. Hall, R. E. Hall, and J. A. Hausman. 1974. "Estimation and
Inference in Nonlinear Structural Models." Annals of Economic and Social Mea-
surement 3:653-666.
Berndt, E. R., and N. E. Savin. 1977. "Conflict among Criteria for Testing Hypotheses
in the Multivariate Linear Regression Model." Econometrica 45:1263 - 1278.
Berzeg, K. 1979. "The Error Components Model: Conditions for the Existence of the
Maximum Likelihood Estimates." Journal of Econometrics 10:99-102.
Bhattacharya, P. K. 1966. "Estimating the Mean of a Multivariate Normal Population
with General Quadratic Loss Function." Annals of Mathematical Statistics
32:1819-1824.
Bhattacharya, R. N., and R. R. Rao. 1976. Normal Approximation and Asymptotic
Expansions. New York: John Wiley & Sons.
Bianchi, C., and G. Calzolari. 1980. "The One-Period Forecast Error in Nonlinear
Econometric Models." International Economic Review 21:201 - 208.
Bickel, P. J. 1975. "One-Step Huber Estimation in the Linear Model." Journal of the
American Statistical Association 70:428 -433.
1978. "Using Residuals Robustly, I: Tests for Heteroscedasticity, Nonlinearity
and Nonadditivity." Annals of Statistics 6:266 - 291.
Bickel, P. J., and K. A. Doksum. 1977. Mathematical Statistics: Basic Ideas and
Selected Topics. San Francisco: Holden-Day.
Bishop, Y. M. M., S. E. Fienberg, and P. W. Holland. 1975. Discrete Multivariate
Analysis, Theory and Practice. Cambridge, Mass.: MIT Press.
Blattberg, R., and T. Sargent. 1971. "Regression with Non-Gaussian Stable Distur-
bances: Some Sampling Results." Econometrica 39:501-510.
Blumen, I., M. Kogan, and P. J. McCarthy. 1955. The Industrial Mobility of Labor as
a Probability Process. Ithaca, N.Y.: Cornell University Press.
Bock, M. E. 1975. "Minimax Estimators of the Mean of a Multivariate Distribution."
Annals of Statistics 3:209 - 218.
Bodkin, R. G., and L. R. Klein. 1967. "Nonlinear Estimation of Aggregate Production
Functions." Review of Economics and Statistics 49:28-44.
Borjas, G. J., and S. Rosen. 1980. "Income Prospects and Job Mobility of Young
Men." Research in Labor Economics 3:159-181.
Bosldn, M. J., and F. C. Nold. 1975. "A Markov Model of Turnover in Aid to Families
with Dependent Children." Journal of Human Resources 10:476 - 481.
Box, G. E. P., and D. R. Cox. 1964. "An Analysis of Transformations." Journal ofthe
Royal Statistical Society ser. B, 26:211-252 (with discussion).
Box, G. E. P., and G. M. Jenkins. 1976. Time Series Analysis: Forecasting and Con-
trol, rev. ed. San Francisco: Holden-Day.
Box, G. E. P., and G. C. Tiao. 1973. Bayesian Inference in Statistical Analysis. Read-
ing, Mass.: Addison-Wesley.
References 481
Breusch, T. S. 1978. "Testing for Autocorrelation in Dynamic Linear Models." Aus-

tralian Economic Papers 17:334-335.
Breusch, T. S., and A. R. Pagan. 1979. "A Simple Test for Heteroscedasticity and
Random Coefficient Variation." Econometrica 47:1287-1294.
1980. "The Lagrange Multiplier Test and Its Applications to Model Specifica-
tion in Econometrics." Review of Economic Studies 47:239-253.
Brillinger, D. R. 1975. Time Series: Data Analysis and Theory. New York: Holt,
Rinehart, and Winston.
Brook, R. J. 1976. "On the Use of a Regret Function to Set Significance Points in Prior
Tests of Estimation." Journal of the American Statistical Association 71:126 -
131
Brown, B. W. 1983. "The Identification Problem in Systems Nonlinear in the Vari-
ables." Econometrica 51:175 -196.
Brown, B. W., and R. S. Mariano. 1982. "Residual-Based Stochastic Prediction in a
Nonlinear Simultaneous System." Analysis Center, The Wharton School, Uni-
versity of Pennsylvania.
Brown, M., and D. Heien. 1972. "The S-Branch Utility Tree: A Generalization of the
Linear Expenditure System." Econometrica 40:737-747.
Brown, P., and C. Payne. 1975. "Election Night Forecasting." Journal of the Royal
Statistical Society ser. A, 138:463-498 (with discussion).
Brown, R. L., J. Durbin, and J. M. Evans. 1975. "Techniques for Testing the Con-
stancy of Regression Relationships over Time." Journal of the Royal Statistical
Society sex. B, 37:149- 192 (with discussion).
Brown, W. G., and B. R. Beattie. 1975. "Improving Estimates of Economic Parame-
ters by Use of Ridge Regression with Production Function Applications." Ameri-
can Journal of Agricultural Economics 57:21-32.
Brundy, J. M., and D. W. Jorgenson. 1974. "Consistent and Efficient Estimation of
Systems of Simultaneous Equations by Means of Instrumental Variables," in P.
Zarembka, ed., Frontiers in Econometrics, pp. 215-244. New York: Academic
Press.
Buse, A. 1982. "Tests for Additive Heteroscedasticity: Some Monte Carlo Results."
Research Paper no. 82-13, Department of Economics, University of Alberta.
Butler, J. S., and R. Moffitt. 1982. "A Computationally Efficient Quadrature Proce-
dure for the One-Factor Multinomial Probit Model." Econometrica 50:761-764.
Carroll, R. J., and D. Ruppert. 1982a. "Robust Estimation in Heteroscedastic Linear
Models." Annals of Statistics 10:429-441.
1982b. "A Comparison between Maximum Likelihood and Generalized Least
Squares in a Heteroscedastic Linear Model." Journal of the American Statistical
Association 77:878-882.
Chamberlain, G. 1982. "Multivariate Regression Models for Panel Data." Journal of
Econometrics 18:5 - 46.
Chamberlain, G., and Z. Griliches. 1975. "Unobservables with a Variance-Compo-
482 References
nents Structure: Ability, Schooling and the Economic Success of Brothers." Inter-
national Economic Review 16:422-429.
Champernowne, D. G. 1953. "A Model of Income Distribution." Economic Journal
63:318-351.
Charatsis, E. G. 1971. "A Computer Program for Estimation of the Constant Elasticity
of Substitution Production Function." Applied Statistics 20:286-296.
Chow, G. C. 1960. "Tests for Equality between Sets of Coefficients in Two Linear
Regressions." Econometrica 28:591-605.
1968. "Two Methods of Computing Full-Information Maximum Likelihood
Estimates in Simultaneous Stochastic Equations." International Economic Re-
view 9:100-112.
1973. "On the Computation of Full-Information Maximum Likelihood Esti-
mates for Nonlinear Equation Systems." Review of Economics and Statistics
55:104-109.
Chow, G. C., and R. C. Fair. 1973. "Maximum Likelihood Estimation of Linear
Equation Systems with Auto-Regressive Residuals." Annals of Economic and
Social Measurement 2:17-28.
Christ, C. F. 1966. Econometric Models and Methods. New York: John Wiley & Sons.
Christensen, L. R., D. W. Jorgenson, and L. J. Lau. 1975. "Transcendental Logarith-
mic Utility Functions." American Economic Review 65:367-383.
Chung, C. F., and A. S. Goldberger. 1984. "Proportional Projections in Limited
Dependent Variable Models." Econometrica 52:531-534.
Chung, K. L. 1974. A Course in Probability Theory, 2d ed. New York: Academic Press.
Clark, C. 1961. "The Greatest of a Finite Set of Random Variables." Operations
Research 9:145-162.
Cochrane, D., and G. H. Orcutt. 1949. "Application of Least Squares Regression to
Relationships Containing Autocorrelated Error Terms." Journal of the American
Statistical Association 44:32-61.
Cooley, T. F., and E. C. Prescott. 1976. "Estimation in the Presence of Stochastic
Parameter Variation." Econometrica 44:167 - 184.
Cooper, J. P. 1972. "Two Approaches to Polynomial Distributed Lags Estimation: An
Expository Note and Comment." The American Statistician 26:32 - 35.
Cosslett, S. R. 1978. "Efficient Estimation of Discrete-Choice Models from Choice-
Based Samples." Workshop in Transportation Economics, University of Califor-
nia, Berkeley.
1981a. "Maximum Likelihood Estimator for Choice-Based Samples." Econo-
metrica 49:1289-1316.
1981b. 'Efficient Estimation of Discrete-Choice Models," in C. F. Mansld and
D. McFadden, eds., Structural Analysis of Discrete Data with Econometric Appli-
cations, pp. 51-111. Cambridge, Mass.: MIT Press.
1983. "Distribution-Free Maximum Likelihood Estimator of the Binary
Choice Model." Econometrica 51:765-782.
References 483
Cox, D. R. 1961. "Tests of Separate Families of Hypotheses," in J. Neyman, ed.,

Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and
Probability, 1:105-123. Berkeley: University of California Press.
1962. "Further Results on Tests of Separate Families of Hypotheses." Journal
of the Royal Statistical Society ser. B, 24:406-424.
1966. "Some Procedures Connected with the Logistic Qualitative Response
Curve," in F. N. David, ed., Research Papers in Statistics, pp. 55-71. New York:
John Wiley & Sons.
1970. Analysis of Binary Data. London: Methuen.
1972. "Regression Models and Life Tables." Journal of the Royal Statistical
Society ser. B, 34:187-220 (with discussion).
1975. "Partial Likelihood." Biometrika 62:269-276.
Cox, D. R., and D. V. Hinkley. 1974. Theoretical Statistics. London: Chapman and
Hall.
Cragg, J. G. 1971. "Some Statistical Models for Limited Dependent Variables with
Application to the Demand for Durable Goods." Econometrica 39:829-844.
Cramer, H. 1946. Mathematical Methods of Statistics. Princeton: Princeton Univer-
sity Press.
Crowder, M. J. 1980. "On the Asymptotic Properties of Least-Squares Estimators in
Autoregression." Annals of Statistics 8:132-146.
Dagenais, M. G. 1978. "The Computation of FIML Estimates as Iterative Generalized
Least Squares Estimates in Linear and Nonlinear Simultaneous Equations
Models." Econometrica 46:1351 1362.
David, J. M., and W. E. Legg. 1975. "An Application of Multivariate Probit Analysis
to the Demand for Housing: A Contribution to the Improvement of the Predictive
Performance of Demand Theory, Preliminary Results." American Statistical As-
sociation Proceedings of the Business and Economics and Statistics Section, pp.
295-300.
Davidson, W. C. 1959. "Variable Metric Method for Minimization." Atomic Energy
Commission, Research Development Report ANL-5990, Washington, D.C.
Davis, L. 1984. "Comments on a Paper by T. Amemiya on Estimation in a Dichoto-
mous Logit Regression Model." Annals of Statistics 12:778-782.
Deacon, R., and P. Shapiro. 1975. "Private Preference for Collective Goods Revealed
through Voting and Referenda." American Economic Review 65:943-955.
Dempster, A. P., N. M. Laird, and D. B. Rubin. 1977. "Maximum Likelihood from
Incomplete Data via the EM Algorithm." Journal of the Royal Statistical Society
ser. B, 39:1-38 (with discussion).
Dempster, A. P., M. Schatzoff, and N. Wermuth. 1977. "A Simulation Study of
Alternatives to Ordinary Least Squares." Journal of the American Statistical
Association 72:77 111 (with discussion).
Dent, W. T., and C. Hildreth. 1977. "Maximum Likelihood Estimation in Random
Coefficient Models." Journal of the American Statistical Association 72:69-72.
484 References
Dhrymes, P. J. 1971. Distributed Lags: Problems of Estimation and Formulation. San

Francisco: Holden-Day.
Diewert, W. E. 1974. "Applications of Duality Theory," in M. D. Intriffigator and
D. A. Kendrick, eds., Frontiers of Quantitative Economics, 2:106-171. Amster-
dam: North-Holland Publishing.
Domencich, T. A., and D. McFadden. 1975. Urban Travel Demand. Amsterdam:
North-Holland Publishing.
Doob, J. L. 1953. Stochastic Processes. New York: John Wiley & Sons.
Draper, N. R., and D. R. Cox. 1969. "On Distributions and Their Transformation to
Normality." Journal of the Royal Statistical Society ser. B, 31:472-476.
Draper, N. R., and H. Smith. 1981. Applied Regression Analysis, 2d ed. New York:
John Wiley & Sons.
Draper, N. R., and R. C. Van Nostrand. 1979. "Ridge Regression and James-Stein
Estimation: Review and Comments." Technometrics 21:451-466.
Dubin, J. A., and D. McFadden. 1984. "An Econometric Analysis of Residential
Electric Appliance Holdings and Consumption." Econometrica 52:345 -362.
Dudley, L., and C. Montmarquette. 1976. "A Model of the Supply of Bilateral Foreign
Aid." American Economic Review 66:132-142.
Duncan, G. M. 1980. "Formulation and Statistical Analysis of the Mixed, Continu-
ous/Discrete Dependent Variable Model in Classical Production Theory." Econ-
ometrica 48:839-852.
Duncan, G. T., and L. G. Lin. 1972. "Inference for Markov Chains Having Stochastic
Entry and Exit." Journal of the American Statistical Association 67:761-767.
Durbin, J. 1960. "Estimation of Parameters in Time-Series Regression Models." Jour-
nal of the Royal Statistical Society ser. B, 22:139 -153.
1963. "Maximum-Likelihood Estimation of the Parameters of a System of
Simultaneous Regression Equations." Paper presented at the Copenhagen Meet-
ing of the Econometric Society.
1970. "Testing for Serial Correlation in Least-Squares Regression When Some
of the Regressors Are Lagged Dependent." Econometrica 38:410-421.
Durbin, J., and G. S. Watson. 1950. "Testing for Serial Correlation in Least Squares
Regression, I." Biometrilca 37:409-428.
1951. "Testing for Serial Correlation in Least Squares Regression, II." Bio-
metrika 38:159-178.
1971. "Testing for Selig Correlation in Least Squares Regression, III." Bio-
metrika 58:1 - 19.
Efron, B. 1975. "The Efficiency of Logistic Regression Compared to Normal Discrimi-
nant Analysis." Journal of the American Statistical Association 70:892- 898.
1977. "The Efficiency of Cox's Likelihood Function for Censored Data."
1982. The Jackknife, the Bootstrap, and Other Resampling Plans. Philadel-
phia: Society for Industrial and Applied Mathematics.
References 485
Efron, B., and C. Morris. 1972. "Limiting the Risk of Bayes and Empirical Bayes
Estimators-Part II: The Empirical Bayes Case." Journal of the American Statis-
tical Association 67:130-139.
1973. "Stein's Estimation Rule and Its Competitors-An Empirical Bayes
Approach." Journal of the American Statistical Association 68:117-130.
1975. "Data Analysis Using Stein's Estimator and Its Generalizations." Jour-
nal of the American Statistical Association 70:311-319.
1976. "Families of Minimax Estimators of the Mean of a Multivariate Normal
Distribution." Annals of Statistics 4:11 -21.
Ehrlich, I. 1977. "Capital Punishment and Deterrence: Some Further Thoughts and
Additional Evidence." Journal of Political Economy 85:741 -788.
Eicker, F. 1963. "Asymptotic Normality and Consistency of the Least Squares Esti-
mators for Families of Linear Regressions." Annals of Mathematical Statistics
34:447-456.
Eisenpress, H., and J. Greenstadt. 1966. "The Estimation of Nonlinear Econometric
Systems." Econometrica 34:851-861.
Engle, R. F. 1984. "Wald, Likelihood Ratio, and Lagrange Multiplier Tests in Econo-
metrics," in Z. Griliches and M. D. Intrilligator, eds., Handbook of Econometrics,
2:775 - 826. Amsterdam: North-Holland Publishing.
Fair, R. C. 1974. "On the Robust Estimation of Econometric Models." Annals of
Economic and Social Measurement 3:667-677.
1976. A Model of Macroeconomic Activity. Vol. 2, The Empirical Model.
Cambridge, Mass.: Ballinger.
1978. "A Theory of Extramarital Affairs." Journal of Political Economy
86:45-61.
Fair, R. C., and D. M. Jaffee. 1972. "Methods of Estimation for Markets in Disequilib-
rium." Econometrica 40:497-514.
Fair, R. C., and W. R. Parke. 1980. "Full-Information Estimates of a Nonlinear
Macroeconometric Model." Journal of Econometrics 13:269-291.
Farebrother, R. W. 1975. "Minimax Regret Significance Points for a Preliminary Test
in Regression Analysis." Econometrica 43:1005-1006.
1980. "The Durbin-Watson Test for Serial Correlation When There Is No
Intercept in the Regression." Econometrica 48:1553 - 1563.
Feller, W. 1961. An Introduction to Probability Theory and Its Applications, vol. I, 2d
ed. New York: John Wiley & Sons.
Ferguson, T. S. 1958. "A Method of Generating Best Asymptotically Normal Esti-
mates with Application to the Estimation of Bacterial Densities." Annals ofMath-
ematical Statistics 29:1046-1062.
Fisher, F. M. 1966. The Identification Problem in Econometrics. New York: McGraw-
Hill.
1970. "A Correspondence Principle for Simultaneous Equation Models."
486 References
Fletcher, R., and M. J. D. Powell. 1963. "A Rapidly Convergent Descent Method for
Minimization." Computer Journal 6:163-168.
Flinn, C. J., and J. J. Heckman. 1982. "Models for the Analysis of Labor Force
Dynamics." Advances in Econometrics 1:35-95.
Fomby, T. B., R. C. Hill, and S. R. Johnson. 1978. "An Optimal Property of Principal
Components in the Context of Restricted Least Squares." Journal ofthe American
Forsythe, A. B. 1972. "Robust Estimation of Straight Line Regression Coefficients by
Minimizing p-th Power Deviations." Technometrics 14:159- 166.
Freund, J. F. 1971. Mathematical Statistics, 2d ed. Englewood Cliffs, N.J.: Prentice-
Hall.
Froehlich, B. R. 1973. "Some Estimates fora Random Coefficient Regression Model."
Journal of the American Statistical Association. 68:329-335.
Fuller, W. A. 1976. Introduction to Statistical Time Series. New York: John Wiley &
Sons.
1977. "Some Properties of a Modification of the Limited Information Estima-
tor." Econometrica 45:939-953.
Fuller, W. A., and G. E. Battese. 1974. "Estimation of Linear Models with Crossed-
Error Structure." Journal of Econometrics 2:67-78.
Gallant, A. R. 1975a. "Nonlinear Regression." The American Statistician 29:73-81.
1975b. "Testing a Subset of the Parameters ofa Nonlinear Regression Model."
1977. "Three-Stage Least-Squares Estimation for a System of Simultaneous,
Nonlinear, Implicit Equations." Journal of Econometrics 5:71-88.
Gallant, A. R., and A. Holly. 1980. "Statistical Inference in an Implicit, Nonlinear,
Simultaneous Equation Model in the Context of Maximum Likelihood Estima-
tion." Econometrica 48:697-720.
Gallant, A. R., and D. W. Jorgenson. 1979. "Statistical Inference for a System of
Simultaneous, Nonlinear, Implicit Equations in the Context of Instrumental
Variable Estimation." Journal of Econometrics 11:275-302.
Gastwirth, J. L. 1966. "On Robust Procedures." Journal of the American Statistical
Gayer, K. M., and M. S. Geisel. 1974. "Discriminating Among Alternative Models:
Bayesian and Non-Bayesian Methods," in P. Zarembka, ed., Frontiers in Econo-
metrics. pp. 49-80. New York: Academic Press.
Ghosh, J. K., and B. K. Sinha. 1981. "A Necessary and Sufficient Condition for
Second Order Admissibility with Applications to Berlcson's Bioassay Problem."
Annals of Statistics 9:1334-1338.
Ghosh, J. K., and K. Subramanyam. 1974. "Second-Order Efficiency of Maximum
Likelihood Estimators." Sankhya ser. A, 36:325-358.
Gnedenko, B. V., and A. N. Kolmogorov. 1954. Limit Distributions for Sums of
Independent Random Variables. Reading, Mass.: Addison-Wesley.
References 487
Godfrey, L. G. 1978. "Testing Against General Autoregressive and Moving Average

Error Models When the Regressors Include Lagged Dependent Variables." Econ-
ometrica 46:1293-1302.
Goldberg, S. 1958. Introduction to Difference Equations. New York: John Wiley &
Sons.
Goldberger, A. S. 1964. Econometric Theory. New York: John Wiley & Sons.
1981. "Linear Regression After Selection." Journal of Econometrics 15:357 -
366.
1983. "Abnormal Selection Bias," in S. Karlin, T. Amemiya, and L. A. Good-
man, eds., Studies in Econometrics, Time Series, and Multivariate Statistics, pp.
67-84. New York: Academic Press.
Goldfeld, S. M., and R. E. Quandt. 1965. "Some Tests for Homoscedasticity." Journal
of the American Statistical Association 60:539-547.
1968. "Nonlinear Simultaneous Equations: Estimation and Prediction." In-
ternational Economic Review 9:113-136.
1972. Nonlinear Methods in Econometrics. Amsterdam: North-Holland Pub-
lishing.
1978. "Asymptotic Tests for the Constancy of Regressions in the Heterosce-
dastic Case." Research Memorandum no. 229, Econometric Research Program,
Princeton University.
Goldfeld, S. M., R. E. Quandt, and H. F. Trotter. 1966. "Maximization by Quadratic
Hill-Climbing." Econometrica 34:541-551.
Goodman, L. A. 1961. "Statistical Methods for the 'Mover-Stayer' Model." Journal of
the American Statistical Association 56:841-868.
1972. "A Modified Multiple Regression Approach to the Analysis of Dichoto-
mous Variables." American Sociological Review 37:28-46.
Gourieroux, C., J. J. Laffont, and A. Monfort. 1980. "Coherency Conditions in Simul-
taneous Linear Equation Models with Endogenous Switching Regimes." Econo-
metrica 48:675-695.
Gourieroux, C., and A. Monfort. 1981. "Asymptotic Properties of the Maximum
Likelihood Estimator in Dichotomous Logit Models." Journal of Econometrics
17:83-97.
Gradshteyn, I. S., and I. M. Ryzhik. 1965. Table of Integrals, Series, and Products.
New York: Academic Press.
Granger, C. W. J., and P. Newbold. 1977. Forecasting Economic Time Series. New
York: Academic Press.
Greene, W. H. 1981. "On the Asymptotic Bias of the Ordinary Least Squares Estima-
tor of the Tobit Model." Econometrica 49:505-513.
1983. "Estimation of Limited Dependent Variable Models by Ordinary Least
Squares and the Method of Moments." Journal of Econometrics 21:195-212.
Grenander, U., and G. Szego. 1958. Toeplitz Forms and Their Applications. Berkeley:
University of California Press.
488 References
Griliches, Z. 1967. "Distributed Lags: A Survey." Econometrica 35:16-49.

Gronau, R. 1973. "The Effects of Children on the Housewife's Value of Time."
Journal of Political Economy 81:S168-S199.
1974. "Wage Comparisons - a Selectivity Bias." Journal ofPolitical Economy
82:1119-1143.
Guilkey, D. K., and P. Schmidt. 1979. "Some Small Sample Properties of Estimators
and Test Statistics in the Multivariate Logit Model." Journal of Econometrics
10:33-42.
Gumbel, E. J. 1961. "Bivariate Logistic Distributions." Journal of the American Sta-
tistical Association 56:335-349.
Gunst, R. F., and R. L. Mason. 1977. "Biased Estimation in Regression: An Evalua-
tion Using Mean Squared Error." Journal of the American Statistical Association
72:616-628.
Gurland, J., I. Lee, and P. A. Dahm. 1960. "Polychotomous Quantal Response in
Biological Asssay." Biometrics 16:382-398.
Haberman, S. J. 1978. Analysis of Qualitative Data. Vol. 1, Introductory Topics. New
York: Academic Press.
1979. Analysis of Qualitative Data. Vol. 2, New Developments. New York:
Academic Press.
Haessel, W. 1976. "Demand for Agricultural Commodities in Ghana: An Application
of Nonlinear Two-Stage Least Squares with Prior Information." American Jour-
nal of Agricultural Economics 58:341 -345.
Halmos, D. R. 1950. Measure Theory. Princeton: D. Van Nostrand.
Hammerstrom, T. 1981. "Asymptotically Optimal Tests for Heteroscedasticity in the
General Linear Model." Annals of Statistics 9:368-380.
Han, A. K. 1983. "Asymptotic Efficiency of the Partial Likelihood Estimator in the
Proportional Hazard Model." Technical Report no. 412, Institute for Mathemati-
cal Studies in the Social Sciences, Stanford University, Calif.
Hartley, H. 0. 1958. "Maximum Likelihood Estimation from Incomplete Data."
Biometrics 14:174 -194.
1961. "The Modified Gauss-Newton Method for the Fitting of Non-Linear
Regression Functions by Least Squares." Technometrics 3:269-280.
Hartley, M. J. 1976a. "The Estimation of Markets in Disequilibrium: The Fixed
Supply Case." International Economic Review 17:687 -699.
1976b. "Estimation of the Tobit Model by Nonlinear Least Squares Methods."
Discussion Paper no. 373, State University of New York, Buffalo.
1976c. "The Tobit and Probit Models: Maximum Likelihood Estimation by
Ordinary Least Squares." Discussion Paper no. 374, State University of New
York, Buffalo.
Harvey, A. C. 1976. "Estimating Regression Models with Multiplicative Heterosee-
dasticity." Econometrica 44:461-465.
1978. "The Estimation of Time-Varying Parameters from Panel Data." An-
nales de l'insee 30-31:203-226.
References 489
1981a. The Econometric Analysis of Time Series. Oxford: Philip Allan Pub-
lishers.
1981b. Time Series Models. Oxford: Philip Allan Publishers.
Hatanaka, M. 1974. "An Efficient Two-Step Estimator for the Dynamic Adjustment
Model with Autoregressive Errors." Journal of Econometrics 2:199 - 220.
1978. "On the Efficient Estimation Methods for the Macro-Economic Models
Nonlinear in Variables." Journal of Econometrics 8:323-356.
Hause, J. C. 1980. "The Fine Structure of Earnings and On-the-Job Training Hypoth-
esis." Econometrica 48:1013-1029.
Hausman, J. A. 1975. "An Instrumental Variable Approach to Full Information
Estimators for Linear and Certain Nonlinear Econometric Models." Economet-
rica 43:727-738.
1978. "Specification Tests in Econometrics." Econometrica 46:1251- 1272.
Hausman, J. A., and D. McFadden. 1984. "Specification Tests for the Multinomial
Logit Model." Econometrica 52:1219 - 1240.
Hausman, J. A., and W. E. Taylor. 1981. "Panel Data and Unobservable Individual
Effects." Econometrica 49:1377 - 1398.
Hausman, J. A., and D. A. Wise. 1976. "The Evaluation of Results from Truncated
Samples: The New Jersey Income Maintenance Experiment." Annals of Eco-
nomic and Social Measurement 5:421-445.
1977. "Social Experimentation, Truncated Distributions, and Efficient Esti-
mation." Econometrica 45:919 - 938.
1978. "A Conditional Probit Model for Qualitative Choice: Discrete Decisions
Recognizing Interdependence and Heterogeneous Preferences." Econometrica
46:403-426.
1979. "Attrition Bias in Experimental and Panel Data: The Gary Income
Maintenance Experiment." Econometrica 47:455 -473.
Heckman, J. J. 1974. "Shadow Prices, Market Wages, and Labor Supply." Economet-
rica 42:679-693.
I 976a. "The Common Structure of Statistical Models of Truncation, Sample
Selection and Limited Dependent Variables and a Simple Estimator for Such
Models." Annals of Economic and Social Measurement 5:475-492.
1976b. "Simultaneous Equations Models with Continuous and Discrete En-
dogenous Variables and Structural Shifts," in S. M. Goldfeld and R. E. Quandt,
eds., Studies in Nonlinear Estimation, pp. 235-272. Cambridge, Mass.: Ballinger
Publishing.
1978. "Dummy Endogenous Variables in a Simultaneous Equation System."
1979. "Sample Selection Bias as a Specification Error." Econometrica
47:153- 161.
1981a. "Statistical Models for Discrete Panel Data," in C. F. Manski and
D. McFadden, eds., Structural Analysis of Discrete Data with Econometric Appli-
cations, pp. 114 - 178. Cambridge, Mass.: MIT Press.
490 References
198 lb. "The Incidental Parameters Problem and the Problem of Initial Condi-
tions in Estimating a Discrete Time-Discrete Data Stochastic Process," in
C. F. Manslci and D. McFadden, eds., Structural Analysis of Discrete Data with
Econometric Applications, pp. 179- 195. Cambridge, Mass.: MIT Press.
1981c. "Heterogeneity and State Dependence," in S. Rosen, ed., Studies in
Labor Markets, pp. 91- 139. Cambridge, Mass.: National Bureau of Economic
Reseach.
Heckman, J. J., and G. J. Borjas. 1980. "Does Unemployment Cause Future Unem-
ployment? Definitions, Questions and Answers from a Continuous Time Model
of Heterogeneity and State Dependence." Economica 47:247 -283.
Heckman, J. J., and T. E. MaCurdy. 1980. "A Life Cycle Model of Female Labor
Supply." Review of Economic Studies 47:47-74.
Heckman, J. J., and S. Polachek. 1974. "Empirical Evidence on the Functional Form
of the Earnings-Schooling Relationship." Journal of the American Statistical As-
sociation 69:350-354.
Heckman, J. J., and B. Singer. 1982. "The Identification Problem in Econometric
Models for Duration Data," in W. Hildenbrand, ed., Advances in Econometrics,
pp. 39-77. Cambridge: Cambridge University Press.
1984a. "A Method for Minimizing the Impact of Distributional Assumptions
in Econometric Models for Duration Data." Econometrica 52:271-320.
1984b. "Econometric Duration Analysis." Journal of Econometrics 24:63 -
132.
Heckman, J. J., and R. J. Willis. 1977. "A Beta-Logistic Model for the Analysis of
Sequential Labor Force Participation by Married Women." Journal of Political
Economy 85:27-58.
Hettmansperger, T. P., and J. W. McKean. 1977. "A Robust Alternative Based on
Ranks to Least Squares in Analyzing Linear Models." Technometrics 19:275 -
284.
Hildreth, C., and J. P. Houck. 1968. "Some Estimators for a Linear Model with
Random Coefficients." Journal of the American Statistical Association 63:584 -
595.
Hill, R. W., and P. W. Holland. 1977. "Two Robust Alternatives to Least-Squares
Regression." Journal of the American Statistical Association 72:1041 - 1067.
Hinkley, D. V. 1971. "Inference in Two-Phase Regression." Journal of the American
1975. "On Power Transformations to Symmetry." Biometrika 62:101-112.
Hoadley, B. 1971. "Asymptotic Properties of Maximum Likelihood Estimators for the
Independent Not Identically Distributed Case." Annals of Mathematical Statis-
tics 42:1977-1991.
Hodges, J. L., and E. L. Lehmann. 1950. "Some Problems in Minimax Point Estima-
tion." Annals of Mathematical Statistics 21:182-197.
1963. "Estimates of Location Based on Rank Tests." Annals of Mathematical
Statistics 34:598-611.
References 491
Hoe!, P. G. 1971. Introduction to Mathematical Statistics, 4th ed. New York: John
Wiley & Sons.
Hoer!, A. E., and R. W. Kennard. 1970a. "Ridge Regression: Biased Estimation for
Nonorthogonal Problems." Technometrics 12:55-67.
1970b. "Ridge Regression: Applications to Nonorthogonal Problems." Tech-
nometrics 12:69-82.
Hoerl, A. E., R. W. Kennard, and K. F. Baldwin. 1975. "Ridge Regression: Some
Simulations." Communications in Statistics 4:105-123.
Hogg, R. V. 1974. "Adaptive Robust Procedures: A Partial Review and Some Sugges-
tions for Future Applications and Theory." Journal of the American Statistical
Holford, T. R. 1980. "The Analysis of Rates and of Survivorship Using Log-Linear
Models." Biometrics 36:299-305.
Horowitz, J. L., J. M. Sparmann, and C. F. Daganzo. 1982. "An Investigation of the
Accuracy of the Clark Approximation for the Multinomial Probit Model." Trans-
portation Science 16:382-401.
Hosek, J. R. 1980. "Determinants of Family Participation in the AFDC-Unemployed
Fathers Program." Review of Economics and Statistics 62:466-470.
Howe, H., R. A. Pollack, and T. J. Wales. 1979. "Theory and Time Series Esti-
mation of the Quadratic Expenditure System." Econometrica 47:1231 -
1248.
Hsiao, C. 1974. "Statistical Inference for a Model with Both Random Cross-Sectional
and Time Effects." International Economic Review 15:12-30.
1975. "Some Estimation Methods for a Random Coefficient Model." Econo-
metrica 43:305-325.
1983. "Identification," in Z. Griliches and M. D. Intrilligator, eds., Handbook
of Econometrics 1:223-283. Amsterdam: North-Holland Publishing.
Hsieh, D., C. F. Manski, and D. McFadden. 1983. "Estimation of Response Probabili-
ties from Augmented Retrospective Observations." Mimeographed Paper, De-
partment of Economics, Massachusetts Institute of Technology, Cambridge,
Mass.
Huang, D. S. 1964. "Discrete Stock Adjustment: The Case of Demand for Automo-
biles." International Economic Review 5:46-62.
Huber, P. J. 1964. "Robust Estimation of a Location Parameter." Annals of Mathe-
matical Statistics 35:73- 101.
1965. "The Behavior of Maximum Likelihood Estimates under Nonstandard
Conditions," in J. Neyman, ed., Proceedings of the Fifth Berkeley Symposium,
1:221-233. Berkeley: University of California Press.
1972. "Robust Statistics: A Review." Annals of Mathematical Statistics
43:1041-1067.
1977. Robust Statistical Procedures. Philadelphia: Society for Industrial and
Applied Mathematics.
1981. Robust Statistics. New York: John Wiley & Sons.
492 References
Hurd, M. 1979. "Estimation in Truncated Samples When There Is Heteroscedas-

ticity." Journal of Econometrics 11:247-258.
Imhof, J. P. 1961. "Computing the Distribution of Quadratic Forms in Normal Vari-
ables." Biometrilca 48:419-426.
Jaeckel, L. A. 1972. "Estimating Regression Coefficients by Minimizing the Disper-
sion of the Residuals." Annals of Mathematical Statistics 43:1449-1458.
James, W., and C. Stein. 1961. "Estimation with Quadratic Loss," in J. Neyman, ed.,
Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and
Probability, 1:361-379. Berkeley: University of California Press.
Jennrich, R. I. 1969. "Asymptotic Properties of Non-Linear Least Squares Estima-
tors." The Annals of Mathematical Statistics 40:633 -643.
Jobson, J. D., and W. A. Fuller. 1980. "Least Squares Estimation When the Covar-
iance Matrix and Parameter Vector Are Functionally Related." Journal of the
American Statistical Association 75:176 - 181.
Johnson, N. L., and S. Kotz. 1970a. Continuous Univariate Distributions - 1. Boston:
Houghton Mifflin.
1970b. Continuous Univariate Distributions-2. Boston: Houghton Mifflin.
1972. Distributions in Statistics: Continuous Multivariate Distributions. New
York: John Wiley & Sons.
Johnston, J. 1972. Econometric Methods. 2d ed. New York: McGraw-Hill.
Joreskog, K. G., and D. Sorbom. 1976. LISREL III Estimation of Structural Equation
Systems by Maximum Likelihood Methods. Chicago: National Educational Re-
sources.
Jorgenson, D. W., and J. Laffont. 1974. "Efficient Estimation of Nonlinear Simulta-
neous Equations with Additive Disturbances." Annals of Economic and Social
Measurement 3:615-640.
Jorgenson, D. W., and L. J. Lau. 1975. "The Structure of Consumer Preferences."
Annals of Economic and Social Measurement 4:49-101.
1978. "Testing the Integrability of Consumer Demand Functions, United
States, 1947-1971." Mimeographed paper.
Judge, G. G., and M. E. Bock, 1983. "Biased Estimation," in Z. Griliches and M. D.
Intrilligator, eds., Handbook of Econometrics, 1:599-649. Amsterdam: North-
Holland Publishing.
Kahn, L. M., and K. Morimune. 1979. "Unions and Employment Stability: A Se-
quential Logit Approach." International Economic Review 20:217-236.
Kalcwani, N. C. 1967. "The Unbiasedness of Zellner's Seemingly Unrelated Regres-
sion Equations Estimators." Journal of the American Statistical Association
62:141- 142.
Kalbfleisch, J. D. 1974. "Some Efficiency Calculations for Survival Distributions."
Biometrika 61:31 - 38.
Kalbfleisch, J. D., and R. L. Prentice. 1973. "Marginal Likelihoods Based On Cox's
Regression and Life Models." Biometrilca 60:267-278.
References 493 .
1980. The Statistical Analysis of Failure Time Data. New York: John Wiley &
Sons.
ICariya, T. 1981. "Bounds for the Covariance Matrices of Zellner's Estimator in the
SUR Model and the 2SAE in a Heteroscedastic Model." Journal of the American
Statistical Asssociation 76:975-979.
Kay, R. 1979. "Some Further Asymptotic Efficiency Calculations for Survival Data
Regression Models." Biometrika 66:91-96.
Keeley, M. C., P. K. Robins, R. G. Spiegelman, and R. W. West. 1978. "The Estima-
tion of Labor Supply Models Using Experimental Data." American Economic
Review 68:873-887.
Kelejian, H. H. 1971. "Two-Stage Least Squares and Econometric Systems Linear in
Parameters but Nonlinear in the Endogenous Variables." Journal ofthe American
Statistical Association 66:373 - 374.
1974. "Efficient Instrumental Variable Estimation of Large Scale Nonlinear
Econometric Models." Mimeographed paper.
Kelejian, H. H., and S. W. Stephan. 1983. "Inference in Random Coefficient Panel
Data Models: A Correction and Clarification of the Literature." International
Economic Review 24:249-254.
Kendall, M. G., and A. Stuart. 1979. The Advanced Theory of Statistics, 4th ed., vol. 2.
New York: Charles Griffin and Co.
Kenny, L. W., L. F. Lee, G. S. Maddala, and R. P. Trost. 1979. "Returns to College
Education: An Investigation of Self-Selection Bias Based on the Project Talent
Data." International Economic Review 20:775-789.
Kiefer, J., and J. Wolfowitz. 1956. "Consistency of the Maximum Likelihood Estima-
tor in the Presence of Infinitely Many Incidental Parameters." Annals of Mathe-
matical Statistics 27:887-906.
Koenker, R. 1981a. "Robust Methods in Econometrics." Bell Laboratories Eco-
nomics Discussion Paper no. 228.
1981b. "A Note on Studentizing a Test for Heteroscedasticity." Journal of
Koenker, R., and G. Bassett, Jr. 1978. "Regression Quantiles." Econometrica
46:33-50.
1982. "Robust Tests for Heteroscedasticity Based on Regression Quantiles."
Econometrica 50:43 - 61.
Koopmans, T. C., and W. C. Hood. 1953. "The Estimation of Simultaneous
Linear Economic Relationships," in W. C. Hood and T. C. Koopmans, eds.,
Studies in Econometric Method, pp. 112-199. New York: John Wiley
& Sons.
Kotlikoff, L. J. 1979. "Testing the Theory of Social Security and Life Cycle Accumula-
tion." American Economic Review 69:396-410.
Koyck, L. M. 1954. Distributed Lags and Investment Analysis. Amsterdam: North-
Holland Publishing.
494 References
Lachenbruch, P. A., C. Sneeringer, and L. T. Revo. 1973. "Robustness of the Linear

and Quadratic Discriminant Function to Certain Types of Nonnormality." Com-
munications in Statistics 1:39-56.
Lai, T. L., H. Robbins, and C. Z. Wei. 1978. "Strong Consistency of Least Squares
Estimates in Multiple Regression." Proceedings of the National Academy of
Sciences 75:3034-3036.
Lancaster, T. 1979. "Econometric Methods for the Duration of Unemployment."
LeCam, L. 1953. "On Some Asymptotic Properties of Maximum Likelihood Esti-
mates and Related Bayes Estimates." University of California Publications in
Lee, L. F. 1977. "Estimation of a Modal Choice Model for the Work Journey with
Incomplete Observations." Mimeographed paper, Department of Economics,
University of Minnesota.
1978. "Unionism and Wage Rates: A Simultaneous Equations Model with
Qualitative and Limited Dependent Variables." International Economic Review
19:415-433.
1981. "Simultaneous Equations Models with Discrete and Censored Vari-
ables," in C. F. Mansld and D. McFadden, eds., Structural Analysis of Discrete
Data with Econometric Applications, pp. 346-364. Cambridge, Mass.: MIT
Press.
1982a. "Health and Wage: A Simultaneous Equation Model with Multiple
Discrete Indicators." International Economic Review 23:199-221.
1982b. "A Bivariate Logit Model." Mimeographed paper, Center for Econo-
metrics and Decision Sciences, University of Florida.
1982c. "Some Approaches to the Correction of Selectivity Bias." Review of
Economic Studies 49:355-372.
Lee, L. F., G. S. Maddala, and R. P. Trost. 1980. "Asymptotic Covariance Matrices of
Two-Stage Probit and Two-Stage Tobit Methods for Simultaneous Equations
Models with Selectivity." Econometrica 48:491-503.
Lerman, S. R., and C. F. Manski. 1981. "On the Use of Simulated Frequencies to
Approximate Choice Probabilities," in C. F. Manslci and D. McFadden, eds.,
Structural Analysis of Discrete Data with Econometric Applications, pp. 305 -
319. Cambridge, Mass.: MIT Press.
Li, M. M. 1977. "A Logit Model of Homeownership." Econometrica 45:1081-1098.
Lillard, L. A., and Y. Weiss. 1979. "Components of Variation in Panel Earnings Data:
American Scientists 1960-70." Econometrica 47:437-454.
Lillard, L. A., and R. Willis. 1978. "Dynamic Aspects of Earnings Mobility." Econo-
metrica 46:985-1012.
Lindley, D. V. 1972. Bayesian Statistics, A Review. Philadephia: Society for Industrial
and Applied Mathematics.
Loeve, M. 1977. Probability Theory, 4th ed. Princeton: D. Van Nostrand.
References 495
McCall, J. J. 1971. "A Markovian Model of Income Dynamics." Journal of the Ameri-
can Statistical Association 66:439-447.
McFadden, D. 1974. "Conditional Logit Analysis of Qualitative Choice Behavior," in
P. Zarembka, ed., Frontiers in Econometrics, pp. 105 -142. New York: Academic
Press.
1976a. "A Comment on Discriminant Analysis 'versus' Logit Analysis."
Annals of Economic and Social Measurement 5:511 -524.
1976b. "The Revealed Preferences of a Government Bureaucracy: Empirical
Evidence." Bell Journal of Economics 7:55-72.
1977. "Quantitative Methods for Analyzing Travel Behavior of Individuals:
Some Recent Developments." Cowles Foundation Discussion Paper no. 474.
1978. "Modelling the Choice of Residential Location," in A. Karlqvist et al.,
eds., Spatial Interaction Theory and Planning Models, pp. 75-96. Amsterdam:
North-Holland Publishing.
1981. "Econometric Models of Probabilistic Choice," in C. F. Manski and D.
McFadden, eds., Structural Analysis of Discrete Data with Econometric Applica-
tions, pp. 198-272. Cambridge, Mass.: MIT Press.
McFadden, D., and F. Reid. 1975. "Aggregate Travel Demand Forecasting from
Disaggregated Behavior Models." Transportation Research Board, Record, no.
534, Washington, D. C.
McGillivray, R. G. 1972. "Binary Choice of Urban Transport Mode in the San Fran-
cisco Bay Region." Econometrica 40:827 - 848.
McKean, J. W., and T. P. Hettmansperger. 1976. "Tests of Hypothesis in the General
Linear Model Based on Ranks." Communications in Statistics A,1:693- 709.
MacRae, E. 1977. "Estimation of Time-Varying Markov Process with Aggregate
Data." Econometrica 45:183- 198.
MaCurdy, T. E. 1980. "An Intertemporal Analysis of Taxation and Work Disincen-
tives." Working Papers in Economics no. E-80-4, The Hoover Institution, Stan-
ford University, Calif.
1982. "The Use of Time Series Processes to Model the Error Structure
of Earnings in a Longitudinal Data Analysis." Journal of Econometrics. 18:83 -
114.
Maddala, G. S. 1971. "The Use of Variance Components Models in Pooling Cross
Section and Time Series Data." Econometrica 39:341-358.
1980. "Disequilibrium, Self-Selection and Switching Models." Social Science
Working Paper 303, California Institute of Technology.
1983. Limited-Dependent and Qualitative Variables in Econometrics. Cam-
bridge: Cambridge University Press.
Maddala, G. S., and F. D. Nelson. 1974. "Maximum Likelihood Methods for Models
of Markets in Disequilibrium." Econometrica 42:1013-1030.
Malik, H. J., and B. Abraham. 1973. "Multivariate Logistic Distributions." Annals of
Statistics 1:588 - 590.
496 References
Malinvaud, E. 1961. "Estimation et prevision dans les modeles Economiques autore-

gressifs." Revue de l'institut international de statistique 29:1-32.
1980. Statistical Methods of Econometrics, 3d rev. ed. Amsterdam: North-
Holland Publishing.
Mallows, C. L. 1964. "Choosing Variables in a Linear Regression: A Graphical Aid."
Paper presented at the Central Region Meeting of the Institute of Mathematical
Statistics, Manhattan, Kansas.
Mandelbrot, B. 1963. "New Methods in Statistical Economics." Journal of Political
Economy 71:421-440.
Mann, H. B., and A. Wald. 1943. "On Stochastic Limit and Order Relationships."
Annals of Mathematical Statistics 14:217-226.
Manski, C. F. 1975. "The Maximum Score Estimation of the Stochastic Utility Model
of Choice." Journal of Econometrics 3:205 -228.
1985. "Semiparametric Analysis of Discrete Response: Asymptotic Properties
of the Maximum Score Estimator." Journal of Econometrics 27:313 - 333.
Manslci, C. F., and S. R. Lerman. 1977. "The Estimation of Choice Probabilities from
Choice-Based Samples." Econometrica 45:1977- 1988.
Manski, C. F., and D. McFadden. 1981. "Alternative Estimators and Sample Designs
for Discrete Choice Analysis," in C. F. Manski and D. McFadden, eds., Structural
Analysis of Discrete Data with Econometric Applications, pp. 2 - 50. Cambridge,
Mass.: MIT Press.
Marcus, M., and H. Minc. 1964. A Survey of Matrix Theory and Matrix Inequalities.
Boston: Prindle, Weber & Schmidt.
Mariano, R. S. 1982. "Analytical Small-Sample Distribution Theory in Econometrics:
The Simultaneous-Equations Case." International Economic Review 23:503 -
533.
Mariano, R. S., and B. W. Brown. 1983. "Asymptotic Behavior of Predictors in a
Nonlinear Simultaneous System." International Economic Review 24:523-536.
Marquardt, D. W. 1963. "An Algorithm for the Estimation of Non-Linear Parame-
ters." Society for Industrial and Applied Mathematics Journal 11:431-441.
Mayer, T. 1975. "Selecting Economic Hypotheses by Goodness of Fit." Economic
Journal 85:877-883.
Miller, R. G., Jr. 1981. Survival Analysis. New York: John Wiley & Sons.
Mizon, G. E. 1977. "Inferential Procedures in Nonlinear Models: An Application in a
UK Industrial Cross Section Study of Factor Substitution and Returns to Scale."
Econometrica 45:1221 - 1242.
Mood, A. M., F. A. Graybill, and D. C. Boes. 1974. Introduction to the Theory of
Statistics, 3d ed. New York: McGraw-Hill.
Morimune, K. 1979. "Comparisons of Normal and Logistic Models in the Bivariate
Dichotomous Analysis." Econometrica 47:957 -976.
Mosteller, F. 1946. "On Some Useful 'Inefficient' Statistics." Annals of Mathematical
Statistics 17:377 - 408.
References 497
Muthen, B. 1979. "A Structural Probit Model with Latent Variables." Journal of the
Nagar, A. L., and S. N. Sahay. 1978. "The Bias and Mean Squared Error of Forecasts
from Partially Restricted Reduced Form." Journal of Econometrics 7:227 -243.
Nakamura, M., A. Nakamura, and D. Cullen. 1979. "Job Opportunities, the Offered
Wage, and the Labor Supply of Married Women." American Economic Review
69:787-805.
Nelder, J. A., and R. W. M. Wedderbum. 1972. "Generalized Linear Models." Jour-
nal of the Royal Statistical Society ser. B, 135:370-384.
Nelson, F. D. 1977. "Censored Regression Models with Unobserved, Stochastic Cen-
soring Thresholds." Journal of Econometrics 6:309-327.
1981. "A Test for Misspecification in the Censored Normal Model." Econo-
metrica 49:1317-1329.
Nelson, F. D., and L. Olson. 1978. "Specification and Estimation of a Simultaneous-
Equation Model with Limited Dependent Variables." International Economic
Review 19:695-709.
Nerlove, M. 1958. Distributed Lags and Demand Analysis for Agricultural and Other
Commodities. Washington, D.C.: U.S. Department of Agriculture.
1971. "Further Evidence on the Estimation of Dynamic Relations from a
Time Series of Cross Sections." Econometrica 39:359-382.
Nerlove, M., D. M. Grether, and J. L. Carvalho. 1979. Analysis of Economic Time
Series: A Synthesis. New York: Academic Press.
Nerlove, M., and S. J. Press. 1973. "Univariate and Multivariate Log-Linear and
Logistic Models." RAND Corporation Paper R-1306-EDA/NIH, Santa Monica,
Calif.
Neyman, J., and E. L. Scott. 1948. "Consistent Estimates Based on Partially Consis-
tent Observations." Econometrica 16:1-32.
Nickell, S. 1979. "Estimating the Probability of Leaving Unemployment." Economet-
rica 47:1249-1266.
Norden, R. H. 1972. "A Survey of Maximum Likelihood Estimation." International
Statistical Revue 40:329-354.
1973. "A Survey of Maximum Likelihood Estimation, Part 2." International
Statistical Revue 41:39-58.
Oberhofer, W., and J. Kmenta. 1974. "A General Procedure for Obtaining Maximum
Likelihood Estimates in Generalized Regression Models." Econometrica
42:579-590.
Olsen, R. J. 1978. "Note on the Uniqueness of the Maximum Likelihood Estimator for
the Tobit Model." Econometrica 46:1211 -1215.
1980. "A Least Squares Correction for Selectivity Bias." Econometrica
48:1815-1820.
Paarsch, H. J. 1984. "A Monte Carlo Comparison of Estimators for Censored Regres-
sion Models." Journal of Econometrics 24:197 - 213.
498 References
Parke, W. R. 1982. "An Algorithm for FIML and 3SLS Estimation of Large Nonlinear
Models." Econometrica 50:81-95.
Pearson, E. S., and N. W. Please. 1975. "Relation between the Shape of Population
Distribution and the Robustness of Four Simple Statistical Tests." Biometrika
62:223-241.
Pesaran, M. H. 1982. "Comparison of Local Power of Alternative Tests of Non-Nested
Regression Models." Econometrica 50:1287- 1305.
Pfanzagl, J. 1973. "Asymptotic Expansions Related to Minimum Contrast Estima-
tors." Annals of Statistics 1:993-1026.
Phillips, P. C. B. 1977. "Approximations to Some Finite Sample Distributions Asso-
ciated with a First-Order Stochastic Difference Equation." Econometrica
45:463-485.
1982. "On the Consistency of Nonlinear FIML." Econometrica 50:1307 -
1324.
1983. "Exact Small Sample Theory in the Simultaneous Equations Model," in
Z. Griliches and M. D. Intrilligator, eds., Handbook of Econometrics, 2:449-516.
Amsterdam: North-Holland Publishing.
Pierce, D. A. 1971. "Least Squares Estimation in the Regression Model with Autore-
gressive-Moving Average Errors." Biometrika 58:299-312.
Plackett, R. L. 1960. Principles of Regression Analysis. London: Oxford University
Press.
1965. "A Class of Bivariate Distributions." Journal of American Statistical
Poirier, D. J. 1978. "The Use of the Box-Cox Transformation in Limited Dependent
Variable Models." Journal of the American Statistical Association 73:284-287.
Powell, J. L. 1981. "Least Absolute Deviations Estimation for Censored and Trun-
cated Regression Models." Technical Report no. 356, Institute for Mathematical
Studies in the Social Sciences, Stanford University, Calif.
1983. "Asymptotic Normality of the Censored and Truncated Least Absolute
Deviations Estimators." Technical Report no. 395, Institute for Mathematical
Studies in the Social Sciences, Stanford University, Calif.
Powell, M. J. D. 1964. "An Efficient Method for Finding the Minimum of a Function
of Several Variables without Calculating Derivatives." Computer Journal 7:115 -
162.
Powers, J. A., L. C. Marsh, R. R. Huckfeldt, and C. L. Johnson. 1978. "A Comparison
of Logit, Probit and Discriminant Analysis in Predicting Family Size." American
Statistical Association Proceedings of the Social Statistics Section, pp. 693-697.
Prais, S. J., and H. S. Houthaldcer. 1955. The Analysis of Family Budgets. Cambridge:
Cambridge University Press.
Pratt, J. W. 1981. "Concavity of the Log Likelihood." Journal of the American Statis-
tical Association 76:103- 106.
References 499
Press, S. J., and S. Wilson. 1978. "Choosing Between Logistic Regression and Discrim-
inant Analysis." Journal of the American Statistical Association 73:699-705.
Quandt, R. E. 1958. "The Estimation of the Parameters of a Linear Regression System
Obeying Two Separate Regimes." Journal of the American Statistical Association
53:873-880.
1982. "Econometric Disequilibrium Models." Econometric Reviews 1:1-63.
1983. "Computational Problems and Methods," in Z. Griliches and M. D.
Intrilligator, eds., Handbook of Econometrics, 1:699-764. Amsterdam: North-
Holland Publishing.
Quandt, R. E., and J. B. Ramsey. 1978. "Estimating Mixtures of Normal Distributions
and Switching Regressions." Journal of the American Statistical Association
73:730-738.
Radner, R., and L. S. Miller. 1970. "Demand and Supply in U. S. Higher Education: A
Progress Report." American Economic Review-Papers and Proceedings
60:326-334.
Rao, C. R. 1947. "Large Sample Tests of Statistical Hypotheses Concerning Several
Parameters with Applications to Problems of Estimation." Proceedings of the
Cambridge Philosophical Society 44:50-57.
1965. "The Theory of Least Squares When the Parameters Are Stochastic and
Its Applications to the Analysis of Growth Curves." Biometrilca 52:447-458.
1970. "Estimation of Heteroscedastic Variances in a Linear Model." Journal
of the American Statistical Association 65:161-172.
1973. Linear Statistical Inference and Its Applications, 2d ed. New York: John
Wiley & Sons.
Reece, W. S. 1979. "Charitable Contributions: The New Evidence on Household
Behavior." American Economic Review 69:142 -151.
Rice, P., and V. K. Smith. 1977. "An Econometric Model of the Petroleum Industry."
Journal of Econometrics 6:263-288.
Roberts, R. B., G. S. Maddala, and G. Enholm. 1978. "Determinants of the Requested
Rate of Return and the Rate of Return Granted in a Formal Regulatory Process."
Bell Journal of Economics 9:611-621.
Robinson, P. M. 1982a. "On the Asymptotic Properties of Estimators of Models
Containing Limited Dependent Variables." Econometrica 50:27-41.
1982b. "Analysis of Time Series from Mixed Distributions." Annals of Statis-
tics 10:915-925.
Rosenberg, B. 1973. "The Analysis of a Cross Section of Time Series by Stochastically
Convergent Parameter Regression." Annals of Economic and Social Measure-
ment 2:399-428.
Rosenzweig, M. R. 1980. "Neoclassical Theory and the Optimizing Peasant: An Econ-
ometric Analysis of Market Family Labor Supply in a Developing Country."
Quarterly Journal of Economics 94:31-55.
500 References
Royden, H. L. 1968. Real Analysis, 2d ed. New York: Macmillan.

Ruppert, D., and R. J. Carroll. 1980. "Trimmed Least Squares Estimation in the
Linear Model." Journal of the American Statistical Association 75:828-838.
Sant, D. T. 1978. "Partially Restricted Reduced Forms: Asymptotic Relative Effi-
ciency." International Economic Review 19:739-747.
Sargent, T. J. 1978. "Estimation of Dynamic Labor Schedules under Rational Expec-
tations." Journal of Political Economy 86:1009-1044.
Sawa, T., and T. Hiromatsu. 1973. "Minimax Regret Significance Points for a Prelimi-
nary Test in Regression Analysis." Econometrica 41:1093-1101.
Scheffe, H. 1959. The Analysis of Variance. New York: John Wiley & Sons.
Schlossmacher, E. J. 1973. "An Iterative Technique for Absolute Deviation Curve
Fitting." Journal of the American Statistical Association 68:857-865.
Schmee, J., and G. J. Hahn. 1979. "A Simple Method for Regression Analysis with
Censored Data." Technometrics 21:417-432.
Schmidt, P., and R. Sickles. 1977. "Some Further Evidence on the Use of the Chow
Test under Heteroscedasticity." Econometrica 45:1293-1298.
Schwarz, G. 1978. "Estimating the Dimension of a Model." Annals of Statistics
6:461 - 464.
Sclove, S. L. 1973. "Least Squares Problems with Random Regression Coefficients."
Technical Report no. 87, Institute for Mathematical Studies in the Social
Sciences, Stanford University, Calif.
Sclove, S. L., C. L. Morris, and R. Radhakrishnan. 1972. "Non-Optimality of Prelimi-
nary-Test Estimators for the Mean of a Multivariate Normal Distribution."
Shiller, R. J. 1978. "Rational Expectations and the Dynamic Structure of Macroeco-
nomic Models." Journal of Monetary Economics 4:1-44.
Shorrocks, A. F. 1976. "Income Mobility and the Markov Assumption." Economic
Journal 86:566-578.
Silberman, J. I., and G. C. Durden. 1976. "Determining Legislative Preferences on the
Minimum Wage: An Economic Approach." Journal of Political Economy
84:317-329.
Silberman, J. I., and W. K. Talley. 1974. "N-Chotomous Dependent Variables: An
Application to Regulatory Decision-Making." American Statistical Association
Proceedings of the Business and Economic Statistics Section, pp. 573-576.
Silvey, S. D. 1959. "The Lagrangian Multiplier Test." Annals of Mathematical Statis-
tics 30:389-407.
1969. "Multicollinearity and Imprecise Estimation." Journal of the Royal
Society ser. B, 31:539-552.
Small, K. A. 1981. "Ordered Logit: A Discrete Choice Model with Proximate Covar-
iance Among Alternatives." Research Memorandum no. 292, Econometric Re-
search Program, Princeton University.
References 501
Small, K. A., and D. Brownstone. 1982. "Efficient Estimation of Nested Logit Models:
An Application to Trip Timing." Research Memorandum no. 296, Econometric
Research Program, Princeton University.
Smith, K. C., N. E. Savin, and J. L. Robertson. 1984. "A Monte Carlo Comparison of
Maximum Likelihood and Minimum Chi-Square Sampling Distributions in
Logit Analysis." Biometrics 40:471-482.
Spitzer, J. J. 1976. "The Demand for Money, the Liquidity Trap, and Functional
Forms." International Economic Review 17:220-227.
1978. "A Monte Carlo Investigation of the Box-Cox Transformation in Small
Samples." Journal of the American Statistical Association 73:488-495.
Srivastava, V. K., and T. D. Dwivedi. 1979. "Estimation of Seemingly Unrelated
Regression Equations: A Brief Survey." Journal of Econometrics 10:15 -32.
Stapleton, D. C., and D. J. Young. 1984. "Censored Normal Regression with Measure-
ment Error on the Dependent Variable." Econometrica 52:737-760.
Stein, C. 1973. "Estimation of the Mean of a Multivariate Normal Distribution."
Technical Report no. 48, Department of Statistics, Stanford University, Calif.
Stephenson, S. P., and J. F. McDonald. 1979. "Disaggregation of Income Mainte-
nance Impacts on Family Earnings." Review of Economics and Statistics 61:354 -
360.
Stigler, S. M. 1973. "Simon Newcomb, Percy Daniell, and the History of Robust
Estimation, 1885- 1920." Journal of the American Statistical Association
68:872-879.
1977. "Do Robust Estimators Work with Real Data?" Annals of Statistics
5:1055-1098.
Strawderman, W. E. 1978. "Minimax Adaptive Generalized Ridge Regression Esti-
mators." Journal of the American Statistical Association 73:623-627.
Strickland, A. D., and L. W. Weiss. 1976. "Advertising, Concentration, and Price-Cost
Margins." Journal of Political Economy 84:1109-1121.
Strotz, R. H. 1960. "Interdependence as a Specification Error." Econometrica
28:428-442.
Swamy, P. A. V. B. 1970. "Efficient Inference in a Random Coefficient Regression
Model." Econometrica 38:311-323.
1980. "A Comparison of Estimators for Undersized Samples." Journal of
Swamy, P. A. V. B., and J. S. Mehta. 1977. "Estimation of Linear Models with Time
and Cross-Sectionally Varying Parameters." Journal of the American Statistical
Taylor, W. E. 1978. "The Heteroscedastic Linear Model: Exact Finite Sample Re-
sults." Econometrica 46:663-676.
1980. "Small Sample Considerations in Estimation from Panel Data." Journal
of Econometrics 13:203-223.
502 References
1981. "On the Efficiency of the Cochrane-Orcutt Estimator." Journal of Econ-

ometrics 17:67-82.
1983. "On the Relevance of Finite Sample Distribution Theory." Econometric
Reviews 2:1 -39.
Taylor, W. F. 1953. "Distance Functions and Regular Best Asymptotically Normal
Estimates." Annals of Mathematical Statistics 24:85-92.
Telser, L. G. 1963. "Least-Squares Estimates of Transition Probabilities," in C. Christ,
ed., Measurement in Economics, pp. 270-292. Stanford, Calif.: Stanford Univer-
sity Press.
Theil, H. 1953. "Repeated Least-Squares Applied to Complete Equation Systems."
Mimeographed paper. The Hague: Central Planning Bureau.
1961. Economic Forecasts and Policy, 2d ed. Amsterdam: North-Holland
Publishing.
1971. Principles of Econometrics. New York: John Wiley & Sons.
Theil, H., and A. S. Goldberger. 1961. "On Pure and Mixed Statistical Estimation in
Economics." International Economic Review 2:65 -78.
Thisted, R. A. 1976. "Ridge Regression, Minimax Estimation, and Empirical Bayes
Method." Technical Report no. 28, Division of Biostatistics, Stanford University,
Calif.
Tobin, J. 1958. "Estimation of Relationships for Limited Dependent Variables."
Toikka, R. S. 1976. "A Markovian Model of Labor Market Decisions by Workers."
American Economic Review 66:821 -834.
Tomes, N. 1981. "The Family, Inheritance, and the Intergenerational Transmission of
Inequality." Journal of Political Economy 89:928 - 958.
Toyoda, T. 1974. "Use of the Chow Test Under Heteroscedasticity." Econometrica
42:601-608.
Tsiatis, A. A. 1981. "A Large Sample Study of Cox's Regression Model." The Annals of
Tsurumi, H. 1970. "Nonlinear Two-Stage Least Squares Estimation of CES Produc-
tion Functions Applied to the Canadian Manufacturing Industries." Review of
Economics and Statistics 52:200-207.
Tuma, N. B. 1976. "Rewards, Resources, and the Rate of Mobility: A Nonstation-
ary Multivariate Stochastic Model." American Sociological Review 41:338 -
360.
Tuma, N. B., M. T. Hannan, and L. P. Groeneveld. 1979. "Dynamic Analysis of Event
Histories." Journal of Sociology 84:820-854.
Uhler, R. S. 1968. "The Demand for Housing: An Inverse Probability Approach."
Review of Economics and Statistics 50:129-134.
Vinod, H. D. 1978. "A Ridge Estimator Whose MSE Dominates OLS." International
Economic Review 19:727-737.
Wald, A. 1943. "Tests of Statistical Hypotheses Concerning Several Parameters When
References 503
the Number of Observations Is Large." Transactions ofthe American Mathemati-

cal Society 54:426-482.
1949. "Note on the Consistency of the Maximum Likelihood Estimate."
Wales, T. J., and A. D. Woodland. 1980. "Sample Selectivity and the Estimation of
Labor Supply Functions." International Economic Review 21:437-468.
Walker, S. H., and D. B. Duncan. 1967. "Estimation of the Probability of an Event as a
Function of Several Independent Variables." Biometrika 54:167 - 179.
Wallace, T. D., and A. Hussain. 1969. "The Use of Error Components Models in
Combining Cross-Section with Time Series Data." Econometrica 37:55 -72.
Wallis, K. F. 1980. "Econometric Implications of the Rational Expectations Hypoth-
esis." Econometrica 48:49-73.
Warner, S. L. 1962. Stochastic Choice of Mode in Urban Travel-A Study in Binary
Choice. Evanston, Ill.: Northwestern University Press.
1963. "Multivariate Regression of Dummy Variates Under Normality As-
sumptions." Journal of the American Statistical Association 58:1054-1063.
Watson, G. S. 1955. "Serial Correlation in Regression Analysis, I." Biometrika
42:327-341.
Welch, B. L. 1938. "The Significance of the Difference between Two Means When the
Population Variances Are Unequal." Biometrika 29:350-362.
Westin, R. B. 1974. "Predictions from Binary Choice Models." Journal of Economet-
rics 2:1 - 16.
Westin, R. B., and P. W. Gillen. 1978. "Parking Location and Transit Demand."
Journal of Econometrics 8:75 - 101.
White, H. 1980a. "A Heteroskedasticity-Consistent Covariance Matrix Estimator and
a Direct Test for Heteroskedasticity." Econometrica 48:817 - 838.
1980b. "Nonlinear Regression on Cross-Section Data." Econometrica
48:721-746.
1982. "Instrumental Variables Regression with Independent Observations."
Econometrica 50:483 -499.
- , ed. 1983. "Non-Nested Models." Journal of Econometrics 21:1-160.
White, H., and G. M. MacDonald. 1980. "Some Large Sample Tests for Nonnor-
mality in the Linear Regression Model." Journal of the American Statistical
White, K. J. 1972. "Estimation of the Liquidity Trap with a Generalized Functional
Form." Econometrica 40:193-199.
Whittle, P. 1983. Prediction and Regulation. Minneapolis: University of Minnesota
Press.
Wiggins, S. N. 1981. "Product Quality Regulation and New Drug Introductions: Some
New Evidence from the 1970's." Review ofEconomics and Statistics 63:615 - 619.
Willis, R. J., and S. Rosen. 1979. "Education and Self-Selection." Journal of Political
Economy 87:S7 - S36.
504 References
Witte, A. D. 1980. "Estimating the Economic Model of Crime with Individual Data."
Quarterly Journal of Economics 94:57-84.
Wu, C. F. J. 1983. "On the Convergence Properties of the EM Algorithm." Annals of
Wu, D. M. 1965. "An Empirical Analysis of Household Durable Goods Expenditure."
Econometrica 33:761 -780.
Zacks, S. 1971. Theory of Statistical Inference. New York: John Wiley & Sons.
Zarembka, P. 1968. "Functional Form in the Demand for Money." Journal of the
Zellner, A. 1962. "An Efficient Method of Estimating Seemingly Unrelated Regres-
sions and Tests for Aggregation Bias." Journal of the American Statistical Associ-
ation 57:348-368.
1971. An Introduction to Bayesian Inference in Econometrics. New York: John
Wiley & Sons.
Zellner, A., D. S. Huang, and L. C. Chau. 1965. "Further Analysis of the Short-Run
Consumption Function with Emphasis on the Role of Liquid Assets." Economet-
rica 40:193-199.
Zellner, A., and H. Theil. 1962. "Three-Stage Least Squares: Simultaneous Estimation
of Simultaneous Equations." Econometrica 30:54-78.
Name Index
Abraham, B., 319 Bellman, R., 416, 459, 474

Abramovitz, M., 193 Belsley, D. A., 264
Adams, J. D., 365 Bennett, J., 307
Adelman, I. G., 420 Benus, J., 252
Aigner, D. J., 67, 69, 123 Bera, A. K., 382
Aitchison, J., 307 Berger, J. 0., 65
Akahira, M., 136 Berkson, J., 275, 278, 280
Akaike, H., 52, 146, 147, 466 Berndt, E. R., 138, 145
Albert, A., 271 Berzeg, K., 214
Albright, R. L., 309, 310 Bhattacharya, P. J., 65
Almon, S., 178, 469 Bhattacharya, R. K., 92
Amemiya, T., 40, 52, 123, 146, 147, 161, Bianchi, C., 262, 263
178, 179, 189, 195, 199, 202, 203, 204, Bickel, P. J., 71, 72, 73, 75, 154, 203, 465,466
205, 206, 211, 215, 217, 218, 222, 245, Bishop, Y. M. M., 316
246, 248, 249, 251, 254, 257, 258, 259, Blattberg, R., 77
264, 269, 278, 285, 286, 295, 296, 298, Blumen, I., 418, 472
299, 300, 302, 303, 304, 305, 306, 321, Bock, M. E., 69, 466
322, 328, 408, 465, 468, 470, 471, 472 Bodlcin, R. G., 141,468
Anderson, J. A., 271 Boes, D. C., 466
Anderson, T. W., 10, 159, 172, 173, 174, Borjas, G. J., 405, 406
180, 183, 187, 216, 221, 235, 238, 239, Boskin, M. J., 383, 421, 424
240, 417, 469, 470, 471 Box, G. E. P., 172, 249, 250, 465
Andrews, D. F., 72, 73, 75 Breusch, T. S., 145, 206-207, 469
Anscombe, F. J., 203 Brillinger, D. R., 159
Apostol, T. M., 470 Brook, R. J., 54
Arabmazer, A., 380, 381 Brown, B. W., 256, 263
Arrow, K. J., 128 Brown, M., 128
Ashenfelter, 0., 365 Brown, R. L., 222
Ashford, J. R., 317, 319 Brown, W. G., 67
Brownstone, D., 304
Baldwin, K. F., 64, 69 Brundy, J. M., 242
Balestra, P., 211, 213, 214, 215 Buse, A., 206
Baltagi, B. H., 211 Butler, J. S., 350
Baranchik, A. J., 63
Barankin, E. W., 125, 433, 470
Bartholomew, D. J., 412, 418 Calzolari, G., 262, 263
Basmann, R. L., 239, 470 Carroll, R. J., 78, 154, 202
Bassett, G., 77, 78, 154, 468 Carvalho, J. L., 159, 178
Battese, G. E., 211,212 Chamberlain, G., 199, 217
Beach, C. M., 176, 190 Champemowne, D. G., 420
Beattie, B. R., 67 Charatsis, E. G., 141
506 Name Index
Chau, L. C., 246 Fair, R. C., 78, 264, 265, 365, 403, 473
Chenery, H. B., 128 Farebrother, N. W., 53, 469
Chow, G. C., 234, 264, 465 Feinberg, S. E., 316
Christ, C. F., 228 Feller, W., 274
Christensen, L. R., 128 Ferguson, T. S., 125
Chung, K. L., 90, 466, 472 Fisher, F. M., 229, 256, 469, 470
Clark, C., 309 Fisher, R. A., 16
Cochrane, D., 189-190 Fletcher, R., 138
Cooley, T. F., 222 Flinn, C. J., 446, 448, 457
Cooper, J. M., 469 Fomby, T. B., 59
Cosslett, S. R., 321, 322, 326, 331, 332, 334, Forsythe, A. B., 77
336, 337, 338, 339, 346, 358, 471, 472 Freund, J. F., 466
Cox, D. R., 147, 249, 250, 251, 278, 307, Froehlich, B. R., 469
449, 450, 465, 466 Fuller, W. A., 159, 161, 178, 195, 202, 211,
Cragg, J. G., 410, 472 212, 238, 239
Cramer, H., 93, 118
Crowder, J. M., 178 Gallant, A. R., 113, 136, 137, 141, 145, 257,
Cullen, D., 395 262
Gastwirth, J. L., 74
Dapnzo, C. F., 309 Gayer, K. M., 49
Dagenais, M. G., 264 Geisel, M. S., 49
Dahm, P. A., 293 Ghosh, J. K., 125, 279
David, J. M., 293 Gillen, P. W., 389
Davidson, W. C., 138 Gnedenko, B. V., 90
Davis, L., 279 Godfrey, L. G., 469
Deacon, R., 293, 294, 471 Goldberg, S., 166
Dempster, A. P., 66, 375 Goldberger, A. S., 25, 360, 367, 380, 381, 472
Dent, W. T., 469 Goldfeld, S. M., 37, 38, 137, 138, 204, 206,
Dhrymes, P. J., 178 207, 256, 266, 354, 470
Diewert, W. E., 128 Goodman, L. A., 315, 316, 417, 418, 419
Dolcsum, K. A., 71, 465, 466 Gourieroux, C., 273, 472
Domencich, T. A., 269 Gradshteyn, I. S., 474
Doob, J. L., 159 Granger, C. W. J., 159
Draper, N. R., 137, 251, 466 Graybill, F. A., 466
Dubin, J. A., 408 Greene, W. H., 368, 472
Dudley, L., 387 Greenstadt, J., 264
Duncan, D. B., 275 Grenander, U., 161
Duncan, G. M., 405, 406, 407, 408, 473 Grether, D. M., 159, 178
Duncan, G. T., 421 Griliches, Z., 178, 217
Durbin, J., 191, 192, 193, 194, 196, 222, Groeneveld, L. P., 444
224, 225, 233 Gronau, R., 388, 473
Durden, G. C., 312 Guilkey, D. K., 317
Dwivedi, T. D., 198 Gumbel, E. J., 319
Gunst, R. F., 66
Efron, B., 62, 63, 64, 67, 135, 283, 284, 455 Gurland, J., 124, 293, 433, 470
Ehrlich, I., 252
Eicker, F., 199 Haberman, S. J., 316
Eisenpress, H., 264 Haessel, W., 258
Engle, R. F., 142 Hahn, G. J., 378
Enholm, G., 394, 395 Hall, B. H., 138
Evans, J. M., 222 Hall, R. E., 138
Name Index 507
Halmos, D. R., 467 Jarque, C. M., 382

Ham, J., 365 Jenkins, G. M., 172
Hammerstrom, T., 203 Jennrich, R. I., 107, 117, 468
Hampel, F. R., 72, 73, 75 Jobson, J. D., 202
Han, A. K., 455 Johnson, C. L., 300
Hannan, M. T., 444 Johnson, N. L., 62, 296, 300, 463, 472
Hartley, H. 0., 140, 141 Johnson, S. R., 59
Hartley, M. J., 375, 403, 472 Johnston, J., 238
Harvey, A. C., 159, 172, 189, 222, 242 Joreskog, K. G., 217
Hatanaka, M., 178, 265 Jorgenson, D. W., 128, 242, 257, 258, 262
Hause, J. C., 217 Judge, G. G., 67, 69, 466
Hausman, J. A., 138, 145, 217, 218, 233,
299, 302, 304, 308, 309, 310, 381, 395
Kahn, L. M., 310,311
Heckman, J. J., 252, 348, 349, 350, 351, 353,
Kakwani, N. C., 198
354, 368, 369, 387, 389, 390, 391, 392,
Kalbfleisch, J. D., 412, 451, 455, 474
395, 396, 397, 401, 402, 412, 446, 448,
Kariya, T., 198, 202
457, 473, 474
Kay, R., 455
Heien, D., 128
Keeley, M. C., 365
Hettmansperger, T. P., 79
Kelejian, H. H., 218, 219, 220, 246, 249
Hildreth, C., 204, 469
Kendall, M. G., 36
Hill, R. C., 59
Kennard, R. W., 56, 60, 64
Hill, R. W., 75, 77
Kenny, L. W., 396, 397, 404
Hinkley, D. V., 222, 251, 465, 466
Kiefer, J., 331, 346, 347
Hiromatsu, T., 53-54
Klein, L. R., 141, 468
Hoadley, B., 117
Kmenta, J., 186, 252
Hodges, J. L., 61, 74, 79, 124
Koenker, R., 72, 77, 78, 154, 207, 468
Hod, P. G., 466
Kogan, M., 418, 472
Hoer!, A. E., 56, 60, 64
Kolmogorov, A. N., 90
Hogg, R. J., 72, 74
Koopmans, T. C., 470
Holford, T. R., 437
Kotlikoff, L. J., 364, 365
Holland, P. W., 75, 77, 316
Kotz, S., 62, 296, 300, 463, 472
Holly, A., 145, 262
Koyck, L. M., 178
Hood, W. C., 470
Horowitz, J. L., 309
Hosek, J. R., 328 Lachenbruch, P. A., 285
Houck, J. P., 204, 469 Laffont, J., 257, 472
Houthakker, H. S., 198, 203 Lai, T. L., 467
Howe, H., 128 Laird, N. M., 375
Hsiao, C., 216, 218, 220, 221, 469,470 Lancaster, T., 444, 445, 446, 448, 474
Hsieh, D., 321, 328 Lau, L. J., 128, 258, 262
Huang, D. S., 246, 388 LeCam, L., 124, 337
Huber, P. J., 71, 72, 73, 75, 105, 468 Lee, L. F., 293, 318, 319, 382, 387, 392, 394,
Huckfeldt, R. R., 300 396, 397, 400, 402, 403, 404, 405
Hurd, M., 379 Legg, W. E., 293
Hussain, A., 210 Lehmann, E. L., 61, 74, 79
Lermann, S. R., 309, 310, 321, 322, 327, 471
Imhoff, J. P., 173, 192 Li, M. M., 280
Lillard, L. A., 216
Jaeckel, L. A., 78, 79 Lin, L. G., 421
Jaffee, D. M., 403, 473 Lindley, D. V., 465
James, W., 56, 61-62 Loeve, M., 466
508 Name Index
McCall, J. J., 418, 419 Newbold, P., 159

McCarthy, P. J., 418, 472 Neyman, J., 120, 155
MacDonald, G. M., 465 Nickell, S., 447
McDonald, J. F., 365 Nold, F. C., 421,424
McFadden, D., 269, 278, 285, 286, 295, 296, Norden, R. H., 118
298, 299, 300, 302, 303, 304, 305, 306,
321, 322, 328, 408, 471, 472 Oberhofer, W., 186
McGillivray, R. G., 283 Olsen, R. J., 373, 387, 396, 397, 398
McKean, J. W., 79 Orcutt, G. H., 189-190
MacKinnon, J. G., 176, 190
MacRae, E., 431 Paarsch, H. J., 383
MaCurdy, T. E., 217, 218, 257, 473 Pagan, A. R., 145, 206-207, 469
Maddala, G. S., 213, 214, 216, 361, 392, Parke, W. R., 264, 265
394, 395, 396, 397, 402, 404 Payne, C., 67
Malik, H. J., 319 Pearson, E. S., 465
Malinvaud, E., 225, 228 Pesaran, M. H., 148
Mallows, C. L., 52 Pfanzagl, J., 135, 468
Mandelbrot, B., 71 Phillips, P. C. B., 173, 238, 260, 266
Mann, H. B., 87, 88, 89 Plackett, R. L., 192, 319, 463, 464
Manski, C. F., 309, 310, 321, 322, 327, 328, Please, W. W., 465
339, 343, 346, 471, 472 Poirier, D. J., 123, 251
Marcus, M., 462 Polachek, S., 252
Mariano, R. S., 238, 263 Powell, J. L., 251, 252, 284, 382, 383
Marquardt, D. W., 140 Powell, M. J. D., 138, 141
Marsh, L. C., 300 Powers, J. A., 300
Mason, R. L., 66 Prais, S. J., 198, 203
Mayer, T., 51 Pratt, J. W., 293
Mehta, J. S.,218, 221 Prentice, R. L., 412, 451, 474
Miller, L. S., 356 Prescott, E. C., 222
Miller, R. G., 412 Press, S. J., 284,317
Minc, H., 462
Minhas, B. S., 128 Quandt, R. E., 37, 38, 120, 137, 138, 141,
Mizon, G. E., 128, 141, 145 204, 206, 207, 222, 256, 266, 354, 403, 470
Moffitt, R., 350
Monfort, A., 273, 472 Radhankrishnan, R., 67, 68, 466
Montmarquette, C., 387 Radner, R., 356
Mood, A. M., 466 Ramsey, J. B., 120
Morimune, K., 179, 310, 311, 318, 319 Rao, C. R., 58, 87, 89, 90, 91, 92, 93, 94,
Morris, C. L., 62, 63, 64, 66, 67, 68, 466 116, 118, 124, 138, 142, 183, 337, 407,
Mosteller, F., 74 463, 464, 465, 466, 469
Muthen, B., 318 Reece, W. S., 365
Reid, F., 278, 286
Nagar, A. L., 242 Revo, L. T., 285
Nakamura, A., 395 Rice, P., 470
Nakamura, M., 395 Robbins, H., 467
Nelder, J. A., 203 Robins, P. K., 365
Nelson, F. D., 381, 382, 389, 396, 397, 398, Robinson, P. M., 380
402 Roberts, R. B., 394, 395
Nerlove, M., 159, 178, 211, 213, 214, 215, Robertson, J. L., 280
216, 317 Rogers, W. H., 72, 73, 75
Name Index 509
Rosen, S., 404, 405 Strickland, A. D., 470

Rosenberg, B., 222 Strotz, R. H., 229
Rosenzweig, M. R., 365 Stuart, A., 36
Royden, H. L., 19, 117 Subramanyam, K., 125, 279
Rubin, D. B., 375 Swamy, P. A. V. B., 218, 221, 242
Rubin, H., 235 Szego, G., 161
Ruppert, D., 78, 154, 202
Ryzhik, I. M., 474 Talley, W. K., 293
Taylor, W. E., 189, 201, 213, 217, 218, 238,
Sahay, S. N., 242 239
Sant, D. T., 242 Taylor, W. F., 125, 470
Sargent, T. J., 77, 141 Telser, L. G., 431
Savin, N. E., 145, 280 Theil, H., 6, 24, 25, 49, 236, 239, 241
Sawa, T., 53-54, 240, 470 Thisted, R. A., 64, 66, 466
Schatzoff, M., 66 Tiao, G. C., 465
Scheffe, H., 30 Tobin, J., 360, 361, 363, 364, 374, 472
Schlossmacher, E. J., 78 Toikka, R. S., 429, 430, 473
Schmee, J., 378 Tomes, N., 396, 398
Schmidt, P., 37, 317, 319, 380, 381 Toyoda, T., 37
Schwartz, G., 79, 466 Trost, R. P., 392, 394, 396, 397, 404
Sclove, S. L., 60, 64, 67, 68, 466 Trotter, H. F., 138
Scott, E. L., 120, 155 Tsiatis, A. A., 450
Segun, I. A., 193 Tsunimi, H., 470
Shapiro, H., 252 Tukey, J. W., 72, 73, 75
Shapiro, P., 293, 294, 471 Tuma, N. B., 444, 474
Shiller, R. J., 178
Shorrocics, A. F., 418,420
Sickles, R., 37 Uhler, R. S., 300
Silberman, J. I., 293, 312
Silvey, S. D., 142, 466 Van Nostrand, R. C., 466
Singer, B., 412, 474 Vinod, H. D., 67
Sinha, B. K., 279
Small, K. A., 304, 306 Wald, A., 28, 87, 88, 89, 118, 142
Smith, H., 137 Wales, T. J., 128, 373, 375, 390, 472, 473
Smith, K. C., 280 Walker, S. H., 275
Smith, V. K., 470 Wallace, T. D., 210
Sneeringer, C., 285 Wallis, K. F., 178
Solow, R. M., 128 Warner, S. L., 282, 283
Sorbom, D., 217 Watson, G. S., 191, 192, 193, 194, 225
Sowden, R. R., 317 Wedderburn, R. W. M., 203
Spannann, J. M., 309 Wei, C. Z., 467
Spiegelman, R. G., 365 Weiss, Y., 216,470
Spitzer, J. J., 252 Welch, B. L., 36
Srivastava, V. K., 198 Wermuth, N., 66
Stapleton, D. C., 472 West, R. W., 365
Stein, C., 56, 61-62, 470 Westin, R. B., 285, 286, 389
Stephan, S. W., 218, 219, 220 White, H., 117, 148, 199, 200, 370, 465
Stephenson, S. P., 365 White, K. J., 252
Stigler, S. M., 71, 75 Whittle, P., 159, 167, 171, 177
Strawderman, W. E., 65 Wiggins, S. N., 365
510 Name Index
Willis, R. J., 216, 348, 349, 350, 351, 353, Young, D. J., 472
404, 405
Wilson, S., 284 Zacics, S., 46, 138, 465
Wise, D. A., 308, 309, 310, 395 Zarembka, P., 251, 252
Witte, A. D., 365 Zellner, A., 24, 49, 197, 241, 246, 465
Wolfowitz, J., 331, 346, 347
Woodland, A. D., 373, 375, 390, 472, 473
Wu, C. F. J., 375
Wu, D., 388
Subject Index
Absorbing state, 421 Autoregressive integrated moving-average

Admissible, 47, 48 (ARIMA) process, 172
Akaike Information Criterion (AIC): as Autoregressive model
solution to general model selection First-order (AR(1)): asymptotic
problem, 146-147; as solution to problem normality of least squares estimator
of selecting regressors, 52; in choosing in, 174-175; autocovariance matrix of,
optimal significance level, 54-55; in 163-164; consistency of least squares
qualitative response model, 280-281, 313 estimator in, 173-174; definition, 162;
Almon lag model, 178-179 maximum likelihood estimators of
Almost sure convergence, 86 parameters, 175-176; optimal
Almost surely uniform convergence, 106 predictor in, 177; spectral density of,
a-trimmed mean, 71, 73, 78 164
Amemiya's least squares and generalized pth order (AR(p)): autocovariance
least squares, 393 matrix of, 167; definition, 167; least
Asymptotic bias, 95 squares estimators of parameters,
Asymptotic distribution, 92 172-173; moving-average representa-
Asymptotic efficiency, 123-125 tion of, 167-168; properites of, 167-170
Asymptotic expectation: definition, 94; Second-order (AR(2)): autocovariance
relationship with limit of expectation and matrix of, 165-167; definition, 164
probability limit, 93-95 Autoregressive model with moving-average
Asymptotic F-test, 37, 38 residuals (ARMA(p,q)): definition, 170;
Asymptotic likelihood ratio test, 37-38 optimal predictor in, 177; spectral density
Asymptotic mean. See Asymptotic expecta- of, 170
tion
Asymptotically normal, 92 Balestra-Nerlove model, 215-216
Asymptotically unbiased, 95 Bayes estimator: generalized, 48; in classical
Autocovariance: definition, 160; derivation linear regression model, 24-26; proper, 47,
in first-order autoregressive (AR( I)) 48
model, 163; derivation in second-order Bayes's rule, 24, 48, 321
autoreg,ressive (AR(2)) model, 165-166; Bayesian statistics, 23-24, 47, 48-49, 456
spectral density as Fourier transform of, Behrens-Fisher problem, 36
160. See also Autocovariance matrix Berkson's minimum chi-square (MIN X2 )
Autocovariance matrix: characteristic roots, estimator

161, 169; general form, 160; in first-order In binary qualitative response model:
autoregressive (AR( I)) model, 163-164; in asymptotic distribution of, 276-277;
second-order autoregressive (AR(2)) chi-square tests based on, 280-281;
model, 166-167; in pth order autoregres- comparison with maximum likelihood
sive (AR(p)) model, 167; relationship estimator, 278-279; definition, 276;
between that of moving-average model inconsistency of, 278
and that of autoregressive model, 171-172 In first-order Markov model, 415
512 Subject Index
Berkson's minimum chi-square (MIN e) Lindeberg-Feller's, 92; Lindeberg-Levy's,

estimator (continued) 91
In multinomial qualitative response Champernowne process, 420
model: asymptotic distribution of, Characteristic function, 91, 149
291; definition, 290-291; in the Characteristic root, 459
Deacon-Shapiro model, 294-295 Characteristic vector, 459
In two-state Markov model with Chebyshev's inequality, 86-87
exogenous variables, 423-424 Chi-square distribution, 463
Best asymptotically normal (BAN) estima- Choice-based sampling, 320, 325-327,
tors, 124-125 331-332, 338
Best linear predictor, 3 Choice-based sampling maximum
Best linear unbiased estimator (BLUE): likelihood estimator (CBMLE): rela-
constrained least squares as, 21, 23; least tionship with random sampling
squares as, 4, 11-13 maximum likelihood estimator
Best nonlinear three-stage least squares (RSMLE), 321-322; case where the
(BNL3S) estimator: comparison with non- density f of the independent variables
linear full information (NLFI) maximum is known and the selection probability
likelihood estimator, 259, 261; computa- Q is known, 329; case where f is
tion of, 258; definition, 257-258 known and Q is unknown, 330; case
Best nonlinear two-stage least squares where f is unknown and Q is known,
(BNL2S) estimator asymptotic inferiority 337-338; case where f is unknown and
to NLLI maximum likelihood estimator, Q is unknown, 333-337
252, 254-255; computation of. 249; Manski-Lerman weighted maximum
definition, 248 likelihood estimator (WMLE):
Best predictor, 3, 39 asymptotic normality of, 324-325;
Best unbiased estimator, 13, 17-20 consistency of, 323-324; definition,
Beta-logistic model, 350-351 322-323; modification of, 327-328
Better (Best) estimator, 8-11, 40 Manski-McFadden estimator (MME):
Bootstrap method, 135 alternative interpretation of, 336;
Borel field, 83 asymptotic normality of, 331;
Bore! measurable, 84, 467 consistency of, 330-331; definition, 330
Bore! sets, 83 Clark's approximation, 309
Box-Cox maximum likelihood estimator, Classical linear regression model: assumption
250-252 of normality, 13, 27; definition, 2; with
Box-Cox transformation: applications, 141, linear constraints, 20-26; with stochastic
252; definition, 249; nonlinear two-stage constraints, 23-26
least squares (N2LS) estimator in, 250 Cochrane-Orcutt transformation, 189-190
Concentrated likelihood function, 125-127
Conditional likelihood function: definition,
Categorical models. See Qualitative response 474; in Nerlove-Press model, 317; in
models nested logit model, 303
Cauchy distribution, 70, 94, 383 Conditional mean squared prediction error,
Cauchy-Schwartz inequality: applications, 11, 39, 40
98, 130; generalized, 326 Conjugate gradient method, 141
Censored regression models, 360, 364. See Consistency, 95
also Tobit models Constrained least squares (CLS) estimator as
Censoring in duration models: left-, 447-448; best linear unbiased estimator (BLUE), 23;
right-, 434 definition, 21-22; in Almon lag model,
Central limit theorems (CLT): for K-depen- 179; relationship with Bayes estimator
dent sequence, 175; Liapounov's, 92; under stochastic constraints, 25-26
Subject Index 513
Convergence, modes of almost sure, 86; Nonstationary (semi-Markov): applica-

almost surely uniform, 106; in distribution, tions, 444-447; definition, 442;
85; in mean square, 85; in probability, 85; heterogeneity in, 445-446; left-censor-
in probability in the generalized sense, ing in, 447-448; proportional hazards
340; in probability uniformly, 106; in form of, 449-455
probability semi-uniformly, 106; logical Stationary: definition, 433-434; duration
relationship among, 86; theorems as dependent variable of regression in,
concerning, 86-89 438-440; maximum likelihood
Covariance estimator. See Transformation estimates of parameters of, 437-438;
estimator relationship to standard Tobit model,
Covariance matrix of error terms, 185, 186 435; relationship with Poisson
Cox's partial maximum likelihood estimator distribution, 436; right-censoring in,
(PMLE): as maximization of the joint 434; when observations are discrete,
density of order statistics, 451-452; 440-442
asymptotic distribution of, 450, 453-455; Durbin's test: definition, 196-197; equiva-
definition, 449; intuitive motivation of, lence with Rao's score test, 469
450-451 Durbin-Watson test: asymptotic distribution
Cox's test: definition, 147-148; in multivar- of test statistic, 191; definition, 191; exact
iate qualitative response model, 319 distribution of test statistic, 191-193; lower
Cramer-Rao lower bound: applicability and upper bounds for, 193-194; when
asymptotically, 123-124; for regression lagged endogenous variables are present,
coefficients, 17-19; for residual variance, 195-196
18, 19-20; general theorem on, 14-17;
generalization of, 336-337 Edgeworth expansion, 93, 135, 173
Cumulants, 91 Efficient. See Better (Best) estimator
EM algorithm: definition and convergence
Decision theory. See Statistical decision theory properties, 375-376; in Tobit model,
Dependent variables. See Endogenous 376-378
variables Empirical Bayes method, 60-61
Determinant as a criterion for ranking Empirical distribution function, 118, 135
estimators, 10-11 Endogenous sampling. See Choice-based
DFP iteration: definition, 138; in nonlinear sampling
simultaneous equations model, 264 Endogenous variables, 2
Diffuse natural conjugate, 49 Enriched samples, 332
Discrete models. See Qualitative response Equilibrium probabilities: definition,
models 415-416; in Boskin-Nold model, 424-425
Discriminant analysis: in binary qualitative Error components model
response model, 281, 282-285; in Three-error components model (3ECM):
multinominal qualitative response model, asymptotic properties of LS and GLS
299-300 in, 209; definition, 208; definition and
Disequilibrium models, 361, 402-403 properties of transformation estimator
Distributed lag models: Almon, 178-179; in, 209-211; estimation of variances
definition, 177-178; geometric, 178 in, 211
Disturbance. See Error term Two-error components model (2ECM):
Dummy variable regression. See Transfor- as between-group and within-group
mation estimator regression equations, 212-213;
Duration, 428, 434-435 definition, 211-212; definition of GLS
Duration dependence, 445-446. See also estimator in, 213; definition of
State dependence transformation estimator in, 212;
Duration models maximum likelihood estimation of,
514 Subject Index
Error components model (continued) In Toikka's three-state Markov model,

213-214; with endogenous regressors, 429-430
217-218; with serially correlated error, In three-error components model, 211
216-217 In two-error components model, 212,213
See also Balestra-Nerlove model Iterated, 186
Error rate, 284 See also Generalized least squares (GLS)
Error term: additive versus multiplicative, estimator
128, 468; assumptions on density function Fixed effects estimator. See Transformation
of, 152; constant variance (homoscedas- estimator
ticity) 2, 127; normality, 2-3, 13, 129; Fourier transform, 160
serial correlation, 128; serial independence, Full information maximum likelihood
2-3, 13 (FIML) estimator. asymptotic equivalence
Estimable, 58 with 3SLS estimator, 242; asymptotic
Exogenous sampling, 320 normality, 233-234; consistency, 231-233;
Exogenous sampling maximum likelihood definition, 231-232; instrumental variables
estimator (ESMLE): definition, 321-322; interpretation of, 233-234
inconsistency in choice-based sampling
model, 324 Gauss-Markov theorem, 11. See also Best
Exogenous variables, 2 linear unbiased estimator
Exponential family density, 124-125, 203, 279 Gauss-Newton method: definition, 139, 140;
Extremum estimators: asymptotic normality for obtaining BNL3S estimator, 261; for
of, 105, 111-112, 114; consistency of, obtaining NL2S estimator, 470; Hartley's
106-111, 114; definition, 105 algorithm, 140-141; Marquardt's alga
rithm, 140; second round estimator,
F-distribution, 464 139-140, 141
F-test, 28-31, 32-35, 38, 136-137, 184 Generalized Bayes estimator, 48
Feasible generalized least squares (FGLS) Generalized classical linear estimator. See
estimator, 186 Two-stage least squares (2SLS) estimator
In linear regression model with Generalized extreme value (GEV) distribu-
heteroscedasticity: as a BAN estimation, 306
tor, in case of general parametric Generalized extreme value (GEV) model,
heteroscedasticity, 202-203; definition, 306-307
in case of constant variance in subset Generalized least squares (GLS) estimator: as
of sample, 200; definition, in case of best linear unbiased estimator (BLUE),
unrestricted heteroscedasticity, 199; 182; definition, in case of known, positive
definition, when variance is a linear definite covariance matrix, 181-182;
function of regressors, 206; exact definition, in case of singular covariance
distribution, in case of constant var- matrix, 185; examples of equivalence with
iance in subset of sample, 201-202 least squares, 182-184; in linear regression
In linear regression model with serial model with unknown covariance matrix,
correlation: asymptotic deviation from 197, 199; in three-error components
GLS in case of lagged endogenous model, 210-211; in two-error components
variables, 194-195; asymptotic model, 212-213. See also Feasible
equivalence with GLS, 189; generalized least squares (FGLS) estimator,
Cochrane-Orcutt transformation as Partially generalized least squares (PGLS)
procedure for calculating, 189-190; ef- estimator
ficiency relative to least squares, 189 Generalized maximum likelihood estimator,
In Markov model when only aggregate 339, 346-348
data are available, 431 Generalized ridge estimators, 61-66
In seemingly unrelated regression (SUR) Generalized two-stage least squares (G2SLS)
model, 198 estimator, 240-241
Subject Index 515
Generalized Wald test: definition, 145; in model, 298-299; nested multinomial logit
nonlinear simultaneous equations model, model as correction for, 300, 302; tests for,
261-262 299, 302
Geometric lag model, 178 Independent variables. See Exogenous
Goldfeld-Quandt estimator, 204-205, 206 variables
Goldfeld-Quandt peak test for homoscedas- Information matrix, 16
ticity, 207 Initial conditions: in Balestra-Nerlove model,
Goodness of fit. See R2 215, 216; in first-order autoregressive
Gumbel's type B bivariate extreme value (AR(1)) model, 163; in Markov chain
distribution, 300 model, 413-414; in panel data qualitative
response model, 352, 353-354; in
Hadamard product, 462 two-error components model with serially
Hausman's specification test: as test for correlated error, 216, 217
independence of irrelevant alternatives Instrumental variables (IV) estimator.
(IA), 299, 302; as test for normality in definition, 11-12; FIML estimator as,
Tobit model, 381; asymptotic properties 223-224; G2SLS estimator as, 241; in two-
of, 145-146; in nonlinear simultaneous error components model with endogenous
equations model, 265 regressors, 217-218; two-stage least squares
Hazard rate: in duration model, 435; in as asymptotically best, 239-240
Tobit model, 472 Iterative methods. See EM algorithm;
Heckman's two-step estimator: in standard Gauss-Newton method; Method of scoring;
Tobit model, 367-372; in Type 2 Tobit Newton-Raphson method
model, 386-387; in Type 3 Tobit model,
390, 392-393; in Type 4 Tobit model, 396;
in Type 5 Tobit model, 402 Jackknife estimator, 135-136
Heterogeneity: in duration model, 445, Jackknife method, 135-136
446-447; in Markov model, 414; in panel Jensen's inequality, 116
data qualitative response model, 349, Jordan canonical form, 459
350-353. See also Mover-stayer model
Heteroscedasticity: constant variance in a Khinchine's weak law of large numbers
subset of the sample, 200-202; definition, (WLLN), 102
198; general parametric, 202-203; in Kolmogorov laws of large numbers (LLN), 90
standard Tobit model, 378, 379-380; tests Koyck lag model, 178
for, 200, 201, 203, 206-207; unrestricted, Kronecker product, 462
198-200; variance as an exponential
function of the regressors, 207; variance as
a linear function of regressors, 204-207 L estimators, 73-74, 77-78
Hildreth-Houck estimator in a heteroscedas- Li, estimators, 72, 73, 77
tic regression model: definition, 204, 205, Lag operator, 162
206; modifications, 205-206, 469 Lagrange multiplier test. See Score test
Milder's inequality, 19 Lagrangian interpolation polynomial, 469
Homoscedasticity, 2 Laplace distribution, 70
Laws of large numbers (LLN): Ithinchine's,
Idempotent, 460 102; Kolmogorov's number 1 and number
Identification: in a linear simultaneous 2, 90; Markov's, 467; strong, 90; weak, 90
equations model, 230; in a nonlinear si- Least absolute deviations (LAD) estimator
multaneous equations model, 256 in classical regression model, 152-154; in
Incidental parameters. See Nuisance standard Tobit model, 382-383. See also
parameters Median
Independence of irrelevant alternatives (IA): Least squares (LS) estimator
as characteristic of multinomial logit In autoregressive model: asymptotic
516 Subject Index
Least squares (LS) estimator (continued) Lebesgue measure, 83, 106, 113, 124, 467
normality in first-order case, 174-175; Lebesgue-Stieltjes integral, 467
asymptotic normality in pth order Liapounov central limit theorem (CLT),
case, 173; consistency in first-order 90-91,92
case, 173-174; small sample proper- Likelihood function. See Concentrated
ties, 173 likelihood function; Conditional likelihood
In classical linear regression model: as function; Maximum likelihood estimator
best linear unbiased estimator (MLE)
(BLUE), 11-13; as best unbiased Likelihood ratio test
estimator under normality of error As test for homoscedasticity, 201
terms, 17-20; asymptotic normality of, In nonlinear regression model, 144, 145
96-98, 99; consistency of, 95-96; Under general parametric hypotheses:
definition, 4-5; equivalence under asymptotic distribution of, 142-144;
normality of, with maximum definition, 142; small sample
likelihood estimator (MLE), 13; properties of, 145
geometric interpretation of, 5-6; mean Under linear hypotheses on a linear
and variance of, 7-8; of a linear model: definition, 28-32, 32-34, 38;
combination of regression parameters, relationship with Wald test and score
7, 58; of a subset of regression test, 144-145
parameters, 6-7; unbiased alternative Limit distribution, 85, 92
to estimate of variance of error terms, 8 Limited dependent variables model. See
In duration model with one completed Tobit model
spell per individual, 439 Limited information maximum likelihood
In geometric lag model, 178 (LIML) estimator: asymptotic distribution,
In linear regression model with general 237-238; asymptotic equivalence with
covariance matrix of error terms: 2SLS estimator, 236, 238; definition,
asymptotic normality in case of 235-236; exact distribution, 238-239;
known covariance matrix, 185; Fuller's modification, 238
asymptotic normality in case of serial Limited information model, 234-235
correlation, 187-188; consistency, Linear regression model, with general covar-
184-185; covariance matrix, 182; ex- iance matrix of error terms, 184
amples of equivalence with GLS, Linear simultaneous equations model,
183-184; inconsistency in presence of 228-229
lagged endogenous variables and serial Lindeberg-Feller central limit theorem
correlation, 194; relative efficiency as (CLT), 90-91, 92
compared with GLS, 182-183 Lindeberg-Levy central limit theorem (CLT),
In Markov model when only aggregate 90-92
data are available, 431 Linear constraints: as testable hypotheses,
In standard Tobit model: biasedness 27; form of, 20; stochastic, 23-24
when all observations used, 367-368; Logistic distribution: Gumbel's bivariate,
biasedness when positive observations 319; Plackett's bivariate, 319; relationship
used, 367 with normal distribution, 269
In Toikka's three-state Markov model, Logit model
429, 430 Binary: definition, 268-269; global
See also Constrained least squares (CLS) concavity of likelihood function, 273
estimator; Generalized least squares Multinomial: as result of utility-maxi-
(GLS) estimator. mizing behavior, 296-297; definition,
Least squares predictor, 39-40 295; global concavity of likelihood
Least squares residuals, 5, 21, 32-33 function, 295-296; independence of
Lebesgue convergence theorem, 117 irrelevant alternatives (IIA) in, 298
Subject Index 517
Logit model (continued) and stationary first-order Markov model,

Multivariate: definition, 314; fundamen- 417; in Hsiao's random coefficients model,
tal difference between multivariate 220-221; in linear regression model with
probit model and, 318 general parametric heteroscedasticity,
Multivariate nested, 313-314 202-203; in linear regression model with
Nested: definition, 300-302, 303; serial correlation, 190-191; in Markov
estimation, 303-304; multiple-level model when only aggregate data are
forms, 305-306 available, 433; in multinomial qualitative
Ordered, 292-293 response model, 287, 288-289; in
Sequential, 310 two-error components model, 213-214; in
Universal, 307 two-state Markov model with exogenous
Logit transformation, 278 variables, 422, 424; inconsistency in
Log-linear model: definition, 314, 315; certain models, 120; iterative methods for
estimation of, 317; saturated, 315; obtaining, 137-139; second-order
unsaturated, 315 efficiency, 124-125, 279; under constraints
Log Weibull distribution, 296 on the parameters, 142. See also Box-Cox
maximum likelihood estimator; Choice-
based sampling; Cox's partial maximum
M estimators, 72-73, 75-77, 105. See also likelihood estimator (PMLE); Exogenous
Extremum estimators sampling maximum likelihood estimator;
Mallow's criterion: definition, 52; optimal Full information maximum likelihood
critical value of F-test implied by, 55 (F1ML) estimator; Generalized maximum
Markov chain model: definition, 412-413; likelihood estimator in qualitative
duration in, 428; estimation of, when only response model; Limited information
aggregate data are available, 430-433; maximum likelihood (LIML) estimator;
homogeneous, 414; hypotheses tests in, Nonlinear full information (NLFI)
417; maximum likelihood estimates of pa- maximum likelihood estimator; Nonlinear
rameters, 417, 422; minimum chi-square limited information (NLLI) maximum
(MIN ,e) estimates of parameters of, 415, likelihood estimator; Probit maximum
423-424; mover-stayer model, 418; likelihood estimator; Random-effect probit
multi-state (Toikka's model), 428-430; maximum likelihood estimator; Random
nonlinear generalized least squares sampling maximum likelihood estimator
estimates of parameters, 414-415; second- (RSMLE); Tobit maximum likelihood
order model, 420; stationary, 414; estimator
two-state (Boskin-Nold model), 421-428 Maximum score estimator: consistency,
Markov matrix, 413 340-343, 344-345; definition, 339,
Markov's law of large numbers (LLN), 467 343-344; relationship with generalized
Maximum likelihood estimator (MLE): maximum likelihood estimator, 347-348
asymptotic efficiency, 123-124; asymptotic Mean squared error, 8, 40
normality, 120-121; consistency, 115-116, Measurable function, 467
118; definition, 115; in autoregressive Measure, 467
moving average model (ARMA (p,q)) Measure space, 467
model, 172; in Balestra-Nerlove model, Median
216; in binary qualitative response model, Population, 148
271-273, 274-275, 278-279; in Boskin- Sample: asymptotic normality, 149-151;
NoId model, 425-427; in classical linear re- asymptotic variance under normal,
gression model, 13-14, 17-19, 21, 37-38, Laplace, and Cauchy distributions,
118-119, 121-123; in duration model, 70-71; consistency, 150; definition, 70,
437-439; in first-order autoregressive 73, 149. See also Least absolute
(AR(I)) model, 175-176; in homogeneous deviations (LAD) estimator
518 Subject Index
Method of scoring: conditions for equiva- Nonlinear least squares (NLLS) estimator:
lence with Gauss-Newton method, 203; asymptotic normality of, 132-134;
definition, 138; equivalence with nonlinear consistency of, 129-130; definition,
weighted least squares (NLWLS) iteration, 127-129; equivalence of, with MLE under
274-275, 289-290; in random coefficients normality, 129; in Markov model when
model, 221; when variance is linear only aggregate data are available, 431, 433;
function of regressors, 206 in standard Tobit model, 372-373, 378;
Mills' ratio, 472 inconsistency in nonlinear simultaneous
Minimax estimator, 47, 48 equations model, 245; under general
Minimax regret, 47 parametric constraints, 144; under linear
Minimini principle, 51 constraints, 136-137
Minimum chi-square (MIN x2) method, 275. Nonlinear limited information (NLLI)
See also Berkson's minimum chi-square maximum likelihood estimator asympto-
(MIN x2) estimator tic covariance matrix of, 254; definition,
MINQUE, 469 252-254; inconsistency of, when normality
Mixed estimator under stochastic constraints, fails to hold, 255; iterative methods for
25 solving for, 254
Mixture of normal distributions, 72, 77, Nonlinear regression model, 127-128
119-120 Nonlinear simultaneous equations models:
Model selection problem, 146 full information case, 255-256; limited
Modified nonlinear two-stage least squares information case, 252-253
(MN2LS) estimator: definition, 254; Nonlinear three-stage least squares (NL3S)
inconsistency when normality fails to hold, estimator: asymptotic normality of, 257;
255 consistency of, 257; definition (Amemiya),
Moore-Penrose generalized inverse, 112, 466 257; definition (Jorgenson and Laffont),
Mover-stayer model, 418-419 257. See also Best nonlinear three-stage
Moving-average (MA) models: autocovar- least squares (BNL3S) estimator
iance matrix of, 171; first-order (MA(1)), Nonlinear two-stage least squares (NL2S)
171-172; spectral density of, 170-171 estimator: asymptotic normality, 247-248;
Multicollinearity, 56, 59 consistency, 246-247; definition, 246, 250;
in case of Box-Cox transformation, 250,
251-252. See also Best nonlinear two-stage
Newton-Raphson method: applications, 141; least squares (BNL2S) estimator; Modified
as method of calculating FIML estimator, nonlinear two-stage least squares (MNL2S)
234; definition, 137; DFP iteration, 138; estimator
for obtaining NLFI maximum likelihood Nonlinear weighted least squares (NLWLS)
estimator, 264; method of scoring as estimatoc in binary qualitative response
means of finding MLE, 138; quadratic model, 274-275; in duration model with
hill-climbing, 138; second-round estima- one completed spell per individual, 440; in
tor, 137, 139 Markov model when only aggregate data
Nonlinear full information (NLFI) maxi- are available, 432, 433; in multinomial
mum likelihood estimatoc asymptotic qualitative response model, 289-290; in
properties, 259, 261; definition, 259; multi-state Markov model with exogenous
inconsistency when error terms are not variables, 428-429; in standard Tobit
normal, 259-260; iterative method, 260-261 model, 372-373, 472; in two-state Markov
Nonlinear generalized least squares (NLGLS) model with exogenous variables, 423. See
estimator: in first-order Markov model, also Method of scoring
415; in Markov model when only Nonnested models, 147-148
aggregate data are available, 431. See also Nonnormality: in standard Tobit model,
Nonlinear weighted least squares 378, 380-382, 383; robust estimation
(NLWLS) estimator under, 70-71
Subject Index 519
Normal discriminant analysis. See Discrimi- response model, 307-310; in multivariate

nant analysis qualitative response model, 317-318; in
Nuisance parameters, 120 Tobit model, 366, 386. See also Probit
model
Probit model
Odds ratio: in Bayesian solution to selection- Binary: definition, 268, 269; global
of-regressors problem, 48-49; in binary concavity of likelihood function,
logit model, 278 273-274
One-factor model in panel data qualitative Multinomial: definition, 307-308;
response model: definition, 352-353; independent versus nonindependent,
random-effect probit maximum likelihood 308; iterative methods for calculating
estimator in, 354; transformation estimators in, 309
estimator in, 353-354 Multivariate: definition, 317, 318;
Optimal significance level, 52-55 fundamental difference between mul-
Order condition of identifiability, 230-231 tinomial logit model and, 318
Order relationship, 90-91 Ordered, 292-293
Overidentification, 231 Sequential, 310
Probit transformation, 277
Panel data qualitative response model: Projection matrix, 14
definition, 348; heterogeneity, 349-353; Proportional hazards model, 449-455
one-factor, 352-354; specification of initial
conditions, 352, 353-354; true state Quadratic hill-climbing, 138
dependence, 349, 351-353 Qualitative response models, 267
Parke's derivative free algorithm, 264 Binary: definition, 268; hypotheses tests,
Partially generalized least squares (PGLS) 280-282; prediction of aggregate
estimator, 199-200 proportion, 285-286
Posterior distribution, 24 Multinomial: definition, 287; hypotheses
Posterior probability, 48 tests, 291-292; ordered, 292-293;
Posterior risk, 47 unordered, 292, 293
Prediction Multivariate, 311
In classical linear regression model: best Quantal response models. See Qualitative
linear predictor, 3; best predictor, 3, response models
39; conditional mean squared
prediction error, 39, 40; least squares
as best linear unbiased predictor, R estimators, 74, 78-79
39-40; unconditional mean squared R2: definition, 6; test for equality with zero,
prediction error, 40 31; Theil's correction, 49-51, 54, 55;
In nonlinear simultaneous equations weakness, 46
model, 262-264 Random coefficients models (RCM's):
In time series model, 177 Hildreth-Houck and error components
Prediction criterion: definition, 51; optimal models as special cases of, 204, 218;
critical value of F-test implied by, 54, 55 Hsiao's model, 220-221; Kelejian and
Pre-test estimator, 67 Stephan model, 218-220; Swamy-Mehta
Principal components, 58 model, 221-222; Swamy's model, 221;
Principal components estimator, 58-60 varying parameter regression model, 222
Probability measure, 82 Random-effect probit maximum likelihood
Probability of correct classification (PCC), 284 estimator, 354
Probability space, 82 Random sampling maximum likelihood
Probit maximum likelihood estimator: in estimator (RSMLE), 321-322, 329
binary qualitative response model, 271, Random variables, 81, 83
273-274; in multinomial qualitative Rank condition of identifiability, 230
520 Subject Index
Rao's score test. See Score test a-algebra, 82

Reduced form equations, 229 Spectral density: conditions for existence and
Regret, 47 continuity, 169; definition, 160-161; of
Reverse least squares estimator, 42, 99-101 autoregressive moving average (ARMA
Ridge regression: generalized ridge estima- (p,q)) process, 170; relationship between
tors, 63-66; relationship with empirical that of moving-average model and that of
Bayes method, 60-61; relationship with autoregressive model, 170-171; relation-
Stein-type estimators, 56; ridge estimators, ship with characteristic roots of autocovar-
60-66; ridge trace method, 60 iance matrix, 161
Riemann integal, 84 Standard linear regression model. See
Riemann-Stieltjes integral, 84-85 Classical linear regression model
Risk, 47 State dependence, 349, 351. See also
Robust estimation: concept of, 70-72; in Duration dependence
regression case, 75-79; L estimator, 73-74, Stationary time series: as infinite sum of
77-78; L 73, 77; M estimator, cycles with random coefficients, 161; defi-
72-73, 75-77, 105; median, 70-71, 73, nition (strictly), 159; definition (weakly),
148-151; R estimator, 74, 78-79; trimmed 159
mean, 73, 78; Winsorized mean, 73. See Statistical decision theory, 46-48
also Least absolute deviations (LAD) Stein's estimator, 61-62
estimator Stein's modified estimator, 63
Stein's positive rule estimator, 63
Sample space, 81-82 Stieltjes integral, 84-85, 467
Schwartz's criterion, 79 Stochastic order relationship. See Order
Score function, 339 relationship
Score test: as test for homoscedasticity, Stratified sampling, 471
206-207; asymptotic distribution in general Strictly stationary time series, 159
case, 142, 144; asymptotic distribution in Strong convergence, 467. See also Almost
nonlinear simultaneous equations model, sure convergence
145; definition, 142; equivalence with Strong laws of large numbers (LLN), 90
Durbin's test, 469; in linear model with Strongly consistent, 95
linear constraint, 144-145; in nonlinear re- Structural change, tests for, 31-35, 36-38
gression model with normality, 144; Structural equations, 229
relationship with Lagrange multiplier test, Student's t distribution, 463
142 Superefficient estimators, 124
Second-order efficiency of maximum Survival models. See Duration models
likelihood estimator, 125. See also Best Switching regression model, 120, 411
asymptotically normal (BAN) estimator
Seemingly unrelated regression (SUR) 1-test, 27-28, 34-35, 36-37, 136, 184, 465
model, 197-198 Test for equality of variances, 35. See also
Selection of regressors: Bayesian solution, Heteroscedasticity
48-49; selection among nested models, 45, Tests of general parametric hypotheses:
52-55; selection among nonnested models, generalized Wald statistic, 145; Hausman's
46-52 specification test, 146-147; likelihood ratio
Separate families of hypotheses, 147 test, 142-145; score test, 142, 144-145;
Serial correlation of error terms: as first-order Wald test, 142-145
autoregressive (AR( I)) process, 188-195; Tests of linear hypotheses: F-test (likelihood
definition, 186; in linear regression model ratio test), 28-31, 32-34, 184, 465;
with lagged endogenous variables, generalization of F-test to nonlinear
194-197; properties of least squares (LS) model, 136-137; generalization oft-test to
estimator in case of, 184-185, 187-188 nonlinear model, 136; 1-test, 27-28, 34-35,
Subject Index 521
Tests of linear hypotheses (continued) Two-stage least squares (2SLS) estimator:

36-37, 184, 465; tests for structural asymptotic distribution, 236-237;
change, 31-35, 36-38 asymptotic equivalence with L1ML
Theil's corrected R 2: appraisal of, 49-51, 55; estimator, 238; definition, 236; exact
optimal critical value of F-test implied by, distribution, 238-239; interpretations,
54 239-240. See also Generalized two-stage
Three-stage least squares (3SIS) estimator, least squares (G2SLS) estimator; Nonlinear
241-242. See also Nonlinear three stage two-stage least squares (N2SLS) estimator
least squares (NL3S) estimator Two-step method of estimation: in multivar-
Tobit maximum likelihood estimator: as an iate nested logit model, 303-304; in nested
equilibrium solution of the EM algorithm, logit model, 314. See also Heckman's
376-378; asymptotic properties, 373; two-step estimator
consistency under serial correlation, 380; Type I extreme value distribution, 296
definition, 373; global concavity of
likelihood function, 373-374; inconsistency Unconditional mean squared prediction
under heteroscedasticity, 379-380; error, 40
inconsistency under nonnormality, 380-381 Underidentification, 231
Tobit model: Amemiya's consistent Uniform convergence of a sequence of
estimator, 374-375; applications, 364, 365, random variables: almost surely uniform,
387-389, 391-392, 394-395, 397-399, 106; in probability semi-uniformly, 106; in
400-408; as result of utility maximizing probability uniformly, 106
behavior, 362; classifications, 383-384; Uniformly smaller risk, 47
consistency of estimators under serial cor-
relation, 378-379, 380; definition, 360; Wald's test: asymptotic distribution in
disequilibrium models, 402-403; general nonlinear simultaneous equations model,
simultaneous equations forms, 385, 145; asymptotic distribution, 142, 144;
392-393, 396-397; global concavity of definition, 142; in linear model with linear
likelihood function, 373-374; inconsistency constraints, 144-145; in nonlinear
of estimators, 378-381; tests for nonnor- regression model with normality, 144
mality, 381-382; Tobin's initial estimator, Weak convergence, 467. See also Consist-
373. See also Heckman's two-step ency; Convergence in probability
estimator; Least absolute deviations (LAD) Weak laws of large numbers (LLN), 90
estimator; Least squares (LS) estimator, Weakly consistent, 95
Nonlinear least squares (NLLS) estimator; Weakly stationary time series, 159
Nonlinear weighted least squares Weibull distribution, 445
(NLWLS) estimator; Probit maximum Weighted least squares (WLS) estimator,
likelihood estimator; Tobit maximum like- 371-372. See also Feasible generalized least
lihood estimator squares (FGLS) estimator, Generalized
Toeplitz form, 160 least squares (GLS) estimator, Nonlinear
Trace, as a criterion for ranking estimators, weighted least squares (NLWLS) estimator
10, 11 Welch's method, 36-37
Transformation estimator: in Balestra- Winsorized mean, 73
Nerlove model, 215-216; in discrete case,
353-354; in Hsiao's random coefficients
model, 220; in three-error components
model, 209-211; in two-error components
model, 212-213
Transition probabilities, 413
Truncated regression model, 360, 364. See
also Tobit models

Advanced Econometrics - 1985 - 1era Edición - Amemiya

Uploaded by

Copyright:

Available Formats

Advanced Econometrics - 1985 - 1era Edición - Amemiya

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

Advanced Econometrics - 1985 - 1era Edición - Amemiya

Uploaded by

Copyright:

Available Formats

What are some of the main features of the book discussed in the preface?

What are some of the main features of the book discussed in the preface?

What prerequisites are recommended for using the book as a textbook?

What prerequisites are recommended for using the book as a textbook?

ADVANCED

"The book provides an excellent overview of mod-

Advanced Econometrics is both a comprehensive

Harvard University Press Cambridge, Massachusetts 1985

This book is printed on acid-free paper, and its binding materials

Library of Congress Cataloging in Publication Data

This book is intended both as a reference book for professional econometri-

1 Classical Least Squares Theory 1

Appendix 1 Useful Theorems in Matrix Analysis 459

1.1 Linear Regression Model

Then we can always write

1.1.4 Matrix Notation

If we want to focus on the columns of X, we can write X =

1.2 Theory of Least Squares

1.2.1 Definition of Least Squares Estimators of ft and a2

Figure 1.1 Orthogonal decomposition of y

(that is, K= 1 and T= 2), the decomposition (1.2.6) can be illustrated as in

1.2.2 Least Squares Estimator of a Subset of #

tor and /.3. 2 is a K2 vector such that ICI K2 = K. Partition X conformably as

X = (X 1 , X2 ). Then we can write X' Xfi = X ' y as

= NMIX2 rIX2M1Y, (1.2.13)

mal number of linearly independent columns of X2, provided that VIMIXI is

1.2.3 The Mean and Variance of /land 62

Using the properties of the projection matrix given in Theorem 14 of Appen-

1.2.4 Definition of Best

DEFINITION 1.2.2. Let 0 and 0* be estimators of a vector parameter 0. Let A

more efficient) than 0* if

Suppose that 02 is known; define another estimator of 0, by

1.2.5 Least Squares as Best Linear Unbiased Estimator (BLUE)

estimator of p 'ft is p ' (X' X)-1X' y. We therefore have a constructive proof of

1.3 Model 1 with Normality

1.3.1 Maximum Likelihood Estimator

E(u' Mu)2 = Eu ' Muu' Mu (1.3.5)

which readily follows from Theorem 2 of Appendix 2 because M is indempo-

1.3.2 Cramer-Rao Lower Bound

Proof Define P = E(0 0)(-6 , Q = E("e exa log Lion and

=E [ea log L] by assumption A

Therefore (1.3.8) follows from (1.3.10), (1.3.11), and assumption B.

respectively. The equivalence between assumptions A and A' follows from

in other words, the operations of differentiation and integration can be inter-

, I 1 f(z, 0 + hh) f(z, 0)

where 0* is between 0 and 0 + h. Next, we can write the last integral as

1.3.3 Least Squares Estimator as Best Unbiased Estimator (BUE)

From (1.3.15) and (1.3.16) we can immediately see that assumption A is

where we have written X as x, inasmuch as it is a vector. Therefore, by Holder's

f Illafl ' (1I-2 Y

1.3.4 The Cramer-Rao Lower Bound for Unbiased Estimators of a 2

1.4 Model 1 with Linear Constraints

1.4.1 Constrained Least Squares Estimator (CLS)

1.4.2 An Alternative Derivation of the Constrained Least

Therefore A = A if the matrix in the square bracket above is nonsingular. But

singularity of X ' X implies nonsingularity of R'X'XR. Because the matrix

1.4.3 Constrained Least Squares Estimator as Best Linear

1.4.4 Stochastic Constraints

= Q(Q'Q)-1c R(R ' X ' XR)-112' X' XQ(Q'Q)-1c.

where fi is given in (1.4.11).