Statistical Analysis of Econometric Models

Author(s): Arnold Zellner

Source: Journal of the American Statistical Association , Sep., 1979, Vol. 74, No. 367
(Sep., 1979), pp. 628-643
Published by: Taylor & Francis, Ltd. on behalf of the American Statistical Association

Stable URL: https://www.jstor.org/stable/2286981

REFERENCES
Linked references are available on JSTOR for this article:
https://www.jstor.org/stable/2286981?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to
digitize, preserve and extend access to Journal of the American Statistical Association

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms

ARNOLD ZELLNER*
In this article, a summary of some research bearing on the statistical
analysis of econometric models is reviewed. Many estimation, testing,
and prediction techniques used in econometrics have just large-
sample justifications. Selected Bayesian inference results relating to
econometric models are reviewed. On the problem of constructing
econometric models, an approach that is a blend of traditional
econometric and modern time series analysis techniques is described.
Many statistical problems requiring further analysis are noted. It is
concluded that better solutions to these problems, better data, more
sophisticated use of economic theory, application of more rigorous
diagnostic checks, including forecasting checks and use of well-
designed simulation experiments, probably will produce improved
macroeconometric models.
KEY WORDS: Econometric statistics; Econometric modeling;
Bayesian econometrics estimation; Prediction; Hypothesis testing.
1. INTRODUCTION
Substantial progress has been made in developing
data, concepts, and techniques for the construction and
statistical analysis of econometric models. Comprehensive
data systems, including national income and product ac-
counts, price, wage and interest rate data, monetary
data, and many other measures, have been developed for
almost all countries. In many cases, annual measure-
ments have been augmented by quarterly and monthly
measurements of a broad array of economic variables. In
recent years, scientifically designed sample surveys have
been employed to expand data bases of a number of
countries. While research continues to improve data
bases, we must recognize that the work that produced
our current, extensive data bases is a major accomplish-
ment in the field of scientific measurement and enables
economic analysts to avoid the charge of "theory without
measurement."
In reviewing the development of concepts for the
statistical analysis of econometric models, it is very easy
to forget that in the opening decades of this century a
major issue was whether a statistical approach is ap-
propriate for the analysis of economic phenomena. For-
tunately, the recognition of the scientific value of
sophisticated statistical methods in economics and busi-
ness has buried this issue. To use statistics in a sophis-
ticated way required much research on basic concepts of
* Arnold Zellner is H.G.B. Alexander Professor of Economics and
Statistics, Graduate School of Business, University of Chicago, 1101
East 58th St., Chicago, IL 60637. Research was financed by NSF
Grant SOC 7305547 and by income from the H.G.B. Alexander
Endowment Fund, University of Chicago. The author is grateful to
David C. Hoaglin, Stephen E. Fienberg, and two anonymous
referees for helpful comments. An earlier version of this article was
presented to the American Statistical Association's meeting in
Chicago, August 1977.
628
This content downloaded from
86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
Statistical Analysis of
Econometric Models
econometric modeling that we take for granted today. It
was necessary to develop fundamental concepts such as
complete model, identification, autonomous structural
relationships, exogeneity, dynamic multipliers, and sto-
chastic equilibrium, to name a few, that play an impor-
tant role in linking statistical analyses and economic
theory.
Many statistical estimation, testing, and prediction
techniques have been developed for use in connection
with many different kinds of econometric models, in-
cluding linear and nonlinear interdependent structural
models, models involving qualitative and quantitative
variables, models with time series complications, models
for combined time series and cross-section data, and
models with random parameters. This research on sta-
tistical techniques and computer programs implementing
them, a joint product of statisticians and econometricians,
has been extremely important in the development of
modern econometric modeling techniques.
Given this past record of solid achievement in the
areas of measurement, concepts, and statistical tech-
niques, it is relevant to ask how current statistical
analyses of econometric models can be improved so as to
yield models with better forecasting and policy-analysis
performance. To answer this question, I shall first try,
in Section 2, to summarize the main features of current
or traditional econometric modeling techniques. Tradi-
tional econometric analyses, like many statistical analy-
ses, tend to concentrate attention mainly on given
models and not on procedures for discovering and re-
pairing defects of proposed models. Section 3 describes
an approach that emphasizes the latter aspect of econo-
metric model construction and is a blend of traditional
econometric techniques and modern time series tech-
niques. While this approach, called structural econometric
modeling time series analysis (SEMTSA), is not a
panacea for all problems, it probably will be helpful in
improving the quality of econometric models. A con-
cluding section considers prospects for the future.
2. THE TRADITIONAL ECONOMETRIC
MODELING APPROACH
In this section, I shall attempt to characterize tradi-
tional econometric modeling techniques, to provide a
? Journal of the American Statistical Association
September 1979, Volume 74, Number 367
Invited Paper
 Zeliner: Statistical Analysis of Econometric Models 629

summary of statistical procedures used in econometric Model-Building Activities:

modeling, and to describe some of the statistical needs of A Schematic Representation

traditional econometric model builders.

Stage 1: Statement of Objectives

Feasibility Studya
Choice of Methodologyb
2.1 Overview of the Traditional Approach

The schematic diagram in the figure represents, in

broad outline, the activities of many econometric Stage 2: Preparatory Work
modelers. Whatever the problem, there is usually a state- Review of Literature

Preliminary Data Analysis

ment of objectives, although, at times, the statement may
not be so clear-cut and specific as could be desired. Some-
times, objectives are so ambitious that, given our present Stage 3: Formulation of
knowledge, data, and techniques, they may be practically Initial Model (MO}
unattainable. The next steps in traditional econometric
modeling involve a review of the theoretical and empirical Mathematical Analysis Statistical Analysis of
Computer Simulation Old and New Data
literature bearing on the objectives of a modeling project,
preparation of a data base, and preliminary data analysis. New Research
The objective of these activities is the formulation of an Findings

initial variant of an econometric model. Unfortunately,

most econometrics and statistics texts are woefully silent
Reformulation of
on the basic methodology of how to formulate an initial Model MO
variant of a model. General prescriptions, such as "use
relevant economic theory" and "formulate as simple a
model as possible," are valuable guidelines. Often the 1st IterationO: Formulation of
relevant economic theory does not yield precise informa- Model M,

tion regarding functional forms of relationships, lag

Mathematical Analysis < Statistical Analysis of
structures, and other elements involved in a stochastic Computer Simulation n Old and New Data
specification of a model. Further, model simplicity has
yet to be defined in a completely satisfactory manner. New Research
Findings
Still, it is worthwhile to emphasize the importance of
using elements of economic theory, other outside informa-
tion, and simplicity in formulating an initial variant of Reformulation of

a model. For example, models that imply unexploited Model M,

profit opportunities probably will be unsatisfactory be-

cause exploitation of such profit opportunities will gen-
2nd Iteration: Formulation of
erally upset properties of the proposed model that
| Model M2 |
contains them.
Once an initial variant of a model, denoted by Mo, has
been formulated, it is traditionally subjected to a number NOTE: For further discussion of this approach to modeling, see Hamilton et al.

of mathematical, statistical, computer simulation, and (1969) and Zellner (1970).

a It is assumed that this study shows the project to be feasible.
judgmental checks. These include simple mathematical b It is assumed that a modeling approach is selected.

checks on the number of equations and number of endog- e The iterative procedure may disclose problems in the original formulation of
goals, feasibility, and methodology so that refining and reformulation of the effort
enous variables, consistency of variables' units of mea- may not be confined solely to the model itself. Also, it is possible that other feed-

surement, conditions for parameter identification, and back loops, not shown in the figure, may be important in the process of converging
on a satisfactory variant of a model.
compatibility with results from mathematical economic
theory. Computer simulation experiments are often em-
ployed to gain information about local and global dy- work is undertaken to reformulate Mo and to produce a
namic and other properties of Mo. Statistical checks in- new variant of the model, M1i. Then M1 is subjected to
volve formal hypothesis testing procedures, forecasting the battery of checks mentioned previously. This process
tests, residual analysis, data evaluation, and other of checks and reformulation continues, using as much new
diagnostic checks. In evaluating the adequacy of Mo, a data as possible, until a satisfactory version of the model
good deal of judgment or prior information is employed,is obtained, satisfactory in the senses that it passes
usually informally. For example, the algebraic signs and diagnostic checks satisfactorily and accomplishes the ob-
magnitudes of parameter estimates are reviewed to jectives of the model-building project.
ascertain whether they are compatible with results pro- In connection with realizing the objectives of a model-
vided by economic theory, by previous studies, and by building project, it is useful to havre formulated as simple
judgmental information. a model as possible. If the objectives require the model
If Mo is found to be inadequate in certain respects, builder to capture much detail, the model probably will

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
630 Journal of the American Statistical Association, September 1979
be large, but with care in model building it can still be
sophisticatedly simple. Large and simple models seem
preferable to large and complicated models. In fact, a
very disturbing feature of some large, complicated models
in the literature is that it is not known whether they have
a unique solution or many solutions.
In the past, model builders have used some or some-
times all the elements of the approach described before,
but generally have not been vigorous enough in applying
the various checks. Mathematical analyses have often
been superficial and incomplete. Simulation experiments
have not been very extensive or well designed in general.
Statistical checks on the quality of data and on specifying
assumptions have not been pursued vigorously enough.
The relationship of models' properties to relevant eco-
nomic theory has not been examined thoroughly in a
number of instances. Finally, many econometric model
builders have not stressed simplicity enough. Some cur-
rently operating econometric models are highly complex
systems of a hundred or more nonlinear stochastic differ-
ence equations with hundreds of parameters that have
to be estimated from highly aggregated time series data.
Failure to take account of Ockham's Razor, the Jeffreys-
Wrinch Simplicity Postulate, and the Principle of
Parsimony in formulating econometric models has had
very serious consequences in much traditional economet-
ric model building. See Jeffreys (1957, 1967) for evidence
of the importance of simplicity in science.
These criticisms of traditional econometric models
have to be tempered, however, because many methodo-
logical techniques needed in a sensible model-building
process are not yet available. Good formal sequential
testing procedures for model construction remain to be
developed. Even for a given structural econometric
model, exact finite-sample tests and optimal finite-
sample estimates and predictors have not been available.
Good or optimal designs for simulation experiments re-
main to be derived. The problems of missing and imper-
fect data have not been completely solved. Tried and
tested economic theory dealing with stochastic markets,
dynamic reactions, and a number of other important
issues has not been available. Thus econometric model
building has been a mixture of economic and statistical
theory and empirical practice. It is probable that such
interaction between theory and practice will produce
improvements in both.
To illustrate elements of recent statistical practice in
traditional econometric model building, I next review
some estimation, testing, and prediction techniques and
provide some indications of current developments and
open problems.
2.2 Statistical Estimation Problems
Learning the values of parameters appearing in struc-
tural econometric models (SEM's) is important in
checking the implications of alternative economic theories
and in using SEM's for prediction, forecasting, and
This content downloaded from
86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
policy making. Thus, econometric research has placed a
heavy emphasis on statistical estimation problems.
2.2.1 Asymptotically Justified Estimation Procedures.
Research since the 1940's has resulted in a greatly en-
hanced understanding of estimation problems associated
with SEM's and a relatively large number of operational
procedures for obtaining consistent, asymptotically
normally distributed, and efficient parameter estimates
for some or all parameters of linear and nonlinear, static,
and dynamic SEM's with serially uncorrelated or serially
correlated errors. These procedures, which are discussed
at length in econometric textbooks and the econometric
literature, include maximum likelihood, two- and three-
stage least squares, 32-class, double 32-class, instrumental-
variable, nonlinear maximum likelihood, nonlinear two-
and three-stage least squares, and other procedures.
Further, many of the parameter estimates produced by
such procedures approximate Bayesian posterior means
of parameters in large samples. A most important result
of this research, aside from providing asymptotically
justified estimation procedures, has been to rule out a
number of proposed inconsistent and/or asymptotically
inefficient estimation procedures. For example, it is well
known by now that misapplication of the classical least
squares (CLS)' estimation procedure to estimate struc-
tural parameters produces inconsistent estimates except
in the very special case of a fully recursive SEM.
Choice among alternative asymptotically justified
estimates has often been made on the basis of ease of
computation. For example, with systems linear in the
parameters, calculation of two- and three-stage least
squares estimates involves just simple algebraic opera-
tions, whereas computation of maximum likelihood esti-
mates involves more complex numerical procedures.
Some current computer packages compute a number of
asymptotically justified estimates and leave the difficult
choice among them to the user. Of course, in truly large
samples, asymptotically equivalent estimates should not
be very far different. If in practice such estimates, based
on a given large sample of data, are radically different,
this may be interpreted as indicating that the asymptotic
properties of different estimates take hold at different
sample sizes or, more likely, that specification errors are
present and affect alternative estimates differently. Un-
fortunately, not much analysis is available on the sensi-
tivity of alternate asymptotically justified estimates to
various kinds of specification errors; one recent paper in
this area is Hale, Mariano, and Ramage (1978). More
systematic analysis of this range of problems and pro-
duction of asymptotically justified estimates that are
relatively robust to specification errors would be welcome
and would serve as a useful additional guide to users in
selecting estimates when the sample size is truly large.
On the other hand, if the sample size is not truly large,
1 Some use the term ordinary least squares (OLS); I prefer classical
least squares (CLS), since the least squares principle is not ordinary.
 Zellner: Statistical Analysis of Econometric Models 631

even if a SEM is correctly specified, various asymptoti- is clearly inadmissible. This is not to say that MSE is the
cally justified estimates of the same parameter can assume only criterion for judging estimators, but it has received
quite different values. considerable attention in this area of research. As stated
Students and others invariably ask for a definition of before, these results have been surprising to many, par-
what constitutes a truly large sample. An easy answer to ticularly those who narrowly emphasize unbiasedness, or
this question is hard to give. The sample size alone is not minimum M\SE, or minimum-variance unbiasedness as
usually all that is relevant. Values of the parameters and criteria for judging estimators or who uncritically accept
features of input or exogenous variables also must be asymptotic justifications. To illustrate that these criteria
considered. Because parameter values usually are un- are inadequate even for the simple case in which a struc-
known and the object of estimation, prior information tural parameter 0 is equal to the reciprocal of a reduced-
about them is needed before one can say with any con- form regression coefficient, ir, that is, 0 1/7r, the ML
fidence what is a truly large sample in connection with the and almost all other asymptotically justified estimation
estimation of a specific SEM. Needless to say, if the procedures would recommend estimating 0 by 0 = 1/*,
sample size is not truly large, the asymptotic justifica- where 7r is the least squares estimator of the regression
tions for estimation and other large-sample inference coefficient 7r. Because *r is normally distributed, 0 is the
procedures become dubious. In a Bayesian context, one reciprocal of a normally distributed variable and hence
can compute the posterior distribution for a parameter does not possess finite moments of any order. Thus 0 has
and check to see that it is approximately normal with infinite risk and is inadmissible relative to quadratic and
posterior mean equal to the maximum likelihood esti- many other loss functions.
mate and posterior variance equal to the relevant element In addition to exact distributional work on the finite-
of the inverse of the estimated Fisher information matrix. sample properties of asymptotically justified estimators,
If so, large-sample conditions have been encountered. research has provided approximations to the moments of
These considerations do not give a justification for using these estimators, surprisingly even sometimes when mo-
the large-sample normal approximation to the posterior ments do not exist. As Anderson (1977) has pointed out,
distribution without computing the exact finite-sample these moment expressions approximate moments of
posterior distribution. truncated Taylor or other series approximations to the
estimators and not moments of the estimators. How im-
2.2.2 Finite-Sample Problems and Procedures. Recogni- portant this distinction is remains to be seen. Further,
tion that large-sample justifications for estimation pro- very fruitful work that uses Edgeworth-Charlier series
cedures do not contain explicit information on how large approximations to the moments and distributions of
a sample must be for them to hold and that practical estimators has been reported by Sargan (1976).
workers often must deal with limited data has prompted Monte Carlo studies also have been employed in an
considerable research on the finite-sample properties of effort to determine the finite-sample properties of alterna-
estimation procedures. A good deal of research has been tive estimators (see Sowey 1973). Generally, these studies
concentrated on obtaining analytically the exact finite- have been marred by an inadequate coverage of the high-
sample distributions of certain asymptotically justified dimensional parameter spaces associated with models,
estimators, for example, maximum likelihood (ML), two- even simple two-equation supply-and-demand models
stage least squares (2SLS), and other estimators for that usually contain about 10 or more parameters. Be-
parameters in relatively simple models. This ingenious cause risk functions of estimators usually intersect, failure
and difficult distributional work unfortunately has shown to examine the entire parameter space can yield mislead-
that the finite-sample distributions of estimators, derived ing and confusing results regarding the dominance of one
in the main from an underlying noncentral Wishart dis- estimator relative to another in terms of, for example,
tribution, are rather complicated and involve a number MSE. Thus, the results of Monte Carlo experiments that
of parameters with unknown values. The latter fact investigate the behavior of estimators over a limited
makes the application of these distributional results to number of points in the parameter space must be con-
concrete problems difficult. This research has shown that sidered very cautiously. See Thornber (1967) for a
asymptotically equivalent estimators have very different valuable illustration of this point.
finite-sample properties. For example, the (limited-in- While much effort has been directed at determining
formation) ML estimator does not possess finite moments the finite-sample properties of given, asymptotically
of any order, and in certain frequently encountered cases justified estimators, relatively little work has been done
the 2SLS estimator does not possess a mean or higher on the problem of producing estimates that have a small-
moments. Also, certain asymptotically unbiased estima- sample justification. Using approximate moment expres-
tors can have serious finite-sample biases. Further, and sions, Nagar (1959) attempted to define an approximate
perhaps surprising, under some conditions, the inconsis-minimal-MSE estimator for structural coefficients within
tent CLS or OLS estimator has a smaller mean squared the 32-class. Unfortunately, his "estimator" depends on
error (MSE) than consistent estimators that possess a parameters with unknown values that have to be esti-
finite-sample second moment. Of course, if an estimator mated to operationalize his estimator. When these
fails to possess a second moment, it has infinite MSE and parameters are estimated, it appears that the "optimal"

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 632 Journal of the American Statistical Association. Seotember 1979
properties of his estimator are vitally affected. His work properties in analyzing a particular set of data is not
provides some evidence that use of a value of X3 less than clear. Further, as many, including Tiao and Box (1975),
one, the value that produces the 2SLS estimate, is prob- have emphasized, the computed value of an optimal point
ably better than the value of one. Analysis by Sawa estimator can be a very bad representation of the in-
(1972) provides the approximate MSE of a Jc-class formation in a given set of data. Likelihood advocates
estimator for a structural parameter of a simple model emphasize the importance of studying properties of likeli-
and points in the same direction, namely, that finite- hood functions, while Bayesians emphasize the desir-
sample MSE usually is lower, and sometimes much ability of studying both likelihood functions and posterior
lower, when a value of 3C < 1 is employed. Sawa (1972) distributions to understand the information content of a
also has reported properties of estimators that are a given sample for possible values of parameters of a
linear combination of the 2SLS and the inconsistent CLS model. For both likelihood advocates and Bayesians, a
estimators. By appropriate choice of the weights, he has point estimate is just a summary measure that does not
obtained approximately unbiased and approximate necessarily convey all or most of the information in a
minimal-MSE estimators. These results do not appear sample regarding parameters' probable values.
relevant for cases in which the second moment of the
2SLS estimator does not exist, and the justification for .2.3 Bayesian Estimation Results. In the past 15
considering a linear combination of a consistent and an years, there has been a growing amount of research con-
inconsistent estimator is not apparent. cerned with developing and applying the Bayesian ap-
Recently, Fuller (1977) has presented modified limited- proach to the problems of estimating values of parame-
information ML and modified fixed 32-class estimators ters in SEM's and other econometric models, and ele-
that have finite moments. Restricting these modified ments of the Bayesian approach have appeared in
estimators to have the same, but arbitrary, bias, he shows econometric textbooks. As is well known, inferences abo
that to order T-2, where T is the sample size, the modified parameters' values, for example, elements of a parameter
ML estimator dominates in terms of approximate MSE. vector 0, are based on the posterior probability density
In almost all the analytical finite-sample work on the function (pdf) for 0,
sampling properties of estimates, problems with time
series complications have not been analyzed, for example, p(O I D, I) = cp( II)t(0 I D) (2.1)
estimates of parameters of models with lagged endog-
where c = [f p(01I)t(OID)d0]-' is a normalizing con-
enous variables and/or serially correlated error terms.
stant, D denotes the data, I denotes the prior informa-
Relatively little effort has been devoted to obtaining good
tion, p (OIl) is the prior pdf, and t(O ID) is the likelihood
finite-sample estimates of error terms' covariance
function. The following points are relevant particularly
matrices. Recent statistical work by Perlman, Eaton, and
for analyses of SEM's.
others certainly seems relevant. It is highly probable
that all, or almost all, the asymptotically justified estima- 1. The posterior pdf in (2.1) is an exact finite-sample
tors mentioned are inadmissible under quadratic loss and pdf, and, hence, large-sample approximations, while
sometimes convenient and useful, are, in principle, not
other loss functions over a wide range of conditions. This
needed. This statement applies to the analysis of all kinds
range of "Steinian" issues has received very little atten-
tion in connection with finite-sample work on structural of models, including static and dynamic SEM's.
parameter estimators' properties. The impact of pre- 2. Use of the prior pdf, p (OII), enables an investigator
testing on the finite-sample properties of the usual to incorporate prior information in an analysis, as much
or as little as he sees fit. Of course, if no sample informa-
structural-coefficient estimators is relatively unexplored.
For example, I have conjectured that the limited- tion were available, as is the case in some low-income
information ML estimator's distribution subject to a countries, prior information would be the only kind of
favorable outcome of the rank test for identifiability information
will available. In connection with SEM's, prior
possess finite moments. In the simple case in which a information must be introduced in some form to identify
structural parameter is the reciprocal of a reduced-form structural parameters. In sampling-theory approaches,
regression coefficient, 0= 1/7r, it is easy to establish the identifying information almost always has been in-
that the ML estimator O' 1/*, subject to the outcome
troduced as exact restrictions on parameter values (e.g.,
of a t test that rejects r = 0, possesses finite moments, setting certain coefficients equal to zero, equivalent to
where * is the least squares (ML) estimate of 7r. Last and using a degenerate or dogmatic prior pdf for these
most basic, the relevance of sampling-theory criteria, parameters in a Bayesian setting). Use of prior pdf's
such as unbiasedness, admissibility, and minimal MSE enables investigators to represent this required prior
of estimators for the analysis of a given sample of data information more flexibly; see Dr6ze (1975) and Kadane
has not been considered adequately in the econometric (1975).
literature. Sampling properties of procedures seem 3. Use of Bayes's Theorem provides the complete
relevant before we take the data in connection with design posterior pdf for parameters of interest and not just a
problems or in characterizing average properties of summary point estimate. If a point estimate is desired,
estimation procedures. The relevance of these average however, it usually can be obtained readily. For example,

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 Zellner: Statistical Analysis of Econometric Models 633

for quadratic loss functions, it is well known that the zero mean vector and g X g pds covariance matrix
mean of the posterior pdf, if it exists, is an optimal pointog= (r-1)/2;r-1.
estimate in the sense of minimizing posterior expected Under the assumptions made earlier, the parameters
loss. II and Zg of the URF system in (2.4) are identified and
4. Generally, Bayesian estimates have very good can be estimated by using Bayesian or non-Bayesian
sampling properties, because they minimize average risk techniques whether or not the structural parameters in
when average risk is finite and are admissible. r, B, and 2 are identified. It has long been recognized
5. In large samples under general conditions, the that, under the assumptions made before, r, B, and 2
posterior pdf, p(0 D, I), assumes a normal form with are not identified and that additional prior information
mean vector equal to the ML estimate of 0 and covariance must be added in order to identify these structural pa-
matrix equal to the inverse of the estimated Fisher in- rameters. Identifying prior information can take various
formation matrix. Thus, in large samples there is a dove- forms. Here we discuss only the case in which it involves
tailing of Bayesian and sampling-theory numerical restrictions that subsets of structural parameters assume
results; however, their interpretation is quite different. zero values. In addition, it is necessary to adopt a
normalization rule for elements of the r matrix. Here
I now turn from the general features of the Bayesian
let all diagonal elements of r be equal to 1. We shall wr
approach to a brief review of some Bayesian estimation
the system in (2.2) with identifying restrictions and
results for the SEM. A representation of the linear (in
normalization rule imposed as
the parameters) SEM is
Yrr = XBr + U (2.6)
Y r = X B + U (2.2)
nXg gXg nXk kXg nXg
Then the restricted reduced-form system is given by
where Y is an n X g matrix of observations on gyendoge-
XBrrr-1 + Urr-1 (2.7)
nous (or dependent) variables and X is an n X k matrix
of observations on k predetermined variables, assumed XH1r + Vr (2.8)
of rank k. Predetermined variables include both exoge-
Obviously, the fundamental f
nous (independent) and lagged endogenous variables. r
to reduce the number of fre
is a g X g structural parameter matrix, assumed non-
by so doing, to provide a m
singular, and B is a k X g matrix of structural parameters.
free structural parameters in rr and Br are identified.
U is an n X g matrix of disturbance or error terms. It
Explicit statements of the conditions for identification of
will be assumed that the rows of U have been inde-
structural parameters are given in econometrics text-
pendently drawn from a g-dimensional normal distribu-
books and other works. Since the free parameters in rr,
tion with zero mean vector and g X g positive definite
Bry and z are identified, their number cannot exceed the
symmetric (pds) covariance matrix 2. Note that if
number of parameters in the URF system in (2.4),
r = ig, the system in (2.2) is in the form of a multi-
namely, kg parameters in 11 and g(g + 1)/2 distinct
variate regression model when X contains no lagged
parameters in the g X g RF covariance matrix Q.
endogenous variables or in the form of a multivariate
The likelihood function for the restricted structural
autoregressive system with input variables when X con-
system in (2.6) is
tains both exogenous (or independent) and lagged
endogenous variables. In the special case r = Ig, 4(rr, analysis
Br, 2; 1 D) oC% I mod {rr I } n 1Z -n/2
of (2.2) from the Bayesian point of view would proceed .exp { -I tr(Yrr - XBr)'( Yrr - XBr) 2;} (2.9)
pretty much along multivariate regression lines if initial
values for the lagged endogenous variables are taken as where o denotes proportionality, D denotes the data,
and mod rr I denotes the absolute value of the Jacobian
given (see Zellner (1971) and the references cited there).
The unrestricted reduced-form (URF) system as- determinant, irr. , for the transformation from the n
sociated with the SEM in (2.2) is given by postmultiply-rows of U to the n rows of Y in (2.6). If the system in
ing both sides of (2.2) by r- to yield: (2.6) is autoregressive, (2.9) is the likelihood function
conditional upon initial values (assumed given). Then,
Y = XBr-P + ur-P (2.3)
from (2.1), the posterior pdf for the free parameters in
or rr, Br, and z is given by
Y = X HI + V (2.4)
nXg nXk kXg nxg p(rr, Br, 2jID, I) o p(rr, Br, 2;II)(irr, Br, 2D) (2.10)
where
II = Br- and V = Ur-, (2.5)
where p (rr, Br, III) is the prior distribution and the
prior information is denoted by I. Given a prior distribu-
with the k X g matrix H being the tion (URF) coefficient
and the likelihood function, the technical problems
matrix and the n X q matrix V the URF disturbance or of analyzing properties of the posterior distribution,
error matrix. The assumptions about the rows of U imply that is, obtaining its normalizing constant, its marginal
that the rows of V can be considered independently distributions, and its moments, remain.
drawn from a g-dimensional normal distribution with In the special case of a fully recursive SEM, F,. is in

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
634 Journal of the American Statistical Association, September 1979
triangular form, implying that stricted
Ir, = multivariate
1, and regression
2 is as-problem
sumed to have a diagonal form. These assumptions
simplify the likelihood function in (2.9) considerably and
also simplify the analysis of the posterior pdf in (2.10)
(see Zellner 1971 for details). The fully recursive case,
however, is a very special case of a SEM. In the general
case, work has concentrated on the analysis of (2.10)
using the likelihood function in (2.9). In several studies,
the posterior distribution in (2.10) has been computed
for a few simple models. Dreze and Morales (1976),
Harkema (1971), Morales (1971), and Richard (1973)
have analyzed it by using several different informative
prior distributions. Except for some approximate results
provided by Zellner (1971) and Zellner and Vandaele
(1975), it is necessary to use numerical integration tech-
niques to analyze features of the posterior distribution.
These approximate results have been used by Mehta and
Swamy (1978) to provide a ridgelike Bayesian estimate.
Kloek and van Dijk (1976) have studied the application
of Monte Carlo numerical-integration techniques in
analyzing posterior distributions. Although more experi-
ence with applications, assessing and using various prior
distributions, and computational procedures would be
useful, past research has yielded results that will be
valuable in obtaining better analyses of given SEM's,
particularly in small-sample cases.
Bayesian research also has focused on limited-informa-
tion analyses, that is, estimation of parameters of a
single equation or of a subset of equations of a SEM.
Complete posterior distributions for these problems have
been obtained and analyzed by Dreze (1972, 1976),
Morales (1971), Reynolds (1977), Rothenberg (1975),
and Zellner (1971). A single equation of the system in
(2.6), say the first, is given by
Yi = Yi Ti + X1 0i + ui (2.11)
nXl nXml mlXl nXkl klXl nXl
where Yi and Y, are components of Y, that is,
Y (y'. YI. Y0) with the variables in Yo not appearing
in (2.11), X1 is a submatrix of X, X = (X1i:X0) with the
variables in X0 not appearing in (2.11), and ul is a sub-
vector of U, U = (ul. Uo) and y, and 51 are parameter
vectors to be estimated. The assumptions introduced
about the rows of U imply that the elements of u, have
been independently drawn from a normal distribution
with zero mean and variance oai. The URF equations for
y, and Yl, a subset of the equations in (2.4), are
(y1 Y1) = X(=li. Hi) + (v1 V1) (2.12)
On postmultiplying both sides of (2.12) by (1. -'l)'
and comparing the result with (2.11), we achieve com-
patibility, given that
=1 = I'lrl + (2.13)
where the zero vector on the right side of (2.13) is
(k - k) X 1, and ui = v - Viyi. From (2.12) and
(2.13), the estimation problem can be viewed as a re-
This content downloaded from
86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
with the struc-
tural parameters yi and 51 involved in the restrictions on
the elements of =i and Hi. A necessary condition for the
identification of yi, gi, and oal is that k - k, > mi. Note
that (2.13) reflects restrictions arising from just the
first equation (2.11) of a system. The information in
restrictions similar to (2.13) associated with other
structural equations is not taken into account in estimat-
ing yi, 51, and all and hence the nomenclature, limited-
information or single-equation analysis.
Previous research has shown that ML, 2SLS, and
3SLS estimates are approximate means of posterior
pdf's for structural parameters under special conditions.
Of course, given complete posterior pdf's for paramaters,
optimal point estimates can be computed, that is,
posterior means for quadratic loss functions and medians
for absolute-error loss functions. A particularly simple
optimal point estimate under a generalized quadratic
loss function can be derived as follows. Upon multiplying
both sides of (2.13) on the left by X = (Xi:Xo), we
obtain
X7c = XIIJYi + X151 = Zlbl (2.14)
where Z1 = (XH1X1) and 6', = (y'1"'1). Take as loss
function,
L = (X=l - Zid1)'(Xi -Zdi)
= (1 - d )'Z'iZ(Bl - di) (2.15)
a generalized quadratic loss function. Given a posterior
pdf for the reduced form parameters =1 and HII, the
posterior expectation of L in the first line of (2.15) can
be evaluated yielding EL = E- lX'X=l-2d'lEZ'iX=l
+ d'iEZ'iZidi, where E is the posterior expectation
operator. Then the value of di that minimizes expected
loss, di*, termed a minimum-expected-loss (MELO)
estimate, is given by Zellner (1978):
di* = (EZ'iZi)-1EZ'iX=i (2.16)
When the system in (2.12) is analyzed under a diffuse
prior for the regression coefficients =1 and Hi and for the
error covariance matrix, the marginal posterior pdf for
(7ci. Hi) is in the matrix Student t form, and, hence, the
expectations in (2.16) are readily available. In this case,
Zellner (1978) has shown that di* is in the form of a
3C-class estimate with a value of 3C that depends on the
sample size and is less than one in finite samples. Also,
di* possesses at least first and second sampling moments
(Zellner and Park 1979). Further, the optimal estimate
of yi is a matrix-weighted average of the 2SLS and CLS
estimates with the weight on the CLS estimate going to
zero as the sample size increases (Zellner 1976). In small
samples, however, the optimal estimate of yi and 51 can
be very close to or exactly equal to the CLS estimate.
Thus, empirical workers who have persisted in their use
of CLS estimates may not be very far from an optimal
estimate in small samples. Further, this averaging of
2SLS and CLS estimates bears some resemblance to the
work of Sawa, mentioned previously; however, the
 Zeliner: Statistical Analysis of Econometric Models 635

weights that Sawa uses and those that associated with The quality of the approximation and finite-sample
(2.16) are different. Last, this point-estimation ap- power functions for widely used large-sample approxi-
proach has been applied to yield a MELO estimate of mate tests are relatively unexplored topics in econometric
parameters appearing in all equations of a system, that research.
iS, ?1, 2, 6..., 8a} where Bi' = (y'i 'i). Also see Mehta
Another topic that has received very little attention in
and Swamy (1978) for some useful Bayesian results for econometric research is the effects of pretests on the
obtaining point estimates of the B*'s that are related to properties of subsequent tests and on estimators' and
ridge-regression results. predictors' properties. That pretesting can vitally affect
Some additional issues regarding the Bayesian ap- properties of estimators is evident from consideration of
proach have been aptly summarized in the following simple cases, for example, Ylt = YY2t + ult and Y2t
remarks by Tukey (1978): = 7r2Xt + U2t. The RF equations for this simple system
It is my impression that rather generally, not just in economet- are Ylt = 7r1xt + vit and Y2t = 7r2Xt + V21 with a = 7ri/r2-
rics, it is considered decent to use judgment in choosing aThe ML estimator for y is j = 7r/72, where *i = Xxtyit/
functional form, but indecent to use judgment in choosing a
Ixt2, i = 1, 2. - does not possess finite moments; however,
coefficient. If judgment about important things is quite all
the distribution of - subject to the outcome of a pretest
right, why should it not be used for less important ones as well?
Perhaps the real purpose of Bayesian techniques is to let us do that rejects 72 = 0, namely, L*21 > CS72 > 0, where c
the indecent thing while modestly concealed behind a formal is a critical value and S,T2 is the standard error associated
apparatus. If so, this would not be a precedent. When Fisher
introduced the formalities of the analysis of variance in the with *2, does possess finite moments.
early 1920's, its most important function was to conceal the fact Work on Bayesian posterior odds ratios for selected
that the data was being adjusted for block means, an important hypotheses relating to SEM's parameters' values is re-
step forward which if openly visible would have been con-
sidered by too many wiseacres of the time to be "cooking the ported in Reynolds (1977). The posterior odds ratio, K12,
data." If so, let us hope the day will soon come when the role for two mutually exclusive hypotheses, H1 and II2, is
of "decent concealment" can be freely admitted. given by

2.3 Hypothesis Testing and SEM's

K12 = 012 X f pi(O)l1( I y)dO f p2(O)f2(Oly)d,
Sampling-theory procedures used for testing hypothe-
(2. 17)
ses relating to structural coefficients' values have in the
main been large-sample procedures, usually large-sample where 012 is the prior odds ratio, and for iX-1, 2, Pi(O)
likelihood-ratio tests or large-sample tests based on the is the prior pdf, and fi(0ly) is the likelihood function. If
Wald criterion, that is, for testing the rank condition for H1 and H2 are mutually exclusive and exhaustive, and
identifiability, overidentifying restrictions on structural H1 is 0 = 00, while H2 is 0 # 00, a pretest estimate that is
parameters, and general linear hypotheses regarding optimal relative to quadratic loss is given by
structural parameters' values. Recent research (Berndt
= P'00 + p2082
and Savin 1975, Savin 1976) has emphasized that as-
ymptotically equivalent testing procedures can produce = 00 + (62 - 0)/(K12 + 1) (2.18)
conflicting results when used in finite-sample situations
with a given nominal significance level. Analysis of as- where K12 = P1/P2 -p/(l - p1), where pi a
ymptotic power functions by Morgan and Vandaele the posterior probabilities on H1 and H2, res
and 02 iS the posterior mean for 0 under H2
(1974) has demonstrated that certain ad hoc testing pro-
pretest estimate, which also can be computed for other
cedures are dominated by standard large-sample testing
kinds of hypotheses, is a neat solution to the pretesting
procedures. Also, finite-sample approximations to the
problem as it relates to estimation. Similar considerations
sampling distribution of the likelihood-ratio test statistic
have received little, if any, attention in the econometric apply in obtaining combined, optimal predictions from
two or more alternative models.
literature.
In special cases, the exact finite-sample distribution of
2.4 Prediction Procedures for SEM's
a test statistic is available. In one such case the null
hypothesis specifies the values of all coefficients of endoge- Several alternative methods for generating predictions
nous variables in an equation, for example, y, = y1O and from SEM's have been discussed in the literature. First,
#ii 0 in yi = Yiyi + X1g, + ui. Conditional on it has been recognized that the URF system Y =XII
yi = 'y?, it is seen that yi -Y1Y1 = XXU1 + +u Viscan inbe fitted by least squares and used to generate
the form of a multiple regression given that X1 does not unrestricted reduced-form predictions (URFP's). Such
contain lagged endogenous variables. In this special predictions will not generally be efficient because re-
situation, test statistics that have exact t or F distribu- strictions on structural coefficients that imply restrictions
tions are available; however, the requirement that the on the elements of II are not reflected in URFP's. Second,
null hypothesis specify values for all elements of Ti is from the restricted SEM, YJi' = XBr + U, we can
quite restrictive. Also, when the system is dynamic, that obtain the restricted reduced-form system Y = XBrPv'
is, when X1 contains lagged endogenous variables, only + Vr, and restricted reduced-form predictions (RRFP's)
approximate large-sample test procedures are available. can be obtained from V'.t = X'firPr-, where x=f is a given

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 636 Journal of the American Statistical Association, September 1979

vector and Ar and 'r are estimated restricted structural-regressive-integrated-moving-average (ARIMA) time
coefficient estimates. Such predictions will be asymp- series and econometric models has been the subject of
totically efficient if Pr and r'r are asymptotically efficient
much research (Leuthold et al. 1970; Cooper 1972;
estimates and if, of course, there are no specification Nelson 1972; Christ 1975; also see Section 3).
errors. If Br and rr are estimated by inefficient but con-
sistent methods, it is not always the case that a predictor
3. THE SEMTSA APPROACH
based on them will be better in large samples than an
URFP (Dhrymes 1973). Last, the partially restricted As mentioned before, much past econometric research
reduced form (PRRF) equations can be used to generate has concentrated on the analysis of given models and
predictions, namely, yi = XIIiyi + X? S + vi, i = 1, 2, yielded relatively little on formal methods for checking
whether formulated models are consistent with informa-
... . g. Estimates of Hi, yi, and 5i along with given vectors
x'1 and X'if yield the PRRFP's P = x' fIiti + X/f i, tion in sample data and for improving models. In addi-
i = 1, 2, ..., g. Since the PRRFP's use more prior in- tion, many time series aspects of econometric modeling
formation than the URFP's in overidentified SEM's, have not been adequately treated. This is not to say that
they will have higher precision in large samples than time series considerations were totally absent from
URFP's. On the other hand, they will not generally be econometric research, but rather that there was no sys-
as precise as RRFP's in large samples when no specifica- tematic synthesis of econometric modeling and time
tion errors are present in the SEM. Approximate expres- series analysis.
sions for the variance-covariance matrix of forecast error Most important in stimulating some econometricians'
vectors are available in the literature for the prediction interest in time series techniques was the good forecasting
procedures mentioned previously. Further, it is apparent performance of simple, univariate time series models
that specification errors can vitally affect relative large- relative to that of large econometric models in the work
sample properties of these predictors. Then, too, only of Cooper (1972) and Nelson (1972). Much earlier,.
limited attention has been given to the problems of pre- Milton Friedman suggested that econometric models'
dicting future values of the exogenous variables in X. forecasts be compared with those of simple, univariate
It has been pointed out in the literature that the RRF "naive" models, a suggestion implemented by Christ
predictor, y'A = XIfP3rr-1 will not in general possess (1951). The relatively good forecasting performance of
finite moments, whereas the other predictors mentioned simple univariate autoregressive (AR) or Box-Jenkins'
will have finite moments in general for the URF predictor ARIMA models surprised econometric model builders.
and in most situations for the PRRF predictor (Knight In theory, a properly specified multiequation econometric
1977). More thorough analyses of alternative predictors' model should yield more precise forecasts than a uni-
finite-sample properties would be most valuable; see variate time series model, since the former incorporates
Schmidt (1977) for Monte Carlo experimental evidence much more sample and prior information. The reasonable
that led him to conclude that "The first main conclusion conclusion, drawn by many from these forecasting studies,
... is that inferences about forecasts are not terribly is that the econometric models considered in these studies
reliable, unless one's sample is fairly large" (p. 1004). probably contain serious specification errors (e.g., see
From the Bayesian point of view, the predictive Hickman 1972). For example, the econometric models
probability function for the URF system, Y = XII + V, may contain incorrect functional forms for relations, in-
is available. Its mean vector is an optimal point prediction appropriate lag structures, incorrect assumptions about
relative to a quadratic loss function. Optimal multistep the exogeneity of variables, incorrect assumptions about
predictions for the URF when it has autoregressive com- error terms' properties, and so forth. Because the rela-
plications have been obtained by Chow (1973). Richard tionship between econometric models and univariate
(1973) has studied predictive pdf's for the SEM and has ARIMA processes was not clearly understood, many
applied some of his results that incorporate restrictions econometricians considered simple time series models to
on structural coefficients in the analysis of small models. be ad hoc, mechanical, alternative models. Further, it
Further work to enlarge the range of prior pdf's used in was not apparent how time series analysis could be used
these analyses and to provide computer programs to to improve properties of SEM's. These issues were taken
perform calculations conveniently would be worthwhile. up in an article by Zellner and Palm (1974) and have
Some other issues that arise in use of econometric since been pursued in a number of other works such as
models for forecasting are (a) procedures for using Evans (1975, 1976, 1978), Palm (1976, 1977), Plosser
judgmental information and econometric models in (1976), Prothero and Wallis (1976), Trivedi (1975),
making forecasts; (b) ways of combining forecasts from Wallis (1976, 1977), and Zellner and Palm (1975).
alternative models (Nelson 1972; Granger and Newbold This research on the SEMTSA approach has, first,
1977); (c) criteria for the evaluation of the accuracy of emphasized that dynamic, linear (in the parameters)
forecasts (Granger and Newbold 1973, 1977); (d) data SEM's are a special case of multivariate or multiple time
quality and forecasting (Zellner 1958); and (e) seasonal series processes, such as studied by Quenonille (1957)
adjustment and forecasting (Plosser 1976). Further, the and others. Second, it has been shown that assuming
relative forecasting performance of univariate auto- variables to be exogenous places important restrictions

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 Zellner: Statistical Analysis of Econometric Models 637

on the parameters of a multiple time series process. Third,
the transfer-function (TF) equation system associated
with a dynamic linear SEM has been derived and shown
to be strongly restricted by structural assumptions. While
the TF equation system had appeared in the econometric
literature earlier under other names, its role in economet-
ric model building had not been emphasized. Fourth, in
the case of random exogenous variables generated by a
multiple time series process, it is possible to derive the
final equations (FE's), associated with the SEM, and
individual FE's are in the form of ARIMA processes of
the type studied by Box and Jenkins (1970) and others.
Thus, as emphasized in the SEMTSA approach, the
Box-Jenkins ARIMA processes are not ad hoc, alterna-
tive (to SEM's), mechanical models but are, in fact,
implied by SEM's (see the studies cited previously for
explicit examples). In addition, assumptions about
structural equations' properties have strong implications
for the forms of FE's and TF's that can be tested.
To make some of these considerations explicit, a
multiple time series process for a p X 1 vector of random
variables Zt (assumed mean-corrected for convenience)
is represented as follows (Quenouille 1957):
H(L) zt = F(L) et t = 1, 2, ..., T (3.1)
PXP PXl PXP PXl
where H(L) and F(L) are finite-order matrix polynomials
(assumed of full rank) in the lag operator L, and et is a
vector of serially uncorrelated errors with zero mean
vector and identity covariance matrix. If, for example,
F(L) is of degree zero in L, that is, F(L) = Fo, with Fo
of full rank, then the error vector in (3.1) is F0et with
zero mean and only a nonzero contemporaneous co-
variance matrix, EF0ete'tF'0 = FoF'o. Other specifica-
tions of F(L) allow for moving-average error terms. For
stationarity I H(L) I = 0 must have all its roots outside
the unit circle, while for invertibility the roots of IF (L) I analysis presented previously to take account of non-
= 0 must lie outside the unit circle.
Upon multiplying both sides of (3.1) by the adjoint
matrix associated with H(L), denoted by H*(L), we
obtain
IH(L) I zt = H*(L)F(L)et (3.2)
a set of FE's for the elements of Zt. Each of the FE's in
(3.2) is in autoregressive-moving-average (ARMA) form
that is, jH(L) zit = 'iet, where IH(L) j is an auto-
1971). These assumptions place restrictions on lag
regressive polynomial, and a'i, a 1 X p vector of poly-
nomial operators, is the ith row of H*(L)F(L). That patterns in equations, serial correlation properties of error
terms, and on which variables appear with nonzero
atet, a sum of moving-average processes, can be repre-
sented as a moving-average process in a single random
variable has been proved in the literature. Thus even
with the general multiple time series process in (3.1),
processes on individual variables will be in the Box-
Jenkins form.
In structural econometric modeling it is usually as-
sumed that some of the variables in Zt are exogenous. (3.5),
Ljet and the probability that errors will be made in
Zt= (y'tj*x't), where yt, a pi X 1 vector, denotes the specifying an initial variant of (3.5) generally will be
vector of endogenous vrariables and Xt, a p2 X 1 vector, high. The solution to this problem is not to discard the
denotes the exogenous variables. Then (3.1) can be
written as
(H1l H12\I'yt\ (F11 F12N(e1t\
(H21 | H22AXt F21 F22i\e2J 3
where the partitioning of H (L) = { Ht}, F (L)- {Fij}
and et has been made to conform to that for Z't = (y't .X'
The assumption that xt is exogenous places the following
restrictions on the matrix lag operators in (3.3):
H21-O, F12 0, and F21-O . (3.4)
On inserting (3.4) in (3.3), we have
Hilyt + H12xt = Flelt, (3.5)
and
H22xt = F22et . (3.6)
The equation system in (3.5) is the dynamic structural-
equation system, while that in (3.6) is the multivariate
ARMA process generating the exogenous variables in xt.
By multiplying both sides of (3.5) on the left by the
adjoint matrix associated with Hi,, denoted by Hi,*, we
obtain the TF system,
l H1 I 3yt + H11*Hl2xt = Hii*Fi1eit . (3.7)
Last, the FE's associated with (3.5) to (3.6) are obtained
by multiplying both sides of (3.6) on the left by H22*, the
adjoint matrix associated with H22, to obtain
I H2x2 l t = H22*F22e2t , (3.8)
and substituting for xt in (3.7) from (3.8) to yield
I H11 I I H22 It =-Hll*Hl2H22*F22e2t
+ IH221Hii*Fileit . (3.9)
Equations (3.8) and (3.9) are the FE's for the variables
in xt and yt, respectively. Each variable has an ARMA
process, as mentioned before. Simple modifications of the
stochastic exogenous variables, such as time trends, and
seasonal or other "dummy" variables, can easily be made.
In structural econometric modeling in the past, workers
have concentrated attention on the SEM given in (3.5).
Economic and other considerations have been employed
to justify the classification of variables into the two cate-
gories, endogenous and exogenous. Further special as-
sumptions regarding the matrices Hi,, H12, and F11 are
required to achieve identification (e.g., see Hannan
coefficients in equations of the system. If the resultant
system is appropriately specified and estimated, it of
course can be used for forecasting, control, and structural
analysis, the traditional objectives of SEM's. It must be
recognized, however, that a large number of specifying
assumptions have to be made to implement the SEM in

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
638 Journal of the American Statistical Association, September 1979

initial variant of (3.5), which may contain much valuable finding remedies for them. When the initial variant of a
information, but to pursue complementary analyses that SEM has been reformulated, its implications for the
can help to identify problems in the formulation of the forms of FE's and TF's can be checked empirically. Also,
initial variant and to suggest appropriate reformulation the roots of FE's and TF's can be calculated, estimated,
of specifying assumptions. Also, it is important that these and examined for reasonableness.
complementary analyses yield useful results along the The SEMTSA approach provides an operational and
way toward obtaining a good SEM; useful synthesis of traditional econometric and time
In the SEMTSA approach, it is suggested that workers series analysis techniques that can produce SEM's with
use economic theory and other outside information to fewer specification errors and better forecasting per-
formulate an initial, tentative form for (3.5). The next formance. As with traditional SEM's, however, some
step involves deducing algebraically the forms of the TF statistical problems associated with the SEMTSA ap-
system in (3.7) and the FE's in (3.9). As is obvious from proach require further research. First, there is the problem
the forms of the TF and FE systems, assumptions re- of determining the forms of the FE's from sample data.
garding the SEM in (3.5) will result in a number of im- Box and Jenkins' well-known suggested techniques based
portant restrictions on TF's and FE's that can be checked on properties of estimated autocorrelation and partial
empirically. For example, from (3.7), (3.8), and (3.9), it autocorrelation functions are helpful in ruling out a
is seen that the AR parts of the FE's and TF's will be number of forms for FE's; however, these techniques are
identical when lag operators do not contain common rather informal. For nested FE models, large-sample
factors. As pointed out in Zellner and Palm (1974), sys- likelihood-ratio tests can be employed to aid in dis-
tems with special features, that is, fully recursive systems criminating among alternative FE models. For nested and
or systems in which Hi, is block-diagonal, will lead to nonnested models, Bayesian posterior odds ratios also
cancellation, and thus the AR parts of FE's and TF's are useful. For example, in discriminating between a
will not be identical. Also, other special assumptions white-noise process and a first-order moving-average
about the forms of H11 arid H12 in (3.5) will result in process, Evans (1978) has shown that the posterior odds
TF's and FE's with different AR lag polynomials (see ratio is a function not only of the first-order sample serial
Zellner and Palm 1975 for an example). Work on examin- correlation coefficient, ri, but also of higher-order sample
ing the implications of specific SEM's for the forms of serial correlation coefficients, r2, r3, . . ., the latter having
FE's and TF's is extremely important in enhancing under- weights that decline as the sample size increases. Be-
standing of SEM's. For example, the effect of changing a cause ri is not a sufficient statistic and because the ri's
variable's classification from exogenous to endogenous are highly correlated in small samples, a large-sample test
on the forms of the TF's and FE's can be easily deter- using just r1 does not use all the sample information and
mined. Also, structural assumptions about lag structures, can lead to erroneous inferences. Posterior odds ratios
properties of structural error terms, and forms of policy also are useful in situations in which roots of AR poly-
makers' control policies all result in strong restrictions on nomials lie on the unit circle, a situation in which it is
TF's and FE's. In addition, Quenouille (1957, Ch. 5) has known that usual large-sample likelihood-ratio tests
provided valuable analysis of the effects of incorrect based on x2 statistics are invalid. Geisel (1976) has re-
inclusion or exclusion of variables, measurement errors, ported work indicating that Bayesian posterior odds
parameters varying with time, nonlinearities, and so on. ratios performed better than variants of Box-Jenkins pro-
When the forms of TF's and FE's associated with a cedures in discriminating among alternative ARIMA
SEM have been derived, the next step in the SEMTSA schemes. Extensions of this work and the early work of
approach is to analyze data to determinie or identify the Whittle (1951) on Bayesian hypothesis testing in time
forms of FE's and TF's to check that the empirically de- aeries analysis would be very valuable. This work also
termined FE's and TF's are compatible with those im- can shed light on the problem of determining the degree,
plied by the tentatively formulated SEM. Of course this if any, of differencing required to induce stationarity.
work not only provides checks on a SEM but also esti- Note that in formulating a posterior odds ratio, station-
mates FE's that can be used for prediction and TF's that arity is not required. Stationarity is required for most
can be used for prediction and control. If the analysis ofuses of sample autocorrelation and partial autocorrela-
the FE's and TF's provides results compatible with the tion functions.
implications of the SEM, the SEM's parameters can be Second, there is the problem of determining the forms
estimated, and it can be used for prediction, control, and of TF's. Important work on this problem for simple
structural analysis. If, as is usually the case, the results TF's has been reported by Box and Jenkins (1970, 1976),
of FE and TF analysis do not confirm the implications ofHaugh and Box (1977), Haugh (1972), Granger and
an initial variant of a SEM, the SEM must be reformu- Newbold (1977), and others. Also, the econometric work
lated. This reformulation process is facilitated consider- on distributed lag models is relevant (e.g., see Aigner
ably by knowing the results of TF and FE analyses. That 1971, Dhrymes 1971, Griliches 1967, and Nicholls,
is, the latter analyses usually indicate specific deficiencies Pagan, and Terrell 1975). Recent work of Sims (1972,
of the initial variant of a SEM, and many times recogni- 1975), Pierce and Haugh (1977), Skoog (1976), Wu
tion of these deficiencies is an important first step in (1978), and others on tests for special recursive structures,

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 Zellner: Statistical Analysis of Econometric Models 639

along with procedures suggested by Box, Jenkins, Haugh, tematic measurement errors and (t is a vector of random
Granger, Newbold, and others, may be useful in checking measurement errors. In this form, the SEM becomes
the assumptions about input variables' properties. In what engineers call a state-variable model. Perhaps use
TF's with several input variables, it may be advisable to of results in the engineering literature would be useful
reduce the number of free parameters by using some of in work with SEM's. Measurement problems are not
the assumptions in the distributed-lag literature regard- insignificant: Initial and subsequently revised figures for
ing coefficients of current and lagged input variables GNP and other important quarterly economic series
(e.g., see Shiller 1973). As with determining the forms of differ considerably, in some cases systematically, and
FE's, it is probable that posterior odds ratios will be found provide different information regarding cyclical turning
useful in discriminating among alternative forms for points (see Zellner 1958). Similar results have been ob-
TF's and in obtaining posterior probabilities associated tained in current work with preliminary and revised
with alternative variants of TF's. figures for quarterly nominal GNP. Revisions in the
Third, there is the problem of obtaining good estimatespreliminary estimates of quarterly GNP amounting to 5
of parameters in FE's, TF's, and SEM's. Currently, to 10 billion dollars are common. For example, the first
various asymptotically justified estimates are available, and subsequently revised figures (in billions of current
and some of these take account of random initial condi- dollars) for GNP in the fourth quarter of 1954 are 361.0,
tions and restrictions implied by the assumptions of 362.0, 367.1, 367.1, 367.7, and 373.4; for the fourth
stationarity and invertibility. The small-sample proper- quarter of 1965, the preliminary and subsequently re-
ties of these asymptotically justified estimates require vised figures are 694.6, 697.2, 704.4, 708.4, 710.0, and
much further investigation, a point also emphasized by 710.0. These figures illustrate an important measurement
Newbold (1976), who writes, "As regards estimation, I problem confronting econometric model builders and
am not sure that uncritical use of maximum likelihood forecasters that has not been adequately treated in the
estimates is justified in small samples without some in-literature.
vestigation of their sampling properties." As pointed out Sixth, aggregation problems have received increased
in Section 2, ML estimators do not in general possess attention in recent work. Articles by Geweke (1976),
good finite-sample properties. These comments imply Rose (1977), Tiao and Wei (1976), and Wei (1976)
that more work to obtain good finite-sample estimates is provide valuable results on temporal and other kinds of
required. Extensions of the valuable work of Box and aggregation in the context of time series models. In work
Jenkins (1970, 1976), Newbold (1973), Tiao and Hillmer by Laub (1971, 1972), Peck (1973, 1974) and Levedahl
(1976), and others on Bayesian estimation of time series
(1976), attention has been focused on economic models
models seem to be possible and can provide additional for individual firms and consumers using panel data and
good finite-sample estimation results. the implications of these microanalyses for aggregate
Fourth, the problems associated with seasonality are dividend, investment, and automobile expenditure func-
important in formulating and analyzing SEM's and yet tions. At the microlevel, discrete decisions, such as
have received relatively little attention. Because seasonalbuy/not buy or change/don't change the dividend rate,
variation accounts for a large fraction of the variation are extremely important. Yet macroformulations of be-
of many economic variables, a proper treatment of havioral relationships that are incorporated in many
seasonality is critical. In much econometric work, sea- SEM's do not properly take account of this discrete
sonally adjusted variables are used with little or no microbehavior and as a result are misspecified. Many
attention to the procedures employed for seasonal ad- estimated investment, dividend, and automobile expendi-
justment and their possible effects on determination of ture functions that are based on partial adjustment
lag structures and other features of SEM's. In the models show long response lags that are spurious and are
SEMTSA approach, Plosser (1976) and Wallis (1976) the result of aggregation over buyers and nonbuyers or
have provided valuable analyses of seasonality in SEM's. corporations that change and those that do not change the
Fifth, the problem of measurement errors in economic dividend rate in a particular quarter. Levedahl (1976) has
time series requires much more attention. It is well shown analytically and empirically that the adjustment
known that a number of economic series are derived coefficient in a partial adjustment model for automobile
wholly or in part from sample surveys. Many statistical expenditures is related to the proportion of consumers
analyses of such data are based on the usually erroneous purchasing a car in a particular period. Because this
assumption of simple random sampling. Analyses that proportion varies considerably over time, the adjustment
take proper account of the designs of sample surveys, coefficient is an unstable parameter, and models fitted
their sampling errors, and possible biases would be most under the assumption that it is stable have obvious
welcome. Further work to consider SEM's subject to problems in forecasting. These findings relating to defects
measurement error would also be valuable. For example, of widely used partial adjustment equations have serious
(3.5) could be formulated in terms of the true values of implications for SEM's that incorporate such equations.
variables, Z't = (y't-.x't). The measured values of vari- Further work on formulating macro-SEM's that takes
ables Zem' = (yam', Xem') could be assumed given by Ztm = account of discrete elements in economic behavior
better
seems very important in obtainiing better models.
Rze + {t, where R is a matrix of coeffcients reflecting sys-

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 640 Journal of the American Statistical Association, September 1979
Seventh, forms of policy makers' control may change
and thus cause instability in lag parameters and other
features of a model, a point emphasized by Lucas (1973).
Analyses using subsamples of data may indicate the
empirical importance of this problem.
Last, time series analysts have identified relatively
simple, low-order ARIMA processes for economic vari-
ables appearing in SEM's. On the other hand, the
ARIMA processes or FE's associated with most SEM's
usually are complicated, high-order schemes. To il-
lustrate, Leuthold et al. (1970) formulated and estimated
a SEM for analysis of hog markets with daily data. They
also identified and fitted ARIMA processes for the daily
price and quantity variables. Their time series model
for price was found to be a simple random walk. As shown
in Zellner (1975), the form of their SEM implies FE's for
price and quantity with AR parts of at least third order,
quite at variance with their random walk finding for the
price variable. It seems that they forced a misspecified
SEM on the data, one that involves the implicit assump-
tion of unexploited profit opportunities in the hog
markets. Indeed Muth (1961, p. 327, fn. 11), in his path-
breaking paper on rational expectation models (i.e.,
models that do not imply unexploited profit opportuni-
ties), writes in connection with a general supply and de-
mand model, "If the production and consumption flows
are negligible compared with the speculative inventory
level, the process [on price] approaches a random walk.
This would apply to daily or weekly price movements of
a commodity whose production lag is a year." Thus,
economic theory provides some support for the empirical
finding that daily hog prices follow a random walk and
that the SEM for the hog markets is probably
misspecified.
Similar considerations apply to the Hendry (1974)
model of the UK analyzed in Prothero and Wallis (1976).
The latter workers identified rather simple ARIMA
processes for variables appearing in Hendry's SEM. The
FE's associated with Hendry's estimated model have
ninth-order AR parts. Prothero and Wallis (1976, p. 483)
apparently take this finding of a ninth-order AR part of
the FE's of Hendry's model seriously and attribute the
relatively low orders of the empirically identified FE's to Christ's (1975) thoughtful and relatively favorable review
"relatively small coefficients of higher powers of L [the
lag operator], which proved difficult to detect in our uni- ing sentence
variate analyses." Also, they state that "the size of the
available sample [42 quarterly observations] has clearly
restricted our ability to detect subtle higher-order
effects." Whether these subtle higher-order effects are
real or are results of specification errors present in the ahead
eight-equation Hendry model is a point that deserves
further attention. In addition, the burgeoning literature
on rational expectation economic models has important
implications for the formulation and analysis of SEM's
(for some examples, see Evans 1975, Grossman 1975,
McCallum 1977, Nelson 1975, Ranson 1974, Sargent and
Wallace 1975, Wickens 1976, and Flood and Garber
1978).
4. CONCLUSIONS
This review of some of the research on SEM's has
emphasized the following major points:
1. Substantial progress has been made in research on
statistical methods for constructing, analyzing, and using
econometric models.
2. There is a serious need for developing and vigorously
applying additional statistical, mathematical, computer
simulation, and economic diagnostic checks of properties
of the SEM's.
3. For given SEM's, more work has to be done to
develop and apply estimation, testing, and prediction
procedures that have finite-sample justifications. In this
connection, the present author and others believe that
Bayesian procedures offer good solutions for many finite-
sample problems.
4. More formal procedures for using prior information
in the analysis of given SEM's are required, a problem
area that can be approached most satisfactorily at
present by use of the Bayesian approach.
5. Most serious is the need for formal, sequential sta-
tistical procedures for constructing SEM's.
6. The synthesis of traditional econometric model-
building techniques and modern time series analysis
techniques, called the SEMTSA approach previously,
probably will lead to improved SEM's, a view of the
present writer, Granger and Newbold (1975), and others.
7. Further use of existing economic theory, such as the
theory of efficient markets (see Fama 1970 for a review
of this theory), and rational expectations theory probably
will yield better SEM's. Having SEM's consistent with
elements of sound economic theory has long been em-
phasized in the econometric literature, and further
attention to this point in current work with SEM's is
critical.
So that this listing of research needs not be construed
as misrepresenting the quality of current U.S. macro-
SEM's that are used to generate quarterly forecasts of
important economic variables, such as GNP, unemploy-
ment, prices, and interest rates, it is relevant to consider
of the forecasting properties of such models. In the open-
of his article, he writes, "Econometric models
of the U.S. economy have been developed to the point
where forecasters who use them can forecast real and
nominal GNP two or three quarters ahead with root mean
square errors of less than one percent, and six quarters
with RMS errors of one to two percent." This
rather optimistic summary statement fails to take ac-
count of the fact that population RMS errors have been
estimated from rather small samples of forecast errors and
hence are not very precise. A confidence interval at a
reasonable level for the population RMS error would
probably be rather broad. Also, the implication of a 1
to 2 percent error for nominal GENP that now exceeds
1.5 trillion (1,000 billion) dollars, is about a 15 to 03

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 Zeliner: Statistical Analysis of Econometric Models 641

billion dollar or greater error, which would be considered on lack of expertise in the practical aspects of modelling that
substantial by most analysts. instances have occurred where well-built univariate models have
done better than poorly built multivariate "econometric" ones.
Further with respect to the very same models that yield
RMS errors of 1 percent or 2, Christ (1975) in the second In closing, it must be concluded from what has been
paragraph of his article writes, "... though the models presented and from Christ's remarks, that, while con-
forecast well over horizons of four to six quarters, they siderable progress has been made in work with SEM's,
disagree so strongly about the effects of important an econometric model as satisfactory as the Ford Model
monetary and fiscal policies that they cannot be con- T has not as yet appeared.
sidered reliable guides to such policy effects, until it can
be determined which of them are wrong in this respect [Received July 1977. Revised April 1979.]
and which (if any) are right." This statement clearly
indicates that at least some, or perhaps all, of the models REFERENCES
that Christ considered (Wharton; Data Resources Inc.;
Aigner, D.J. (1971), "A Compendium on Estimation of the Auto-
Bureau of Economic Analysis; St. Louis; Fair; Liu-Hwa; regressive Moving Average Model From Time Series Data,"
Hickman-Coen; and University of Michigan) probably International Economic Review, 12, 348-369.
contain serious specification errors. Anderson, T.W. (1977), "Asymptotic Expansions of the Distribu-
tions of Estimates in Simultaneous Equations for Alternative
Next, Christ (1975, p. 59) writes,
Parameter Sequences," Econometrica, 45, 509-518.
In general, it appears that subjectively adjusted forecasts using Berndt, E., and Savin, N.E. (1975), "Conflict Among Criteria for
ex ante exogenous values are better than the others. It is no Testing Hypotheses in the Multivariate Linear Regression Model,"
surprise that subjective adjustment helps. It may surprise some Discussion Paper No. 74-21 (rev.), Dept. of Economics, Uni-
that the use of actual exogenous values does not help, and some- versity of British Columbia.
times hinders. But there is likely to be some interaction, in the Box, G.E.P., and Jenkins, G.M. (1970), Time Series Analysis, Fore-
sense that if a forecaster feels that the preliminary forecast casting and Control, San Francisco: Holden-Day (rev. ed., 1976).
turned out by his model is unreasonable, he may both adjust (1976), "Discussion of the Paper by Dr. Prothero and Dr.
the model and change his ex ante forecast of the exogenous Wallis," Journal of the Royal Statistical Society, Ser. A, 139,
variables, in order to obtain a final forecast that he thinks is 493-494.
more reasonable. Chow, G.C. (1973), "Multiperiod Predictions From Stochastic
Difference Equations by Bayesian Methods," Econometrica, 41,
109-118 (reprinted in (1975), Studies in Bayesian Econometrics and
Christ's conclusion that "subjectively adjusted forecasts
Statistics in Honor of Leonard J. Savage, eds. S.E. Fienberg and
... are better than the others" (emphasis his) underlines A. Zellner, Amsterdam: North Holland Publishing Co., 313-324).
the importance of using prior information carefully in Christ, C.F. (1951), "A Test of an Econometric Model for the United
States, 1921-1947," in Conference on Business Cycles, New York:
preparing forecasts. His statement that use of the actual
National Bureau of Economic Research, 35-107.
values, rather than the anticipated values, of exogenous (1975), "Judging the Performance of Econometric Models
variables "does not help" is indeed surprising. In this of the U.S. Economy," International Economic Review, 16, 54-74.
connection, it should be appreciated that the subjective Cooper, R.L. (1972), "The Predictive Performance of Quarterly
Econometric Models of the United States," in Econometric Models
adjustments often take the form of adjusting the values of Cyclical Behavior (Vol. 2), ed. B.G. Hickman, New York:
of intercept terms in equations of a model. Because Columbia University Press, 813-926.
equations often are formulated in terms of nonstationary Dhrymes, P.J. (1971), Distributed Lags: Problems of Estimation and
Formulation, San Francisco: Holden-Day.
variables and may be considered as local approximations, (1973), "Restricted and Unrestricted Reduced Forms:
adjustments to intercept terms and slope coefficients will Asymptotic Distribution and Relative Efficiency," Econometrica,
be needed when values of the variables move away from 41, 119-134.
Dreze, J.H. (1972), "Econometrics and Decision Theory," Econome-
sample values. In such situations, thoughtful adjustment trica, 40, 1-17.
of intercept terms is a partial step in the direction of (1975), "Bayesian Theory of Identification in Simultaneous
obtaining better results; but, because it is partial, there Equation Models," in Studies in Bayesian Econometrics and Sta-
tistics, eds. S.E. Fienberg and A. Zellner, Amsterdam: North-
is no assurance that use of actual rather than anticipated Holland Publishing Co., 159-174.
values of exogenous variables will produce better results (1976), "Bayesian Limited Information Analysis of the
in general. Simultaneous Equations Model," Econometrica, 44, 1045-1076.
, and Morales, J.A. (1976), "Bayesian Full Information
Last, from the information on models' forecast errors Analysis of Simultaneous Equations," Journal of the American
that Christ has assembled, it appears that SEM's for Statistical Association, 71, 919-923.
both nominal and real GNP outperform univariate Evans, P. (1975), "Time Series Analysis of a Macromodel of the
U.S. Economy, 1880-1915," H.G.B. Alexander Research Founda-
ARIMA schemes for these variables in terms of estimated
tion, Graduate School of Business, University of Chicago.
RMS errors. As pointed out before, many have recog- (1976), "A Time Series Test of the Natural-Rate Hypothesis,"
nized, implicitly or explicitly, that a correctly specifiedH.G.B. Alexander Research Foundation, Graduate School of
Business, University of Chicago.
multiequation SEM should, in theory, perform better (1978), "Time-Series Analysis of the German Hyperinfla-
in forecasting than a univariate ARIMA process. For tion," International Economic Review, 19, 195-209.
example, Box and Jenkins (1976, p. 493) comment, Fama, E.F. (1970), "Efficient Capital Markets: A Review of Theory
and Empirical Work," Journal of Finance, 25, 383-417.
If the question is whether a set of univariate [ARIMA] models Flood, R.P., and Garber, P.M. (1978), "An Economic Theory of
of this kind which takes no account of relationships between Monetary Reform," unpublished manuscript, Dept. of Economics,
the series describes a set of related time series better than the University of Virginia.
corresponding multivariate [econometric] model then pre- Fuller, W.A. (1977), "Some Properties of a Modification of the
dictably the answer must be "No." It is a sobering commentary Limited Information Estimator," Econometrica, 45, 939-953.

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 642 Journal of the American Statistical Association, September 1979

Geisel, M.S. (1976), "Box-Jenkins or Bayes?," report of research to Graduate School of Industrial Administration, Carnegie-Mellon
the Econometrics and Statistics Colloquium, University of University.
Chicago. McCallum, B.T. (1977), "Price-Level Stickiness and the Feasibility
Geweke, J. (1976), "The Temporal and Sectoral Aggregation of of Monetary Stabilization Policy With Rational Expectations,"
Seasonally Adjusted Time Series," paper presented to the NBER- Journal of Political Economy, 85, 627-634.
CENSUS Conference on Seasonal Analysis of Economic Time Mehta, J.S., and Swamy, P.A.V.B. (1978), "The Existence of Mo-
Series, September 1976 (published in (1978), Seasonal Analysis of ments of Some Simple Bayes Estimators of Coefficients in a
Economic Time Series, ed. A. Zellner, Washington, D.C.: U.S. Simultaneous Equation Model," Journal of Econometrics, 7, 1-13.
Government Printing Office, 1978, 411-427). Morgan, A., and Vandaele, W. (1974), "On Testing Hypotheses in
Granger, C.W.J., and Newbold, P. (1973), "Some Comments on the Simultaneous Equation Models," Journal of Econ.ometrics, 2, 55-
Evaluation of Economic Forecasts," Applied Economics, 5, 35-47. Morales, J.A. (1971), Bayesian Full Information Structural Analysis,
P (1975), "The Time Series Approach to Econometric Model New York: Springer-Verlag.
Building," paper presented to the Seminar on New Methods in Muth, J.F. (1961), "Rational Expectations and the Theory of Price
Business Cycle Research, Federal Reserve Bank of Minneapolis, Movements," Econometrica, 29, 315-335.
November 1975. Nagar, A.L. (1959), "The Bias and Moment Matrix of the General
(1977), Forecasting Economic Time Series, New York: 3C-Class Estimators of the Parameters in Simultaneous Equations
Academic Press, 1977. and Their Small Sample Properties," Econometrica, 27, 575-595.
Griliches, Z. (1967), "Distributed Lags-a Survey," Econometrica, Nelson, C.R. (1972), "The Prediction Performance of the FRB-
35, 16-49. MIT-Penn Model of the U.S. Economy," American Economic
Grossman, S. (1975), "Rational Expectations and the Econometric Review, 62, 902-917.
Modeling of Markets Subject to Uncertainty: A Bayesian Ap- (1975), "Rational Expectations and the Estimation of
proach," Journal of Econometrics, 3, 255-272. Econometric Models," International Economic Review, 16, 555-561.
Hale, C., Mariano, R.S., and Ramage, J.G. (1978), "Finite Sample Newbold, P. (1973), "Bayesian Estimation of Box-Jenkins Transfer
Analysis of Misspecification in Simultaneous Equation Models," Function-Noise Models," Journal of the Royal Statistical Society,
Discussion Paper No. 357, Dept. of Economics, University of Ser. B, 35, 323-336.
Pennsylvania. (1976), "Discussion of the Paper by Dr. Prothero and Dr.
Hamilton, H.R., Goldstone, S.E., Milliman, J.W., Pugh III, A.L., Wallis," Journal of the Royal Statistical Society, Ser. A, 139,
Roberts, E.R., and Zellner, A. (1969), Systems Simulation for 490-491.
Regional Analysis: An Application to River-Basin Planning, Nicholls, D.F., Pagan, A.R., and Terrell, R.D. (1975), "The Estima-
Cambridge, Mass.: MIT Press. tion and Use of Models With Moving Average Disturbance
Hannan, E.J. (1971), "The Identification Problem for Multiple Terms: A Survey," International Economic Review, 16, 113-134.
Equation Systems With Moving Average Errors," Econometrica, Palm, F. (1976), "Testing the Dynamic Specification of an Econo-
39, 751-765. metric Model With an Application to Belgian Data," European
Harkema, R. (1971), Simultaneous Equations: A Bayesian Approach, Economic Review, 8, 269-289.
Rotterdam: Rotterdam University Press. (1977), "On Univariate Time Series Methods and Simul-
Haugh, L.D. (1972), "The Identification of Time Series Interrela- taneous Equation Econometric Models," Journal of Econometrics,
tionships With Special Reference to Dynamic Regression," un- 5, 379-388.
published doctoral dissertation, Dept. of Statistics, University of Peck, S.C. (1973), "A Test of Alternative Theories of Investment
Wisconsin, Madison. Using Data From the Electric Utilities Industry," unpublished
,and Box, G.E.P. (1977), "Identification of Dynamic Regres- doctoral dissertation, Graduate School of Business, University of
sion (Distributed Lag) Models Connecting Two Time Series," Chicago.
Journal of the American Statistical Association, 72, 121-130. (1974), "Alternative Investment Models for Firms in the
Hendry, D.F. (1974), "Stochastic Specification in an Aggregate Electric Utilities Industry," Bell Journal of Economics and Manage-
Demand Model of the United Kingdom," Econometrica, 42, 559- ment Science, 5, 420-458.
578. Pierce, D.A., and Haugh, L.D. (1977), "Causality in Temporal
Hickman, B.F. (ed.) (1972), Econometric Models of Cyclical Behavior Systems: Characterizations and a Survey," Journal of Economet-
(Vols. 1 and 2), New York: Columbia University Press. rics, 5, 265-294.
Jeifreys, H. (1957), Scientific Inference (2nd ed.), Cambridge: Plosser, C.F. (1976), "A Time Series Analysis of Seasonality in
Cambridge University Press. Econometric Models With Application to a Monetary Model,"
(1967), Theory of Probability (3rd rev. ed.), London: Oxford unpublished doctoral dissertation, Graduate School of Business,
University Press. University of Chicago.
Kadane, J.B. (1975), "The Role of Identification in Bayesian (1976), "Time Series Analysis and Seasonality in Econometric
Theory," in Studies in Bayesian Econometrics and Statistics, eds.
Models," paper presented to the NBER-CENSUS Conference on
S.E. Fienberg and A. Zellner, Amsterdam: North-Holland Pub- Seasonal Analysis of Economic Time Series, September 1976
lishing Co., 175-191. (published in (1978), Seasonal Analysis of Economic Time Series,
Kloek, T., and van Dijk, H.K. (1976), "Bayesian Estimates of ed. A. Zellner, Washington, D.C.: U.S. Government Printing
Equation System Parameters: An Application of Integration by Office, 365-393).
Monte Carlo," Report 7622/E, Econometric Institute, Erasmus Prothero, D.L., and Wallis, K.F. (1976), "Modelling Macroeconomic
University, Rotterdam. Time Series," Journal of the Royal Statistical Society, Ser. A, 139,
Knight, J.L. (1977), "On the Existence of Moments of the Partially 468-486.
Restricted Reduced-Form Estimators from a Simultaneous- Quenouille, M.H. (1957), The Analysis of Multiple Time-Series, New
Equation Model," Journal of Econometrics, 5, 315-321. York: Hafner Publishing Co.
Laub, P.M. (1971), "The Dividend-Earning Relationship: A Study Ranson, R.D. (1974), "Money, Capital, and the Stochastic Nature of
of Corporate Quarterly Panel Data, 1947-65," unpublished Business Fluctuations," unpublished doctoral dissertation, Grad-
doctoral dissertation, Graduate School of Business, University of
uate School of Business, University of Chicago.
Chicago. Reynolds, R. (1977), "Posterior Odds for the Hypothesis of Inde-
- (1972), "Some Aspects of the Aggregation Problem in the pendence Between Stochastic Regressors and Disturbances,"
Dividend-Earning Relationship," Journal of the American Sta- H.G.B. Alexander Research Foundation, Graudate School of
tistical Association, 67, 552-559. Business, University of Chicago.
Leuthold, R.M., MacCormick, A.J.A., Schmitz, A., and Watts, Richard, J.F. (1973), Posterior and Predictive Densities for Simul-
D.G. (1970), "Forecasting Daily Hog Prices and Quantities: A taneous Equations Model, New York: Springer-Verlag.
Study of Alternative Forecasting Techniques," Journal of theRose, D.E. (1977), "Forecasting Aggregates of Independent ARIMA
American Statistical Association, 65, 90-107. Processes," Journal of Econometrics, 5, 323-346.
Levedahi, J.W. (1976), "Predictive Error of the Stock Adjustment Rothenberg, T. (1975), "sBayesian Analysis of Simultaneous Equa-
Model," HI.G.B. Alexander Research Foundation, Graduate tions Models," in Studies in Bayesian Econometrics and Statistics
School of Businless, University of Chicago. in Honor of Leonard J. Savage, eds. S.E. Fienberg and A. Zellner,
Luceas, R.E. (1973), "Econometric Policy Evaluation: A Critique," Amnsterdam: North-Holland Publishing Co., 405-424.

This content downloaded from

86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
 Zellner: Statistical Analysis of Econometric Models 643
Sargan, J.D. (1976), "Econometric Estimators and the Edgeworth
Approximation," Econometrica, 44, 421-448.
Sargent, T.J., and Wallace, N. (1975), "Rational Expectations, the
Optimal Monetary Instrument and the Optimal Money Supply
Rule," Journal of Political Economy, 83, 241-254.
Savin, N.E. (1976), "Conflict Among Testing Procedures in a
Linear Regression Model with Autoregressive Disturbances,"
Econometrica, 44, 1303-1315.
Sawa, T. (1972), "Finite-Sample Properties of the 3C-Class Estima-
tors," Econometrica, 40, 653-680.
(1973), "The Mean Square Error of a Combined Estimator
and Numerical Comparison With the TSLS Estimator," Journal
of Econometrics, 1, 115-132.
Schmidt, P. (1977), "Some Small Sample Evidence on the Distribu-
tion of Dynamic Simulation Forecasts," Econometrica, 45, 997-
1005.
Shiller, R.J. (1973), "A Distributed Lag Estimator Derived From
Smoothness Priors," Econometrica, 41, 775-788.
Sims, C.A. (1972), "Money, Income and Causality," American
Economic Review, 62, 540-552.
(1975), "Exogeneity and Causal Ordering in Macroeconomic
Models," paper presented to the Seminar on New Methods in
Business Cycle Research, Federal Reserve Bank of Minneapolis,
November 13-14, 1975.
Skoog, G.R. (1976), "Causality Characterizations: Bivariate, Tri-
variate and Multivariate Propositions," Staff Report No. 14,
Federal Reserve Bank of Minneapolis.
Sowey, E.R. (1973), "A Classified Bibliography of Monte Carlo
Studies in Econometrics," Journal of Econometrics, 1, 377-395.
Thornber, H. (1967), "Finite Sample Monte Carlo Studies: An Auto-
regressive Illustration," Journal of the American Statistical Associa-
tion, 62, 801-808.
Tiao, G.C., and Box, G.E.P. (1973), "Some Comments on 'Bayes'
Estimators," The American Statistician, 27, 12-14 (reprinted in
(1975), Studies in Bayesian Econometrics and Statistics in Honor
of Leonard J. Savage, eds. S.E. Fienberg and A. Zellner, Amster-
dam: North-Holland Publishing Co., 620-626.
Tiao, G.C., and Hillmer, S. (1976), "Seasonal Adjustment: A
Bayesian View," paper presented to the 13th meeting of the
NBER-NSF Seminar on Bayesian Inference in Econometrics,
November 1976.
Tiao, G.C., and Wei, W.S. (1976), "Effect of Temporal Aggregation
on the Dynamic Relationship of Two Time Series Variables,"
Biometrika, 63, 513-523.
Trivedi, P.K. (1975), "Time Series Analysis Versus Structural
Models: A Case Study of Canadian Manufacturing Behavior,"
International Economic Review, 16, 587-608.
Tukey, J.W. (1976), "Discussion of Granger on Seasonality," paper
presented at the NBER-CENSUS Conference on Seasonal Analy-
sis of Economic Time Series, September 1976 (published in (1978),
Seasonal Analysis of Economic Time Series, ed. A. Zellner, Wash-
ington, D.C.: U.S. Government Printing Office, 50-53).
Wallis, K.F. (1976), "Seasonal Adjustment and Multiple Time Series
Analysis," paper presented to the NBER-CENSUS Conference on
Seasonal Analysis of Economic Time Series, September 1976
DAVID A. BELSLEY and EDWIN KUH*
Zellner's article marks a significant step in econometric
writings, for it is one of the first pieces by a maj or
econometrician that recognizes the full spectrum of the
* David A. Belsley is Professor, Department of Economics, Boston
College, Chestnut Hill, MA 02167, and Visiting Professor, MIT
Center for Computational Research in Economics and Management
Science. Edwin Kuh is Professor of Economics, Sloan School of
Management, Massachusetts Institute of Technology, Cambridge,
MA 02139, and Director of the MIT Center for Computational
Research in Economics and Management Science.
This content downloaded from
86.59.13.237 on Mon, 11 Dec 2023 22:48:56 +00:00
All use subject to https://about.jstor.org/terms
(published in (1978), Seasonal Analysis of Economic Time Se
ed. A. Zellner, Washington, D.C.: U.S. Government Printing
Office, 347-357).
(1977), "Multiple Time Series Analysis and the Final Form
of Econometric Models," Econometrica, 45, 1481-1497.
Wei, W.S. (1976), "Effects of Temporal Aggregation in Seasonal
Time Series Models," paper presented to the NBER-CENSUS
Conference on Seasonal Analysis of Economic Time Series,
September 1976 (published in (1978), Seasonal Analysis of Eco-
nomic Time Series, ed. A. Zellner, Washington, D.C.: U.S. Govern-
ment Printing Office, 433-444).
Whittle, P. (1951), Hypothesis Testinq in Time Series Analysis,
Uppsala: Almqvist and Wiksells Boktryckeri AB.
Wickens, M.R. (1976), "Rational Expectations and the Efficient
Estimation of Econometric Models," Working Paper No. 35,
Faculty of Economics, Australian National University.
Wu, D.M. (1978), "Causality Test and Exogeneity Test," unpub-
lished manuscript, Dept. of Economics, University of Kansas.
Zellner, A. (1958), "A Statistical Analysis of Provisional Estimates of
Gross National Product and Its Components, of Selected National
Income Components, and of Personal Saving," Journal of the
American Statistical Association, 53, 54-65.
(1970), "The Care and Feeding of Econometric Models,"
Selected Paper No. 35, Graduate School of Business, University
of Chicago.
(1971), An Introduction to Bayesian Inference in Economet-
rics, New York: John Wiley & Sons.
(1974), "Time Series Analysis and Econometric Model Con-
struction," invited paper presented to the Conference on Applied
Statistics, Dalhousie University, Halifax, Nova Scotia, 1974
(published in (1975), Applied Statistics, Amsterdam: North-
Holland Publishing Co., 373-398).
(1976), "A Note on the Relationship of Minimum Expected
Loss (MELO) and Other Structural Coefficient Estimates," un-
published manuscript, H.G.B. Alexander Research Foundation,
Graduate School of Business, University of Chicago.
(1978), "Estimation of Functions of Population Means and
Regression Coefficients Including Structural Coefficients: A
Minimum Expected Loss (MELO) Approach," Journal of Eco-
nometrics, 8, 127-158.
and Palm, F. (1972), "Time Series Analysis and Simul-
taneous Equation Econometric Models," paper presented to the
Olso Econometric Society Meeting, 1972 (published in (1974),
Journal of Econometrics, 2, 17-54).
(1975), "Time Series and Structural Analysis of Monetary
Models of the U.S. Economy," Sankhya, Ser. C, 37, 12-56.
, and Park, S.B. (1979), "Minimum Expected Loss (MELO)
Estimators for Functions of Parameters and Structural Coeffi-
cients of Econometric Models," Journal of the American Statistical
Association, 74, 185-193.
- and Vandaele, W. (1975), "Bayes-Stein Estimators for k-
Means, Regression and Simultaneous Equation Models," in
Studies in Bayesian Econometrics and Statistics in Honor of Leona
J. Savage, eds. S.E. Fienberg and A. Zellner, Amsterdam: North-
Holland Publishing Co.
Comment
actual practice of econometric model building-including
its trial-and-error (or iterative) aspects that are often
viewed askance by econometric theorists. Implicit in
Zellner's broad-gauged view is a realization that the
classical statistical techniques, in their purist sense, pro-
a Journal of the American Statistical Association
September 1979, Volume 74, Number 367
Invited Paper