Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Econometrics as Pioneering in Nonexperimental Model Building

Author(s): Herman O. Wold


Source: Econometrica, Vol. 37, No. 3 (Aug., 1969), pp. 369-381
Published by: The Econometric Society
Stable URL: http://www.jstor.org/stable/1912787 .
Accessed: 22/05/2014 09:03

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
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.

The Econometric Society is collaborating with JSTOR to digitize, preserve and extend access to Econometrica.

http://www.jstor.org

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
E CO METR
VOILUME37 JUlY.1969 NUMBER 3

ECONOMETRICS AS PIONEERING IN NONEXPERIMENTAL


MODEL BUILDING'

BY FIERMAN0. WOLD

Econometrics is seen as a vehicle for fundamental innovations in scientific method.


above all in the development of operatixe forecasting procedures in nonexperimenntal
situations. On the one hand there is the evolution relative to experimental situations, for
experiments yield reproducible knowledge and new results can therefore be used for fore-
casting as soon as they are established. Within nonexperimnentalforecasting, on the other
hanid,there is the evolution from the largely passive forecasts that are typical in demography
over to the operative forecasts in terms of instruments and targets that are the aspiration,
difficult to materialize with success, of the multirelational forecasting models of modern
econometr-ics.The econometric developmenitsconfirm the general experience that scientific
evolution is never a straightforward affair. To solve new types of problems requires a
reor-ientationof the analysis, sometimes a sharp turn fronmthe old tracks, and the sharper
the turn the more confused and controversial is the accompanying debate. As briefly
reviewed in my Address, econometrics has been pathbreaking in several fundamental
dievelopmentsconnected with forecasting, incluLdingthe transition from exact to stochastic
relations, fi om descriptive to explanatory analysis, and from unirelational to mnultirelational
model building.

INTRODUCTION

THE ECONOMETRIC SOCIETYis coming of age. It can look back over a third of a
century since its birth: it so happens that these thirty-three years are precisely
the middle third of our century. I shall take this piece of number magic as an excuse
for a discourse on scientific evolution. Let us take a look at econometrics as a
whole; let us focus on broad perspectives. How does econometrics fare in the
turbulent stream of developments'? The moderni world dodges along under the
joint impact of population explosions, nuclear explosions, and educational
explosions. Where does econometrics stand. and where are we going'?
Speaking broadly, econometrics serves the purposes of operations anialysis.
From the general point of view of scientific method, operations analysis is a
synthesis of experimental and nonexperimental model building. The operations
under analysis involve some element of manipulation, a feature that can be seen
as part of a controlled experiment. The reactions to the manipulations are uncon-
trolled, however, and thereby make the overall process nonexperimental. At the
micro level of econometrics the operations at issue are under the management of
firms and institutions. At the macro level the operations constitute society-wide
' This paper was given as the PresidenitialAddress at the December, 1966 meeting of the Econometric
Societv in San Francisco.

369

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
370 HERMAN 0. WOLD

economic policy. The experimental features of econometric operations analysis


enter into the design of and the carrying out of the micro and macro operations
in play. The nonexperimental features include the interplay between the operations
and the actions and reactions of the free market and of the autonomous forces in
the national or international economy.
For the founders of the Econometric Society the first and foremost aspiration
was the quantification of economics-the integration of economic thinking with
mathematical and statistical analysis. That this goal has been reached is clear.
It is also clear that powerful developments influenced the direction of this trend.
Since the middle of the nineteenth century, economic statistics has been flowing
from industrialized countries in ever broader streams. Since the 1910's, with a
few forerunners around the turn of the century, the accumulating data has been
exploited for econometric analyses of more and more varied kinds. Specifically,
one might mention the National Bureau of Economic Research, founded in 1920,
and Wesley Mitchell's pioneering studies of lags and leads in the dynamics of the
business cycle.
Econometrica,from its first volume in 1933, became a vehicle for rapid progress
of the theory and application of econometrics. Model building soon became the
key word for coordinating theoretical and empirical research. The term "model"
in this usage is a joint theoretical-empirical construct, where a clear cut distinction
is made between theoretical-hypothetical structure and empirical observations,
and where the observations are used, on the one hand, to estimate the parameters
of the hypothetical construct, and on the other, to test its empirical validity.
The methods of econometric model building have developed hand in hand with
the operative use of the models in applied work. I shall focus below on three
broad aspects of econometric model building and operations analysis. The first
aspect is the gradual evolution in the construction of unirelational models and
the succeeding bold rise in aspiration levels by building multirelational models.
The second aspect has a still wider perspective, namely econometrics as a spearhead
in the nonexperimental methods of model building. The third aspect, which
provides the main incentive for my address, deals with the lessons that can be
drawn from econometrics to promote interdisciplinary developments of the
social sciences.

Around 1900 econometric model building invaded a vast virgin territory, and
the developments soon formed a pattern in accordance with the general rule that
scientific evolution proceeds from simpler to more complex achievements. In the
first phase, roughly up to the birth of our Society, econometric models were
unirelational. More specifically, the main proving ground of unirelational econ-
ometric methods was demand analysis: the quantitative assessment of demand
functions and demand elasticities. These early developments extended over more
than thirty years from 1907, when Rodolfo Benini estimated the price elasticity
of the Italian demand for coffee to be 0.47 (the first empirical demand elasticity
on record), to 1938, when Henry Schultz published "The Theory and Measurement

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 371

of Demand. " This book was the first comprehensive treatment of the main problems
of demand analysis, including the much discussed problem of "choice of regres-
sion," although Schultz left this fundamental problem open. Several features
combined to give demand analysis a leading part in the initial phase of econ-
ometrics. First was the great operational need for quantitative knowledge about
demand functions and elasticities. Second, empirical data were available that could
be exploited for the analysis. Third, the theoretical approaches in the area were
long established, including Cournot's notion of the demand function (1838);
Marshall's notion of demand elasticity with respect to price (1890); the theory of
consumer demand as based on the notions of cardinal utility (Gossen, 1854;
Jevons, 1871) or ordinal utility (Pareto, 1906; Slutsky, 1915). The methods of
unirelational model building soon spread to other areas of econometrics, often
giving rise to new techniques-in particular the analysis of cost functions and
Cobb-Douglas' assessment of production functions.
The Harvard Business Barometer, the famed forecasting model of the 1920's,
broke down and disappeared after it failed to forecast the stock market crisis in
1929 and the ensuing great depression. The insufficiency of unirelational models
for purposes of economy-wide forecasting was revealed in a glaring light. The
time was ripe for more advanced approaches. Multirelational models rapidly
cam-eto the foreground in econometric research, and the 1930's became a melting
pot for fundamental innovations. Jan Tinbergen launched his epoch-making
macroeconomic models for the Netherlands (1937) and the United States (1939).
Within a few years two other approaches followed in the form of multirelational
models. These, too, were pathbreaking innovations: Wassily Leontief's input-
output models (1941) and Trygve Haavelmo's approach (1943), initiated and
developed under the auspices of the Cowles Commission. For a while these
approaches met with scepticism similar to that voiced by nonquantitative econ-
omists against the forerunners in econometrics at the beginning of the century.
On this score, Lord Keynes' critical review of one of Tinbergen's earliest macro-
economic models (Economic Journal, 1939) may be cited. The multirelational
macro models have come to stay, however, as witnessed by the broadening flow
of macroeconomic model building through the 1950's and 1960's all over the globe.
This work has even received encouragement from the Pope, in connection with the
Study Week on macroeconomic operations analysis organized in 1963 by the
Pontifical Academy of Sciences [3]. Achievements specially notable are the Dutch
and United States macroeconomic models associated with the names of Pieter de
Wolff and Daniel Suits. These models have been in operative use for annual fore-
casting for more than twelve years and both have an impressive record of successful
forecasts [4]. At our San Francisco meeting many papers and entire sessions are
devoted to the most recent developments in multirelational model building.

2
My brief survey of the econometric evolution from unirelational to multi-
relational model building has come to an end. I shall now turn to what is emphasized

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
372 HERMAN 0. WOLD

in the title of my address, namely that econometrics belongs to the wider area of
nonexperimental model building and is pioneering in this area.
We shall thus be concerned with experimental versus nonexperimental model
building. This is a dualism that goes to the core of the scientific method. The main
point I wish to comment on is that nonexperimental methods are valid and
indispensable tools in the scientific workshop. Recognition is due-and perhaps
overdue-for nonexperimental analysis as an active partner to the method of
controlled experiments, and for the integration of both experimental and non-
experimental approaches in the scientific method.
Experiments give reproducible knowledge. From the point of view of model
building the information given by the experiment is a part or the whole of a model,
or a cognitive model as we may say to emphasize the epistemological aspects of
model building. A key point in the present context is that reproducible knowledge
is predictive knowledge. Hence, experimental models are predictive. Within the
framework of experimental design, the outcome of an experimental replication
can be forecast, be it a piece of descriptive knowledge, such as the measurement of a
physical constant, or a piece of explanatory knowledge, such as the cause-effect
relationship established by a stimulus-response experiment.
The development of nonexperimental models in econometrics has involved a
great many problems at all levels, from the general foundations of the scientific
method to special technical problems about the various models at issue. Reference
is made below to two key matters of debate at the fundamental level.
The first matter concerns the scientific evolution from deterministic (exact,
residual-free) models to stochastic models. This is a generalization that is indis-
pensable in empirical work in order to account for the statistical scatter that in
practice blurs the relationships of the model. The generalization is radical, and by
no means trivial. We note a parting of the ways, inasmuch as the statistical scatter
is sometimes apparent, being due to observational errors, and sometimes genuine,
as is the case when it is due to biological or other inherent variability in the observed
phenomena, or when it is due to factors that have been neglected in the relations
under analysis. For this and other reasons the problems involved in the generaliza-
tion have been subject to persistent and partly confused debate, as is witnessed by
the two problem areas known as "the choice of regression" and "simultaneous
equations." These are two veritable stumbling blocks on the road of nonexperi-
mental model building.
The second issue concerns the distinction between causal and noncausal fore-
casting models. To some extent this is a matter of aspiration levels in model
building. Again, this is a topic that is by no means trivial.
My limited purpose in this context is to bring home two points. First, the prob-
lems at issue are of fundamental relevance for the construction of nonexperimental
models for purposes of operations analysis. Second, econometrics has played a
pioneering part in posing these problems and in establishing principles and
methods of general scope for their treatment. Its influence is very broad, serving as a
pioneer in the wide field of the social sciences, and in the still wider realm of
nonexperimental model building.

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 373

As to the evolution from deterministic to stochastic model building, the salient


point is that mathematical operations with deterministic models may or may not
carry over to corresponding operations with stochastic models. Trouble arises
when an operation is in current use in deterministic models but does not carry
over. In the "choice of regression," the crucial operation is inversion.Let us consider
the simple case of a deterministic relation between two variables y, z:
(1) y =-:1?1 + p1Z Z=2 ?+ 2Y

with (2 = - (11, f2 = 1/fll. The sign indicates that the two versions of the
relation are identical in the sense of formal inversion (in a graph they are repre-
sented by one and the same line). For example, we may on the one hand want to
deal with demand y as a function of price z, or on the other hand, price z as a
function of demand y. The corresponding stochastic specification gives us two
regression relations, namely
(2) E(ylz) = oc' + /3z 0 E(zly) = oc'2+ '2Y
with /2 # 1/f' . We have here written the two regressions in terms of predictors (a
convenient synonym for conditional expectations). The signs X, # indicate that
inversion is an operation that does not carry over to stochastic relations specified
in terms of predictors (in a graph the two regressions give different lines). The
dualism (2) was recognized as a key feature of regression analysis from its very
beginning [Galton, 1886]. In econometrics "the choice of regression" was posed
as a problem by P. Mackeprang [1906], who used the twin regressions to estimate
the price elasticity of coffee as 0.42 and 0.83, respectively, and he left the choice
between the two estimates open. The emergence of two regression lines instead of
just one estimate of the demand relation was regarded more or less as a statistical
misfortune. The problem was subject to persistent debate in econometrics through
the 1920's and 1930's. For Schultz (in the work referredto above) the problem was
still open, inasmuch as he reported the twin estimates for all of his price elasticities.
Linked with the second issue, Mackeprang's problem dissolves if we specify the
operative use of the model in causal terms. A demand relation is a cause-effect
relationship with price as causal variable and demand as effect variable; that is, the
relation tells us how demand is influenced by a change in price, assuming that other
influencing factors remain constant. On this hypothesis, the answer to the problem
is given by the regression of demand upon price, not price upon demand, just as
the regression of response upon stimulus provides a consistent estimate for the
relationship between stimulus and response in a controlled experiment. The main
difference is that, in general, we must use multiple regression to take into account
influencing factors other than price, whereas simple regression will suffice in a
randomized stimulus-response experiment. True, it is relevant to ask the inverse
question of how price is influenced by demand, but to answer this question we
must model a price mechanism. According to current economic theory, the price
mechanism involves both consumers' demand and producers' supply, and is not
simply the inverse of the demand relation. To sum up, "the choice of regression" is
a choice between causal models, and the choice between the models settles the

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
374 HERMAN 0. WOLD

choice between the regressions. And if we ask not only for the demand relation, but
also for the price mechanism, the question carries over from unirelational to
multirelational models.
In the problem of "simultaneous equations," there are two crucial operations:
inversion, and the solving of implicit equation systems. This problem area has been
much debated over the last twenty years. This is not the place to enter upon details,
since the problem is still under debate, and it would abuse the privilege of a
presidential address to set forth an argument when there is no opportunity for
discussion. Nor is it necessary for my limited conclusion in this passage, namely
that the problems at issue first were posed and analyzed in an econometric setting.
It will serve my purpose, however, just to state how the operational problems of
"simultaneous equations" pose themselves. The following formulas give, on the
left, the deterministic version and, on the right, the corresponding stochastic version
in terms of predictors. For the sake of simplicity we consider the special case when
all relations of the structural form are behavioral. The symbol y stands for a
vector y where the component variable yi, occurring in the left hand member, is
suppressed.
The structural form is:
(3a-b) y = f,B + Fz, E(yly, z) = /By+ Fz.
The reduced form is:
(4a-b) y = (I - )-'Fz, E(ylz) = (I - f-)'Fz.
The structural form, after rearrangement so that each current endogenous vari-
able occurs just once in the left hand member, becomes
(5a-b) y' = ,T'y'+ F'z, z) = f'y' + F'z.
E(y'137',
The passage from structural to reduced form involves the solving of an implicit
equation system; the rearrangement of the structural form involves inversion of
one or more behavioral relations. These operations, as specified on the left, are
formally legitimate in deterministic models. Only to a limited extent do the opera-
tions carry over to the stochastic models. For example, if we assume (4b), relation
(3b) can, in general, be satisfied only if matrix ftis of the subdiagonal type that is
characteristic of Tinbergen's type of models, known as causal chain systems.
For the Haavelmo type of models, interdependentsystems, the matrix ftis more
general. This makes the situation more complicated. Here the passage from (3a) to
(4a) carries over from (3b) to (4b) if the structural form is respecified in line with
Henry Theil's two stage method of estimation. Let n denote the number of equa-
tions. If we write
(6a) Yi*= E{yiE[(I- ft<'Fz]i} = [(I - f,) 'Fz]i (i = 1,... , n)
for the predictors of y given by the reduced form, and respecify the structural form
as
(6b) y= ?+ z +
where ftand F are numerically the same matrices as before and ? stands for the

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 375

respecified residuals, then in general it will be possible to satisfy both (6a) and
(6c) E(yjy*,z) = fl,y + Fz.
These brief comments will suffice to indicate that causal chains and interdepend-
ent systems are genuine innovations from the general point of view of scientific
method. In principle, the models are of general scope in nonexperimental model
building. Some of the ensuing problems are tricky and difficult, and research is still
in active development. This is particularly so with regard to the estimation tech-
niques for interdependent systems, an important problem area that I cannot enter
upon in this address (see [2]).
Proceeding, we come to the second broad area of debate, causal versus noncausal
forecasting models. For a start I shall comment upon forecasting models quite
generally, and then turn to the special case of causal models.
Briefly stated, nonexperimental models are based on past observations, and the
model sums up regularities observed in the past. In a forecasting model the past
regularities are assumed to persist in the future, and the systematic comparison
between the forecast generated from the model and the actual course of events is
the ultimate touchstone for testing the validity of the model. Forecasting is to
nonexperimental model building as replications are to controlled experiments.
From the point of view of operations analysis, forecasting plays the part of key
tool as well as key purpose in model building. In this respect too, and again with
reference to unirelational as well as multirelational model building, econometrics
has pioneered in the social sciences. More specifically, forecasting enters into three
phases of model building in the nonexperimental workshop. First, provisional
forecasts are extracted from the model under construction and compared with
actual developments as a check and as a guide for improving the model. Second,
the finished model generates the forecasts that it is intended to provide, and in due
course the comparison with actual developments shows whether or not the model is
a good one ("the proof of the pudding is in the eating"). Third, a nonexperimental
forecasting model as a rule needs revision from time to time, owing to dynamic
features that have not been covered by the model, and the errors in the forecasts are
indicators for the need of such revision.
As to this last point, the macroeconomic models designed for annual forecasting
may now be in a danger zone. After the Second World War Western economies
formed a Keynesian pattern with unemployment and other unused capacities of
industrial production, and adopted economic policies providing substantial
production increases accompanied by little or no inflation. In recent years the
employment ceiling has been reduced or nearly reached in several countries, and
a more or less marked inflation has set in. Will the forecasting models to which I
have referred continue to be successful if the inflationary tendencies spread and
become more severe? Or will the breakdown of the Keynesian pattern subsist,
and call for a complete remoulding of the forecasting models?
The debate on the causal aspects of forecasting models has an epistemological
background. Since the days of Galileo and Hume, causation and related concepts
have been under a cloud of philosophical scepticism. In 1914 the scepticism was

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
376 HERMAN 0. WOLD

emphatically voiced by Bertrand Russell, with the result that from the 1920's well
into the 1950's causal concepts were practically taboo in large areas of the social
sciences, especially in sociology. The half truth, "causation is predictability
according to a law," is typical of this period. In recent years (say in the course of
some ten years), causal analysis has been resuming its central place in explanatory
models, including the models of operations analysis. The lost son is greeted back
in many quarters, perhaps most clearly in sociology and econometrics.2
In econometric model building the operative use of forecasting models has been
instrumental in clearing up the haze around predictive versus causal aspects of the
models. Specifically, I understand that recent developments have led to a general
consensus about the following point: if a forecasting model is to be applied for pur-
poses of economic policy in terms of instruments and targets in the sense of Jan
Tinbergen [1956], the relations between instruments and targets must be specified
as (hypothetical) cause-effect relationships.
I shall stop here for a moment to compare econometric and demographic
forecasting models. The mortality forecasts of actuarial science are an early
instance of socioeconomic prediction. Their operative success is witnessed by the
flourishing development of life insurance companies since the beginning of the
eighteenth century. The first population forecasts date from the nineteenth century,
and equally time-honored is the experience that they are successful with regard
to mortality, whereas fertility developments have proved very nasty by fooling
the efforts of forecasters. In comparison with econometric forecasting, the salient
point in this context is not the degree of success of the forecast, but the difference
in aspiration levels with regard to the operative use of the forecasting models.
Population forecasts are passive in the sense that they do not involve instruments
and targets. To paraphrase in terms of the ceteris paribusassumption of forecasting
models (that the phenomena subject to forecasting are governed by the same general
rules in the future as in the past), the rules of the demographic models involve no
specification of instruments or targets in Tinbergen's sense. In Tinbergen's
models, and many other econometric models, the ceteris paribus assumption
includes rules of the game specified in terms of instruments and targets. This
assumption is of a cause-effect nature.
The situation may be summarized as follows (and again I believe that there is a
more or less complete consensus about these points):
(i) The cause and effect specification of individual parameters of a forecasting
model is formally in line with a stochastic specification of the relations in terms of
predictors, and is conceptually a distinct additional hypothesis which may or may
not be realistic and may or may not be adopted.
(ii) Causal specification of a forecasting model is not needed if none of the
variables in the model is subject to manipulation or control in the sense of opera-
tions analysis. The cognitive basis of the forecast is then nothing but the ceteris
2 See, for example, H. M. Blalock, Jr., Causal inferences in nonexperimentalresearch, Chapel Hill:
Univ. of North Carolina Press, 1964, and my review of this book in Econometrica,33 (1965), 879 f, with
references to H. Simon (1954), G. Eklund (1959), H. Wold (1952); in this last reference I have now taken
the opportunity to correct a printing error in the dating.

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 377

paribus assumption that the general conditions of the past will remain in the
future.
(iii) Causal specification of a forecasting model is a necessity if one or more of the
variables in the model are subject to manipulation or control in the sense of
operations analysis.
(iv) Causal specification of a forecasting model is a necessity if the future pattern
of variability is foreseen to involve changes relative to the past. If the changes are
in the nature of innovations relative to the past, the forecasting model requires a
relaxation of the ceteris paribus assumption on which it is based.3 The occurrence
of such innovations, especially when they are man made, creates a specific incentive
for forecasting. Major innovations relative to the past set a ceiling on forecasting
by scientific methods.

3
In this remaining section, we shall plunge into very deep waters. Econometrics
is but one member of the social sciences. When it comes to nonexperimental model
building and operations analysis for social purposes in general, the problems span
an interdisciplinary field and call for interdisciplinary research and collaboration.
What lessons can then be drawn from the pioneering contributions of econometrics
-lessons about the reach and limitation of these contributions? As background
for my comments in this direction, three problem areas will be mentioned.
The first area concerns juvenile delinquency.Delinquency is counteracted by an
elaborate social system, including criminal laws, correction by enforced treatment,
punishment by imprisoning, etc. Is the system efficient from the point of view of
social-wide operations analysis? How can the system best be improved?
The second area deals with the promotionof research.Research and its practical
results are a key factor of social progress. How efficient is the organization of
research from the point of view of social-wide operations analysis?
The third area concerns the distributionof income and wealth. In Western culture
the distribution of income and wealth follows the traditions of laissez faire, today a
modus vivendi which is modulated to a large extent by the economic policy of
national governments, including an elaborate system of income taxes, inheritance
taxes, etc. What can be said about the existing distribution of income and wealth
from the point of view of socioeconomic operations analysis?
Modern social sciences have a pressing challenge to tackle these and other urgent
socioeconomic problems in the spirit of social-wide operations analysis. A first
comment is, of course, that even the most successful of the existing macroeconomic
models are only a small beginning if seen in the perspective of such wide and
complex problems. But a beginning has been made, and econometrics shows the
way in two fundamental respects-model building and forecasting. Models are
needed that come to grips with the problems in terms of causes and effects, ends
and means of social-wide operations. Forecasting plays the part of key tool as
well as key purpose of the model building.
3 See [1] for a study of this situation.

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
378 HERMAN O. WOLD

A second obvious comment is that model building and forecasting are general
principles only. The specific causal hypotheses embodied in econometric models
give little or no guidance when it comes to socioeconomic model building outside
the economic sphere. Model building is gaining momentum in the social sciences
by and large, especially in sociology. But at all levels there are serious difficulties
that must be mastered before reaching the goal of operationally useful forecasting
models. As to the second problem area, research organization at large, a demand-
supply approach calls for measurement of input and output. Whereas input into
research is relatively easy to measure in terms of the input of work hours, equip-
ment, material, and other items, the output of research is notoriously difficult to
quantify and assess. Referenceis made to the "cost-benefit" approach; the measure-
ment difficulties at issue show up the very term "benefit."
To repeat, the problem areas cited are only a sample from a multitude of society-
wide problems that call for interdisciplinary collaboration in the construction of
causal models. Human knowledge is like rays of light; the various social sciences
are specialized, like the spectral components of light; and what is needed in
interdisciplinary collaboration is a resynthesis into full white light. In historiog-
raphy the characteristic spectral line is the chronological aspect of the social
phenomena; in statistics it is the numerical aspect; in geography it is the regional
aspect; in demography it is the population aspect; and in economics it is human
needs in relation to technology and material resources available to meet the needs.
That none of the social sciences is self contained in its characteristic line of the
spectrum is clear. All through the social sciences there is a considerable amount of
interdisciplinary synthesis in concepts, observations, and models.
Of special interest is the situation in sociology. Its characteristic line in the
scientific spectrum is the small groups aspect of social phenomena. At the same
time sociology has another program, sociology at large, where the aim is to provide
an integration or general synthesis of the social sciences. In the sociological
literature the small groups aspect has an overwhelming dominance over sociology
at large. This is an inheritance from an earlier phase of sociology, when a reaction
set in against highly speculative tendencies in the direction of sociology at large.
In recent years, however, sociology at large has gained momentum. In its comeback
it can draw some general guidance from model building in econometrics. Time i$
now ripe to pose the problems of sociology at large in terms of social-wide
operations analysis, using the quantitative methods of model building, the con-
struction of models in terms of socioeconomic units and institutions (typical
examples being the family, the business firm, the trade union, the school, the
government agency, and so on), the specification of the ends and means of the
model in terms of instruments and targets, and the coordination of causal
hypotheses and predictive aspects of the model so that forecasting plays the dual
role of operative purpose and test of validity of the model.
The three problem areas may again be quoted for illustration. For example,
to pose the problem about the distribution of income and wealth in an operationally
meaningful way, it is necessary to specify ends and means by way of an integrated
society model in the sense of sociology at large. This is not to say that a mammoth

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 379

model is needed. On the contrary, to judge from the experience of econometric


model building, the development of successful forecasting models resembles the
history of aviation: the first flying machines of the brothers Wright were small
things, and only gradually, as the problems were mastered step by step, did the
machines become bigger and more sophisticated.
The question is: how many basic forces must an integrated society model include
to be able to fly and give forecasts of operative use? Physics at present works with
two basic forces, gravitation and electro-magnetism. The socioeconomic force
field includes economic, ethical, artistic, and political forces, and such forces in
a great many ramifications and combinations. Will four basic forces do for a start?
How many relationships will be needed to come to grips with the key features of
the socioeconomic developments?
The socioeconomic forces are directed by individuals. Hence individual
psychology interferes with sociology at large. There is a need to develop the model
building in this opposite direction, microanalysis, so as to include the driving
forces at the plane of individual behavior and personal psychology. Here again
there is urgent need for further development. For one thing, personal psychology
is as yet relatively undeveloped when it comes to predictive testing and operative
use of the various approaches. For example, the Swedish psychiatrist Sjobring
has created a system of psychological taxonomy that seems to be realistic and
useful. Neither Sjobring nor other personal psychologists, however, have attempted
to record the psychological profiles of a group of individuals or a social community,
in order to explore the extent to which psychological profiles of individuals provide
operationally useful information for purposes of forecasting individual behavior,
much less as a means of coming to grips with psychological factors for purposes
of model building on a social-wide scale.
Sjobring's system has four dimensions or forces. Together with the basic
forces of sociology at large, this makes a minimum of eight forces. The number,
being much larger than the two basic forces of physics, is in full accordance with
what has often and rightly been said, namely that biological problems on the
whole are of a higher order of difficulty and complexity than the problems of the
natural sciences, and that socioeconomic phenomena are of a still higher order of
difficulty and complexity.

CONCLUSION

In conclusion I should like to adduce some general comments on the present


outlook on socioeconomic model building at large.
Many people will agree that developments of econometrics in this century
have been rapid and even spectacular. At the turn of the century, Western economic
and socioeconomic policy was dominated by the principle of laissez faire. Society-
wide model building was not at all in demand and economic research took little
or no interest in earlier approaches in this direction, such as Quesnay's table,
so that in due course econometric model building had to start more or less from
scratch. Hence I could finish my address on an optimistic note by stating that if

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
380 HERMAN 0. WOLD

progress over the last fifty years has been good, it is reasonable to expect that
econometrics will continue to grow and more or less automatically yield results
in the direction of interdisciplinary, integrated, society-wide models. It is question-
able, however, whether such optimism is warranted. It is easy to point to two main
forces behind modern socioeconomic progress, namely administrative planning
and scientific innovations. It is equally easy to conclude that a sound coordination
of these two forces is indispensable for steady progress. But are we actually on the
road to such a coordination?
Speaking broadly it would seem that administration and science can handle
small scale problems quite well, whereas the picture is entirely different when it
comes to the big and burning social problems. Part of the story is that administra-
tion and research in our modern society are institutionalized, and efficient
coordination between institutions is no easy or automatic matter. For one thing,
university research on the whole is oriented towards small scale problems, projects
suitable for master degrees, Ph.D. degrees, and promotional work in the university
career. For such purposes it is risky to take up large problems, let alone burning
social problems. On the administrative side, the institutional treatment of social
problems has reached some sort of modus vivendi, be it that the state of things
today often has marked streaks of modus morendi, and that it is sometimes a
euphemism to call it stable, and to call it a modus. Nonetheless, planning has a
stability of its own, for good and evil, and no doubt we are far from a social system
that makes for automatic coordination of planning and research, let alone an
optimal coordination. In passing, to avoid misinterpretation, I wish to state
explicitly that I do not see research as a panacea, and still less do I believe that
scientists can ever take over the functions of politicians and administrators.
In this matter it is easy to point to extremes. On the one hand cases exist in
which planning and research coordinate very well; the results are impressive and
progress is rapid. On the other hand, cases exist in which the manifest results are
obviously very poor. Everybody knows that successful results come more easily
when the problems involved are mainly of a technological nature, as with satellites
and moon ships. At the other extreme we usually find problems of an essentially
social nature, such as the clashes between races in our present day society. There is
a wide intermediate area in between the two extremes where planning and research
proceed in some modus vivendi. Nobody knows how far it is from one extreme to
the other. The point I wish to make in this context is that when it comes to large
social problems, the coordination between planning and research is not automatic.
On the contrary, the larger the problems and the more institutionalized the planning
and research, the more pronounced will be tendencies toward inner frictions and
conflicts in coordination. Both planning and research have their vested interests,
personal and material, and the institutional rules of the game are very different.
To venture a conjecture, it seems that the frictions are the rule rather than the
exception, unfortunately, with the result that the average efficiency of coordination
is not far above the lower extreme.
To finish with a positive formulation, I think there is plenty of room for increasing
the overall efficiency of coordination between planning and research. The aims of

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions
ECONOMETRICS 381

my address will be fulfilled in a main part if it can stimulate researchers and


planners to be on watch for frictions, overt or hidden, and to find ways and means
for improving coordination.

University Institute of Statistics, Uppsala, Sweden

REFERENCES

[1] MOSBAEK,E. J., ANDH. WOLD:"Estimation vs. Operative Use of Parameters in Models for Non-
Experimental Situations," forthcoming in Economie et Econometrie, Vol. 1, No. 1. (Cahiers de
'I.S.E.A., Paris.)
[2] -: InterdependentSystems. Structure and Estimation. Amsterdam, North-Holland Publishing
Co. In press.
[3] SALVIUCCI, P., ED.: "La Semaine d'Etude sur le R61e de l'Analyse Econometrique dans la Formula-
tion de Plans de Developpement," Scripta Varia, 28, Vatican City, Pontifical Academy of
Sciences, 1965.
[4] WOLD, H., G. H. ORCUTT, E. A. ROBINSON,D. SUITS, AND P. DE WOLFF: Forecasting on a Scientific
Basis. Lisbon, Gulbenkian Institute of Science, 1967.

This content downloaded from 195.97.171.78 on Thu, 22 May 2014 09:03:11 AM


All use subject to JSTOR Terms and Conditions

You might also like