452 PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA -

IR

properties: one does not design a model in order that it should have
minimum delay; one does not design a model so that it will be an
eo 1pso predictor. One designs a model so that it represents something
 that one believes to be true about the real world. After the
model has been designed, and not before, one asks: What properties
does it have? And what properties can I give the estimates of the
model by using different estimating techniques?
Now, indeed, there are always several ways in which a given
real phenomenon can be represented in a model. Given two of these
which we believe to be equally valid in terms of economic theory,
we would of course choose the one which had better properties in
terms of forecasting or in terms of the properties which the estimation
 techniques appropriate to the model will have. Frequently, however,
 that is not in fact the primary question. In general, we are
not indifferent between two representations and the choice between
models does not come on the question: « Will one be recursive and
the other simultaneous? » The crucial question is rather that of
which model represents what we believe to be true about the real
world and which model and estimation technique is appropriate to
the use which we intend to make of the results.

THEIL

A question of clarification. Do you intend to split up a large
equation system into subsystems and construct principal components
of the predetermined variables in each subsystem?

FISHER

Professor THEIL wishes to know whether I recommend splitting
up a large equation system into subsystems and constructing principal
 components of the predetermined variables in each subsystem.
This is not what I recommend, although similar procedures have in

61 Fisher - pag. 68