386 
PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA - 
25 
of them raise problems of varying magnitude for structural 
astimation. 
As is well known, there are now a fairly large number of 
alternative estimation techniques available for such estimation. 
Such methods fall into classes which differ in the assumptions 
made or amount of information taken into account. They also 
have different properties. In general, a great deal is known 
concerning the asymptotic properties under ideal conditions of 
most of these estimators; rather less is known of small sample 
properties; and a very few results are available on their relative 
robustness — the relative degree to which they stand up to 
such things as multicollinearity, specification error, and serial 
correlation in the disturbances of the model. 
This paper begins by reviewing the known properties of 
the principal estimators in the context of economy-wide models. 
We observe that the features of such models mentioned above 
make the use of even the best of such estimators rather suspect 
in its original form, while the size of such models makes them 
literally unavailable when time series of lengths usually 
encountered are the data involved. This leads naturally to 
estimation using instrumental variables in some form, and the 
second half of the paper is devoted in one way or another to 
exploring the question of how appropriate instrumental vari- 
ables should be chosen. It is argued that this is best done 
through continual application of the a priori structural infor- 
mation which governs the formulation of the entire model in 
the first place, rather than through relatively arbitrary statistical 
devices. 
1.2. Classification of Estimators 
For our purposes, the estimators which have been proposed 
for structural estimation may be divided into three classes. The 
first of these consists of ordinary least squares and its generali- 
‘6] Fisher - pag. 2