100
PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA - 28
not be remarkably important as the small sample variances
of such estimators are infinite in some cases. Ordinary least
squares certainly does have the property of finite small sample
variances under ordinary conditions, however defective it may
be for other reasons. Ordinary least squares estimates of the
reduced form equations may therefore be appropriate ones to
consider if one is willing to assume that serial correlation is
animportant.
Note that this is not quite the same as the situation as
regards structural estimation already discussed. In that con-
text several strong assumptions have to be nearly satisfied in
order to justify the use of ordinary least squares. In the present
context, only the assumption of no serial correlation must be
approximately satisfied; if it is, the remaining argument against
ordinary least squares is the one of lack of asymptotic effi-
ciency and this may be by no means decisive in a world of
relatively small samples (18).
In practice, however, ordinary least squares estimation of
the reduced form of a large economy-wide model is simply
incapable of accomplishment. If all lagged endogenous vari-
ables are treated as predetermined, the number of exogenous
and predetermined variables in any but the most aggregative
economy-wide model is simply too large to permit this type of
estimation in the presence of the relatively low number of obser-
vations ordinarily available.
2. FULL-INFORMATION ESTIMATORS
We now discuss the class of full-information estimators out
of what is perhaps the natural order. because it is relatively
(5) All of our discussion of the effects of serial correlation has over-
looked the existence of estimation techniques designed precisely to deal
with that problem. See for example JoHNsTON [15, pp. 192-195] and
THEIL (32, pp. 219-225]. All of these techniques, however, assume that
there are no lagged endogenous variables in the model, and we have prin-
cipally been concerned with the problems raised bv serial correlation when
there are such lagged variables
61 Fisher - pag. 16
