TR

PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA -

28

4.4. Robustness

Thus neither large nor small sample properties under ideal
conditions provide much guide for the choice of an estimator
from the limited-information class in the present state of knowledge.
 This is not entirely the case as regards the robustness
of such estimators, however. KLEIN and NAKAMURA have
shown that as a consequence of the stochastic nature of % in
limited-information maximum likelihood, that estimator is more
sensitive to multicollinearity than are the other members of the
k-class (*). In the absence of other criteria, these seem grounds
for abandoning limited-information maximum likelihood in
practice in favor of some other limited-information estimator.
There seem to be no very strong reasons, however, for
choosing among the limited-information estimators other than
maximum likelihood. The paper on robustness just mentioned
indicates that these do not differ among themselves as regards
this property (*!). Since two-stage least squares is the easiest
of these estimators to compute and since it does provide a natural
 generalization of ordinary least squares in the presence
of theoretically given normalization rules (2), it seems natural
to choose it in the present state of our knowledge.

5. NEAR-CONSISTENCY, BLOCK-RECURSIVE SYSTEMS, AND THE
CHOICE OF ELIGIBLE INSTRUMENTAL VARIABLES

5.1. Introduction

In this section we begin the discussion of the choices of
predetermined instrumental variables which are available and

(°°) KLEIN and NAKAMURA [16].
(*') See also Fisuer [8] for proof that the same is true as regards sensitivity
 to specification error.
(32) See Cuow [7].

61 Fisher - pag. 24