.68
PONTIFICIAE ACADEMIAE SCIENTIARVM SCRIPTA VARIA - ©
Secondly, it is of course possible that Professor WoLD’s J? can
be less than 1. This can happen if structural change takes the form
of only a small change in the parameters but a large downward
change in the variance of the disturbance terms.
My next point is that Professor WoLp has been careful to avoid
an error which is occasionally made in the literature, namely that
least squares provides an unbiased forecast of the dependent variable.
That is false unless the least squares parameters are themselves un-
biased and this is not the case if, for example, a lagged dependent
variable appears on the right-hand side of the equation. Professor
Worp, however, has not said this although he has said something
which sounds like it. What he has said is that least squares is a
consistent estimator of an unbiased predictor and this, as he has
shown, is true.
Finally, it seems to me that Professor Wozp is unduly worried
about what he calls the danger of over-fitting in the reduced form.
The circumstance in which the reduced form cannot be estimated
by ordinary least squares because there are too few observations
relative to the number of exogenous variables in the model, is quite
a common one in dealing with large econometric models. This is not
of great consequence as a fundamental matter, however, because
one then uses intrumental variables methods which drop some of
the exogenous variables for purposes of estimation. No difficulty
of principle arises, although there is then a problem of how one
ought to choose the instrumental variables to be retained. This
is a question which I cover in my paper.
[HEIL
I. Regarding the difficulty of over-fitting in the reduced form,
Messrs. T. KLoEK and L.B.M. MENNES formulated a procedure (in
a recent issue of « Econometrica ») which is designed to handle
this problem, which is indeed serious when the number of pre-
determined variables is not small compared with the number
"2] Wold - pag. 54