Page 684 -

The ADL Model and Generalized Least Squares in Lag Operator Notation 683

introduces bias; however, because there are few coefficients, the variance of the estima-
tor of the dynamic multipliers can be small. In contrast, estimating a long distributed lag
model using GLS produces less bias in the multipliers; however, because there are so
many coefficients, their variance can be large. If the ADL approximation to the dynamic
multipliers is a good one, then the bias of the implied dynamic multipliers will be small,
so the ADL approach will have a smaller variance than the GLS approach with only a
small increase in the bias. For this reason, unrestricted estimation of an ADL model
with small number of lags of Y and X is an attractive way to approximate a long distrib-
uted lag when X is strictly exogenous.