The Relative Power of Zero-Padding When Testing for Serial Correlation Using Artificial Regressions.

Artificial regression allows a simple and flexible test of serial correlation with many virtues that promote it, in principle, to a position of dominance. But it has a serious small-sample problem: successively truncated lagged residual regressors reduce limited d. of f. twice, simultaneously reducing T and increasing K. It is therefore with interest that one learns one can pad-out the truncated residuals with zeros and the test remains asymptotically valid. Of course, asymptotic virtues are small comfort with limited d. of f., so one wonders about the small-sample effectiveness of this zero-padding procedure. The Monte Carlo study presented here addresses this issue: in a sample of size 20, there appears little, if any, gain to zero-padding and, indeed, in the most common cases, zero-padding results in marginally reduced power. Citation Copyright 1996 by Kluwer Academic Publishers.