A Note on Exact Tests for Serial Correlation

Abstract

A transformation of the OLS residual vector to achieve a desired covariance structure, proposed earlier by Durbin, is shown to be capable of substantially changing the OLS residual vector even when that vector already has nearly the desired covariance structure. This may explain its substantially inferior performance in Monte Carlo comparisons with the transform proposed by Abrahamse and Koerts. A new transform, involving only small alterations in Durbin's procedure, is shown to avoid the defect of the Durbin transform.