Tightening Durbin-Watson bounds

Abstract

The null distribution of the Durbin-Watson statistic, used in testing for serial correlation in regression disturbances, depends on the explanatory variable values. Although modern computing power makes it feasible to obtain exact critical values in any specific case, bounds to critical values used to be widely employed and sometimes still are. Much research was motivated by the ambiguity resulting from a test statistic falling inside the bounds, but one very simple approach to tightening bounds was overlooked. It is worth recording, if only for its pedagogic interest and possible applicability to other problems.