An approximation to the null distribution and power of the Durbin—Watson statistic

Abstract

It is computationally complex to obtain the exact distribution of the Durbin—Watson statistic for detecting autocorrelation in regression residuals. Moreover, the distribution cannot be tabulated once and for all because of its dependence on the regressors. It is found that the distribution can be well approximated by an appropriate. Pearson distribution which is computationally simple, easily programmed and seems to be valid when the regressors are commonly observed economic time series. It is also shown that the power of the statistic is relatively insensitive to such choices of the regressors and to a variety of correlation structures of the regression residuals.