Robustness to non-normality of the errors of both the null distribution and power of the Durbin-Watson (DW) test is examined in linear regression models with and without an intercept. Nominal critical regions (i.e., critical regions assuming the errors to be normal) are obtained exactly by numerical integration following Imhof (1961), and the actual critical regions when the errors are non-normal are obtained by simulation with 10,000 replications. A variety of non-normal errors and regressors are examined. It is found that the tests against the alternatives of positive lag-1 autocorrelation are relatively robust except in the case of regressions without an intercept and the regressors are stationary. In this case, the tests reject more often than that suggested by their nominal levels if the error distributions are highly skewed. The tests against the alternatives of negative lag-1 autocorrelation are found to be non-robust to skewed error distributions and in these cases, the tests reject far less often than that suggested by their nominal levels. In contrast, the powers are relatively robust. There is some evidence of minor losses in power but in most cases powers are increased if the errors are non-normal. In general, for a given error distribution, the power of the DW test is higher in regressions without an intercept as compared to powers with an intercept.
展开▼
