I am performing 5-year rolling regressions over a 90-year period with monthly data and wanted to test for autocorrelation in the residuals. A Durbin-Watson test on the full sample leads to the conlusion that there is a positive 1st-order autocorrelation, while a DW test on the first window leads to the acceptance of the null hypothesis.
So please forgive me this neophyte question, but should I perform the DW test on the full sample, or on each window?
In this later case if there is AC for just a few windows, do I perform a GLS regression for each to be consistent? If there is no AC it should provides the same estimates than OLS anyway, even if OLS is more effective when this condition is not violated, isn't it?
