I am running a time-series regression. The Durbin-Watson statistics is very close to 2. In such a situation, would it still be better to use Newey-West standard errors, or is it ok to use OLS standard errors?
1 Answer
First, I would recommend to use a software package that not only reports the Durbin-Watson test statistics but also a p-value. That might give you more of an indication how close or far from 2 the statistic actually is. Furthermore, you may consider other tests for autocorrelation, e.g., Breusch-Godfrey etc. And if there is no evidence for autocorrelation (or heteroscedasticity) then the OLS standard errors are probably fine.
A pragmatic approach could also be to simply try whether using Newey-West standard errors makes a relevant difference. If the autocorrelations in the residuals are small, then Newey-West should lead to very similar results anyway.
-
1$\begingroup$ +1. Regarding the last paragraph: In small samples, as in below 50 or so observations, Newey West standard errors often give oversized tests though. $\endgroup$– KOEApr 4, 2015 at 19:49
-
2$\begingroup$ Yes, good point. With all HC and HAC standard errors one should also keep in mind that they are less efficient if the standard assumptions do hold. $\endgroup$ Apr 4, 2015 at 21:46
-
$\begingroup$ Thanks very much. The sample size is above 100 observations. I've checked Newey-West standard errors and they are indeed not too far from HC standard errors. $\endgroup$ Apr 6, 2015 at 10:18