In a time-series OLS regression of daily stock returns on a set of explanatory variables I want to test whether the error terms are auto-correlated before I decide to use Newey-West standard errors (heteroskedasticity is present).

Using the Breusch-Godfrey test, I have to specify a number of lags that shall be considered in the test for serial correlation.

  • Is there any intuition (maybe an economic one?) for considering more than first order serial correlation?
  • If yes, is there any theory or empirical finding that can be used to choose a reasonable lag length up to which I have to check?

Thank you very much for your help!!

  • $\begingroup$ Use back ticks ` for code, do not italicize. Also, related question: "Breusch-Godfrey Test and the length of the lag, p". $\endgroup$ – Richard Hardy Jun 26 '16 at 18:10
  • $\begingroup$ Thanks for your reply, I edited the formatting. I also noticed the question (and upvoted it) but as it is quite old and still unanswered I thought I'd provide some additional information on my specific setting in the hope someone could help out here. $\endgroup$ – Juliett Bravo Jun 26 '16 at 18:30
  • 1
    $\begingroup$ Since you are using stock returns, I dont see why you want to include more than say, 10 lags... In any case, the test is not really that general, I would use NeweyWest standard errors in any case just to be sure. The gain of doing so is significant (i.e. robust standard errors), there is no real cost. $\endgroup$ – Repmat Jun 28 '16 at 11:10
  • $\begingroup$ Thanks for your reply @Repmat! I have a variety of regressions (about 20) that use the same set of controls and only differ in their explanatory variable. I wanted to check each regression for heteroskedasticity and serial correlation. In case only the former is present I wanted to use White standard errors, if both is present Newey-West. Would you consider this as too cumbersome? And is there really no cost in doing so? Because there are some recognized papers that "only" use White, so I wonder how they decided to not use Newey-West. $\endgroup$ – Juliett Bravo Jun 28 '16 at 12:22
  • 1
    $\begingroup$ In general, I can't picture any modern paper that would accept you not using robust standard errors - especially once you go into time series (unless you are doing some very specific simulation or similar). Both HC and HAC estimator are only valid asymptotically, I suppose that the HC estimator is a bit more efficient (since you only have one level of uncertainty) but from a practional point of view it should not matter (which is what I meant to imply) $\endgroup$ – Repmat Jun 28 '16 at 12:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.