I had a time series model with 5 time series variables and it's a model that's reputed in the literature for having autocorrelation problems. Why when I use standard OLS var-covar matrix only 2 variables are (highly) significant but with Newey West varcovar matrix 4 of the variables are (highly) significant (1% level)?


1 Answer 1


There may be several effects. First, you may be hitting negative autocorrelations somewhere down the road. Second, most sandwich-like estimators are biased down, so with short samples (say of total length 10-15), you may have notably larger biases with the smaller data set.

If there are wild swings in significance, you need to proceed conservatively: report both, but act only on the evidence that's supported in both specifications.

  • $\begingroup$ I have 1000 time series observations and it's my HAC Newey West that is giving me significance across the board (you said they were "biased down"? Or did you mean SEs were biased down and not the test statistic?) $\endgroup$
    – user14281
    Oct 3, 2012 at 14:04
  • $\begingroup$ The standard errors are biased down. $n=1000$ should not give you trouble. Look at the ACF/PACFs to see if you can spot any negative autocorrelations of the residuals. May be your model is simply misspecified, and then there is little you can do about it other than adding some extra terms, lags, etc. $\endgroup$
    – StasK
    Oct 3, 2012 at 14:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.