Autocorrelation

[Brian Conway]

I'm running dwatson test (estat dwatson) after a regression and the DW statistic is 1.26. Since this falls outside the 1.5-2.5 range, it suggests the presence of AC. When I ran a breusch-godfrey test (esttat bgodfrey) on the same regression, it was not significant and thus suggests the absence of AC. I was wondering which test should I rely on? Thank you.

[Andrew Musau]

Since this falls outside the 1.5-2.5 range

Those limits look suspect (being round numbers). What is your sample size and how many RHS variables do you have?

[Brian Conway]

N=43 and number of IVs is 6. Thank you for your help.

[Andrew Musau]

For those values, \(d_L= 1.23\) and \(d_U= 1.83\), so you have \(d_L \leq d \leq d_U\) and you have inconclusive evidence for positive serial correlation. You should proceed "as if" you have positive serial correlation and use robust (Newey-West) standard errors if estimation is by OLS.

[Brian Conway]

Thank you very much for your help Andrew.

[Brian Conway]

I'd like to ask your help. In one regression, DW=.74 (3 IVs and 43 obs). Because this is well outside the 1.5-2.5. range for adjudicating DW, I concluded AC is present. As a check on this, I also ran bgodfrey, which was significant. Yet, when I look at .74 in relation to the table above, it gives 1.3683-1.666 as the range for adjudicating DW, which would suggest AC is not present. How should I adjudicate AC in this case? Thank you.

[Andrew Musau]

If all your RHS variables are exogenous, use the Durbin Watson test.