Autocorrelation and heteroskedasticity

[Brian Conway]

I'm running time series regressions with a small dataset (about 50 obs) and wondering about correcting for auto correlation and heteroskedasticity in the same regression model. I came across past literature suggesting that it is best to correct for auto correlation first and then heteroskedasticity: thus, in a regression (where Durbin-Watson statistic was .79 and the hettest was significant and chi2=6.34) I tried to correct for auto correlation using <prais x1 x2 x3, corc> and then ran <regress x1 x2 x3, vce(robust)> to correct for heteroskedasticity. I was wondering if this is correct? Thank you for any help you can give.

[George Ford]

prais permits vce(robust). The prais takes care of AC and the robust the hetero.

[Brian Conway]

Thank you very much George.

[Jeff Wooldridge]

With only 50 observations, you might want to avoid GLS. You can use the newey command with a small lag, probably two or three.

[Brian Conway]

Thank you very much Jeff. I had come across the newey option [https://www.montana.edu/cstoddard/56...orrelation.pdf (please see second last page)] but was bit hesitant to use as wasn't sure whether works well with small N. Would you still recommend?

[Brian Conway]

I was wondering if newey command with a lag of 2 is ok to use for yearly data (60 obs in one case, about 40 obs in another)? Thank you.

[Andrew Musau]

See #4 for some rules of thumb: https://www.statalist.org/forums/for...lation-command

[Brian Conway]

I have a data source published in volume 1999 and the data reported in it relate to 31 December 1997. I’d like to check that I’m correct in assuming that the data relates to 1997 (and not 1998)? Thank you.