xtpcse produces different coefficients

[Hovhannes Nahapetyan]

Hi all,

I hope someone can help me understand why xtpcse produces different coefficients. I specify the following serial correlations:

xtpcse DV IV(s) i.country (country dummies) - this one assumes no autocorrelation

xtpcse DV IV(s) i.country (country dummies), corr(ar1) - this one assumes common AR(1)

xtpcse DV IV(s) i.country (country dummies), corr(psar1) pairwise - this one assumes panel-specific AR1

Of course I should get different standard errors and P-values since each regression assumes certain type of autocorrelation (first one imposes no autocorrelation) but that should not affect coefficients for IV(s). However, each model produces different IV coefficients in addition to different standard errors and P-values (the later of course expected)


Thanks

[Andrew Musau]

If I recall correctly, xtpcse switches from OLS to Prais–Winsten estimation with serial correlation of type AR(1). Therefore, if the estimators are different, you should not expect the same coefficient estimates. You should further read

help xtpcse

[Hovhannes Nahapetyan]

Hi Andrew

Thank you for your response. Prais–Winsten estimation is a FGLS estimator and Beck in his 2001 article ("Time-Series–Cross-Section Data: What Have We Learned in the Past Few Years?" ) argues against using FGLS ( Park-Kmenta) and instead recommends using OLS. So it's somewhat surprising that xpcse gives that option of Prais–Winsten estimation unless the latter does not have the FGLS estimation issues criticized by Beck in the above mentioned paper. Would you be able to comment on this issue.

Thanks

[Andrew Musau]

I have not read the Beck paper, but Prais–Winsten estimation is quite a standard procedure to deal with AR(1) type serial correlation in a linear model. Yes, the procedure is FGLS.

[Hovhannes Nahapetyan]

Thank you very much Andrew