GLS random-effect model with robust estimates

[Taha Nazari]

Hello all,

I am following a procedure discussed in a paper to analyze my longitudinal data. I have to apply a generalized least squares (GLS) random-effects model with robust errors.
I checked the -xtgls- command, but it has no option for neither random effects nor robust errors. I checked the forum to see if this issue was discussed before or not, and I encountered a comment by Carlo Lazzaro saying that:

GLS is a particular case of OLS when the residual distribution shows heteroskedastcity.In Stata, it simply implies to replace default with -robust- standard errors

I was wondering that Generalized least squares model is the same as -xtreg [varlist], re vce(robust)-?
If no, how I can achieve random-effect model with robust errors using -xtgls-?

[Carlo Lazzaro]

Taha:
-xtgls- was conceived for T>N panel datasets (unlike -xtreg- that was developed for N>T panel datasets).
As you mention, it does not return -re- but has options for different error structure (see -xtgls- helpfile) that make the clustered-robust standard error redundant.

[Joro Kolev]

There are different schools of thought of whether calculating robust variance post GLS makes sense or not. One school holds what Carlo said. Another school says that you might specify a GLS weighting matrix, but nobody can guarantee that you have specified it correctly, hence robust variance is desirable and needed post GLS too.

Once upon the time back in 2014, I wrote a paper on how to calculate robust variance post -xtgls-. I was not fortunate with the referee at Stata Journal, he did not understand the issue, and my paper was rejected.

I am attaching the paper which describes how to calculate robust variance post -xtgls-.

[Taha Nazari]

Thanks to both Carlo and Joro.

Carlo:
I will definitely check your paper. That would be a huge help.

[Carlo Lazzaro]

Taha:
as far as the paper is concerned, it is not Carlo, but Joro!