GLS versus Robust

[amin khan]

Respected Members,
There is mix discussion about the correction of heteroskedasticity and auto-correlation.. some scholar suggest for Robust whereas other asking to apply GLS. would any member please explain which option is best and why ?

[Carlo Lazzaro]

Amin:
first step: are you dealing with a large N, small T panel dataset or the other way round?

[amin khan]

Carlo Lazzaro, i have panel data-set. i am observing 30 firms data for 10 years for my study. total number of observation is 300

[Carlo Lazzaro]

Amin:
you are dealing with a large N, small T panel dataset.
Your friend is -xtreg- with robust or clustered standard errors (if you suspect heteroskedasticity and/or autocorrelation).
Please note that, unlike -regress-, under -xtreg- both -vce(robust)- and -vce(cluster clusterid)- options do the same job.

[amin khan]

Carlo Lazzaro, actually i am confused about selection of Robust method and GLS. i used both method but with robust, my result become insignificant. a supervisor also suggested me to run GLS to correct Hetroskedasticity and autocorrelation.my question can i run GLS instead of robust?

[Carlo Lazzaro]

Amin:
supervisors' suggestions should be carefully considered.
As an aside, -xtgls-; -xtpcse- and -xtscc- (see http://www.stata.com/bookstore/micro...ata/index.html for furher details) seems to advised for small N, large T panel datasets, too (i.e., the opposite that you have).
The lack of statistical significance of your results (by the way: I do not think that statistical significanc should be the goal of inference) may also be due to a limited sample size.

[amin khan]

Thank you for sharing your knowledge Carlo Lazzaro

[Carlo Lazzaro]

Errata corrige: my reply #6 is miswritten: -xtreg- applies to large N, small T panel datasets (as in your case) , whereas -xtpcse- and the like are advised for small N, large T panel datasets. Sorry for the mishap and thanks to Jol Blanche for pointing this out.

[Jesse Wursten]

Amin:
you are dealing with a large N, small T panel dataset.
Your friend is -xtreg- with robust or clustered standard errors (if you suspect heteroskedasticity and/or autocorrelation).
Please note that, unlike -regress-, under -xtreg- both -vce(robust)- and -vce(cluster clusterid)- options do the same job.

Just a very minor remark, but robust and cluster only return the same result if you cluster on the panel identifier. E.g.

webuse abdata
xtset id year

*** Same standard errors
xtreg n w k, r
xtreg n w k, cluster(id)

*** Different standard errors
xtreg n w k, r
xtreg n w k, cluster(ind)

[Carlo Lazzaro]

Good point, Jesse.