GLS or FE models?

[Guest]

Hi everyone!
I have a strongly balanced panel dataset (N=50, T=8) and I've estimated two models:
1) FE model:

xtreg y x1 x2 x3, fe

In this case I have evidence of time-fixed effect, so I have modified the model as follow:

xtreg y x1 x2 x3 i.year, fe

2) GLS model:

xtgls y x1 x2 x3, igls panels(heteroskedastic)

The easy question is: if I have evidence of entity and time fixed effects, which one should I choose?
Because I know that FE controls for these factors...but GLS? it controls or exists commands to control for these factors?
Thanks in advance for your kind help!

[Carlo Lazzaro]

student90 (please, see FAQ #6 on preferred registration requirements on this forum, Thanks);
you report a large N, small T panel dataset, whereas -xtgls- is better suited for long panels (i.e., small N, large T).
I would recommend you to stick with -xtreg- (assuming that your depvar is a linear one, a detail that was left unreported in your post).
Eventually, please consider that -xtset panelid year- will automatically take time effect into account; hence, despite including -i.year- in the right hand side of the equation, you might find no time trend at all (or a statistical significance for a limited number of years only),

[Guest]

Thanks a lot Carlo! I'm sorry I was imprecise
Yes in my case the depvar is linear.
I used i.year because running a testparm (to see if dummies for all years are equal to zero) I found that they are jointly significant (six of them are significant P>|t| = 0.000...only two are not). I'm following the proceeding suggested on http://dss.princeton.edu/training/Panel101.pdf (slide 31).

I knew that FE takes into account time-invariant effects, but what about unobserved effects that vary across time, but not across individuals?
If I've understood properly, the initial command xtset countries year automatically takes into account these effects?

Best regards,
Giovanni

[Andrew Musau]

Hi Giovanni

What Carlo implies is that there should not be a big difference between these two regressions and not your interpretation.

xtreg y x1 x2 x3, fe
xtreg y x1 x2 x3 i.year, fe
Reverting to your question, fixed effects controls for time invariant unobserved effects since these are wiped out by the fixed effects transformation (by taking deviations from the firm means). If you consider that the unobserved effect is time varying, then the effect will not be wiped out - and your estimates will be biased as with OLS.

Therefore, with time varying unobserved effects, you will need to consider other methods such as instrumental variables or some form of quasi-experimental variation to avoid this bias. There is a lot of literature on this - especially as it relates to the estimation of the effect of education on earnings (where the unobserved effect is ability)- which you should look at.

[Carlo Lazzaro]

As far as the theoretical background implied in his post is concerned, Giovanni may also want to take a look at the following Paul Allison's textbook http://www.stata.com/bookstore/fixed...els/index.html.

[Guest]

Thanks a lot for all the suggestions!!! They were really useful!

[Kenny Putri]

Hi All
I want to ask something related to this. So, basically I have long panel dataset. When I checked the result for hausman test, it shows that I have to use FE. But also I have several problems like multicol, autocorr, and heteroscedastic, and when I do robust test, it shows bad result (only one variable is significant). But when I try to fix it with xtgls the result was good. And now I'm confused whether I have to use xtgls or xtreg depvar indepvar, fe ro. because I heard that xtgls is just same with RE. So, what should I do? sorry for my bad english

[Carlo Lazzaro]

Kenny:
replies in this thread can be useful for your probelm, too.
As an aside, I would not judge a statistical approach good or bad solely on the ground of a statistical significant outcome.
For the future, I would recommend you to start a new thread. Thanks ans welcome to the list.