xtgls measure of fit

[Steffen Heinig]

Hi Statalist Forum,

I have another quick question regarding panel data sets.
Someone made me aware of using xtgls instead of xtreg, fe.
I have a panel data set with 80 regions over 44 quarters. As I have read other posts here, is this already considered as big id and small t? And if so, would xtgls unable to help?

Second, as far as I have understood this, xtgls would allow to include heteroskedasticity as well as AR1 for each panel, which is good for a data set with many different regions.

I am using:
xtgls $ylist $xlist, panels(hetero) corr(ar1) force

Nevertheless, this above method will not provide any information about the Goodness of fit (AIC and BIC, as well as Likelihood).
I tried estat ic, with the following result:


. estat ic
likelihood information not found in last estimation results
r(321);

However, when I remove the specifications form the regression command a likelihood is reported, unfortunately with homoskedasticity assumption. Is there another way to get any measure out of my command?

Are the advantageous of xtgls so much better than xtreg, fe???

Picture enclosed with an output.

Many thanks for some help.
Kind regards,
Steffen

[gabriela callo]

Hi Statalist forum,

Nowadays, we are handling a similar question. We are using xtgls $ylist $xlist, panels(hetero) corr(ar1) force, and we would like to obtain the AIC and BIC criteria. Is there any possibility to calculate the AIC and BIC for this model?
Thanks in advance for the answer.
Best regards,

Gabriela CQ

[Rich Goldstein]

according to the help file, "estat ic and lrtest are available only if igls and corr(independent) were specified at estimation." see

help xtgls postestimation

added in edit: clarification - these 2 options are needed to make the -xtgls- estimator a max likelihood estimator and AIC and BIC are only available for MLE's

[gabriela callo]

Thanks Rich,

Our data presents autocorrelation, should I use corr (independent) even knowing the autocorrelation of our data?

Actually, I am looking for a way to choose the best model. I was using AIC and BIC until xtgls. Is there other way to compare models without using MLE, or, is there a way to calculate the MLE for xtgls, using panels(hetero) corr(ar1)?

I really appreciate the answer.

Best regards,

Gabriela CQ

[Rich Goldstein]

gls is not an MLE so, as far as I know, the only way to get an MLE via -xtgls- is as I said in #2 above; there are other ways to compare models, but, as with AIC and BIC, each has strengths and weaknesses and I have no idea what is important to you (e.g.,are you concerned with comparing predictions or are you concerned with the stability of the estimates as the data change or are you worried about a particular assumption and want your estimates robust to that, or. ...?