Am 08.06.2015 um 17:45 schrieb JOSE FRANCISCO PERLES RIBES:
"estat ic (the command Stata uses to provide AIC and BIC) is not
appropriate after xtreg with the be , pa, or re (random effects) option"....
However, I have just checked that Gretl provides and AIC and BIC values
after random effects and between effects estimation.
Somebody has a clue because Stata warns about this AIC and BIC criteria
with this models??? Are the values provided by Gretl reliables to
compare this kind of models or there are a better criteria to compare
them?? Could be this issue due for different estimation methods in Stata
and Gretl for panel data???
I can only offer a partial explanation.
First, the standard RE estimator is not MLE. But the information
criteria are based on (maximized) likelihoods. The BE estimator could be
interpreted as MLE I guess, but for a different transformation of
dependent variables; maybe that's why Stata doesn't give you the
likelihood (and thus the ICs) to avoid misunderstandings and wrong
Now why does gretl give you the AIC/BIC? I'm not 100% sure but one
justification in the RE case would be that it's reasonably close to MLE
most of the time. Can somebody else comment?
For BE my first reaction would be as above: It's not comparable, even
though it's fine on its own.