st: Comparing model-fits of multi-level models with the AIC when estimating with REML

[Schmid Samuel <Samuel.Schmid@unilu.ch>]

Hi,

I am using -xtmixed- in order to fit two-level models. There are only 20 ca= ses on the macro-level, so I decided to estimate the models using REML (res= tricted maximum likelihood).

Now, I would like to compare the model-fits using the Aikaike Information C= riterion (AIC). Given that I use REML, is it still sensible to compare the = AIC for models that differ in their fixed-effect specifications? Or is this=  not appropriate for the same reason that I cannot do a Likelihood-Ratio-Te= st to compare models with different fixed-effects specifications when using=  REML?

If it is not appropriate to do this, what are my alternatives to get a usef= ul comparison? I know there are constructions / approximations of R2 which = measure the 'proportion of explained variance' for each level of the models= . But this measure and its original notion seem problematic in the context = of multi-level models, so I would like to get a more suitable device for co= mparison.

Thanks a lot for your help!

Samuel

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/