glmmPQL model selection

Emmanuelle TASTARD <tastard <at> cict.fr> writes:

Hi,
IÂ’m sorry, I know that it is a recurrent question but I have not been
able to find the response in the Rhelp archives.
I think my data require the use of the glmmPQL function but I do not
know how to make the model selection. Since the AIC and log-likelihood
are apparently meaningless, how can we select the parameters for a model
and compare the models to find which one fits best the data?  

I think your choices are (1) use the estimated standard errors/p-values
of the fixed effects to decide whether to include them in the model
or (2) if you really need likelihood-based tests, use lmer.  
(Model selection for variance parameters is a can of worms, see
Pinheiro and Bates.)  Also remember that *all* methods for this kind
of model are approximations, it's just a question of which ones are
more accurate (generally and in particular situations).

That's just my best guess, someone else may have better advice ...

   cheers
     Ben Bolker