[R] glmmPQL, log-likelihoods issue

[Douglas Bates]

Andrew Perrin <clists at perrin.socsci.unc.edu> writes:

> Sorry for my ignorance, but could you explain a little further?  I'm
> guessing from your response that this makes the log-likelihood that is
> quoted by glmmPQL a poor measure of model fit. Are there are statistics
> that would be better for reporting model fit?

You could try GLMM from the lme4 package instead.  It has two methods
of fitting the model - PQL and Laplace.  The parameter estimates from
PQL should be similar to those from glmmPQL (it's essentially the same
algorithm) but the log-likelihood reported by GLMM is that evaluated by the
Laplacian approximation.

If you choose method="Laplace" then GLMM does the PQL fit followed by
further iterations to optimize the (second-order) Laplacian
approximation to the log-likelihood.  This takes longer to fit but
should be more accurate. The log-likelihood from this fit should be
greater than that from the PQL fit for the same model.  These
log-likelihood can be compared between models.