R help - Dec 2005 - generalized linear mixed model by ML

Dear All,

I wonder if there is a way to fit a generalized linear mixed models (for
repeated  binomial data)  via a direct Maximum Likelihood Approach. The
"glmm" in the "repeated" package (Lindsey), the
"glmmPQL" in the  "MASS" package (Ripley) and
"glmmGIBBS"  (Myle and Calyton) are not using the full maximum
likelihood as I understand. The "glmmML" of Brostrom uses the
"full maximum likelihood" by approximating the integral via  Gauss-
Hermite  quadrature. However, glmmML is only valid for the random intercept
model and the binomial family must be represented only as  binary data. Does the
lmer do the work?

The reason for asking that is:

McCulloch (1997): "Unfortunately, PQL methods can perform poorly relative
to Maximum likelihood"
Agresti (2002) p. 524 "........but where possible we recommend using ML
rather than PQL".


Thanks a lot,

Abderrahim 

	[[alternative HTML version deleted]]

Abderrahim Oulhaj <abderrahim.oulhaj <at> pharmacology.oxford.ac.uk>
writes:
> 
> Dear All,
> 
> I wonder if there is a way to fit a generalized linear mixed models (for
repeated  binomial data)  via a direct
> Maximum Likelihood Approach. The "glmm" in the
"repeated" package (Lindsey),
the "glmmPQL" in the 
> "MASS" package (Ripley) and "glmmGIBBS"  (Myle and
Calyton) are not using the
full maximum likelihood as I
> understand. The "glmmML" of Brostrom uses the "full maximum
likelihood" by
approximating the integral
> via  Gauss- Hermite  quadrature. However, glmmML is only valid for the
random
intercept model and the
> binomial family must be represented only as  binary data. Does the lmer do
the
work?
> 
  Hmmm.  I will be interested to hear what others have to say on
this topic.  

* lmer() in the lme4 package (new version of nlme) can in
fact do GLMMs with a choice of different
integration methods (PQL is the default but not the only choice).

* GLMMGibbs [sic] was actually
using a full likelihood rather than an approximation, but was
a Bayesian rather than a ML approach [GLMMGibbs is now in the
"Devel" section of CRAN, apparently because of various unresolved
compilation/installation problems.]

  If you want to fit temporal correlation, as well as 
individual random effects, you may be out of luck: a GEE
model will probably be your best bet in that case.  When I asked
about the possibility of incorporating temporal/spatial correlation
structures like those in nlme into lme4, Doug Bates said that he
wanted to work first on getting the basic framework of the package
really solid [can't blame him at all, and of course honor and
glory to him for putting so much work into these tools in the first place]

   good luck,
    Ben Bolker