[R-sig-ME] lme4a, glmer and all that

[Douglas Bates]

This is a good news - bad news announcement.  As Ben Bolker indicated
in a recent thread, the lme4a version of the lme4 package is
undergoing rapid development and the capability of fitting generalized
linear mixed models with glmer has recently been added.

These capabilities are changing rapidly and some aspects of the model
fits are yet incomplete.  For example, the standard errors of the
fixed-effects parameters are not yet updated for the final parameter
estimates so don't pay too much attention to those standard errors.
Also, you do need to call glmer (i.e. don't try calling lmer) to fit a
GLMM.  In the past the structures for lmer and for glmer were the same
with some flags indicating which type of model was active.  Now they
are different.

So the good news is that the brave and foolhardy can try installing
that version, once we settle the issue of a dependency on an
unreleased version of the Matrix package (Martin, can we back that out
for the time being?).

The bad news is that I can't reproduce the results from a GLMM fit
using the currently released lme4 package and I am suspicious that the
results from the current version of lme4 are the inaccurate ones.
Although I'm still trying to verify this (and I would appreciate help
- see below), I would recommend that you do not quote results from a
GLMM fit with version 0.999375-32 of the lme4 package.

Having others trust the results from this software and then find out
that these are wrong is about the worst thing that I can imagine
happening in my work.  If this is the case I sincerely apologize.

I would ask two things.  If this is an error in the lme4 software,
please do not phrase it as "an error in R" and try to correct others
if they claim this.  There is often confusion between the R system and
environment and the results from individual packages.  If there is an
error then it is an error in the package.  In particular it would be
an error in the derivation that I did and the code that I wrote and I
take full responsibility for it.  However, I do not want this misstep
to be phrased as "the results from R can't be trusted".

The second thing I would ask is if others (in particular, those who
don't mind installing a package that is undergoing rapid development)
can cross-check results from the two versions against other software
to which you may have access, such as Stata or SAS.  You can either
post to the list or send results to me privately.

Thanks for your understanding.  We will try to clear this up as soon
as possible.