[R] Possible bug in lmer nested analysis with factors

Yan Wong h.y.wong at leeds.ac.uk
Tue Sep 20 12:04:33 CEST 2005

On 18 Sep 2005, at 16:04, Douglas Bates wrote:

> You are correct that good documentation of the capabilities of lmer
> does not currently exist. lmer is still under active development and
> documentation is spread in several places.  The vignette in the mlmRev
> package explores some of the capabilities of lmer.  Also see the
> examples in that package.

Yes. Thanks for this, and indeed for the development of the package.
I'm currently trying to do GLMMs (binary response), so I thought that
I should learn mixed modelling using a library with these capabilities.

> You are correct that the denominator degrees of freedom associated
> with terms in the fixed effects is different between lme and lmer.
> ...
> Some arguments on
> degrees of freedom can be made for nested grouping factors but the
> question of testing fixed effects terms for models with partially
> crossed grouping factors is difficult.

Would it not be possible to recognise when the model is fully nested,
and make this a special case? I was imagining using lmer as a
replacement for lme, so finding that they differ in this way came
as some surprise. When learning to use a new, relatively undescribed
routine, I usually try to see if I can reproduce known results. This
is where I was coming unstuck when trying to reproduce lme results
using lmer.

I suspect that many people (I know of one other in my group) will use
lmer as a drop-in replacement for lme specifically for its GLMM
capabilities rather than for its partial nesting. I realise, however,
that this might not be your priority.

> This area could be a very fruitful research area for people with
> strong mathematical and implementation skills.

That's not me, I'm afraid. I am only just working through Chapter 1
of your (excellent) "mixed effects models in S" book.

> There are already some facilities for lmer models such as mcmcsamp and
> simulate which can be used for evaluating the posterior distribution
> of a single coefficient or for a parametric bootstrap of the reference
> distribution of a quantity like the likelihood ratio statistic for a
> hypothesis test.

This, again, is beyond me at the moment. But I do hope that someone
else can respond to the call, especially for "textbook" as well as
more complex examples of lmer usage.

Best wishes

Yan Wong

More information about the R-help mailing list