[R-sig-ME] glmer overdispersion correction, family = binomial

[Paul Johnson]

On Fri, Mar 4, 2011 at 10:31 PM, Colin Wahl <biowahl at gmail.com> wrote:
> I do not feel as though I have a strong grasp of what a glme model is doing,
> and I am pretty lost when it comes to concepts like the "variance-covariance
> structure." Please feel free to suggest sources for understanding the
> mechanisms underlying glmms and/or "variance-covariance structure" if, in
> the following questions, you dont feel like being something of an
> instructor.
>

Wow, is this a great thread or what?

Here's my  "teacher" advice. You started doing ANOVA, now you're here,
sorta like wandering in from the wilderness into the opening of a dark
cave.   Right now, your comments make me think you are sticking your
sword into the darkness hoping to spear an apple that might be hung
from a string.

Take two steps back from the specific problem you are wrestling with
and try to understand random effects models in some other context.
Take a few days, do something else that is not so vital to your
research.

If you have not done it yet, You might investigate the way a Poisson
count model can exhibit "overdispersion", in which case it is common
to add a particular kind of additional individual level random error
(log Gamma) and it turns into a NegativeBinomial distribution.
There's a very good book (possibly even great) by Scott Long on
modeling Categorical and LImited Dependent variables.  I just noticed
yesterday in CRAN that there are several zero-inflated poisson
packages I had not seen before. This one, "ZIGP" has 2 competing
hypothesis tests to choose between fitted models, the vuong (which I
had heard of before) and clarke (which is new to me).   The pscl
package also has those models, plus the hurdle model.

The other thing worth looking at is the MCMCglmm package, partly
because it has nice detailed vignettes and the code is pretty easy to
decipher, but also because it is Bayesian, and you'll need to put some
of that into your head to understand the posterior analysis suggested
by glmer.  The author of MCMCglmm is a frequent contributor here.
Another package that I've learned from is glmmML, which. like lme4, is
for generalized linear models with random effects and clustering.
glmmML has an interesting function "glmmboot", which offers one  way
of assessing "clustering" effects in mixed models.

After you stretch your brain out a bit in a different direction, it
will be easier to understand what you need to do in the specific
context of your problem. Some of the questions you are asking about
are too specific for us to answer because we are not intimate with
your subject material.

Well, that's my bit!

-- 
Paul E. Johnson
Professor, Political Science
1541 Lilac Lane, Room 504
University of Kansas