R help - Apr 2005 - Summary: GLMMs: Negative Binomial family in R

Here is a summary of responses to my original email (see my query at the
bottom).  Thank you to Achim Zeileis , Anders Nielsen, Pierre Kleiber  and Dave
Fournier who all helped out with advice.  I hope that their responses will help
some of you too.

 *****************************************
Check out
glm.nb() from package MASS fits negative binomial GLMs.
*****************************************

For known theta, you can plug negative.binomial(theta) into glmmPQL()
for example. (Both functions are also available in MASS.)

Look at package zicounts for zero-inflated Poisson and NB models. For
these models, there is also code available at
  http://pscl.stanford.edu/content.html
which also hosts code for hurdle models.
*****************************************

Consider using the function supplied in the post:
https://stat.ethz.ch/pipermail/r-help/2005-March/066752.html
for fitting negative binomial mixed effects models.

*****************************************

Check out these recent postings to the R list:
http://finzi.psych.upenn.edu/R/Rhelp02a/archive/48429.html
http://finzi.psych.upenn.edu/R/Rhelp02a/archive/48646.html
*this refers to the  random effects module of AD Model Builderthat can be called
from R via the driver functon glmm.admb(). Their example problem fits the model
with a negative binomial. The function can be downloaded from
http://otter-rsch.com/admbre/examples/nbmm/nbmm.html

************
***********
My Original Query

Greetings R Users!

I have a data set of count responses for which I have made repeated observations
on the experimental units (stream reaches) over two air photo dates, hence the
mixed effect.  I have been using Dr. Jim Lindsey's GLMM function found in
his
"repeated" measures package with the "poisson" family.

My problem though is that I don't think the poisson distribution is the
right
one to discribe my data which is overdispersed; the variance is greater than
the mean.  I have read that the "negative binomial" regression models
can
account for some of the differences among observations by adding in a error
term that independent of the the covariates.

I haven't yet come across a mixed effects model that can use the
"negative
binomial" distribution.

If any of you know of such a function - I will certainly look forward to hearing
from you!  Additionally, if any of you have insight on zero-inflated data, and
testing for this, I'd be interested in your comments too.  I'll post a
summary
of your responses to this list.

Best Regards,
Nadele Flynn, M.Sc. candidate.
University of Alberta