Skip to content
Prev 8620 / 20628 Next

overdispersion in GLMMs (Alejandro Mart?nez Abra?n)

Elizabeth and Alejandro--

We discuss the over-dispersed Poisson mixed model (with per-observation 
random effect) in the following tutorial:

Atkins, D. C., Baldwin, S., Zheng, C., Gallop, R. J., & Neighbors, C. 
(in press). A tutorial on count regression and zero-altered count models 
for longitudinal substance use data. Psychology of Addictive Behaviors.

which you can find along with data and R code:

http://depts.washington.edu/cshrb/newweb/statstutorials.html

Hope that helps.

cheers, Dave

 > 1. How can I estimate overdispersion in a Poisson GLMMM?
 >
I usually fit overdispersed Poisson (or Binomial) GLMMs by adding a
unique identifier for each observation, then adding that unique ID as
a random term.  You can "test" how overdispersed the model is by
looking at the standard deviation assoicated with that random effect,
or by comparing the fit of a model with the overdispersion term to one
without it.

Alternatively, you can look at overdispersion due to random effects of
individual, plot, etc, using the same basic procedure.

I think I got the idea from the Gelman et al. Bayesian stats text.  I
would be curious to know if others do this also.

 >
 > 2. I am trying to run a quasipoisson GLMM using the lmer function and the
 > lme4 library but I get a
 > warning stating that "glmer cannot deal with quasi error families? 
Any tip?
 > because I have seen this done.
 >
My understanding is that this functionality has been removed, since it
is +/- redundant with the approach used above, but less naturally
linked to the mixed model framework.


-- 
****************************
Elizabeth E. Crone
Senior Ecologist, Harvard Forest
Harvard University
Petersham MA 01366
office: (978)756-6145
main: (978)724-3302
cell: (406)531-3498
FAX: (978)724-3595
email: ecrone at fas.harvard.edu