[R-sig-ME] Over-dispersion and lmer

mikhail matz matz at mail.utexas.edu
Tue Dec 20 18:59:26 CET 2011


Hey Ricardo - 

I faced this problem myself very recently, and there is a recent exchange on the subject which seems to be an easy solution. Look up R-sig-mixed-models Digest, Vol 47, Issue 53; thread by David Stainbrook. Briefly, you can effect modeling with negative binomial distribution (which takes care of the overdispersion) on counts data by using lognormal poisson distribution. You just have to include a random effect corresponding to every observation in your data (that is, add something like (1|index) to the formula) and specify poisson family (which will by default imply log-link).  You will get a warning ("Number of levels of a grouping factor for the random effects is *equal* to n, the number of observations." ) but this does not seem to matter. There is even a reference to a paper laying down the theory: Elston et al, Parasitology (2001), 122, 563-569.

cheers

Misha Matz



More information about the R-sig-mixed-models mailing list