[R-sig-eco] Observation-level random effects to account for overdispersion

[Dixon, Philip M [STAT]]

All,

I am a real fan of using observation-level random effects to account for overdispersion with count data.  Using the default log link, an observation-level normally-distributed random effect models the data as a convolution of a Poisson and log-normal distribution.  The negative binomial is the convolution of a Poisson and gamma distribution.  For all practical purposes, the log-normal and gamma distribution are indistinguishable.  Computationally, the normal model is much easier especially if you have additional normally distributed random effects.  The sum of normals is easy to model; the sum of some normals and a log gamma is not.

Philip Dixon