[R-sig-ME] glmmadmb Negative binomial dispersion parameter

[Paul Johnson]

Hi Alain,

>> I forgot to mention that the more and more I look at the NB GLM(M) the less I like it. You should only go for a NB GLM(M) is the cause of the overdispersion is large variation. If there is something else that is causing overdispersion (e.g. non-linear patterns, zero inflation, missing covariate, wrong link function), then the parameter theta is going to consume that information.

I agree that the overdispersion term will mop up all the “unexplained" variation, whether this is due to bad explanation (= a poorly specified model, e.g. nonlinearity, ZI, etc) or just the fact that even good explanations have limits, and we should of course be careful to avoid bad explanations. However in my experience of biological count (or any) data, even if we take care to fit the model carefully there will generally be a substantial amount of unexplained (and unexplainable) variation, which is why I tend to include an overdispersion term as a matter of course, whether using negative binomial or Poisson-lognormal GLMMs. This doesn’t stop us looking for zero-inflation etc. E.g. if there’s real zero inflation, a NB GLMM with ZI should still fit better that a NB GLMM without ZI.

>> And I still need to see the first data set for which the NB GLM(M) gives predictions with decent confidence intervals.

Concerning — I’d be interested in knowing more. How are you calculating the CIs? Are you talking about NB GLM(M)s in general, or glmmADMB? 

All the best,
Paul