[R] negative binomial lmer

Douglas Bates bates at stat.wisc.edu
Sat Jul 29 17:25:45 CEST 2006

On 7/28/06, Gregor Gorjanc <gregor.gorjanc at bfro.uni-lj.si> wrote:
> Ben Bolker <bolker <at> ufl.edu> writes:
>
> ...
> >  I haven't tried it, but you could also consider using
> > a Poisson-lognormal (rather than neg binomial, which is Poisson-gamma)
> > distribution, which might make this all work rather well
> > in lmer:
> >
> > www.cefe.cnrs.fr/esp/TBElston_Parasitology2001.pdf
>
> Actually it is very simple
>
> lmer(y ~ effA + (1 | effB), family=quasipoisson)
>
> i.e. this fits the following model
>
> y_ijk ~ Poisson(\lambda_ijk)
> log(lambda_ijk) = \mu + effaA_i + effB_ij + e_ijk
> effB_i ~ Normal(0, \sigma^2_b)
> e_ijk ~ Normal(0, \sigma^2_e)
>
> Gregor

I would advise checking the results from lmer against those from
another way of fitting this model or the negative binomial model.
There may be a problem in the way that lmer handles the scale
parameter.  I haven't checked  generalized linear mixed models with a
scale parameter as extensively as I have checked those without a
separate scale parameter (the binomial and Poisson families).  If
anyone can provide me with an example of such a model and sample data
(preferably off-list) I would appreciate it.