Diana Virkki <d.virkki <at> griffith.edu.au> writes:
>
> Hi all,
>
> I am having some trouble running GLMM's and using model averaging with
> QAICc.
>
> Let me know if you need more detail here:
> I am trying to run GLMM's on count data in the package glmmADMB with a
> negative binomial distribution due to overdispersion. The dispersion
> parameter has now reduced to 2.679 for the global model (from a dispersion
> parameter of 27.507 with a poisson distribution), and I
> am not sure if this
> is still considered too high for running the models?
A dispersion parameter of 27 probably indicates something wonky
about the original data. I'm also surprised by a dispersion parameter
not close to 1 for the fitted NB model (as the NB model should in
principle take care of most of the overdispersion -- the mean square
of the Pearson residuals might be slightly different from 1, because
the NB shape/overdispersion parameter is calculated by ML, but this
is still a suspiciously large value).
>
> I would like to try to use QAICc's for model selection and model
averaging
> with the package MuMIn. I have so far been able to produce a QAICc output
> only for the models. I read that model averaging with QAICc can be done in
> MuMIn but cannot find the syntax to get these outputs, including the model
> weightings, parameter estimates, confidence intervals, and relative
> variable importance.
Can't help you there. In my experience MuMIn can only model-average
the wide range of model types it knows about, but there could easily
be features I don't know about.
> Any advice would be greatly appreciated. As well as if there are other
> potential better options for dealing with the overdispersion.
You probably need to look at your data more carefully -- do the
model fits seem reasonable? Are there big outliers, or zero-inflation,
or ... ?
If you are using glmmADMB for mixed model fitting, I would suggest
follow-ups go to r-sig-mixed-models at r-project.org ...
Ben Bolker