dave fournier otter at otter-rsch.com
Sun Nov 1 00:13:08 CET 2009

> > My questions are:
> >
> > 1) How can I extract the AIC values from m1?
> >
> > 2) Are the AIC values comparable between the two models (i.e. can I
> compare
> > them for model selection)?
> >
> > 3) For m2, the true estimates for the fixed effects can be calculated by
> > exp(returned estimate). Is this true for m1 too, or does the negative
> > binomial distribution require a different conversion?
> >

The simplest way is to use the save.dir option as in
if(nchar(pkg)) library(pkg, character.only=TRUE)})
random=~Site, group="Year", data=ex1m,
family="nbinom",zeroInflation=TRUE,save.dir="c:/dir1")
random=~Site, group="Year", data=ex1m,
family="nbinom",zeroInflation=TRUE,dir="c:/dir2")

The parameters together with estimated standard deviations will be found
in c:/dir1 and c:/dir2 in the file nbmm.std. the -log-likelihood is in
the file nbmm.par. that can be used for a likelihood ratio test on the
significance of adding the parameter T to the model.

The file nbmm.rep contains the predicted mean for each observation in
the second column of the list.  I believe this is what you mean by
"exp(returned estimate)".

The standard glmmADMB parameterizes the variance as mu*(1+mu/alpha)
However for you data it turns out that parameterizing the variance as
mu*tau produces a much better fit to the data.  The  log-likelihoods are
about -1530 and -1465 so you should use this formulation.

Interestingly it turns out that adding T  signficantly improves the fit
for the former parametrization of the variance but not for the latter
parametrization.

I think this illustrates the importance of estimating the overdispersion
within the model rather than by ad hoc quasi-likelihood hacks. It also
illustrates the utility of employing tools like AD Model Builder
which make it easy to  modify the form of your model rather than being

--
David A. Fournier
P.O. Box 2040,
Sidney, B.C. V8l 3S3