[ADMB Users] Comparing non-nested models?

[Ben Bolker]

On 11-10-27 11:35 AM, dave fournier wrote:
> Why can't you use the likelihood ratio?  The zero inflated Poisson is
> a limit of zero inflated NB.  Presumably the log-likelihood for
> the NB is quite a bit higher if you suspect over inflation.
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users

  Exactly.
  You can use the likelihood ratio test; if you're being careful you
need to account somehow for the fact that the null hypothesis (zero
overdispersion) is on the boundary, so the LRT is not technically
applicable -- but in fact in this case it's conservative, and the
general solution is to halve the observed p-value. There's some
discussion of this on http://glmm.wikidot.com/faq (I think) and in the
Bolker et al 2009 TREE paper.

  I will check anova() to see what's wrong with printing the models.

  With a 25 deviance-unit difference between the models the details
of your test are pretty much moot.

  For what it's worth the same description applies to testing for
zero-inflation (it is a nested model, but the null hypothesis is on the
boundary (zero-inflation=0).

  The likelihood ratio test should be asymptotically efficient, but I
think there may be specific tests of zero-inflation (and perhaps
overdispersion) that are more powerful for finite-size samples ...

  These questions seem to have strayed a bit from ADMB, and might be
more appropriate on an R list (r-help, r-sig-ecology if the problems
are ecological, stack overflow, r-sig-mixed-models if they're mixed ...)