[R-sig-ME] model comparison in glmmADMB

[Ben Bolker]

RH Gibson, School Biological Sciences <Rachel.Gibson at ...> writes:

> 
> I am trying to compare two models to see whether the random effect is 
> significant, using the anova command I get the following output:
> 
> > anova(m7,m8)
> Analysis of Variance Table
> 
 [snip]

>   NoPar  LogLik Df -2logQ P.value
> 1    12 -2427.4
> 2    11 -2427.9 -1  -0.96
> Warning message:
> In pchisq(q, df, lower.tail, log.p) : NaNs produced
> 
> Why might this be happening? Can I still use the log likelihood ratio as a 
> comparison method? And more generally is this an appropriate test of the 
> random effect?

  anova.glmmadmb is not smart enough to put the terms in order of
increasing complexity, and it expects you to do so.  This will be
clarified/fixed in the next version.

  More generally, LRTs of random effects are OK but not great; they are
known to be conservative (see http://glmm.wikidot.com/faq both for ways
to test random effects and for questions as to why you would be wanting
to test the significance of a random effect in the first place).