[R] lmer, p-values and all that

[Ben Bolker]

Michael Grant <michael.grant <at> colorado.edu> writes:

> 
> 
> Dear Help:
 
> I am trying to follow Professor Bates' recommendation, quoted by
> Professor Crawley in The R Book, p629, to determine whether I should
> model data using the 'plain old' lm function or the mixed model
> function lmer by using the syntax anova(lmModel,lmerModel).
> Apparently I've not understood the recommendation or the proper
> likelihood ratio test in question (or both) for I get this error
> message: Error: $ operator not defined for this S4 class.

  I don't have the R Book handy (some more context would be extremely
useful!  I would think it would count as "fair use" to quote the
passage you're referring to ...)
 
> Would someone be kind enough to point out my blunder?

  You should probably repost this to the r-sig-mixed-models at r-project.org
mailing list.

  My short answer would be: (1) I don't think you can actually
use anova() to compare likelihoods between lm() and lme()/lmer()
fits in the way that you want: *maybe* for lme() [don't recall],
but almost certainly not for lmer().  See http://glmm.wikidot.com/faq
for methods for testing significance/inclusion of random factors
(short answer: you should *generally* try to make the decision
whether to include random factors or not on _a priori_ grounds,
not on the basis of statistical tests ...)

  Ben Bolker