[R] Mixed effects model with binomial errors - problem

Ben Bolker bolker at ufl.edu
Wed Sep 10 21:00:23 CEST 2008

RFTW <l.temarvelde <at> nioo.knaw.nl> writes:

> ok, the model does run now! 
> but, dont i need the method="ML" when i want to compare this model with a
> reduced model  using anova(model1, model2)? The R-Book tells me that REML is
> not good for that (p.635)

  maximum likelihood ("ML") is the default [and only!] estimation
method for generalized linear mixed models (GLMMs, or in other
words lmer with a "family" argument) in lmer.  It is correct that
you shouldn't compare REML fits with different fixed effects, but
unfortunately it's almost as bad (not theoretically wrong, but
very unreliable for small to moderate sample sizes) to use likelihood
ratio tests (as in anova(model1,model2)) for these tests.

   See http://www.zoo.ufl.edu/bolker/glmm_rev-26aug.pdf for more
information ...

  Ben Bolker

More information about the R-help mailing list