[R] Evaluating the significance of the random effects in GLMM

[Ben Bolker]

Gabriela Agostini <gabrielaagostini18 <at> gmail.com> writes:

> 

[snip]

> I am working with GLMM using the binomial family
> I use the following codes
> 
> I dropped no significant terms, refitting the model and comparing the
> changes with likelihood:
> 
> G.1<-lmer(data$Ymat~stu+spi+stu*sp1+(1|ber),data=data,family="binomial")
> G.1b<-lmer(data$Ymat~stu+spi+(1|ber),data=data,family="binomial")
> 
> anova (G.1,G.2)
> 
> But, when I want to evaluate the significance of random effect (1|ber)
> I cannot use a likelihood-ratio test, probably because the link
> function of both models is different.
> 

  A couple of minor comments: 
* you should probably use Ymat rather than data$Ymat
as your response, it will make post-processing easier
* in your first model do you really mean stu*sp1 rather than stu*spi?
* since A*B is equivalent to A+B+A:B, your first model specification is
equivalent (assuming you really meant stu*spi) to stu*spi OR stu+spi+stu:spi.
This won't change your answers but will be clearer to experienced R users.

  I don't understand why anova() won't work in this case.  At least
for the example you've shown us, it should.  The link functions aren't
different.  

  Please (1) follow up to r-sig-mixed-models at r-project.org and (2)
try to provide a little more information: a reproducible example if
possible (http://tinyurl.com/reproducible-000).

  PS the section in http://glmm.wikidot.com/faq may provide some
useful background on testing random effects.