[R-sig-ME] gls for generalized linear model (ChenChun)

[Ben Bolker]

ChenChun <talischen at ...> writes:

> 

 [snip]

>  * REML does nothing in glmer()--> So REML 
> is not used in estimating the random effect for GLM? (Sorry I am new
> to mixed models)

  Correct, see http://glmm.wikidot.com/faq#reml-glmm

> * are you sure it makes sense to test the statistical significance of 
> the random effect?--> My experiment is
> to estimate the survival of animals, which the survival experiment 
> is conducted with a group of animals
> (ni) in the ith experiment, i.e. animals are clustered by expID. 
> However, for some experiments, the
> animals are coming from the same basket. That's why basketID is 
> also used as a random effect.

 [snip]

> The model output shows that the estimated variance between baskets is
>  almost zero, indicating that maybe
> there is no basket effect, or it can be neglectable. 
> What's why I would like to test whether including the
> random effect of basket significantly improves the model fitting. 
> If not and the variance is small, I can
> leave out the basketID term. Do you think this is reasonable?

[snip]

See http://glmm.wikidot.com/faq#random-sig ... Although opinions vary,
I generally feel that one should leave random effects in the model
as long as they are (1) suggested by the experimental design
and (2) not messing up the model fit.