gls for generalized linear model

ChenChun <talischen at ...> writes:

[snip]

2013/9/16 ChenChun <talischen at ...>

Dear R users,

I am fitting a GLMM model on survival:

fit1 <- lmer(alive ~ treatment + (1 | expID), 
 family = binomial, data = Data, REML = TRUE)

I would like to test whether the random effect is significant. 
Normally for a linear model, I could test it
against a model without random effect using gls, for instance
gld(response ~ variable, data=...,
method="REML"). However, it seems that gls does
not support the generalized 
linear model (family =
binomial). May I ask how I can test the random effect in this case?

As I recall, in the current CRAN version of lme4 (<1.0), the deviances
from glmer() are not commensurate with glm(), but in the new version (>=1.0)
they are, so you could compare the deviances of 

fit1 <- glmer(alive ~ treatment + (1|expID), family=binomial, data=Data)

and

fit0 <- glm(alive ~treatment, family=binomial)

(at the moment there is no automatic anova() method that can handle
this, although this could be added).

 One comment and one question:

* REML does nothing in glmer()
* are you sure it makes sense to test the statistical significance of
the random effect?