I do agree that the gold standard is simulations as suggested above.
There are examples in the "worked example" section of glmm.wikidot.com .
These simulations can be very slow -- hopefully (??) Doug Bates will be
able to work out an mcmcsamp() scheme for GLMMs soon, and we can see
if that (much faster) approach generally agrees with null simulations ...
If you look at Doug's schedule for July (www.stat.wisc.edu/~bates) you
will understand why he has been missing in action recently. In my
"spare time" - when not working on a major report finished last week
and a grant proposal due this week and in preparing these various
courses - I am working on the draft of a book about lme4 because if I
don't get some part of that manuscript to John Kimmel soon he is going
to be very rude to me in Rennes.