[R-sig-ME] The simple burning questions about lmer and their answers

Page E. Van Meter vanmete7 at msu.edu
Thu Oct 30 00:11:06 CET 2008

Hello all,
I have been teaching myself how to use generalized linear mixed effect 
models in R. While I find that the message boards like this one are 
extremely helpful, they have a "time travel" issue. I am finding answers 
to my burning questions that are years old and I am having trouble 
determining if the issues raised in 2006 are still issues today. So I 
wanted to put all of the hot topics I have been reading about, and the 
overly simplified answers I have gleaned in one thread. I hope others 
may add their burning questions and the answers they use to this thread 
as well?? And always the hope for comments:

Q1: Should I use ML or REML to compare the fixed effects in a set of 
nested models using likelihood-ratio testing (LRT)?
A1: It is much better to use ML to compare fixed effects in your model.

Q2: Should I use ML or REML to compare the random effects in a set of 
nested models using likelihood-ratio testing (LRT)?
A2: To test the worth of random effects in your model, REML is better. 
HOWEVER, if using lmer with a family specification to call glmer (family 
= binomial or poisson) you are automatically using ML. REML is not an 
option for generalized linear mixed effect models using lmer.

Q3: Ok, so how do I asses the value of random effects in my generalized 
linear mixed effect model?
A3: <insert cricket sounds here>
I gather that your experimental design should be the major indicator 
that a random term is needed in your model specification. But sometimes 
showing that the random effect is valuable (perhaps in a case where none 
of your fixed effects are worth anything) is, in itself, an interesting 
result.  Help?

Q4: So what is the hoopla about p-values for fixed effects in lmer?
A4: Short answer is that p-values are now included in the current 
incarnation of lme4 (can't find documentation of this but I see them so 
they must be there right?). The long answer about why you should be 
hesitant to use them and more hesitant to NEED them was provided by Doug 
here :

Page E. Van Meter
Michigan State University
Department of Zoology
vanmete7 at msu.edu

More information about the R-sig-mixed-models mailing list