[R-sig-ME] Likelihood test for random effect in a GLMM

hadley wickham h.wickham at gmail.com
Fri Nov 28 18:02:22 CET 2008


On Fri, Nov 28, 2008 at 10:49 AM, Benjamin Dantzer <bendantzer at gmail.com> wrote:
> I am doing very similar analyses and get similar results (enormously low
> P-values) when I compare a GLMM to GLM using the LRT.
>
> However, from my earlier search for an answer to the question such as yours,
> I thought that glmms  (glmer) are fit using ML and glm ONLY via REML?  So
> that a LRT is not appropriate. But I cannot provide a complete answer here.
>
> See the following:
>
> https://stat.ethz.ch/pipermail/r-sig-mixed-models/2008q3/001180.html

In general, I think it's a *VERY* bad idea to compare log-likelihoods
produced by different modelling functions.  Unless you are completely
sure that the likelihoods are defined in the exactly same way (i.e.
that the same constants have been dropped, or not), you are unlikely
to get meaningful results.

I think that the model that the original poster provided could more
clearly be written as:

fleaburden ~ (sex + width + sess + Nhat + alt) ^ 2

(andcodeismucheasiertoreadifyouusespaces)


Hadley

-- 
http://had.co.nz/




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