[R-sig-ME] Likelihood drops on adding random effect

[dave fournier]

This discussion reminds me of some reanalysis
I did a while back.

     https://stat.ethz.ch/pipermail/r-sig-mixed-models/2011q4/006953.html


The results were that the Laplace approximation may not be accurate enough
for this type of model and that AGHQ seems to be broken on lme?.

Perhaps that is the source of your problem. I would advise reanalyzing
with AD Model builder's random effects package. It can deal with crossed 
effects
by extending the Laplace approximation with various types of importance
sampling.