[R-sig-ME] Random vs. fixed effects

[dave fournier]

It appears that the real problem here is not with the data, but with the
estimation procedure.  When fitting the model by maximum likelihood both
lme and AD Model builder give an estimate of about 3.8 (true value 4)
for the std dev of the RE's.  When fitting the data by reml
AD Model builder gets an estimate of about 4.4.  I think that the
reml estimation is more numerically challenging than ml and that lmer's
algorithm is just not up to the job. You can see the ADMB code etc.
at

         http://groups.google.com/group/admb-users/t/83d371e04bbe48c9