[R-sig-ME] lme4a, glmer and all that

[Jeffrey Evans]

Thanks Doug,

I guess I'll forge ahead with lme4 for now.

I have another question for you about overdispersion, but I'll put that in a
separate thread.

Thanks,
Jeff

-----Original Message-----
From: dmbates at gmail.com [mailto:dmbates at gmail.com] On Behalf Of Douglas
Bates
Sent: Thursday, June 24, 2010 11:30 AM
To: Mitchell Maltenfort
Cc: Jeffrey Evans; r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] lme4a, glmer and all that

On Thu, Jun 24, 2010 at 8:38 AM, Mitchell Maltenfort <mmalten at gmail.com>
wrote:
> Haven't seen anything new since his March 4 post indicating that a 
> check vs Stata agreed with lme4, not lme4a.

Sorry, I should have followed up more publicly on that posting.  There was a
difference in the results between lme4 and lme4a and I was concerned that
the problem was in lme4.  It turns out that the problem was in lme4a and has
now been resolved.

Having said that, I have now encountered examples where lme4a converges to a
different and better optimum than does lme4.

All of the maximum likelihood estimation methods for mixed models end up
doing some kind of numerical optimization procedure.  One of the changes in
lme4a is the use of the bobyqa optimizer from the minqa package, as opposed
to the nlminb optimizer in the stats package.  I feel that the bobyqa
optimizer is more effective and often faster than nlminb (although not
always).

One of the big differences between linear mixed models and generalized
linear mixed models is the number of parameters in the general optimizer
problem.  In linear mixed models one can "profile out" the fixed-effects
parameters and produce a much easier optimization problem.  For generalized
linear mixed models profiling out the fixed-effects produces only an
approximate minimum.  In the example from Dave Atkins fitting a Poisson GLMM
that has been discussed on this list recently the differences were minimal
and solving the reduced problem was much faster (17 seconds versus 300
seconds) than the full optimization problem.  However, in the examples from
the 2007 JSS paper by Doran, Bates, Bliese and Dowling
(http://http://www.jstatsoft.org/v20/i02) I have seen a substantially better
minimum deviance using the full optimization than using the reduced
optimization.

Bottom line is that the results from lme4 should be ok but lme4a, when I get
it all sorted out, can do better.

> On Wed, Jun 23, 2010 at 7:17 PM, Jeffrey Evans 
> <Jeffrey.Evans at dartmouth.edu> wrote:
>> Can anyone provide a status update on Doug Bates' comment from March 
>> about doubting parameter estimates from glmer in lme4?
>>
>> A) Which version is suspect - version 32, it seems?
>> B) Des version 33 resolve this issue?
>>
>> Many thanks,
>>
>> Jeff Evans
>> Dartmouth College
>>
>> In March Doug Bates wrote:
>> "Two further comments. It is only the results from fitting 
>> generalized linear mixed models with the current lme4 that I have 
>> cause to doubt. The results from linear mixed models do check out. "
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list 
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>