[18649] in s-news-athena
[S] Summary#2: Generalized linear mixed models
daemon@ATHENA.MIT.EDU (Melanie Wigg)
Fri Jul 16 16:16:43 1999
From: Melanie Wigg <mbwigg@navigator.math.uwaterloo.ca>
Message-Id: <199907162012.QAA20643@navigator.math.uwaterloo.ca>
To: s-news@wubios.wustl.edu
Date: Fri, 16 Jul 1999 16:12:20 -0400 (EDT)
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Hello,
After posting my summary regarding fitting generalized linear mixed
models, I received a few more responses. Jonathan Wand,
Max Kuhn, and Douglas Bates all suggest using PROC NLMIXED in
SAS v7. Their original emails are included below.
Thanks to everyone who responded!
Melanie
--------------------------------------------------------------------------
Jonathan Wand:
--------------
I don't see anyone reference to PROC NLMIXED in the responses. It is
experimental in SAS v7. Works using quadrature and I have had success with
some simple models.
I am also working on some code to implement McCulloch (jasa 1997) MCEM
algorithm in S+.
--------------------------------------------------------------------------
Max Kuhn:
---------
If you do pursue the SAS approach to GLMM's and you have SAS V7, you may want to
check out the experimental NLMIXED procedure. The difference between the two is
that the GLIMMIX macro computes variances on the scale of the linear predictor,
but NLMIXED computes them on the "natural" scale of the data. The latter uses a
nonlinear model with different error distributions (such as normal, gamma,
binomial..) and the former computes a "linear-like" mixed model on the linear
predictor.
Anyway, this is better outlined in the documentation on NLMIXED on the SAS web
site. See http://www.sas.com/service/techsup/v8sugipapers.html for model
details.
If you do have version 7, there are updated dll's for the procedure. I can send
them to you if you wish.
--------------------------------------------------------------------------
Douglas Bates
-------------
As one of the developers of the NLME library, I thought I would
contribute my bit to this discussion.
Your respondents are quite correct - the NLME 3.0 library does not
have facilities for generalized linear mixed models (although I would
take exception to Ming Ji's statement "BTW, GLMM is far more
general and difficult than NLME." - that is simply not true).
During a recent visit, Bill Venables took the opportunity to educate
us regarding generalized linear models. He and Jose' Pinheiro and I
have come to the conclusion that the hooks to implement the Penalized
Quasi-Likelihood (PQL) method for GLMM's are there in the NLME
library. It would not be trivial but it would be possible to combine
pieces of the glm and lme functions to do PQL estimation. Neither
Jose' nor I will likely have time to do that in the near future. Even
if we did have time, it might not be the best idea as the PQL method
gets some bad press in the GLMM literature due to biases.
I think the preferred method for estimating GLMM's is that in the new
PROC NLMIXED of SAS version 7. That PROC used an adaptive Gaussian
quadrature for the integration with respect to the random-effects that
is part of the definition of the likelihood. Interestingly enough,
the methods in this SAS PROC are based on a paper by Pinheiro and
Bates about approximations to the log-likelihood for the nonlinear
mixed-effects model although, at the time we wrote that, we were not
thinking of generalized linear mixed models.
There are some problems with implementing a method like the adaptive
Gaussian quadrature in S. Within each iteration of an optimization
there are parts of the calculation that are not easily vectorized. It
is likely that a language like omega (http://www.omegahat.org/) would
be more effective there.
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
