I read with considerable interest Joseph Coveney's comments below re (P)QL
approaches and the other contributions following on from Tom Steichen's
e-mail today re Stata 9.
For those of us in the bio-medical sphere, categorical outcomes (eg, alive
/ dead) are _very_ important and , in the general outcomes literature, they
are currently associated with large data sets (250000 observations and
upwards are not at all unusual, whether these be from administrative or
purpose built data-bases). Although we can use "xtlogit", unlike "gllamm"
there is no way (?) to access empirical Bayes' estimates whilst using
"xtlogit" as with "gllapred" (and "xtreg").
This is somewhat frustrating, as, although it may be useful to demonstrate
with "xtlogit" that there is previously undiagnosed "heterogeneity", the
literature using multilevel models for (binary) outcomes proceeds to use
these EBE to illustrate important aspects of the model (the classic paper,
some time ago now, is H. Goldstein and D. J. Spiegelhalter. League tables
and their limitations: Statistical issues in the comparisons of
institutional performance. Journal of the Royal Statistical Society A 159
(Part3):385-443, 1996.)
In Stata, this has also been demonstrated very nicely by Sophia Rabe-Hesketh
in the 3rd Edition of "A handbook...." and the current "GLVM" book.
Hence I understand the comments by SamL and, perhaps, the sense of
disappointment that a module for "non-linear mixed effects models" was not
first cab off the rank in Stata 9, although I do not presume to know (i) the
reasons for this in official Stata or (ii) what Stata 9 may hold for
"gllamm" and the much awaited "Multilevel and Structural Equation Modeling
for Continuous, Categorical, and Event Data" (Sophia Rabe-Hesketh, Andrew
Pickles, and Anders Skrondal, Stata Press; Expected publication date: June
2005).
BTW, I will certainly be purchasing Version 9.
John Moran
E-mail: [email protected]
-----Original Message-----
From: Joseph Coveney [mailto:[email protected]]
Sent: Wednesday, 6 April 2005 12:24 PM
To: Statalist
Subject: re: Re: st: RE: Stata 9 announcement
SamL wrote (excerpted):
I am also hoping that David is wrong about stata leaving it to gllamm to
work for non-linear mixed models, although this is clearly the short-term
situation. Gllamm is a godsend, but it is also very difficult to be sure
one is actually estimating what one wants to estimate--I've noticed
several queries on statalist that have that flavor. This may be because
gllamm is not integrated into stata and thus may not follow stata
conventions. I am not sure. Maybe gllamm does follow the conventions
closely, but that it is so flexible that it is difficult to document.
Whatever the reason, it seems very slow and very opaque.
At any rate, even if gllamm runs faster, nothing will substitute for
having a module for non-linear mixed models that is actually written and
supported by stata. It is my sincere hope that in a few releases (10?
11?) such will be the case. Multi-level modelling of discrete outcomes is
pretty mainstream now, for good or ill. It is my hope it will become
mainstreamed into stata 10 or stata 11 as well.
[edited]
----------------------------------------------------------------------------
----
There are quasilikelihood approaches to fitting multilevel generalized
linear
models that are faster than -gllamm- and that might be more suitable than
-gllamm- for certain circumstances, but these have their own foibles. In
addition, -gllamm- is *much* more flexible and capable than your average
penalized quasilikelihood implementation. Even so, quasilikelihood methods
should become more accessible with Mata.
But -gllamm- is not the general-purpose nonlinear mixed effects
model-fitting
routing that SAS's PROC NLMIXED is, at least not yet. (You can get marginal
nonlinear models now using -ml- with -cluster()-; this kind of approach will
become more user-friendly with Stata 9's new -nl, cluster()-.) SamL's
comment
that the list gets queries by users unsure of what they're getting with
-gllamm-
might be taken to heart by StataCorp: it's probably better to get
-xtmixed-
into the hands of Stata users and see how much support is needed with the
simpler case before venturing into nonlinear mixed effects models. And
after
users are accustomed to the steps to be taken when encountering models that
settle toward negative variance components (which appear to be constrained
to
zero by the parameterization that -xtmixed- uses) in order to fit the data,
then Stata Corp can proceed to devote resources to embellish -xtmixed- with,
say, a bigger smorgasbord of covariance structures (if there's enough user
interest--what's already there will likely handle anything I'd ever
confront)
or denominator degrees of freedom estimations.
Joseph Coveney
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
