[S] gee

Vincent Carey 525-2265 (stvjc@channing.harvard.edu)
Tue, 7 Apr 98 19:08:15 EDT

Apropos gee():

1. The revised version 4.13 (these are the sources used
for the windows port provided by Brian Ripley) has now been
submitted to statlib. This is the last revision of the
Cgee-based GEE solver that will be provided to statlib. In
the unlikely event of future changes, such will be posted to
s-news and to my web page as patches to that shar file.
Version 4.13 may not yet be on statlib, but can be retrieved
from the Web on the page cited just below. My plan to discontinue
revision of the Cgee-based solver is based on my desire
to support a much more flexible implementation to be described
below. Because the flexible implementation to be described below
is considerably slower than the Cgee-based solver, I do not
propose to replace the latter solver at this time.

2. Although I do not plan to actively revise the
Cgee-based solver in the future, I will, under reasonable
conditions, discuss its performance/problems "privately"
with users. This is in keeping with the list policy
discussed in Brian Ripley's recent posting. If important
discrepancies are discovered, users seeking my advice should
prepare minimal test datasets that evoke the discrepancy/problem
and e-mail them to me with explanatory comment. Users should
not expect contributors to rationalize upon demand discrepancies
between alternative implementations of procedures. As
information on software problems of public significance
emerges, the list will be notified so that users may take
precautions or make minor revisions as required.

3. Some discrepancies between GEE implementations in
SAS (PROC GENMOD and two IML versions), Stata, and
my contribution for S-plus have been described to s-news
and in several entries on the web page


[It is my understanding that there is a SAS patch
that corrects the problem arising with exchangeable
correlation parameter estimation when many
singleton "clusters" are present. An entry on the
aforementioned web site provides test data to determine
whether or not this patch has been applied to any
given SAS installation.]

Numerical discrepancies are not necessarily evidence of
implementation errors.

Users of gee in Oswald should note a problem in the
documentation of version 3.2, page 53 (November
1997), in which gee is invoked with an "offset="
option. That is an obsolete way of handling offsets
in GLM formulae, and the fit reported is effectively
a zero-offset fit. The correct result ensues
upon adding "+ offset(tex(log(obs.period)))" to
the regression model formula.

The Oswald developers are aware of this matter.
I mention it here because the item was recently brought
to my attention as a "SAS vs. S-plus discrepancy".
It is rather a documentation-coordination/
bulletproofing problem.

4. The "GEE construction kit" for Splus, described
briefly at the 1998 ENAR "GEE intro" session, allows
user-defined scale and working correlation estimators,
as well as user-defined link and variance functions.
These aspects of the GEE are coded by the user in S-plus.
(Various components are "pre-coded" with the kit to
emulate the functionality of the Cgee-based solver,
but these may always be overridden by the user without
any need for recompilation or reloading.)
No dynamic loading is necessary, although some
calculations will use a dynamically loaded C-language
patterned matrix library if such is present. (C
source for this library is provided with the kit.)
This GEE solver is slower than 4.13, but is appropriately
user-configurable and involves much less source code.
I plan to make a "test version" of this kit available
on the aforementioned web site later this week, and
I will notify s-news of its availability at that time.

(Thomas Lumley's user-configurable GEE solver is
available for Xlispstat as volume 1 #3 of J Stat Software
[http://www.stat.ucla.edu/journals/jss/]; numerical
agreement of these implementations has not been inspected.)

Finally, I would like to thank Professor Brian
Ripley for handling many gee-related queries on s-news,
and for making the port to Windows.

Vincent J. Carey, Ph.D.
Assistant Professor of Medicine (Biostatistics)
Harvard Medical School
181 Longwood Avenue voice: (617)-525-2265
Boston MA, 02115 USA fax: (617)-731-1541

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