David Airey wrote:
Anybody know if the Stata 8 command xtgee handles multiple categorical
dependent variables? I tend to use commands which allow me not to
think
about "link functions". Is the logit link a generalized logit link
(for
nominal multinomial regression)? I'm thinking of buying the GEE book
at
the bookstore, but I can't tell from the TOC whether this is covered.
I've managed to get gllamm to run an appropriate model thanks to a
previous post. I'd like to compare results to a multinomial GEE.
------------------------------------------------------------------------
--------
According to Page 103 of J. W. Hardin & J. M. Hilbe, _Generalized
Estimating
Equations_, (Boca Raton: Chapman & Hall/CRC, 2003), "The SUDAAN
package is the
only one of the four packages used in the text that has support for
fitting
this model." The four packages are SAS, S-Plus, Stata and SUDAAN.
Ah. Thanks for reading that sentence to me! I had previously read a
paper on GEE:
Horton & Lipstiz (1999) Review of software to fit generalized
estimating equations regression models. The American Statistician, May
1999, Vol. 53,
which suggested this, but the versions covered were old.
Have you considered -mlogit, cluster()- as an alternative? I suspect
that it
would be like SUDAAN's GEE with an independent working correlation
structure,
and as such ought to give similar results to what you would would get
with
SUDAAN, with some reduction in efficiency, the degree depending upon
just how
far from reality the independent working correlation is in the
particular case.
I had not, but mainly because I was not clear on what it does in
comparison to the alternatives. I will get clear, since this option is
scattered throughout the Stata estimators.
The authors (James Hardin and Joe Hilbe) illustrate fitting of the
model with
SUDAAN using an artificial dataset. It would be interesting to see
just how
close results from -mlogit, cluster()- are to what they got from
SUDAAN (with
an exchangeable working correlation structure), but that particular
dataset
isn't apparently included in the downloadable zip file of datasets
used in the
book.
If I can help here I'll post any comparative results down the road.
I strongly recommend the book, by the way. It has not only
background, but
also practical pointers and methods (for example, diagnostics and
residuals,
systematic selection of a working correlation structure, how to do a
roll-your-
own [fixed] working correlation structure in Stata) to make it well
worthwhile
to anyone contemplating a population-averaged GEE approach.
Sold! I have a weakness for buying books and borrowing them from the
library.
-Dave
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/
