Yulia Marchenko wrote:
Jeph Herrin <[email protected]> asks:
> I'm trying to understand why, contrary to the documentation, these two do
> not
> give the same results:
>
> . glm rate exposed, family(binomial) link(logit) robust
> . xtgee rate exposed, family(binomial) link(logit) robust /*
> */corr(independent) i(id)
>
> The -glm- converges nicely, the -xtgee- gives an error:
>
> estimates diverging (missing predictions)
> r(430);
>
> One feature is that the dependent variable is a proportion, not 0/1;
For the parameter estimates from -glm- to be equivalent to those
from -xtgee-,
the -irls- option needs to be specified with -glm-. This is because, by
default, -glm- uses maximum likelihood optimization to obtain parameter
estimates.
I would like to see Jeph's data to determine why -xtgee- did not converge.
I
have emailed him privately and will report any conclusions to the list.
--------------------------------------------------------------------------------
I didn't realize that -xtgee- can accept a proportion for the response
variable with the binomial family as can -glm-.
In the dummy example below, correspondence between -xtgee- and -glm- seems
to have more to do with using -cluster()- with -glm- than with using the
IRLS
algorithm.
Joseph Coveney
clear
set more off
set seed `=date("2006-08-22", "ymd")'
set obs 200
generate byte treatment = uniform() > 0.5
generate byte repetition = ceil(uniform() * 6)
generate int pid = _n
expand repetition
generate byte positive = uniform() > 0.5
xi i.repetition
xtgee positive treatment _I*, i(pid) family(binomial) ///
link(logit) corr(independent) robust nolog
glm positive treatment _I*, family(binomial) link(logit) ///
robust irls nolog
glm positive treatment _I*, cluster(pid) family(binomial) ///
link(logit) nolog
exit
*
* 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/
