st: cryptic convergence error message in xtgee/xtlogit

[Kenneth Flamm <kflamm@mail.utexas.edu>]

I am attempting to run a GEE logit model on a binary data set consisting of 
about 26,000 clusters observed at 3 points in time.
It runs fine with an independent working correlation structure. When I make 
any assumption about the off-diagonal terms not being zero, however 
(exchangeable,unstructured,AR(1), whatever) I get an error message that 
looks like

Iteration 1: tolerance = 33.763565
estimates diverging (missing predictions)
r(430);

I have tried everything I can think of: I have 50 dummies, in addition to 
various continuous variables, and have dropped all cases where one of the 
50 dummies is associated with binary outcomes in the single digits, on the 
theory that small cell counts may be the problem. No go, still get the same 
error message.

I have tried tricks to make sure that I am not on the boundary of my 
parameter space: replicating my sample 1000 times, then adding an 
artificial sample once where I have reversed the 0/1 outcomes to ensure 
sure I am in the interior. Same problem.

I have even made sure that it is not a numerical problem in Stata 9.0. Ran 
the same problem in SAS 9 Proc Genmod and get an even more cryptic error 
message (estimation error)!

Has anyone seen this error message, know what it means, and have any ideas 
about corrective action that might work? Any good ideas would be most 
appreciated.

                                                  Regards,

                                                        Ken Flamm

*
*   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/