st: xtmelogit--any quick way to identify variance components near zero?

["Clyde Schechter" <clyde.schechter@einstein.yu.edu>]

I'm running a series of analyses using xtmelogit.  The data set has a
little over 5,000 observations, and there are three levels of random
effects (including a random slope at the one level).  Even using the
laplace option, each of these analyses takes about 10 hours to run.  Fair
enough -- it is what it is.

But what's particularly frustrating is that for some of the analyses, one
or more of the variance components apparently is zero or very close to
zero.  So the analysis gets hung up as the estimate of the log-sigma tries
to march down to negative infinity and goes on "forever."  (I cut my
losses at 25 iterations.)

I tried picking a random sample of about 500 observations from the data
set to test the model--that runs in a tolerable amount of time, but it
didn't pick up the problem: in the subset the variance component was
pretty small, but not small enough to prevent convergence in a reasonable
number of iterations. (FWIW, from a theoretical perspective it is very
surprising that these variance components are turning out so small--the
best I can think of to explain it is that the fixed effects in the model
are accounting for much more of the anticipated higher-level variation
than I expected them to--when I omit some of those fixed effects the model
is much better behaved.)


So, does anybody know of a reasonably quick way to diagnose very small
variable coefficients in xtmelogit in advance of trying to run it?  Thanks
in advance for any help.


Clyde Schechter, MA MD
Associate Professor of Family & Social Medicine
Albert Einstein College of Medicine, Bronx, NY, USA



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/