Statalist The Stata Listserver


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: st: xtlogit problem


From   jlinhart@stata.com (Jean Marie Linhart, StataCorp LP)
To   statalist@hsphsun2.harvard.edu, mguerin@ovc.uoguelph.ca
Subject   Re: st: xtlogit problem
Date   Wed, 17 May 2006 10:27:50 -0500

Last week, Michele Guerin <mguerin at ovc dot uoguelph dot ca> asked
about a problem with the convergence of an -xtlogit- model with Stata
SE 9.2.  The model converged with 9.1, then did not converge when we
made the switch to 9.2.

Michele's model is one in which the random effect (sigma) parameter is
effectively zero. This means that lnsig2u heads to negative infinity.

A little bit of investigation quickly showed the change causing the
problem was the switch from saving the log-likelihood out as a float
(in 9.1) to a double in 9.2.  This switch, with more accuracy in the
log likelihood, allowed -ml- to take a bigger step, pushing lnsig2u
closer to negative infinity.  Unfortunately, this also pushed the
negative hessian over the line from just barely invertible to not
invertible, resulting in "not concave" messages and non-convergence of
the model.

The good news is that with a sigma so close to zero, there is no
reason to be using a random effects model at all.  Michele can
estimate using -logit- and obtain the same results.

I will include below the results from Stata 9.1 (where the -xtlogit-
converges), the equivalent -logit- model showing the same parameter
results, and the results from Stata 9.2 where I force the -xtlogit- to
quit after 3 iterations, again with the same parameter results except
for lnsig2u which is heading to negative infinity.  I have not
included the iteration logs for these.

Hope this helps,

--Jean Marie
jlinhart@stata.com

******************************************************************



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index