Re: st: changes in -xi: xtmixed- command from Stata 10 to 11

[rgutierrez@stata.com (Roberto G. Gutierrez, StataCorp)]

Morten Vejs Willert <amklinik@gmail.com> writes:

> I have been running an -xi: xtmixed- command like the following in Stata 10
> for a large part of the analyses for my PhD:

> xi: xtmixed [outcomevariableX] i.time*i.random || id: i.time, mle cov(un)

> In Stata 10 I would usually get my results after 3 or 4 iterations.  Now,
> after upgrading to Stata 11 the program just keeps doing more and more
> iterations, without resolving the issue and giving the results of the
> analysis. Does anybody the reason for this? I like some of the new features
> of Stata 11, but will be forced to downgrade to Stata 10 if this issue can
> not be resolved

The most likely cause of such discrepancies in behavior from Stata 10 to Stata
11 are the refinements we made to the -ml- optimizer for Stata 11.  These
improvements to -ml- have, almost universally, made it more convergent across
all models and data, but occasionally you can find a model/data combination
where the opposite occurs.  Such phenomena tend to be more prevalent with
-xtmixed- and the other mixed-model commands because of the difficulty that
can arise from estimating at the boundary of the parameter space.  That is,
for these commands it is easier to find a model/data combination that is just
on the knife-edge of convergence, thus exhibiting a change in behavior from
Stata 10 to Stata 11.  Most of the time, models that did not converge in Stata
10 now converge in Stata 11, but sometimes the opposite can occur.

If Morten wishes, he can email me his data and model and I'll see if we I find
a way for Stata 11's -xtmixed- and -ml- to work better with it.  That way, he
can then take advantage of -xtmixed-'s new features.

--Bobby
rgutierrez@stata.com
*
*   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/