Xiao Chen <[email protected]> asks about the differences in estimation results
after -xtlogit, re- in Stata 9 and Stata 10 and compares these results to the
new -xtmelogit- command.
> I noticed some changes to xtlogit for the random effect models from stata 9 to
> stata 10. Now the results using xtlogit in stata 10 are slightly different
> from using xtlogit in stata 9. I wonder if someone could explain what the
> implication of this difference is.
As of Stata 10, all -xt- random-effects estimation commands use a new default
numerical method of approximating the likelihood function. The new method is
described in [XT] xtlogit on p. 216 and corresponds to the
-intmethod(mvaghermite)- option of -xtlogit-. The differences Xiao observes
are most likely due to this.
> Now Stat 10 has xtmelogit for general random effect models, not just random
> intercept models. Comparing xtmelogit and xtlogit for random intercept model,
> it seems that there is some small numeric difference as well. Would it be
> possible to match the results from the two commands by adding some options to
> either xtlogit or xtmelogit?
By default, -xtlogit, re- approximates the likelihood by using the mean and
variance adaptive Gauss-Hermite quadrature method (-intmethod(mvaghermite)-)
with 12 integration points (-intpoints(12)-). -xtmelogit- uses the adaptive
Gauss-Hermite approximation based on the mode and its curvature with 7
integration points (see Methods and Formulas sections of [XT] xtmelogit and
[XT] xtlogit for details). In order to obtain more similar results from
-xtlogit, re- and -xtmelogit-, Xiao should increase the number of integration
points to 12 in -xtmelogit- and request -intmethod(aghermite)- approximation
method with -xtlogit-. For example,
. xtlogit ..., re intmethod(aghermite)
and
. xtmelogit ..., intpoints(12)
-- Yulia
[email protected]
*
* 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/
