I made a change to ghk2(), which led to a minor change in -cmp-, and
which can cause (hopefully small) changes in -cmp- results.
-cmp- breaks observations that require the GHK algorithm into
identically-censored groups. So in a trivariate probit problem, all
observations with 0 outcomes for all 3 equations are one group, those
with 0,0,1 are in another, and so on. Before the latest version, when
computing the likelihoods for each group of observations, -cmp- started
at the beginning of the Halton/generalized
Halton/Hammersley/pseudorandom sequence of pre-computed draws. Now it
uses a different part of the sequence for each block, which prevents
spurious simulated correlations in errors across observations. In
general, this should improve convergence behavior.
To allow this change, ghk2() now requires a new argument that specifies
where to start in the sequence. Missing (".") is interpreted as 1. This
means that calls to ghk2() will need to be edited to work with the new
version.
Update with "ssc install cmp, replace" and "ssc install ghk2, replace".
--David
*
* 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/
