st: How does Stata’s ml handle missing log-likelihood values?

[<kirin_guess@yahoo.com.tw>]

Dear Statlist,

I’m programming a maximum-penalised-likelihood model by mata-based type d0 
evaluator together with “ml max” in Stata.
The penalty term is the log-determinant of the Hessian matrix of the 
unpenalised log-likelihood.
I’d like to know how the ml handles missing log-likelihood values.

I know that between any two iterations displayed in the screen, Stata 
actually computes a lot of log-likelihood values (lnf).
I guess Stata does this for computing the first and second-order derivatives 
of lnf, right?
My question is that, what if some of those log-likelihood values are 
missing?
I encountered this problem because Stata sometimes returned a negative 
determinant during computing numerical derivatives,
and made the penalty term as well as the log-penalty-likelihood become 
missing.
Although the model always converged and returned estimates in the end,
I’m not sure whether those estimates were valid.
So I’d like to know how the ml handles missing log-likelihood values when 
Stata computes numerical derivatives,
and what I should do to deal with this problem (if it is a problem.)

In addition, I found that the estimates returned by “ml max, search(on)” 
sometimes were
quite different from the estimates returned by “ml max, search(off)”.
(Both of commands can achieve convergence criteria.)
Do the missing log-likelihood values during the derivative computation cause 
these unstable estimates?

Thank for your helps.


Kirin 

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