st: Why does specifying "ml search" results in different estimates?

[<kirin_guess@yahoo.com.tw>]

Dear Statalists,

I’m compiling a model by using the command, “ml”, and the method is 
“d0”. (I don’t have analytical formulae of the score and hessian, so I 
can’t use “d1” or “d2”. The method “lf” is not suitable for my 
model.)

When maximising the log-likelihood, I found that specifying “ml search” 
resulted in different estimates, and the differences of some estimates were 
quite obvious. I’ve tried to specify more rigorous convergence criteria, 
but that didn’t solve the problem. Hence I’m seeking your advice. What 
would be the causes of this problem? How to solve/minimise it? Is it always 
better to use “ml search”, if I don’t care about the extra computation 
time?

Any suggestions are welcome, and thank you very much in advance.

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/