Nishant Dass <[email protected]> inquired about the speed of -mprobit-
compared to -mlogit-.
> When I try to do a multinomial analysis, I have to rely on
> -mlogit- because -mprobit- takes about 10 minutes for each
> "iteration" (and doesn't converge even after running for 8
> hours). Then, when I try to instrument one of the
> independent variables, I have no choice but to use
> -ivprobit-.
-mprobit- is slow when compared to -mlogit- since it uses quadrature to
approximate the multivariate normal distribution. Moreover, it only computes
the first order derivatives of the log likelihood analytically (in -ml- terms
it has a d1 likelihood evaluator) so the second derivatives are approximated
numerically. All compuations are done in ado an Mata. In comparison, -mlogit-
is implemented in C and computes the second derivatives analytically.
If you run the document example, the sysdns3 data, -mprobit- is about 900 times
slower than -mlogit-. My guess is you have a large data set with many
alternatives so Stata's version of -mprobit- may not be a feasible routine to
use, or you need a faster computer. Note, that -mprobit- is using Newton-
Raphson (-ml, technique(nr)-) and you can get a faster feedback in the iteration
history if you use technique(bfgs), but the overall computation time will
probably be the same.
With that said, if you still feel there may be a problem with the -mprobit-
algorithm feel free to contact me at personally and, if possible, send the
dataset or give me the specifics of your problem and I will simulate one.
-Rich
[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/
