> ------------------------------
>
> Date: Tue, 12 Dec 2006 06:52:14 -0800 (PST)
> From: Kam Kup <[email protected]>
> Subject: st: Convergence when using MVNP: Prof Jenkins?
>
> This is addressed to Prof Jenkins and anyone else who
> has included large-dimensional calculations mvnp
> calculations in a likelihood evaluator:
>
> I've been using mvnp in a maximum likelihood
> estimation and am finding that convergence is
> extremely slow. I am doing a selection model with
> multiple cases each involving an mvnp calculation of
> different degrees. The three cases each have two-,
> three-, and four-variate calls to mvnp, each with an
> independent Cholesky matrix.
>
> Because some of the cases involve several more error
> terms that are independent of those in the mvnp
> calculation, I also have to add several ln(normal(-x))
> terms to the calculation.
>
> Convergence is extremely slow and the elements of the
> Cholesky matrix take on some outlandishly large values
> in the process.
>
> I am using ml rather than d0 since the log likelihood
> is simply a sum of the individual likelihoods. I'm
> also using technique(nr dfp) as you did for a
> different problem in your 2006 paper.
>
> Prof Jenkins, has this happened in your experience, or
> do you think there is something wrong with the
> likelihood evaluator?
>
I have no idea whether your program for the likelihood evaluator is OK
or not (and in case you were wondering, sorry, but I don't have time
to look at it).
In my experience, run time is higher,
* the more equations there are
* the more draws
* the larger the sample size
* the poorer the starting values
* the less well identified the model is
The last two issues make reliable convergence more difficult too.
Ensure that you are using the plug-in (default) not ado code for
-_gmvnp-
Consider using Halton rather than pseudo-random draws, and explore use
of antithetics
I suggest feeding in starting values from univariate probits (i.e.
ignoring selections)
good luck
Stephen
-------------------------------------------------------------
Professor Stephen P. Jenkins <[email protected]>
Institute for Social and Economic Research
University of Essex, Colchester CO4 3SQ, U.K.
Tel: +44 1206 873374. Fax: +44 1206 873151.
http://www.iser.essex.ac.uk
Survival Analysis using Stata:
http://www.iser.essex.ac.uk/teaching/degree/stephenj/ec968/
Downloadable papers and software: http://ideas.repec.org/e/pje7.html
*
* 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/
