In discussing this with Rich Gates at Stata, he cautiously suggested
that rather than modifying -ml-, one could move the work to the
evaluator program. I could write what I call a "pseudo-d2" evaluator
that computes the scores analytically and the Hessian numerically (based
on the varying analytical scores) while using the linear form
restriction to save time on the latter. That's what -cmp- now does. In
my testing, pseudo-d2 (with Newton-Raphson) comfortably beats both lf
and d1, by factors on the order of 2-4 relative to the best alternative.
Is this an argument for a new evaluator type in -ml-? With a new type, I
wouldn't have needed to do this extra work.
--David
David Roodman
Research Fellow
1776 Massachusetts Avenue NW
Washington, DC 20036
Center for Global Development
Independent Research and Practical Ideas for Global Prosperity
Phone: 1 202 416-0723
*
* 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/
