Thanks Jeff, I was not aware that BFGS and DFP did not need the
equation-level scores. I still think it would be a nice addition to
-ml- if it was possible to pass it the coefficient/variable-level
scores instead of the equation-level scores. This would make it
possible to use e.g. the robust and cluster options with likelihoods
that do not meet the linear-form restrictions.
Best wishes
Arne
On 05/08/06, Jeff Pitblado, StataCorp LP <jpitblado@stata.com> wrote:
Arne Risa Hole <arnehole@gmail.com> has a question/suggestion
regarding the quasi Newton techniques in the -ml- command:
> I was wondering if anyone knows whether StataCorp has considered
> extending the functionality of the -ml- command to allow the
> alternative optimisation algorithms (BHHH etc) to be used to maximise
> likelihoods that do not meet the linear-form restrictions.
> My understanding is that at the moment the only way these algorithms
> can be used is to pass the equation scores to -ml- as g1, g2 etc. and
> that this only works if each observation contains a score.
I believe Arne is referring to BHHH, BFGS, and DFP. Of which, only BHHH
requires equation-level scores from the 'd1' and 'd2' likelihood evaluators
(BHHH works automatically with 'lf' evaluators, and not at all for 'd0'
evaluators).
BFGS and DFP do not require equation-level scores; they work directly with the
gradient vector to build up their own version of a Hessian-type matrix.
> One useful extension would be to allow the possibility of passing the
> values of the likelihood differentiated with respect to the
> coefficients (instead of the equations/ index functions) directly to
> -ml-. Then the observations containing the scores for the groups of
> observations could be identified using "if==`last'", or something
> similar.
> I have used this approach along with matrix accum to create the outer
> product of gradient matrix. Setting `negH' equal to this matrix
> produces a "homemade" BHHH procedure which seems to work fine, but it
> would be nice to also be able to use the other algorithms.
Arne has devised a way around -ml-'s above restriction on BHHH by making the
likelihood evaluator produce the OPG.
--Jeff
jpitblado@stata.com
*
* 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/
*
* 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/