On Thu, Sep 17, 2009 at 3:22 PM, Partha Deb <[email protected]> wrote:
> Also, while EM might be able
> to deal with multimodalities better than ML (esp. Stata's -ml-), there is
> nothing inherent about EM that makes it "better".
If anything, EM is actually worse than the brute force -ml- when it
comes to multimodality and false convergence declaration. (I am
surprised that I have to disagree with Imbens :)) There are known
examples when EM converges to saddlepoints of the likelihood
(-difficult- option would probably be able to pull maximization out of
it, however). McLachlan & Krishnan
(http://www.amazon.com/Algorithm-Extensions-Wiley-Probability-Statistics/dp/0471201707)
give an example of this behavior (at least I remember one in the first
edition, it would be strange if they took it away in the second
edition). Also, since neither the likelihood nor its gradient are
computed, the algorithm has to rely only on -tolerance- (in terms of
Stata's -maximize-), which is the worst convergence criterion to work
with.
--
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
