Re: st: an estimation method question

[Maarten buis <maartenbuis@yahoo.co.uk>]

--- On Wed, 17/3/10, Xiang Ao wrote:
> I am thinking whether we can do a maximum likelihood with
> some constraints by observation.  That is, assume each
> share conforms to beta distribution, for example, then
> within each firm, the shares add up to one.  Would that
> be valid?

That distribution is called a Dirichlet distribution. The
basic version is implemented in -dirifit-. In your case 
you want to allow for different number of categories and
add the possibility that variables may have different 
values across categories, but add the constraint that 
the effects are constant across categories.

Before you start on such a programming excersise it is 
good (i.e. absolutely crucial) to know what to expect. 
So you need to fit a simpler model using existing 
software that is halfway reasonable. Going with the
-xtgee- or -glm- suggestion I proposed earlier would 
be such a way. It won't enforce the sum to one 
constraints, but in my experience that is almost never
a problem: if the model is halfway reasonable then the
predicted proportions will almost always add up to a
number very close to 1. Adding that constraint later
on, will thus add only a little bit of extra 
information, and thus make the standard errors a bit 
smaller, but I would not expect much more improvement 
than that. So you might want to consider if such small 
gains are worth the effort of programming a new 
procedure. The smaller standard error aren't the only 
pay-off though: programming is fun!!(but try selling 
that to your employer...)

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


      

*
*   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/