Re: st: maximum likelihood estimation similar to Poi 2002

["Brian P. Poi" <brian@poiholdings.com>]

On 3/28/2011 4:46 PM, Alex Olssen wrote:
> Thanks a lot for pointing out e(ll), that should help a lot.
>
> I just remembered that I had one other problem with -nlsur-
>
> It has to do with the number of parameters that I can estimate.  I
> have 35 years of annual observations on 4 shares.  I want to use 12
> regressors in each of 3 equations - one dropped for estimation.
>
> -nlsur- doesn't like this.  But to my mind this should be possible.
> Dropping one equation for estimation I no longer have any cross
> equation restrictions and I estimate as if by OLS equation by
> equation.  Clearly with 35 observations I can estimate a simple model
> with 12 regressors by OLS.  Alternatively think of the simple
> multivariate case.
>
> Further more in R I can estimate a model with 3 equations with 12
> regressors each using 35 years of annual observations on 4 shares.

-nlsur- uses a simple counting rule to determine identification, and it 
tends to be conservative because it doesn't consider cross-equation 
restrictions (or, equivalently, that the same parameter appears in 
multiple equations).  In certain situations it may be possible to fit 
models with fewer observations than -nlsur- thinks you need -- and in 
those cases -ml- is one alternative.  However, based on my experience 
with AIDS and similar models, the more observations and variation in the 
data, the better.  Usually when people have non-convergence issues in 
fitting these models, it often ends up being a matter of not having 
enough data to identify all the parameters.

   -- Brian Poi
   -- brian@poiholdings.com
*
*   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/