Dear Philipp, many thanks for your response. I've checked the journal reference
and you are right all the discussion is framed within nested models. I guess I
should be more specific about my query:
My data is a household panel survey out of which I estimate my main dependent
variable (my regressor of interest). Then I take this generated regressor into
a probit model where my intuition tells me that since the sample is large
enough (around 17,000 observations) there should not be an important problem of
efficiency, but I still wanted to check.
In a second type of model, the same predicted regressor is instrumented and
then taken into a probit, in some sort of two-stage fashion. Given that my
generated regressor is actually the dependent variable in the IV regression
(and not an instrument in itself) I wonder if it should not cause any problem
for further inference as all source of variation is left for the residuals in
the equation and hence I can proceed with the probit afterwards. If you/anyone
think this is not really the case then I wonder if carrying out a Maximum
Likelihood IV Probit (ie, run both the IV regression and the probit all in one
step to achieve efficiency) could solve part of the problem, but still need to
find a solution for correcting the Standard Errors.
Again any clarification on this would be very helpful.
Alejandro
In message <[email protected]> [email protected] writes:
> Aljandrop,
>
> off-list, since I am not really sure.
>
> > Am trying to estimate a Probit Model where my main regressor is based on
> > predictions from another regression. Is there anything I should do to
adjust
> > for the standard errors given that am using an estimated dependent variable?
>
> There is a recent issue of Political Analysis (Volume 13, Number 4,
> Autumn 2005), a "Special Issue on Multilevel Modeling for Large Clusters."
>
> If I remember correctly, one of the things almost all authors argue is
> that you do not need to adjust for standard errors. But, as the title of
> the special issue indicates, this advice is tailored to situations in
> which you have large clusters (lots of level-1 observation per level-2
> unit).
>
> Also, you may want to check out GLLAMM, in case you want to fit the
> model differently.
>
> HTH,
> Philipp
>
> *
> * 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/
>
--
Alejandro de la Fuente
Department of International Development/QEH
University of Oxford, Mansfield Road, Oxford OX1 3TB
Tel: 01865 281836
*
* 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/
