Re: st: 2 stage with Tobit and Poisson (GMM)

[Austin Nichols <austinnichols@gmail.com>]

Andrea Rispoli <andrea.rspl@gmail.com>:
Sounds reasonable to me.  Probably some reviewer will ask you to
produce IV diagnostics (id, weak IV, overID, test of endog) using
linear IV since that is where the theory is most developed.  You can
transform your variables using e.g. ln(1+x) to get approximately the
same functional form (not quite, but close enough for comparison,
probably) and to get good weak identification stats from -xtivreg2-,
you will need to use not your predicted values but the original
instruments.

On Thu, Mar 1, 2012 at 7:43 AM, Andrea Rispoli <andrea.rspl@gmail.com> wrote:
> Dear Statalisters,
> I am running a panel two stage model. In the first stage the dependent
> variable is continuous and truncated therefore a tobit model would be
> appropriate.
> In the second stage the dependent variable is a count variable
> therefore a poisson regression would be appropriate.
>
> One possibility would be to use ivreg and assume linearity for both
> stages. However Angrist 2001 suggests that when the first stage is non
> linear and the second is [linear], a solution is to use the appropriate model in
> the first stage (e.g. Probit, Tobit), obtain fitted values and them
> use them as instruments in the ivreg.
>
> Since in my case also the second stage is non linear, I was
> considering the Tobit model in the first stage, obtain fitted values
> and use them as instruments in a exponential panel regression with
> endogenous regressors (using GMM) in the second stage. Would this be
> correct or is [am I missing something]?
*
*   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/