Re: st: estimating treatment effect of a binary endogenous regressor on binary outcome

[Jorge Eduardo Pérez Pérez <jorge_perez@brown.edu>]

It looks like your instrument may be weak. What is your instrument? Is
it binary? What is the F-statistic on the first stage relationship
between the instrument and the endogenous regressor?

Reducing your control group to increase the number of treated
observations to control observations is ill advised in general. IV has
a finite sample bias and you may just be increasing the bias.

What is your benchmark to say that your estimates of 10 or 0.8 are too high?



On Nov 13, 2013 4:24 AM, "Erkan Duman" <erduman@sabanciuniv.edu> wrote:
>
> Hello.
> I am working on the impacts of having a migrant at home on the school
> attendance of children. My regressor is an endogenous binary variable
> and the outcome is also binary. I have searched a lot to find an
> estimator which will consistently estimate the treatment effect.
> Chiburis et al. (2012) paper recommends bivariate probit or IV2SLS.
> However, bivariate probit requires strong assumptions (bivariate
> normal errors) and my specification does not satisfy this assumption
> which results in severly biased estimates. IV2SLS gives a treatment
> effect over 1 (actually around 10). In my study, the treatment
> receivers only constitute around 2 percent of the sample. I believe
> this low treatment probability causes problems in estimating the
> treatment effects because when I reduce the size of the control group
> and come up with a treatment group of 16% of the sample, the treatment
> estimate fits in (0,1) range ; still it is too high around 0.80. I
> have encountered a similar thing to my problem as "rare events" in the
> literature, but doesn't seem to solve my problem or maybe I am wrong.
> Also in http://www.stata.com/meeting/chicago11/materials/chi11_nichols.pdf
> some semi-parametric estimators are suggested by  Austin Nichols which
> do not require bivariate normal errors but weaker assumptions.
> However, I am not familiar with semi-parametric or nonparametric
> estimators. Can anyone help me with an appropriate semiparametric
> estimator and its stata command?
> I believe there is a solution to this estimation problem and I hope
> someone will help me.
> Thanks.
> Best regards.
>
> --
> Erkan Duman
> Graduate student - PhD
> Faculty of Art and Social Sciences
> Sabancı University
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/faqs/resources/statalist-faq/
> *   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/