Re: st: RD design using -probit- and -margins-

[Austin Nichols <austinnichols@gmail.com>]

If z really takes on 25 possible values, then why not compare z==0 to
z==-1 then extend to a linear model with z in {0,1} or z in {-1,-2}
before going to a quadratic in z?  This is the idea of RD, using a
small range of possible assignment variable values to compare like to
like, and extending (usually via nonparametric regression) to the
boundary to extrapolate to the one point where the actual and
counterfactual conditional mean outcomes are both observable.

If you have a sharp design, the dummy t==(z>=0) captures the jump at
the boundary, and that is the only coef of interest, not any of the
interactions.

On Fri, May 25, 2012 at 9:07 AM,  <altruist81@gmx.de> wrote:
> Dear Stata Users,
>
> I follow a regression discontinuity design (RD) where I have
> y=outcome [0,1]
> t=treatment [0,1]
> z=running variable [-12, -11, ... , 11, 12]
>
> and t=1 if z>=0.
>
> I fit several models of varying flexibility, i.e.
>
> -probit y i.t##c.z-
>
> and
>
> -probit y i.t##c.z##c.z-
>
> and
>
> -probit y i.t##c.z##c.z##c.z-
>
> I want to use -margins- to give me an interpretable estimate of the local treatment effect, i.e. the effect of t switching from 0 to 1 on y at z=0.
>
> My questions:
> 1. how do i have to specify -margins- ?
>
> 2. I have different outcomes of interest. In some cases 8more so with the more flexible models, -margins- would tell me the efect is not estimable. What does this suggest and is there any way around the problem?
>
> Many Thanks!
> Darjusch
>
> P.S. I am aware of -rd- , but prefer a parametric approach in this setting.
*
*   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/