Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# Re: st: marginal effects in biprobit and average treatment effect in switching probit

 From Austin Nichols <[email protected]> To [email protected] Subject Re: st: marginal effects in biprobit and average treatment effect in switching probit Date Fri, 15 Jun 2012 12:27:53 -0400

```Monica Oviedo <[email protected]>:
The relevant code appears on slide 14 of 46 and is prefaced by the text:
*
How do we calculate the marginal effect of treatment after biprobit? Three
"obvious" approaches: use -margins-, use -predict- to get probabilities, or use
binormal() with predicted linear indices. The last is more correct, but all
should give essentially the same answer.
*
Evidently, the -margins- approach is the least correct.
If results differ, you should prefer the other approaches, but
read the references on interpretation of various ATE estimates.

initiate a new query.

I doubt -biprobit- will work well for the model with an endogenous interaction.
You have a single instrument z for two endogenous variables y2 and x*y2.

biprobit (y1=x y2 xy2) (y2=x z)
try
ivreg2 y1 x (y2 xy2=z xz)
(ivreg2 is on SSC) and read the references on weak instruments
diagnostics in the -ivreg2- help file.

On Thu, Jun 14, 2012 at 6:45 AM, Monica Oviedo <[email protected]> wrote:
> Dear Statalist:
>
>  I'm estimating the effect of an endogenous dichotomous variable y2 on a
> dichotomous variable y1 using a recursive biprobit model:
>
> biprobit (y1=x y2) (y2=x z)
>
> Where z is the exclusion restriction. I'm interested in the marginal effect
> of y2 on y1, which I think is:
>
> E[y1/y2=1] - E[y1/y2=0]
>
> I did what Austin Nichols suggested in this thread (namely, the conditional
> prob of Y1=1 given y2=1 less the conditional probability of Y1=1 given y2=0,
> letting y2=1 and y2=0 in turn for each observation, and then averaging over
> observations). In addition, I followed the procedures sugested by him in
> this file:
> http://www.stata.com/meeting/chicago11/materials/chi11_nichols.pdf
>
> This is:
>
> margins, dydx(y2) predict(pmarg1) force
>
> I think the latter is correct for estimating what I need. However, I get
> very different results from both procedures (in the first case a marginal
> effect of 0.08 vs a marginal effect of 0.45 using the second way).
>
> What is the difference between both procedures? Is it supposed that they
> estimate the same effect?
>
> A final question is if biprobit is well suited for estimate the following:
>
> biprobit (y1=x y2 x*y2) (y2=x z)
>
> This is, if  there is any problem when an interaction term between the
> endogenous variable y2 and a continous x is added.
>
> Regards,
>
> Monica Oviedo

*
*   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/
```

• References: