Re: st: biprobit with endogenous regresor and marginal effects

[Austin Nichols <austinnichols@gmail.com>]

Lina C <linacs81@gmail.com> :
The point of that message (with perhaps the implied threat of
incorrect results from -mfx- or -margins-) was that many different
marginal effects are possible, and one can manipulate the predicted
probabilities and samples over which one averages differences in
predicted probabilities to estimate the desired population/dgp
quantity, but one has to carefully specify the estimand first; see
also
http://stata.com/meeting/chicago11/materials/chi11_nichols.pdf

On Wed, Aug 17, 2011 at 3:50 AM, Lina C <linacs81@gmail.com> wrote:
> Hello,
>
> I'm estimating a bivariate probit model with an endogenous regressor
> (that is instrumented). I have been following the threat (see below)
> where it is suggested to calculate manually the marginal effects given
> that -mfx neither -margins would work to estimate the marginal
> effects.
>
> Y1 = Y2 X
> Y2 = Z X
>
> I have two questions:
> 1. I have calculated the marginal effects as proposed below. However,
> they have a different sign to the original coefficient of the
> biprobit. I was wondering if this could happen or if there has been
> some improvement to calculate the marginal effects in those cases in
> Stata.
>
> 2. Is it common for the standard errors using the iprobit to be large
> in comparison to running a 2SLS with 2 binary outcomes?
>
> Thank you.
>
> Lina.
>
>
>
> -----Original Message-----
> From: [hidden email]
> [mailto:[hidden email]] On Behalf Of Austin Nichols
> Sent: segunda-feira, 1 de Fevereiro de 2010 18:03
> To: [hidden email]
> Subject: Re: st: marginal effects in biprobit and average treatment effect
> in switching probit
>
> Paula Albuquerque <[hidden email]>:
> -mfx- is not really appropriate (and -margins- will refuse to participate).
> You need to fully specify what margin you are thinking about to get a
> sensible marginal effect. Try calculating the marginal effect of X using
> predictions after your -biprobit- and after the user-written command.
> E.g.you might want to calculate the conditional prob of Y=1 given X=1
> less the cond prob of Y=1 given X=0, letting X=1 and X=0 in turn for
> each observation, and then averaging over observations.
>
> sysuse nlsw88, clear
> g x=race==1
> g y=married
> biprobit (y=x grade south smsa) (x=industry occupation union)
> mfx
> * margins, dydx(*)
> g wasx=x
> replace x=1
> predict p1a, p11
> predict p1b, p10
> predict p1c, p01
> predict p1d, p00
> g p1=p1a/(p1a+p1c)
> replace x=0
> predict p0a, p11
> predict p0b, p10
> predict p0c, p01
> predict p0d, p00
> g p0=p0b/(p0b+p0d)
> replace x=wasx
> drop wasx
> g dp=p1-p0
> su dp
*
*   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/