Margins Command after Biprobit Estimation

[Rene Maric]

Dear all,
I am estimating an instrumental variable regression with a dummy outcome variable Y, a dummy endogenous variable X, and a continuous instrument Z. When estimating a linear 2SLS model:

  
 ivreg 2sls Y (X = Z)

I obtain an estimate of 3, which is unreasonable given that Y and X are dummy variables. I ascribe this to the relative low treatment probability (only 10% have X=1) in my sample and the low sample size of around 4000 (See Chiburis 2011).

I therefore turn to estimating a bivariate probit model:

  
 biprobit (Y = i.X) (X = Z )

However, I am now confused about the correct command to obtain the marginal effect of X turning 1 on Y.

Either I type:

  
 margins, dydx(X) force post

Or:

  
 margins, dydx(X) force post predict(pmarg1)

Both commands give me very different marginal effects, which command (if any of the two) would give me the effect I am interested in?

[Rene Maric]

Just to clarify, the effect i am interested in is the following:

P[Y=1|X=1,Z] - P[Y1=1|X=0,Z]

[Mustafa Coban]

Hi Rene,

biprobit is in your case not the approriate command, since you want to fit a recursive bivariate probit model. I've written a new package rbiprobit that handles your case and provides postestimation commands to compute treatment effects

ssc install rbiprobit

Applied to your example, the syntax would be

rbiprobit Y = , endog(X = c.Z)
rbiprobit tmeffects, tmeff(ate)