Bookmark and Share

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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: marginal effects in biprobit

From   Janice Compton <>
Subject   st: marginal effects in biprobit
Date   Mon, 13 Jun 2011 15:22:25 -0500

I am using BIPROBIT (Stata9) to estimate an IV model of the type
BIPROB (Y X1 X2 X3) (X1 X2 X3 Z)
where Y, X1 and Z are all binary variables.

I have a comment on an old post and then a question. First, to find marginal effects, I started by using the program found here:

The description provided sounds like Average Treatment Effect, but I was getting negative results where I had positive coefficients. After a lot of searching, I discovered that ATE is found by either of the following procedures:

gen wasx1=x1;
replace x1=1;
predict p1, pmarg1;
replace x1=0;
predict p2, pmarg1;
replace x1=wasx1;
gen ATE1=p1-p2;
sum ATE1;


predict xb1, xb1;
scalar b_x1=x[x1]
gen ate2=0;
replace ate2=norm(xb1+ b_x1) ? norm(xb1) if X1==0;
replace ate2=norm(xb1) ? norm(xb1 ? b_x1) if X1==1;
sum ATE2;

These give the same results. I think the program from the above post looks like a reasonable thing to do but I?m not sure what the interpretation is. Can anyone explain that?

My question relates to the pmarg1 estimate.  If I run the following:
gen wasZ=Z;
replace Z=1;
predict check1, pmarg1;
replace Z=0;
predict check2, pmarg1;
sum check1 check2;

I find that check1 and check2 are identical, so values of the instrument do not effect pmarg1. This suggests to me that pmarg1 is estimating the predicted probability of Y at observed values of X1 and not estimated values of X1, as I had expected. Does anyone know how to estimate the predicted probability of Y conditional on values for the instrument, Z?

Thanks very much for your consideration.

Janice Compton

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index