Hi all,

I am running an ivprobit model along the lines of the following: "ivprobit y x2 x3 x4 x5 x6 x7 x8 (x1 = x9 x10 x11 x12)"

When running this model as a simple probit model, the coefficient on x1 (the endogenous variable) is positive and statistically significant. However, as I had hypothesized, when using ivprobit, the coefficient on x1 is negative and significant. However, when I run "margins, dydx(x1) predict(pr)," the value ends up being positive.

This is the value I would have expected for the margin on x1 if I had run a simple probit model rather than an ivprobit model. In fact, the results on the "margins, dydx(x1) predict (pr)" command after the probit and ivprobit models are nearly identical.

After reading through several threads here, it appears that recent versions of stata updated margins after ivprobit to "account for endogeneity" without much more explanation. Indeed, when I use version control, I get a negative value after running "margins, dydx(x1) predict (pr)"

I'm wondering if, essentially, the newer version of margins is calculating the margins for the original variable (in a model that also adds the instruments as additional covariates) whereas the older version was calculating the margins for the instrumented version. Does anyone know if that is true? I want to know the margins for the instrumented version, so if I am right then I would just use the output from the version controlled version of margins.

To get a sense for what I mean, you can run the following and get exactly what you would expect if what I am saying is true:

webuse laborsup, clear
probit fem_work other_inc fem_educ kids male_educ
margins, dydx(other_inc) pred(pr)
ivprobit fem_work fem_educ kids (other_inc = male_educ)
margins, dydx(other_inc) pred(pr)
version 13.1: margins, dydx(other_inc) pred(pr)

Compare the results. Simple probit yields a coeff of -0.038 on other_inc; in the IV approach this changes to -0.054. Thus, one would expect margins on other_inc to be negative after probit and negative (but with slighlty greater absolute value) after ivprobit.

However, margins pr after probit yields -0.00978 and after ivprobit also yields -0.00978. After the version control, however, you get -0.014, which is about what you would expect (again if you want the effects of the instrumented version and not the original one).

Maybe I'm not 100% right here because in my case I have more than one instruments and the results for margins after probit vs ivprobit are not 100% exact, but they are very close. Either way, my question remains, if I want to estimate the marginal effect of x1 on the probability of y being 1, after accounting for endogeneity via the IV approach, should I just be using the version controlled version of margins?

Thanks so much for any insights!
Rajan