Suppose I have the following model:


$D=1(x'\alpha+\gamma Q+u>0)$

and I want to estimate $\gamma$, the error terms $e,u$ are jointly normal.

If $e$ and $u$ are correlated, can I still estimate $\gamma$ consistently by running a probit using the second equation only?

Why or why not?

If I cannot, how could I estimate $\gamma$ in this case?


1 Answer 1


If $u$ is normally distributed conditional on $x$ and $Q$, then you could estimate through probit the second equation. Because $Q$ depends on $e$, and $e$ is correlated with $u$, this assumption likely fails (that's your point?). Also, note that if $u \sim N(0, \sigma^2)$ then probit really estimates $\frac{\gamma}{\sigma}$. (i.e., the ratio of the coefficient to the standard deviation of the error).

If the equation with $Q$ as an outcome had a right-side variable that is not in the equation with $D$ as the outcome, then that variable could serve as an instrument and you could perform IV Probit. At present, there isn't such a variable, and thus IV probit isn't an option.

  • $\begingroup$ Thank you so much, Michael! This is very helpful. $\endgroup$ Jul 14, 2022 at 3:05
  • $\begingroup$ Could you give me some reference on the ivprobit that you mentioned? $\endgroup$ Jul 14, 2022 at 3:21
  • 1
    $\begingroup$ The documentation of the Stata command describes the method: stata.com/manuals/rivprobit.pdf $\endgroup$ Jul 14, 2022 at 6:43
  • $\begingroup$ Thanks, Michael! $\endgroup$ Aug 10, 2022 at 2:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.