Heckprob twostep?

[Birte Olsen]

Dear Statalist,

I am running a Heckprob model, and I wanted to do a twostep estimator instead of the Maximum likelihood estimator. However, I could not find any options for twostep with the Heckprob command as opposed to the Heckman model ... If anyone could inform me whether this is a STATA issue or a methodological issue (and possibly why heckprob cannot be run in two steps), that would be greatly appreciated!

[Mike Lacy]

I'm not sure about the details of estimation methods and assumptions here, but perhaps the -twopm- user-contributed model (-ssc describe twopm-) might be helpful here.

[Jeff Wooldridge]

There is no easily implementable two-step estimator for probit like there is for the linear model under the usual assumptions. Let's just say one would never implement it because it's just as complicated as the joint MLE.

In the linear case, there are some advantages to using a two-step method. But not really in nonlinear cases.

You will sometimes see the inverse Mills ratio from the first-stage probit added to the second stage probit. However, unless one is using that method to estimate an average treatment effect, the coefficient magnitudes and average partial effects won't be easily interpretable. I'd go with the joint MLE.

[Birte Olsen]

Dear Mike and Jeff (If I may),

Thank you for the response! That makes very much sense; I will go with the joint MLE approach!

As for the twopm, I am already doing a logit model following the two-part model approach as a robustness check! Thank you for the suggestion either way!