heckprob estimation

[Melissa Pf]

Dear Stata User's,

I want to estimate a heckman selection model for nonresponse analysis of survey data.
Since my dependent variable is binary I decided to go for the bivariate probit model with sample selection.
I figured out the code I have to use is heckprob based on the work of Van de Ven and Van Praag 1981.
However, from the manual and the help page of heckprob it does not get clear to me which estimation procedure is used. In the heckman command you can chose between full ML estimation and the two-step procedure.

From my research so far and the article from Van de Ven and Van Praag 1981 I think the two-step procedure is applied where first the selection equation is estimated and a lambda is calculated which is then inserted in the outcome equation. Is that the case or is this only valid for modeling continuous variables? If yes, why does the output not show a value of Lambda like it is the case in the heckman command?

I already looked a lot for the solution (in books and the internet) and did not find a clear description.
I hope you can help me with that.

Kind regards,
Melissa

Reference: Ven, W. V. d., and B. V. Praag (1981): \The Demand for Deductibles in Private Health Insurance: A Probit Model with Sample Selection," Journal of Econometrics, 17(2), 229 - 252.

[Phil Bromiley]

If you read the full documentation including the Methods and Formulas , you'll see it is a maximum likelihood estimation. You can also tell it is maximum likelihood rather than two step because the examples who iteration logs.

[Melissa Pf]

Thanks for your answer!

Unfortunately another question came up when using these kinds of models. Maybe you can help me with that as well.

I want to compare the results from the heckprob model to those of some other adjustments (mostly weighting methods). Besides the regression coefficients from the probit regression I also want to compare the estimated means/ proportions from the different methods. The heckprob postestimation manual shows a lot of different possible predicted probabilities and I'm not quite sure which one I should use to "predict" the missing Y values and hence get an estimate for the proportion of my Y variable which is comparable to the others.

Thank you for any recommendations!

[Jeanne Kelly Ruiz]

Hi!!

I would like to show u the output of my heckprob. I am not sure if it has problems because the log pseudo is too high and the wald chi2 p value is too low, although, i see that many of my variables are significant and the wald test of independence is ok (P value under 0.05). Could you see any problem?

Thanks