Marginal effects after bivariate probit and parameter homogeneity test.

[Dario Barboni]

Hello people,

I am replicating an empirical economic study in which the authors conduct a bivariate probit regression with marginal effects at the sample means. They report two columns with marginal effects (one for each of the 2 dependent variables). The results show opposite effects for the two DV, which confirms their theory. In addition, they run a parameter homogeneity test for each independent variable to check if the differences between the effects are significant.
If I run a biprobit in stata, I also get 2 tables: One for each of both DV. However if I run mfx afterward, I get only ONE table. I also tried all 4 combinations, i.e. p(00), p(10), p(01), p(11) but all effects are completely different compared to the study.

Does anyone have an idea what the authors did in econometric terms? How do I get the two coefficient tables as two mfx tables? And what could be the command for a parameter homogeneity test?

Maybe it helps to have a bit of the theoretical background:

They use a biprobit because of the following reason: The data comes from a questionnaire in which every observation is a firm (i.e. the data is at firm level). Each firm has a dummy variable for "patent application" and for "trade secrecy". Both can be 1 or 0, independently, so there are 4 possible combinations. The theoretical model that is tested, however, considers one innovation that can either be patented or kept secret (i.e. the model is at innovation level/product level). At first they run a probit regression where "patent application" is the DV. However, as in the sample there are firms that have a value of 1 for patenting AND secrecy, they conduct a biprobit. The results show opposite effects of the independent variables on patenting and secrecy, which is a likable result that does not contradict the results of the first probit.

I know this is a tough and long question but I don't know who to ask anymore.

Thank you very much and best regards!

[John Mullahy]

Dario: I am not completely certain what you are hoping to estimate, but my best guess from your question is that you are looking for properties of the marginal distributions of y1 and y2, not their joint distribution. The p00, p01, p10, and p11 predictions in margins capture the marginal effects on the joint probabilities. I suspect you might be looking for either

biprobit (y1 x1 x2) (y2 x1 x2)
margins, dydx(*) predict(pmarg1) predict(pmarg2)

or

biprobit (y1 x1 x2) (y2 x1 x2)
margins, dydx(*) predict(xb1) predict(xb2)

[Dario Barboni]

John Mullahy , thank you very much. The results look much better now. Does anyone have an idea regarding the parameter homogeneity test?

[Dimitriy V. Masterov]

If you add the post option to the margins command, you can then use test command to do hypothesis tests the output. The coefl option will also be helpful here to figure out what Stata call the AMEs. This should allow you to do the PHT.

[Gabriel Temesgen]

Dears,
I am conducting research on WTP (Contingent Valuation Method) using double bounded price elicitation method. I think the best model is the Bivariate probit model. When I run the model by inserting the initial bid as an independent variable, I got nothing. it iterates many times and says not concave. What I feel as first was, (the initial price for using the technology is birr 500 for users and birr zero for non-users) whihc means the dependent variable and the initial bid variable become perfectly the same. how can I solve the problem and run the model? maybe, If you need also my data and do file, I can attach you here.
waiting for. regards

[jomike]

Dear John Mullahy,
I would like to ask if the code you posted above on31 Jul 2019, 12:24 estimates the average marginal effect in biprobit models and therefore is doing exactly the same as the code posted by Jeffrey Wooldrige on 27 Apr 2017, 09:00, which you can find under this link: https://www.statalist.org/forums/for...ther-than-e-b?
I am interested in estimating average marginal effects in biprobit models.
Would be very helpful to know!
Best,
Johanna

[John Mullahy]

Johanna: Jeff Wooldridge's code is set up to handle the case where y2 is a right-hand-side variable in the structural equation for y1. The code I presented was for the simpler case where neither y1 nor y2 appears as a RHS variable in the other's structural model.

[jomike]

Dear John, thanks a lot for your immediate answer- very helpful to know!