Instrumental Variable Probit using Panel Data

[Kat De Chavez]

Hi All,

Can someone let me know how I can implement probit IV (first stage is OLS, second stage is probit) using panel data?

Is there a single command or a series of steps to do this? I know that ivprobit command is for cross-sectional data.

Thanks.

[Carlo Lazzaro]

Kat:
welcome to the list.
Can't you switch to a pooled ivprobit?

[Kat De Chavez]

Carlo, I have a panel data structure so using ivprobit will not be appropriate.

[Carlo Lazzaro]

Kat:
I got that you have panel data structure. That's why I mentioned pooled ivprobit, not ivprobit.
Unfortunately, I have no idea of other possible solutions (and the lack of others' replies seems to confirm the difficulty of your task, so far).

[Carlo Lazzaro]

Kat:
I got that you have panel data structure. That's why I mentioned pooled ivprobit, not ivprobit.
Unfortunately, I have no idea of other possible solutions (and the lack of others' replies seems to confirm the difficulty of your task, so far).

[Jeff Wooldridge]

Carlo is correct: You can apply ivprobit to panel data. You would want to use the "cluster(id)" option, where "id" is the cross sectional identifier, to obtain valid inference. If you are worried about correlation between the unobserved effects and explanatory variables you can use a correlated random effects approach. Papke and Wooldridge “Panel Data Methods for Fractional Response Variables with an Application to Test Pass Rates,” Journal of Econometrics 145, 121-133, July 2008, shows how to use a two-step control function method. But you can also use ivprobit in one step.

[Cooper Felix]

Carlo is correct: You can apply ivprobit to panel data. You would want to use the "cluster(id)" option, where "id" is the cross sectional identifier, to obtain valid inference. If you are worried about correlation between the unobserved effects and explanatory variables you can use a correlated random effects approach. Papke and Wooldridge “Panel Data Methods for Fractional Response Variables with an Application to Test Pass Rates,” Journal of Econometrics 145, 121-133, July 2008, shows how to use a two-step control function method. But you can also use ivprobit in one step.

Dr. Wooldridge,

I wanted to make sure I fully understand what you suggested, so basically for panel data with binary dependent variable (DV) and continuous endogenous variables (x1 x2 x1*z3), one can specify the following with ivprobit to obtain valid inference:

 ivprobit DV controls (x1 x2 x1*z3 = z1 z2 z1*z3), vce(cluster firm_ID) first

where controls and z3 are exogenous variables; z1 and z2 are instruments;

Is there any other command that I can use for robustness check? Thanks!

[高佳]

Carlo is correct: You can apply ivprobit to panel data. You would want to use the "cluster(id)" option, where "id" is the cross sectional identifier, to obtain valid inference. If you are worried about correlation between the unobserved effects and explanatory variables you can use a correlated random effects approach. Papke and Wooldridge “Panel Data Methods for Fractional Response Variables with an Application to Test Pass Rates,” Journal of Econometrics 145, 121-133, July 2008, shows how to use a two-step control function method. But you can also use ivprobit in one step.

Dear Dr. Wooldridge,
I came across a similar problem like that in your paper (Papke and Wooldridge, 2008) with the only difference that the dependent variable is binary instead of fractional.
action_i,t = a_i + X_it,1 + X_it,2 + e_i,t

in which i, t are subscripts stand for individual and time, respectively. action_i,t is a dummy variable equals 1 if individual i take certain action at t and equals 0 otherwise;
It's reasonable to suspect that X_it,1 is endogenous, we thus find an instrument z_it for it.

action_i,t takes a disproportionally larger number of 0 (around 3/4), and X_it,1 (also contains a large proportion of 0) is the number of people lining behind the focal customer waiting to check out at a supermarket.

The question is, can I use the method you and Pake proposed in your 2008 JE paper directly in my setting? If not, what kind of modification do you suggest?

Thanks so much, not just for this particular question but also for all the econometrics I learned from your fantastic books and journal articles.