Estimating an instrumental variable model - endogenous variable is binary and DVs are continuous

[Bruno Hoepers]

Hi all. I have a cross-sectional dataset with information on the presence or absence of a Brazilian TV network's signal (globo_coverage), for each one of the 3,659 localities in Brazil in 1982.

I want to estimate the effect of having access to Globo Network's signal on the vote for a Brazilian political party in the general election of 1982 in the country, for several offices (i.e., federal deputy, state deputy, governor and senator).

The problem: variable "globo_coverage" is probably endogenous. I need to run an instrumental variable model. This endogenous variable is binary and all my dependent variables are continuous. Estimation through 2SLS using the command ivregress seems wrong.

Variables I want to use as instruments: distance to the state capital in km (dist_cap_est) and geographical latitude.

Could someone provide me advice on how to proceed here?

Thank you.

PS: please let me know if you need more info about the dataset.

[George Ford]

https://journals.sagepub.com/doi/pdf/10.1177/1536867X1401400301

[Bruno Hoepers]

Very interesting command and paper, George. Thanks for bringing it to my attention.

[Joro Kolev]

You can use TSLS, Bruno. It is appropriate for your context.

A little more fancy procedure would be if you use a first stage probit or logit, generate the predicted probabilities, and use these predicted probabilities as instruments.

Note, you do not plug in the predicted probabilities directly, you use them as instruments.

[Bruno Hoepers]

Thank you for your answer, Joro. I definitely need to think more about this project and the analyses with these data.