This paper compares asymptotic and finite sample properties of linear IV and
bivariate probit in models with an endogenous binary treatment and binary outcome.
The results provide guidance on the choice of model specification and help to
explain large differences in the estimates depending on the specification chosen.