Perhaps due to a surfeit of mincepies I'm blanking on the following question:
How does one mechanically get an ATE plus standard error out of a fitted switching regression ('Tobit-5') model?
A worked example using the R package
sampleSelection would be ideal.
I suppose this is the practical version of this question, but if I could operationalise that I wouldn't be asking.
OK, so it seems (from some old Wooldridge slides) that one can average the covariates and then multiply them by the difference between the vector of coefficient estimates in the control model and the coefficient estimates from the treatment model. But how does that work when covariates cannot be coherently averaged, e.g. when they are dummy variables?