Again, related to the heckprob (Heckman probit selection) command,
1. How to test if I have omitted variables issue in the heckprob model,
especially when I don't have any information about those omitted variables?
Or, put it more properly, I have a bunch of variables thrown into my
heckprob equation, however, I am not sure if I omitted any, even worse, the
variables that I have are the only ones, and I just want to show that the
variables that I have suffice and I don't have omitted variable problem.
2. Question Two is about some common criticism of Heckman model that it is
sensitive to model specifications. It seems to me hard to conduct Hausman
test of the heckprob model as I am not sure what kind of model specification
is a consistent and yet less efficient one. One solution is just to use
Hausman test to test the selection equation (compare estimates from the
heckprob and a single probit equation of the selection process) and then
argue that since the estimates for the selection is consistent, we should
expect the estimates in the main equation is also consistent? Also, what if
Rho is significantly different from zero for models with more variabless,
while not different from zero if I drop some of the variables (I saw a
similar posting before, but didn't see any answers yet).
3. Again, how to get a scalar measure of the heckprob model (like R-square).
I posted this question days ago. The real problem is that, for example, if I
want to use McFadden's R2, what's the null model (the null model with two
intercepts only for both the main and selection equations will not be
estimated, or shoudl I use the null model when the selection process is
fully specified, while only have an intercept for the main model?). Need