cmp: bivariate tobit w/ bootsrapped SE's

[justin winikoff]

Hello,
I am beginning a project in which I'd like to use a bivariate tobit model where both dependent variables are left censored at 0. I've stumbled across cmp, which seems like my best option.
The model I'd like to run is something along the lines of:

cmp (y1= x1 x2) (y2= x1 x3), ind("cond(y1>0, $cmp_cont,$cmp_left)" "cond(y2>0, $cmp_cont,$cmp_left)")

However, x1 is estimated via a first stage and predicted values are used. I believe the standard errors must be bootsrapped. Just toying around with the cmp command and a bivariate probit with sample data, it did not appear vce(bootsrap, reps(100)) worked. It said "invalid vce option". Is there a way to bootsrap?

Moreover, will I be able to get the marginal effects from a tobit model (of the observed, not latent variable) using the typical margins notation for tobit models?

Thank you!

[David Roodman]

You should be able to incorporate the first stage directly into the estimation, as a third equation. cmp is very flexible that way.
If you still want to bootstrap you can do it with the bs/bootstrap prefix command.

Yes, you can still do marginal effects, though mostly working with one equation at a time. margins doesn't receive the relationships between the equations very well, so it's hard to get, e.g. the effect of a variable in one equation on outcomes of another.