using cmp for estimating a bivariate probit model with random effects

[Elin Vimefall]

Hi

I’m estimating a bivariate probit analyzing what determines the child’s probability to go to school and the probability to work.
I have one endogenous variable (exp) and to control for this I use x3 as an instrument.
I also want to include random effects at the household level because there can be more than one kid from each household.
I use the program cmp (Roodman 2011):
cmp (exp = x1 x2 x3 || household: )(work = x1 x2 exp || household: ) (school = x1 x2 exp || household: ), cluster(district) ind($cmp_cont $cmp_probit $cmp_probit)
However my model doesn’t seem to run (my computer has been working on it more than a day).
Can someone tell me what I’m doing wrong (it works without the random effects)? Or if there is any better way to control for the fact that there can be several children in the same family.

Best regards /Elin Vimefall

Roodman D. (2011). Fitting fully observed recursive mixed-process models with cmp. The Stata Journal 2011, 11(2), pp 159-206.

[David Roodman]

Hi Elin,
In principle, cmp can fit that model, and you have specified it correctly. However, it is very demanding to have a three dimensional normal random effect whose cross-correlations are unconstrained, on top of an individual-level error of the same form, and with two variables heavily censored (probit). By "demanding" I mean that it may take a very large sample for convergence to be achieved. I would experiment with the covariance() option to make one level or the other uncorrelated, as in cov(independent unstructured). Also try breaking the model into two models, one for school and one for work. The theoretical gain from combining them is efficiency. But you can still estimate them separately.
--David