cmp routine

[Dario Maimone Ansaldo Patti]

Dear all,

I am trying to fully understand the functioning of cmp routine. I would like to estimate a triangular system like:

y= cons +beta*x1+gamma*y2

y2=cons+delta*z1

I have done it using cmp and then I tried to replicate the estimates using the procedure described in http://www.stata.com/support/faqs/st...es-regression/

The results are comparable but the S.E. that I obtain using the cmp routine are a bit different (they are smaller, particularly for the exogenous regressor x1). Nothing changes in terms of the significance of the estimated parameters, anyway. However, I am curious to understand why I observe such a slight difference. Do you have any hint?

Best

Dario

[David Roodman]

cmp uses Maximum Likelihood. The example at http://www.stata.com/support/faqs/st...es-regression/ does not. So I wouldn't expect them to give the same answer.

If you add x1 to the y2 equation, then you'll be doing LIML with cmp and standard 2SLS with, say, ivregress. Those will match, since in general LIML and 2SLS are the same in exactly identified systems, i.e., ones in which the number of instruments = the number of instrumented variables. If you add instruments, then the two will diverge.

[Dario Maimone Ansaldo Patti]

Thank you very much David.