Thanks a lot for your help and sorry for my late replay (was out of town).
I tried to compute the two-step procedure you suggested for a binary probit
and then compare results with heckprob command. Estimations seem reasonable
although with the two-step procedure I am loosing some observations with
respect to the ml estimation. At this stage I don�t think I fully
understand how the heckprob command works and how the missing observations
of the latent probit equation are treated (?)
However, in this two step procedure I guess I also need to compute new
standard error and rho once I introduce the inverse mills ratio. Do you
think this is correct?
Many thanks again for your precious help.
Best wishes
Laura
Laura Serlenga wrote:
ls> i am relatively new to stata but, since i am currently working on
ls> migration issues (intention to return and remittance) with
cross-sectional
ls> data. i am finding it very useful. However, I need to run a Heckman
ls> selection model with an ordered probit in the outcome equation. I saw a
ls> couple of messages on the list dated 2003, does anybody know whether
there
ls> has been any further development on the issue? would you suggest any
other
ls> solution rather than modifying the heckprob.ado file?
One solution is to run a probit selection model based on inclusion or
exclusion from the observed variable (y_1), calculate the inverse mills
ratio as a correction factor, and then incorporate that correction
factor into the ordinal probit model for the outcome you are modelling
(y_2).
For example,
probit y_1 xvars ...
capture drop hat
capture drop invmills
predict hat if e(sample), xb
gen invmills = exp(-.5*hat^2)/(sqrt(2*_pi)*normprob(hat))
oprobit y_2 xvars ... invmills
--
Kind regards,
Ian
-------------------------------
Ian Watson
Senior Researcher
acirrt, University of Sydney
NSW, 2006, Australia
