st: sample selection and ordered probit estimation with opsel. Results depend on using robust or otherwise.

[Massimiliano Volpi <maxvolpi@hotmail.com>]

Dear Dr. De Luca and Stata-listers,

I am using opsel, a command that Giuseppe De Luca (copied in) wrote 'Estimation of ordered response model with sample selection'; The Stata Journal (2011) vol. 11(2) pp 213-239.

I am working on CIS (Community Innovation Survey) data and I tried to estimate my model both with and without the options "robust" and "weight" (I use analytical weights, as the data are from a survey and each observation represents a certain number of companies).

However, the test on the presence of sample selection changes dramatically, depending on whether I specify the option robust or not.
Without the option robust, I can reject the independence of the selection (probit) and the outcome (ordered probit) equation. So, I have evidence that sample selection is indeed at work.
But, if I specify the "robust" option, then I can almost invariably reject the presence of sample selection.

Adding weights on its own (i.e. not combined with the option "robust") makes little difference. But, again, 'weight' and 'robust' would imply lack of sample selection.

Could you or anyone on the list please clarify what the option 'robust' does in this setting? And, more generally, whether it is reasonable to use it when working with a sample selection model; I am assuming it should just make my model more robust to heteroskedasticity, but potentialliy hiding sample selection issue seems far more serious than loosing efficiency...)



Best regards

Max Volpi
Natural Environment Research Council
 		 	   		  
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/