When an error in a selection equation is correlated with regression equations, then these correlation measures have to be corrected in the likelihood function, otherwise the results will be biased.

      Here are a couple of procedures one can use.
1) nlmixed
2) qlim

      here is a link for the further reference ( first 2 pages)
http://siteresources.worldbank.org/DEC/Resources/movestay.pdf