Your were right, there are some PSUs with only missing values for y. This is
due to the nature of the problem that -svyheckman- tries to address, i.e. we
do not observe y but for those whose selection equation: gamma*z + n >0.
However, the regression on the main equation (y = beta*x + lambda*mills + u)
should take only those values of y that are observed, hence facing the
single PSU problem. This is true when we follow a two-stage procedure.
However all svy estimators follow a--pseudo--ML method (except svyregress
and svyivreg); although I can see how -svyheckman- is taking a larger sample
than -svyreg- I cannot see how this estimates can be meaningful given the
single PSUs detected by the two-step procedure (or by svyreg y x).
