Hello Statalisters.
I am using probit with selection (heckprob) for the first time. I'm having
a
lot of fun (as usual when working with Stata!), but I'm concerned that I
am
misinterpreting my findings and/or using the technique incorrectly. So I
have a couple of quick (well, OK maybe not so quick) questions. I would
greatly appreciate any help.
I am working with data from a survey on which we unfortunately were able
to
achieve only an 18.2% response rate for the 2 outcomes of interest in our
current analysis. (This was not unexpected for our population--low
response
rates have become the norm for this group, although the 18% was
disappointing.) We have a total of 6412 observations, with 5246 censored
and
1166 uncensored. The selection model does contain one variable (gender)
which is not included in the outcome model (not significant in outcome
model). Rho is not significant. Wald chisq(18)=95.05, p =0.0000.
My first question is this: is there a minimum response rate or proportion
of
uncensored observations under which Heckman or similar types of methods
are
not appropriate? And if so, what would that be?
I am quite confused about using and interpreting interaction effects in
heckprob. There are several significant interactions in my outcome model.
One of these interactions is significant in predicting selection as well.
It
seems sort of incorrect to have the same interaction term in both
equations
although I can't figure out why I think that. If an interaction term is
significant in the selection equation alone and in the outcome equation
alone, is it more appropriate to include it in the selection equation when
running heckprob?
If it would help I could reproduce the output here.
Thank you so much...
Margaret Tyler
University of Iowa
*
* 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/
