whatif shouldn't take very long with 4k obs.
the usually multiple imputation story is that you compute your quantity of
interest, whatever it is (including in or out of the convex hull), and you
combine the inferences from the separate sets. you might not wind up with
a 0/1 imputation, but that's still a reasonable approach I think. on the
other hand, I don't know any research at all on this subject, so maybe you
can figure out and propose a good solution!
Gary
On Thu, 4 May 2006, Justin Ryan Grimmer wrote:
Hey Gov2001,
Does anyone get a sense of what the relevant convex hull is when we impute
missing data? Is the relevant convex hull the convex hall that arises from
the entire set of imputed data, or is the crucial test if the counterfactual
lies within the convex hull of each the imputed data sets? What if a
counterfactual lies outside of the convex hull in some data sets (or is
extreme by other measures) but is more reasonable in other imputed data sets?
Obviously this latter question will be more likely if there is a great deal
of variation between data sets.
Also, how long should the whatif command take on a data set of about 4000
observations?
Thanks and have a good night
Justin