I'm glad to see there is a Stata command that does what Clyde wants.
But, I still wonder about his last concern: "The most cogent
objection I can think of to this proposal is that perhaps I shouldn't
be using MI at all in these situations, or, perhaps, that my
imputation models need some serious re-thinking when the imputations
show this kind of irregularity." If you are getting imputations that
give bizarre results is it ok to just drop them?
At 11:38 AM 4/9/2013, you wrote:
I'm pretty sure the answer to this question is no: I've scoured the
MI manual and can't find it.
Let me explain what I'm looking for and why I think such a command
would be useful.
Working with a data set that has a great deal of missing
information, I sometimes end up over-generating imputations. That
is, my initial estimate of the number of imputations I will need
ends up too high. It would be nice to be able to tell Stata to
simply drop some of them, rather than having to start over and
generate a smaller number. You might ask what the harm is in
keeping them all. Well, the more imputations you have, the longer
each MI analysis takes, and if there are many analyses, the time
starts to add up.
There is another reason I would like to see a command like
this. When the data include difficult patterns of missingness, such
as pairs of variables V1 and V2 that are never both non-missing in
the same observation (so their imputations are entirely dependent on
yet other variables and there is no direct information about their
associations), my experience has been that the imputed data sets
sometimes contain imputed values that are widely outside the range
of the observed data, even by a few orders of magnitude. MI
analyses that include these data sets sometimes produce results that
are clearly absurd.
And there is a third situation. When doing chained imputations,
there is the issue of the burn-in. Again, we have to make an
initial guess and see what happens. Sometimes, I will generate a
large number of imputations using a long burn-in because earlier
attempts with a shorter burn-in don't converge to stable
estimates. Maybe I have 75 imputations after a burn-in of 30. If I
study the convergence carefully, I may find that there are 3 or 4
imputations where convergence has not been achieved, but the rest
look OK. It would be nice to be able to tell Stata to drop those 3
or 4--after all, 75 is just a round number and in most circumstances
71 or 72 imputations will serve nearly as well.
While the -mi estimate- command itself permits specification of a
numlist with the numbers of the imputations I want to use, having to
repeatedly specify this list is a nuisance. It would be simpler to
be able to drop those imputations, re-save the data set, and then
move on to analysis.
The most cogent objection I can think of to this proposal is that
perhaps I shouldn't be using MI at all in these situations, or,
perhaps, that my imputation models need some serious re-thinking
when the imputations show this kind of irregularity.