your intuition is right. to be fully consistent, you'd need to impute for
each strata separately. But imputation doesn't need to nail every
obervation exactly, since you're only affecting the missing values and not
the observed values. probably you'll be ok if your quantity of interest
doesn't vary as a function of the missing data over strata, and in
practice this will matter if the relationship is strong. you could try
running amelia within each strata and combining them to see, if you have
enough observations within each.
Gary
On Tue, 3 Aug 2004, Matthew Allan Vile wrote:
I know that multiple imputation requires you to
include at least every
variable in your final estimation model (including interactions) and
variables not included but later added to the dataset will have their
"relationship with other variables in the dataset biased (normally
towards zero)."
However, what happens when you include a variable in the dataset during
imputation and then use that variable to stratify the regressions (as
with the "by variname:" option in Stata)? Effectively, this
stratification mimics an interaction by allowing coefficient values to
vary across regressions within the sample; are the results of this kind
of procedure still BLUE (providing other regression presumptions are
met).
Matthew Vile
Senior Data Analyst, GII