Dan and Sarah:
I have dealt with a similar problem only in my case
individuals were nested within surveys. I ran -ice-
separately for each survey and combined the imputed
files afterwards. Because I wanted to include
interaction with each annual dummy and gender, I
actually rand -ice- separately for each survey,
year, gender combination.
One thing to keep in mind is to make sure you have more
completely observed cases than variables in your imputation
model for each call to -ice-. -ice- will not complain if you
don't, but it won't converge, and it won't tell you that it
hasn't converged, since convergence is a tricky issue with
-ice- (you can run -ice- with the trace option and lots of
cycles and look at a lineplot of mean imputed value versus
cycle, and the result should show no trend if it converged).
In my case, before I found out what was going on, the results
were really of. It took me a while to find out what was going
on, especially since I first estimated the model using
interactions with all the dummies. Not enough observations
isn't the first thing you think about if you have a 100,000+
observations...
HTH,
Maarten
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting adress:
Buitenveldertselaan 3 (Metropolitan), room Z434
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
--- Sarah A. Mustillo wrote:
> I was waiting for someone else to reply, but haven't seen any cross
> the list. When I use ICE, I impute within each subject. I have no
> citations on this or anything to support that approach, it just
> intuitively seemed to me that it could be problematic to impute across
> the entire dataset when the obs are correlated. So, I reshape the data
> to wide and impute as one observation per person. Whether this actually
> makes a difference is probably questionable. At some point I'll try it
> both ways and compare...
Dan MacNulty wrote:
> I have two questions with respect to imputation and the program ICE. In
> situations where data are subject-specific, e.g. clustered within
> patient_ID, is it recommended that one impute within each subject or is
> it sufficient to impute across the entire dataset? If the former, can
> ICE be implemented to impute missing values within each subject, and if
> so, how?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
