RE: st: multimple imputation in steps??

["Ploutz-Snyder, Robert (JSC-SK)[USRA]" <robert.ploutz-snyder-1@nasa.gov>]

Thanks Martin,

I've come across several references to ice, including references to it in the Stata im manual itself.  I assumed that Stata's im programming incorporated the essence of ice, but I don't know that for sure.

Can someone give me the run-down on how ice differs from mi??

r

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Maarten buis
Sent: Wednesday, April 07, 2010 3:07 PM
To: statalist@hsphsun2.harvard.edu
Subject: Re: st: multimple imputation in steps??

--- On Wed, 7/4/10, Ploutz-Snyder, Robert (JSC-SK)[USRA] wrote:
> Consider a dataset where x1-x5 are separate but correlated
> variables that measure "factor1"  and further that
> X6-x10 all tap into "factor2."  
> 
> Using mi impute in a very general sense, I could simply
> impute all missing items in one step by using
> 
> mi impute x1-x10
> 
> But that ignores the shared variance structure of the
> factors (nothing on the right side of the impute equation).
> 
> Instead I would like to generate factor scores prior to
> imputations, which will result in factor1 and factor2 scores
> for subjects who are missing no data.  Then the second
> step would be to use mi impute to impute the FACTOR scores
> from the items loading on the factors, so that the factor
> loading structure is preserved in the imputed data. 
> This assumes that the factor structure for subjects with any
> missing data match the factor structure for complete
> subjects, but I'm willing to assume that.
> 
> So following the factor analysis and predict statements, I
> could use
> 
> mi impute mvn factor1=x1-x5
> 
> ...and then separately 
> 
> mi impute mvn factor2=x6-x10
> 
> But this results in different samples in the m>0
> imputations, and I can't use mi estimate commands.  
> 
> 
> What is the solution here?  How can we use mi impute
> that preserve the underlying covariance structure among the
> data, i.e. allows different modeling of missing data based
> on what we know about the data a-priori, but then also use
> all of the m>0 imputations in estimation commands
> following imputation?

I would use -ice- to simultaneously impute x1-x10 _without_ 
the factor variables. This will preserve the correlation
structure, and you can use -factor- afterwards to estimate
your factor variables, while including a maximum of the 
observed information.

Hope this helps,
Maarten 

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


      

*
*   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/

*
*   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/