Consolidating multiply imputed data into a single dataset

[amsavage]

If one uses multiple imputation in STATA to derive 5 complete datasets, could the newly imputed values simply be averaged to create a substitute value? EG. 'mi impute mvn' is used to generate 5 possible values for one missing value. Can I average those 5 values and use the result by itself in further analysis?

I would like to use MI in STATA to create a complete dataset that I can then export for use in other software (exporting to SPSS; generating a covariance matrix + summery stats for use in AMOS). I am very restricted as to which software I can use for what purpose as I'm working with restricted data in a secured lab. I don't intend to average any test statistics, I realize that would be problematic, I just want to create a single complete dataset. Thank you so very much for your assistance!!!

 

[spunky]

Can I average those 5 values and use the result by itself in further analysis?

in general, 'no'. if you merely average the data sets and then use the averaged data sets for further analyses, you're assuming there is no uncertainty related to the missingness of the data... which means your standard errors will be wrong (too small) and your type 1 error would be too high. you need to use Rubin's rules for missing data to accurately average your estimates and obtain the proper standard errors.

you're using STATA? ditch AMOS (which is a horrible little piece of software anyway) and use the GLLAM package to fit your SEM needs:

http://www.gllamm.org/

in general, using Rubin's rules to combine multiply-imputed datasets (particularly the variance ones) is not easy unless you know what you're doing.