05-21-2015 11:46 AM
I just got the "MI of missing data with SAS" book last night. I ran code from section 6.3 (for arbitrary missing data, categorical > 2 groups, using FCS discrim).
Everything worked fine. I have the imputed data from all of the iterations,.., etc., but what do I use in the original dataset. Is there a way to kick out the original dataset with these data inserted into it?
05-21-2015 01:31 PM
I am now under the impression that you will not typically get a single imputed dataset (with the pooled values). Instead you get the n imputed datasets and a pulled estimate of those datasets. For my example, that was the percentages in each of the groups for the missing categorical variable. I think that is the answer to my question.
Other scenarios where the purpose may be inferential statistics, you would run the procedure with all of the datasets and get a pooled estimate for the statistic.
If there are any other possibilities people know of I would appreciate hearing their thoughts!
05-22-2015 08:38 AM
As you figured out, PROC MI creates stacked versions of the original data, where the various replicates are indicated by distinct values of the _IMPUTATION_ variable. You use the BY _IMPUTATION_ statement in your favorite procedure to analyze each imputed value, then use PROC MIANALYZE to aggregate the results. PROC MIANALYZE produces a point estimate for your statistic and a standard error that corrects for the added variance due to imputation.