Hi, I'm a n00b to the forum so went ahead and mined the forum and scoured the web.
I ran a large scale survey with ~50 survey, attitudinal items. Participants in the survey were given the option the respond "Don't know" to the survey rather than forced response. I'd like to test some assumptions on whether replacing missing data helps or hurts the analysis, i.e., mean, median imputation are fine but not ideal when there may be larger patterns on a respondent or survey item level. All the items are coded on 5 point agreement scales with the Don't know option that has been recoded as missing. There are underlying constructs with about 4-5 items per construct. Any help on steps to follow after imputation has been complete would be helpful. There is a package in R called "missRanger" that the data scientists are using but wanted to explore a SAS version of multiple imputation using PROC MI or another solution.
(What I'm getting stuck on is what to do after PROC MI has created N datasets, i.e., do I analyze each imputed dataset on its own? Or can I get estimates averaged from all N imputed datasets?).
Thanks, Greg
