[EMAIL PROTECTED] writes: >Like Rod, I cannot understand why NSIs will not, shall not, or cannot use >MI when providing public use data sets.
I have to share one story here. Years ago I was taking one of Don Rubin's workshops. At the time I prepared a sample data set which, while not public, was used by several government agencies. As the very issue of using MI was being discussed I raised my hand and went into how I had asked several of the users if they wanted me to use MI in the imputations and all of them had said they would not use it, and it would only confuse them. Don was puzzled as to the why, and urged me to try to continue to educate my clients. At that moment I heard the person behind me coughing. I turned around and it was precisely one of the government analysts who used my database. I had not even been aware that she was taking the workshop. I profusely apologized lest she have taken anything I had said personally, and then asked her if now that she had herself been exposed to MI would she use it if I provided it. She said, "Definitely NOT!". The reality was that the people using the data set were economists charged with coming up with a best estimate, regardless of how precise. They were not interested in variances, but had economic models to be run. If the best estimate was poor, that was too bad, it still had to be used. Having one official imputation provided by me made their job easier. I mention this story, because whereas a multiply imputed data can be helpful to one set of users, they can be at times useless to some and confusing to others. I suspect some agencies are driven by the users they have most contact with.