firstname.lastname@example.org helpfully replied:
>Thanks for chiming in David and Paul, Howard, and Ya for your replies.
Always happy to cause problems and sow the seeds of destruction.
Think of me as your own personal Eris. :-) :-)
>I want to see how imputation works in this case with a high missing
>rates. The original dataset is much larger and the missing pattern is
>actually got from that dataset ( I am aware of the non-monotone
>missing pattern) . I want to use the "complete" data from the same
>large dataset but drop some values to deliberately reproduce the
>missing pattern and see how imputation would work.
>I know that these complete cases may not be comparable to those with
>actual missing values. But I think it is still a way that can show the
>good and bad of different imputation methods.
Okay, now this is a good idea. I like it. Demonstrating the effects of
imputation methods using a real data set and fixed missing value patterns
is nice. (Although you'll have to caveat that the pattern of missings may
affect the results, particularly when you have so many missings which
fall in a monotone pattern.)
>Would someone provide more comments or suggest a better approach?
Not me. I like the approach so far.
I hope you're going to contrast single and multiple imputation, as well as
comparing multiple imputation methods. Since you have so many
monotone-pattern records, you have the option of letting PROC MI
fill in the monotone missings before filling in the non-monotones through
David L. Cassell
3115 NW Norwood Pl.
Corvallis OR 97330
Express yourself instantly with MSN Messenger! Download today - it's FREE!