if you do the ROC on out of sample data, which of course isn't imputed,
then you can do this without any problem. you can also see whether
imputing in sample will help forecast.
if you're doing it all in sample, then i think the only thing you can do
is to do the ROC on the basis of the fully observed data on Y after the
imputation, and without it and compare.
Gary
On Tue, 13 May 2003, Ryan Thomas Moore wrote:
I think I understand what ROC curves indicate for binary choice models:
whether, for many values of C, one model always, sometimes, or never
predicts 0/1's correctly.
Now, can ROC curves compare an imputed dataset to a non-imputed one? The
idea would be this: for the same model, more information allows you to
better predict the 0/1's than less information. Does this make sense? Is
it trivial?
Thanks,
Ryan