05-26-2015 04:04 PM
So, do I have to worry about the "Quasi-Complete separation" message when running a PROC LOGISTIC, if I just care to know which variables are significant? I don't care about their estimates. Also, fyi, all my predictor variables are categorical and and are collapsed as much as it makes sense for them to be.
05-27-2015 06:47 AM
It depends on where the separation is. When there is quasi-complete separation, usually one variable will have very odd confidence intervals. Those shouldn't be used. But the other ones can. See this paper by Paul Allison