# How to deal with separation in logistic regression? [duplicate]

I'm running a binary logistic regression on 15 independent variables for 180 observations in Stata (version 11). This I do for four different groups, i.e. four dependent variables. For three, it works fine, but the fourth gives me the warning message that one IV "predicts success perfectly" and that the IV is "dropped and 96 obs not used". This IV is a dummy (and considered important for this one out of the four groups). This leaves me with 84 observations. In comparison with the other three estimations, the SE seem to be a bit inflated, but p-values are fine (if I drop the problematic IV, the SE go down, but the p-values go up). The LR chi squared shows significance at the 10% level and the pseudo R squared is 0.191.

I suppose this is a case of (quasi-)complete separation. My questions are

1. if I can still use the regression estimates after such a warning, because Stata dropped the problematic cases (=all of the 84 the observations, for which the dummy exists or is coded 1).
2. or if I should drop the problematic IV (which would be unfortunate, even though I would have a the original sample size of 180 then)
3. and if neither 1. or 2., how I can deal with this problem.

I know that 180 observations are not much for a model with 15 IVs. However, I'm hoping for some advice on how to solve the problem other than reducing an overfitted model - especially specific Stata knowledge would be of great help.