🚀 go-pugleaf

I echo what Dick Darlington and others said.  The only time collinearity is a
DEFINITE problem is when

1. you are predicting on new data
       and
2. the new data has different amount or type of collinearity from the
   training data

or

3. if you believe individual statistics on correlated variables (i.e.,
   if you don't do a pooled test)

Davis et al. (Modern Statistical Methods in Chronic Disease Epidemiology,
Wiley, New York, 1986, eds. Moolgavkar and Prentice) discuss how to get
collinearity statistics from the covariance matrix of the parameter
estimates in logistic models.

Multicollinearity approaches the SAS Type II, III, IV sums of squares in
terms of time spent by statisticians on overrated problems.




Frank Harrell
Division of Biometry
Duke University Medical Center
feh@biostat.mc.duke.edu