Logistic Regression In Cartoons

Abstract

This paper contains discussions of issues associated with logistic regression where the issues are easier to understand when presented graphically than when they are presented using formulas. The topics discussed are: odds are not probabilities or percents; why use 1 / (1 + exp(-x)); complete and partial separation; why do we use the logit (What do the estimates in SAS printouts mean); the effect of beta values on the percent curve; oversampling – why and how; the ROC curve; the C statistic.