Weight of Evidence Coded Variables for Binary and Ordinal Logistic Regression - Bruce Lund, Magnify Analytic Solutions
Weight-of-evidence (WOE) coding of nominal or discrete variable is widely used when preparing predictors for usage in binary logistic regression models. This practice is especially prevalent in credit risk modeling. When using WOE coding an important preliminary step is binning (or coarse classing) of the levels of the predictor to achieve parsimony without giving up predictive power. One approach to binning is to select levels for collapsing so as to minimize the loss of information value (IV) at each step. WOE and binning will be discussed for binary logistic models. Next, the concepts of WOE, binning, and IV are extended to ordinal logistic regression. Lastly, (for binary models) guidelines for assignment of degrees of freedom for WOE-coded predictors within a fitted logistic model are discussed. The assignment of degrees of freedom bears on the ranking of logistic models by use of SBC or AIC. All computations in this talk are performed by using SAS® and SAS/STAT®. SAS programs related to the talk will be provided by the presenter.
About the Presenter
Bruce Lund has used SAS for over 30 years. For the past 15 years, he has worked at Magnify Analytic Solutions of Detroit, Wilmington, and Charlotte. Before Magnify, he was the customer database manager at Ford Motor and a mathematics professor at University of New Brunswick, Canada. He has a mathematics PhD from Stanford University.