Screening, Transforming, and Fitting Predictors for the Cumulative Logit Model - Bruce Lund, Magnify Analytics

The cumulative logit model is a logistic regression model where the target has 2 or more ordered levels. If only 2 levels, then the cumulative logit is the binary logistic model. Predictors for the cumulative logit model might be “NOD” (nominal, ordinal, discrete) where typically the number of levels is under 20. Alternatively, predictors might be “continuous” where the predictor is numeric and has many levels. This talk discusses methods that pre-screen and transform both NOD and continuous predictors before the stage of model fitting. Once a collection of predictors has been screened and transformed, the talk discusses predictor variable selection for model fitting. One focus of this talk is determining when a predictor should be allowed to have unequalslopes. If unequalslopes are allowed, then the predictor has J‑1 distinct slopes corresponding to the J values of the target variable. SAS® macros are presented which implement the screening and transforming methods. Because this talk is 20 minutes, some topics will be summarized and reference papers will be given. Familiarity with PROC LOGISTIC is assumed.

About the Presenter

Bruce Lund is a data miner and predictive modeler with 30 years of experience in direct marketing in the automotive industry including the development of in-market models, long-term value models, and loyalty/defector models for automotive clients. For the past 15 years, Bruce has been a consultant for Magnify Analytic Solutions, a division of Marketing Associates of Detroit MI, Wilmington DE, and Charlotte NC. Before Magnify, he was customer database manager at Ford Motor Company and a mathematics professor at the University of New Brunswick in Canada. He has a PhD in mathematics from Stanford University. During the past years he has been a frequent contributor to SAS user groups.