Rice Analytics

Automated Reduced Error Predictive Analytics

The Reduced Error Logistic Regression (RELR) algorithm has special advantages in segmentation applications where you may wish to have an accurate predictive model for different smaller segments. RELR works with very high dimensional problems involving many variables and interactions and produces very accurate interaction effects and tends not to produce spurious interaction effects.  An interaction effect implies that a regression effect is differentially related to different segments.  For example, if old people buy a product on warm days and young people buy it on cold days, then you have a differential regression effect in the young and old segments that is expressed as a significant AGE x TEMPERATURE effect in RELR results.  Other regression modeling algorithms, such as Stepwise Regression, have difficulty with interaction effects, as Stepwise may miss such an effect or return many such spurious interaction effects or return regression coefficients with opposite signs to the real causal effect which would not allow interaction effects to be interpreted. 

Because RELR is able to select such accurate interaction effects when they exist in its parsimonious Parsed variable selection, there will be no need to build separate predictive models for separate segments.  Instead, RELR's predictive segmentation will allow one to automatically segment using the regression results when such segments truly differ in regression effects. When segments do not differ, there will not be corresponding interaction effects in its parsimonious Parsed RELR variable selection.  Instead of building large numbers of models for separate artificially clustered segments, you can build one model with Parsed RELR and it will tell you in which segments different regression effects will occur.  You can read about the stability and the correct signs of RELR regression coefficients in effects that include interaction effects in our 2008 JSM paper that can be downloaded from the Papers and Presentations page of this website.

An advantage to Reduced Error Logistic Regression is that it works very well to generate stable regression coefficients with correct signs even with interaction effects, so you can use such interaction effects from the regression model for predictive segmentation where regression effects are different across segments. 

Machine Learning  Segmentation  Consumer Surveys  Predictive Modeling  Risk Management