Software

Reduced Error Logistic Regression (RELR) has been featured over the past few years through invited and contributed sessions at major conferences such as M2007, JSM 2008, and MWSUG 2008. RELR represents a new generation of reduced error predictive modeling software that successfully models and subtracts error out of the predictive model. RELR's non-arbitrary error model is based upon the Extreme Value Error properties of logistic regression; it yields strikingly reduced error in the logit coefficients. RELR's ability to model and subtract error is in stark contrast to approaches such as Penalized Logistic Regression that deal with error by arbitrarily smoothing or shrinking regression coefficients.

In direct comparisons to standard predictive modeling approaches that include Penalized Logistic Regression (PLR), RELR models have had significantly reduced average squared error in the validation sample. The fact that RELR's validation sample average squared error has been so much lower than Penalized Logistic Regression's (p<.0001) at a sample size of 1000 training observations is especially remarkable because RELR does not use the validation sample to determine its error model, whereas PLR's smoothing model uses the validation sample data to arrive at its estimates.

Given a large number of multicollinear variables, an often used rule of thumb is that Penalized Logistic Regression requires a sample size of 75,000 observations for a model to have predictive validity. RELR has been able to achieve such predictive validity with a sample size that is significantly less than 1000 observations in some instances.

Reduced error usually translates directly into substantial ROI. For example, large savings in primary and secondary data costs can be attained through RELR, much better fit can be obtained in risk models through RELR, and reduced error in RELR's variable importance measure means that its Parsed^TM variable selection gives simple and accurate models with much less likelihood of costly spurious variables. Case studies can be found at www.RiceAnalytics.com. Potential customers may also enroll there through January 31, 2009 for a Free ROI Evaluation for their application.

RELR is available as a set of stored SAS Macros with various scalability and Graphical User Interface options that include the ability to snap into SAS Enterprise Miner as an extension node.

SAS and Enterprise Miner are trademarks of SAS Institute. Parsed^TM is a trademark of Rice Analytics. Rice Analytics is a member of the SAS Alliance Partnership.