An extension of regression trees to generate better predictive models

Abstract

For situations where the data are drawn from reasonably homogeneous populations, traditional methods such as multiple regression typically yield insightful analyses. For situations where the data are drawn from more heterogeneous populations, decision tree approaches, such as Classification and Regression Trees (CART) and Generalized, Unbiased, Interaction, Detection, and Estimation (GUIDE), are more likely to recognize idiosyncratic subpopulations and interactions automatically. In contrast to CART, however, GUIDE yields models with better predictive performance for each subpopulation. This article extends the idea of GUIDE to handle analysis of covariance-type problems. This article compares GUIDE modeling to various decision tree methods and to multiple regression. The article identifies and discusses the relative advantages and disadvantages of multiple regression, CART, and GUIDE. GUIDE produces quality or reliability models that exhibit greater predictive accuracy than multiple regression or CART for complex, highly diverse populations. Also, GUIDE is readily applicable to many other areas, such as repairability and maintainability settings involving both qualitative and quantitative variables. A small case study of an engineered wood product, medium-density fiberboard, is presented to illustrate the application of GUIDE. Accepted in 2005 for a special issue on Reliability co-edited by Hoang Pham, Rutgers University; Dong Ho Park, Hallym University, Korea; and Richard Cassady, University of Arkansas.