Data Mining with Stepwise Regression
Author:
Robert Stine
Department of Statistics                                         
Wharton School, University of Pennsylvania    

Abstract:

This talk will discuss using stepwise regression to fit models to very 
large datasets.  The goal of this talk is to show that with several 
modifications, stepwise regression becomes an easy and effective tool 
in data mining large datasets.

The standard version of stepwise regression has several pitfalls that
 one must avoid.  Three modifications to standard regression fix the 
most serious aspects of these problems: (1) use interactions to capture 
non-linearities and cope with missing data, (2) use Bonferroni-type 
procedures to select which variables to include in the model, and 
(3) use an adjusted critical value to accommodate
high-leverage points.  The talk will describe each of these three
 modifications and explain why each is necessary. 
 With all three of these modifications in place, stepwise regression 
becomes a data-mining procedure well-suited to almost any data set.