James G. Andress’s research while affiliated with Booz Allen Hamilton and other places

This article compares the predictive ability of models developed by two different statistical methods, tree analysis and regression analysis. Each was used in an exploratory study to develop a model to make predictions for a specific marketing situation. In situations where there are large sample sizes and where the data are subject to interaction, nonlinearities or causal priorities, there is reason to believe that tree analysis will be superior to regression analysis. This study provided support for the use of trees rather than regression for exploratory research in one such situation. Two predictive models were developed from a sample of 2,717 gas stations. Each used conventional decision rules - following what were thought to be the typical rules-of-thumb. Predictions were then made on the gasoline volume for 3,000 stations (virgin data). The predictions generated by the AID model were substantially more accurate than those generated by the regression model. We do not, by any means, regard the problem of trees vs. regression as a closed issue. More evidence is required from other situations to test out the generality of the hypothesis tested in this study. Our impression is that the ideal type of data analysis will employ both trees and regression concurrently.

Two models were developed to predict the average weekly volume for 3,000 gas stations. The model derived from a tree approach was found to be superior to that from a regression approach. The superiority of trees was significant in both a practical and a statistical sense.