A Direct Approach to Predicting Discretized Response in Target Marketing

Abstract

A problem that occurs frequently in direct marketing is the prediction of the value of a discretizing function of a response variable. For example, to target consumers for a coupon mailing, a retail chain may want to predict whether a prospective customer's grocery expenditures exceed a predetermined threshold. The current approach to this prediction problem is to model the response variable and then apply the discretizing function to the predicted value of the response variable. In contrast with this “indirect” approach, the authors propose a “direct” approach in which they model discretized values of the response variable. They show theoretically that the direct approach achieves better predictive performance than the commonly used indirect approach when the response model is misspecified and the sample size is large. These two conditions are commonly met in direct marketing situations. This result is counterintuitive because the direct approach entails “throwing away” information. However, although both the discretized response data and the continuous data provide biased predictions when a misspecified model is used, the lower information content of the discretized variable helps the bias be smaller. The authors demonstrate the performance of the proposed approach in a simulation experiment, a retail targeting application, and a customer satisfaction application. The key managerial implication of the result is that the current practice of restricting attention to models based on the indirect approach may be suboptimal. Target marketers should expand the set of candidate models to include models based on the proposed direct approach.