Measuring Cross-Category Price Effects with Aggregate Store Data

• Authors: Inseong Song;Pradeep K. Chintagunta
• Venue: Management Science
• Year: 2006

Our objective is to understand the cross-category effects of marketing activities using aggregate store-level scanner data. For this, we provide a framework derived from household utility maximizing behavior which assumes that a household chooses the “bundle” of products with the highest utility. We use a second-order Taylor series approximation to an arbitrary utility function to represent bundle utility. Aggregate sales or shares in each category are derived under the assumption that households are heterogeneous in their preferences and in their sensitivities to marketing activities. Our estimation accounts for potential price endogeneity in demand. Using store-level scanner data on four product categories---liquid laundry detergents, powdered laundry detergents, liquid fabric softeners, and sheet fabric softeners---we find evidence for a complementary relationship between liquid softeners and both forms of detergents. We also find that the magnitude of cross-category elasticities are brand specific, i.e., different brands in a category have a different price impact on the demand for a brand in another category. The results have implications for retailers in terms of the potential need for cross-category management, as well as for manufacturers such as Procter & Gamble that participate in all four categories. We compare our model with a log-log regression specification on three criteria---estimated elasticities, hold-out sample predictions, and retailer cross-category pricing. We find that the proposed model produces more reasonable estimates relative to the log-log model; it predicts better and is more useful for pricing purposes. Further, in a simulation study, we show that our proposed model can recover the elasticities from a data-generating process that simulates household-level joint outcomes across categories even after these data have been aggregated to brand-level shares within each category. By contrast, the log-log regression model is unable to do so.