Treating Zero Brand Sales Observations in Choice Model Estimation: Consequences and Potential Remedies
Published 10/1/2008
Author: Richard A. Briesch, William R. Dillon, and Robert C. Blattberg
View this content
Executive Summary
The estimation of consumer demand models to assess the impact of price and promotion on the basis of store and/or household (HH) panel data continues to be of interest to both academics and practitioners. Even a casual review of the extant literature in this area shows that most of the effort has focused on developing estimation techniques capable of accommodating specific effects—for example, heterogeneity and endogeneity. Recently, however, several studies have attempted to shift attention to other putative problems related to data preparation decisions and their impact on parameter bias.
This research contributes to this recent literature by focusing on the consequences of and potential remedies for incorporating zero brand sales observations when calibrating choice models on the basis of HH panel data. The authors show that including or excluding zero brand sales observations can bias price elasticities, even in cases in which the data are generated by a multinomial logit model. To support this finding, the authors provide analytical, simulation, and empirical evidence to suggest that zero brand sales observations can lead to substantial bias when the model does not account for zero brand sales. Specifically, this research demonstrates that (1) zero brand sales are pervasive—that is, they are more the rule than the exception when analyzing HH panel data; (2) price coefficients will be biased upward or downward depending on how the zero brand sales observations are treated (included or excluded); (3) it is always sage to remove structural zeros (i.e., observations in which the brand was not in distribution) before estimating choice model parameters; (4) a relatively straightforward investigation of the incidence of “missingness” by stockkeeping unit (SKU) and brand as well as SKU/brand promotional activity may suggest the extent of the problem; and (5) the problems related to the presence of nonstructural zero brand sales do not necessarily go away if SKU-/day-level data are rolled up to the brand-/week-level, especially for smaller-volume share brands.
To assist choice modelers, this research develops two alternative model forms—selection and mixture—that can better handle zero brand sales observations at both the SKU and the brand levels. Both model forms perform better than the multinomial logit model. In general, the selection model is preferable in two respects: First, this model form provides the best in-sample and out-of-sample fits and hit rates. Second, this model form does not suffer from the estimation challenges that are likely to plague the pattern-mixture model.
Biography
Richard A. Briesch is Associate Professor of Marketing in the Cox School of Business at Southern Methodist University. He received his PhD in Marketing from Northwestern University. His primary area of research is the general area of modeling consumer decision making. His articles have appeared in Journal of the American Statistical Association, Journal of Consumer Research, Marketing Science, Journal of Retailing, and other leading academic journals. He won the William R. Davidson Award for the best paper in Journal of Retailing. Over the past ten years, Professor Briesch has also consulted for several national and multinational firms.
William R. Dillon is Herman W. Lay Professor of Marketing and Professor of Statistics in the Cox School of Business at Southern Methodist University, where he also serves as Senior Associate Dean. He received his PhD in Marketing and Quantitative Methods from City University of New York. He has published more than 40 articles in the general areas of segmentation, positioning, market structure, and issues related to the use of latent class/mixture models and covariance structure models. His articles have appeared in Journal of Marketing Research, Journal of the American Statistical Association, Marketing Science, Management Science, and Journal of Marketing. He has also published four books, two of which have appeared in the prestigious Wiley series on probability and statistics. He currently serves on the editorial review boards of Journal of Marketing Research and Marketing Letters. Over the past 20 years, Dillon has consulted for several national and multinational firms. In 1994, he cofounded Marketing and Planning Systems, a Boston-based marketing research consulting firm.
Robert C. Blattberg is Polk Brothers Distinguished Professor of Retailing in the Kellogg Graduate School of Management at Northwestern University. He directs the Center for Retail Management, where he also serves as the Chief Analytical Officer of Information Resources. His primary research is in the areas of database marketing, sales promotions, pricing, and retailing. His articles have appeared in the Journal of Marketing Research, Management Science, Marketing Science, Econometrica, Journal of Marketing, Journal of Direct Marketing, and other leading academic journals. He has coauthored Sales Promotions (Prentice Hall) and Customer Equity (Harvard Business Press). Professor Blattberg has consulted for a wide variety of firms, including American Express, Kroger, Best Buy, Rite Aid, IRI, and A.T. Kearney. He won both the John D.C. Little Award for best paper in Marketing Science and the Robert B. Clarke Award from the Direct Marketing Educational Foundation as Educator of the Year.
J Marketing Research, Volume 45, Number 5, October 2008
View Table of Contents.