Treating zero brand sales observations in choice model estimation: Consequences and potential remedies

Abstract

This research focuses on the consequences of and potential remedies for incorporating zero brand sales observations when calibrating choice models on the basis of household panel data. As the authors show both analytically and empirically, 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. The authors show that an appropriate choice of a model form (i.e., specification) is influenced by the incidence of "missingness" and the mechanism presumed to be causing a brand to have no purchases in a given store/week. The authors propose two model forms that may perform well in the presence of zero brand sales observations and investigate their relative performance on the basis of simulated data and two household panel data sets. © 2008, American Marketing Association.