Modeling Preference and Structural Heterogeneity in Consumer Choice

Consumer heterogeneity is fundamental to the marketing concept, providing the basis for market segmentation, targeting and positioning, as well as micro-marketing. Substantial effort has already been devoted to incorporate heterogeneity in brand choice models. However, most of the research in this area has focused on differences in preferences or tastes across consumers. In contrast, limited attention has been given to the possibility that consumers might also differ in the process they follow when making choices. Failure to account for either form of consumer heterogeneity may lead to misinterpretations of market structure and market segments, as demonstrated in our simulations. Our main research objective in this paper is to account for these two forms of consumer heterogeneity with an integrated model. We develop a choice model that simultaneously identities consumer segments on the basis of their preferences, response to the marketing mix, and choice processes. This choice model is a finite mixture of nested logit models that incorporates the mixture of multinomial logits as a special case. One main limitation of the now popular multinomial logit model is that choices by each consumer are assumed to be Independent from Irrelevant Alternatives. As a consequence, that model predicts that, within each segment, a brand has exactly the same cross-elasticities upon every competitor. In contrast, our mixture of nested logits allows for violations of the IIA property within each segment, leading to more realistic patterns of brand competition, where one brand draws differently from each competitor, depending on the choice process used by members of the segment. In addition to the flexible combination of choice processes and preference structures, our model also incorporates a latent class of “hard-core loyal” consumers who are expected to always buy the same brand, and are thus insensitive to price and promotions. Application of our model to household scanner data in the peanut butter category led to four types of segments: (a) hard-core loyals, who do not respond to price and promotions. (b) brand-type, for which the brand choice precedes the choice of product form (creamy vs. crunchy). (c) form-type. Members of these segments first choose the product form and then decide for the brand (Peter Pan, Jif, Skippy, or Store brand). (d) IIA-type. Members of these segments make choices according to the IIA property. Our results showed a large (14%) hard-core loyal segment. We also identified three segments of the brand-type, one segment of the form-type and three IIA-type segments. All these segments also differ in their preferences for brands and product forms. One useful feature of our model is that it allows for cross-elasticity structures with non-proportional draw, even when computed within a homogeneous segment. Most importantly, the proposed model allows for more complex cross-elasticity structures within a segment, rather than constraining the elasticities to any particular pattern a priori. These cross-elasticity patterns imply distinct competitive environments within each segment, leading to different price and promotion strategies in each segment. For example, a promotion by Jif-crunchy to a brand-type segment is more likely to cannibalize on shares of the creamy form of the same brand, than to draw from competitors. If offered to a form-type segment, the same promotion is more likely to draw shares from competitors. Because of its finite-mixture formulation, our proposed model shares the same limitations of this category of models. For example, maximum likelihood estimation of the model may lead to local optima, thus requiring a multiple number of random starts, in order to increase the chances of finding the model that best fits to the data. Our model also requires specification of the number and type of segments in each particular application. Several criteria are available to select the number of segments. However, the specification of choice processes must be based on a heuristic procedure, or on the basis of prior knowledge.