Modeling Preferences for Common Attributes inMulti-Category Brand Choice∗Vishal P. Singh†, Karsten Hansen‡, Sachin Gupta§Decmber 19, 2003AbstractCharacteristics based discrete choice models of demand have been used ex-tensively in both economics and marketing. The basic endeavor in these modelsis to view products as bundles of characteristics, with consumer preferences de-fined over this characteristics space. In the context of brand choice in packagedgoods categories, Fader and Hardie (1996) show the advantages of this approachin terms of parsimony as well as model fit. More importantly, this modelingapproach lends itself to important applications such as predicting demand fornew products. In this paper, we propose a multi-category brand choice modelthat is based upon the conceptualization that the intrinsic utility for a brand isa function of underlying attributes or characteristics, some of which are commonacross categories. Our general premise is that preferences for attributes that arecommon across categories are likely to be correlated. Further, we project the un-observed component of preferences for attributes and sensitivities to marketingmix variables to a lower dimensional space of unobserved factors. The factorsare interpretable as unobservable household “traits” that explain similarity inchoice behaviors across categories. Since the traits transcend categories, we canuse household specific factor estimates derived from purchasing in existing cat-egories to predict preferences for attributes in new categories. The proposed∗The authors are listed in reverse alphabetic order and contributed equally. They acknowledgehelpful comments of seminar participants at Northwestern, Stanford, Cornell, Tilburg, and the Uni-versity of Washington. They thank the following for generously sharing the data used in this study:ACNielsen, Kraft Foods, Information Resources, and David Bell.†[email protected], GSIA, Carnegie Mellon University.‡[email protected], Kellogg School of Management, Northwestern University.§[email protected], Johnson Graduate School of Management, Cornell University.1model is applied to household panel data for three closely related snack cate-gories, and for two less related food categories. We find strong correlations inpreferences for product attributes such as brand names and low-fat or fat-free.In two cross-category targeting applications we demonstrate that these high cor-relations in product attribute preferences across categories imply that (1) onecan use the model estimates to improve forecasts of preferences for an attributein a new category, and (2) that one can score potential targets for a new productin an existing category based on prospects’ probability of choice.1 IntroductionA vast literature in marketing studies brand choice behavior within individual prod-uct categories. A consistent empirical finding across a number of categories is thatconsumer heterogeneity in brand preferences and responsiveness to marketing mix vari-ables explains a substantial part of the variation in brand choices of households. Inother words, consumers are very different from one another within each category. Animportant question that has intrigued marketing researchers is whether a household ex-hibits similarities in it’s choice behavior across seemingly disparate categories. Stateddifferently, are consumers’ buying behavior and sensitivities to marketing mix variablesdetermined primarily by household-specific factors or primarily by category-specificcharacteristics? While this question has long been of interest (see for example earlywork by Blattberg, Peacock, and Sen 1976), appropriate methodologies to address theissue have only recently been developed (Ainslie and Rossi 1998). Using data fromfive product categories, Ainslie and Rossi find substantial correlations in price, displayand feature sensitivity of households. Similarly, Erdem (1998) and Erdem and Winer(1999) find that consumers’ preferences for a brand name are correlated across cat-egories. These findings have sparked interest in the development of multi-categorychoice models, i.e., models in which consumer preferences for brands and their respon-siveness to marketing activities in each category have a joint distribution that allowscorrelatedness across categories.The problem of finding and explaining generalities in consumers’ buying propensi-ties across product categories is of intrinsic academic interest. For example, discoveryof empirical patterns can trigger development of new theories of consumer behavior. Atthe same time it has important implications for marketing practice. Consumer pack-aged goods processing is a highly concentrated business, with only a few companiesaccounting for a very large share of the overall global market (Rogers 2001). Each of2the major manufacturers, such as Procter and Gamble, Unilever, General Mills, andKraft, owns brands in a vast number of product categories. The downstream customersof these manufacturers are large supermarket chains, who are multi-category firms aswell. For these businesses, developing an understanding of consumers’ preferencesor traits that transcends product categories can be a source of strategic advantage.Potential application areas include umbrella branding and advertising (Erdem 1998,Erdem and Sun 2002), brand equity and its extendability (Park and Srinivasan 1994),cross-category promotions (Chintagunta and Haldar 1999), new product targeting (thispaper), and so forth.Our focus in this paper is on the use of observed household purchase data to an-alyze preferences in multiple categories. In packaged goods markets, availability ofgood household panel data from marketing research providers such as ACNielsen andInformation Resources Inc., or from retailers’ frequent shopper programs, have facil-itated such an endeavor. In industries like consumer durables or services, detailedbehavioral data may not be available, hence approaches based on stated intentions orpreferences might be more suitable. One such technique –conjoint analysis – has typ-ically been used to measure attribute preferences within categories, but there may bean opportunity to extend the approach to measure correlatedness of preferences acrosscategories.However, analysis of consumers’ choices in disparate product categories poses dif-ficult modeling challenges. A fundamental issue that must be addressed is how tocorrelate