Modeling Multiple Category Brand Preference with Household ...

Modeling Multiple Category Brand Preference with Household Basket Data

GARY J. RUSSELL

University of lowa

WAGNER A. KAMAKURA

University ol Pittsburgh

Household basket datu contain important information about the structure of brand preferences both

within and across product categories. This research exploits the information in long-run basket .VIM-

mar-~ dutu to segment consumers with respect to brand preferences. The approach provides insights

into the competitive structure of brands within euch product catcgon: and identifies potentiul synergies

acm.cs product categories. The model is upplied in an analysis of retailer and national brand numc

preferencesforfourpaper

goods ccrtegorirs. We discuss implicationsforjoint promotion, product bun-

dling and product assortment decisions.

INTRODUCTION

Category management is increasingly advocated as a marketing strategy concept which has the potential to revolutionize retailing (Harris and McPartland, 1993). Simply put, category management is the notion that retailers should evaluate specific brands both in terms of contribution to product category profitability and in terms of contribution to overall store profitabililty (Nielsen, 1992). From a marketing theory perspective, category management builds strategies which integrate three basic marketing concepts: consumer preference segmentation, demand substitutability (within product categories) and demand complementarity (across product categories). The focus is on understanding the underlying determinants of consumer behavior and planning marketing activities which exploit the way that consumers assemble a basket of purchases.

Gary J. Russell is Associate Professor of Marketing at the College of Business Administration. University of Iowa, 108 Pappajohn Business Administration Building, Iowa City, IA 52242. Wagner A. Kamakura is Thomas Marshall Professor of Marketing at the Katz Graduate School of Business, University of Pittsburgh. 318 Mervis Hall. Pittsburgh. PA 15260.

Journal of Retailing, Volume 73(4), pp. 439461, ISSN: 0022-4359 Copyright 0 1997 by New York University. All rights of reproduction in any form reserved.

439

440 Cross Category Marketing Strategy

Journal of Retailing Vol. 73, No. 4 1997

From the retailer's perspective, an impo~~t implication of category management is the c~rdination of marketing strategies (product assortment, pricing and promotion) across product categories (Blattberg, 1989; Blattberg and Neslin, 1990; Zenor, 1994). Our research addresses the central role of cross-category preferences in the design and implementation of these strategies. In particular, knowledge of preference correlations is essential in evaluating the profitability of joint product promotions. If preferences for two complements (or inde~ndent) products are negatively correlated across the consumer population, then product bundling {selling two products as one unit at a single price} can be a profitable strategy (Nagle, 1987; Braden, 1993). In contrast, if preferences for a set of complementary (or independent) products are positively correlated across the consumer population, the retailer may be able to target a particular consumer segment by only promoting one product in the set. The increased store traffic by this segment may then increase the sales of all products in the set.

Knowledge of multiple category product preference patterns also allows the retailer to predict the likely composition of market baskets. Loblaws, a major Canadian grocery retailer, studies the make-up of consumer market baskets to understand the cross category product preferences of consumers who typically buy its President's Choice store label (Goldberg, Urban, and Wertz, 1995). Such information is used in both product assortment and store layout decisions. Catalina Marketing Coloration, a retailing indust~ consultant, sells point-of-purchase electronic couponing systems which can be programmed to issue coupons in a given category based upon the consumer's purchase behavior in other categories (Catalina Marketing, 1997). The coupon program is designed to encourage consumers to increase the number of categories considered on a shopping trip. In essence, the system allows the retailer to impIement a cross-category marketing strategy based upon correlated category usage patterns.

Measuring Multiple Category Preferences

Our research develops an approach for analyzing multiple category brand preference using long-run purchase summaries on consumer market baskets. Consistent with the logic of Grover and Srinivasan's (1987) approach to market structure, we build a consumer preference segmentation model and use the model output to study patterns in brand preferences. Our goal is the development of a marketing research tool which uncovers interesting patterns of preference correlation across product categories and thereby provides a basis for identifying potentially profitable cross-category marketing oppo~unities.

The proposed approach has two important features. First, the analysis is based upon a general model of multiple category demand. In contrast to earlier work in the market structure literature (Urban, Johnson, and Hauser, 1984; Grover and Srinivasan, 1987; Kamakura and Russell, 1988; Ramaswamy and DeSarbo, 1990; Russell and Kamakura, 19941, demand relationships among brands are unrestricted: brands can be substitutes, complements or independent. This feature is essential in any model of multiple-catego~ demand.

Modeling Multiple Category Brand Preference with Household Basket Data

441

It allows our model to represent consumer preferences for all brands of interest to the retailer.

Second, our procedure only requires information on brand consumption. Marketing mix variables (e.g., price and promotion data) are not used in this analysis. Although these variables affect the composition of market baskets on a week by week basis, we show analytically that a representation of brand preference segmentation can be recovered without explicit knowledge of marketing mix activity. Because our procedure places light demands on data collection, it can be used when marketing mix data is incomplete or missing altogether. Examples of such data sources include A.C. Nielsen wand panels (in which the consumer records purchases by scanning UPC product codes at home) and consumption data obtained from a retailer's "buyer club" participants. Thus, the procedure can be used when conventional panel data models are impractical.

Overview

We begin our discussion by formulating a general consumer purchase model for brands in multiple product categories. We then specialize the model to the analysis of long-run purchase volumes of brands in multiple product categories observed in a consumer panel. To illustrate the potential of the approach, we study consumer preferences for national and retailer brand name extensions across four paper goods product categories. Our results show that the cross-category pattern of preferences for national brand names are considerably different from the cross-category pattern for retailer brand names.

A MODEL OF MULTIPLE CATEGORY PURCHASING

On a typical shopping trip, consumers make multiple purchases, both within and across product-category boundaries. Clearly, category incidence plays an important role; consumers rarely purchase products in all available categories on the same shopping trip. In addition, category incidence is conditional upon a number of factors such as household inventory and marketing mix activity (both within and across categories), which are not always observable. Thus, we seek a model which simultaneously accounts for category incidence and product choice over a predetermined time period.

Household Purchase Model

Our model is developed in terms of purchase quantity. Let X [h (s), i(m), t] be the volume (in equivalent units) of brand i purchased by household h during week r. This notation emphasizes that each household h is a member of some consumer segment s (s = 1,2, ... S). Moreover, each brand i can be placed in a known category m. For example, in our empirical

442

Journal of Retailing Vol. 73, No. 4 1997

work, m represents one of four paper-goods categories: toilet paper, paper towels, facial tissue or paper napkins.

We assume that X 11%(s), i(m), t) is distributed as a Poisson random variable with mean

h[h(.s), i(m), t] = h[h, m, t]cY[s, i(m)

t]

(1)

where

h ih, m, tl is a time varying category purchase rate (which includes seasonality and

inventory effects),

a[s, i(m)1 is the stable intrinsic preference of segment s for product i(m), and P[i(m), fl is a time varying attractiveness of product i(m) due to marketing mix

activity.

That is, the mean is decomposed in terms of two time-va~ing components (category purchase rate and marketing mix activity) and one stable component (intrinsic product preference). Because the goal of this research is a representation of product preferences, our primary interest is the estimation of the segment-specific product preferences a[.~, i(m)]. The Poisson distribution used here has found considerable application in models of consumer purchase behavior (e.g., Ehrenberg, 1972; Lenk, Rao, and Tibrewala, 1993). Although the Poisson assumption makes the model more tractable, the key equation of our market basket model (equation (5) below) does not depend upon distributional form.'

We also assume that, conditional upon the means h[h(s), i(m), t], the observed purchase quantities X[h(s), i(m), t] are statistically independent with respect to household h(s), product i(nrj and time f. This conditional independence assumption is commonly made in the context of repeated observations on individuals {Fahrm~ir and Tutz, 1994, section 5.2). It is important to recognize that this assumption &es not rule out correlations among the observed X[h(s), i(nz), t]. Rather, conditional independence requires that any correlations in the observed quantities X[h(s), i(m), t] be induced by correlations among the purchase quantity mruns h/z(s), i(m), t¬ by correlations among the purchase quantity residuals X[h(s), i(m), t] - h[h(s), i(m). r] (see Gelman et al., 1996, section 5.2). In general, a retailer should expect to observe correlations among purchase quantities due to: (a) inter-category correlations in purchase rates h[h, m. t], (b) inter-segment correlations in product preferences a[~, i(m)], and (c) inter-brand correlations in marketing mix activities P[ i(m), t]. Our model allows for correlations of this sort.

Model Characteristics

Our model formulation has three desirable features for multiple-category analysis. First, a proportional intensity relationship is used to model the means in equation (1). The base rate of category purchasing h[ h, m, f] is muItiplicatively adjusted for both intrinsic preference a[.~, i{m)] and marketing activity P[ i(nz), t]. Although the marketing activity compo-

Modeling Multiple Category Brand Preference with Household Basket Data

443

nent is independent of household (or segment), the model does not imply that all consumers are equally affected by marketing activity. Note that the derivative

13h(h(s), i(m), t)/ap[i(m), t] = h[h, m, t]a[s, i(m)]

depends both upon the household and the consumer segment. In terms of the absolute change in the overall purchase rate A[h(s), i(m), t], our framework implies that consumers with a higher propensity to buy the category (large h[h, m, f]) or the product (large a[s, i(m)]) will be more affected by changes in marketing mix activity. Thus, the model allows for a natural interaction in category base rate, product preference and marketing activity.2

Second, the model does not force products to be substitutes. Looking over time, the model views the quantities purchased of different products i(m) as correlated time series where the pattern of correlations depends upon the time varying means a[h(s), i(m), t]. Because this approach allows for simultaneous purchasing of different products, it can accommodate complementary and independent products. This feature allows the model to accommodate all items in a consumer's market basket.

Third, the model implies that category incidence depends upon the household's pattern of intrinsic preferences. Using standard results on the Poisson distribution, it is easily

shown that Fi,,X[ h( s), j( m, t)], the total volume purchased in category m by household h

in week t, also follows a Poisson distribution with mean Ej ,h[h(s),j(m),

t]. Conse-

quently, the probability of purchasing category m is an increasiig function of

where the summation is over all products in category m. Intuitively, category incidence is more likely when marketing mix activity makes a high preference brand more attractive (i.e., a[s,j(m)] and Pu(m), t] are simultaneously large for some brandj.) This structure is similar to a nested logit formulation (cf. Bucklin and Gupta, 1992) in which category incidence depends upon category purchase propensity h[h, m, t] and a preference-weighted index of current category attractiveness CjE,a[s, j( m)] prj( m), t].

ANALYZING MARKET BASKET INFORMATION

Multiple category preference patterns can be understood by analyzing the inter-segment pattern of intrinsic product preferences a[s, i(m)]. Clearly, we could obtain this information first by specifying the components of the purchase rate decomposition (equation (1)) and then estimating a fully-specified demand system using weekly marketing mix information. However, when the researcher's primary goal is to recover estimates of the a[ s, i(m)], it is much simpler to work solely with long-run summaries of household basket data. In this section, we use the general multiple category purchase model to justify this simpler approach.

 ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download