Brand inertia in U.S. household cheese consumption.

The determinants of consumer demand have been the core of a robust research area at the macro and micro levels (Deaton and Muellbauer 1980; Campbell and Deaton 1989). Understanding demand at the aggregate level is key to economic growth, while that at the consumer or micro level has been of substantial interest to firms, researchers, and policy makers. The interest of oligopolistic firms has centered on demand for individual brands or varieties of a product, and consumers' willingness to switch among brands. Consequently, policy makers have focused on the exercise of pricing or market power, especially when consumer demand at the product or brand level is inelastic. For researchers, the existence of a critical mass of consumers for each variety or brand has been a key assumption of the monopolistic-competition models of industrial organization and economic growth (Krugman 1980; Barro and Sala-i-Martin 1999).

Recent advances in information technology have facilitated the compilation of large scanner and household databases, which have contributed to the better understanding of consumer behavior. Such advances include better supply chain management practices and the realization of scale economies through information obtained at the wholesale and retail levels. Retailers have played an important role in these advances, but their increased role in product promotion and marketing strategies has recently come under scrutiny in the context of antitrust policies (Federal Trade Commission 2003; Klein and Wright 2007). Thus, it is important to understand consumers' brand choices, which will help test the presence or absence of a critical mass of loyal customers. Moreover, it will help identify whether demand for products (attributes or services) is inelastic at the brand level, which create opportunities for the exercise of market power.

Prior studies have modeled brand choices using household and scanner data to provide guidance on firms' pricing and promotion decisions (Guadagni and Little 1983; Colombo and Morrison 1989; Capps 1989; Heien and Wessels 1990; Russell and Kamakura 1994). As Chintagunta, Kyriazidou, and Perktold (2001) note, understanding the determinants of consumers' purchase behavior is important since their willingness to switch brands affects demand elasticities and the degree of competition in oligopolistic markets with differentiated products (Bayus 1992). Along the same line, some studies have analyzed the competitive effect of a new product or brand introduction at the retail level (Hausman and Leonard 2002; Bonfrer and Chintagunta 2004). The focus of the latter is on how retail prices of existing products and of overall product category are affected by the introduction of a new brand.

Our study contributes to and extends the above literature by modeling brand choices as outcomes of consumers' dynamic utility maximization (Heckman 1981; Chintagunta, Kyriazidou, and Perktold 2001). In such optimization, we account for the role of past brand experience of households to investigate the extent of inertia (persistence) or variety-seeking behavior embodied in current choices. (1) Heckman (1981) and others note that households who have experienced an event in the past are more likely to experience the same event in the future. Such dependencies can occur when either preferences or constraints relevant to future choices are affected by experiencing an event or households differ in certain unmeasured variables, which affect the probability of experiencing an event. We avoid the latter instance by controlling for household demographic and locational characteristics unlike the above studies that focus mostly on brand prices and competition. We also verify the structure of the stochastic component of our empirical model while identifying inertia or variety-seeking behavior in households' brand choices. In modeling brand choice, we condition for households' past retail store choices. For instance, we can identify if brand-inert consumers would be willing to shop in several stores to purchase the preferred brand. Furthermore, we investigate whether demand for products (attributes or services) is inelastic at the brand level, which has implications for the exercise of market power.

The ACNielsen Homescan database is used to sample continuously cheese-purchasing U.S. households during 1998-2003 for estimating discrete-choice models of brand choices. The three major cheese categories in the sample are cheddar, shredded, and sliced American cheese. Our data are grouped into subsamples corresponding to four regions--Northeast, Central, Southeast, and West--which vary in terms of brand choices available to consumers. For example, Tillamook cheese, more frequently found in the western states, does not have a major presence in eastern U.S. states. In each region, purchase observations are compiled for top brands, which account for a majority of subsample cheese consumption in each category. Each regional brand model includes household demographic characteristics (size, education, and income) and location dummies. The maximum likelihood procedure is used to estimate the parameters of the brand choice models.

A Model of Brand Choice

Most studies analyzing brand choice employ an empirical model, where the systematic component of a consumer's utility from a brand is assumed to be a linear function of prices and household/individual characteristics (Guadagni and Little 1983). A variable measuring previous brand purchases is often introduced to identify consumers' loyalty to brands. However, the theory of brand decision making has received limited attention. Pollak (1970) is one of the earliest authors describing static, short-term, and dynamic utility functions, where the latter incorporated habit formation. Deaton and Muellbauer (1980), Heckman (1981), and Deaton (1992) are some key contributions to dynamic utility theory, which lead to micro-level demand models. Within the class of micro-level models, dynamic discrete demand models have often been employed to understand consumers' preferences for varieties and attributes within a product category (Chintagunta, Kyriazidou, and Perktold 2001; Seetharaman 2004). Here, the general framework is to set up consumers' utility in each period as dependent on the consumption of a product category, which is available in alternative forms (brands), observed and unobserved heterogeneity (e.g., product quality, household characteristics), and the choice made by consumers in the previous period. Then, consumers' maximization of expected discounted utility over an infinite horizon yields the optimal sequence of purchase decisions--a discrete choice.

We draw on the contributions noted above, especially Heckman (1981), Chintagunta, Kyriazidou, and Perktold (2001), and Seetharaman (2004), in specifying our dynamic, discrete-choice models of brand choice. We begin with consumers' brand-purchase decisions in successive intervals of time. Each consumer has M (m = 1, 2, ..., M) brands to choose from. Let v(m, i, t) be the expected (lifetime) utility that arises if the i-th consumer chooses the m-th brand at time t. The expected utility is a function of all relevant variables including demographic characteristics (e.g., household size, education, and income). The i-th consumer chooses the m-th brand at time t if:

(1) v(m,i,t) > v(n,i,t) m [not equal to] n; m, n = 1, 2, ..., M.

For any other brand n [not equal to] m, let the difference in expected utilities between choosing the m-th and n-th brand be:

(2) V(m, i, t | n) = v(m,i, t) - v(n, i, t) for all n [not equal to] m,

where V is conditional on n. Suppose the difference in utilities can be decomposed into two components: observed by the researcher, [bar.V](m, i, t | n), and the unobservable, [epsilon](m, i, t). Then, the difference in expected utilities can be rewritten as

(3) V(m, i, t | n) = [[bar.V](m, i, t | n) + [epsilon] m, i, t).

We record whether or not the i-th consumer chooses the m-th brand at time t by introducing a dummy variable d(m, i, t), which takes the value of 1 when the m-th brand is purchased and 0 otherwise. Thus, d(m, i, t) = 1 if V(m, i, t | n) > 0, while d(m, i, t) = 0 if V(m, i, t | n) [less than or equal to] 0. To make this model tractable, we follow Heckman (1981) in letting the difference in utilities, V(m, i, t | n), take the following functional form:

(4) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]