Quant Mark Econ DOI 10.1007/s11129-011-9117-0
Investigating brand preferences across social groups and consumption contexts Minki Kim · Pradeep K. Chintagunta
Received: 1 February 2010 / Accepted: 12 December 2011 © Springer Science+Business Media, LLC 2011
Abstract Using a unique dataset on U.S. beer consumption, we investigate brand preferences of consumers across various social group and context related consumption scenarios (“scenarios”). As sufficient data are not available for each scenario, understanding these preferences requires us to share information across scenarios. Our proposed modeling framework has two main building blocks. The first is a standard continuous random coefficients logit model that the framework reduces to in the absence of information on social groups and consumption contexts. The second component captures variations in mean preferences across scenarios in a parsimonious fashion by decomposing the deviations in preferences from a base scenario into a low dimensional brand map in which the brand locations are fixed across scenarios but the importance weights vary by scenario. In addition to heterogeneity in brand preferences that is reflected in the random coefficients, heterogeneity in preferences across scenarios is accounted for by allowing the brand map itself to have a discrete heterogeneity distribution across consumers. Finally, heterogeneity in preferences within a scenario is accounted for by allowing the importance weights to vary across consumers. Together, these factors allow us to parsimoniously account for preference heterogeneity across brands, consumers and scenarios. We conduct a simulation study to reassure ourselves
M. Kim (B) Korea Advanced Institute of Science and Technology (KAIST), Daejeon, South Korea e-mail:
[email protected] P. K. Chintagunta University of Chicago Booth School of Business, Chicago IL, USA e-mail:
[email protected]
M. Kim, P.K. Chintagunta
that using the kind of data that is available to us, our proposed estimator can recover the true model parameters from those data. We find that brand preferences vary considerably across the different social groups and consumption contexts as well as across different consumer segments. Despite the sparse data on specific brand-scenario combinations, our approach facilitates such an analysis and assessment of the relative strengths of brands in each of these scenarios. This could provide useful guidance to the brand managers of the smaller brands whose overall preference level might be low but which enjoy a customer franchise in a particular segment or in a particular context or a social group setting. Keywords Heterogeneous brand preferences · Social group and consumption context · Random coefficients logit model JEL Classification C51 · D12 · L66 · M3
1 Introduction In this paper, we use a unique dataset on beer consumption to investigate consumers’ brand preferences. Our data include choices of brands across consumers and over consumption occasions. These consumption occasions are uniquely associated with a “social group”—whether the consumption occurs alone, with friends, relatives etc.—and with a “context”—watching TV at home, at a dance club, etc. The objective then is to estimate the distribution of consumers’ brand preferences across these various scenarios (each scenario is a consumption context—social group combination). In principle, if a large number of panelists repeatedly consumes various brands of beer at each of the social group/context scenarios, we would be able to estimate the distribution of consumers’ preferences using a random coefficients choice model using only the observations corresponding to each of these scenarios. In practice, we may not have enough information or a sufficient number of observations for each individual social group/context combination that would allow us to estimate heterogeneous preferences for each scenario (as is true in our data). Consequently, it would become necessary to “share” information across the various scenarios. How can this be accomplished? We assume that the distribution across consumers of brand preferences for the J brands in a given scenario s (s = 1, 2, ..., S) can be decomposed into two components. The first component is the standard random coefficients logit model that accounts for preference heterogeneity across the J brands. This component is common across scenarios. To allow for differences in preferences between scenarios, we first pick a “base” scenario (say, s = 1) and then parsimoniously account for the deviations in preferences under each
Investigating brand preferences across social groups and consumption contexts
of the other scenarios from those for that base scenario.1,2 We assume that the (J − 1) × (S − 1) deviations of mean preference parameters can be parsimoniously represented by projecting these deviations onto a low dimensional (unobserved) “attribute” space. In particular, the (J − 1) × (S − 1) deviations can be decomposed into (J − 1) × D locations of brands in D-dimensional space (with the Jth brand located on the origin) as in Elrod and Keane (1995), Chintagunta (1994), etc.; and D × (S − 1) attribute weights that consumers assign to the D dimensions in each of the S − 1 scenarios. Parsimony is achieved if D (J − 1) × (S − 1)/(J + S − 2), i.e., typically when there are a large number of brands and/or scenarios. For example, with 11 brands and 25 scenarios (as is the case with our empirical application), the “full” model would require estimating 240 preference deviation parameters whereas with the proposed approach and D = 2, we estimate only 68 such parameters. Further, additional (to the random coefficients) heterogeneity across consumers within a scenario (other than the base scenario) can be accommodated in one of several ways: (i) by allowing for a discrete distribution of the (J − 1) × D location matrix across a finite number of “segments” (as a continuous distribution will require a large number of parameters); (ii) by allowing the D × 1 vector of attribute weights for each scenario to have a D-variate normal distribution across consumers; or (iii) by combining (i) and (ii). Once again, parsimony is achieved since the heterogeneity distribution is restricted in dimensionality to D rather than to S − 1. In cases (i) and (iii), we have a hybrid continuous-discrete model of heterogeneity (the discrete component coming from (i) and the continuous components from the standard random coefficients distribution as well as from the distribution of importance weights) whereas in case (ii) we only have a combination of two continuous distributions. A key advantage of this specification is that in the absence of information on the consumption scenarios, our model reduces to the standard random coefficients logit model. By sharing information across scenarios, our data allow us to extend extant random-coefficients choice models to identify scenario-specific preference parameters as well. By estimating the distribution of brand preferences for the different scenarios, managers will be able to use this information to understand the following. First, if a brand is dominant across all scenarios, then it means that the brand is fortressed well against competitive encroachment across the different contexts within which those competitors might position themselves. Second, and more importantly, for a brand whose overall preferences are small, such an analysis
1 In
other words, the preference distribution from the random coefficients logit model is also the distribution of preferences for the base scenario. 2 Allowing for the full set of random coefficients across all S − 1 remaining scenarios would require estimating (J − 1) × (S − 1) additional mean preference parameters and (S − 1) × J × (J − 1)/2 additional covariance parameters for the heterogeneity distribution. Note that with J brands there are only J − 1 parameters since one of the mean preference parameters is normalized to 0.
M. Kim, P.K. Chintagunta
could reveal niches, i.e., scenarios within which is might have strong preferences. This information can then be used by the brand to reinforce its image, provide the basis for creative campaigns and for finding “adjacent” scenarios to which it could turn its attention to next. These are insights that will not be available if the scenario information is ignored. The output of our analysis therefore, enables us to assess the relative strengths of the various brands in each of the context/social group scenarios. As a brand that has a low overall market share might nevertheless be preferred by consumers in a specific scenario, we look for such situations in our results. These results could provide useful guidance to the smaller brands as to the kind of copy/imagery that they might want to use in their commercials such that the copy/image could resonate with consumers in those scenarios, i.e., to facilitate the positioning function. Note that our methodology facilitates such an analysis even if the data on specific brand—scenario combinations are sparse. Further, given the within scenario heterogeneity we account for in our analysis, we are able to provide some assessment as to the specific demographic groups as well as geographic markets where preferences for a brand within a scenario are high. In this way, our model and analysis also provide inputs into the targeting function for each brand. Finally, by including advertising as a driver of consumer choices in our analysis, we are able to assess how sensitive consumers would be to advertising, relative to their sensitivities to the advertising of the other brands. The remainder of this paper is organized as follows. In the next section, we briefly describe related research. This is followed by a section on the data that are available to us. We then specify our model that takes into account the key features of our data. The subsequent sections relate to our empirical results and a general discussion.
2 Related research Marketing researchers have for long been interested in studying the drivers of households’ brand choices. This interest has typically manifested itself in analyzing scanner panel data to identify the impact of marketing activities of firms as well as characteristics of households on the choices made by households in a product category (see, e.g., the long literature in marketing beginning with Guadagni and Little 1983 and leading up to the recent survey by Allenby and Rossi 2009). More recently, researchers have also been interested in studying the extent to which these choices of households may be influenced by the context within which the product is consumed (e.g., is the brand of beer chosen by a consumer depend on whether the consumption occurred in a bar or at home) as well as the social situation associated with the consumption occasion (e.g., whether the beer was consumed when the person is alone or with a close friend, a partner, etc.). While the behavioral literature on this is extensive (Belk 1974, 1975; Fennell 1978; Dickson 1982; Ratneshwar and Shocker 1991; Bettman et al. 1998; Carlson and Bond 2006, etc.), quantitative
Investigating brand preferences across social groups and consumption contexts
studies are more limited. Three studies, among others, typify the type of research in this area. Yang et al. (2002) look at how consumers’ preferences for beer brands vary across, what they refer to as, objective environmental conditions and motivating conditions. Specifically, they study the impact on consumers’ preferences of the interactions between product attributes (such as brand name, flavor, aroma, number of calories) and different motivating conditions (such as being thirsty, wanting to relax, etc.) and on how these preferences then drive consumers’ choices across different environments such as drinking alone at home watching TV, drinking with others at home eating dinner, drinking at a party at a friend’s house, etc. While the data are collected via a survey and conjoint task, the authors are able to investigate the variation in preferences across consumers since they elicit multiple choices from the same respondent. This study represents an early and important first step in integrating the information on consumption “context” to understand intra- and inter-person variation in brand preferences. Our current study shares many aspects of similarity with that of Yang et al. In particular, we look at the same product category and the same general notion of better understanding consumer heterogeneity as it relates to social groups and contexts in which consumption takes place. Different from them however, our interest is to focus on the brands and to understand how consumers differ in their preferences for the brands in the different social groups and context scenarios. There are two studies related to ours that do look at unobserved attribute space and the locations of brands in this space alongside the variations in preferences across consumers due to different uses and usage occasions, etc. The first is the study by Lee et al. (2002) which attempts to identify multiple ideal points for a household using only revealed preference, i.e., purchase data from that household. The idea is that the observed purchase sequence comes from households switching between ideal points where each ideal point could correspond to a usage occasion (and thus to a different social group or context scenario in our discussion above). Since these ideal points are located in the same (unobserved) attribute space as the brands, consumers have different brand preferences depending upon which of the different ideal points is activated on a given purchase occasion. Further, as the actual consumer motivations at each purchase occasion are unknown, the data represent a “mixture” of purchase behavior corresponding to the different ideal points. Given the nature of data however, one cannot associate each ideal point with a specific scenario let alone allow brand locations to vary across consumers. Yet this study represents an important building block to acknowledging the importance of accounting for preference variation across usage occasions. Building on the Lee et al study, DeSarbo et al. (2008) propose a new clusterwise multiple-ideal-point model and methodology that allows different consumer segments to have different ideal points corresponding to the 5 situations in which they might consume over-the-counter analgesics (overall, headaches, fever/cold, etc.). Since these situations correspond to the different social group/context scenarios described in our data, the DeSarbo et al study is
M. Kim, P.K. Chintagunta
closely related in its objectives to what we are focused on in this study. At the same time, given their data, the researchers have to assume that the relative brand locations remain the same across the different customer segments. Thus while preferences in the different situations do vary across segments, all consumers agree on the relative locations of brands in attribute space. By contrast, we use the attribute space to characterize deviations in consumer preferences from those under a base scenario. Different from them (and from Lee et. al.) we also do not employ an ideal point model but rather focus on preferences (or more specifically, the deviations in preferences from those in a “base” scenario) as in Elrod and Keane (1995). Thus our research can be seen as combining elements of the Yang et al. (2002) study with elements of the Lee et al and the DeSarbo et al studies.
3 Data For our empirical analysis, we construct a unique dataset that integrates multiple data sources. The panel data on household consumption of beer including the detailed demographic information on the respondents is provided by Research International. Research International conducts extensive surveys on a quarterly basis to collect information on U.S. household consumption and purchasing behaviors of alcoholic beverages. In each quarterly survey, approximately 6,000 households are selected to participate as panelists. While the panelists from one survey to the next may change, the demographic distribution of each survey is controlled to remain predominantly unchanged. Panelists are requested to respond to survey questions about their alcoholic beverage consumption and purchasing behaviors during the 7 days prior to the date of the survey. Thus we refer to the respondents as panelists only because they typically provide information on multiple purchases and/or consumption occasions during the 7 day period. The questions provide detailed information on the respondents’ beer consumption including their brand choices, associated social groups at time of consumption, context information within which consumption occurs and marketing-related factors associated with every consumption occasion—the role of prices and product availability. In our analysis, we focus our attention on the 10 biggest beer brands by share in the sample that account for 95% of all purchases by the households in the data collection period. We collect all the other brands chosen by these panelists into a single composite 11th brand. We use data over a period of two years, from January 2006 to December 2007 for our estimation. By only keeping the panelists that made at least four consumption decisions over the study period, we are left with 1,839 panelists in the 191 DMAs accounting for 19,739 choice observations in our sample. To investigate the predictive power of the model estimates, we split these panelists into an estimation sample and a holdout sample. Our holdout sample consists of 389 panelists, 20% of the whole sample, making 4,124 consumption decisions.
Investigating brand preferences across social groups and consumption contexts Table 1 Descriptive statistics for top 10 beer brands Brand number
Brand name
1 2 3 4 5 6 7 8 9 10 11
Bud Light Budweiser Miller Lite Coors Light Corona Extra High Life MGD Heineken Natural Light Busch Composite Total
Frequency sample 4,261 4,309 2,689 2,206 726 1,599 1,243 654 1,116 888 48 19,739
Share (%) sample 21.59 21.83 13.62 11.18 3.68 8.1 6.3 3.31 5.65 4.5 0.24 100
Share (%) overall samplea 24.11 20.68 14.19 10.9 6.21 6.09 5.53 4.21 4.32 3.36 0.39 100
Share (%) U.S. marketb 29.0 17.7 12.7 11.95 5.86 3.5 2.36 3.65 6.39 4.34 2.42 100
a Note
that these top 10 brands account for approximately 57.47% of shares in the overall sample for 2 years, from January 2006 to December 2007
b Source:
Beer marketer’s INSIGHTS (http://beerinsights.com). In terms of shipments (including exports), these brands account for 64.15% which is the average of year 2006 and 2007
In Table 1, we provide descriptive statistics for the top 10 brands and one composite brand from all 19,739 consumption occasions. Note from Table 1 that the top 3 brands control over 60% of the market among these panelists. To see how the numbers match up to the overall shares for these brands across all households (and not just the ones that had at least 4 consumption decisions) as well as to national shares, we provide those numbers in columns 5 and 6 in Table 1. While there are a few differences (Budweiser being a prominent one), we see that our selected sample of households is, by and large, quite representative both of the overall sample and of the national profile. Demographic variables play an important role in brand choice decisions by consumers. This leads us to consider the following six demographic variables available in the data: age of respondents (Age); a dummy for nonCaucasian respondents (Non-white); a dummy for respondents living in rural areas (Rural); a dummy for male respondents (Sex); a dummy for married respondents (Married); a dummy for respondents who are currently working (Work).3 In Table 2, we provide summary statistics of the demographic variables. Table 2 seems to indicate that the sample is slightly skewed towards nonwhite drinkers (who account for about a third of beer consumers in the population). Overall, there does appear to be a fair bit of variability in the demographic variables and we will exploit this when trying to understand
3 We
divide the age by 100 and take the demeaned value of age in estimation.
M. Kim, P.K. Chintagunta Table 2 Summary statistics of demographic variables Variable
Number of people
Mean
Std. dev.
Min
Max
Age Non-white Rural Sex Married Work
1,839 1,839 1,839 1,839 1,839 1,839
48.873 0.389 0.192 0.699 0.599 0.662
14.561 0.488 0.394 0.459 0.490 0.473
21 0 0 0 0 0
95 1 1 1 1 1
heterogeneity in brand preferences across consumers. In addition to the above variables, our data also contain information on the social group and the context that is uniquely associated with each consumption occasion recorded by the panelist. It is this information that enables us to identify heterogeneity in preferences across various situations. We classify the social groups and the contexts each into five categories, which generates 25 possible social group/context scenarios. We then analyze the consumption observations corresponding to each of the 25 possible social group/context combinations. Descriptive statistics for social group and context scenarios are given in Tables 3 and 4, respectively. Tables 3 and 4 show that over half of all consumption occasions fall in the “alone” category. Further, the most frequently stated context is relaxing and watching TV. The other social groups and contexts seem to be evenly associated with beer consumption. One advantage of understanding the preferences of consumers across these various scenarios is that it can provide input to marketing managers regarding e.g., the settings of commercials that either emphasize current preferences or that direct consumers to other contexts and social groups within which a brand of beer can be consumed. Table 5 summarizes Tables 1, 3, and 4 arraying all combinations of brand choices with social group/context scenarios for every consumption occasion. Table 5 highlights the need for sharing information across scenarios. Note that certain cells in the above table are sparsely populated implying that obtaining brand preferences for each of the 25 scenarios is likely to pose a challenge. Our approach below shares information across scenarios, brands and consumers to obtain preference estimates. With the type of consumption data available to us, a question that arises is whether variation in the social group/consumption contexts is largely a
Table 3 Social group associated with consumption Number
Categories for social group
Frequency
1 2 3 4 5
Alone With a very close person With relatives With one or two people With group of friends
11,083 3,003 1,761 1,648 2,244
Total
19,739
Share(%) 56.15 15.21 8.92 8.35 11.37 100
Investigating brand preferences across social groups and consumption contexts Table 4 Context associated with consumption Number
Categories for context
Frequency
1 2 3 4 5
Relaxing, watching TV at home Eating a meal at home Working or doing a hobby at home Dancing club and sports bar Local pub
12,852 2,244 1,787 1,067 1,789
Total
19,739
Share(%) 65.11 11.37 9.05 5.41 9.06 100
cross-sectional phenomenon, i.e., some beer drinkers always consume under certain circumstances whereas others consume under different circumstances, or whether it is a combination of both cross-sectional as well as time series variation within a particular consumer. Accordingly, we construct a scenario switching matrix (the scenario under which a consumer drinks beer at time t given the consumption scenario in (t − 1)) from the data. Given the dimensions of this matrix, we do not report it here, it shows that there is indeed switching across these scenarios which indicates that preferences are being identified due to variation both within and across consumers. Although not addressed here, a question of interest is whether the strong state dependence effects documented in recent research (Dube et al. 2009) can, at least partially, be attributed to persistence in consumption scenarios for a customer over time. Our survey data also contain information on marketing-mix related factors that might have motivated a specific brand being consumed during a particular occasion. Respondents provide information on whether or not price or availability was a factor in a brand being chosen. We label these variables as “motivational factors” and include them in our analysis. Including these factors will enable us to isolate the preferences for the products after controlling for the effects of these marketing variables. Low price was a motivator in about 6% of cases whereas availability was cited as the main reason for the choice in 4% of observations.
Table 5 Brand choices per social group/context scenario Brand
5 Categories of social groups 1 2 3 4 5
Total
5 Categories of context 1 2 3 4
5
Bud Light 2,203 655 455 425 523 4,261 2,692 466 377 273 453 Budweiser 2,643 508 332 332 494 4,309 2,929 432 399 234 315 Miller Lite 1,327 589 164 221 388 2,689 1,621 280 243 199 346 Coors Light 1,142 413 225 179 247 2,206 1,343 237 212 141 273 Corona Extra 335 96 113 68 114 726 488 76 41 61 60 High Life 998 224 137 118 122 1,599 1,091 264 114 43 87 MGD 796 153 101 95 98 1,243 810 174 100 57 102 Heineken 340 71 77 78 88 654 425 75 39 40 75 Natural Light 741 169 78 50 78 1,116 770 161 151 5 29 Busch 529 120 75 79 85 888 657 72 106 11 42 Composite 29 5 4 3 7 48 26 7 5 3 7 Total 11,083 3,003 1,761 1,648 2,244 19,739 12,852 2,244 1,787 1,067 1,789
M. Kim, P.K. Chintagunta
Since geographic information on the households’ locations (i.e., designatedmarket-area) are available in the survey data, we combine the consumption data with aggregate, market-level advertising information provided by TNS Media Intelligence (TNS MI) reports. TNS MI collects weekly media advertising expenditures (such as those on cable TV, Network TV, consumer magazines, newspaper, radio, outdoor advertising, Internet, etc.) at the designatedmarket-area (DMA) level. One issue we face is that the Network TV advertising expenditures are common across consumers in all markets. Yet, Network TV is viewed to a different extent by consumers in different DMA markets. Therefore, we transform the Network TV expenditures to the DMA level based on DMA-level viewership data. The viewership data are available from the Nielsen Station Index in which Nielsen Media Research estimates the total number of TV households in the U.S. and the number of TV households in each of the 210 Designated Market Areas (DMA). We allocate total Network TV dollars to each DMA in proportion to that DMA’s television viewing households relative to the total number of television viewing households in the U.S. As we use Network TV expenditures as a measure of national advertising, we construct the local advertising expenditures by combining the expenditures on local newspaper, radio, cable TV, Internet, magazines, and outdoor advertising at the DMA level. Since we know the DMA of each of our consumers, we can match up the local advertising expenditures to each customer based on his/her location. Notice that TNS MI collects the DMAlevel advertising information for about 100 large DMAs at the brand level but expresses the total amount of local advertising expenditures on all other DMAs as one aggregated value. In order to make the most out of TNS MI data, we divide this value by the number of DMAs not shown in the data. For instance, if TNS MI collects the advertising expenditures of Bud Light for 96 DMAs at the weekly level in 2007, the value for all other DMAs is equally divided across the remaining 95 DMAs at the weekly level. We present descriptive statistics on the advertising expenditures—national and local in Table 6. Table 6 Descriptive statistics on national/local advertising Brand
National advertising expenditures Total U.S. dollars (000) Year 2006 Year 2007
Local advertising expenditures Total U.S. dollars (000) Year 2006 Year 2007
Bud Light Budweiser Miller Lite Coors Light Corona Extra High Life MGD Heineken Natural Light Busch Composite
103511.5 50198.4 57043.6 74500.5 9005.4 340 9664.7 7806.8 0 2340 350.3
16030.3 11828.3 8205.7 24476.5 12529.3 2446.5 19969.4 33818.7 85.1 1341.8 5775.5
98697.3 50443.1 45950.2 66496.1 13222.3 13742 0 10594.5 0 165.1 1691
21591.5 11701.4 13763.7 25282.6 12080.8 5384.8 7119.8 39814.2 113.6 1342.1 6738.7
Investigating brand preferences across social groups and consumption contexts Table 7 Product attributes of beer brands
Brand
Alcohol (%)
Calories/12 oz
Bud Light Budweiser Miller Lite Coors Light Corona Extra High Life MGD Heineken Natural Light Busch Composite
4.1 5.0 4.2 4.2 4.6 4.7 4.7 5.0 4.2 4.6 4.75
110 145 96 102 148 143 152 150 95 133 126.5
The table shows that the large brands do invest more on advertising than do the smaller brands. Further, some of the smaller brands, e.g., Natural Light do not spend at all on national advertising whereas others such as Miller Genuine Draft (MGD) significantly cut spending in 2007. To resolve the endogeneity problem associated with both local and national advertising expenditures, we adopt a control function approach used in Petrin and Train (2006, 2010). We use the average advertising expenditures spent in the other DMAs as an instrumental variable for the DMA-level local advertising expenditures (as in Hausman 1996; Nevo 2001) since local advertising expenditures are likely driven by local demand conditions. Coming up with an instrument for Network TV advertising is more of a challenge. We use the sweeps rating periods as an instrument for Network TV advertising expenditures. The intuition here is that during the sweeps periods, advertisers are more likely to want to advertise since the networks air higher quality shows. This is unlikely to be correlated with the demand for any specific product category (in our case, beer).4 Apart from the survey data, we also collect information on product attributes—alcohol content (%) and calories/12 oz as possible drivers of preferences and report these data in Table 7.
4 Model formulation In our empirical analysis, we focus on understanding the heterogeneity in brand preferences across social group and consumption context scenarios. We describe the formulation of our proposed model in this section. Recall that our main objective is to parsimoniously represent the preferences across these scenarios given the limited amount of data we may have for any individual
4 Sweeps
months are November, February, May and July in which Nielsen collects viewing information based on seven-day diaries filled out by sample households in many television markets in the United States.
M. Kim, P.K. Chintagunta
scenario. Thus we need to be able to share information across scenarios in the formulation. Denote by U ijdt consumer i’s (i = 1, 2, . . . , N) indirect utility for brand j ( j = 1, 2, . . . , J) on consumption occasion t (t = 1, 2, . . . , Ti ) in DMA d. We can decompose U ijdt as: U ijdt = aij∗ +
S
αijs Bist + Z i υ j +
s=2
M
imt γmj + X1,ijdt (β1,i + ζ Z i )
m=1
+ X2, jβ2,i + ξ jdt + ijdt
(1)
where aij∗ ( j = 1, 2, . . . , J − 1) is the intercept term representing the intrinsic preference of consumer i for brand j (for the “base” scenario s = 1); αijs denotes the consumer i’s time-invariant deviation in intrinsic preference for brand j for the sth social group/context scenario from the preference in the base scenario; Bist is an indicator variable which takes the value of 1 if the sth social group/context scenario is associated with person i’s consumption occasion t; Z i is the 1 × H vector of demographic variables of consumer i; υ j represents the H × 1 vector of the effect of these demographic variables on preference for brand j; imt is the 1 × M indicator vector that takes the value of 1 if consumption occasion t of the person i is associated with a marketing-mix related motivational factor (price or availability); γmj is the M × 1 vector of the effect of motivational factors on preference for brand j; X1,ijdt is the 1 × F1 vector of marketing variables, e.g., advertising, associated with brand j corresponding to consumption occasion t of the person i residing in DMA d; β1,i is the F1 × 1 vector of person-specific response parameters associated with the covariates X1,ijdt ; ζ represents the F1 × H matrix of the effect of the demographic variables on the covariates X1,ijdt ; X2, j is the 1 × F2 vector of product attributes of brand j; β2,i is the F2 × 1 vector of personspecific response parameters associated with the covariates X2, j; ξ jdt is brand j’s unobserved portion of mean utility varying with d and t; and ijdt is the error term. Finally, ijdt is the unobserved (to the researcher) component of utility that we assume follows a type-I extreme value distribution. Now, we discuss how we parsimoniously represent the deviations in intrinsic preferences across the S − 1 scenarios, i.e., the αijs in Eq. 1. Denote the J × 1 vector of consumer i’s preference deviations for the J brands for the sth social group/context scenario as αis = [αi1s , αi2s , . . . , αi Js ] . Assuming a factor structure for these preference deviations, we are able to decompose the brand preference deviations for each social group/context scenario s where s = 2, . . . , S as:5 αis = Ai · wis
5 There
(2)
are several identification issues associated with these types of models. We return to these issues later.
Investigating brand preferences across social groups and consumption contexts
where Ai is a J × D matrix of the “locations” of the J brands on the Ddimensional map for person i; and wis is an D × 1 vector of (possibly negative) importance weights for the D dimensions corresponding to the sth social group/context scenario. Equation 2 can be shown as: ⎤ ⎡ a1 i1 αi1s ⎢ αi2s ⎥ ⎢ 1 ⎢ ⎥ ⎢ ai2 ⎢ .. ⎥ ⎢ . ⎢ . ⎥ ⎢ . ⎢ ⎥ ⎢ . ⎢ αijs ⎥ = ⎢ 1 ⎢ ⎥ ⎢ aij ⎢ .. ⎥ ⎢ ⎢ . ⎣ . ⎦ ⎣ . . αi Js ai1J ⎡
2 ai1 . . . ai1D
⎤
⎥ ⎡ 1 ⎤ 2 wis . . . ai2D ⎥ ai2 ⎥ ⎥ .. ⎥ ⎢ .. ⎢ wis2 ⎥ . ⎥ . ⎢ · ⎥ ⎢ . ⎥ ⎥ aij2 . . . aijD ⎥ ⎥ ⎣ .. ⎦ .. .. ⎥ wisD . . ⎦ ai2J . . . aiDJ
(3)
where aij = aij1 , aij2 , . . . , aijD , j = 1, 2, . . . , J, denotes the jth row of the matrix Ai . Then, the utility function can be written as
U ijdt = aij∗ + aij
S s=2
wis Bist + Z i υ j +
M
imt γmj
m=1
+ X1,ijdt (β1,i + ζ Z i ) + X2, jβ2,i + ξ jdt + ijdt
(4)
Note that the above formulation is quite general in the sense that we allow for the locations and the attribute weights to be consumer-specific. It is worth noting that the model specification in the literature that uses a factor structure to represent preferences in a single social group/context scenario assumes the matrix of locations to be common across consumers whereas the importance weights to vary across them (e.g., Elrod 1988; Chintagunta 1994; Elrod and Keane 1995). This restriction is necessary to ensure identification. In our case, as we show below, the availability of data across multiple scenarios allows us share information across these scenarios to estimate heterogeneous location and importance weight parameters. 4.1 Properties of the proposed model specification (1) The formulation in Eq. 1 nests the standard random coef f icients model. Recall from the above equation that the term aij∗ is common across all the S scenarios. We assume that aij∗ follows a multivariate normal distribution across consumers, i. Now, consider the special case of our model where Bist = 0 for all scenarios, s > 1. Our formulation in Eq. 1 is then a standard random coefficients choice model with a continuous distribution of heterogeneity across consumers (see for example Allenby and Rossi 1999). In this case, we are not using the information contained in the consumption scenarios in
M. Kim, P.K. Chintagunta
order to distinguish consumers’ differential preferences in these scenarios. The standard random coefficient model without the social group/context information can be written as follows. U ijdt = aij∗ + Z i υ j +
M
imt γmj + X1,ijdt (β1,i + ζ Z i ) + X2, jβ2,i + ξ jdt + ijdt
m=1
(5) where
i = aij∗ ( j = 1, 2, . . . , J − 1) , β1,i , β2,i ∼ MV N ( , ) .
= a∗j ( j = 1, 2, . . . , J − 1) , β1 , β2 ; is a (J − 1 + F1 + F2 ) dimensional covariance matrix. (2) Model in which mean preferences are the same across scenarios but variance in brand preferences is dif ferent across scenarios. To see how such a model is nested within the general specification in Eq. 1, we can write the preference deviation parameters as follows αis = A · wis
wis ∼ MV N 0, σs2 I D
(6)
In the above equation, A as before, is a J × D matrix of the “locations” of the J brands on the D-dimensional map but which is common across all households; and wis is an D × 1 vector of (possibly negative) importance weights for the D dimensions corresponding to the sth social group/context scenario. Further, we impose a specific structure on these importance weights assuming them to follow a multivariate normal distribution with a zero mean (represented by the D × 1 vector of zeros, 0) and a covariance matrix which is the D−dimensional identity matrix scaled by a scenario-specific variance term. The restriction to the identity matrix stems from the identification conditions laid out in Elrod (1988) and by Elrod and Keane (1995). Given the mean zero assumption, the mean preference for brand j is scenario s is still a∗j ( j = 1, 2, . . . , J − 1) for all scenarios. However, the variance in preferences for brand j in scenario s(> 1) is now given by: jj + σs2 a jaj where jj is the variance to the jth row and jth column of the matrix corresponding
1 2 D and a j = a j , a j , . . . , a j , denotes the jth row vector of the matrix A. In the absence of identification restrictions, this model requires the estimation of an additional J × D parameters for the matrix A and S − 1 parameters for σs2 as compared to the simple random coefficients model in 1).
Investigating brand preferences across social groups and consumption contexts
(3) Model in which mean preferences and variances in brand preferences are dif ferent across scenarios. This model relaxes the assumption on the mean preferences across scenarios imposed by the specification in model 2) above. Here, αis = A · wis
wis ∼ MV N ws , σs2 I D
(7)
where ws now denotes the D × 1 mean vector of the distribution of importance weights for the sth scenario. Given this assumption, the mean preference for brand j is scenario s is now a∗j + a jws ( j = 1, 2, . . . , J − 1) for scenarios 2 through S. The variance in preferences for brand j in scenario s(> 1) is still given by: jj + σs2 a jaj as in the above specification. Compared to the model in 2) above, this specification requires the estimation of an additional (S − 1) × D parameters. (4) Incorporating additional heterogeneity in the brand location matrix, A. Recall that the matrix A is common across all consumers and all scenarios. One can imagine relaxing this assumption to provide further flexibility to the distribution of preferences across scenarios without significantly increasing the number of parameters to estimate. In particular, we can allow the parameters in the matrix A to vary across households following a discrete distribution with
supports or market “segments.” This will yield the following distribution for the preference deviation parameters: αis = Aiπ · wis
wis ∼ MV N ws , σs2 I D Aiπ = Aπ , π = 1, 2, ..,
(8)
This model is now a hybrid discrete-continuous model and requires the estimation of an additional J × D × ( − 1) location parameters as compared to the specification in 3) above. Further, we need to estimate ( − 1) additional parameters for the sizes of the segments. In sum, our model formulation takes the (J − 1) × (S − 1) additional mean preference parameters and (S − 1) × J × (J − 1)/2 additional covariance parameters required for estimating a random coefficients model for each scenario and reduces this to an additional (J × D × ) + (S − 1) × (D + 1) parameters. As long as D and are small, this will result in a more parsimonious model specification than the full model. Further identification restrictions will lower the number of parameters to be estimated. In particular, only a (J − 1) × D − 1 matrix of brand locations is identified as one of the brands needs to be located at the origin (to accommodate translational invariance, Elrod 1988) and another along one of the axes (rotational invariance). This drops the number of parameters to (((J − 1) × D − 1) × ) + (S − 1) × (D +
M. Kim, P.K. Chintagunta
1) + ( − 1). Suppose we have 11 brands (J), 25 scenarios (S), a 2-dimensional map (D) and 3 supports for the distribution of locations ( ), the full model would require the estimation of 240 mean deviation parameters and 1,320 covariance matrix parameters (= total of 1,560 parameters). Our proposed model by contrast estimates only an additional 131 parameters as compared to a single standard random coefficients choice model estimated across all 25 scenarios. This is close to a 12-fold reduction in the number of parameters to be estimated. We formulate the likelihood for our proposed model specification as follows. First, as we assume a type I extreme value distribution for the error term ijt , we obtain the multinomial logit model (McFadden 1973) for our choice probabilities. As we observe data at the household level over Ti occasions for household i where each consumption occasion t is uniquely associated with a scenario s, we can express the conditional probability Pijt of household i choosing brand j on consumption occasion t assuming a consumption scenario S s, i.e. r=2 wr Birt = ws , as: Pijt
M exp aij∗ + aij ws + Z i υ j + m=1 imt γmj + X1,ijdt (β1,i + ζ Z i ) + X2, j β2,i + ξ jdt = J M ∗ exp aig + aig ws + Z i υg + imt γmg + X1,igdt (β1,i + ζ Z i ) + X2,g β2,i + ξgdt g=1
(9)
m=1
for estimation, we let {θi , πi } denote the set of parameters varying across consumers, which reflects unobserved heterogeneity, where θi = { i , wis } and πi = {Ai }. We assume that θi is a realization of the random variable (vector) θ and πi is a realization of the random vector π which have multivariate distributions G(θ) and H(π ) across consumers, respectively. Conditional on {θi , πi }, the likelihood function for consumer i can be written as: Li|θi ,πi ⎧ ⎪ ⎪ ⎪ ⎪ ⎪ J Ti ⎪ ⎨
⎛
⎞ϑ ⎫ ijt ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎬
⎟ ⎜ ⎜ exp a∗ + a S ws B + Z υ + M γ + X ⎟ ⎜ ij s=2 ist i j 1,ijdt (β1,i + ζ Z i ) + X2, j β2,i + ξ jdt ⎟ m=1 imt mj ij ⎜ ⎛ ⎞⎟ = ⎜ J ⎟ ⎪ S M ⎟ ⎪ t=1 ⎪ j=1 ⎜ ⎪ ∗ +a ⎝ ⎪ exp ⎝aig ws Bist + Z i υg + imt γmg + X1,igdt (β1,i + ζ Z i ) + X2,g β2,i + ξgdt ⎠⎠ ⎪ ig ⎩ g=1 s=2 m=1
⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎭
(10)
where T i is the number of brand choices made by person i during the sample period, Bist takes the value of 1 if the sth social group/context scenario is associated with person i’s consumption occasion t, ϑijt = 1 if person i chooses brand j on occasion t and ϑijt = 0 otherwise. Since the unobserved product characteristics ξ jdt might be correlated with the marketing variables X1,ijdt , we resolve the potential endogeneity bias by following the control function approach suggested by Petrin and Train (2006, 2010).6
6 We
run the first stage regression using the instrumental variables suggested in the data section. After obtaining the predicted marketing variables and accordingly the residuals, we use them to create control functions for the individual level data.
Investigating brand preferences across social groups and consumption contexts
Note that our estimation is based on the random effects model with a “hybrid” discrete-continuous heterogeneity distribution which approximates G(θ) by a normal distribution and H(π ) by a discrete distribution with a finite number of supports , Aπ , π = 1, 2, .., with associated probability masses ϕ(πk ), k = 1, 2, . . . , such as k=1 ϕ(πk ) = 1. Hence, the distribution of π is determined empirically from the data and the number of support vectors is determined by using a stopping-rule procedure based on the Bayesian information criterion (BIC, Allenby 1990). To obtain the unconditional likelihood for consumer i, we first integrate over the normal distribution G(θ ) of the random variable θ. " (11) Li|πi = Li|θ,π dG(θ ) θ
Then, for the random variable π , we integrate over the discrete distribution H(π ) with a finite number of supports : Li =
Li|πk · ϕ(πk )
(12)
k=1
As we compute the unconditional likelihood by integrating the conditional likelihood function over this multivariate, “hybrid” discrete-continuous distribution above, the sample likelihood function is obtained by multiplying the unconditional likelihood function across households, L=
N
Li
(13)
i=1
where N is the total number of households in the sample. We obtain the estimates for our model parameters by maximizing the logarithm of this sample likelihood function. While our model specification is sufficiently general we believe, an alternative model specification can be to treat the scenario (i.e., SGCO) information as additive as follows.7 aijs = aij∗ + I SG δijs + ICO λijs (14)
where δijs ∼ MV N δ js , δ and λijs ∼ MV N λ js , λ . Note that aij is consistent with the standard random coefficient model; δ js is of dimension (J − 1) ∗ (SG − 1) where SG is the number of social groups; I SG is the 1 × SG indicator vector that takes the value of 1 if consumption is associated with one of social groups SG; δ is of dimension (J − 1) ∗ (SG − 1) × (J − 1) ∗ (SG − 1); λi is of dimension (J − 1) ∗ (CO − 1) where CO is the number of contexts; ICO is the 1 × CO indicator vector that takes the value of 1 if consumption is associated with one of consumption contexts CO; and λ is of dimension (J − 1) ∗ (CO − 1) × (J − 1) ∗ (CO − 1). Compared to our proposed hybrid
7 We
thank an anonymous reviewer for pointing out this alternative specification.
M. Kim, P.K. Chintagunta
model specification, this approach is not parsimonious as it requires many more parameters to estimate. Note that even if we rule out the possibility of the brand preferences for a specific consumption scenario s varying across consumers by defining aijs = aij∗ + I SG δ js + ICO λ js , this still requires more number of parameters than our hybrid model. To check whether our model can recover the true parameter values in an empirical setting, we simulate the choice data given the model parameters (estimated below), and the observed data structure. We consider a simple 1 segment version of our model where we further restrict our attention to consumers who only consumed the top 5 brands under the 4 largest SGCO situations. We generated 30 replications of the data. We find that the estimator does a very good job in recovering the underlying model parameters, with even the largest deviations being less than 1% from the true values. While it is likely that the error will increase as we add more parameters (our actual estimated model has many more parameters), we feel that the ability of the estimator to recover the model parameters even in this simplified setting is nevertheless reassuring. Complete details are available from the authors.
5 Empirical results As our base scenario, we chose the one that has the largest number of observations associated with it. In particular we chose the SG = 1; CO = 1 scenario—drinking alone, relaxing watching TV at home. This scenario has 8,047 of the total 19,739 observations in the data. Thus the results from our base scenario apply to the single largest group of observations in our data. Before we present the empirical results for our proposed model specification, we provide the following table to show fit measures such as log-likelihood (LL) and Bayesian Information Criterion (BIC) across various model specifications. We provide a model fit comparison, both in-sample and out-of-sample, from the various model specifications in Table 8. Model I is the standard random-coefficients model that does not use information on the social group/context scenarios; the second, model II, is model I with additive effects of social group and context scenarios (see Eq. 14). Model III is the proposed hybrid model. For our data we find that a 2-dimensional map (D = 2) with a 3-segment heterogeneity distribution for the map locations ( = 3) fits the data the best using the BIC criterion. Our results in Table 8 show a large improvement in the fit from using the proposed specification both in-sample (using the BIC criterion) and out-of-sample (using the predictive log-likelihood criterion). Table 9 provides the parameter estimates and their standard errors for the proposed specification. After controlling for other factors driving choices, it appears that consumers in the base scenario (i.e, when drinking alone relaxing and watching TV at home), on average have the highest intrinsic preference for Budweiser, followed by Corona, Miller Lite, Bud Light and Heineken—in that order. Note that this is in contrast with the actual shares in the sample as evidenced
Investigating brand preferences across social groups and consumption contexts Table 8 Model fit comparisons: in-sample and out-of-sample Model Specifications for in-sample fits (Number of people=1450, Number of observations=15615) I. Standard random coefficient model without social group & context (SGCO) informationa −LL Number of parameters BIC 10208.555 171 22068.284 II. Additive model with 5 social group and 5 context informationb −LL Number of parameters BIC 10093.513 251 22610.678 III. Hybrid discrete-continuous heterogeneity model exploiting 25 SGCO scenariosc −LL Number of parameters BIC 8452.392 302 19820.891 Out-of-sample fits (Number of people=389, Number of observations=4124) Model specifications Model I Model II Model III aA
−LL 3175.990 3213.303 2524.424
BIC 7775.483 8516.075 7562.871
random coefficient model is estimated with a full variance-covariance matrix
b Compared to the random coefficient model, we further estimate 80 more parameters for scenario
specific brand preferences in an additive way, i.e. (J−1)∗ (SG−1) + (J−1)∗ (CO−1) c This is the proposed model specification which is a random coefficient model along with the map structure. With social group & context (SGCO) information, brand map locations vary across segments and scenario specific attribute weights also vary across consumers Note: We only provide the model fits for the 3-segment models in this table as increasing the number of segments does not improve the BIC. The 2-dimensional location parameters are identified in this specification
by the information in Table 1 where the top 5 brands are Budweiser, Bud Light, Miller Lite, Coors and High Life. The strong preferences for Corona and Heineken despite their small shares are testament to the brand equity for these products. Next, looking at the covariance matrix of preferences for the base scenario, we find that the preferences for Bud Light are positively correlated only with those for Coors Light, indicating that the main competition for the former brand is from the latter. Coors Light on the other hand, competes with Bud Light, but also with Miller Lite and Corona. Similarly, Budweiser appears to be competing with all brands with the exception of Bud Light, Coors Light and Heineken. Thus Anheuser–Busch seems to have differentiated its two main flagship brands Budweiser and Bud Light in the minds of their consumers. In terms of the demographic effects, we note from Table 9 that Bud Light is preferred by younger, more rural, married males; whereas Heineken is preferred by young, urban, unmarried males. Thus, there appears to be considerable variation in the demographic profiles of each brand. Finally, from Table 9 we see that motivational factors like prices and availability only influence the smallest brand—Busch by raising its share. In a similar vein, using our results, we can compute the preferences under each of the other scenarios that we analyze in the paper. Table 10 provides the details of the advertising effects and the impact of the product characteristics.
Busch
Natural Light
Heineken
MGD
High Life
Corona Extra
Coors Light
Miller Lite
Budweiser
Variance in preferences Bud Light
Mean preferences
0.7467 (0.0704) −3.4529 (0.1668) −0.9939 (0.063) 1.6854 (0.0864) −0.7456 (0.0509) −2.9701 (0.1459) −2.4752 (0.1228) −0.4914 (0.0514) −2.8637 (0.1418) −3.0590 (0.1498)
3.1786 (0.1281)
Bud Light
−3.4529 (0.1668) 21.2157 (0.3998) 11.5524 (0.2698) −7.7338 (0.2115) 14.7152 (0.3004) 21.3931 (0.2977) 9.4377 (0.2372) −4.2201 (0.2795) 16.0158 (0.2819) 22.1974 (0.3104)
5.4338 (0.3184)
Budweiser
−0.9939 (0.063) 11.5524 (0.2698) 12.3114 (0.3377) 4.0236 (0.2693) 18.2368 (0.3027) 10.6211 (0.3034) 1.6593 (0.2142) −5.5578 (0.246) 10.3528 (0.2618) 13.4796 (0.3024)
3.2430 (0.1818)
Miller Lite
1.6854 (0.0864) −7.7338 (0.2115) 4.0236 (0.2693) 42.5247 (0.6836) 2.2864 (0.4059) −27.2089 (0.4271) 6.1256 (0.3359) −15.1734 (0.5317) −5.3159 (0.4129) −8.7941 (0.4065)
−1.5182 (0.1798)
Coors Light
−0.7456 (0.0509) 14.7152 (0.3004) 18.2368 (0.3027) 2.2864 (0.4059) 32.7390 (0.5423) 18.2649 (0.375) −0.4421 (0.2846) −4.1610 (0.4489) 15.1824 (0.3462) 22.7178 (0.3858)
3.5220 (0.2129)
Corona Extra
Table 9 Parameter estimates and (standard errors) from the proposed hybrid model
−2.9701 (0.1459) 21.3931 (0.2977) 10.6211 (0.3034) −27.2089 (0.4271) 18.2649 (0.375) 82.1092 (0.9679) 30.0169 (0.4848) −46.0524 (0.7392) −10.0003 (0.5574) 7.8501 (0.5625)
−1.0281 (0.2799)
High Life
−2.4752 (0.1228) 9.4377 (0.2372) 1.6593 (0.2142) −6.1256 (0.3359) −0.4421 (0.2846) 30.0169 (0.4848) 23.1481 (0.4355) −14.7188 (0.6297) −1.2898 (0.4065) −4.5911 (0.3664)
−1.2266 (0.2570)
MGD
−0.4914 (0.0514) −4.2201 (0.2795) −5.5578 (0.246) −15.1734 (0.5317) −4.1610 (0.4489) −46.0524 (0.7392) −14.7188 (0.6297) 138.4970 (1.5005) 62.9001 (0.7719) −4.8827 (0.7265)
0.6378 (0.3158)
Heineken
−2.8637 (0.1418) 16.0158 (0.2819) 10.3528 (0.2618) −5.3159 (0.4129) 15.1824 (0.3462) −10.0003 (0.5574) −1.2898 (0.4065) 62.9001 (0.7719) 94.7590 (1.4224) 38.2430 (0.6556)
−0.5312 (0.1882)
Natural Light
−3.0590 (0.1498) 22.1974 (0.3104) 13.4796 (0.3024) −8.7941 (0.4065) 22.7178 (0.3858) 7.8501 (0.5625) −4.5911 (0.3664) −4.8827 (0.7265) 38.2430 (0.6556) 88.3976 (1.4394)
−2.7097 (0.2040)
Busch
M. Kim, P.K. Chintagunta
Low Price
Motivational Factors Availability
Work
Married
Male
Rural
Non-white
Demographic effects Age
−1.4564 (0.0952) −1.6674 (0.1088)
−7.3092 (1.4961) 0.9438 (0.1098) 2.9203 (0.1021) 0.6427 (0.0929) 0.8791 (0.058) −1.029 (0.2293)
−1.9248 (0.1138) −0.814 (0.1241)
−4.4287 (1.5118) 1.3311 (0.1124) −0.8674 (0.0757) 2.7547 (0.0812) −1.6862 (0.0566) −2.6272 (0.2355) −1.7203 (0.117) −2.4198 (0.1072)
0.6862 (1.4629) 0.0392 (0.1136) −0.3192 (0.1118) −1.0351 (0.0896) 0.7378 (0.0542) −0.5169 (0.2227) −1.5651 (0.1395) −1.4089 (0.0989)
−1.5884 (1.5466) −0.0358 (0.1136) 1.9801 (0.1262) −0.6141 (0.1165) 1.2391 (0.0649) 0.1038 (0.2387) −0.3928 (0.1872) −0.9806 (0.1449)
−13.6102 (1.486) 2.2758 (0.1023) −1.8221 (0.0861) −1.3652 (0.0948) 1.0216 (0.0623) −1.4811 (0.2339) −4.3827 (0.0919) −4.4259 (0.1043)
6.2712 (1.4732) 4.1521 (0.1386) −2.3129 (0.0999) 0.0002 (0.0983) 0.0828 (0.0869) −3.8437 (0.2336) −3.2201 (0.1146) −4.1957 (0.095)
−0.9832 (1.5351) 2.4112 (0.1404) −3.5681 (0.1024) 2.1687 (0.1057) −0.2546 (0.0645) −2.3774 (0.2384) −5.1816 (0.2115) −1.2885 (0.1012)
−9.7421 (1.5151) 4.2368 (0.1429) −6.4287 (0.3613) 0.9113 (0.0909) −0.541 (0.0902) −2.7949 (0.2332) −2.4198 (0.1166) −5.1327 (0.097)
−0.8568 (1.5154) −1.6047 (0.1276) 4.610 (0.1492) 0.9102 (0.1046) −0.8165 (0.0823) −4.252 (0.2352)
0.8237 (0.1684) 1.633 (0.1101)
−12.5265 (1.5639) −1.5606 (0.166) 4.1261 (0.0891) 0.8461 (0.0952) −0.6303 (0.0785) −2.9453 (0.261)
Investigating brand preferences across social groups and consumption contexts
M. Kim, P.K. Chintagunta Table 10 Parameter estimates and (S.E.) for marketing variables and brand characteristics Local Ads Mean
−4.0082 (0.3885)
Network TV Ads
Alcohol
1.3046 (0.0595)
Variance-covariance Local Ads
2.0241 −3.9895 (0.3632) (0.3749) Network TV Ads −3.9895 9.7542 (0.3749) (0.4605) Alcohol 3.6067 −6.506 (0.4045) (0.5075) Calorie −4.6014 9.723 (0.7396) (1.1953) Demographic information for marketing variables Age Non-white Rural Male Interactions Local Ads Network TV Ads
−3.7113 (0.5778) −1.4270 (0.2276)
0.8403 (0.5500) −0.3137 (0.0635)
24.2181 (2.2385) −2.9235 (0.4866)
Calorie
−1.1289 (0.2709)
0.3162 (0.1294)
3.6067 (0.4045) −6.506 (0.5075) 9.0504 (0.850) −7.6993 (1.2067)
−4.6014 (0.7396) 9.723 (1.1953) −7.6993 (1.2067) 10.8843 (2.8137)
2.4804 (1.0992) 0.2053 (0.0555)
Married
Work
2.8948 (1.0411) −1.3501 (0.0680)
1.0934 (0.3410) −0.8219 (0.0732)
First, note that the characteristics in the data only vary by brand so these effects are being identified only off 11 observations. Second, while the intercept for the local ads is negative, the overall effect is positive at the “average” demographic profile. Table 10 indicates that a higher alcohol content lowers the choice of a brand whereas a higher calorie content is consistent with higher preferences. We also find that older consumers respond less to advertising whereas rural consumers are more sensitive to local advertising. Further, males tend to have a higher advertising sensitivity. In Fig. 1, we provide the locations of the brands along each of the D = 2 locations that capture the deviations of preferences from those under the base scenario for each of the = 3 segments. First, we see from the table that there is considerable heterogeneity in the preference deviation from the base scenario across the 3 segments. This implies that the mean preferences for brands are also likely to vary across the 3 segments. Next, from Fig. 2, we see that the importance weights also vary across the difference scenarios (the importance weights are along the X- and Y- axes whereas the variance component corresponding to each scenario is on the Z-axis). Together with the results in Fig. 1, this indicates that (a) preferences are different across consumers in a discrete way corresponding to the 3 segments and (b) that preferences are different across scenarios—the main point of this paper. Note however, that we also estimated the covariance matrix of preferences for the base scenario in Table 9 and an additional variance component for each scenario in Fig. 2. This implies that the covariance matrix of preferences are also scenario-specific following a continuous distribution. Together with the heterogeneity from the discrete distribution, our formula-
Investigating brand preferences across social groups and consumption contexts 7
MGD
6
Size of Segment 1: 0.2756
5 4 3
High Life
2 1 0 −6
−4
−2
Corona Extra Miller Lite
Budweiser
0
−1
Natural Light
Bud Light
Composite
Busch
2
4
6
8
Coors Light
−2
Heineken −3 −4 6
Budweiser
MGD 4
High Life 2
Natural Light Busch −14
−12
Corona Extra −10
−8
−6
−4
Composite 0
−2
0
2
4
6
−2
Bud Light
Coors Light −4 −6
Miller Lite
−8
Heineken
−10
Size of Segment 2: 0.2694
−12 1
Composite
Busch
Miller Lite
0 −4
−3
−2
−1
High Life
0
1
2
3
4
5
−1
Bud Light Corona Extra Coors Light
−2 −3
Heineken
Budweiser −4 −5
Natural Light
−6
MGD
−7
Fig. 1 Two-dimensional maps for 3-segment model
Size of Segment 3: 0.4551
M. Kim, P.K. Chintagunta
Fig. 2 Importance weights and variance component varying across the different scenarios
tion allows for a very flexible discrete-continuous heterogeneity pattern across consumers and scenarios. Once again, our identification comes from having data across the scenarios while pooling choice information across them. We reiterate however, that the resulting preference distribution while flexible is not as flexible as estimating a separate covariance matrix of preferences for each scenario since we do not have sufficient data for each scenario. Based on the variances represented along with the Z-axis in Fig. 2, there appear to be differences across scenarios in terms of the amount of within scenario heterogeneity that is present. We see that the variance is high in scenarios that involve eating a meal at home (SG = 2 & CO = 2; SG = 4 & CO = 2; SG = 5 & CO = 2) whereas the variance appears to be low in several scenarios involving consumption with a group of friends (SG = 5; CO = 3, 4, 5). Using the mean brand preferences computed from the estimates in Table 9, and Figs. 1 and 2, we next compute, based on average preferences in each scenario, the ranking of each brand within each scenario. We then present this ranking for a subset of scenarios in Table 11. We also distinguish this ranking for each of the 3 estimated segments in the data. Table 11 reveals several interesting findings. Overall, there appears to be considerable heterogeneity in preferences across scenarios as well as across consumer segments. First, we note that each of the brands—Bud Light, Budweiser, Miller Lite, Corona and Heineken—is the top ranked brand in at least one scenario and in at least 1 segment. Second, the Heineken brand stands out in the sense that while its overall share in the market is small (coming in at number 7 for our particular sample), for the customers in segment 2 it ranks in
Segment 1 SG=2 & CON=2 SG=2 & CON=4 SG=3 & CON=4 SG=4 & CON=2 SG=4 & CON=3 SG=5 & CON=2 Segment 2 SG=2 & CON=2 SG=2 & CON=4 SG=3 & CON=4 SG=4 & CON=2 SG=4 & CON=3 SG=5 & CON=2 Segment 3 SG=2 & CON=2 SG=2 & CON=4 SG=3 & CON=4 SG=4 & CON=2 SG=4 & CON=3 SG=5 & CON=2
Selected scenarios
4 3 4 1 4 5
1 4 2 2 2 1
3 1 2 1 2 2
4 3 4 5 4 5
1 2 1 3 1 1
Budweiser
3 2 1 3 2 3
Bud Light
2 3 3 4 4 3
2 1 3 3 1 2
1 1 3 4 3 2
Miller Lite
5 4 4 7 3 4
5 5 5 7 6 6
6 5 5 6 6 6
Coors Light
9 8 6 10 9 9
10 10 9 10 10 10
5 4 2 2 1 1
Corona Extra
Table 11 Brand preference rankings per social group/context scenario (3 segments)
4 6 7 8 8 7
8 7 8 8 8 7
7 6 8 8 9 8
High Life
7 5 8 2 5 6
7 8 7 6 7 8
2 7 6 5 5 4
MGD
11 10 10 9 10 10
3 2 1 1 3 4
10 10 10 10 10 10
Heineken
8 7 9 6 6 8
9 9 10 9 9 9
8 9 7 7 8 7
Natural Light
6 9 5 5 7 5
6 6 6 4 5 3
9 8 9 9 7 9
Busch
10 11 11 11 11 11
11 11 11 11 11 11
11 11 11 11 11 11
Composite
Investigating brand preferences across social groups and consumption contexts
M. Kim, P.K. Chintagunta Table 12 Average demographic profiles of the 3 segments
Demographic variable
Segment 1
Segment 2
Segment 3
Age Non-white Rural Sex (Male) Married Work
47.295 40.892 % 16.357 % 72.119 % 61.710 % 70.632 %
48.737 38.947 % 17.193 % 70.877 % 61.754 % 68.421 %
49.301 38.616 % 19.196 % 68.192 % 59.710 % 65.960 %
the top 3 for 5 of 6 scenarios in Table 11. This attests to the brand’s effective “niche” positioning in the US marketplace. As we see later in Table 13, segment 2 consumers tend to live in big cities such as New York, Houston, LA and Atlanta. Another brand with niche status seems to be Corona. The brand is ranked very low across scenarios in segments 2 and 3. However, for segment 1, under scenarios where consumers are working at home with 1 or 2 people or having a meal with a group of friends, Corona is the highest ranked brand. This information is likely very useful for brand managers who might want to reinforce the imagery corresponding to these scenarios in their advertising. Alternatively, given the high ranking for some other scenarios in segment 1, the brand might want to emphasize those situations to increase its market share. Importantly, our analysis in Table 11 indicates that even lowly ranked brands like Busch can identify situations—segment 2 when eating a meal with a group of friends—where the brand has relatively high preferences (ranked number 3). The combination of the information on the geographic locations of these segments along with the social group and consumption context that the brand is most closely linked with provides the basis for better targeting of scarce marketing resources for these brands. A question then arises is the following: In order to implement a targeted advertising campaign, it needs to be the case that local advertising elasticities of the brands in question allow for such campaigns. While not reported here, we find that Heineken and Coors have the largest advertising elasticities. So at least in the case of Heineken, focusing on its top segment and in specific scenarios seems like a reasonable
Table 13 Proportion of consumers belonging to each segment in top 10 DMAs Designated market area
Segment 1
Segment 2
Segment 3
Number of panelists
New York Chicago Los Angeles St. Louis Houston Denver Philadelphia Boston (Manchester) Cleveland-Akron (Canton) Atlanta
25.42% 18.00% 25.00% 15.22% 5.13% 23.08% 35.29% 18.18% 15.15% 12.90%
30.51% 12.00% 20.83% 8.70% 33.33% 12.82% 14.71% 18.18% 21.21% 19.35%
44.07% 70.00% 54.17% 76.09% 61.54% 64.10% 50.00% 63.64% 63.64% 67.74%
59 50 48 46 39 39 34 33 33 31
Investigating brand preferences across social groups and consumption contexts
strategy. Table 12 provides the demographic profile of each segment. Segment 3 appears to be the largest segment. While there are some differences across the groups, e.g., segment 1 customers are younger, urban, male, employed and tend to be more non-white, these differences are not very large across segments. Table 13 on the other hand, does point to differences in composition across cities—New York having the smallest proportion of segment 3 customers, followed by LA; whereas Philadelphia has the largest proportion of segment 1 customers. Houston has the largest proportion of segment 2 customers. The information in Tables 12 and 13 can help marketers fine tune the targeting strategies for their specific brands.
6 Conclusions In this paper, we set out to understand the preferences of consumers across various social group and context related consumption scenarios. Given the paucity of data on some of these scenarios, we extended the extant literature on continuous random coefficients logit brand choice models to allow us to understand the nature of these preferences as well as to accommodate heterogeneity in preferences across consumer segments. To accomplish this, we augment the standard random coefficients model across consumers for a “base” scenario with a parsimonious representation of the deviations in preferences of products across scenarios from this base scenario. The latter representation is based on projecting the deviations onto locations on a low dimensional “map” along with scenario-specific weights for the dimensions of the map. Additional heterogeneity in preferences across consumers and scenarios is accommodated by allowing the locations to be different across consumer “segments” and scenario-specific importance weights to be heterogeneous across consumers. We also incorporate the effects of advertising in our analysis. Our empirical results from the beer category revealed the following insights. The data suggest that our proposed model fits the data better than a standard random coefficients logit model with a continuous heterogeneity distribution. The improvement in fit is even after we penalize it for the larger number of parameters estimated in the proposed model. Further, the proposed model also fits the data better than an alternative approach to accounting for the social group and context information. This suggests that marketers should indeed consider how consumers’ preferences can vary across social group and consumption context scenarios when designing their marketing programs. Focusing on our base scenario—drinking alone, relaxing watching TV at home—we are able to estimate the intrinsic brand preferences as well as the nature of competitive effects across brands in this scenario. We find strong preferences for Corona and Heineken despite their small shares. We find that Bud Light “competes” with Coors Light, although the latter competes with Bud Light, but also with Miller Lite and Corona. Motivational factors like
M. Kim, P.K. Chintagunta
lower prices and better availability only increase the share of the smallest Busch. Importantly, we find considerable heterogeneity in preferences across scenarios—providing a justification for the improved fit of the model that allows for such heterogeneity as well as emphasizing the need to account for such heterogeneity in marketing applications. Our results appear to be particularly encouraging for smaller brands like Heineken as customers in cities like New York, Houston, LA and Atlanta rank it in the top 3 brands for several scenarios. Not only does it attest to the brand’s effective “niche” positioning in the US marketplace; but combined with the higher sensitivity to local advertising for that brand, it suggests an approach for effective and efficient marketing of this brand. Similarly, Corona also appears to have a niche status with consumers. Allowing for brand preferences to vary across scenarios therefore, provides useful inputs to the brand managers of these products. Indeed, even for the smallest Busch brand, we find a scenario in which its relative preferences are high. In summary, we propose an extension to the extant literature on estimating brand maps from consumer data to allow for differential preferences of consumers across social group and context scenarios. We obtain several implications for the brands competing in the beer marketplace. We hope future research can further enhance our understanding of these preferences that can enable managers to better understand and to better market their brands. Acknowledgements We thank Karen Garvin and David Miller of Research International and Rafael Alcaraz (now at Hersheys) and Gary Fehlhaber at MillerCoors for the data. The authors also thank Sha Yang for her feedback, and the Editor and 2 anonymous QME reviewers for their extensive comments.
References Allenby, G. M. (1990). Hypothesis testing with scanner data: The advantage of bayesian methods. Journal of Marketing Research, 27, 379–389. Allenby, G. M., & Rossi, P. E. (1999). Marketing models of consumer heterogeneity. Journal of Marketing Research, 30(2), 171–182. Allenby, G. M., & Rossi, P. E. (2009). Bayesian Applications in Marketing. Working Paper, University of Chicago, Booth School of Business. Belk, R. W. (1974). An exploratory assessment of situational effects in buyer behavior. Journal of Marketing Research, 11, 156–163. Belk, R. W. (1975). Situational variables and consumer behavior. Journal of Consumer Research, 2(3), 157–164. Bettman, J. R., Luce, M. F., & Payne, J. (1998). Constructive consumer choice processes. Journal of Consumer Research, 25(3), 187–217. Carlson, K. C., & Bond, S. D. (2006). Improving preference assessment: Limiting the effects of context through pre-exposure to attribute levels. Management Science, 52(3), 410–421. Chintagunta, P. K. (1994). Heterogeneous logit model implications for brand positioning. Journal of Marketing Research, 31, 304–312. DeSarbo, W. S., Atalay, A. S., Lebaron, D., & Blanchard S. J. (2008). Estimating multiple consumer segment ideal points from context-dependent survey data. Journal of Consumer Research, 35, 142–153.
Investigating brand preferences across social groups and consumption contexts Dube, J. P., Hitsch, G., & Rossi, P. E. (2009). State Dependence and Alternative Explanations for Consumer Inertia. Working Paper, University of Chicago, Booth School of Business. Dickson, P. R. (1982). Person-situation: Segmentation’s missing link. Journal of Marketing, 46, 56–64. Elrod, T. (1988). Choice map: Inferring a product-market map from panel data. Marketing Science, 7, 21–40. Elrod, T., & Keane, M. (1995). A factor-analytic probit model for representing market structure in panel data. Journal of Marketing Research, 32, 1–16. Fennell, G. (1978). Consumer’s perception of the product use situation. The Journal of Marketing, 42, 38–47. Guadagni, P. M., & Little, J. D. C. (1983). A logit model of brand choice calibrated on scanner data. Marketing Science, 2, 203–238. Hausman, J. (1996). Valuation of new goods under perfect and imperfect competition. In T. Bresnahan, R. Gordon (Ed.), The Economics of New Goods, Studies in Income and Wealth (Vol. 58). Chicago: National Bureau of Economic Research. Lee, J. K. H., Sudhir, K., & Steckel, J. H. (2002). A Multiple ideal point model: Capturing multiple preference effects from within an ideal point framework. Journal of Marketing Research, 39, 73–86. McFadden, D. (1973). Conditional logit analysis of qualitative choice behavior. In P. Zarenmbka (Ed.), Frontiers of Econometrics. New York: Academic Press. Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica, 69(2), 307–342. Petrin, A., & Train, K. (2006). Control Function Corrections for Omitted Attributes in Dif ferentiated Product Models. Working Paper, University of Minnesota, Twin Cities. Petrin, A., & Train, K. (2010). A control function approach to endogeneity in consumer choice models. Journal of Marketing Research, 47(1), 3–13. Ratneshwar, S., & Shocker, A. (1991). Substitution in use and the role of usage context in product. Journal of Marketing Research, 28, 281–295. Yang, S., Allenby, G. M. & Fennell, G. (2002). Modeling variation in brand preference: The roles of objective environment and motivating conditions. Marketing Science, 21(1), Winter, 14–31.