الگوهای منطقه ای و طبقه ای در رفتار مصرف کننده : افشای روندها

We offer a method of analysis that allows for an “unbundling” of the data to a disaggregate household level, and then “rebundling” it in a manner designed to identify patterns and relationships which are otherwise masked. Applying the method in the context of ‘healthy’ products and using census block group level data, we study consumption over several categories in two locations. The analyses show that studies involving geographically dispersed data must test for, and take into account when required, conceptually sound spatial effects in order to accurately assess impact. We also show that while both location and the product category have a significant impact on the proportion of healthy products purchased, the degree to which consumers choose healthier alternatives is a function of the category as well as the location. Finally, we provide preliminary evidence from survey data that supports the variations we find, and further explores attitudinal differences as well. There are rich implications for retailers in that new products introduced to benefit from popular trends (such as ‘healthy’ alternatives) may not succeed for all categories or locations. Retailers would benefit from understanding the spatial, demographic and attitudinal effects that play into consumption behavior, and such effects can be better understood when studying choice at the category and region level. Finally, public policies aimed at promoting healthier purchasing habits may have greater impact if special attention is given to specific categories and regions.

Quantitative research studying consumption behavior in marketing has often explored empirical and theoretical issues using aggregate level data, usually from households or retailers. In both cases, the data may be aggregated in several ways. The common practice has been to collect data at the individual or store level and subsequently sum it over various stores, chains and households/neighborhoods for aggregate level analyses – often some form of regression. This gives rise to two issues. Firstly, any trends or consumption patterns that are specific to a product category over neighborhoods may become unobservable at the aggregate level when data is thus summed. For example, the baby food market in 2006 saw a fairly flat sales growth from 2005 to 2006 and only a 3.1% increase in 2007. Over these years, new product introductions were also fairly flat in the baby food category. A closer look at the disaggregated data reveals however, that in the same time period, the growth of organic baby food sales saw a 16.4% jump followed by a 21.6% jump in 2007 (Martinez 2010). Note that the popularity of the ‘organic’ attribute in baby food did not spill over to potato chips or other snack foods (Cabbage Patch Kids, 2007 and Zegler, 2006). Thus, even as manufacturers and retailers recognize consumer demand for certain types of products, analysis using the traditional aggregations over products or categories may fail to offer insights which would otherwise be observed. The second issue is methodological in nature. Traditionally, such data analysis rarely required information of a spatial nature (that is, relating to the ‘spatial arrangement’ or the distance between observation points). While such a spatial thrust is quite common to modeling in the study of epidemiology or agriculture, it has received far less attention in marketing (Allaway et al., 2003, Bradlow et al., 2005 and Gonzalez-Benito et al., 2005). In fact, models using distance based data such that spatial autocorrelation is explicitly accounted for, may not have been an option in the past, when the geographical detail required was lacking in the kind of data used typically by marketing researchers. However, as more and more research that studies patterns of behavior over smaller unit areas is undertaken and data is collected from adjoining units, there may very well be effects such as word-of-mouth that lead to patterns in variables such as satisfaction (Mittal, Kamakura, & Govind 2004), resulting in autocorrelation across these areas as a function of distances from each other (note that we undertake a more detailed discussion of the specific factors leading to spatial effects here in the subsequent ‘Modeling’ section). To further complicate matters, a typical marketing data set may also include variables such as product characteristics that are independent of such spatial correlations and impervious to aggregation issues. It becomes essential then, to use a methodology that accommodates such spatial and non-spatial variables so as to avoid potential bias and misleading results. Indeed, if some trends in the market are only reflected in certain product categories, and if the diffusion of this trend has spatial properties, clearly there is a need for a change in the way such data is analyzed. We thus propose a regeneration of the data by first disaggregating consumption data at the overall household level into consumption at the specific category levels (see Fig. 1). Once the category level sales have been determined, an aggregation by category over census block groups1 will yield the data set required to study the impact of the different variables on category consumption. It should be noted that this data set is readily available to retail chains – indeed, they collect this data – the issue arises only with respect to aggregating it in a different way (that is, by category and by census block group). This methodology of disaggregation and then ‘re-aggregation’2 of data begins to address the issue of how to use the volumes of scanner data now available to the retailer and can lead to critical implications for retail strategy in several areas such as line extensions, product distribution and promotional strategy.We apply our methodology to a particularly appropriate trend that has gained momentum over the last 10 years, perhaps as a function of shifting demographics; that is, the health trend. Consumer concern with health and healthy products is at an all time high with various entities involved in this arena. As government agencies and medical associations announce their recommendations for improving quality of life, manufacturers and retailers respond with new product introductions that purportedly follow those recommendations. It should be noted that what constitutes a ‘healthy’ alternative is of course, not always agreed upon. For example, the use of artificial sweeteners in diet drinks may not be seen as a healthy alternative to the regular product. We argue, however, that for individuals concerned with the caloric aspect of consumption, it is indeed touted by the various entities including the media, as a better alternative. We thus focus on the consumer's point of view – that is, is there an alternative available in a given category that offers a ‘healthier’ – in this case, a lower calorie or fat content – version? While the absolute ‘healthiness’ of these choices may be debatable, they clearly indicate an effort on the part of the consumer to choose what they perceive as a less harmful or ‘healthier’ version of a product. Moreover, with federal guidelines in place regarding the labeling of such products,3 not only do consumers have a clear signal in choosing products, we also have a standard benchmark for tagging the ‘healthy’ alternatives for inclusion in the product definition. (For the purposes of this paper, then, please note that we refer to the lower calorie or lower fat version of the product as the ‘healthy’ alternative.) Whether the consumer is responding to these messages, or even receiving them, however, is an issue that is much debated by researchers and practitioners alike, and needs further exploration. Thus, while such products appear to proliferate in the market place, researchers also show that this proliferation does not necessarily translate into purchasing and consumption (Balasubramanian and Cole, 2002 and Krozup et al., 2003). National health trends, in fact, show an increased incidence of diseases such as diabetes and obesity over the same time period (Amann 2004). With so many sources, including managers, manufacturers and government agencies, propagating health benefits, it is somewhat of a surprise that there is so little evidence supporting a stronger trend towards healthy consumption (Thompson 2004). It is significant to note, however, that much of the research conducted to date in this area has used either aggregate level data or has been based on experimental or self reported survey data,4 as opposed to actual purchase data. While such surveys have proved invaluable in targeting a specific audience (by distributing questionnaires strategically), and obtaining attitudinal and perceptual input, the information regarding consumption is often of little use in reflecting actual purchase behavior for a variety of reasons. First, since the responses often rely entirely on the respondent, the questions may elicit responses that are too general or even erroneous (Stobbe 2006). Second, since such responses are often seen as a reflection on themselves, they may well be subject to bias arising from issues of impression management (Leary and Kowalski 1990). Finally, as we posit earlier, consumption data collected at an aggregate level may very well mask any differences in category level patterns. This would certainly help explain not only the disparity between the increasingly persistent messages regarding healthy diets and the simultaneous increase in the prevalence of consumption related diseases, but also the very marginal success that public policy has enjoyed in its very broad effort to impact public choice behavior. In this context then, if the current trend in healthy consumption does not spill over to other categories or neighborhoods, then the introduction of healthy alternatives must be considered more strategically than has been the practice so far. Several interesting research questions arise as a result. Firstly, are consumers actually purchasing healthier products? Secondly, if some of them do actually consume more healthy food options, is such behavior relegated to certain food categories so that the aggregation of data then masks any changes in consumption patterns? Finally, could consumption within categories then be a function of ‘geodemographics’ – that is, a combination of the neighborhood characteristics and demographic variables (Johnson 1989) – requiring perhaps an explicit incorporation of correlations across areas to appropriately model consumption without bias? Resolving such issues regarding consumption behavior will have a significant impact not only in terms of our understanding of consumer consumption at the micro level but also in terms of category and area manufacturer and retailer strategy. Results show that consumers differ significantly not only in their purchasing patterns across neighborhoods but across categories as well. Some categories (such as milk) seem to elicit stronger trends towards healthy consumption while others (such as ice cream) show a significantly weaker pattern of purchasing alternatives that are deemed healthier. Furthermore, not only do we see differences in patterns across the nature of locations, but the degree to which we find such differences also varies across categories. It is interesting to note that as posited while this difference shows up to only a small extent when considering aggregate level purchases, it is accentuated when the analysis is done at the category level. We thus contribute to the literature in several ways. Firstly, we offer insights with respect to data aggregation issues. That is, if studying the data at the usual aggregate levels (usually store or area) masks category and location specific impacts, then clearly the data needs to be analyzed at a different level of aggregation in order for underlying patterns to be observed. We thus offer a means of re-aggregating store level data to offer specific category and neighborhood level insights into consumption behavior. Note that this method also serves to avoid the bias in parameters caused by combining data at multiple levels of aggregation (Jong, Ruyter, & Lemmink 2004) – the final data we use is in fact at a uniform level of aggregation. Secondly, we show that when observation points are in close physical proximity, there is a correlation pattern among them that is a function of the distance between two observations. This is particularly true when we consider sales from contiguous geographic units (census block groups in this case) rather than from households as is typical of most marketing studies. This makes it imperative to consider the spatial nature of observations as is traditionally done in fields such as epidemiology or geography which deal with such contiguous data. We thus offer a conceptual argument for a specific spatial process – by accommodating this we avoid estimation bias and potentially misleading results. Finally, in an exploratory attempt to unearth some of the attitudes and reasons behind such variations, we conduct a survey that collects data regarding the issues of our study. Firstly, we explore the basic differences in attitudes towards health issues that across locations, and secondly, we explore the possibility of word of mouth effects which may be at the root of the spatial effects we discover. We find that our analysis strongly supports our findings of neighborhood differences discussed above, which in turn is explained in their responses to attitudinal questions. Furthermore, there is also some evidence that word of mouth effects could very well play a role in the purchase of healthy products. Clearly, there is significant scope for research combining data from actual consumption as determined from the scanner data, and attitudinal data as obtained from surveys. In the next section, we discuss the conceptual foundation of the spatial processes to be accommodated and the context within which our study takes place. We then discuss our modeling framework and how these two issues impact it. We follow this with the methodology as well as the nature of the data set being used. Next comes a discussion of results and a section on managerial and other implications. We finally conclude with a section on the limitations as well as possible extensions of our work.

In conclusion, we offer a method of analysis that allows for a process of using disaggregate household level data, and then aggregating it in a manner designed to identify patterns and relationships which are otherwise masked. The analyses here also show that studies involving geographically dispersed data can apply the appropriate spatial econometric tools to test for their presence (as opposed to using an ad hoc method). This issue of identifying the correct spatial process is an important one, since each spatial specification produces different interpretations and consequently, different policy implications. Our technique of using category specific data at the census block level, confirming the presence of spatial effects accounted for by a lagged spatial model and finally using the appropriate model to estimate parameters, offers one way of achieving a greatly improved fit. By using census block level data instead of zip codes, as has previously been done, we are able to offer analysis at a finer level of distinction. This also enables us to use the census data collected at the block level (Gauri, Pauler, & Trivedi 2009) and offers the benefit of using a much larger data set in order to estimate our model. Finally, our preliminary work with survey data offers support for our results, and more importantly, provides rich insights into the reasons and attitudes for such purchasing patterns. This area of research may benefit from a more in depth study of the level of aggregation – both category and neighborhood level – that would be sufficient to identify patterns of healthy purchasing behavior. Too much disaggregation would produce unnecessary noise while too little may mask patterns. Note that in our case, as displayed in Fig. 1, while we do aggregate the data in a specific way so as to arrive at the required data set, the analysis itself is done only at a single level – the census block group level. If we are, however, to study the issue of modeling simultaneously at several levels, this approach would have to be modified in order to accommodate a multilevel or nested model formulation using techniques such as Hierarchical Linear Modeling (see Jong et al. 2004, among others). We leave this issue for future researchers to explore. Another area rich with potential for future research is a more in depth exploration of the attitudes towards healthy consumption not just across a wider set of consumers, but also at a category specific level. While we explore some of these issues independently in the survey analysis reported here, clearly there is significant scope for research combining data from actual consumption as determined from the scanner data, and survey data. Finally, this study offers plausible explanations for why public policy has seen only marginal success in its effort to impact public choice behavior. If the impact is measured at an aggregate level, any differences in purchasing patterns that are category specific will tend to get masked. Healthy consumption in one category does not imply healthy consumption in all categories, and certain elements of healthy consumption may need greater emphasis in certain neighborhoods. As a result, any effective policy at the very least must be implemented by area and may prove even more effective by concentrating on a few select categories. Such a decomposed method of studying consumption would also allow managers to understand consumer behavior at a finer, more segmented level, thereby allowing a marketing strategy that could be tailored to category and consumer specific segments. Such a strategy would benefit not only manufacturers and retailers in drawing a greater response from the relevant segments, but also public policy makers constantly seeking ways to improve consumer decision making.