Patterns of Product Assortment and Price-Cost Margins across the Food Retailing Landscape

Transcription

1 Patterns of Product Assortment and Price-Cost Margins across the Food Retailing Landscape Chenarides, Lauren 1 ; Jaenicke, Edward C. 2 ; Volpe, Richard J. 3 1 Department of Agricultural Economics, Sociology and Education, Penn State University, State College, PA, 16803, U.S., lec201@psu.edu 2 Department of Agricultural Economics, Sociology and Education, Penn State University, State College, PA, 16803, U.S., ecj3@psu.edu 3 Department of Agribusiness, CalPoly, San Luis Obispo, CA, 93407, U.S., rvolpe@calpoly.edu Paper prepared for presentation at the EAAE-AAEA Joint Seminar Consumer Behavior in a Changing World: Food, Culture, Society March 25 to 27, 2015 Naples, Italy

2 Abstract In a late-modern society where people and societies may be at increased risk from economic crashes, environmental catastrophes, and other consequences of increased globalization, consumers who reside in food deserts may be particularly vulnerable. These consumers, already burdened with poor food availability, might actually face additional hardships such as purchasing foods with higher price mark-ups or facing limited or less healthy product assortments. Current research generally shows that traditional supermarkets have chosen not to enter food deserts because of insufficient underlying profit-maximization conditions and, as a result, mini-mart and convenience stores are disproportionally prevalent. However, no research clearly documents the geographic patterns of product assortments or price-cost margins in or near food deserts that might result from store location decisions. This paper, which investigates food access and underserved areas in the U.S., differs from existing research by explicitly examining food retailers decisions that might exacerbate or mitigate consumer impacts of residing in food deserts. Topic: "New trends and directions in food consumption" Keywords: food retailing, food deserts, micro data, food inequality This project was supported by the Agriculture and Food Research Initiative Competitive Grant No from the USDA National Institute of Food and Agriculture.

3 Introduction The USDA's Economic Research Service estimates that 23.5 million people reside in areas with severe food access limitations, and nearly half of these households are low-income (Ver Ploeg, 2010). These food deserts are defined as neighborhoods without access to fresh, healthy, and affordable food (AMS, 2014). Several definitions exist for the term food desert, yet researchers collectively agree on two common standards that identify these underserved areas: (i) households living in these communities are unable to access food retailing outlets that sell healthy and affordable foods, and (ii) the demographic concentration within these communities are disproportionately comprised of socioeconomically disadvantaged households (e.g., Beaulac, Kristjansson, and Cummins, 2009; Bitler and Haider, 2010; Ver Ploeg, 2010). Due to the nuanced variability of the term food desert among researchers, other terminology may also be used to describe underserved communities, such as areas that are susceptible to access problems or areas with low food availability. Throughout this paper, we will be incorporating the following terminology: food access is used to describe the allocation and affordability of food linked with consumer preferences (i.e., used as an indication of whether an area may be associated with food deserts) and food availability is used to describe the availability of food supply (i.e., from supermarkets and large grocery stores, or the availability of affordable, nutritious food). The goal of this paper is to investigate the impact of poor food access by specifically looking at product assortment and price mark-ups as a result of retailer's short-run strategic decisions. The deficit of research that examines the retailer's perspective could result in the mischaracterization of the full set of hardships stemming from food deserts. For example, consumers who reside in underserved areas already burdened with poor food availability might actually face additional hardships, such as purchasing foods with higher price mark-ups or being restricted to purchase items from a smaller set of product assortment. Once a retailer's perspective is included, a number of straightforward, but currently unanswered, research questions emerge: Do firms located in underserved communities have higher markups and/or offer lower product assortments? What is the relationship between observed product assortments (or price-cost margins) and market-level demographic attributes? Despite the importance of this topic and questions such as these, little or no research exists that documents how the distribution of consumer types and geographic patterns are associated with these marketing strategy outcomes. This paper differs from existing research by

4 explicitly examining food retailers decisions that might exacerbate or mitigate consumer impacts of residing in underserved areas in the U.S. Existing research on food deserts mainly focuses on consumer well-being, health outcomes, and purchasing behaviors. The density of non-traditional food outlets within food deserts is high, and current research claims that consumers are faced with higher prices, scant product selection, and potentially a lack of nutritious products (Alviola et al., 2013; Blachard and Matthews, 2007). Current studies examine how food access differences may contribute to health disparities, linking higher BMI and obesity with the prevalence of convenience stores and fast food restaurants (Chen et al., 2009; Larson et al., 2009; Ver Ploeg, 2010 & 2012). While consumer impacts from limited access have been at the forefront of many research agendas, it is important to also recognize the evolution of the food retailing landscape. No longer comprised of individually owned and operated grocery stores, the food retailing industry is concentrated with large-scale chain supermarkets and supercenters, and now even other types of store channels (e.g., drug stores) have begun offering food products. Although the efficiencies in transportation and other operational procedures have aided the distribution of food across the U.S., food insecurity has increased (EFSNE, 2013). In underserved areas, high entry costs make it undesirable for chain grocery stores to enter these markets. Reflecting entry, exit, and equilibrium concepts, some economic research investigates the behavior of the food retailing industry and product differentiation to show that the food retailing landscape is the equilibrium outcome of certain supply and demand factors (Bonanno et al., 2012; Ellickson, 2007). In these models, extreme food unavailability, or food deserts, is one such equilibrium outcome, where the most significant indicators of food deserts and other landscape outcomes is the uneven dispersion of consumer types. Beginning with a national-level investigation of the U.S. food retailing landscape, this paper is aimed at identifying geographic patterns of price-cost margins and product assortment, highlighting those within food deserts. After calculating census tract-level product assortments and price-cost margins, we then specify a reduced-form model that aims to estimate the effect of market-level characteristics, including demographic data, on our calculated margins and assortments. We find that a socio-demographic characteristics consistently play a role in explaining levels of product assortment, while areas with high portions of the population living in food deserts have a negative and significant effect on product assortment.

5 Related Literature Food Access The USDA's 2009 Report to Congress finds that households across the country face barriers that prevent them from accessing stores that sell affordable nutritious foods (Ver Ploeg, 2010). While this report outlines several poignant findings, it is not the first instance when the issue of food insecurity has been documented or addressed. Since the 1960s, researchers have investigated the effects on individuals of living in areas that lack stores selling affordable nutritious food, now commonly referred to as food deserts (Alviola et al., 2013; Alwitt and Donley, 1997; Beaulac et al., 2009; Kyureghian et al., 2013; Ver Ploeg, 2010; Zenk et al., 2005). Much of this research focuses on how the lack of food availability affects consumer health outcomes and purchasing behavior. In areas with poor food access, convenience stores have higher densities and inevitably serve as non-traditional food stores. These stores have higher prices, less product selection, and fewer nutritious products (Alviola et al. 2013; Blanchard and Matthews 2007). The local food environment has been linked with households' diets, suggesting that changes in the environment may spur positive changes in shoppers' nutritional choices (Moore and Diez Roux, 2006). Researchers are still investigating how poor food access may be linked with obesity, diabetes, and other health-related conditions; however, patterns of health concerns (e.g., adult obesity and diabetes) and poor food access may be interrelated. Current research also aims to link consumer demographics in underserved areas to the retail food environment and shopping behaviors. For example, research shows that income has a statistically significant positive effect on fruit and vegetable purchases as well as average store size (Dunkley et al. 2008; Kyureghian et al. 2013). Numerous studies across the U.S. consistently find that supermarkets are less prevalent in poorer areas, while fast-food restaurants appear in more concentrated numbers (Alwitt and Donley, 1997; Moore and Diez- Roux, 2006). The combination of a high density of fast-food stores and the migration of supermarkets to suburban areas may in fact contribute to the disparities in choices among households living in underserved communities. Compared to convenience or corner grocery stores, supermarkets sell a wider range of higher-quality products at lower prices, yet many households living in poor access areas do not have the transportation means to drive to the nearest supermarket. Households that are better-off may reside in low-income areas; however, these households are more likely to own a car, so traveling to a supermarket outside of their

6 immediate neighborhoods is not considered outside of their means, and are therefore able to escape the food desert in which they live (Ver Ploeg, 2010). With the recent funding cuts to the Supplemental Nutrition Assistance Program (Food Stamps), many low-income households across the U.S. who have benefited from government subsidies will be affected. Despite these cuts, the current administration cites limited food access as a focal point in rising obesity rates (Larson et al., 2009; Lee, 2012). In communities that do not have access to supermarkets or other traditional food outlets, households are burdened with the task of shopping at convenience stores and fast food chains to meet their dietary needs, thereby contributing to an unhealthy lifestyle. In addition, the average cost of healthier market baskets using the USDA s Thrifty Food Plan (TFP) in underserved communities are found to be higher compared to the average cost of a standard market basket resulting in an extra 35% of a household s annual food budget (Jetter and Cassady, 2006). In a subsequent report to Congress, the USDA finds few changes to food access (i.e., the opening of new supermarkets) over a recent four-year period from 2006 to 2010 (Ver Ploeg, 2012). According to the 2012 USDA report, the number of supermarkets between 2006 and 2010 actually decreased, while supercenters and large grocery stores moderately increased. The report highlights supermarket availability as an indicator of food security, implying that consumers who have access to supermarkets will be able to better meet the dietary needs of their household. Some states, such as Pennsylvania, directly test this assumption via a pilot study through the Pennsylvania Fresh Food Financing Initiative that evaluates the impacts of opening a new supermarket in Philadelphia. The study finds that, although there is increased access, shoppers do not markedly change the amount of fruit and vegetable consumption (Cummins et al., 2014). This finding supports an earlier national-level study, which indicates that the density of supermarkets in urban areas does not have a significant effect on household fruit and vegetable consumption (Kyureghian et al., 2013). This recent evidence suggests that accessibility alone is not necessarily a solution to addressing concerns about diet and nutrition. From the consumer's perspective, food access issues are a multidimensional problem. As more research is conducted, the results indicate that other forces may exist within underserved communities that are preventing households from incorporating higher-quality products into their market baskets. Rather than qualifying this issue as one dictated by a lack of access, researchers have suggested that it might be an issue of ease of access (Lee, 2012). While low food availability within the food retailing landscape may indeed be one force, the larger question may be how accessibility and marketing outcomes interact to influence

7 consumer purchasing behavior. Given the current food retailing landscape, a goal of our research will be to investigate what other forces may be impacting consumers' decisions by looking correlations between the geographic dispersion of consumer demographics on product assortments and price-cost margins chosen by food retailers. The Food Retailing Sector Over the past few decades, the introduction of new food retailer formats such as supercenters and club stores has significantly changed the landscape of the food retailing industry. For example, Wal-Mart supercenters have had one of the fastest growing grocery departments. In 1999, Wal-Mart ranked number five in total U.S. grocery sales and, as of 2011, became the top grocery retailer in the U.S. and Canada (Kaufman et al., 2000; Supermarket News, 2013). The supercenter s rise over the traditional food outlets can be attributed in large part to the innovation of automated distribution and procurement systems (Ellickson, 2004). Although innovation brings with it sunk costs, the investment in new technology for large chain stores means reduced costs in the long-run due to better tracking mechanisms of their inventory, as well as the expectation that stores could offer more products to their patrons. The line of current economic research that investigates the competitive behaviors of food retailers shows that the food retailing landscape is the equilibrium outcome of supply and demand factors (e.g., Shaked and Sutton 1987; Sutton 1991; Ellickson 2007; Ellickson and Grieco 2013; Bonanno, et al. 2012). Extreme food unavailability in a localized market, i.e., a food desert, is one such equilibrium outcome. In markets where access is limited, retailers may not have an incentive to overcome such high fixed costs and therefore choose to locate in markets with more stable demand. If the demand potential is low, retailers may not be willing to participate in certain economies. In these equilibrium models, in particular those which highlight food accessibility, the most significant indicators of food deserts and other landscape outcomes is the uneven dispersion of consumer types (e.g., Ellickson, 2007; Bonanno, 2012). Within underserved markets, access to larger food stores often entails higher transportation costs, especially for low-income individuals. Research examining economizing practices for poor households has indicated that low-income shoppers try to take advantage of volume discounts and product promotions; however, given the higher search costs they face, these households are often unable to take advantage of the benefits of shopping at larger format stores, such as supermarkets and discount merchandisers, which tend to locate in the suburbs or higher-income areas (Leibtag and Kaufman, 2003). Low-income households,

8 whose presence is more concentrated in rural and urban regions, are faced with shopping at the smaller food stores where food prices tend to be higher (USDA, 1997). Despite the increased differentiation of store formats over the past three decades, the food retailing industry is still dominated by a small number of large-scale food retail stores, leaving a small portion of the market to be dispersed among smaller fringe grocery stores. As a result, the food retailing industry has separated itself into two separate tiers: a highquality segment and a low-quality segment (Ellickson, 2004). If high-quality firms want to gain profits, they must rely on less price sensitive consumers, so a fewer number of these stores will enter low-income markets (Bonanno and Lopez, 2009). On the other hand, the lowquality stores, which can be characterized as glorified convenience stores, do not invest in the same technologies as supermarkets or large chain retailers, and, unlike the high-quality firms, continue to grow in number as the market grows. In food deserts, where the presence of lowquality stores is more prevalent, these fringe stores may face lower entry costs due to their smaller size. Consequently, due to space limitations, the portion of selling space dedicated to product assortment of fresh fruits and vegetables is compromised (Ver Ploeg, 2010). The homogeneity of the food retailing landscape in food deserts may compound the hardships faced by these households. Models of Product Assortment and Price-Cost Margins The significance of product assortment within the context of food retailing can be seen in the literature that examines consumer preferences for variety in retailing, and, by extension, research on store choice models. In this literature, it is less common to find consistent specifications of product assortment; however, analysis looking at assortment is most often done at the category level (Julander 1992; Briesch et al. 2009; Hwang et al., 2010; Mantrala et al., 2009). Traditionally, product assortment is considered strictly as the number of Stock Keeping Units (SKUs) or Universal Product Codes (UPCs) offered within a product category (Broniarczyk and Hoyer, 2010). Early research finds that consumers generally prefer a store that offers several alternatives of the same type of product, as consumers are able to spread their fixed costs of shopping over a larger number of goods (Baumol and Ide, 1956). However, more recent research concludes that more is not always better. With the availability of store-level scanner data and store characteristics data, alternative dimensions of product assortment can be measured. These measures can account for the selling space of a retailer, the amount of space dedicated to certain product categories, similarity of items, and the number of size options available for each brand (Briesch, et al., 2009). Expanding on the

9 existing assortment research, our analysis will follow similar logic in specifying alternative definitions of product assortment, accounting for the presence, variety, and selling space of food products. Extensive research can be found in the industrial organization literature on price-cost margins (PCM) analysis. Beginning with models of Structure-Conduct-Performance (SCP) as seen in the work of Demsetz (1973), Clark, Davies and Waterson (1984), and Cotterill (1986), PCMs were originally used as outcome measurements of market concentration and market power. However, the criticisms of the reduced-form SCP technique led to the introduction of a more empirically-driven approach known as NEIO, or new empirical industrial organization. The SCP paradigm often left unresolved the issue of whether PCMs were a result of market power or market efficiency. For many researchers, the NEIO method provides a more analytical technique at estimating the relationship between structure and performance. The NEIO approach is an econometrically-focused method, aimed at estimating parameters that measure the degree of competition within a single-product industry (Bresnahan, 1989). The drawback remains, however, that PCMs are often not directly observable, so cost functions must be estimated using inferences from supply-side profitmaximization relationships and market conduct. Examples of this line of empirical research can be seen in the work of Villas-Boas (2007) and Draganska et al. (2007) who use equilibrium models to recover economic margins. Alternatively, economists have relied on estimating marginal cost by incorporating a conjectural variation parameter that may affect margins estimates depending on the industry-specific market-behavior assumptions, as in Azzam (1997) and Lopez et al. (2002). Rather than extend this theory or attempt to estimate PCMs, our analysis will take a much more direct approach. Using newly acquired price data from U.S. military commissaries, price mark-ups can be calculated as a straightforward difference between wholesale prices paid by U.S. military commissaries and prices paid at food retailers. The use of the wholesale prices in the commissary data, in particular, linked to the retail scanner data, we believe, is an innovative empirical solution to the problem of identifying unreported wholesale prices and, by extension, identifying price-cost margins for nationally-branded products 1. The use of these data, as well as store-level scanner data, supports a more complete 1 This analysis does not make any assertions about the role of price-cost margins within food retailers strategic decisionmaking processes; rather, our goal is to measure the impact that market-level and demographics have on observed (i.e., shortrun) price-cost margins.

10 picture of the price-cost relationship with extensions and applications in the marketing, health, and policy sectors. Data Our analysis utilizes four main streams of data: (1) weekly U.S. military commissary sales data via EmpowerIT, (2) weekly store-level scanner data from food retailers via Information Resources, Inc. (IRI) InfoScan, (3) store characteristics data via Nielsen s TDLinx, and (4) a compilation of publically available data sources that represent components of the market and demographic landscape. In this section we describe the data sources and how we use this data to construct key variables for our analysis. U.S. Military Commissary Data Central to our analysis is the use of data on U.S. military commissary prices available for the years Operated by the Defense Commissary Agency (DeCA), commissaries are supermarkets located on military bases that sell food and other household items at cost. As stated on the DeCA s web page, a five percent surcharge is added at checkout to the patron s grocery bill, before coupons are deducted. Eligible shoppers include active-duty military, guard and reserve members, retirees and qualified family members (DeCA, 2012). Originally collected by EmpowerIT, the U.S. military commissary data present weekly average prices and total sales for national brand UPCs sold at U.S. military commissaries. In addition, UPClevel descriptions are provided, including food categories, manufacturer, brand, segment, commodity, and department. Using the DeCA s stated surcharge information, observed prices in the data may be treated as a direct corollary to wholesale prices after removing the 5% mark-up. We use these data in combination with IRI s InfoScan store-level scanner data to construct our monthly average aggregate price-cost margins. UPCs are unique product codes that are used to link with the UPCs in IRI. Store-Level Scanner Data In addition to the EmpowerIT commissary data, we use the IRI retailer-level scanner data (InfoScan ) to construct our dependent variables. The InfoScan dataset is a representative sample of food retailers across the U.S. that report average weekly prices, sales information, and product descriptions for food items sold. Information is collected for each UPC purchased during a given week.

11 TDLinx Store Characteristics Data We rely on Nielsen s TDLinx panel dataset over the same time span to describe various store characteristics in each market. This dataset contains additional information not available in IRI s InfoScan panel about a sample of food retailers, such as annual revenue estimates, square footage, number of registers/checkouts, and number of employees. We use these variables to construct census tract-level aggregations for the average number of check out registers (Registers) and average sales volume per square foot (Efficiency), each weighted by format-level revenue estimates. In addition, our analysis includes controls for presence of non-traditional food stores and industry concentration. We use the TDLinx panel data to calculate the number of convenience stores (Convenience) and supercenters (Supercenters) within each county. Finally, to capture information about food retailer concentration at the census-tract level we use the four-firm concentration ratio for food retailers (CR4). Publically Available Data Beginning with the USDA s Food Access Research Atlas (FARA), a census tract-level data set that identifies areas with poor food access, we select four specifications of low food access (two binary and two continuous) to include in our analysis. The FARA defines low food access as being far from a supermarket, supercenter, or large grocery store ( supermarket for short). If a significant portion of the population in each tract is far from a supermarket, then that tract is considered to have low-access (USDA, 2013). The four specifications of food access we present in our analysis are: 1. (FoodDesert-Orig) The original food desert measure. Identified by areas that are lowincome 2 and low-access, where low-access is measured using a 1 mile demarcation in urban areas and 10 miles in rural areas, according to the proximity to the nearest supermarket (FoodDesert-NoVehicle) Accessibility is more than proximity to a store. This measure incorporates vehicle access. Identified by areas that are low-income and low-access, as well as areas where more than 100 households have no access to a vehicle and are more than ½ a mile from the nearest supermarket. 2 Low-income tracts are tracts such that the poverty rate of 20% or higher, or the median family income is less than 80% of the median family income for the state or metro-area 3 According to this measure, an estimated 18.3 million people live low-income and low-access tracts in 2010.

12 3. (FoodDesert-%-1mi) Rather than use a binary indication, we categorize census tracts according to the percentage of individuals living with low-income and at least 1 mile from the nearest supermarket. This measure correlates to the original food desert measure; however, does not make a distinction between rural or urban census tracts. 4. (FoodDesert-%-NoVehicle) Similar to its binary counterpart, we identify census tracts according to the percentage of housing units living at least ½ mile from the nearest supermarket and report having no access to a vehicle. For each of these measures, tracts are defined as low-access if the aggregate number of people in the census tract with low-access is at least 500 or the percentage of people in the census tract with low-access is at least 33 percent (USDA, 2013). Finally, a tract is identified as having low vehicle access if at least 100 households are more than ½ mile from the nearest supermarket and have no access to a vehicle; or at least 500 people or 33 percent of the population live more than 20 miles from the nearest supermarket, regardless of vehicle access (USDA, 2013). The maps in Figures 1 and 2 correspond to the four specifications for the state of Pennsylvania. We augment the FARA with supplemental data from several other sources (see Table 1 for complete list of variables) across various spatial units. Due to the availability limitations of archived public data, not all of our sources may cover the same date range of We do not think this presents an issue for several reasons. As an alternative to the decennial census, which only becomes available every ten years, we use the American Community Survey (ACS) five year estimates over the period to define median income (MedInc), median age (MedAge), median house value (HouseVal), percentage of population with a bachelor s degree (BachDegree), and percent of population with Hispanic ethnicity (Hispanic). Similar to the decennial Census, data from the five year estimates include demographic information at the census tract level, and therefore allow us to control for smaller geographic differences in demographics. Other sources we report in our model may not align year-to-year with our specified date range due to availability of archived data, choosing to use updated data as data for the years specified may have become obsolete, or using lagged years for certain variables. Our final data set is a three-year panel of these sources plus the quantifications of price mark-ups and product assortment derived from the military commissary data, InfoScan, and TDLinx by census tract.

13 Figure 1: FoodDesert-Orig (Left) & FoodDesert-NoVehicle (Right)

14 Figure 2: FoodDesert-%-1mi (Left) & FoodDesert-%-NoVehicle (Right)

15 Table 1. Summary Statistics for the Econometric Variables (Page 1 of 2) Variables Definition Mean St. Deviation Source Agency Year(s) Spatial Unit Product Assortment & Price-Cost Margins ASSORT1 (Refer to Data Section) 1,671 1, InfoScan Information Resources, Inc Census Tract ASSORT2 (Refer to Data Section) InfoScan Information Resources, Inc Census Tract MARGINS1 (Refer to Data Section) (Unavailable) EmpowerIT Defense Commissary Agency Census Tract MARGINS2 (Refer to Data Section) (Unavailable) EmpowerIT Defense Commissary Agency Census Tract Food Access Measures- Binary FoodDesert-Orig FoodDesert-NoVehicle A low-income tract with at least 500 people or 33 percent of the population living more than 1 mile (urban areas) or more than 10 miles (rural areas) from the nearest supermarket, supercenter, or large grocery store. A low-income tract in which at least one of the following is true (i) at least 100 households are located more than ½ mile from the nearest supermarket and have no vehicle access; or (ii) at least 500 people or 33 percent of the population live more than 20 miles from the nearest supermarket, regardless of vehicle availability Food Access Research Atlas USDA ERS 2010 Census Tract Food Access Research Atlas USDA ERS 2010 Census Tract Food Access Measures- Continuous FoodDesert-%-1mi FoodDesert-%-NoVehicle Percentage of individuals in an urban tract with low income and living more than 1 mile from the nearest supermarket, supercenter, or large grocery store. Low income is defined as annual family income at or below 200 percent of the Federal poverty threshold for family size. (Quantiles reported) Percentage of housing units located at least ½ mile from the nearest supermarket, supercenter, or large grocery store and reporting no access to a vehicle. (Quantiles reported) Food Access Research Atlas USDA ERS 2010 Census Tract Food Access Research Atlas USDA ERS 2010 Census Tract Demographic Variables Urban Urban census tract Census U.S. Census Bureau 2010 Census Tract POP Total Population 4,251 1, Census U.S. Census Bureau 2010 Census Tract MedInc ($) Median Income 53,250 24, American Community Survey U.S. Census Bureau Census Tract MedAge Median Age American Community Survey U.S. Census Bureau Census Tract HouseVal ($) Median House Value 175, , American Community Survey U.S. Census Bureau Census Tract BachDegree (%) Percentage of the population reported as having completed a American Community Survey U.S. Census Bureau Census Tract bachelor's degree. Hispanic (%) Percentage of the population reported as having hispanic ethnicity American Community Survey U.S. Census Bureau Census Tract Infrastructure Variables Hospitals The number of establishments known and licensed as general medical and surgical hospitals primarily engaged in providing diagnostic and medical treatment (both surgical and nonsurgical) to inpatients with any of a wide variety of medical conditions. (NAICS Index #622110) County Business Patterns U.S. Census Bureau 2009 County FIPS Schools The number of establishments primarily engaged in furnishing academic courses and associated course work that comprise a basic preparatory education. (NAICS Index #611110) County Business Patterns U.S. Census Bureau 2009 County FIPS

16 Table 1. Summary Statistics for the Econometric Variables (Page 2 of 2) Variables Definition Mean St. Deviation Source Agency Year(s) Spatial Unit Product Bus_RailAssortment & Price-Cost Margins The number of establishments primarily engaged in operating local and suburban commuter rail systems PLUS the number of establishments primarily engaged in operating local and suburban passenger transportation systems using buses or other motor vehicles over regular routes and on regular schedules within a metropolitan area and its adjacent nonurban areas. (NAICS Index # & #485113) County Business Patterns U.S. Census Bureau 2009 County FIPS Retail-Market Characteristics CR4 Concentration ratio of top 4 largest food retailers TDLinx Store Characteristics Data Nielsen County FIPS Efficiency ($/sq.ft.) The total estimated sales volume divided by the store's reported 23,805 16, TDLinx Store Characteristics Data Nielsen Census Tract total square footage weighted by share of each store's annual commodity value. Registers Average number of checkout registers weighted by share of each TDLinx Store Characteristics Data Nielsen Census Tract store's census tract-level annual commodity value. Convenience The number of convenience stores within each county TDLinx Store Characteristics Data Nielsen County FIPS Supercenters The number of supercenters within each county TDLinx Store Characteristics Data Nielsen County FIPS DrugOnly Takes a value of 1 if a drug store is reported as the only food retail outlet within the census tract InfoScan Information Resources, Inc Census Tract

17 Methods Product Assortment The first dependent variable we have chosen to investigate is retailers average product assortment. This variable can be identified and estimated by exploiting the detailed store-level scanner data. To capture various characterizations of assortment, we propose two methods: ASSORT1 measures the average number of brands sold at a retailers in one census tract during the year and reflects the breadth of product assortment. To calculate this variable, we weight the total number of brands at each store by its census tract-level expenditure share. Previous research finds that shoppers prefer more brands when choosing a store, so this first measure of product assortment reflects the availability of options available to customers. ASSORT2 reflects the census tract average number of unique UPCs per brand sold at a individual stores in each month. The average number of UPCs per brand is weighted according to the expenditure share of each store within the census tract. Each brand sold at a food outlet may have a different number of UPCs, so while stores may carry similar brands across markets, this specification will capture the variety of products offered within each brand. It can be considered to be measure of assortment depth. Price-Cost Margins Unlike other studies that infer marginal costs from estimated structural models, we make use of the U.S. military commissary data to compute wholesale prices for national brands sold at commissaries and then extrapolate these wholesale prices to nearby stores. Specification of PCMs within this context presents two major challenges. The first challenge is quantifying average PCMs. To arrive at these PCMs, we match national-brand food products sold in commissaries with the same products sold in surrounding retailers by UPC, such that neighboring retailers fall within a distance of 30 miles. We calculate wholesale prices and price-cost margins for the neighboring non-commissary retailers by removing the 5% surcharge from the observed prices. Therefore, for each UPC sold at a given retailer, price-cost margins are calculated accordingly 4 : PCM itw st S,c t C = Price s t Price ct Price st (1) 4 This calculation is analogous to the Lerner Index.

18 for product (UPC) i, in market (census tract) t, for week w, given that store s in market t falls in the set of stores (S) surrounding commissary c also in market t. To ensure manageability of the data set, the monthly PCMs are calculated as the weighted average mark-up for each UPC: Avg_diff% itm st S,c t C = (PCM it1 Q it1 ) + (PCM it2 Q it2 ) + (PCM m it3 Q it3 ) + (PCM m it4 Q it4 ) (2) m m Q itm Q itm Q itm Q itm Our PCM data set includes all matching nationally-branded UPCs for each commissaryretailer pair, such that the neighboring retailer falls within 30 miles of the commissary. The second identified challenge is arriving at an aggregate census tract-level PCM. Although the mark-up data is restricted to national brand products, the variability of UPCs offered at different food retail outlets may still differ across markets. Ultimately, our preferred method for developing census tract-level PCMs will rely on a market basket of individual projects first calculated at the store level and then aggregated to the census tract level. However, for now, we employ the much simpler method of basing our census track-level measures of PCM on a single product that is widely purchased. More specifically, we calculate a census-tract PCM measure for refrigerated yogurt, which is generally considered to be a nutritious product (MARGIN1). For comparison s sake, we also calculate a census tract PCM measure for potato chips, another widely purchased product but generally considered to be highly caloric and not very nutritious (MARGIN2). To construct mark-ups at the census tract-level, we have summarized our procedure below: Step 1: Using the original UPC-level data from InfoScan and EmpowerIT, we match weekly observed prices of UPCs from U.S. commissaries with neighboring food retailers to generate a percent difference in price between food retailer and wholesale (commissary) prices. Step 2: (Monthly aggregation) Using weekly the results from Step 1, we calculate the monthly weighted average of mark-ups by share of units sold in each week. Step 3: (Census tract aggregation) We isolate the UPCs for the top purchased (i) refrigerated yogurt and (ii) bagged potato chip brands, then, using the share of total expenditures by store, we calculate a weighted average of mark-ups at the census tractlevel.

19 Econometric Model We develop a reduced-form model that represents the relationship of demographic and market-level characteristics with product assortment and price-cost margins at the census-tract level. Formally, the model to be estimated for market t in month m is: Assort itm = α 0 + α 1t (FoodDesert t ) + α X X tm + α M M tm + α Z Z tm + μ Atm (3) Margin itm = β 0 + β 1t (FoodDesert t ) + β X X tm + β M M tm + β Z Z tm + μ Btm (4) where Assort is the average level of product assortment across retailers in each census tract, Margin is the average price-cost margin for each census tract, X are census tract-level sociodemographic variables, M are census tract infrastructure characteristics, such as number of hospitals (Hospitals), Z are census tract-level characteristics of the retail markets, and A and B are error terms. To consider the case that census tract-level error terms are corrected with neighbors error terms, we investigate the following spatial error specification: μ Atm = λ A Wμ Atm + ε A, and μ Btm = λ B Wμ Btm + ε B where ε A ~N(0, σ 2 A I n ) and ε B ~N(0, σ 2 B I n ), and W is a weighting matrix. Following the approach outlined by LeSage (1998) and Aneslin (1999), we begin by estimating and testing for the spatial relationships of these variables using a Moran's I or Local Moran's I. These test statistics are used to identify the presence of spatial autocorrelation by assessing the extent to which patterns are spatially random. If the null hypothesis cannot be accepted, then there is evidence of spatial clustering (Light and Harris, 2012). The Local Moran s I is calculated as follows: I t = (y t Y ) w tj (y j Y ) where y t represents the attribute level for market t, Y is the mean attribute level across all markets, and w tj represents the weighting value between markets t and j. For our analysis, each market is represented by a U.S. census tract (for the contiguous U.S.). Attribute levels (y t ) for census tract t are the levels of product assortment or price-cost margin outcomes j t

20 (Assort and Margins). Spatial weighting matrices are constructed according to distance (in miles) between census tracts. LeSage (1998) and Anselin (1999) provide guidance on how to construct spatial weighting matrices. After testing for spatial autocorrelation among the product assortment or price-cost margin variables, we use the same procedure with the Local Moran's I for food deserts using the USDA's Food Access Research Atlas and compare these results with the Local Moran s I values from the marketing strategy analysis. Finally, we estimate a bivariate Local Moran s I using each marketing strategy variable and the indicator for food desert. The statistic for the bivariate Local Moran s I is similar to the univariate Local Moran s I, and is calculated as follows: I t = (y t Y ) w tj j t (FoodDesert j FoodDesert ) where we replace the weighted attributes with a measure of FoodDesert. The bivariate statistic compares the levels of product assortment or price-cost outcomes in census tract t with the presence of a food desert in a neighboring census tract j, rather than the level of the same marketing strategy outcome in census tract j. Testing for and identifying these types of spatial relationships can illuminate emergent patterns and strategic interaction between agents (Anselin, 1999). If patterns of irregular product assortments or PCMs can be linked with geographic locations of food deserts, then this could suggest market spillover effects in how neighboring firms compete. Estimation Results and Discussion With key variables defined, it becomes possible to estimate reduced-form economic models of product assortment and price-cost margins as functions of demographic and market-related variables. Due to temporary and unforeseen data access complications, results in this paper are unavailable for price-cost margins, as well as estimates from the spatial econometric model. However, we have applied these methods to product assortment for the state of Pennsylvania for the year We estimate equation (3) for the two specifications of Assort using ordinary least squares (OLS), and our results are presented in Table 2. In addition, for each measure of product assortment, we run five separate models. This first model is a baseline, where we only include socio-demographic variables (X). The second model adds a single retail-market characteristic (Z), a concentration measure (CR4), to the baseline. The third model uses the same X variables from the baseline as well as information on county level infrastructure (M).

21 The fourth model combines elements from the first three, such that we estimate each product assortment specification on socio-demographic, CR4, and infrastructure variables. The final model includes the socio-demographic and infrastructure parameters, as well as information on multiple retail-market characteristics (Z) at the census tract-level. Finally, for each of these models, we vary the FoodDesert parameter using the four definitions described above. We begin with the results for the baseline model. We expect that the sociodemographic variables selected will have a positive and significant effect on product assortment, as supported by the literature on quality competition in food retailing. In other words, we expect that levels of product assortment will increase with higher population, wealth, and education levels. We see that for both specifications of assortment, except for Hispanic and ln(medinc) with only one instance of significance, the baseline results confirm these expectations. The results for models two through four indicate that demographics, especially ln(pop), ln(medage), ln(houseval), and urban markets, consistently have positive and significant effects on product assortment, regardless of the specification. Markets where retail concentration is higher (i.e., higher levels of CR4) show a negative and significant effect on product assortment. This result suggests that markets with higher concentration and less competition among retail stores yield a smaller selection of product assortment. In areas identified as food deserts, this result would imply that individuals residing in these areas may not in fact have the same selection of products as areas with higher levels of food retailer competition. On the other hand, infrastructure plays a mixed role on product assortment outcomes. The presence of a higher level of public transportation is shown to have a negative and significant effect on both specifications of product assortment, as well as higher levels of hospitals on the number of unique brands in a market (ASSORT1). Moreover, when these infrastructure variables are added to the model, the impact of the (continuous) food-desert variable is even stronger. This somewhat surprising result suggests that localized areas with stronger presence of hospitals and public transportation are not conducive for high-assortment food stores. The final set of estimation results includes additional controls for the retail-market environment. In these models, we again find that socio-demographic factors play a positive and significant role, although age is shown to no longer be significant. A higher number of checkout registers (Registers) is positive and significantly related to higher levels of product assortment. This result could suggest that more registers signify more customer traffic, which may be related to an increase in the number of products available to consumers. Conversely,

22 Table 2. Preliminary Regression Results (Page 1 of 10) Dependent Variable Assort1 Model 1: Baseline Binary 1 Binary 2 Continuous 1 Continuous 2 Estimate St. Err. Estimate St. Err. Estimate St. Err. Estimate St. Err. Food Access Measures FoodDesert-Orig (150.11) FoodDesert-NoVehicle (112.49) FoodDesert-%-1mi ** (28.39) FoodDesert-%-NoVehicle * (24.94) Demographic Variables Urban *** (115.80) *** (115.95) (130.52) *** (115.05) ln(pop) *** (92.50) *** (92.65) *** (93.09) *** (92.61) ln(medinc) (153.58) (155.15) * (155.84) (153.30) ln(medage) ** (244.10) ** (245.10) ** (243.34) ** (243.48) ln(houseval) *** (111.59) *** (111.75) *** (112.37) *** (112.57) BachDegree * (911.79) * (911.93) * (910.65) * (912.58) Hispanic (419.26) (420.93) (419.27) (423.18) Concentration Variable CR4 Infrastructure Variables Hospitals Schools Bus_Rail Retail-Market Characteristics Efficiency Registers Supercenters Drug Only Constant *** ( ) *** ( ) *** ( ) *** ( ) R-squared N *** Coefficient is statistically significant at the 0.01 level. ** At the 0.05 level. * At the 0.10 level.

23 Table 2. Preliminary Regression Results (Page 2 of 10) Dependent Variable Food Access Measures FoodDesert-Orig FoodDesert-NoVehicle FoodDesert-%-1mi FoodDesert-%-NoVehicle Demographic Variables Urban ln(pop) ln(medinc) ln(medage) ln(houseval) BachDegree Hispanic Assort2 Model 1: Baseline Binary 1 Binary 2 Continuous 1 Continuous 2 Estimate St. Err. Estimate St. Err. Estimate St. Err. Estimate St. Err (0.14) 0.03 (0.10) ** (0.03) ** (0.02) 0.35 *** (0.11) 0.35 *** (0.11) 0.21 * (0.12) 0.36 *** (0.10) 0.28 *** (0.08) 0.28 *** (0.08) 0.31 *** (0.08) 0.29 *** (0.08) 0.08 (0.14) 0.09 (0.14) 0.15 (0.14) 0.07 (0.14) 0.60 *** (0.22) 0.60 *** (0.22) 0.63 *** (0.22) 0.62 *** (0.22) 0.62 *** (0.10) 0.62 *** (0.10) 0.58 *** (0.10) 0.58 *** (0.10) (0.83) (0.83) (0.83) (0.83) (0.38) (0.38) (0.38) (0.38) Concentration Variable CR4 Infrastructure Variables Hospitals Schools Bus_Rail Retail-Market Characteristics Efficiency Registers Supercenters Drug Only Constant R-squared N *** (1.41) *** (1.46) *** (1.40) *** (1.42) *** Coefficient is statistically significant at the 0.01 level. ** At the 0.05 level. * At the 0.10 level.

24 Table 2. Preliminary Regression Results (Page 3 of 10) Dependent Variable Food Access Measures FoodDesert-Orig FoodDesert-NoVehicle FoodDesert-%-1mi FoodDesert-%-NoVehicle Demographic Variables Urban ln(pop) ln(medinc) ln(medage) ln(houseval) BachDegree Hispanic Concentration Variable CR4 Assort1 Model 2: Baseline + Concentration Variable Binary 1 Binary 2 Continuous 1 Continuous 2 Estimate St. Err. Estimate St. Err. Estimate St. Err. Estimate St. Err (153.87) (114.43) ** (29.18) ** (25.41) *** (124.47) *** (124.46) * (138.16) *** (123.75) *** (94.83) *** (95.04) *** (95.37) *** (94.95) (156.15) (157.79) * (158.68) (155.86) ** (249.06) ** (250.00) ** (248.34) ** (248.58) *** (117.62) *** (117.82) *** (118.17) *** (118.57) (924.71) (924.58) * (923.67) * (925.39) (423.56) (425.61) (423.38) (427.63) * (322.49) * (322.56) (322.82) * (322.02) Infrastructure Variables Hospitals Schools Bus_Rail Retail-Market Characteristics Efficiency Registers Supercenters Drug Only Constant R-squared N *** ( ) *** ( ) *** ( ) *** ( ) *** Coefficient is statistically significant at the 0.01 level. ** At the 0.05 level. * At the 0.10 level.