Food Access and Food Choice: Applications for Food Deserts

Transcription

1 Food Access and Food Choice: Applications for Food Deserts Final Report Research Innovation and Development Grants in Economics (RIDGE) Grant # Gayaneh Kyureghian Research Assistant Professor Department of Food Science and Technology The Food Processing Center University of Nebraska-Lincoln Rodolfo M. Nayga, Jr. Professor and Tyson Endowed Chair Department of Agricultural Economics and Agribusiness University of Arkansas, Adjunct Researcher Norwegian Agricultural Economics Research Institute Azzeddine Azzam Professor Department of Agricultural Economics University of Nebraska-Lincoln Parts of this report have been previously published in manuscripts (Kyureghian and Nayga 2012(a), Kyureghian, Nayga and Bhattacharya 2012, Kyureghian and Nayga 2012(b)) with detailed results of our project. We thank Ms. Suparna Bhattachrya for research assistance. 1

2 I. Introduction Poor food choices have been shown to contribute to the rise of major chronic diseases, including overweight and obesity (Centers of Disease Control and Prevention (CDC)). Consequently, the Dietary Guidelines for Americans, 2010, emphasizes the need to shift food intake patterns to a more plant-based diet that emphasizes nutritious food, such as fruits and vegetables. Despite these efforts, only 42% and less than 60% of Americans meet the recommendations for as fruit and vegetable consumption, respectively. In academic and policy circles, as well as in the public eye, the local food environment has been associated with food choices and diet-related health consequences. Limited food access is considered especially worrisome for underserved, predominantly low-income areas, which are believed to be disproportionately subject to health and income disparities (Bitler and Haider 2011). The Food, Conservation, and Energy Act of 2008, refers to an area in the United States with limited access to affordable and nutritious food, particularly such an area composed of predominantly lower-income neighborhoods and communities (Sec Study and Report on Food Deserts, The Food, Conservation, and Energy Act of 2008, The United States Department of Agriculture, June 18, 2008) as food deserts. In February 2010, the Obama Administration proposed a $400 million Healthy Food Financing Initiative (H.R. 3525: Healthy Food Financing Initiative) that would eradicate food deserts by improving food access. Several states have launched policy efforts to increase access to healthy food. The concern in policy circles is that there may be insufficient availability and affordability of healthy food in these areas that may cause poor dietary choices. The literature findings in various disciplines of social science, marketing and nutrition, has addressed the issue of food access and choice from distinct, albeit overlapping angles (Larson, Story and Nelson 2009, Beaulac, Kristjansson and Cummins 2009, Blanchard and Lyson 2002, Sharkey, Horel and Dean 2010, Michimi and Wimberly 2010, Staus 2009). The empiric evidence from these disciplines lacks consensus in whether the food deserts exist and why. Findings from the studies on food deserts are quite diverse. For instance, Blanchard and Lyson (2002) found that residents of food deserts (considered non-metropolitan areas where people travel longer distance) are 23.4% less likely to consume the recommended level of fruits and vegetables (F&V) compared to those in non-food desert areas. Rose and Richards (2004) 2

3 examined the effects of limited access to supermarkets on the amount of F&V purchases. They measured access using three variables: distance to store, travel time to store, and car ownership. They concluded that limited access was negatively associated with the purchases, although the effect on fruits was not statistically significant. Michimi and Wimberly (2010) also found that while the odds of meeting the dietary recommendations concerning F&V consumption decreases as distance to large and medium-size supermarkets increases in metropolitan areas, they did not find similar association in non-metropolitan areas for any size of supermarket. Interestingly, in metropolitan areas they did not find significant relationship for large, medium and small supermarkets combined. Sharkey, Horel and Dean (2010), on the other hand, showed that underserved or low vehicle neighborhoods actually had better special access to a good variety of F&V in six Texas rural counties. Pearson et al. (2005) found no evidence of associations between the distance to the nearest supermarket and the difficulty of grocery shopping with either fruit or vegetable consumption. Bodor et al. (2007) considered not only distance to a store and the store concentration ratio, but also in-store food availability and found that the availability of fresh vegetables in the vicinity was positively related to vegetable intake, while fruit consumption was not associated with fresh fruit availability. In a review of literature on disparities in access to healthy food, Larson, Story and Nelson (2009) reported that although the majority of studies suggest a direct relationship between the presence of supermarkets and meeting the dietary guidelines for F&V, especially for African American adults, no such evidence was found for the youth. Likewise, in a systematic review of food deserts, Beaulac, Kristjansson and Cummins (2009) reported mixed results concerning the availability and quality of healthy foods in disadvantaged areas. A comprehensive review and analysis of the empirical literature on food deserts can be found in Bitler and Haider (2011). The gaps in the literature on whether food deserts exist appear to be related to the inconclusive evidence on the linkage between accessibility and food choice due to data limitations and methodological weaknesses. The data requirements to determine food access are many. One of these limitations is the variety of forms and categories of available foods such as produce, dry grocery, dairy, etc., in fresh, canned, frozen, juiced or dried form, in different sizes of packages, etc. Since there is no consensus in the literature concerning a specific food or a food group indicative of diet quality, data with reasonable coverage of a fairly representative group of foods is essential. The food desert literature typically focuses on fresh fruits and vegetables, 3

4 perhaps due to short shelf life (Blanchard and Lyson 2002, Sharkey, Horel and Dean 2010, Michimi and Wimberly 2010, etc.). Another data requirement is the adequate coverage of the food retail source, such as supermarkets, convenience and grocery stores, restaurants and other away from home sources, farmers markets, pick-yourself farms, etc. The focus of food desert literature has been on supermarkets as the retail outlets with the adequate assortment of healthy foods and affordable prices (Report to Congress 2009). This raised issues of non-adequate representation of the retail access environment, particularly when it concerns food away from home availability (Bitler and Haider 2011). In addition to the issues mentioned above, a common shortcoming of the primary data used in food desert research is the inadequate geographic coverage, typically at county or community level. While some secondary level data sets ameliorate this problem, they are plagued by issues raised above nonetheless. A widely criticized issue is the choice of the measure of food access in the literature. The distance to the nearest store(s) or the density of stores in the market area are the most common measures adopted in the literature (Hellerstein, Neumark and McIrerney 2008, Bitler and Haider 2011). The latter raises the issue of the choice of the appropriate geographic area as the relevant market, such as the census tract, zip code, cluster of zip codes, county, state, etc. (Hellerstein, Neumark and McIrerney 2008, Bitler and Haider 2011). The concept of the food desert also hinges upon whether it is an absolute (no food retail outlet in the area of reference) or a relative (fewer food retail outlets than in other areas) concept. The latter in turn raises the question of adequateness or sufficiency of food availability. There are several different definitions of food deserts, such as a distance of 10 miles or more to the nearest grocery store in rural areas, and 1 mile or more in urban areas, etc. Several other multidimensional definitions (e.g. by USDA, CDC, etc.) take into account not only the distance, but also the income level, commuting time, vehicle ownership, etc. in the reference area. The choice of the specific definition depends on the research question or purpose. For example, while the USDA definition is designed to capture the linkage between food availability and food choice, the CDC definition is more concerned by the linkage between food access and health consequences, such as overweight and obesity rates in the area. The different definitions mentioned above do not always overlap (Liese, 4

5 Battersby and Bell 2012), thereby creating variation in the evidence due to the specific research objectives and, therefore, the choice of food desert definition. Overall, it appears that the focus of much of the previous research is on supply side factors, creating an implicit underlying assumption that food deserts as a supply-side market failure and therefore motivating policy intervention to correct such market inefficiencies. But the contradictory empiric evidence in the previous literature about such complex phenomena as food deserts highlights the need for a more comprehensive approach. In this research project we analyze and interpret factors affecting the associations between food access, affordability and food choices. We consider both supply- and demand-side factors that may give rise or, at the least, compound the adverse dietary and health effects associated with food deserts. This research steps in to fill the aforementioned gaps in the literature. We focus on several staple healthy and unhealthy food groups mentioned in the literature, with an emphasis on fruits and vegetables. The food access measure in this project is the food store density at the county level. The research questions we seek to answer are (i) whether the availability of different types of food retail outlets affects the probability of patronizing that particular type of outlet for purchasing fruits and vegetables; (ii) whether food access or affordability or a combination thereof plays a major role in purchasing fruits and vegetables; (iii) whether household-level heterogeneity confounds the true effects of increased access to supermarkets; and (iv) whether the demands for 10 major food groups (healthy and unhealthy) are elastic or responsive to a proportional increase in supermarket availability. To explore these hypotheses we utilize national-level purchase data on all kinds of at home food purchases (Nielsen HomeScan Panel data), which overcomes most the primary data level shortcomings mentioned above. We use food availability data from the Census Bureau that covers food at home and away from home sources to depict as complete a picture of retail environment as possibly. The results of this research will help to design appropriate policy interventions to address heterogeneous strata disproportionately affected by inadequate food access. The rest of this report is organized as follows. The data sources and issues are discussed in detail in Section II. In Section III we formally test the linkage between the availability of supermarkets and the probability of patronizing supermarkets to purchase F&Vs. 5

6 II. Data Data for this project were obtained from four sources: the Nielsen HomeScan; County Business Patterns, U.S. Census Bureau, Population Estimates, U.S. Census Bureau; and Standard Reference 24, National Nutrient Database, USDA. We draw on 2005 and 2007 County Business Patterns and Population Estimates, U.S. Census Bureau, to delineate the food retail environment and the population/area estimates for the geographical units in our analysis. The food accessibility data include the number of establishments of the following store formats: supermarkets and other grocery stores (North American Industry Classification System (NAICS) code 44511), price clubs (NAICS code ), convenience stores (NAICS code 44512), specialty food stores (NAICS code 4452), full-service restaurants (NAICS code 7221) and limited-service eating places (NAICS code 7222) for approximately 3153 counties 1. In selecting the above food retail sector, we made a point to include all the food retail channels where people obtain food, a shortcoming in the previous literature (see A Report to Congress, Economic Research Service, USDA, 2009). These variables, adjusted for MSA or county level population and area, obtained from Population Estimates, U.S. Census Bureau, were used to create the retail store and restaurant densities per 1000 households per 100 square miles (hereafter referred to as the density variables) for each MSA/county for the reference year. There was a high proportion of missing data in the density variables. About 52% of all counties had all five density variables reported; therefore ignoring the counties with missing values would drastically reduce the sample size and possibly bias the results. To ameliorate this problem, we resorted to using missing data imputation methods. We utilized two types of imputations: last-value-dependent imputation and Markov-Chain Monte Carlo (MCMC) multiple imputations (Xu et al., 2008; Kyureghian et al., 2011). In the case of the last-value-dependent imputation, we obtained the time-series data for each one of NAICS codes mentioned above starting from , iteratively estimated a sequence of least squares regressions for each isolated NAICS industry, then used the estimated parameters and values imputed in the previous iteration to impute or fill in the data for counties with missing data points for the reference year. While this method capitalizes on the past values of the same variable and is logically appealing, 1 We refer to these food retail outlets as Supermarkets, Clubs, Convenience, Specialty, FS and QS, respectively. 2 Data prior to 1998 had a different industry classification system SIC. Although U.S. Census Bureau does provide a matching of 2002 NAICS to 1987 SIC for retail trade, the matching for the five industries were not unambiguous, and therefore were not considered appropriate for this imputation step. 6

7 it has two major drawbacks: it disregards the cross-sectional interdependence between food retail outlets at each point of time, and leaves a substantial portion of the missing data not filled in due to the absence of past data for the particular county. The MCMC multiple imputation method draws pseudorandom draws from the joint distribution of all five NAICS numbers for 2007 until it forms a Markov-Chain that converges to a target distribution. The MCMC method imputed or filled in all the missing values thereby motivating our choice of this method of imputation 3. We align the information on food access with actual household purchase data from the Nielsen panel from the same areas or counties. Nielsen, one of the largest commercial supplier of scanner data, started collecting in-home household scanner data in The panel members, selected from all 48 contiguous states, are supplied with handheld scanners to scan Universal Product Codes (UPCs) of all purchases and to upload this information on a weekly basis. The data are categorized in five datasets by food type: frozen foods, produce and meat products with UPCs, random-weight products without a UPC, dairy products, dry grocery products, and alcohol and cigarettes. Each record in the data set contains a household identification number, purchase date, a set of variables that combined provide a complete description of each product (product type variables), quantity purchased, price, etc. The dataset contains detailed information about both panel demographics (household size and composition, age, education attainment, employment status, race and ethnicity of male and female household heads, income, marital status, area of residence, etc.) and purchase information (price, promotion, purchase date, store type, etc.) 4. The data concerning the store type are organized into grocery, drug, mass merchandiser, supercenters, clubs, convenience and other stores. Grocery stores are stores selling food and nonfood items, including dry grocery, canned goods and perishable items, with annual sales volume 3 The MCMC methods rely on the assumption that the missingness is at random (MAR): the occurrence of missingness does not depend on the values of missing data. The County Business Patterns, U.S. Census Bureau, explains missingness as non-response by corporations. Based on the facts that the reported data on business establishments are aggregated geographically by counties, and that the unit of the source of missingness (corporations) and the unit of the reported data (counties) are distinct and completely independent, we assume that MAR is satisfied. 4 To identify observations corresponding to different food purchases we follow the procedure for the Quarterly Food-at-Home Price Database by ERS, USDA. We gratefully acknowledge Dr. Jessica Todd s help with SAS codes. 7

8 of $1M and more 5. A mass merchandiser is a retail outlet that primarily sells nonfood items but does have some limited nonperishable food items available. A supercenter is an expanded mass merchandiser that also sells a full selection of grocery items. A warehouse club is a membership store that sells packaged and bulk food and nonfood items. Convenience stores are small format stores selling high convenience items such as beverages, snacks and limited grocery items. Examples are conventional convenience and military stores, gas stations and kiosks. To get some sense of how the store breakdown is constructed in the Nielsen classification system, Safeway is classified as a grocery store, Rite Aid as a drug store, Wal-Mart as a supercenter, Target as a mass merchandiser, Costco as a club store, and Seven Eleven as a convenience store (Broda, Leibtag and Weinstein 2009). The socio-demographic variables in the model include race/ethnicity, marital status, education, employment, price, and Poverty Income Ratio (PIR). PIR is the ratio of household income to poverty threshold issued by the U.S. Department of Health and Human Services for each year. Households with PIR less than 1.35, from 1.35 to 1.85, from 1.85 to 2.50, from 2.50to 4.00 and greater than 4.00 are combined in income groups Income 1, Income 2, Income 3, Income 4 and Income 5, respectively. 5 TD Retail Trade Channel and Sub-Channel Overview.doc, Copyright 2011, The Nielsen Company. All rights reserved. Rev. 02/

9 III. Food Store Access, Availability, and Choice When Purchasing Fruits and Vegetables The existing literature on household s choice of stores does not typically account for both household and store characteristics. Dong and Stewart (2012) use the wealth of literature on product brand choice and reconcile it with their data on household characteristics to study the effects of consumer heterogeneity and habits on store type choice. They model the household choice of store types by using three groups of variables store and market variables, such as price, promotion and seasonality; past shopping variables, such as number of shopping occasions by households in each type of store and loyalty renewal; and demographic variables. The authors find that household demographics and past shopping behavior can both influence choice behavior. Our aim in this study is to examine the effects of density of different types of food stores on the likelihood that households will purchase F&V in a specific type of store. In other words, we propose to estimate the probability of patronizing each store types to purchase F&V conditional to the availability of both in home and away from home food retail establishments. We hypothesize that the retail food environment, along with marketing, store-level and sociodemographic factors, plays a significant role in explaining store type choice decisions when purchasing F&V. We use non-linear multinomial logit method to model this association. To address the potential endogeneity of food retail density variables, we use the corresponding lagged values for each county (Courtemanche and Carden 2011). Model Following the existing body of literature (e.g., Guadagni and Little 1983), we start with setting up the model of the household utility function. For household, the utility of buying food in store type at shopping occasion is expressed as: ( ) where is a store type specific parameter, variable accounts for seasonality in store choice, and are market- or store-level variables, such as price or promotion. The last term in the utility function,, has been referred to as the household loyalty variable, referring perhaps to the subject matter in the past research brand loyalty (Guadagni and Little 1983, Fader and Lattin 1993, Dong and Stewart 2012). This is basically the term that captured the cross-sectional 9

10 household heterogeneity in the earlier literature, drawing from past purchasing behavior. Guadagni and Little (1983), for example, used a weighted average of past purchases, with a heavier weight placed on the most recent period. Fader and Lattin (1993) suggested an improvement of this model by using draws from Dirichlet distribution, modified to capture the non-stationarity in choice behavior, to construct the loyalty term. By this assumption, the household choice from among J store types follows a Dirichlet distribution with a PDF ( ) ( ) ( ) ( ) ( ) ( ) where ( ) is the gamma function, and. The expected probability of the store type is expressed by ( ) ( ) where are store-specific parameters. By this definition the household choice only depends on store-level factors. Fader and Lattin (1993) suggested updating the expected probabilities by the number of choice occasions, thereby making the probabilities household-specific. Define to be equal to 1 if household chose store type at shopping occasion, then (3) is updated accordingly as: ( ) ( ) Since the total number of shopping occasions are Dong and Stewart (2012) hypothesized that household characteristics are important factors in explaining choice behavior and modified (4) to capture the full-spectrum of household characteristics: ( ) ( ) where. Following Dong and Stewart (2012), this last term helps to capture the household heterogeneity better than the number of past purchase occasions. Here and are store-specific parameters, and is a vector of household demographic variables. We use 10

11 this household choice mechanism, modified to include retail environment or store density variables along with demographic variables: ( ) There is no clear theoretical distinction nor is there any empirical evidence from the past literature as to where the density variables should appear whether (a) in the household loyalty measure, and therefore enter the choice model (1) indirectly or implicitly through the household loyalty factor, or (b) in the model directly or explicitly, alongside the store-level variables of price and promotion. The choice depends in part on the research question and whether we believe that the dominating effect in determining the probability of a household patronizing a particular store type is the household loyalty to that type of store (affected by the retail environment) or the availability of that particular (and other) type of stores in the household residence area. In this case we rely upon the empirical model to guide the choice. Based on a set of fit statistics we opted for the implicit model in (a). Fader and Lattin (1993) and Dong and Stewart (2012) paid special attention to incorporating non-stationarity in their models. We find that while non-stationarity is likely in modeling brand choice, it is not likely to be a problem in a store choice, let alone a store type choice. In fact, Dong and Stewart (2012) find no evidence of non-stationarity in their store choice model. Therefore, we proceed to defining a store type choice multinomial logit model as: ( ) where the second equation follows from (1). We use non-linear multinomial logit method to estimate model (7) (McFadden 1973, Chintagunta, Jain, and Vilcassim 1991, Fader, Lattin, and Little 1992). Data and Summary Statistics In this study, we use the 2006 Nielsen HomeScan household-level data to account for the consumer behavior. The Nielsen consumer panel for 2006 consists of 37,794 households. When purchasing F&V, groceries and supercenters are the most frequented types, accounting for 11

12 approximately 72% of all purchases. Price clubs are the third most frequented store, but have the highest price 36 cents per 100 g or almost $1.63 per lb. Grocery and convenience stores are next in line, with higher price offerings. The promotional status of the price is captured by promotion variable. Interestingly, drug stores offer a disproportionately high rate of discounts on F&V (i.e., F&V are on sale 54% of the time). Grocery stores offer over a third of their produce at a discounted price. Seasonality variables indicate higher levels of F&V sales towards the end of the year and this is consistent across all store types. Market and store level variables are presented in Table 1. Table 1. Means and Standard Deviations of the Marketing Variables by Food Store Type. Store Type Shopping Frequency Price ( / 100 g) Promotion Season 1 Jan-Mar Season 2 Apr-June Season 3 July-Sep Grocery (0.45) 0.31 (0.31) 0.36 (0.48) 0.11 (0.32) 0.22 (0.41) 0.30 (0.46) Drug (0.08) 0.25 (0.22) 0.54 (0.50) 0.18 (0.38) 0.21 (0.41) 0.25 (0.43) Mass (0.12) 0.23 (0.68) 0.18 (0.38) 0.19 (0.39) 0.22 (0.41) 0.26 (0.44) Supercenter (0.34) 0.25 (0.21) 0.12 (0.32) 0.14 (0.35) 0.22 (0.41) 0.29 (0.45) Club (0.26) 0.36 (0.45) 0.06 (0.23) 0.14 (0.35) 0.24 (0.43) 0.31 (0.46) Convenience (0.05) 0.30 (0.23) 0.16 (0.36) 0.16 (0.36) 0.21 (0.41) 0.30 (0.46) Other (0.23) 0.26 (0.28) 0.16 (0.36) 0.15 (0.36) 0.23 (0.42) 0.30 (0.46) Notes: Numbers in parentheses are standard deviations. Household demographic variables indicate that approximately 67% of households were married, with 9% and 3% of households with an African American or Asian American head, respectively. 50% and 52% of female and male household heads are employed and 64% and 56% of them have educational attainment of some college and higher, respectively. Approximately 24% of households have at least one child. The data also include information about household income categories. We calculate a continuous measure of income Poverty Income Ratio (PIR), by assigning individual incomes equal to the midpoint of the category, and then adjusting to the poverty thresholds by household size 6. The names, descriptions, means and standard deviations of the variables are reported in Table 2. 6 The poverty thresholds are issued in The 2006 HHS Poverty Guideline by the US Department of Health and Human Services. 12

13 Table 2. Description and Summary Statistics of Variables Used in Analysis. Variable Food Environment Variables Mean Super_2005 (NAICS 44511): # of supermarkets and grocers per 100 sq miles Clubs_2005 (NAICS ): # of price clubs per 100 sq miles Convenience_2005 (NAICS 44512): # of convenience stores per 100 sq miles Specialty_2005 (NAICS 4452): # of specialty stores per 100 sq miles FS_2005 (NAICS 7221): # of full-service restaurants per 100 sq miles QS_2005 (NAICS 7222): # of limited-service eating places per 100 sq miles Std Household variables PIR Child: = 1 if at least 1 child under 18 Female Education: = 1 if female head education level is some college or more Male Education: = 1 if female head education level is some college or more Female Employment: = 1 if female head employed Male Employment: = 1 if male head employed Married: = 1 if household head married Black: = 1 if household head is African American Asian: = 1 if household head is Asian Results The marginal effects of the variables from the estimation of (7) are presented in table 3. As indicated above, we report the results from the model where food retail density variables enter the utility function indirectly, through (6). In the interpretation of the estimates of the food retail density variables, we are particularly interested in the marginal effect of supermarkets, which include most of large grocery stores, mass merchandisers and supercenters (as defined by U.S. Census Bureau), since they potentially offer the affordability, assortment and other 13

14 Table 3. Marginal Effects from the Non-Linear Multinomial Logit Model Variable Grocery Drug Mass Supercenter Clubs Convenience Other Predicted Probability Marketing Variables Season ** (0.0012) (0.0005) ** (0.0006) (0.0019) * (0.0015) (0.0003) Season ** (0.0010) ** (0.0005) (0.0006) (0.0017) ** (0.0013) (0.0002) Season ** (0.0009) ** (0.0004) ** (0.0005) ** (0.0015) ** (0.0011) (0.0002) Price ** * ** ** ** ** ** (0.0016) Price Deal ** (0.0010) Household Demographic Variables PIR ** Child ** Female Education ** Male Education ** Female Employ ** Male Employ (0.0014) Married (0.0014) Black ** Asian ** Food Environment Variables Supermarkets ** Clubs ** Convenience ** Specialty ** FS ** QS Log Likelihood 1,983,620 AIC 1,983,926 BIC 1,983,760 Sample size 1,187,149 (0.0008) ** (0.0005) ** ** ** ** (0.0051) (0.0051) ** (0.0002) ** ** ** ** (0.0011) ** (0.0005) ** ** ** ** ** (0.0037) (0.0037) (0.0002) ** ** ** ** ** (0.0027) ** (0.0017) ** ** (0.0003) ** ** (0.0002) ** (0.0002) (0.0082) (0.0082) (0.0002) ** (0.0006) ** ** ** (0.0006) * (0.0002) ** ** (0.0017) ** (0.0016) ** ** (0.0003) ** (0.0002) ** (0.0002) ** (0.0002) (0.0730) (0.0730) ** (0.0003) ** (0.0004) ** ** ** ** ** ** (0.0003) (0.0002) ** ** * ** (0.0006) (0.0006) ** ** ** * ** ** ** (0.0003) ** (0.0003) ** ** ** ** ** (0.0003) (0.0003) ** ** ** ** ** ** Notes: Marginal errors are calculated at mean values of variables. Standard errors are in parentheses. * indicates significance at 5% level, ** indicates significance at 1% level. 14

15 characteristics often cited in the literature as necessary for improving diets (e.g., they offer a varied range of F&V). In general the predicted probabilities reported in table 3 preserve the order and are close in magnitude to observed frequencies reported in table 2. As mentioned above, unlike the loyalty measures in the previous literature, we incorporated the store access variables as well. The results indicate that the number of supermarkets and grocery stores (NAICS 44511) and clubs (NAICS ) have negative impact on the probability of patronizing supercenters and grocery stores and clubs for purchasing F&Vs. Supermarkets, in fact, have negative access on all types of large retailers. These outcomes should not be interpreted as a decrease in the probability of patronizing these types of stores as a result of an increase in the number of these stores. They merely indicate that the probability of purchasing F&Vs from these types of stores is decreased. A likely explanation is that supermarkets are typically a less expensive source of all kinds of food in general, not only F&Vs (U.S. Department of Agriculture, Economic Research Service (USDA ERS) 2009), therefore giving rise to possible substitution away from F&Vs to some other food groups. This result means that the number of supermarkets, which includes most of large grocers, mass merchandisers and supercenters (as defined by U.S. Census Bureau), is negatively associated with the probability of patronizing these stores to purchase F&V. This is in line with findings of Kyureghian, Nayga and Bhattacharya (2012), Kyureghian and Nayga (2012), Beaulac, Kristjansson and Cummins (2009), and Michimi and Wimberly (2010) that generally demonstrate mixed or no association between the availability of these stores and purchase and consumption of F&Vs. Unlike supermarkets and clubs, an increase in convenience stores translates into an increase in the probability of patronizing convenience stores to purchase F&Vs. This indicates that households highly value convenience, which may also explain the large negative impact convenience stores have on the probability of shopping at supercenters: a 1-unit increase in the number of convenience stores reduces the probability of shopping in supercenters by approximately 1 percentage point. The number of specialty stores (bakery, produce and butcher stores, etc.) has mixed effects on different types of stores a negative impact on the probabilities of shopping at mass merchandiser, supercenter and grocery stores, but impacts positively the probability of purchasing F&Vs in club and convenience stores, possibly due to the extreme 15

16 heterogeneity of this group. Contrary to the public belief the limited-service restaurants (NAICS 7222) actually increase the likelihood of purchasing F&Vs from nearly all types of stores. The marketing type variables, like price and price deals, defined as the unit price and the promotional status of the price, have mixed effects on patronizing different types of stores. For example, price is negatively associated with patronizing supercenters and mass merchandisers, which is in line with these stores being perceived as lower-priced than other types (table 1). Price deals, on the other hand, increase the probability of purchasing F&Vs from higher-priced store types, such as grocery, drug, mass merchandiser and other stores. The marginal effects of seasons 1to 3 on probabilities of patronizing grocery stores in these seasons relative to season 4 (Oct.-Dec.) are large and positive. These effects on other types of stores are mixed and sometimes insignificant. The results of household-level variables show that income is positively associated with patronizing higher-priced grocery, drug and club stores, echoing previous findings in literature (Dong and Stewart 2012, Staus 2009). The marginal effect of income is negative for all other types of stores. The presence of children in a household has large positive effect on lower-priced (supercenters) and high-volume (clubs) store types. Household head education attainment has negative and positive impact on the odds of shopping in supercenter and club stores, respectively. Household head employment and marital status have no discernible impact on store choice. Supercenters are noticeably less patronized by non-whites, with Asians preferring clubs more 7 and convenience stores less than whites. African American households demonstrate strong preference of convenience stores and less preference of clubs compared to white households. Concluding Remarks Household store choices could depend not only on store marketing characteristics and household demographic characteristics, but also on physical availability of different types of retail stores. The role of the latter in affecting the probability of patronizing a specific type of food store, when purchasing fruits and vegetables, is the focus of this study. Our results generally suggest that availability of supermarket and club types of food stores is inversely related to the likelihood 7 The stores both in the Nielsen and Census Bureau classifications are classified by size and assortment, thereby making it hard to discern a clear delineation which store types might include the ethnic stores. 16

17 of patronizing these specific types of food stores when purchasing fruits and vegetables. This finding has important policy implications given the attention that the accessible and affordable food retail environment (i.e., supermarkets) has attracted recently in relation to improving dietary quality and reducing obesity rates in the United States. The finding that the availability of convenience stores does in fact induce higher probability of purchasing fruits and vegetables from this type of store is equally intriguing and important. The disproportionately large negative effect of convenience stores on the likelihood of patronizing a supercenter indicates that when it comes to shopping for produce the households value convenience more than larger assortment and affordability typically found at supercenters. This finding suggests perhaps a whole new direction of policy intervention emphasizing reliance on smaller, more flexible store types. This reliance will take advantage of already proliferation dollar and other convenience store network hereby allowing the market mechanism to provide some of the solutions to the access problem and alleviate the burden on tax payers. Future studies should that would research the effect of food access and availability on likelihood of purchasing other types of food in different types of food stores will contribute to fully understanding the issue and designing appropriate remedies. 17

18 IV. The Effect of Food Store Access and Income on Household Purchases of Fruits and Vegetables: A Mixed Effects Analysis Given these concerns raised in the literature, it is important to realize that food availability affects food choice not only through physical access, but also through price and income. Unless these effects are accounted for, the results will likely be spurious. For this reason, supercenters and supermarkets have received much attention primarily due the price affordability and wide assortments of F&V they typically offer (Larson, Story and Nelson, 2009; Larsen and Gilliland, 2009) and due to the market power they exert in influencing market price (Broda, Leibtag and Weinstein, 2009; Courtemanche and Carden, 2011; Hausman and Leibtag, 2007; Hausman and Leibtag, 2004). Broda, Leibtag and Weinstein (2009) examined the consumer behavior in food demand across different store chains, store types and household and zip code characteristics. They used household-level purchase data to debunk several popularly-held beliefs. For example, while it may be the case that supermarkets do not locate in poorest neighborhoods, poorer households do not appear to have limited access to supercenters and they do not pay more for identical foods either due to limited access or market power exercised by the traditionally lowpriced retailers in underserved areas. The authors demonstrated that even though supercenters, mass merchandisers and even drug stores have significantly lower prices than traditional groceries, poor households combine the convenience in shopping nearby with the large volume of shopping from low-priced stores in a way that renders them not worse off than their richer counterparts. Despite the significant contribution of this paper to the empirical literature, failure to account for the retail environment along with other zip code characteristics limits the findings in a significant way. The potential availability of stores is bound to be a major driver for consumption patterns and should be taken into consideration as well. Courtemanche and Carden (2011) did this precisely in their manuscript by researching the effects of supercenters (i.e., Walmart in this case) on health outcomes, obesity in particular, through decreased food prices. They examined the endogeneity of store location decision and effectively estimated the effect of Walmart stores on the rise in obesity. Hausman and Leibtag (2007) demonstrated similar downward trend in prices when a supercenter move into a neighborhood and pointed out the social benefits associated with encouraging such entry. 18

19 Our study builds on previous findings by associating actual consumer behavior (similar to Broda, Leibtag and Weinstein, 2009) with neighborhood retail food availability (Courtemanche and Carden, 2011; Hausman and Leibtag, 2007; Hausman and Leibtag, 2004; and Michimi and Wimberly, 2010). Our aim is to model the individual and interaction effects of income and food access on actual purchases of F&V by households. The specific objectives of this study are threefold. First, we reconcile the gap between food access and actual purchase behavior by using a national purchase data set that has detailed household purchase and demographic information. In this paper, we improve upon the existing literature by analyzing the actual shopping patterns of households by explicitly isolating the effects of food access from the effects of income constraints, while addressing the data limitations mentioned above. Another improvement over the literature is our use of a wider definition of food access to encompass all retail outlets for both food at home and away from home. Second, we use hierarchical data analysis methods to account for possible clustering effects due to income or food access, which is an improvement over the methodology used in past studies. Finally, we conduct variance decomposition to describe the magnitude or the proportion of the contribution these two factors have on the variability of F&V purchases. The findings in this research will improve understanding of how food access issues interact with income levels in influencing purchases of F&V at the household level. To our knowledge, no other known study has examined this issue in the past using detailed household purchase, demographic, geographic, and food store access data. The focus on F&V is also noteworthy considering the need to improve the quality of diets, not to mention the high obesity rates in the U.S Model Following Courtemanche and Carden (2011) and Broda, Leibtag and Weinstein (2009), we postulate a model that estimates the impact of the retail food availability on F&V consumption. The F&V purchase is therefore modeled as a function of own-price, income, demographic variables, and store availability (Courtemanche and Carden, 2011). The choice of mixed effects modeling is motivated by the nature of the data. The observations in the data set are weekly purchases of F&V by households. The observations are completely nested in households, which in turn are partially nested in different income groups and in MSAs/counties with different food 19

20 access levels. The desired mixed effects model therefore involves a hierarchy of 288,884 weekly purchases of F&V by 52,943 households (the number of observations per household varies from 1 to 53, with a mean and median of approximately 27), residing in 3141 counties, clustered in 2311 MSAs. In matrix notation the mixed model is specified as where Y is the variable of interest, X is a vector of fixed covariates, Z is a vector of random effects, and is a vector of disturbances. The random effects in Z have mean and variance represented as E [ ] = [ ] and Var [ ] [ ] where G is the variance-covariance matrix for the random effects that controls for among group variations, and R is a block diagonal matrix of variance for the residual that allows within group variation in the model. The above approach of modeling covariance enables us to account for heteroskedasticity and correlations in the variables. The empirical models we designed to test the hypotheses set forth in the introduction capitalize on the richness of the mixed effects modeling to make inference using our hierarchical data. Four models were specified with two different food retail density variables for the two dependent variables. The models are (8) where is the logarithmic transformation of the dependent variable the ratio of actual and recommended servings of F&V (ratio) for household, in week ; is a vector of fixed effects: household specific demographic variables, price, season and region; is a vector of random effects: the scaled number of Supermarkets and PIR; and are fixed and random effect parameters, respectively; and is the idiosyncratic error term. We define a second model to estimate the effects of Supermarkets and income, only with a full set of density variables - Supermarkets, Convenience, Specialty, FS and QS: (9) 20

21 where is a vector of random effects: the scaled numbers of - Supermarkets, Convenience, Specialty, FS and QS and PIR. Other variables are defined as above. The alternative model specifications are (10) (11) where is the logarithmic transformation of the dependent variable the amount of F&V servings (level) purchased by household, in week ; is a vector of random effects: Supermarkets and PIR in equation (10) and a full set of the food access - the scaled numbers of Supermarkets, Convenience, Specialty, FS and QS and PIR (equation (11)). Other variables are defined as in (8) above. A total of 4 models are estimated. In the analysis, no restrictions were imposed on the variance-covariance matrix for the residual R. In other words, residuals are modeled as homoskedastic. The variance-covariance matrix for the random effects, G, was set as a block-diagonal matrix with the blocks identified by levels of income/access interaction variables, differentiated by metropolitan area status for each household. Data and Summary Statistics We employ four data sets in our analysis: the Nielsen HomeScan; County Business Patterns, U.S. Census Bureau, Population Estimates, U.S. Census Bureau; and Standard Reference 24, National Nutrient Database, USDA. We draw on 2007 County Business Patterns and Population Estimates, U.S. Census Bureau, to delineate the food retail environment and the population/area estimates for the geographical units in our analysis. For the purchase data we use 2008 Nielsen HomeScan panel data. The purchase of a food items in Nielsen is captured by a quantity variable expressed by ounces or fluid ounces. Following Nevo (1997), the reported ounces and fluid ounces were expressed in terms of serving sizes. This was a convenient transformation facilitate the aggregation of the quantities of different types of produce (canned, fresh, frozen, etc.) and relating them to the dietary guidelines. In the first step of this conversion process, the 21

22 observations in the Nielsen data set were divided into two food groups fruits (fresh, frozen, canned, dried, juice) and vegetables (fresh, frozen, canned, and juice) that came from the frozenproduce-meats and dry grocery data sets. Three key variables used to uniquely identify each produce item are: product group (e.g. fresh produce), product module (e.g. fresh fruit remaining), and product (e.g. lemon or mango, etc.) The reference data source of the serving sizes and refuse rates (i.e., the ratio of the skin, stone, and any other inedible parts that are discarded prior to eating in the total weight) for each produce item is the USDA National Nutrient Database for Standard Reference (SR 24). In a few cases where one-to-one matching between the products from the two data sets was not possible, alternative measures were taken: 1) higher level of aggregation (i.e. product modules or product groups) were considered, 2) weighted average of existing products were taken (e.g. melons in Nielsen matched as average weight of all types of melons in SR 24), or 3) different types were considered (e.g. under product module- fruit refrigerated, a specific product citrus salad was matched as fruit salad, canned in SR 24). Using two models, we estimate the associations of food access (i) on quantities of F&V purchased and (ii) on the extent households meet the dietary recommendations concerning F&V. The dependent variables we used in our analysis are therefore expressed as (i) number of servings purchased (in the remainder of the paper we refer to this as level ), and (ii) the ratio of the actually purchased to the recommended numbers of servings of F&V (we refer to this as ratio ) per household per week. The recommended numbers of servings by gender and level of physical activity are available from the Centers for Disease Control and Prevention (CDC). Since the levels of physical activity for household members are not available in the Nielsen panel, we considered 5 a day as the recommended servings for F&V for every household member. Finally, we aggregated the data by week and by broad food group (F&V). In the subsequent regression analysis, we use a logarithmic transformation of the dependent variables, along with random effects, to satisfy the normality requirement for a mixed model specification (Searle, Casella and McCulloch, 1992; Littell et al., 2006). The socio-demographic variables in the model include race/ethnicity, marital status, education, employment, price, and Poverty Income Ratio (PIR). A detailed description of these variables and summary statistics are provided in Table 4. 22

23 Table 4. Variable Descriptions and Summary Statistics Variable Name Description Mean (Std. Dev) Ratio of purchased and recommended servings Percent of the purchased number of F&V servings in recommended number, per household per week (21.68) Purchased servings Number of servings purchased, per household per week 8.42 (11.82) Price of F&V The weighted average price, per serving per week (64.43) Supermarkets Number of supermarkets and large groceries per 1000 households per 100 square mile in each MSA (886.43) Convenience Number of convenience per 1000 households per 100 square mile in each MSA ( ) Specialty Number of specialty stores per 1000 households per 100 square mile in each MSA (448.03) Full-Service Restaurants (FS) Number of full-service restaurants per 1000 households per 100 square mile in each MSA ( ) Quick-Service Restaurants (QS) Number of limited-service eating places per 1000 households per 100 square mile in each MSA ( ) Season1 Months in Jan-Mar 0.25 (0.43) Season2 Months in Apr-Jun 0.24 (0.42) Season3 Months in Jul-Sep 0.23 (0.42) Season4 Months in Oct-Dec 0.28 (0.45) Region1 East 0.16 (0.37) Region2 Central 0.26 (0.44) Region3 South 0.38 (0.49) Region4 West 0.20 (0.40) PIR Poverty Income Ratio = midpoint of category adjusted to poverty thresholds* by household size 4.23 (2.65) Household Size Household Size 2.40 (1.23) Married A binary variable that takes a value of 1 if married and 0 otherwise (single, widowed, divorced/separated) 0.68 (0.47) White A binary variable that takes a value of 1 if white and otherwise (black, oriental, other) Black A binary variable that takes a value of 1 if black and 0 otherwise (white, oriental, other) *The poverty thresholds are issued in The 2008 HHS Poverty Guideline by the US Department of Health and Human Services. (0.34) 0.07 (0.26) 23