Obesity, Fast Food, and Grocery Stores: Evidence from Geo-referenced Micro Data Susan E. Chen∗1 , Raymond J.G.M. Florax1,2 , and Samantha D. Snyder1 1
2
Purdue University, West Lafayette, USA VU University Amsterdam, The Netherlands. This Draft: February 9, 2009
Abstract In this research we provide unique quantitative estimates of the effect of proximity to fast food and grocery stores on obesity. Our empirical model combines geo-referenced micro data on access to fast food restaurants and grocery stores with data about salient personal characteristics, individual behaviors, and neighborhood characteristics such as zoning and crime. We create a “local food environment” for every individual utilizing 21 mile buffers around a person’s home address. Local food landscapes are potentially endogenous due to spatial sorting of the population and food outlets. The BMI of individuals living close to each other are likely (spatially) correlated because of (un)observed individual and neighborhood effects. The biases associated with endogeneity and spatial correlation are handled with spatial econometric instrumental variables and general method of moments techniques. Our policy simulations focus on reducing the density of fast food restaurants, or alternatively increasing access to grocery stores. We account for spatial heterogeneity in both the policy instruments and the home location of individuals, and consistently find small and significant effects for the hypothesized relationships between fast food and grocery store density and individual BMI values.
∗
Contact Author: Department of Agricultural Economics, Purdue University, 403 W. State Street, West Lafayette, IN 47907–2056, USA. Telephone: 765-494-7545. Fax: 765-494-9176. E-mail:
[email protected].
1
Introduction
Effective public health intervention depends critically on identifying and understanding the health related behaviors that cause obesity. While individual level factors such as biological (genetic) and socio-economic conditions have been shown to be associated with obesity, there is a growing literature on the role the food environment plays in the prevalence of obesity in communities. The two main questions addressed in this literature are: does close proximity to fast food restaurants make people obese and/or does lack of access to retail grocers contribute to people’s weight? The dominant research finding is that lack of access to grocery retailers is positively associated with obesity rates (Morland et al., 2006), while the evidence for access to fast food is not so clear (Cummins and Macintyre, 2006). Some studies document a positive correlation between obesity and fast food access (Maddock, 2004; Chou et al., 2004), but others do not (Burdette and Whitaker, 2004; Jeffery et al., 2006). While the studies cited above have been persuasive in arguing that there is a correlation between some subset of these features, the findings are not causal because they do not take into account the importance of choice. People select where they want to live based on some subset of neighborhood characteristics and individual preferences. Although simple regression analyses and bivariate correlations are reflective of the current state of the literature, they do not account for the selection issues and the resulting endogeneity of the food landscape variables identified above. The food landscape variables may be endogenous because where a person chooses to locate may be driven by underlying preferences, which in turn may also be correlated with factors driving obesity. Oftentimes these factors are unobserved. Ignoring this endogeneity does cause bias in the effect of the food landscape on obesity. Other shortcomings of the work in this area are the aggregate nature of the areal unit under consideration, and the lack of employing appropriate spatial econometric techniques. Confidentiality restrictions often prohibit the release of geographical identifiers in publicly available health surveys and economic surveys.1 As a result, many of the studies that examine the relationship between the food landscape and BMI have had to use arbitrarily designated regions (e.g., census tracts or counties), which are largely based on administrative compatibility (for example: Maddock, 2004; Chou et al., 2004; Morland et al., 2002, 2006; Moore and Diez Roux, 2006). The problem with using arbitrarily designated neighborhoods and large areal units is that these samples may lead to biased results because of ecological fallacy and the assumption that people do not shop outside of their census tract. These ‘spillover’ behaviors can only be adequately taken into account by allowing for spatial dependence across administrative units. This study estimates a reduced form model for the determinants of BMI.2 We are able to overcome some of the limitations of previous studies by using two unique data sources that include geographical identifiers for individuals and all retail food establishments, along with demographic, economic, and health data for a group of citizens living in Indianapolis, Indiana, in 2005. Unlike in previous analysis, these unique datasets allow us to create local 1
For example, two key national surveys that could be used to address the question under study are the Behavioral Risk Factor Surveillance Survey and the Census of Retail Trade. These surveys, however, only release information at the census tract level. 2 The underlying behavioral model for BMI has been laid out in Philipson and Posner (2003) and Chou et al. (2004).
1
food landscapes for each individual. We are also able to include individual level neighborhood variables for other neighborhood characteristics such as crime. Arguably, our analysis controls for neighborhood characteristics which have been ignored in previous studies, and we are able to isolate the effect of fast food and chain grocers on individual health. The second contribution to the literature that our research provides stems from the consideration that individuals select where they want to live based on neighborhood amenities—the food landscape being one subset of these amenities. To control for individual neighborhood selection we use an instrumental variables approach based on city zoning regulations. The commercial zoning instrument that we use affects where fast food restaurants locate but we assert, is uncorrelated with other unobserved determinants of BMI. As a final contribution to the literature we argue that dependence across individuals should be accounted for. Given the inherently spatial nature of the dataset it is likely that observations are not independent across space due to (unobserved) social network ties among individuals or shared (unobserved) neighborhood characteristics across individuals living in proximate neighborhoods. We therefore use instrumental variables and generalized method of moments techniques recently developed in spatial econometrics (Kelejian and Prucha, 2007) to account for spatial dependence and heteroskedasticity. This approach provides an even stronger test for the effect of access to fast food restaurants and chain grocers on individuals’ BMI. Recent policy action to influence the food environment, particularly among populations deemed most ‘at risk’ for obesity (e.g., minority and low-income groups) aims to restrict the number of fast food restaurants. In South Los Angeles, ‘health zoning’ has recently been proposed; the ordinance would put in place a two year moratorium on new fast food restaurants. The goal of another proposed law in California is to increase the availability of nutritious foods, particularly in underserved areas (Abdollah, 2007). Drawing on past policies focused on limiting liquor store licenses in response to alcohol abuse problems, and not withstanding that these policies have weathered lawsuits challenging their constitutionality, municipalities and other local governing bodies are considering similar laws focused on where and how fast food restaurants operate (Mair and Teret, 2005). Our research focuses on informing policies such as these by quantifying the estimated effects on BMI of access to fast food restaurants and grocery stores, while explicitly accounting for the spatial variability of the effects.
2
Background
There is increasing evidence that there are clear geographical patterns in where grocery and food retailers locate. Morland et al. (2002), Moore and Diez Roux (2006), and Zenk et al. (2005) found that food retailer patterns follow closely with residential distributions of income, percent minority populations and other neighborhood characteristics. Grocery stores tend to be located in white, affluent neighborhoods while fast food restaurants are disproportionately located in lower income neighborhoods (Block et al., 2004). One reason suggested for this growing disparity in access to food retailers is the consolidation of large grocery chains over the last 30 years. Chung and Myers (1999) incorporated food prices in their research on the grocery retailer environment. They found that price discrepancies are starkest between chain and nonchain retailers, with the latter charging significantly higher prices. And, in accordance with
2
the previous studies, larger chain grocery stores were less likely to locate in inner city, lower income neighborhoods. At the same time the research on obesity prevalence has suggested that there is evidence of spatial clustering of obese individuals. Both at the state level and more disaggregated geographical levels such as the county or census tract, researchers have found that people with similar BMI cluster together(Mobley et al., 2004; Eid et al., 2006). This clustering phenomenon supports the hypothesis that contextual factors, either social (e.g. crime, peer effects) or physical (food landscape), are affecting the health of the individuals that live within these neighborhoods. It has, therefore, been suggested that limited access to food retailers offering affordable and healthy options has lead to an increase in the prevalence of obesity particularly in urban neighborhoods with low income and/or predominantly minority residents (Cummins and Macintyre, 2006). The growing literature on the role that the food landscape plays in the obesity epidemic focuses largely on the consumption of food away from home, (in particular fast food restaurants), and not on retail grocers. Chou et al. (2004), using individual level data from the Behavioral Risk Factor Surveillance Survey, statewide counts of restaurants, and restaurant expenditures data found that residents of states with a higher number of restaurants tended to have higher BMI. In addition, lower prices at restaurants, (and actually food prices, in general), were correlated with higher BMI. While this analysis provides valuable insight into the contributors to obesity and overweight, limitations arise due to the less specific state wide proxies for food consumption behavior. Other notable studies in this literature use less aggregated data to study the association between access to retailers and BMI (Morland et al., 2002, 2006; Rose and Richards, 2004; Jeffery et al., 2006). The work of Morland and colleagues used census tract level data to examine the relationship between food access, consumption, and obesity and overweight. Morland et al. (2002) defines an individual’s local food environment as the number and type of food retailers within the census tract where the person resides. They find that for Blacks in the study, fruit and vegetable consumption increase by 32 percent for each additional supermarket located in their census tract. Additional work by Morland and colleagues divides food retailers into three categories: supermarkets, grocery stores, and convenience stores. They find that lower prevalence of obesity and overweight is associated with the presence of supermarkets, while higher prevalence rates are associated with the types of stores characterized by less healthy dietary options (Morland et al., 2006). There are two main limitations of Morland’s studies. Both employ the census tract as the definition of an individual’s market for food. Again, census tracts vary in size and oftentimes their boundaries do not reflect any substantively significant delineation. The problem with using arbitrarily designated neighborhoods and large areal units is that these samples may create spurious relationships originating from ecological fallacy and boundary issues. The ecological fallacy argument stems from the fact that we are inferring characteristics of smaller areal units (call them neighborhoods), based on the aggregate level data available at the census tract level. For example, the number of grocery stores, median income, race or any other aggregate statistic at the census tract level may be biased and not accurately reflect what is truly happening in the smaller geographical neighborhoods within the census tract. In addition, individuals do not confine their retail activities to the census tract where they live. It may in fact be closer 3
to shop at a grocery store in another census tract if you live close to the border. The second shortcoming is that these studies do not account for unobservable neighborhood effects, which may bias the effect of grocery store access on healthy eating and BMI.3 There are studies that use disaggregated individual level data to study the relationship between access to various retailers, dietary choices and health outcomes. Jeffery et al. (2006) studied a set of survey respondents in Minnesota linking fast food restaurant consumption and health outcomes. Jeffery and colleagues found that BMI tends to increase with fast food meals, however, there was no significant relationship between proximity to fast food restaurants and either consumption of fast food meals or higher BMI. Rose and Richards (2004) used data from the federally funded Food Stamp Program to assess the impact of retail access on fruit and vegetable consumption. This study is different from the preceding ones in that it had information on the actual fruit and vegetables consumed by food stamp participants as well as information on the retail outlet where they made their food purchases. They used the distance and travel time to this store to quantify access. Their result, that distance matters when choosing to consume healthy food at home is an important one. While both of these studies were novel in their use of individual data to study this problem, these studies did not account for both the (unobserved) environmental and (unobserved) social factors that affect people’s eating choices. The relationship between food consumption and a health outcome is a complex one. Previous research has shown that the food consumption decision is affected by food availability. More recent novel research by Christakis and Fowler (2007) has also suggested that obesity can be spread through social networks. Using data from the Framingham Heart Study over a period of 30 years, they found that individuals are far more likely to be obese if their friends and family (social networks) are also obese. They maintain that friends, colleagues, and family affect a person’s perceptions of weight and eating habits. In other words, prevailing norms about how much to eat, exercise, and weigh affect our decisions on food choices, physical activity, and body image. While our study does not address the social network aspect of the obesity epidemic explicitly, our spatial econometric modeling approach allows us to consider the effect of social networks, proxied by the behavior of our neighbors, on body mass index. This study is laid out as follows. We first outline the research methods and the data that we use for this study. We then present empirical estimates of the effect of access to fast food and chain grocers on BMI using models that account for both the sorting of people into neighborhoods and the spatial spillover effects across people. Using the results from our models, we then simulate the marginal effects for two experiments that are of interest to policymakers. The first experiment examines the effect of setting a density limit on the number of fast food restaurants in high density fast food areas. The second policy experiment increases the number of chain grocers in ”at risk” neighborhoods that are particularly vulnerable to the obesity epidemic. 3
Both Morland et al. (2002) and Morland et al. (2006) use a random effects model which accounts for unobservables at the census tract level but assumes that the unobservables are uncorrelated with the error term.
4
3
Research Design, Data and Methods
The data for this study was gathered from an amalgam of sources. The individual health data was taken from the Adult Obesity Needs Assessment telephone survey which was conducted by the Marion County Health Department from February through June of 2005. These self reported data include age, sex, education, income, labor force participation, physical activity, along with weight and height information. Unlike most health surveys, these data included the location of the home address of the respondent.4 The foodscape data were obtained from the Marion County Health Department’s health safety inspection records in 2005. These records included the name, and location of all food retailers in Marion County. Using this data, construction of a representation of the ‘foodscape’ for the county was possible by geocoding the addresses both automatically and manually with ArcMAP 9.1. Retail food stores were classified into four categories: large chain grocer, small grocery store, convenience store, and specialty store. Restaurants were classified into fast food and sit down. Fast food restaurants were selected based on the North American Industry Classification System (NAICS) definitions of Limited-Service Eating Places by the US Census Bureau. Data on neighborhood characteristics were based on geocoded crime data (Indianapolis Metropolitan Police Department, 2007), and zoning regulations (Indiana Spatial Data Portal, 2008). The unique geographic detail of the data allowed us to create individual-specific landscapes for food, criminal activity and zoning. Using the x and y coordinates for each respondent and the x and y coordinates of food retailers, crimes committed and zoning information, distance and density measures were calculated for each respondent within a half mile buffer of where they lived. A half mile was chosen because empirical data suggests that people in the US do not travel distances further than this for shopping or even commuting purposes (Agrawal and Schimek, 2007). Alternative studies in this literature have also used 12 mile as the food market radius in urban areas (Rose et al. (2007), for example). In addition, the geographical scale of Indianapolis, and sensitivity analysis on different measures of market diameter were tested to determine the optimal market radius. A half mile seemed the most appropriate for this urban environment. Figure 1 shows an excerpt of the data. This localized view outlines explicitly how we go about defining these individualized markets for food for each person. We have chosen two people arbitrarily and drawn circles of a 21 mile radius to outline how we come up with our measures. In both cases, there are no chain grocers within 12 mile of where they live. In the second case, there is 1 fast food outlet within 21 mile. We repeat this exercise with the data on crime to a get a count on the number of crimed committed within 12 of each person. We also use GIS zoning maps to estimate the percent of land zoned non-residential within a 12 mile buffer of each persons residence. The sample was restricted to consist of adults between the ages of 21 to 75 years. Implausibly high and low BMI’s were deleted from the original sample. The ultimate sample consists of 3,550 respondents. The descriptive statistics are reported in Table 1. The sample is predominantly white and female. Approximately 58 percent of those interviewed were women and about 30 percent of the respondents classified themselves as Nonwhite. The average age was ap4
To be more precise, the exact address was suppressed but the closest intersection to home was documented.
5
proximately 47 years. Approximately 65 percent had pursued some post-secondary education. Almost 21 percent lived in a household that earned an annual income less than 200 percent of the Federal Poverty Level (FPL), as defined by 2003 standards of the U.S. Department of Health and Human Services. A 200 percent of the FPL income cutoff was used based on research conducted by the National Academy of Sciences and later augmented by the National Center for Children in Poverty which suggest that a household actually requires about twice the FPL to meet basic needs (Cauthen and Fass, 2008). The behavioral variables used in this analysis were weekly physical activity, physical activity on the job, and smoking habits. On average, respondents engaged in vigourous physical activity 3 days out of the week. Where vigorous physical activity is described as an activity that you did for at least 10 minutes which required harder than normal effort e.g. heavy lifting, aerobics, or fast cycling. Just over 41 percent of the respondents reported that their job keeps them physically active and almost one quarter of the sample currently smokes. In terms of the ’foodscape’, the average number of large grocers within 21 mile of a respondent’s residence suggests that most respondents do not have a large chain grocer within their neighborhood. On average there were at least two fast food restaurant within 1/2 half mile of where a respondent lived. The neighborhood variables that affect health outcomes are serious crimes committed within 1/2 mile of where a person lives. A serious crime was defined to be those crimes classified as a attempted or accomplished rape, homicide, robbery of residence or assault on a person. The mean number of serious crimes committed within 1/2 of a respondent’s residence was 41. Since people will choose the neighborhood based on amenities such as safety, this variable is included in our regression to control for residential choice.
4
Empirical Framework
We estimate a reduced form production function for health where health is measured by BMI. To operationalize this model we assume that the BMI of individual i living in community j (or more general at location j, identified by coordinates x and y) is a function of prices, income and demographic characteristics: hij = x0ij β + p0ij γ1 + ωij ,
(1)
where hij is the health outcome of interest which is usually BMI or a dichotomous variable to indicate obesity, p0ij a vector of individual, location-specific counts for the number of grocers and fast food restaurants within a 1/2 mile radius, x0ij a series of demographic characteristics, and ωij an error assumed to be independently and identically distributed. Absent price competition and assuming product homogeneity, p0ij can be seen as a price vector, because it is inversely related to transportation cost. Note that equation (1) differs from the majority of studies in the literature on the effect of the food environment on health because it considers both fast food and grocery stores in the same model. The previous literature focused on either fast food or grocery stores while not taking into consideration that an increase in BMI is a result of cumulative exposure to both food deserts (no groceries) and food swamps (too much fast food). Equation (1) is flexible in allowing BMI to be affected through both these channels. 6
If individuals select where they want to live based on neighborhood amenities—the food landscape being only one subset of these amenities—then estimation of equation (1) will yield biased results for γ1 . One way to solve this problem is to try to control for the heterogeneity across individuals, x0ij , as well as the neighborhoods where they live, n0ij : hij = x0ij β + p0ij γ1 + n0ij γ2 + ωij .
(2)
Including controls for both individual and neighborhood heterogeneity can solve the endogeneity problem if the selection is based on observables. If selection, as is likely the case, is based on both observable and unobservable neighborhood and individual characteristics then estimates of γ1 in equation (2) will still be biased. We argue that in particular the number of fast food restaurants within an individual’s local food landscape of a half mile radius around the individual’s residence is endogenous. When selecting a neighborhood, individuals may make location decisions based on proximity to fast food restaurants, if they value the combination of services and convenience that fast food restaurants provide. The 1/2 mile radius describes the region within which an individual is willing to travel to purchase food quickly and conveniently, presumably under a time constraint. Because it is less likely that planned grocery store trips are subject to the same conveniencedriven impulses, we instrument only for fast food restaurant locations. A valid instrumental variable will only affect BMI through its effect on fast food locations, will not itself be affected by BMI, and will be highly correlated with BMI. We argue that the amount of land that is zoned non-residential within a 12 mile radius of a respondent’s residence is a valid instrument for fast food. Using zoning maps for Indianapolis, we construct this measure by calculating the percent of non-residentially zoned property within a 21 mile neighborhood radius of where a person lives. This creates individualized zones for each person in our sample. Preliminary tests indicate that fast food is endogenous in the BMI equation. We will therefore use an IV approach to account for the endogeneity of fast food. The instrument that we use is the percent of land that is zoned non-residential within 1/2 mile of a persons residence. Percent zoned non-residential within 12 mile is highly correlated with density of fast food outlets within a 12 mile of where a person lives. The regression results for the first stage of the instrumental variable (IV) estimation are presented in the Appendix along with other tests to support the use of the instrumental variable. Our initial tests also reveal that the individual observations are correlated across space. This may be caused by, for instance, social networks which are partly formed on the basis of spatial proximity, or by shared local environmental characteristics among individuals living in proximate neighborhoods. Extending (2) to allow for both types of (spatial) correlation, and including all individuals in the sample, we obtain in matrix notation: h = λW h + Xβ + P γ1 + N γ2 + , = ρW + µ,
(3)
where W is an (n × n) matrix defining who is a neighbor of whom by means of 0–1 values, and µ an error assumed to be independently (and identically) distributed. Erroneously omitting spatial correlation would create both bias and inefficiency in the estimated parameters (Anselin, 2006). The spatial weights matrix is typically standardized so that the sum of each row equals 1, which implies that the spatially lagged version of h contains the average h value of the neighbors. 7
The spatial version of equation (2) presented as (3) has several interesting features that will have major implications for the policy recommendations that can be derived from our model. First, the model in (3) now contains a spatially lagged dependent variable. In order to see how this affects the interpretation of our model it helps to rewrite (3) as: h = (1 − λW )−1 [Xβ + P γ1 + N γ2 + ε],
(4)
where (1 − λW )−1 is the spatial multiplier. The spatial multiplier can be written as I + λW + λ2 W 2 +..., where W contains the neighbors of an individual, W 2 the neighbors of the neighbors, and so forth (Pace and LeSage, 2007). Effectively, this implies that a person’s BMI is not only determined by his or her own characteristics in terms of X, P and N , but also by a person’s location in terms of the average values of X, P and N of the neighbors, the neighbors of the neighbors, etc. For instance, if vigorous physical activity has a negative effect on BMI an environment in which your neighbors are also active aggravates this effect (assuming lambda is positive). These kinds of spillover and feedback effects may be due to social network effects (e.g., imitation behavior, peer effects) or to shared neighborhood characteristics (e.g., availability of a park). Second, equation (4) is very general because it accommodates both spatial spillover and feedback effects in terms of BMI, and allows for potentially spatially correlated omitted variables represented by the error terms. Third, in terms of estimation the specification in equation (4) is not entirely straightforward, because the spatially lagged dependent variable is obviously endogenous, and the spatially correlated errors create a non-spherical error variance-covariance structure. We follow the estimation theory for spatial ARAR models developed recently by Kelejian and Prucha (2007). They propose a combination of instrumental variables and generalized method of moments techniques. As outlined in Arraiz et al. (2007) the estimation procedure takes place in a series of steps. In the first step we use the spatial two stage least squares estimator to estimate equation 3 ignoring the spatially correlated errors. Subsequently, we use the estimated residuals of the first step in a spatial general moments (GM) estimator to obtain an estimate of ρ. With that estimate in hand we apply a Cochrane-Orcutt transformation and re-estimate the model with spatial two stage least squares. The asymptotic variance-covariance matrix for this estimator is derived in Kelejian and Prucha (2007) under the assumption that the innovations are heteroskedastic. Finally, it is obvious from the specification given in equation (4) that the marginal effect of policy-induced changes in X, P or N are not simply equal to β, γ1 or γ2 . Effectively, the marginal effect depends on the spatial location of the policy-induced change, and the resulting spillover and feedback effects (Pace and LeSage, 2007). For example, the marginal effect on BMI of improving the food landscape through ’health zoning’ no longer solely depends on the strictness of the zoning policy measure alone, but also on the location where the zoning is implemented. This suggests that identifying and targeting optimal geographic Policy Implementation Areas (PIAs) as we will call them, can provide a more cost effective means of public health intervention.
8
5
Results
We start by estimating the BMI equation as outlined in equation (1) and equation (2) above. In these specifications BMI is assumed to be a function of individual demographic, behavioral, and neighborhood observed characteristics. The results from the estimation of these Ordinary Least Squares models are outlined in Table 2. Column (1) reports the results of a naive OLS model that only accounts for the food environment. These results suggest that having one more fast food restaurant increases BMI by 0.06 points while having one more large grocery decreases BMI by 0.34 points. Only chain grocers are significant at the 95% confidence interval. When individual heterogeneity is controlled for as outlined in equation (1), the magnitude of the associations between fast food and large grocers changes but the results are still insignificant for fast food (reported in Column (2)). The demographic variables included in this model have the predicted sign and significance reported in the literature (Mokdad et al., 2003). For example, BMI is negatively correlated with education and income, associated positively with being nonwhite, there is a quadratic relationship between BMI and age, and increased physical activity and smoking decrease BMI. In column (3) we provide an estimate of equation (2) where we attempt to control for sorting using an observable neighborhood variables defined as the amount of crime within a 12 mile radius of a person’s residence. We conjecture that the choice of where to live, (measured in our analysis by fast food), is strongly correlated with the level of crime. While crime may not ”belong” in the BMI equation, by putting it in there, crime can proxy and absorb the bias caused by the endogeneity of fast food. Adding this control function does not result in any change in the magnitude or significance of the food landscape variables. In addition, the number of serious crimes committed or attempted within 1/2 mile of where you live has no effect. As discussed earlier, there is reason to believe that the food environment is still endogenous in equation (2). Empirical evidence suggests that people usually drive to grocers even if they live in low income neighborhoods (Clifton, 2004). Consequently, while access and distance may enter into the price of food, the actual time cost of going an extra mile for grocery food is lower. On the other hand, when people want to eat quickly, they will go to the closest restaurant that they can find. As such, distance is very important when choosing among restaurants. It is, therefore, likely that the fast food environment is endogenous in the BMI equation. We, therefore, instrument the number of fast food restaurants using individualized residential zoning patterns. The first stage of the IV regression is presented in Table 7 in the Appendix. There is a strong positive correlation between non-residential zoning within a 21 radius of where a person lives and the number of fast food restaurants in that same radius. Our IV would not be valid if the people who lived in highly non-residential areas were different in observable and unobservable ways from those that do not. While we cannot investigate whether these two groups differ in unobservable ways we can examine whether they differ in observable ways. These results are presented in Table 8 of the Appendix and show that the group means are either similar in magnitude and if not, do not significantly differ from each other. In addition to the selection effect, it is also likely that observations are not independent across space because of either (unobserved) social network ties or shared (unobserved) neigh9
borhood characteristics across individuals living in proximate neighborhoods. To explore this issue we investigate a simple ad-hoc specification where we estimate a random effects model assuming the composite error term can be decomposed into a neighborhood effect measured at the census tract level and an additional random disturbance so that ωij = ui + µij . The random effects estimates are presented in column 2 of Table 3. They are not significantly different from the OLS estimates which are presented in column 1 for comparison purposes. One reason why spatial correlation in the error terms does not seem to matter in this context is because neighborhoods and markets are defined on a much smaller spatial scale than a census tract.5 To explore this issue further we abandon the ad-hoc specifications and the assumption that markets are defined at the census tract level (i.e. clustering at the census tract level) and we undertake a more systematic analysis of the spatial structure of the data. In order to assist with the visualization of individuals and to operationalize the spatial econometric model, specifically the spatial weights matrix, we transform individuals into Thiessen polygons based on their location. The transformation assigns every point in space to the nearest point for which we have an actual observation. The Thiessen polygons are a way to operationalize the model because they transform point data into space data. As Figure 2 shows, we now have contiguity of individuals across space.6 For the regressions with the spatial effects we used first and second order queen contiguity to define who is a neighbor of whom.7 The average number of neighbors is about 20. Preliminary exploration of the data was conducted using the following standard Exploratory Spatial Data Analysis (ESDA) techniques. First, a global Moran’s I was computed for respondent BMI across the entire sample space. Similar to a correlation coefficient, values for Moran’s I typically fall between -1 and +1, indicating strong negative to strong positive spatial correlation of observation values. For this data, the global Moran’s I was 0.019 with a p-value of 0.0004. A pictorial representation of the statistically significant clustering pattern of individuals in the data using Local Indicators of Spatial Association (LISA) is also presented in Figure 2. This map shows that there is clustering of relatively low BMIs in the northern central, northeastern, and south southeastern parts of the county with areas of high BMI flanking the center of the county to the east and west as well as south. We use Lagrange Multiplier (LM) tests for spatial dependence to determine whether the suggested ARAR specification in equation (3) is potentially adequate. The full results of the 5
One should note that the random effects specification does not really account for spatial correlation based on a distance decay pattern among all individuals. Effectively, the errors are correlated within census tracts, but uncorrelated between census tracts. The spatial correlation structure is, therefore, due to the incorporation of spatial heterogeneity by means of random neighborhood effects rather than spatial dependence among observations due to proximity. 6 Thiessen polygons are only needed because they are easier to visualize, and to derive the weights matrix on the basis of contiguity. They are not equivalent to aggregating the data into areal units. Other ways to operationalize this model would be to use distance measures. If we use (inverse) distance, the people in the city center get a disproportionately large number of neighbors because of the relatively large cut-off distance. 7 Queen contiguity is defined as neighbors who share either a common border or a vertex with the index person. If we consider a checkerboard, first order Queen contiguity refers to the immediate neighbors of an area, e.g., area i has 8 neighbors j1-j8. Second order contiguity refers to the neighbors of these 1st order neighbors. E.g., for the same case, there are 16 second order neighbors of the first order j1-j8.
10
spatial diagnostics are presented in Table 4. Comparison of the LM values and significance levels was conducted as suggested by Anselin et al. (1996). The diagnostic tests suggests an ARAR spatial process. In addition the heteroskedasticity tests suggest that the model specification should allow for heteroskedasticity. Estimation was conducted using code that we wrote implementing the GMM and IV techniques described above using the R2.7.1 software.8 The results of the spatial ARAR model as outlined in equation (3) are reported in Table 3, column 3. There are three interesting findings from this specification. First, the effect of fast food and chain grocers are both significant. As explained earlier, the coefficients of β cannot be interpreted as partial effects since they are impacted by the spatial multipliers as outlined in equation (4). In order to estimate partial effects, we have to select particular geographical areas or PIAs where the health policy will be implemented and simulate the effects of the policy change. This exercise will be conducted in the next section. The sign of β is still, however, relevant and shows that increasing fast food has a positive effect on BMI while increasing chain grocers have a negative effect on BMI. The second interesting finding is that the value of ρ is both significant and negative. A significant ρ value suggests that accounting for unobservable factors, (social network ties or unobserved neighborhood effects), is important. Finally, a significant and positive value of λ suggests that the BMI of an individual’s neighbors has an indirect effect on her own BMI. The positive sign of λ also suggests that the explanatory variables of my neighbors affects my BMI in the same direction as the parameter estimate β. We now turn to a discussion the simulations where we will estimate and present the average partial effects.
6
Policy Experiments
Municipal ordinances have traditionally been used to limit the number of liquor stores or pornographic outlets within cities. More recently, however, cities have looked at zoning regulations as a means to improve health outcomes, especially in disadvantaged neighborhoods. Some cities have targeted the “bad” and instituted mandatory nutritional information postings at fast food restaurants, moratoriums on the openings of new fast food restaurants, or implemented outright bans on the use of unhealthy trans fats in food preparation. Other municipalities, however, have focused on the “good” by providing increased access to healthy foods in neighborhoods. The parameter values estimated earlier can now be used to simulate the two policy experiments described. For the “bad” scenario we simulate the effect of restricting the density of fast food in areas that are currently overserved. An area is considered to be overserved if it has more than 6 fastfood restaurants per KM 2 . In the “good” simulation we investigate the implications of providing better access to healthy food, by identifying areas that have more than 40% of the population below the Federal Poverty Level and more than 40% with less than a high school diploma. In both the “good” and the “bad” scenario, we use data gathered online as part of the SEDAC project developed at Columbia University.9 The SEDAC project has data on Marion 8
R is available at http://cran.fhcrc.org/bin/windows/base/old/2.7.1/. The R code used to implement this estimator is available upon request. 9 Shape files for the SEDAC project are available and can be downloaded from
11
county at the grid cell level (1km x 1km rasters) with sociodemographic and economic characteristics attached for each cell. We use the geographical information on food establishments to attach the density of grocery stores and fast food to each one of these grid cells as well. We then use the grids to identify our Policy Implementation Areas (PIAs) for the “good” and “bad” policies. There were originally 563 fast food restaurants in our dataset. In the first policy scenario when we remove one fast food restaurant from high density fast food areas, we decrease the number of fastfood restaurants by 15. When we recalculate the number of fast food restaurants within a 1/2 radius of each person, 178 people are directly affected by the policy change i.e. the count of fast food in their neighborhood decreases. The marginal effects of restricting fast food density is presented in Figure 3 and summarized in Table 5. Since the marginal effect of the policy change is different for each individual depending on their geographic location we report average direct, indirect and total effects of restricting access to fast food outlets in selected PIAs in Table 5. The direct effect reported in Table 5, column 1, is the average partial effects on only those individuals whom are directly affected by the policy experiment divided byall individuals in our sample. The mean direct effect is negative and significant. Restricting fast food density to six restaurants per KM 2 decreases average BMI by -0.01 for the total sample. If instead we find the average over only the 178 people who live in the PIAs and have their access to fast food restricted then the average partial direct effect for this subgroup is larger in absolute value at -0.219.10 The average indirect effect represent the spatial spillover effects or the effect on the people who neighbor the neighborhoods that were subject to the density restriction on fast food. Restricting the density of fastfood also has a small negative average indirect marginal effect on BMI of approximately -0.031 BMI points. The total marginal effect of targeting fast food is the sum of the marginal direct and indirect effects. Restricting access to fastfood has total marginal effect of about -0.042 on BMI. It is also instructive to look at the spatial distribution of the total average effects displayed in Figure 3. It is clear from this diagram that the the change in policy has a different effect on each person depending on how far they live from the neighborhood where the policy is introduced. Moreover, the spatial econometric model, (because it essentially links people across space with a spatial multiplier), makes it clear that the effect of a new policy that attempts to change the food environment in one location can have ripple effects across space and thereby affect BMI in neighboring locations. As a result, even individuals who do not have a fast food or grocery outlet in their local food environment (i.e., the 1/2 mile buffer) are affected by the spatial distribution of these outlets through the spillover effects. The spatial heterogeneity of the partial effects and the diffusion process that occurs is evident by the shading patterns in Figure 3. It is also clear from examining this map that the actual location of the PIA is also important because the marginal impact changes according to where the policy is implemented. In the second policy scenario we increase access to healthy foods by locating grocery stores in disadvantaged neighborhoods. This policy increases the number of grocery stores from 94 to 107. This change directly increases access for 74 people in our sample. The average marginal http://sedac.ciesin.columbia.edu/usgrid/. 10 This is just a rescaling of the direct effect by 3, 550/178 where the denominator is the number of people who actually live in the PIA.
12
effects for increasing access to healthy foods, are displayed in Table 6. All three statistics reported in Table 6 are significant at the 95% confidence interval. The direct effect for those individuals who live in the PIA , (averaged over the total sample), is -0.012. Once again, if we average only over the subset of people who live in the PIA the direct effect is -0.576. The indirect effects are -0.035 indicating that people who are located in neighborhoods where a policy change is not implemented also benefit from more grocery stores in proximate neighborhoods. The total effect of increasing access to healthy food is then -0.048. In Figure 4 we once again see the same spatial diffusion pattern from the neighborhoods where the policy change is implemented to the surrounding neighborhoods. As described earlier, the PIAs have the largest partial effects and from there the marginal effects diffuse and decay as we move away from the PIAs.
7
Conclusion and Discussion
This study differs from the previous literature because of the spatial nature of the analysis. It shows how the magnitude of the marginal effects depends on the exact location of the individuals for which marginal (policy) effects are determined. It also outlines how spatial econometric methodology may be used as a useful tool for local policymakers who want to understand how specific neighborhood policies can impact the health of the local community, and whether these policies affect (spillover into) neighboring communities. Our results suggest that past attempts to explore the relationship between the food landscape and obesity are hindered by sample selection issues. Further analysis of the data also indicates that spatial dependence shows up in BMI values and in the regression error term. When selection and spatial dependence is accounted for in our estimation the marginal effects of changing access to either fast food or chain grocers is significant but small. When we translate our results from BMI to a more widely understood measure such as pounds, the total average effect of improving grocery store access in targeted disadvantaged neighborhoods is less than a pound decrease (calculated using a population of 5’6” females that weighs 160 pounds) on average for the entire Marion county. Restricting access to fast food, has effect of the same magnitude. If you live in the area where the grocery stores is placed, however, the effects are much larger and translate into less than a 5 pound decrease in weight. For fast food it is about a 2 pound decrease for the people who live in the areas where fast food is restricted. A cornerstone of urban renewal is to develop the kind of infrastructure that creates and maintains prosperous neighborhoods. Oftentimes, this includes incentives to attract large chain groceries or to increase the healthy food offerings of existing small grocery and convenience stores. The analysis conducted in this paper demonstrates that policies that target either the “good” or the “bad” aspects of the food landscape have a relative negligible impact on the entire county. For the people who live in the local neighborhoods where the policy takes place, however, the policy change has a more significant impact.
13
References Abdollah, T. (2007). A strict order for fast food. Los Angeles Times September 10,2007. Agrawal, A. W. and P. Schimek (2007). Extent and correlates of walking in the USA. Transportation Research Part D: Transport and Environment 12D(8). Anselin, L. (2006). Spatial econometrics. Palgrave Handbook of Econometrics: Econometric Theory, 901–969. Arraiz, I., D. Drukker, H. Kelejian, and I. Prucha (2007). A spatial Cliff-Ord-type model with heteroskedastic innovations: Small and large sample results. Unpublished manuscript. Block, J., R. Scribner, and K. DeSalvo (2004). Fast food, race/ethnicity, and income A geographic analysis. American Journal of Preventive Medicine 27 (3), 211–217. Burdette, H. and R. Whitaker (2004). Neighborhood playgrounds, fast food restaurants, and crime: relationships to overweight in low-income preschool children. Preventive Medicine 38 (1), 57–63. Chou, S., M. Grossman, and H. Saffer (2004). An economic analysis of adult obesity: results from the Behavioral Risk Factor Surveillance System. Journal of Health Economics 23 (3), 565–587. Christakis, N. A. and J. H. Fowler (2007). The Spread of Obesity in a Large Social Network over 32 Years. New England Journal of Medicine 357 (4). Chung, C. and S. Myers (1999). Do the poor pay more for food? An analysis of grocery store availability and food price disparities. Journal of Consumer Affairs 33 (2), 276–296. Clifton, K. (2004). Mobility Strategies and Food Shopping for Low-Income Families: A Case Study. Journal of Planning Education and Research 23 (4), 402. Cummins, S. and S. Macintyre (2006). Food environments and obesity-neighbourhood or nation? International Journal of Epidemiology 35 (1), 100–104. Eid, J., H. Overman, D. Puga, and M. Turner (2006). Fat City: Questioning the Relationship Between Urban Sprawl and Obesity. Unpublished manuscript. Indianapolis Metropolitan Police Department (2007). Uniform Crime Reporting [online]. Indianapolis Metropolitan Police Department, updated February 20, 2008 [cited May 5, 2008]. Jeffery, R., J. Baxter, M. McGuire, and J. Linde (2006). Are fast food restaurants an environmental risk factor for obesity. International Journal of Behavioral Nutrition and Physical Activity 3 (2). Kelejian, H. and I. Prucha (2007). Specification and estimation of spatial autoregressive models with autoregressive and heteroskedastic disturbances. University of Maryland, forthcoming in Journal of Econometrics. 14
Maddock, J. (2004). The Relationship Between Obesity and the Prevalence of Fast Food Restaurants: State-Level Analysis. American Journal of Health Promotion 19 (2), 137–143. Mair, Julie Samia, M. W. P. and S. P. Teret (2005). The use of zoning to restrict fast food outlets: A potential strategy to combat obesity. The Center for Law and the Public’s Health at Johns Hopkins and Georgetown Universities. Mobley, L., E. Finkelstein, O. Khavjou, and J. Will (2004). Spatial analysis of body mass index and smoking behavior among WISEWOMAN participants. Journal of Women’s Health (Larchmt) 13 (5), 519–28. Mokdad, A., E. Ford, B. Bowman, W. Dietz, F. Vinicor, V. Bales, and J. Marks (2003). Prevalence of Obesity, Diabetes, and Obesity-Related Health Risk Factors, 2001. JAMA 289 (1), 76–79. Moore, L. and A. Diez Roux (2006). Associations of neighborhood characteristics with the location and type of food stores. American Journal of Public Health 96 (2), 325–331. Morland, K., A. Diez Roux, and S. Wing (2006). Supermarkets, Other Food Stores, and Obesity: The Atherosclerosis Risk in Communities Study. American Journal of Preventive Medicine 30 (4), 333–339. Morland, K., S. Wing, and A. Diez Roux (2002). The Contextual Effect of the Local Food Environment on Residents’ Diets: The Atherosclerosis Risk in Communities Study. American Journal of Public Health 92 (11), 1761–1768. Morland, K., S. Wing, A. Diez Roux, and C. Poole (2002). Neighborhood characteristics associated with the location of food stores and food service places. American Journal of Preventive Medicine 22 (1), 23–29. Pace, R. and J. LeSage (2007). Interpreting spatial econometric models. Working Paper . Philipson, T. and R. Posner (2003). The Long-Run Growth in Obesity as a Function of Technological Change. Perspectives in Biology and Medicine 46 (3), S87–S107. Rose, D. and R. Richards (2004). Food store access and household fruit and vegetable use among participants in the US Food Stamp Program. Public Health Nutrition 7 (8), 1081–8. Zenk, S., A. Schulz, B. Israel, S. James, S. Bao, and M. Wilson (2005). Neighborhood racial composition, neighborhood poverty, and the spatial accessibility of supermarkets in metropolitan detroit. American Journal of Public Health 95 (4), 660–667.
15
A
Tables And Figures
Table 1: Summary statistics Variable Body Mass Index Obese Overweight Healthy Weight No. of Fast Food - 12 mile No. of Large groceries - 12 mile Nonwhite Female Age Less than 200% of the FPL More than high school Vigorous Physical Activity per Week Physically Demanding Job Smoker Number of serious crimes within 12 mile Percent Zoned Non-residential
16
Mean 27.675 0.272 0.355 0.362 2.033 0.354 0.303 0.583 47.011 0.207 0.645 2.682 0.413 0.259 41.406 0.322
Std.Dev. 6.099 0.445 0.479 0.481 2.86 0.722 0.46 0.493 14.231 0.405 0.479 2.348 0.492 0.438 43.86 0.214
Table 2: Linear Regression Results Variable No. of Fast Food - 12 mile No. of Large groceries -
1 2
mile
(1) 0.055 (0.046) -0.303* (0.181)
(2) 0.056 (0.044) -0.167 (0.175) 1.035** (0.224) -0.357* (0.204) 0.403** (0.048) -0.004** (0.000) 1.207** (0.265) -0.875** (0.224) -0.319** (0.044) -0.611** (0.213) -1.321** (0.233)
27.669** (0.126) 0.0002 3550
19.907** (1.129) 0.0635 3550
Nonwhite Female Age Age Squared Less than 200% of the FPL More than high school Vigorous Physical Activity per Week Physically Demanding Job Smoker Number of serious crimes within
1 2
mile
Constant R-squared N * p < 0.10, ** p < 0.05.
17
(3) 0.052 (0.044) -0.159 (0.176) 0.972** (0.229) -0.343* (0.204) 0.404** (0.048) -0.004** (0.000) 1.158** (0.268) -0.854** (0.225) -0.320** (0.044) -0.603** (0.214) -1.339** (0.233) 0.003 (0.002) 19.777** (1.133) 0.0637 3550
Table 3: Results of Models that Control for Spatial Dependence Variable No. of Fast Food -
1 2
mile
No. of Large groceries -
1 2
mile
Nonwhite Female Age Age Squared Less than 200% of the FPL More than high school Vigorous Physical Activity per Week Physically Demanding Job Smoker Number of serious crimes within
1 2
mile
Constant
(1) OLS 0.052 (0.044) -0.159 (0.176) 0.972** (0.229) -0.343* (0.204) 0.404** (0.048) -0.004** (0.0005) 1.158** (0.268) -0.854** (0.225) -0.320** (0.044) -0.603** (0.214) -1.339** (0.233) 0.003 (0.002) 19.777** (1.133)
λ (wy) ρ (we) R-squared N
0.0637 3550
(2) (3) RE ARAR 0.054 0.201** (0.045) (0.100) -0.156 -0.481* (0.182) (0.283) 0.944** 0.506** (0.244) (0.218) -0.347* -0.336* (0.194) (0.186) 0.405** 0.393** (0.047) (0.046) -0.004** -0.004** (0.0005) (0.0005) 1.136** 1.059** (0.304) (0.294) -0.819** -0.641** (0.238) (0.233) -0.318** -0.302** (0.044) (0.043) -0.614** -0.576** (0.209) (0.207) -1.349** -1.368** (0.228) (0.219) 0.003 -0.001 (0.003) (0.002) 19.730** -0.821 (1.099) (2.942) 0.753** (0.097) -0.651** (0.131) 3550
Standard errors robust to clustering at the census tract level. * p < 0.10, ** p < 0.05.
18
3550
Table 4: Spatial Diagnostics Test Heteroskedasticity Random Coefficients Breusch-Pagan test Koenker-Bassett test Spatial dependence* Moran’s I (error) Lagrange Multiplier (lag) Robust LM (lag) Lagrange Multiplier (error) Robust LM (error) Lagrange Multiplier (SARMA)
Value
p-Value
248.218 100.154
0.000 0.000
2.03 5.269 5.156 3.365 3.252 8.521
0.042 0.022 0.023 0.067 0.071 0.014
*For Queen First and Second Order Contiguity
19
#
#
## #
Figure 1: Local Food Environment
# ## # # # #
### # ###
ò ÿÿ
# ##
# #
#
##
# ## # ## # # ò ÿ### # # # ÿ##ÿ
#
## ## # ##
# # #
#
# # # # ##
#
## #
#
# ###
ÿ
## ÿ#ÿÿ
#
ÿò ÿÿò ÿÿ###ÿ##
#
#
ÿ##
ÿ #ÿ# #
#
# # # ## # #
# ### ## ##
#
# # #
#
# ## ## # #
# # ## ## #
## ### ##
# #
# ##
# # # #
#
# ## ## #
ÿ ò ÿ #### ÿ####
#
#
#
#
## ###
# # ###
#
#
#
#
## #
BMI
# ### ##
#
#
# # ###
#
##
# # #
#
###
#
## # ## #
# # ## # #
ÿ ÿÿ ÿÿ ÿò # # ## #
# #
## #
## #
# #
ÿò #ÿ # ÿÿ###### # #ÿ ## # ÿÿò # # ÿ
# ##
overweight
#
obese
# # # ## ## ## # #
ÿ
Chain store Fast food outlet
Median family income
#
#
#
#
18625 - 33750 33751 - 45769
# ####
# #
#
# #
45770 - 58684
ÿ
## #
# #
20
healthy
#
ò
# # # ###
#
#
58685 - 78990 78991 - 140217
Figure 2: Local Indicators of Spatial Association for BMI
21
Total Impact of the Good Scenario
Figure 3: Marginal Effects of Changing the Fast Food Environment on BMI
under -0.0677 -0.0677 - -0.0226 -0.0226 - -0.0104 -0.0104 - -0.0058 -0.0058 - -0.0031 -0.0031 - -0.0015 -0.0015 - -5e-04 over -5e-04
Table 5: Marginal Effects of Restricting Access to Fast Food in Targeted Areas on BMI Average Direct Effect* Estimated Mean -0.011 Simulated Standard Error 0.002 Simulated z -5.852
22
Average Indirect Effect -0.031 0.007 -6.161
Total Average Effect -0.042 0.006 -8.816
Total Impact of the Good Scenario
Figure 4: Marginal Effects of Increasing Access to Healthy Food in At Risk Areas on BMI
under -0.066 -0.066 - -0.0068 -0.0068 - -0.0014 -0.0014 - -4e-04 -4e-04 - -1e-04 -1e-04 - 0 0-0 over 0
Table 6: Marginal Effects of Increasing Access to Healthy Food in At Risk Areas on BMI Average Direct Effect* Estimated Mean -0.012 Simulated Standard Error 0.002 Simulated Z -6.737
23
Average Indirect Effect -0.035 0.008 -4.131
Total Average Effect -0.048 0.010 -4.650
B
Appendix Table 7: First Stage of IV Specification Variable No. of Large groceries 12 mile Nonwhite Female Age Age Squared Less than 200% of the FPL More than high school Vigorous Physical Activity per Week Physically Demanding Job Smoker % of Land Zoned Non-Residential Constant R-squared N * p < 0.10, ** p < 0.05.
24
1 2
mile
Estimate 2.2595** (0.1850) 0.0671 (0.1112) 0.0296 (0.0775) -0.0280 (0.0194) 0.0002 (0.0002) 0.0840 (0.1216) -0.0165 (0.0907) -0.0182 (0.0148) 0.1493 (0.0914) -0.0569 (0.0869) 2.2742** (0.4141) 1.2379** (0.4610) 0.4128 3550
Table 8: Further Tests of Validity of the Instrumental Variable Variable High Low t-statistic Body Mass Index 27.975 27.448 2.545 Nonwhite 0.350 0.268 5.237 Female 0.592 0.575 1.006 Age 46.447 47.446 -2.067 Less than 200% of the FPL 0.250 0.175 5.511 More than high school 0.608 0.673 -4.008 Vigorous Physical Activity per Week 2.646 2.704 -0.727 Physically Demanding Job 0.430 0.400 1.801 Smoker 0.279 0.244 2.387 1.579 Number of serious crimes within 12 mile 42.811 40.457 Low and High Represent people who have less than the mean and more than the mean percent around them zoned non-residential respectively.
25
