National Poverty Center Gerald R. Ford School of Public Policy, University of Michigan www.npc.umich.edu Residential Environments and Obesity: What Can We Learn about Policy Interventions from Observational Studies? Jeffrey D. Morenoff, University of Michigan Ana V. Diez Roux, University of Michigan Theresa Osypuk, University of Michigan Bendek Hansen, University of Michigan This paper was delivered at a National Poverty Center conference. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of the National Poverty Center or any sponsoring agency. Residential Environments and Obesity: What Can We Learn about Policy Interventions from Observational Studies? Jeffrey D. Morenoff Ana V. Diez Roux Theresa Osypuk Ben Hansen University of Michigan This paper was prepared for the National Poverty Center’s “Health Effects of Non-Health Policy” Conference in Bethesda, Md., February 9-10, 2006. The motivating question for this chapter is whether policy initiatives aimed at changing physical and/or social features of the residential environment would have measurable impacts on health.1 It has become an increasingly popular view in the field of public health that physical and social features of the residential environment can affect health either directly (e.g., through contaminants in the air or water supply) or indirectly, by influencing behaviors related to health (e.g., physical activity, food intake substance use, utilization of medical care) and/or psychosocial dispositions (e.g., stress and social relationships that help cope with stress) that may be related to health through more complex causal chains (2002). Unfortunately, there are few studies examining the health impacts of policy initiatives that specifically target residential environments. Still, an increasingly large body of observational studies, most of them conducted by social epidemiologists and sociologists, demonstrates that there is considerable variation in mortality and other health outcomes across local geographic areas (Diez Roux 2001; Morenoff and Lynch 2004) and that this variation is related to area socioeconomic status (e.g., poverty rates and other measures of disadvantage and affluence) and appears to be at least partly independent of personal measures of socioeconomic position.. All of this suggests that health is adversely affected by living in impoverished neighborhoods (Robert 1998). However, this research has been criticized on the grounds that people self-select into neighborhoods on the basis of characteristics that are not assessed directly in observational studies on health or that are assessed with considerable measurement error, and that neglecting these selection factors can lead to biased (and perhaps inflated) estimates of the causal effects of neighborhood residence on health (Kling, Liebman, and Katz 2005; Oakes 2004). Another important criticism is that most studies examining the link between residential 1 In this chapter we use the terms “residential environment” and the term “neighborhood” interchangeably but prefer the former because the geographic boundaries used to demarcate residential environments usually do not attempt to represent neighborhoods as perceived by the residents. 1 environments and health have relied on measures of socioeconomic status (usually taken from the census) to explain area variation in health, but they have not identified specific causal mechanisms, especially those relating to features of the physical or social environment that can be manipulated through interventions. In this chapter we consider the question of how interventions that target residential environments might affect health outcomes related to obesity by examining data from a recent study of adult health in Chicago neighborhoods. We focus on obesity because the rapid rise in obesity rates in the United States over the past two decades (Mokdad et al. 1998) cannot be explained solely in terms of biological/genetic factors (French, Story, and Jeffery 2001; Hill and Peters 1998) and because prior theory and research has identified specific features of residential environments that might be causally linked to physical activity and food intake, both proximal determinants of obesity. Obesity is also one of the few health outcomes for which there is experimental evidence of a connection to residential environments. In the Moving to Opportunity study (U.S. Department of Housing & Urban Development 2003), which was conducted in five major metropolitan areas, adults who were randomized to the “experimental” treatment group (receiving a voucher to move to low-poverty (<10% poverty rate) neighborhoods and housing counseling) exhibited significantly (p<.05) lower obesity rates when observed at four to seven years following randomization, compared to the in-place control group (receiving no vouchers or counseling).2 Despite this important experimental study, systematic evidence is still lacking on 2 Those randomized to the Section 8 treatment group (receiving a voucher good for moving to any neighborhood, regardless of neighborhood poverty) also exhibited lower rates of obesity than control group at follow-up, although the difference was only marginally significant. There was also some evidence of parallel changes in nutrition and physical activity, which could contribute to the changes observed in obesity. The Experimental group exhibited significantly higher consumption of fruits and vegetables versus controls at follow-up, although the Section 8 group did not; conversely, the Section 8 group reported higher rates of moderate physical activity versus controls, whereas the Experimental group did not (U.S. Department of Housing & Urban Development 2003). 2 the extent to which body mass and obesity vary across residential contexts, making the current study an important contribution to this literature. Our discussion of residential environments and obesity proceeds in three stages. First, we present results from a descriptive analysis of how obesity-related outcomes and social disparities in obesity vary across residential environments. Second, we consider specific aspects of the built environment that may account for these differences across residential neighborhoods in obesity and physical activity. Finally, we reflect on some of the criticisms that have been launched at observational studies of residential environments and health and suggest some directions for future research. Residential Environments and Obesity in Chicago We begin with two basic questions about residential environments and obesity: Do obesity and related behaviors vary significantly across neighborhoods, and are racial/ethnic and educational disparities in obesity largely the result of the different residential environments in which these groups live? There has been very little research to date on whether obesity varies systematically across residential environments due to the lack of appropriate data sources. We analyze data from a new study designed explicitly to investigate the role of residential environments, in conjunction with individual and household factors, in affecting adult health: the Chicago Community Adult Health Study (CCAHS). In this study, face-to-face interviews were conducted and physical/biological measurements were made on a stratified, multistage, probability sample of 3105 adults, aged 18 and over and living in 343 neighborhood clusters (NCs) within the city of Chicago between May, 2001 and March, 2003. One individual per household was interviewed, and the response rate was 71.82%. The NCs were defined for a 3 previous study (Sampson, Raudenbush, and Earls 1997) as aggregations of census tracts (the typical NC consists of two census tracts) with meaningful physical and social identities and boundaries.3 All data in the ensuing analysis are weighted to take account of the different rates of selection as well as household size and differential coverage and nonresponse across NCs.4 The obesity-related outcomes that we analyze – direct measurements of body mass index (BMI) and waist size, and survey-based measures of exercise, walking, and fruit and vegetable intake – are all defined in Table 1.5 To assess the extent to which obesity-related outcomes vary across residential environments, we rely on the intra-cluster correlation (ICC), ρ, a commonly used method for assessing contextual variation, which is defined as the fraction of variance in an outcome that lies between neighborhoods. Table 2 presents intra-cluster correlation coefficients for each outcome broken down by gender. We first consider the unadjusted ICCs, which we estimated via an unconditional random effects model in which persons vary around a neighborhood mean and neighborhoods vary around a grand mean. This model is written as follows: Yij = γ00 + u0j + rij for person i in neighborhood j, with rij ~ N(0, σ2) and u0j ~ N(0, τ00), and where ρ = τ00 / (τ00 + σ2). We estimated ρ on each outcome first for the full sample (3105 persons nested within 343 NCs), and then separately for women (1870 persons nested within 341 NCs) and men (1235 3 The neighborhood data in CCAHS were collected in collaboration with the Project on Human Development in Chicago Neighborhoods (PHDCN) and they included a second wave of the PHDCN Community Survey and Systematic Social Observation. 4 The weighted sample matches the 2000 Census population estimates for the city of Chicago in terms of age, race/ethnicity and sex. The weights also take into account different rates of subsampling for final intensive interview completion efforts. 5 For the purposes of this chapter, we treat the measures of exercise, walking, and fruit/vegetable intake as continuous variables. However, we realize that this entails some potentially problematic assumptions and are pursuing analyses in which these variables are treated categorically. 4 persons nested within 325 NCs), to take into account gender differences in the extent to which obesity-related outcomes vary across residential environments. For comparison, in educational and organizational research, ICCs tend to range between five and 20 percent (Hox 2002; Snijders and Bosker 1999), but for health outcomes, ICCs tend to be smaller (Morenoff 2003), rarely surpassing 10 percent. It is also important to keep in mind, as Duncan and Raudenbush (1999) demonstrate, that small ICCs can still translate into moderate or large sizes for the main effects of specific neighborhood-level attributes. Table 2 shows that the unadjusted ICCs are relatively high for all of the obesity-related outcomes. In the analysis of the full sample, the ICCs for BMI, waist size, exercise, and walking are all over nine percent, indicating that there is substantial between-area variation in all of these outcomes. We found a somewhat lower ICC in the full sample for fruit and vegetable intake (6.69), a single survey item that is likely measured with substantial error. A second finding from Table 2 is that the ICCs tend to be larger for women than for men, especially in the case of BMI, waist size, and exercise. For example, 17.78 percent of the variance in unadjusted BMI among women lies between neighborhoods, but the comparable ICC for men is only 7.42 percent, and this gender difference is even larger in the case of waist size. These results suggest that women’s health may be more sensitive to factors in the residential environment than men’s health. It is not clear from the extant literature why the proportion of neighborhood variance might be larger for women, but there is evidence from prior studies that socioeconomic disparities in obesity tend to be wider among women than among men (Sobal and Stunkard 1989). A variety of social and biological factors that vary by gender could contribute to the greater social patterning of BMI seen in women generally. 5 Since allocation of persons to neighborhoods is decidedly non-random, depending on household income, family size, ethnicity, and a host of other personal characteristics that are also correlates of health outcomes, ρ, unadjusted for these personal characteristics, represents the combined contribution of the causal effects of interest plus the variation in the outcomes associated with variation in the personal characteristics that select persons into neighborhoods. A better indicator of neighborhood variance can be obtained by estimating a two-level model in which neighborhood random effects are adjusted for person-level background characteristics.6 The value of ρ defined on these adjusted random effects provides an unbiased estimate of the amount of within neighborhood clustering under linear model assumptions and under the strong assumption that all relevant covariates have been included in the model. Because both of these assumptions are potentially problematic, we interpret the adjusted ICCs as more refined descriptive indicators of neighborhood variance in obesity-related outcomes that persists after purging the effects of measured person-specific characteristics. Table 2 shows that the ICCs for BMI and waist size decline after adjusting for personlevel covariates, but that substantial neighborhood variance remains for both outcomes among women (the adjusted ICC for women is 8.98 for BMI and 9.56 for waist size). Interestingly, exercise and walking are not affected by the adjustment for person-level correlates, which 6 To the extent that the predictors of health at the neighborhood level are correlated with person-level covariates, such an adjustment will produce a downward bias in the contribution of neighborhoods (Bingenheimer and Raudenbush 2004). This bias can be removed in a three-step procedure. First, we regress each outcome on a set of personal background characteristics including sex, age, race/ethnicity, immigrant generation, education, and income, with neighborhood fixed effects. This provides so-called “within-neighborhood” estimates of the effects of personal background characteristics that are purged of any correlation with neighborhood characteristics. Second, we adjust each outcome for individual background characteristics using the estimates of the coefficients from the fixed effects model, as follows: y ij* = y ij − βˆ w′ X ij , where βˆ w′ are the “within-neighborhood” coefficients for the vector of covariates, X, as estimated from the fixed-effects regression model described in the first step, and the yij* are the adjusted values of each outcome. Third, we decompose the variance in each adjusted outcome by estimating an unconditional random effects model and calculating an adjusted ρ as the fraction of variation in the adjusted outcome that lies between neighborhoods. 6 suggests that if residential environments do influence these behaviors, the causal mechanisms are probably not strongly related to the demographic and socioeconomic composition of local areas. The adjustment for person-level covariates accounts for a substantial fraction of the neighborhood variance in fruit and vegetable intake among women. Among men, however, the adjusted ICC for fruit and vegetable intake is slightly larger than the unadjusted ICC, implying that the confounding influences we control for were actually suppressing neighborhood variance.7 Decomposing Racial/Ethnic and Educational Disparities in Obesity-Related Outcomes Our second question is whether racial/ethnic and educational disparities in body mass are themselves driven by the differences across residential environments in which these groups tend to live. One of the primary reasons for the current popularity of multilevel approaches to studying health in social epidemiology is the preoccupation with socioeconomic and racial/ethnic disparities in health that persist after controlling for an array of individual-level factors, leading many scholars to speculate that the segregation of social groups into different residential environments may explain some of these disparities (Morenoff and Lynch 2004; Williams and Collins 1995). Prior studies have found significant social disparities in obesity. For example, studies of socioeconomic position and weight or weight gain in the United States and other 7 The “within neighborhood” coefficients from the regressions predicting fruit and vegetable intake among men revealed one surprising result that could account for this suppressor effect: men with annual household incomes of less than $5,000 report eating more fruits and vegetables than men with household incomes of $50,000 or more. One possible explanation of the suppressor effect is that despite the propensity for low income men to consume more fruits and vegetables, the residential environments in which low-income men live generally discourage fruit and vegetable intake (e.g., fruits and vegetables are not widely available in stores or local restaurants). In this case, adjusting for person-level income among men would reveal neighborhood variance in fruit and vegetable intake that was previously suppressed. [NOTE TO REVIEWERS: Another possible explanation for this suppressor effect is that our income variable still has not yet been cleaned, and it is likely that some of the men who report making less than $5,000 a year may be miscoded on income. We are working on cleaning this variable and will re-run our analysis when it is done.] 7 developed countries have shown that obesity is inversely associated with socioeconomic position, but more so for women than for men – some studies show a positive relationship between socioeconomic position and obesity for men (Sobal and Stunkard 1989). There are also substantial racial-ethnic disparities in weight and weight gain, with non-Hispanic blacks and Hispanics manifesting higher levels than whites, again especially in the case of women (add cites). No prior study has examined the extent to which neighborhood differences account for these social disparities in obesity. Using the CCAHS data, we decompose racial/ethnic and socioeconomic disparities in body mass into their within- and between-neighborhood components to assess how much of these disparities stem from the segregation of social groups across areas as opposed to body mass disparities that persist within geographic areas. We define the total health disparity between two groups as the ordinary least squares (OLS) estimate of the association between a covariate (e.g., ethnic group membership or education) and the health outcome. Raudenbush and Bryk (2002: 137) show that the total disparity (βt) is a weighted average of within-context (βw) and betweencontext (βb) components.8 For clarification, βw represents the expected difference in Y between two people who live in the same neighborhood but are of different social groups (defined by X), whereas βb represents the expected difference between the mean of Y in two neighborhoods that differ in one unit on the mean of X (e.g., proportion black, proportion Hispanic, or proportion of 8 We decompose disparities between groups defined by race/ethnicity and education by estimating the following hierarchical model for each group contrast: Yij = γ 00 + γ 01 X ⋅ j + γ 10 ( X ij − X ⋅ j ) + u0 j + rij , where Yij is the body mass index for person i in neighborhood j, X is a dummy variable defining a particular group contrast, X ⋅ j is the mean of X within neighborhood j, γ01 is βb, the effect of between-neighborhood differences in X, and γ10 is βw, the pooled-within-neighborhood disparity between the groups. 8 people with less than 12 years of education). Following Raudenbush and Bryk (2002), we define the “contextual” effect of being in a certain social group (βc) as the difference between the neighborhood-level relationship and the person-level effect, such that βc = βb – βw.9 Conceptually, the contextual disparity represents the expected difference in outcomes between two individuals who share the same group status (as defined by X) but who live in neighborhoods differing by one unit in the mean of X (e.g., proportion black, proportion Hispanic, or proportion of people with less than 12 years of education). We decompose disparities in BMI for comparisons of blacks to whites, Hispanics to whites, and people with less than 12 years of education to people with 16 or more years of education, and we display the results in Table 3. Because of the gender differences we observed in the variance of BMI across neighborhoods, we stratify the analysis in Table 3 by gender. One noticeable pattern is that the total disparities are larger among women than they are among men. For example, black women have BMIs that are 4.66 points higher than those of white women, while the black-white difference among men is only 1.25 points (which is still statistically significant).10 Hispanics have higher BMIs than whites, but the gap is wider among women (3.58 points) than it is among men (2.08). Also, the BMI difference between women with less than 12 years of education and those with 16 or more years of education is 4.14 points, but among men there is no significant difference between educational groups. Among women, the between-neighborhood component is quite large for all three group disparities, and although relatively large within-neighborhood disparities are also present, we find significant contextual 9 The contextual effect, βc, can also be estimated directly using the following model: Yij = γ 00 + γ 01 X ⋅ j + γ 10 ( X ij − X ⋅⋅ ) + u0 j + rij , where γ01 is βc and γ10 is βw (Raudenbush and Bryk 2002). 10 The mean BMI for the full sample is 28.60 with a standard deviation of 6.77. According to CDC definitions (http://www.cdc.gov/nccdphp/dnpa/obesity/defining.htm), a BMI of 25.0-29.9 is considered “overweight,” and a BMI of 30 or more is considered “obese.” 9 disparities between Hispanic women and white women and between women with less than 12 years of education and women with 16 or more years of education. This means that when Hispanic women share the same neighborhoods with white women, or likewise, people in the low education group live in the same neighborhoods as people in the high education group, the BMI gap between groups diminishes. It also means that women have higher BMIs when they live in neighborhoods with greater shares of Hispanics and people with less than 12 years of education. A more refined decomposition uses a set of multivariate prediction models presented in Table 4, stratified once again by gender. We first run OLS models to derive estimates of total disparities between racial/ethnic and educational groups after controlling for other individuallevel covariates (e.g., age, income, and immigrant generation). We then add to these models census-based contextual measures of the proportion of people in the neighborhood who are white, Hispanic, and who have 16 or more years of education, and we estimate the model with a random intercept.11 The neighborhood-level coefficients from the random intercept models represent contextual disparities that have been adjusted for other individual- and neighborhoodlevel covariates. The OLS models in Table 4 show that the BMI gaps between blacks and whites and between Hispanics and whites are slightly smaller but still significant after controlling for other covariates (compare the OLS estimates in Table 4 to the total disparities in Table 3) though still significant, and that they continue to be higher among females compared to males. Educational disparities in BMI are only significant among women and they are substantially smaller after controlling for other covariates; the BMI gap between women with less than 12 11 In Table 3 we estimated contextual effects for blacks, Hispanics, and people with less than 12 years of education, but here we use the proportions of whites and people with greater than 16 years of education because the latter specification presents fewer problems with multicolinearity – proportions of blacks, Hispanics, and people with less than 12 years of education are highly intercorrelated. 10 years of education and 16 or more years of education is only 2.35 points, compared to 4.14 points in Table 3. The random effects models in Table 4 show that both women and men who live in neighborhoods with larger shares of people who have 16 or more years of education tend to have significantly lower BMI, net of other individual- and neighborhood-level controls. Moreover, accounting for the contextual effect of education further reduces the withinneighborhood BMI gap between the lowest and highest educational group (it is only 1.82 points among females in Table 4). Men who live in neighborhoods with higher proportions of whites tend to have higher BMIs, but this relationship is not significant among women. Among males, living in a neighborhood with more whites is associated with higher BMI, although living in a neighborhood with more highly-educated people in the neighborhood is independently associated with lower BMI.12 The Hispanic composition of a neighborhood is not significantly related to BMI among either women or men (it had been significantly related to BMI among females in Table 3 but this effect did not hold up after introducing the person- and neighborhood-level controls in Table 4). Thus far we have demonstrated that a variety of obesity-related outcomes – BMI, waist size, exercise, walking, and fruit and vegetable intake – vary across neighborhoods, even after adjusting for individual-level demographic and socioeconomic factors that may account for some of the selection of people into neighborhoods, and that this variation is stronger among women than it is among men. We also found that contextual variation in BMI is strongly related to the educational composition of neighborhoods – but not to neighborhood racial composition – and that educational disparities in BMI appear to be highly contingent upon neighborhood context. 12 We note as a caveat that there is a high correlation (r = 0.68) between proportion white and proportion of people with 16 or more years of education. When the contextual education variable is removed from the model, the coefficient of proportion white remains positive but becomes non-significant (the point estimate decreases to 1.11, but the standard error remains fairly stable at 1.04). So the significant association between neighborhood proportion white and BMI among men could be the artifact of an over-specified model. 11 In other words, when we compare individuals with different educational backgrounds who live in the same neighborhoods, we find that the educational gap in BMI is much smaller than it is when we compare people with different educational backgrounds who live in neighborhoods that are also stratified by education. These findings suggest that there may be a neighborhood component to obesity and physical activity and that neighborhoods may be especially key for understanding educational disparities, but we have yet to identify specific mechanisms that might account for such a relationship and that would be amenable to policy interventions. Next we explore what some of these mechanisms might be. The Role of the Built Environment One of the reasons we chose the case of obesity to illustrate the connections between health and residential environments is that it is easy to conceive of tangible ways in which neighborhoods might influence how physically active a person is and the quantity and type of food a person eats. Much of the recent scholarship on obesity and residential environments has focused on the role of the so-called “built environment,” including physical structures and other elements of human-made environments, such as stores, parks and recreational facilities, street grids, transportation structure, land use, and other features of urban design (Frank and Engelke 2005). If the built environment around where one lives does influence patterns of physical activity and eating, then we would expect these effects to manifest themselves rather quickly, even if a person only recently moved into the neighborhood, because behaviors can potentially be altered very quickly in response to environmental conditions. However, we might not observe an immediate direct effect of neighborhood conditions on body mass because it may take prolonged exposure to a set of risky or protective environmental conditions before the 12 cumulative effect of changed behavioral patterns becomes large enough to alter a person’s body mass. Not surprisingly, most research on the built environment and health have focused on patterns of physical activity and travel behavior rather than direct links to obesity per se. In a recent overview of this literature, Frank and Engelke (2005) identify three key dimensions of the built environment that influence physical activity and travel behavior: proximity, connectivity, and design. Proximity refers to the availability of common destinations (e.g., places of work, stores, restaurants, and recreational facilities) within a short distance of one’s residence and is conceptually related to measures of population density, housing/building density, and the mixture of land uses (e.g., residential, commercial, and industrial). Areas with higher density tend to offer a greater number of destinations within close proximity to a person’s house, and areas where land uses are mixed tend to offer both a greater range and number of destinations. Connectivity refers to the availability of transportation linkages between destinations and can be studied by counting the number of road intersections or transit stations in a given place, by calculating the travel distance between locations on a street grid, or by tracking the time it takes to travel between locations via public transit systems. Planners are also interested in how design features of streets and buildings influence behavioral proclivities and travel patterns. For example, some urban planners (Frank and Engelke 2005; Rapoport 1987) argue that areas with greater detail in their built environments – including multiple small buildings, architectural details on building facades, and elements of the streetscape and sidewalk such as trees, mailboxes, benches, and lampposts – are more attractive places to walk, bike, and exercise. A parallel body of work has begun to examine differences across neighborhoods in what has been termed “the local food environment”, i.e., the availability and cost of healthy foods. 13 Studies have documented differences across neighborhoods in the number and types of food stores available (Moore and Diez Roux 2006; Morland, Wing, and Diez Roux 2002) as well as in the availability of healthy foods (Horowitz et al. 2004). Although there is compelling evidence that many features of the built environment differ across neighborhoods, empirical evidence linking these features to health behaviors and obesity remains scant. Our aim in this section is to assess some potential mechanisms through which the built environment may be linked to obesity. Because of the limited number of measures currently available in CCAHS, we focus exclusively on indicators of the proximity of common destinations that might be linked to either exercise or walking, including the presence of parks, playgrounds and other public open spaces, the proximity of stores where people can buy food (e.g., grocery stores or convenience stores), the pervasiveness of mixed commercial/residential land use and detached single-family homes in the neighborhood, and population density.13 We expect that people will exercise (and possibly walk) more in places with parks and playgrounds compared to neighborhoods where such facilities are absent. We also expect that areas with more mixed land use should encourage walking, because of the presence of more stores, and possibly exercise, because commercial land use could include private health clubs, and mixed land use might create safer environments for exercising outside because of the number of eyes on the street (Jacobs 1961). For similar reasons, we predict that neighborhoods with greater population density should encourage walking because more people are likely to be seen on the street and the places people can walk to will be more closely situated (these factors might also encourage exercise). To the contrary, we expect that people will walk less in neighborhoods with more 13 The measure of detached single-family home also taps into an element of urban design, namely the scale of buildings and spacing between them. 14 single-family detached homes because such housing tends to increase the distances between different uses and feature more disconnected street networks. We also expect that people who live in higher socioeconomic status neighborhoods will tend to be more active because affluent neighborhoods are likely to benefit residents in ways not captured by the measures we use here, such as offering greater access to private health clubs, more places to buy healthy foods (such as fresh produce), and more safe places to exercise.14 Affluent neighborhoods may also have design features that draw people outside to walk (such as architectural detail on buildings and more trees and decorative streetscapes). For different reasons, we also expect that living in a poor neighborhood may necessitate more walking, especially if people living in these areas are less like to own cars, rely on walking or public transportation for basic transport, and if such neighborhoods are located near transportation hubs.15 As a result, we expect that neighborhood socioeconomic status will not strongly predict walking because factors that encourage walking in affluent and poor neighborhoods will offset one another. The CCAHS data provide a unique opportunity to investigate the link between the built environment and obesity. In addition to the physical measurements of height, weight, and waist size and the survey-based measures of physical activity, the CCAHS also collected contextual data on Chicago neighborhoods from a number of sources. Of primary interest to us here is the so-called Systematic Social Observation (SSO), in which trained observers rated (by filling out a 14 We chose not to include measures of public safety (e.g., crime and disorder) in this analysis, even though they are likely to affect physical activity, because they suggest interventions that are different from those targeted at changing the built environment (although we also recognize that intervening to change the built environment can also be a strategy for reducing crime (e.g., Jeffery 1977)). 15 Unfortunately, the CCAHS does not measure of car-ownership, so this measure of neighborhood socioeconomic status is likely capturing unobserved heterogeneity among individuals in addition to unobserved contextual factors (Oakes 2004). 15 coding sheet) 1664 of the 1672 blocks in which potential survey respondents resided.16 The observer walked twice around the entire block – the first time walking along the “inside” block faces and the second along the “outside” block faces (a block face is one side of a street segment).17 For the purposes of this analysis, we created NC-level measures of the built environment from the SSO of five characteristics of the built environment by calculating the proportion of block faces within each NC that have mixed commercial and residential land use, detached single family houses, a store where food can be purchased (a grocery store/supermarket, green grocer/delicatessen, or convenience store), and a restaurant (either takeout or sit-down), and also constructing a measure of whether any block face within the NC has a park, playground, or other recreational facilities.18 In Table 5, we present results from linear multilevel random effects models of exercise, and walking.19 Here again, we stratify the analysis by gender. In addition to the variables from the SSO described above, we also considered the census-based measure of the proportion of adults with 16 or more years of education and population density (the ratio of the NC’s population to its area measured in squared kilometers). The models we present in Table 5 include only the contextual variables that were significantly related to the outcome in question in 16 To ensure that interviewers can make reliable and valid observations, we conducted a pilot study during Fall 2001 in which two raters made independent observations of the same block at the same time for 80 blocks, one in each of the focal NCs (Morenoff, Raudenbush, and House In preparation). In the second stage of the SSO, during Fall 2002 and Winter 2003, we collected observations from a single rater on almost every block in Chicago that contained a CCAHS survey sample address. 17 A total of 13,251 block faces were rated, and these were nested within 1,379 census block groups, 703 census tracts, and 343 NCs. 18 The SSO protocol asked raters whether any “recreational facilities, parks or playgrounds” were present on the block face. Private health clubs were not enumerated as recreational facilities, so to avoid conceptual confusion, we refer to this measure as parks and playgrounds, dropping the reference to recreational facilities. 19 We again recognize that linear models are problematic for these outcomes and are pursuing further analysis that treats our measures of exercise and walking categorically. 16 a least one of the gender groups.20 We include the contextual education measure in all of the models because even when it was not significant we wanted to control for any residual effects of neighborhood socioeconomic status. In the first set of models, we see that both women and men who live in neighborhoods with more highly educated people are likely to exercise more frequently. Having a park or playground in the NC is positively associated with exercise for women, but there is no significant association among men. It is plausible, although speculative, that women visit parks and playgrounds more frequently, especially if they have children, and thus would be more likely to use these facilities for exercising. The proportion of block faces in a neighborhood with mixed land use is positively associated with exercise for men but not for women, which is in the predicted direction, although the interaction between gender and mixed land use was unexpected and more difficult to interpret. The next set of models shows that neighborhood educational composition is not significantly related to walking for either men or women, as we predicted. Women tend to walk less when they live in neighborhoods with more detached single-family homes and they walk more frequently in areas where more block faces have food stores, but neither of these factors predict the frequency of walking among men. This finding is consistent with our expectation that proximity food stores would be a more important predictor of walking for women than for men. Men who live in areas with higher population density tend to walk more, but density is not a factor for women.21 20 We added each SSO variable separately to a model containing all of the individual-level covariates shown in Table 5 and the neighborhood-level measure of educational composition. [NOTE TO REVIEWERS: We intend to refine this analysis by possibly collapsing some of the SSO measures into an index of walkability. We also intend to explore issues of the geographic scale of these effects, but feel that this issue is beyond the scope of the current paper.] 21 Part of the reason that density is not significant among women may be that its effect is mediated by the prevalence of detached homes and proximity to food stores. When these variables are removed from the model, the point 17 These results provide suggestive, though far from conclusive evidence that features of the built environment relating to the proximity of common destinations, population density, and mixed land use influence physical activity, albeit in gender-specific ways. The gender interactions suggest that the proximity of common destinations is especially important for women, who are more likely to walk in neighborhoods that offer greater access to food stores and more likely to exercise in neighborhoods that offer greater access to parks and playgrounds. Moreover, women tend to walk less in neighborhoods where more of the block faces are dominated by detached single-family homes, which usually makes common destinations less proximate. This analysis also suggests that moving beyond census-based measures of neighborhood demographic and socioeconomic composition and incorporating measures of mutable features of the built environment that are more theoretically linked to health and health behaviors is a promising path for future research. Systematic social observation is a reliable method for collecting data on the built environment that could be fairly easily incorporated into the design of in-home surveys by asking interviewers to rate the block around each respondent’s house (Morenoff, Raudenbush, and House In preparation). It is especially cost effective when the sample design is clustered within relatively small geographic areas – especially when multiple respondents live on the same block – which is often true of not only local studies such as this one in Chicago, but many nationally representative studies as well. Another advantage of using systematic social observation is that it can support analysis of residential environments at multiple levels of geographic scale, including more micro-levels of the residential environment such as the block face, street, or block on which a person lives. Although an exploration of estimate for population density among women increases dramatically (with little change to the standard error) and becomes marginally significant. 18 geographic scale is beyond the scope of this chapter, it is important to note that some measures of the built environment are more theoretically relevant to health at a very small geographic scale and others more relevant at larger geographic scales. For example, whether a person lives on a street and block that is dominated by detached single-family homes may be more influential to determining how much that person walks than the neighborhood-based measure we use of the proportion of block faces in a neighborhood that have detached homes.22 One limitation of the CCAHS design for conducting systematic social observation of neighborhoods is that it is based on a sample of blocks from each NC (the blocks on which sample members resided) and thus does not provide complete geographic coverage of the built environment in any given neighborhood. However, there are secondary sources of data, which could be appended to survey data such as CCAHS, that provide complete geographic information on features of the built environment including the location of businesses (e.g., food stores, restaurants, liquor stores, bars, health clubs, health care facilities), parks and playgrounds, housing, land use patterns, and street networks. Such data could be used to compute accessibility indices that measure the number of recreational facilities, food stores, or restaurants within a given band of geographical distance around a person’s house. The analysis we have presented in this chapter is just an illustration of how collecting data on theoretically relevant features of the built environment can advance the current state of the literature on residential environments and health and also help pinpoint where to target future policy initiatives or interventions aimed at improving public health. Few studies have examined how community-targeted policies are associated with physical activity (Dannenberg et al. 2003) or obesity. Nonetheless, many scholars and public policy advocates call for land development 22 Block-, street-, and block-face-level measures from the CCAHS SSO were not yet available at the time we conducted this analysis. 19 practices that cultivate pedestrian-oriented communities to promote physical activity, such as locating goods and services within walking distance of residences (Satariano and McAuley 2003), and changing zoning laws and land use regulations that currently inhibit the development of mixed-use neighborhoods with compact development and well-connected streets (Frank and Engelke 2005; Schilling and Linton 2005). Several programs have been launched in the past few years with the potential to influence obesity through modification of the built environment (Brisbon et al. 2005; Orleans et al. 2003; Sallis et al. 2004).23 Analyzing the outcomes of these initiatives would certainly help inform future policy recommendations, as would observational studies that collect data on features of the built environment that such initiatives seek to alter and link them to health outcomes. The literature on the built environment and health is still in its early stages, as there have been relatively few attempts to date to design studies that collect data on both the health and behaviors of individuals and their residential context, without giving short-shrift to the latter by relying solely on census-based measures or individual-level perceptions of the neighborhood. Thus, strategies for improving the measurement of residential environments represent one promising development in recent research, but improving measurement does not address all of the shortcomings of observational studies. We now turn to the wider debate about what we can learn about the causal relationship between residential environments and health from observational studies. What Can We Learn about Neighborhood Effects on Health from Observational Studies? 23 The initiatives include the Active Living by Design program funded by the Robert Wood Johnson Foundation (Brisbon et al. 2005; Orleans et al. 2003; Sallis et al. 2004), the Design for Active Living project sponsored by the American Society of Landscape Architects, the Smart Growth Network stimulated by the US EPA and several nonprofit and government organizations, STEPS to a Healthier US funding project sponsored by the US Department of Health and Human Services (Brisbon et al. 2005), Active Community Environments Program (ACEs) sponsored by the CDC (Sallis et al. 2004). 20 As we mentioned at the outset, critics of observational studies on residential environments and health are skeptical about the designs and methods used to identify so-called “neighborhood effects” in this literature. Some critics (e.g., Oakes 2004) argue that observational studies of neighborhood effects are based on deeply flawed designs that render them of little value to social epidemiology or to policy makers interested in improving public health. Most of the criticisms that have been aimed at studies of neighborhood effects apply to all types of observational studies in which individuals are not randomly assigned to treatment groups, but the non-random nature of the residential sorting/selection process seems to provoke particular concern among critics of observational studies of neighborhood effects. In this section we underscore some of the limitations of observational studies but also highlight why we believe observational designs of neighborhood effects are still relevant to policy makers. With this aim in mind, we now turn to a deeper consideration of some of the criticisms of observational research on residential environments and health. A major identification problem in neighborhood effects research is that “social stratification confounds comparisons” (Oakes 2004: 1938). The issue here is that when researchers introduce background attributes of individuals into multilevel models in an effort to control for the selection of people into neighborhoods, they run the risk of explaining away all or most of the true contextual variance in health if these attributes also segregate people into distinctly different types of residential environments. We agree that residential segregation does confound attempts to study certain types of neighborhood effects because adequate comparisons do not exist. For example, if the only neighborhoods in a city that had toxic waste facilities were predominantly low-income neighborhoods, and if income groups were highly segregated from one another in the city, then it would be very difficult (or impossible) to separate the effect of 21 living near a toxic waste facility from the effect of having a low income on health. To estimate such a contextual effect, there would need to be a substantial number of neighborhoods both near and far from toxic waste facilities in which people with high and low incomes live side-by-side. Oakes argues that the search for such comparisons is futile, and even where relevant comparisons exist he is prone to dismiss them because “these people are exceptions to the rule and should not be given that same level of statistical credence as the majority” (2004: 1938). As a result, he concludes that analysts of observational data on neighborhoods and health find themselves in a Catch-22: If they do not control for key individual-level attributes that govern the selection process then their estimates are biased due to residual confounding; but when they do introduce such controls, they “must necessarily make adjustments until there is nothing for the neighborhood-level variables to explain” (2004: 1938). We differ with the assessment that social stratification inevitably confounds comparisons on two points. First, the argument that the between-neighborhood variance component (τ00 in our notation) approaches zero when individual-level covariates are added to multilevel models (to control for selection factors) relies too heavily on the neighborhood variance component as an indicator of neighborhood effects. Studies that lack the power to detect between-neighborhood variance may still have sufficient power to detect fixed effects of specific neighborhood attributes (Diez Roux 2004; Snijders and Bosker 1999), so that reducing the neighborhood variance component to near zero does not necessarily imply that the study cannot detect neighborhood effects. Also, in some cases, the neighborhood variance component can also be biased downward by the omission of confounding covariates from the model (Diez Roux 2004). The upshot is that the best way to test a hypothesis about a neighborhood effect is to test a 22 hypothesis about a specific neighborhood attribute, as guided by prior theory, rather than testing a more diffuse hypothesis about the between-neighborhood variance component. We also take issue with the conclusion that confounding due to stratification is a pervasive and universal problem in neighborhood effects studies. Instead, we adopt a more pragmatic view of the problem and contend that observational studies on neighborhood effects should not be rejected out of hand as lacking adequate comparisons but rather evaluated empirically to determine whether the data can support the type of counterfactual comparisons that are necessary for causal inference (e.g., Harding 2003). We expect that the tendency of stratification to confound comparisons will be most pervasive in studies that seek to make inferences about the effects of neighborhood socioeconomic or racial/ethnic composition, particularly in social contexts where people are highly segregated residentially along these dimensions. However, some of the more interesting research questions are about neighborhood “mechanisms” that are found in a variety of neighborhood structural settings, such as attributes of the built environment that may influence physical activity. In these cases, we are more optimistic about the prospects for observational studies to find individuals with similar background attributes who live in neighborhoods with and without the relevant treatment condition. As an illustration of the types of comparisons we have in mind, we use our data to examine the distribution of individual-level education and race/ethnicity within categories of neighborhood environments as defined by educational composition and racial ethnic composition in Table 6, and by characteristics of the built environment in Table 7. In Table 6 we divide neighborhoods into four categories of educational composition, based on the 2000 census measure of the percentage of adults who have 16 or more years of education (0-4.9 percent, 5- 23 14.9 percent, 15-34.9 percent, and 35 percent or more), and four categories of racial/ethnic composition (75 percent or more white, 75 percent or more black, 75 percent or more Hispanic, and a residual category that we call “mixed”). The table shows the distribution of individuallevel education and race/ethnicity within each category of the neighborhood educational composition and racial/ethnic composition typologies. The presence of some cells with small sample sizes (e.g., below 50 people) indicates how difficult it is to disentangle the contextual and individual-level effects of either race or education in a highly segregated city such as Chicago, although the problem is more severe in with race than it is with education.24 For example, in making comparisons across categories of neighborhood education, we run into sparse data when we attempt to compare people with high and low education living in neighborhoods at either extreme of the distribution of educational composition. However, over half of the sample members who have 16 or more years of education live in neighborhoods in which less than 35 percent of the people have 16 or more years of education, which is where the majority of people from the lower education groups live. There are many neighborhoods in Chicago where people of high and low education groups live side-by-side, but by definition this is not the case at the extremes of the neighborhood education distribution. We also find adequate representations of whites, blacks, and Hispanics across the distribution of neighborhood educational composition, with one exception: there are very few whites residing in neighborhoods at the low end of the educational distribution. Stratification limits comparisons more in the case of neighborhood racial/ethnic composition, as 79 percent of the blacks in our sample live in neighborhoods where at least 75 percent of the people are black, and very small numbers of whites and Hispanics live in these areas. Although a substantial number of whites and Hispanics live in so-called “mixed” 24 The sample does not contain adequate representations of people in the non-Hispanic other race categories (mainly of Asian and Native American descent) to support sub-group analyses on this group. 24 neighborhoods, and a non-trivial share of blacks live there as well, there is generally less support in the data for making comparisons among people of different races who share the same neighborhoods than there is in the case of education. In Table 7, we examine the distribution of education and race/ethnicity across neighborhoods with different distributions of built environment “treatment” conditions. Although people with 16 or more years of education and whites are disproportionately represented in neighborhoods with parks or playgrounds, this does not limit our ability to compare people across educational and racial/ethnic groups who live in neighborhoods with and without parks and playgrounds. We also divide neighborhoods into tertiles based on the proportion of block faces that have a grocery or convenience store and the proportion of block faces that have mixed land use (commercial and residential). We find that there are ample representations of people with different levels of education and of different racial/ethnic backgrounds living in neighborhoods that span the spectrum of built environmental conditions. In fact, we find an overrepresentation of Hispanics and people with less than 12 years of education – both of whom are at significantly higher risk of obesity, according to Table 4 – living in neighborhoods with the highest prevalence of food stores and the most block faces with mixed use. Thus, there is much more support in the data for making comparisons that can disentangle individual-level education and race/ethnicity from neighborhood treatment conditions when those treatment conditions are based on characteristics of the built environment rather than the socioeconomic or racial/ethnic composition of the neighborhood. In summary, we agree with the notion that unaccounted for differences in the characteristics of residents (the selection problem, or “social stratification confounds comparisons”) are a major challenge in estimating neighborhood effects from observational data. 25 But we also believe that the magnitude of this problem can be examined empirically, for example by investigating the relationships between individual-level and neighborhood-level variables similarly to what we do in Tables 6 and 7. This approach will allow researchers to judge the confidence they have in the effect measures reported from a particular dataset. Other approaches such as propensity scores, which can be used to explicitly match on selection-related factors and therefore ensure comparability on individual-level predictors, are also being increasingly used in studies of neighborhood effects in order to examine the robustness of measures estimated using traditional regression approaches. Yet another approach is to actively search out exceptions to the rule, i.e. “naturally” occurring situations where individual and neighborhood-level factors are uncoupled or where there is variability in neighborhood-level attributes of interest even within neighborhoods of similar socioeconomic composition. Not all low-income neighborhoods are the same, and within large cities for example, it is often possible to find similarly low income neighborhoods that differ in features of the built environment. This uncoupling of neighborhood socioeconomic composition and other neighborhood attributes has become more apparent as observational research on neighborhood effects has moved from initial investigation of the “contextual” effect of neighborhood SES to much more specific investigations of the neighborhood attributes that are hypothesized to be related to health outcomes. Another promising avenue in observational studies of neighborhood effects is the evaluation of “natural” experiments, i.e. the investigations of changes in health outcomes that occur as a result of a specific change in neighborhood environments, such as the opening of a new public space, the implementation of a new transportation policy, or the opening of a farmer’s market. Neighborhoods change constantly, but these changes are rarely evaluated for their health effects. The documentation of the health 26 impact of these changes would strengthen causal inferences regarding neighborhood effects and would also provide evidence of high relevance to policy makers. These types of evaluations will require the partnering of researchers with community groups, public agencies, and policy makers. Skeptics may argue that however hard we try observational studies will never yield firm evidence of neighborhood effects or evidence that is directly relevant for policy. It has been argued, for example, that only randomized trials will solve the conundrum of neighborhood effects research and will provide evidence that is relevant in the “real world”. But rather than calling generically for more experimental studies, it is important to critically consider whether randomized trials can be implemented in this area, what these trials would look like, and what we could actually learn from them. A first key requirement of experimental studies is that we have a clear specification of the intervention or treatment that we want to test. There is considerable debate on this topic among neighborhood health effects researchers. Would the ideal experiment be moving poor people to non-poor neighborhoods (as the MTO study has done) or intervening on neighborhoods themselves? And if the latter, what specific interventions should be considered as highest priorities? Unfortunately, prior research on neighborhood effects cannot offer much guidance on this question, for it has not considered treatment conditions that are amenable to policy interventions. It is precisely in this context, when there is still considerable uncertainty on the treatment to be tested, that observational studies conducted in varying places and populations and with varying designs and measures are likely to be most useful, while controlled interventions offer less value relative to cost (Phillips 2001). We argue that this is exactly the current stage of research on neighborhood health effects.25 Although much work to date has 25 For methods to identify the onset of the next stage of research, at which most observational studies diminish in value relative to randomized or natural experiments, see (Greenland 2005). 27 established an association between neighborhood disadvantage/affluence and health, more observational studies are needed on mutable features of residential environments that could in the future be subjected for randomized community trials. A second key consideration pertains to the logistics of implementing such a trial, once the treatment has been defined. We believe the most appropriate design for assessing neighborhood characteristics’ effects on health is by randomly assigning neighborhoods to interventions, not the random assignment of individuals to neighborhoods. When individuals are randomly assigned to neighborhoods (or allowed to select their own neighborhood within certain parameters, as was the case in MTO), researchers have no means of determining which features of the new neighborhoods brought about a confirmed effect. But even assuming randomization of neighborhoods is feasible, there are major challenges in terms of the power of these designs. Studies that randomize changes to neighborhoods will typically have very little power, unless many neighborhoods are randomized or additional assumptions are brought to bear. Suppose a trial with intervention assigned to 10 of 20 communities, each containing 1,000 study participants. In classical randomization inference, the n of such a study is 20, not 20,000 (Gail et al. 1996). Other modes of inference increase the effective sample size, but at the expense of adding assumptions about functional forms or even assumptions that key variables have not been omitted (Gail, Wieand, and Piantadosi 1984) – the same type of assumptions for which observational studies are often faulted. The analyst of randomized community trial thus confronts a dilemma between maintaining the purity of the approach, at the cost of distinguishing only the strongest of effects from statistical noise, and seeking more power through the use of the same fallible, if potentially quite credible, assumptions that are characteristic of observational studies. 28 Another instance in which assumptions become necessary in experimental designs is when the treatment condition being randomized co-varies with other potential influences on health because of chance associations between the treatment condition and other features of neighborhoods. As an example, imagine an intervention that sought to create “active living” communities (http://www.activeliving.org/), perhaps by building mixed-use development projects in selected neighborhoods that were matched on socioeconomic status. Setting aside the practical concerns about how feasible such an undertaking would be and how many communities could be selected for the treatment group, we want to point out the potential, even in a randomized design, for the realized assignment of treatments to co-vary with other neighborhood factors that could affect health. For example, even though the treatment and control groups in this hypothetical experiment were matched on socioeconomic status, the treatment neighborhoods could, by chance, be located next to areas with higher socioeconomic status than those surrounding control group neighborhoods, and this could encourage more walking in treatment areas because there are more stores and other common destinations within walking distance. Conversely, if the areas surrounding a treatment neighborhood were high in crime, this might discourage walking in the active living community. The point is that even in a randomized community design where treatment and control communities were matched on socioeconomic status, there is still the potential that confounding influences could creep into the design – especially when the number of cases being randomized is small, as it often is in community trials – and the need to control such influences leads back to models that introduce parametric assumptions.26 26 Of course, there is the possibility of a nonparametric analysis that doesn't depend on a model for Y given X, but these will often have less power. 29 Our intention in raising these concerns is not to squelch interest in conducting randomized community trials. To the contrary, we feel that they hold great potential for advancing the literature on neighborhood effects and for improving public health. Rather, our aim is to cast doubt on the idea that randomized community trials are a panacea for all the problems that currently plague the literature on neighborhood effects. Observational studies can help identify effective treatment conditions that could later become the subject of randomized trials, especially if more attempts are made to measure mutable features of neighborhood environments that are conceptually most relevant to health. But in the more immediate term, we believe many of the criticisms leveled at multilevel observational studies could be addressed through more careful designs and methodological approaches, which would include paying more explicit attention to the ability of the data to support the types of comparisons necessary for causal inference (the notion of common support), designing more and better measures of the factors that sort people into residential environments, and focusing on treatment conditions that are not inherently endogenous (i.e., they are not brought about by the same factors that influence health), such as those associated with the built environment. 30 References Bingenheimer, Jeffrey B. and Stephen W. Raudenbush. 2004. "Statistical and Substantive Inferences in Public Health: Issues in the Application of Multilevel Models." Annual Review of Public Health 25. Brisbon, Nancy, James Plumb, Rickie Brawer, and Dalton Paxman. 2005. "The asthma and obesity epidemics: The role played by the built environment--a public health perspective." Journal of Allergy and Clinical Immunology 115:1024-1028. Dannenberg, Andrew L, Richard J Jackson, Howard Frumkin, Richard A Schieber, Michael Pratt, Chris Kochtitzky, and Hugh H. Tilson. 2003. "The impact of community design and land-use choices on public health: A scientific research agenda." American Journal of Public Health 93:1500-8. Diez Roux, Ana. 2004. "Commentary: Estimating Neighborhood Health Effects: The Challenges of Causal Inference in a Complex World." Social Science & Medicine 58:1953-1960. Diez Roux, Ana V. 2001. "Investigating neighborhood and area effects on health." American Journal of Public Health 91:1783-9. Duncan, G. J. and S. W. Raudenbush. 1999. "Assessing the Effects of Context in Studies of Child and Youth Development." Educational Psychologist 34:29-41. Frank, Lawrence D. and Peter Engelke. 2005. "Multiple Impacts of the Built Environment on Public Health: Walkable Places and the Exposure to Air Pollution." International Regional Science Review 28:193-216. French, Simone A., Mary Story, and Robert W. Jeffery. 2001. "ENVIRONMENTAL INFLUENCES ON EATING AND PHYSICAL ACTIVITY." Annual Review of Public Health 22:309-335. Gail, Mitchell H., Steven D. Mark, Raymond J. Carroll, Sylvan B. Green, and David Pee. 1996. "On Design Considerations and Randomization-based Inference for Community Intervention Trials." Statistics in Medicine 15:1069--1092. Gail, Mitchell H., S. Wieand, and S. Piantadosi. 1984. "Biased estimates of treatment effect in randomized experiments with nonlinear regressions and omitted covariates." Biometrika 71:431--44. Greenland, Sander. 2005. "Multiple-bias Modelling for Analysis of Observational Data." Journal of the Royal Statistical Society, Series A: Statistics in Society 168:267-306. Harding, David J. 2003. "Counterfactual Models of Neighborhood Effects: The Effect of Neighborhood Poverty on Dropping Out and Teenage Pregnancy." The American journal of sociology 109:676 (44 pages). Hill, James O. and John C. Peters. 1998. "Environmental Contributions to the Obesity Epidemic." Science 280:1371-1374. Horowitz, CR, KA Colson, PL Hebert, and K. Lancaster. 2004. "Barriers for buying healthy foods for people with diabetes: evidence of environmnetal disparities." American Journal of Public Health 94:1549-1554. House, James S. 2002. "Understanding Social Factors and Inequalities in Health: 20th Century Progress and 21st Century Prospects." Journal of Health and Social Behavior 43:125143. Hox, Joop. 2002. Multilevel Analysis: Techniques and Applications. Mahwah, NJ: Lawrence Erlbaum Associates. Jacobs, Jane. 1961. The Death and Life of Great American Cities. New York: Random House. 31 Jeffery, C. Ray. 1977. Crime Prevention Through Environmental Design. Beverly Hills, CA: Sage Publications. Kling, Jeffrey, Jeffrey Liebman, and Lawrence Katz. 2005. "Experimental Analysis of Neighborhood Effects." Washington, DC: NBER Working Paper. Mokdad, A.H., M.K. Serdula, W.H. Dietz, B.A. Bowman, J.S. Marks, and J.P. Koplan. 1998. "The Spread of the Obesity Epidemic in the United States." Journal of the American Medical Association 282:1519-1522. Moore, Latetia V. and Ana V. Diez Roux. 2006. "Associations of Neighborhood Characteristics With the Location and Type of Food Stores." Am J Public Health 96:325-331. Morenoff, Jeffrey D. 2003. "Neighborhood Mechanisms and the Spatial Dynamics of Birthweight." American Journal of Sociology 108:976-1017. Morenoff, Jeffrey D. and John Lynch. 2004. "What Makes a Place Healthy? Neighborhood Influences on Racial/Ethnic Disparities in Health Over the Life Course." Pp. 406-449 in Critical Perspectives on Racial and Ethnic Differences in Health in Late Life, edited by N. B. Anderson, R. A. Bulatao, and B. Cohen. Washington, D.C.: National Academies Press. Morenoff, Jeffrey M., Stephen Raudenbush, and James S. House. In preparation. "Systematic Social Observation by Survey Interviewers: A Methodological Evaluation." Morland, K, S Wing, and Ana Diez Roux. 2002. "Neighborhood characteristics associated with the location of food stores and food service places." American Journal of Preventative Medicine 22:23-29. Oakes, J. Michael. 2004. "The (Mis)estimation of Neighborhoods Effects: Causal Inference for a Practicable Social Epidemiology." Social Science & Medicine 58:1929-1952. Orleans, C. T., M. K. Kraft, J. F. Marx, and J. M. McGinnis. 2003. "Why are some neighborhoods active and others not? Charting a new course for research on the policy and environmental determinants of physical activity." Ann Behav Med 25:77-9. Phillips, Carl V. 2001. "The economics of `more research is needed'." International Journal of Epidemiology 30:771-776. Rapoport, Amos. 1987. "Pedestrian Street Use: Culture and Perception." in Public Streets for Public Use, edited by A. Moudon. New York: Columbia University Press. Raudenbush, Stephen W. and Anthony S. Bryk. 2002. Hierarchical linear models: applications and data analysis methods. Thousand Oaks: Sage Publications. Robert, S. A. 1998. "Community-level socioeconomic status effects on adult health." Journal of Health and Social Behavior 39:18-37. Sallis, James F., Lawrence D. Frank, Brian E. Saelens, and M. Katherine Kraft. 2004. "Active transportation and physical activity: opportunities for collaboration on transportation and public health research." Transportation Research Part A: Policy and Practice 38:249268. Sampson, Robert J., Stephen W. Raudenbush, and Felton Earls. 1997. "Neighborhoods and violent crime: a multilevel study of collective efficacy." Science 277:918-924. Satariano, William A. and Edward McAuley. 2003. "Promoting Physical Activity Among Older Adults: From Ecology to the Individual." American Journal of Preventive Medicine 25:184-192. Schilling, J. and L. S. Linton. 2005. "The public health roots of zoning: in search of active living's legal genealogy." Am J Prev Med 28:96-104. 32 Snijders, Tom A. B. and Roel Bosker. 1999. Multilevel Analysis: An Introduction to Basic and Advanced Multilevel Modeling. Thousand Oaks, CA: Sage Publications. Sobal, Jeffery and Albert J. Stunkard. 1989. "Socioeconomic Status and Obesity: A Review of the Literature." Psychological Bulletin 105:260-275. U.S. Department of Housing & Urban Development, Office of Policy Development and Research, L. Orr, J.D. Feins, R. Jacob, E. Beecroft, L. Sanbonmatsu, L.F. Katz, J.B. Liebman, and J.R. Kling. 2003. "Moving to Opportunity for Fair Housing Demonstration Program Interim Impacts Evaluation." Department of Housing and Urban Development, Washington, DC. Williams, David R. and Chiquita Collins 1995. "US Socioeconomic and Racial Differences in Health: Patterns and Explanations." Annual review of sociology, pp. 349 (38 pages). 33 Table 1. Definitions of Obesity-Related Outcome Variables Variable Definition Body Mass Index (BMI) BMI = (body weight in kilograms)/(height in meters)2 - Height was measured using a tape measure while respondents stood in a doorway in stocking feet. - Weight was measured in stocking feet and street clothes (less sweaters or other heavy over-garments) using digital scales. Waist Size Waist size was measured with a tape measure in inches. Exercise This scale was constructed from survey questions that asked respondents (1) whether they are currently confined to a bed or chair for most or all of the day because of their health, (2) how many days a week do they do light or moderate leisure activities other than walking or working around the house for at least 10 minutes that cause only light sweating or a slight to moderate increase in breathing or heart rate, (3) when they do light/moderate leisure activities, do they generally do them for 20 minutes or more, (4) how many days a week do they do vigorous activities for at least 10 minutes that cause heavy sweating or large increases in breathing or heart rate, and (5) each time they do vigorous activities, do they generally do them for 20 minutes or more? The scale was coded as follows: - 0 = Never: Individuals who said they (a) never engage in light-moderate leisure activities, (b) never engage in vigorous activity, or (b) were confined to a bed or chair. - 1 = Light: Individuals who engage in (a) light-moderate physical activity once a week or less regardless of duration, (b) light-moderate physical activity 2-3 times per week for less than 20 minutes, or (c) vigorous activity once per week or less for less than 20 minutes. - 2 = Light-moderate: Individuals who engage in (a) light-moderate activity 2-3 times per week for more than 20 minutes, (b) light-moderate activity 4 or more times per week for less than 20 minutes, or (c) vigorous activity once per week or less for more than 20 minutes. - 3 = Moderate-heavy: Individuals who engage in (a) light-moderate activity 4 or more times per week for more than 20 minutes, or (b) vigorous activity 2-3 times per week regardless of duration. - 4 = Heavy: Individuals who engage in vigorous activity 4 or more times per week regardless of duration Walking This survey-based measure is based on the following question: “On the average over the past year, how many days a week do you walk continuously for 20 minutes or more, either to get somewhere or just for exercise or pleasure?” (1) Never, (2) Less than once a week, (3) Once a week, (4) 2-3 times a week, (5) 4-5 times a week, (6) Almost every day. Fruit/Vegetable Intake This survey-based measure is based on an open-ended response to the following question: “How many servings of fruit or vegetables do you usually eat in a day? (A serving is a cup of fruit or vegetable juice or a half cup of raw or cooked vegetables or fruits. Include juices and all types of raw or cooked fruits and vegetables.)” Table 2. Intra-cluster Correlation (Percentage of Variance Between Neighborhoods) for Obesity-Related Outcomes Before and After Adjusting for Individual-Level Covariates: CCAHS Unadjusted ICC Adjusted ICC Full Full Outcome Sample Females Males Sample Females Males BMI 10.06 17.78 7.42 6.32 8.98 5.78 Waist Size 11.33 19.54 4.96 5.82 9.56 3.89 Exercise 10.68 15.86 10.94 9.16 15.32 9.01 Walking 9.02 10.44 8.79 9.73 10.61 10.87 Fruit/Vegetable Intake 6.69 8.47 7.38 4.92 5.31 7.91 Table 3. Decomposing Racial/Ethnic and Socioeconomic BMI Disparities: CCAHS OLS Random Intercept "Contextual" BetweenWithinTotal Disparity Neighborhood Neighborhood Disparity (β t) Disparity (β b) Disparity (β w) (β c = β b - β w) Group Comparison in BMI Coef (SE) Coef (SE) Coef (SE) Coef (SE) Females Non-Hisp black vs. Non-Hisp white 4.66 (0.48) ** 4.63 (0.55) ** 4.10 (1.12) ** 0.53 (1.24) Hispanic vs. non-Hisp white 3.58 (0.47) ** 4.77 (0.66) ** 2.30 (0.70) ** 2.47 (1.00) * <12 yrs educ vs. 16+ yrs educ 4.14 (0.52) ** 5.05 (0.71) ** 2.76 (0.77) ** 2.29 (1.04) * Males Non-Hisp black vs. Non-Hisp white Hispanic vs. non-Hisp white <12 yrs educ vs. 16+ yrs educ **p<.01; *p<.05 1.25 (0.47) ** 2.08 (0.43) ** 0.82 (0.48) 1.04 (0.52) * 1.61 (0.63) * 0.81 (0.61) 2.33 (1.17) * 2.51 (0.67) ** 0.76 (0.78) -1.29 (1.28) -0.90 (0.94) 0.05 (0.97) Covariates Person-Level Age Table 4. Individual- and Neighborhood-Level Predictors of BMI for Females Females Males OLS Random Effects OLS Random Effects Coef (SE) Coef (SE) Coef (SE) Coef (SE) a Age squared Non-Hispanic black Hispanic Other race 1st generation immigrant 2nd generation immigrant <12 years of education 12-15 years of education Income < $5,000 Income $5000-$9,900 Income $10,000-$29,999 Income $30,000-$49,999 Missing data on income 0.09 (0.01) ** 0.08 (0.01) ** 0.07 (0.01) ** 0.07 (0.01) ** -0.04 3.67 3.19 -1.13 -0.96 0.05 2.35 1.58 -0.91 0.77 0.74 -0.23 -0.76 -0.03 3.07 2.52 -1.02 -1.07 -0.14 1.82 1.15 -1.05 0.71 0.58 -0.27 -1.13 -0.03 1.45 2.74 -0.24 -0.46 1.18 -0.09 0.22 -1.31 -0.77 -0.61 -1.08 -0.02 -0.02 2.09 2.99 0.32 -0.62 0.92 -0.51 -0.25 -1.30 -0.66 -0.47 -1.03 -0.10 (0.00) (0.55) (0.57) (0.85) (0.51) (0.67) (0.62) (0.47) (0.98) (0.88) (0.63) (0.58) (0.59) ** ** ** ** ** Neighborhood-Level (from Census) Proportion White Proportion Hispanic Proportion 16+ Years of Educ Intercept n ** ** ** * ** * * (0.01) (0.49) (0.58) (0.62) (0.61) (0.69) (0.55) (0.46) (0.80) (0.90) (0.52) (0.52) (0.57) ** ** ** * -0.12 (0.98) 0.98 (0.93) -3.41 (1.43) * 26.33 (0.51) ** 27.33 (0.60) ** 1,870 **p <.01; *p <.05 a (0.01) (0.70) (0.62) (0.80) (0.52) (0.66) (0.65) (0.50) (0.98) (0.88) (0.61) (0.56) (0.57) Coefficients and standard errors have been multiplied by 10. (0.01) ** (0.76) ** (0.64) ** (0.64) (0.63) (0.72) (0.58) (0.47) (0.85) (0.91) (0.53) (0.53) (0.57) 2.64 (1.11) * -0.03 (0.99) -3.74 (1.25) ** 27.79 (0.46) ** 27.79 (0.59) ** 1,235 Table 5. Individual and Neighborhood-Level Predictors of Physical Activity by Gender Exercise Walking Females Males Females Males Covariates Coef (SE) Coef (SE) Coef (SE) Coef (SE) Neighborhood-Level Vars from Census Proportion 16+ Years of Educ 1.08 (0.26) ** 0.94 (0.26) ** 0.57 (0.31) -0.22 (0.40) Population Density 0.00 (0.13) 0.40 (0.15) ** Neighb-Level Vars from SSO Prop BFs with Mixed Comm/Resid Use -0.19 (0.31) Presence of Recreational Facilities 0.26 (0.09) ** Prop BFs with Detached Homes Prop BFs with Grocery/Conv Store Person-Level Age 0.78 (0.27) ** 0.06 (0.09) -0.41 (0.19) * 1.48 (0.55) ** -0.25 (0.34) 0.59 (0.79) 0.02 (0.05) -0.13 (0.02) ** -0.28 (0.03) ** -0.03 (0.03) Age squared Non-Hispanic black Hispanic Other race 1st generation immigrant 2nd generation immigrant <12 years of education 12-15 years of education Income < $5,000 Income $5000-$9,900 Income $10,000-$29,999 Income $30,000-$49,999 Missing data on income 0.00 0.03 0.25 0.19 -0.31 -0.14 -0.22 -0.04 0.01 -0.16 -0.18 0.09 -0.24 0.00 0.37 0.27 0.16 -0.25 -0.17 -0.13 -0.05 -0.39 -0.95 -0.28 -0.21 -0.37 -0.01 -0.18 -0.12 -0.49 -0.15 0.04 0.24 0.20 0.32 0.17 0.35 0.36 0.39 Intercept 2.09 (0.11) ** a (0.00) (0.11) (0.13) (0.32) (0.12) * (0.12) (0.12) (0.09) (0.18) (0.17) (0.12) (0.11) (0.12) * ** ** * * 2.31 (0.13) ** **p <.01; *p <.05 a (0.00) (0.12) (0.14) (0.21) (0.16) (0.16) (0.16) (0.11) (0.23) (0.20) (0.13) (0.13) (0.15) Coefficients and standard errors have been multiplied by 10. (0.00) (0.13) (0.18) (0.32) (0.18) (0.17) (0.16) (0.11) (0.22) (0.20) (0.15) (0.15) (0.16) ** * * * 4.24 (0.14) ** 0.00 -0.06 -0.07 -0.09 -0.08 0.28 0.30 0.21 0.29 -0.28 0.18 0.08 -0.07 (0.00) (0.17) (0.19) (0.33) (0.21) (0.23) (0.21) (0.16) (0.28) (0.34) (0.19) (0.19) (0.19) 4.29 (0.18) ** Table 6. Unweighted Frequencies of Individuals within Categories of Neighborhood-Level Educational and Racial/Ethnic Composition by Person-Level Education and Race/Ethnicity %Adults in NC with 16+ Years of Educ NC Racial/Ethnic Comp Typology 75%+ 75%+ NonNonIndividual-Level Hisp Hisp 75%+ Row Categories Hisp Total 0-4.9% 5-14.9% 15-34.9% 35%+ White Black Mixed Education Less than 12 years 150 425 179 38 32 273 169 318 792 12-15 years 177 757 469 173 156 635 145 640 1,576 16 or more years 26 159 230 322 152 151 36 398 737 Race/Ethnicity Non-Hispanic white Non-Hispanic black Hispanic Non-Hispanic other 12 219 120 2 232 663 428 18 384 277 193 24 355 81 61 36 290 4 33 13 30 981 41 7 52 18 275 5 611 237 453 55 983 1,240 802 80 Column Total 353 1,341 878 533 340 1,059 350 1356 3,105 Table 7. Unweighted Frequencies of Individuals within Categories of Neighborhood-Level Built Environment Measures by Person-Level Education and Race/Ethnicity Park or Tertiles of % Block Faces in Tertiles of % Block Faces in Playground in NC with Grocery or NC with Mixed NC? Convenience Store (Comm/Resid) Use Individual-Level Row Categories No Yes 0-4% 4-12% 12%+ 0-13% 13-30% 30%+ Total Education Typology Less than 12 years 240 552 240 293 259 243 249 300 792 12-15 years 549 1027 607 600 369 599 524 453 1576 16 or more years 182 555 206 338 193 214 252 271 737 Race/Ethnicity Non-Hispanic white Non-Hispanic black Hispanic Non-Hispanic other 273 402 272 24 710 838 530 56 329 510 195 19 399 500 293 39 255 230 314 22 243 636 159 18 378 388 223 36 362 216 420 26 983 1240 802 80 Column Total 971 2,134 1,053 1,231 821 1,056 1,025 1,024 3105