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