Neighbourhood food, physical activity, and educational environments and black/white

disparities in obesity: a complex systems simulation analysis

Author(s): Mark G Orr, George A Kaplan and Sandro Galea
Source: Journal of Epidemiology and Community Health (1979-) , September 2016, Vol.
70, No. 9 (September 2016), pp. 862-867
Published by: BMJ

Stable URL: https://www.jstor.org/stable/44018496

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

BMJ is collaborating with JSTOR to digitize, preserve and extend access to Journal of
Epidemiology and Community Health (1979-)

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms
 Neighbourhood food, physical activity, and
educational environments and black/white
disparities in obesity: a complex systems simulation
analysis
Mark G Orr,1 George A Kaplan,2 Sandro Galea3

► Additional material is ABSTRACT informed by the search for independent causes may
published online only. To view mislead by oversimplification. For example, over
Background Multiple approaches that can contribute
please visit the journal online
the life proposed.
to reducing obesity have been course, people move between neighbour-
These policies
(http://dx.d0i.0rg/l 0. 1 1 36/jech-
2015-205621). hoods and change
may share overlapping pathways, and behaviours.
may have Many of these influ-
encecreating
unanticipated consequences, determinants of BMI and they feedback on
considerable
1 Biocomplexity Institute of
Virginia Tech, Virginia
each other, creating
complexity. Aiming to illuminate the ause
dynamic
of pattern that further
agent-based
Polytechnic Institute and State changes with residential
models to explore the consequences of key mobility. Residential
policies, this
University, Arlington, Virginia, paper simulates the effectsmobility changes neighbourhood
of increasing composition, and
neighbourhood
USA
these changes
availability of good food stores, in composition,
physical activity in turn, change
department of Epidemiology,
access to neighbourhood
infrastructure and higher school quality on resources
the with resultant
reduction
University of Michigan, Ann
Arbor, Michigan, USA changes
of black/white disparities in bodyin individual
mass behaviours,
index thereby
(BMI)affecting
in
3Dean of Boston University the USA. inflows and outflows from the neighbourhood, and
School of Public Health, changing social influences and norms that may
Methods We used an agent-based model, with
Boston, Massachusetts, USA
parameters derived from the empirical literature, which affect diet and physical activity. In short, this is a
Correspondence to included individual and neighbourhood characteristics dynamic system with many characteristics influen-
Mark Orr, Virginia
over the life course as determinants of behaviours cing each other. It is a multilevel system with cross-
Bioinformatics Institute, thought to impact BMI. We systematically varied the level interdependence and feedback, people with
Virginia Polytechnic Institute heterogeneous characteristics, and interdependence
strength of the 3 policy interventions, examining the
and State University, 900
impact of 125 different policy scenarios on black/white between individuals, with change over time. It is a
N. Glebe Rd., Arlington, VA
22203, USA; [email protected]. BMI disparities. system potentially characterised by many positive
edu Results In the absence of any of these policies, black/ feedback loops, allowing considerable dynamic,
white BMI disparities generally increased over time. non-linear behaviour. All these are characteristics
Received 3 March 2015
However, we found that some combinations of these that suggest the use of an agent-based model.
Revised 3 February 2016
Accepted 16 March 2016 policies resulted in reductions in BMI, yielding decreases Embracing the dynamic, complex, multifactor
Published Online First in the black/white BMI disparity as large as a 90%. and multilevel nature of a system holds promise
15 April 2016 Conclusions Within the structure of relationships from both an analytical and conceptual perspective.
captured in this simulation model, there is support for Computer-based simulation models make it pos-
the further use of agent-based simulation models to sible to begin to model the complexities inherent in
explore upstream policies as plausible candidates for the the obesity system. For example, agent-based mod-
reduction of black/white disparities in BMI. These results elling techniques can be designed to capture the
highlight the potential insights into important public key characteristics of the obesity system: multiple
health problems, such as obesity, that can come from levels of analysis, interdependence between levels,
uniting the systems science approach with policy heterogeneity in agents/actors, and interdependence
analysis. within levels (eg, between individual people in a
social network). In short, the agent-based model-
ling approach captures both the microlevel detail
BACKGROUND (similar to microsimulation models), and the inter-
Obesity is a serious and growing problem for the individual and social network processes in a
public's health,1 with considerable impact on popu- dynamic framework. Further, in terms of policy
lation health. The Foresight group's2 systems map and intervention, agent-based models can test
of the drivers of obesity illustrates the many path-what-if scenarios in a way that is transparently
ways over different and interactive levels, and withrelated to key policy levers.
multiple feedback loops, that are implicated in the In this paper, we use an agent-based model to
provide a proof-of-concept of the use of such a
obesity system. It is increasingly recognised that to
model to explore what might happen to body mass
reduce obesity, researchers, practitioners and policy
makers need to confront a complex dynamicindex (BMI) disparities between blacks and whites
<8> CrossMark system.3 in the USA under conditions where multiple pol-
Most of the extant literature focuses on inde- icies are implemented in parallel. It is important to
To cite: Orr MG, pendent determinants of obesity. The search forestablish the relative effects of various policies on
Kaplan GA, Galea S. J independent causes is widely pursued because it BMI. However, and perhaps even more import-
Epidemiol Community Health has the potential to suggest targeted and focused antly, many of the multiple overlapping and inter-
2016;70:862-867. interventions. However, research and interventions acting pathways that potentially mediate the effects

862 Orr MG, et al. J Epidemiol Community Health 2016;70:862-867. doi:10.1 136/jech-201 5-205621 BMJ

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms
of these policies on BMI and disparities in BMI are affected by Policy variables and their pathways
these policies, making it difficult to understand the effects of The key policy variables, good food stores, physical activity
one if not considering the other. In addition, these separate and infrastructure and school quality were conceptualised as
overlapping pathways are part of a complex system in which it neighbourhood-level variables. Figure 1 illustrates the chain
is possible that certain combinations of policies may emerge as from the policy variables to BMI in our model. We describe the
particularly efficacious in reducing BMI black/white disparities. logic behind our model below, with details in the online supple-
Therefore, we used a complex systems approach that captured mentary material.
some of the inherent complexity of the obesity system and Good food stores reflected the number of supermarkets per
allowed us to conduct simulations of what-if policy scenarios 10 000 people in a neighbourhood. The link between neigh-
specific to the implementation of multiple policies. bourhood income level and good food stores was derived from
a study that investigated neighbourhood characteristics and
access to different types of food stores.5 We note that a recent
METHODS systematic review indicates some uncertainty as to the impact of
Overview of the model and simulations neighbourhood food environment on obesity.6
The population of agents in our model represented the eco- School quality was designed to reflect neighbourhood-level
nomic and racial distributions of black and non-Hispanic whites student-teacher ratios. Although not the only determinants of
in the 100 largest metropolitan statistical areas in the USA. For school quality, student-teacher ratio is associated with the eco-
the purposes of the simulation, there were 64 neighbourhoods, nomic returns from increasing education, and some evidence
each with 25 households, with the racial/ethnic and economic suggests other benefits as well.7-9 The amount of variation in
distributions of these neighbourhoods matched to empirical school quality was derived from a study that established its
data sources.4 Please see the online supplementary material for empirical variation.10 School quality directly affected the
details on these procedures. number of years an agent attended school; an agent's education
Policy simulations were run for 100 time steps, approxi- level was a function of both school quality and the number of
mately 2.5 generations, during which agents in the model were years of school attendance.
born, changed neighbourhoods (residential mobility), went to Physical activity infrastructure reflected the varied environ-
school at age 6 years, got jobs after they left school, retired at mental correlates of physical activity as measured in the Physical
age 65 years, had one child at age 25 years, and died at age- Activity Neighborhood Environment Survey (PANES).11
specific rates reflective of the US population. In the process, Specifically, in our model, physical activity infrastructure
these agents exhibited health behaviours (smoking, diet, exer- reflected the seven-point neighbourhood environment index (0-
cise) that produced health outcomes (BMI, cardiovascular 6) derived from PANES in prior work.12
disease and death). These health behaviours were dependent In the model, the most proximal drivers of BMI were healthy
on the agent's education level, neighbourhood school quality, diet and activity, our measures of diet behaviour and physical
neighbourhood physical activity infrastructure, access to activity, respectively. Healthy diet was designed to reflect the
healthy foods and the behaviours of others. The agents were quality of nutrient intake per the Healthy Eating Index-2005.13
connected via social ties (to represent social networks) and We derived the relation between good food stores and healthy
social norms affected the extent of social influence. The para- diet from a study on the relation of healthy food in nearby food
meters in the model were, wherever possible, based on pub- stores and a healthy diet.14 The relation between education level
lished reports, as described below and in the online and healthy diet was established using a study on the relation
supplementary material. At initialisation of the model, distribu- between education level and fruit and vegetable consumption.15
tions of BMI, income, education, diet and physical activity Activity was meant to reflect the number of MET-hours/week
levels were based on national data (National Health Interview for each agent. The relationship between activity infrastructure
Survey (NHIS), census and other sources) as indicated in the and activity was established using data on the relation between
online supplementary appendix. the built environment and activity levels12; the link between

Figure 1 Diagram of agent-rule -I- : Physical

structure related to agent body mass
index (BMI). The light-green boxes /"""
jr ' structure j
depict neighbourhood-level variables;
light-blue depicts agent-level variables.
The dotted-green circle highlights that
/
an agent's diet and activity behaviours
are also a function of direct social ties
/ X + Quali<y ' ' : . iL /->. Activity I I '
/ ' : Education ' M

' Level 1 1 Y
to other agents in the agent's social
network. The red plus and minus signs
Neighborhood ; ^ _ / / '. V,/
Income y School ļ // 1 I
express the nature of the causal rule; 1 /fc + ¡ Attendance I / ļ 1 + I

' ' ( X / ' I HDUy j ' J

positive is direct, negative is inverse.

' '_ Household

Income
^ y/ n+_ /
Í"yL
W
_
Good '

+ I Stores j social
Network

Orr MG, et al. J Epidemiol Community Health 20 1 6;70:862-867. doi: 10.11 36/jech-201 5-20562 1 863

This content downloaded from

172.56.20.163 on Sun, 11 f on Thu, 01 Jan 1976 12:34:56 UTC
All use subject to https://about.jstor.org/terms
education level and activity was derived directly from the 2007 by increases in BMI for whites. Second, we calculated the
National Health Interview Survey (NHIS).16 average BMI Disparity Index for each policy by policy strength,
averaging across all simulation time steps. Third, we plotted the
Outcome variable BMI Disparity Index for all 125 conditions at time steps 20, 30,
As shown in figure 1, BMI was a direct function of diet and 40, 60 and 100 using a heat map. In the heat map, we present
activity. Our assumption regarding the influence of healthy diet the percentage of change from baseline of BMI Disparity Index
on BMI was chosen to match the cross-sectional relationship defined as baseline minus time step X (positive values mean that
between the Healthy Eating Index (HEI-05) and BMI.17 The the BMI Disparity Index lessened). Baseline was defined as the
influence of activity on BMI was derived from a meta-analysis BMI Disparity Index at the beginning of the simulations under
on the effects of activity on weight loss.18 the condition in which none of the three policies were imple-
The primary outcome is the average difference in BMI mented. To give a sense of change over time, the heat map pre-
between blacks and whites (BMI Disparity Index). Each of the sents five time steps (20, 30, 40, 60, 100). Finally, we regressed
125 policy conditions (defined below) was simulated 20 times. the BMI Disparity Index onto the three policies using linear
Each of the 20 simulations, within a condition, produced a regression, separately for each of the time steps used in the heat
time-series of BMI for each race (ie, a BMI value for each time map analysis (each of the 125 policy conditions was a data
step from time steps 1 to 100). For each policy condition and point). In this analysis, the policies were measured on a continu-
each race, we calculated the average of the 20 BMI time-series ous scale in SD units to reflect policy strength - the value for
separately by race. Finally, we computed the BMI Disparity the no-policy condition was the average number of SDs across
Index for each policy condition by subtracting the average white policies for the lowest 20% of neighbourhoods, calculated
time-series from the average black time-series, thus producing a empirically from the simulation. For each time step, we ran two
BMI Disparity Index across time steps in the model, that is, a analyses. First, all three policies were simultaneously entered
time-series of the BMI Disparity Index. without interaction terms. Then, we entered all two-way and
three-way interactions between policies. If the results of the
Model analysis and experimental design interaction model explained significantly more variance in the
Policy manipulations BMI Disparity Index, we present both analyses for
We ranked the neighbourhoods on each policy variable, and the interpretation.
bottom 20% in each policy domain were targeted for the inter-
vention. We created 125 separate conditions, each representing RESULTS
a unique combination of the three policies and five policy
The percentage of the 125 conditions for which BMI increased
strengths. We varied the strength of each policy manipulation by over time was 6% for blacks and 3% for whites. Thus, reduc-
five levels. The strength of the policy manipulations in our simu-
tions in the BMI Disparity Index (shown in the analyses below)
lation experiments varied from one to four SDs from the were not due to increased BMI for whites.
average (which was computed empirically within the agent-
Table 2 presents the average change across all time steps in
based model). For example, given a policy manipulation of one
the BMI Disparity Index, by policy strength for each of the pol-
SD in school quality, we first computed the average and SD of
icies. The three policies each had an impact on BMI disparities,
school quality at initialisation of the simulation. Then, we com-
and there was a graded relationship between policy strength and
puted the level of school quality that is equal to one SD above
the extent to which disparities were reduced, with the greatest
the mean (mean plus 1 SD), and assigned this value to all of the
reduction, in absolute and relative terms, associated with phys-
neighbourhoods in the lowest 20% of school quality. Table 1
ical activity infrastructure.
shows the values of these manipulations in real terms.
Figure 2 presents the heat map analysis. This analysis provides
Under the conditions of multiple simultaneous policies, the
a graphic overview of the percentage of change in the BMI
targeting operated independently among the policies. Policies
Disparity Index from baseline across all policy combinations for
were implemented on start of the simulation and remained in
time steps 20, 30, 40, 60 and 100. We present this analysis, in
force for its full duration.
part, to make the design matrix of the policy experiment more
concrete and, in part, to provide an overview of how the BMI
Statistical analysis
Disparity Index varied both across policy combinations and over
The analyses were conducted in four steps. First, we calculated
time. From this analysis, it is clear that a variety of policy com-
the percentage of the 125 conditions for which the value of binations led to decreases in the BMI Disparity Index.
BMI at the end of the simulation was higher than at the begin-
Furthermore, most of the conditions in which the physical activ-
ning separately for blacks and whites. We used this to rule out
ity infrastructure policy was strong showed a large reduction in
the possibility that reductions in BMI disparities were generated

Table 2 Average percentage change in body mass index (BMI)

Table 1 Strength of policy manipulations in real terms
Disparity Index across policy strength over all time steps in the
Policy strength simulation

Variable No 1 2 3 4 Policy strength

Good food stores 0.5 1.0 1.3 1.7 2.4 Variable No 1 2 3 4

Walkability 3.5 4.2 4.5 4.8 5.1 Good food stores 26.9 30.5 32.8 34.5 35.8
School quality 25.6 14.7 11.7 9.3 7.4
Physical activity infrastructure -2.4 22.5 35.0 45.5 59.9
Numbers reflect the following, good food stores = School
number quality
supermarkets per18.0 27.4 32.3 37.8 44.8
10 000
residents, walkability = neighbourhood environment index, school quality = student/
teacher ratio. Average percentage change for all conditions = 32.1.

864 Orr MG, et al. J Epidemiol Community Health 2016;70:862-867. doi: 10.1 136/jech-201 5-205621

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms
Figure 2 Percentage change in the
body mass index (BMI) Disparity Index
across conditions by time steps. The
heat map showing the percentage
change in BMI Disparity Index from
baseline across all 125 conditions, by
time steps 20, 30, 40, 60, 100. The
metarows represent each time step,
the metacolumns represent school
quality strength. The microrows/
columns represent the strength of
good food stores and physical activity
infrastructure, respectively.

the BMI Disparity Index. The same was not true for the good related to the direct, shorter term effects of improved school
food stores and school quality policies. Finally, the overall quality as an agent progresses through 12 years of school, and
pattern of results seemed to vary somewhat over time. At most the other related to indirect, longer term effects of neighbour-
times, an increase in physical activity infrastructure reduced the hood change via a feedback loop that included school quality,
BMI Disparity Index, whereas strength of the school quality education level, household income, neighbourhood income, and
policy seems to have become more important over time. The
impact of the good food stores policy was weakest and did not
change much over time. Visual analysis of the heat map only Table 3 Associations between good food stores, physical activity
provides a somewhat heuristic view of the simulation results. infrastructure, and school quality and the BMI Disparity Index
Next we present a more rigorous statistical analysis of these
Time step
data.
20* 30* 40* 60* 100 100t
Table 3 shows the regression analysis at time steps 20, 30, 40,
60 and 100. Across the time steps, each of the three policy manip- Good food -0.010 -0.008 -0.007 -0.007 -0.008 -0.003
ulations was independently (p<0.05) related to a decrease in the stores (GFS)
BMI Disparity Index when the interaction terms were not Physical -0.046 -0.056 -0.060 -0.058 -0.049 -0.045
included in the analysis. The effects on the BMI Disparity Index activity
were strongest for the physical activity infrastructure policy com- infrastructure
(Al)
pared to the other two. Furthermore, the coefficient for the
school quality policy increased systematically as time steps School quality -0.002 -0.010 -0.016 -0.024 -0.034 -0.027
(SQ)
increased. Only at time step 100 did the interaction model fit
GFSxAl NS
better than the non-interaction model. In that case, the inter-
GFSxSQ -0.003
action model indicated that the effects of the good food stores
AlxSQ NS
policy was dependent on the school quality policy, and vice versa.
GFSxAlxSQ NS
Taken together, these results can be summarised as follows.
First, all the policies had some effect, with physical activity Values in the table are lin
p<0.05.
infrastructure showing the strongest effect. Second, the effects *The nested interaction model did not explain significantly more variance compared
of the school quality policy were time-dependent, becoming to the non-interaction model.
tThis is the interaction model for time step 1 00.
stronger as the simulation progressed. We hypothesise that the
BMI, body mass index; NS, not significant.
effects of school quality may have reflected two time delays, one

Orr MG, et al. J Epidemiol Community Health 2016;70:862-867. doi: 1 0. 1 1 36/jech-20 1 5-20562 1 865

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms
residential mobility. Third, the effects of the good food stores in unemployment, changes in the real estate market, changes in
and school quality policies showed some signs of a multiplica- corporate practice regarding targeted advertising for food pro-
ducts, which would potentially interact with the policies we
tive interaction, but only very late in the simulation, the nature
of which can be described as the effect of school quality as modelled in unanticipated ways. Similarly, our model does not
slightly dependent on good food stores. include several other population-level changes that are always in
flux over time, such as changes in population age structure and
DISCUSSION the distribution of race/ethnicity. Fourth, the policy interven-
These results indicate that an agent-based model can be used to are, in some cases, heroic, involving large changes from
tions
encompass a number of the determinants of BMI, and they existing policies. Fifth, the model is, at its core, dependent on
suggest that neighbourhood-focused upstream policies might parameter estimates and temporal processes, where possible,
reduce disparities in BMI between blacks and whites. This is in
gleaned from the existing literature. Given the absence of exten-
accord with the large literature on the important role ofmeta-analyses in each domain, we chose best estimates for
sive
upstream determinants of population health.19-21 these based on our assessment of the literature. We include an
A key finding from our simulations is that reducing racial extensive
dis- appendix that provides considerable transparency as to
parities in BMI using upstream policy may take time, andthe thebases for estimating, in some cases speculatively, the sources
time signatures of different policies may differ. This isfor notthese estimates. It would be reassuring to be able to conduct
necessarily surprising, but simply reinforces the theme that extensive
even sensitivity analyses for each of the parameters in the
strong social policies may affect population health slowly.22 model,
To but the number of such analyses can increase exponen-
tially with the number of parameters, thereby making such an
some extent, this may reflect the process by which policies move
through the population. For example, the impact of improving exploration desirable but infeasible. While it is not possible to
school quality on BMI would be expected to be slower than predict how the results will change as the literature in these
intervening on the food or physical activity environment, areasas it evolves and improves, our simulation approach does make
requires cohorts of students to move through the educational it possible to periodically update conclusions. Relatedly,
system. In fact, table 2 suggests that over time, the impact of
although there are existing frameworks and methods for model
validation (eg, see refs. 25 and 26 for two opposing but useful
school quality is accelerating, whereas the impact of improving
the neighbourhood physical activity environment does not viewpoints), these are difficult to apply to abstract,
change much over time. Thus, the effects of upstream policies proof-of-concept models, as the choice of validation targets are
on population health and health disparities may manifest over not aobvious - for example, our model does not include gender,
variety of time windows. and has artificial geography, so it is not entirely clear what
would be a valid BMI target.
While part of the rationale for simultaneously examining mul-
Given these limitations which, presumably, resulted in our
tiple policies was the possibility that we might find potentiation
model showing a smaller black/white disparity in BMI than
or interference among them, we found little evidence for inter-
action, with interactions only appearing after 100 time stepsexpected
of from the literature, the results of this modelling exercise
the model. While this evidence for interaction is interesting,should
one not be interpreted literally, but as a proof-of-concept that
using a systems approach to study obesity disparities has great
must be very wary of projecting so far into the future, as many
of the parameter values that drive the simulation may change promise, an indication of the level of complexity and specifica-
over time. tion required for such a model, and as a tool for highlighting
Although our model was based on hundreds of variables gaps and in available empirical data to inform parameters. In fact,
parameters that were, where possible calibrated to existing there
data is considerable strength that comes from being able to sim-
sources, the model is intended as a proof-of-concept ultaneously
and simulate the impact of three upstream policies on dis-
remains an abstract representation of what drives black/white parities in BMI in a format that allows policy effects to evolve
over generations, with a combination of policies that are unlikely
obesity disparities in the USA, thus the results cannot be inter-
preted for policy purposes, but are only suggestive. There areto bea simultaneously implemented in a population-based trial,
number of caveats. First, the scope of the model is relativelyand with processes such as individual preferences and residential
narrow. It does not incorporate many elements such as macro- mobility explicitly represented. While we believe that the use of a
economic factors, the food production industry and market,systems
the science tool, such as agent-based modelling, has much to
media, healthcare, and it excludes some important agent-level offer, the enterprise is still very much in its infancy. With further
decision-making mechanisms, such as attitudes, beliefs and developments, and with cross-fertilisation from other
approaches, we may be able to move forward in understanding
intentions. Second, some of the structures in the model are sim-
the critical challenges to population health and health disparities
plified in a way that does not afford an accurate representation
of the US population - for example, agents have no gendersuch and as those represented by the obesity epidemic.
only a very simple social network structure, a relatively small
preferential attachment network that is known to capture some
but not all qualities of human social networks.23 An alternative
for future research similar to what we presented here would be
to use synthetic populations that capture key aspects of the US
population,24 and to include more information regarding key
Considerable evidence suggests that education and access to
issues such as a focus on changes in weight in children and ado-
healthy food and physical activity are related to body mass
lescents, and inclusion of other racial/ethnic groups (eg, Latino).
index (BMI), and to differences in levels of BMI between
Third, the time-horizon of the policy scenarios was very long.
population subgroups. Efforts to leverage these drivers of BMI
While the BMI Disparity Index is essentially eliminated for
must contend with the inherent complexity of what is
some policy combinations after 20 years, our model did not
considered the obesity system, a complex web of interacting
capture long-term and large-scale population changes over time
social and behavioural processes.
in social and economic conditions-for example, rapid changes

866 Orr MG, et al. J Epidemiol Community Health 201 6;70:862-867. doi: 1 0. 1 1 36/jech-201 5-20562 1

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms
 9 Mishel L, Rothstein R, eds. The class size debate. Washington DC: Economic Policy
Institute, 2002.
10 Eide E, Showalter MH. The effect of school quality on student performance: a
quantile regression approach. Econ Lett 1998;58:345-50.
1 1 Sallis JF, Kerr J, Carlson JA, et al. Evaluating a brief self-report measure of
A proof-of-concept that agent-based simulations can increase
neighborhood environments for physical activity research and surveillance: physical
our understanding of the complexities and uncertainties in activity neighborhood environment scale (PANES). J Phys Act Health
complex social health problems. Specifically, this study explored 2010;7:533-40.
the potential to reduce disparities in body mass index between 12 Sallis JF, Bowles HR, Bauman A, et al. Neighborhood environments and physical
activity among adults in 1 1 countries. Am J Pre v Med 2009;36:484-90.
blacks and whites in the USA via large-scale
13 Guenther PM, Reedy J, Krebs-Smith SM, et al. Evaluation of the Healthy Eating
neighbourhood-focused interventions that target quality of lndex-2005. J Am Diet Assoc 2008;108:1854-64.
education and access to healthy food and physical activity 14 Franco M, Diez Roux AV, Nettieton JA, et al. Availability of healthy foods and
infrastructure. dietary patterns: the Multi-Ethnic Study of Atherosclerosis. Am J Clin Nutr
2009;89:897-904.
1 5 Krebs-Smith SM, Cook A, Subar AF, et al. US adults fruit and vegetable intakes,
1989 to 1991: a revised baseline for the Healthy People 2000 objective. Am J
Public Health 1995;85:1623-9.
Funding Institute for Integrative Health; Network on Inequalities, Complex Systems,
16 Statistics NCfH. Data File Documentation , National Health Interview Survey, 2007
and Health (grant no. HHSN276200800013C); Robert Wood Johnson Foundation
(machine readable data file and documentation). Hyattsville, MD: National Center
(grant no. 60466).
for Health Statistics CfDCaP, 2008.
Competing interests None declared. 17 Gao SK, Beresford SAA, Frank LL, et al. Modifications to the Healthy Eating Index
and its ability to predict obesity: the Multi-Ethnic Study of Atherosclerosis. Am J
Provenance and peer review Not commissioned; externally peer reviewed.
Clin Nutr 2008;88:64-9.
18 Miller WC, Koceja DM, Hamilton EJ. A meta-analysis of the past 25 years of weight
REFERENCES loss research using diet, exercise or diet plus exercise intervention. Int J Obes Relat
1 Glickman D, Parker L, Sim U, et al. Accelerating progress in obesity prevention: Metab Disord 1997;21:941-7.
solving the weight of the nation. National Academies Press, 2012. 19 Galea S, eds. Macrosocial determinants of population health. New York: Springer,
2 Vanderbroeck P, Goossens J, Clemens M. Tackling obesities: future choices-building 2007.
the obesity system map. UK: GOfS, 2007. 20 Health CoSDo. Closing the gap in a generation: Health equity through action on the
3 Hammond RA. Complex systems modeling for obesity research. Pre v Chronic Dis social determinants of health. Geneva: World Health Organization, 2008.
2009;6:A97. 21 Kaplan GA, Everson SA, Lynch JW. The contribution of social and behavioral
4 Massey DS, Gross AB, Shibuya K. Migration, segregation, and the geographic research to an understanding of the distribution of disease: a multilevel approach.
concentration of poverty. Am Sociol Rev 1994;59:425-45. In: Smedley BD, Syme SL, eds. Promoting health: intervention strategies from social
5 Moore LV, Diez Roux AV. Associations of neighborhood characteristics with the and behavioral research. Washington DC: National Academy Press, 2000:31-55.
location and type of food stores. Am J Public Health 2006;96:325-31. 22 Sterman JD. Learning from evidence in a complex world. Am J Public Health
6 Cobb LK, Appel U, Franco M, et al. The relationship of the local food environment 2006;96:505-14.
with obestiy: a systematic review of methods, study quality and results. Obesity 23 Newman M. Networks: an introduction. New York: Oxford University Press, 2010.
(Silver Spring) 2015;23:1331-44. 24 Kumar S, Grefenstette JJ, Galloway D, et al. Policies to reduce influenza in the
7 Card D, Krueger AB. Does school quality matter? Returns to education and the workplace: impact assessments using an agent-based model. Am J Public Health
characteristics of public schools in the United States. J Political Economy 2013;103:1406-11.
1992;100:1-40. 25 Windrum P, Fagiolo G, Moneta A. Empirical validation of agent-based models:
8 Card D, Krueger AB. School quality and the return to education. In: Burtless G, eds. alternatives and prospects. J Artif Soc Soc Simulation 2007; 10:8.
Does money matter? The effect of school resources on student achievement and 26 Carley K. Validating computational models. Pittsburgh, PA: Carnegie Mellon
adult success. Washington DC: Brookings Institute, 1996:118-19. University, 1996.

Orr MG, et al. J Epidemiol Community Health 2016;70:862-867. doi: 1 0.1 136/jech-201 5-205621 867

This content downloaded from

172.56.20.163 on Sun, 11 Jul 2021 19:29:36 UTC
All use subject to https://about.jstor.org/terms