Social network analysis and agent-based modeling in social epidemiology

Department of Public Health, University of Oxford, Oxford, UK. .
Epidemiologic Perspectives & Innovations (Impact Factor: 1.58). 02/2012; 9(1):1. DOI: 10.1186/1742-5573-9-1
Source: PubMed


The past five years have seen a growth in the interest in systems approaches in epidemiologic research. These approaches may be particularly appropriate for social epidemiology. Social network analysis and agent-based models (ABMs) are two approaches that have been used in the epidemiologic literature. Social network analysis involves the characterization of social networks to yield inference about how network structures may influence risk exposures among those in the network. ABMs can promote population-level inference from explicitly programmed, micro-level rules in simulated populations over time and space. In this paper, we discuss the implementation of these models in social epidemiologic research, highlighting the strengths and weaknesses of each approach. Network analysis may be ideal for understanding social contagion, as well as the influences of social interaction on population health. However, network analysis requires network data, which may sacrifice generalizability, and causal inference from current network analytic methods is limited. ABMs are uniquely suited for the assessment of health determinants at multiple levels of influence that may couple with social interaction to produce population health. ABMs allow for the exploration of feedback and reciprocity between exposures and outcomes in the etiology of complex diseases. They may also provide the opportunity for counterfactual simulation. However, appropriate implementation of ABMs requires a balance between mechanistic rigor and model parsimony, and the precision of output from complex models is limited. Social network and agent-based approaches are promising in social epidemiology, but continued development of each approach is needed.


Available from: Peter Scarborough
Social network analysis and agent-based
modeling in social epidemiology
Abdulrahman M El-Sayed
, Peter Scarborough
, Lars Seemann
and Sandro Galea
The past five years have seen a growth in the interest in systems approaches in epidemiologic research. These
approaches may be particularly appropriate for social epidemiology. Social network analysis and agent-based
models (ABMs) are two approaches that have been used in the epidemiologic literature. Social network analysis
involves the characterization of social networks to yield inference about how network stru ctures may influence risk
exposures among those in the network. ABMs can promote population-level inference from explicitly programmed,
micro-level rules in simu lated populations over time and space. In this paper, we discuss the implementation of
these models in social epidemiologic research, highlighting the strengths and weaknesses of each approach.
Network analysis may be ideal for understanding social contagion, as well as the influences of social interaction on
population health. However, network analysis requires network data, which may sacrifice generalizability, and causal
inference from current network analytic methods is limited. ABMs are uniquely suited for the assessment of health
determinants at multiple levels of influence that may couple with social interaction to produce population health.
ABMs allow for the exploration of feedback and reciprocity between exposures and outcomes in the etiology of
complex diseases. They may also provide the opportunity for counterfactual simulation. However, appropriate
implementation of ABMs requires a balance between mechanistic rigor and model parsimony, and the precision of
output from complex models is limite d. Social network and agent-based approaches are promising in social
epidemiology, but continued development of each approach is needed.
Keywords: Complex systems, epidemiology, social factors, social class, ethnicity, modeling, networks
Social epidemiology and systems thinking
Social epidemiology is concerned with the social variation
in, and the s ocial determina nts of the distribution of
health and disease [1]. This branch of epidemiology is
fundamentally interested in the influences of social fac-
torssuch as individual attributes (i.e., social class and
ethnicity) [2,3]; behaviors (i.e., diet and physical activity)
[4,5]; construct s of social interactio n (i.e. , soc ial supp ort
and social cohesion) [6]; contextual influences (i.e.,
neighborhoods and regions) [5,7]; and the influences of
the allocation of individua ls in space (i.e. race/income
segregation) on the distribution of health and disease in
populations [8,9]. By patterning exposure to disease risk
factors, social factors themselves become fundamental
determinants of health [10]. Social epidemiology has
made great strides during the past two decades. However,
as the field grows, it is becoming readily apparent that
some of the its tools may be limiting. In particular, the
reductionist linear models that are the lingua franca of
epidemiologic analyses are limiting in important ways.
First, the dynamics of populations, in terms of health
and disease, emerge from the behaviors and interactions
of the heterogeneous individu als that c omprise them. In
this way, interaction undergirds many of the mechanisms
that mediate the social production of health and disease.
These interactions may operate both on the macro-scale
between social exposures acting at multiple levels, and on
the micro-scale between individuals within populations.
Interactions challenge the current epidemiologic toolkit
in several ways. As social factors may interact in complex
ways to determine health and disease risk, the current
risk factor approach to epidemiology, which emphasizes
decontextualized, independent effect measures for expo-
sures may not be appropriate [11,12]. For example,
* Correspondence:
Department of Public Health, University of Oxford, Oxford, UK
Full list of author information is available at the end of the article
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
© 2012 El-Sayed et al; licensee BioMed Central Ltd. This is an O pen Access article distributed under the terms o f the Creative Commons
Attribution License (http://c nses/by/2.0), which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.
Page 1
studies have demonstrated that the relation between eth-
nicity and health indicators may be modified by ethnic
density in the area of residence; this observation has been
docum ented for he alth indicators including adverse birth
outcomes [13,14], asthma [15], psychopathol ogy [16],
suicide [17-19], and mortality [20]. This observation chal-
lenges our current approaches because it suggests that
the relation between ethnicit y and health may be hetero-
gen eous, and th at conceptualizing this re lationship inde-
social variab ility in health may be mediated by the degree
and nature of social interaction within and between social
groups. In this regard, several studies have sho wn that
social interaction may transmit non-infectious disease
outcomes [21-24]. Furt hermore, research about the
health influences of social interactions suggests that
population-level modes of social interaction, such as
social cohesion, social capital, and social support, may
shape population health and disease distribution
[6,25,26]. Ultimately, however, social interaction does not
lend itself to the reductionist analytic paradigm that we
employ, as potentially important social interactions
between individuals in a population violate the central
assumption of independence of observations in regres-
sion approaches.
Second, p opulation dynamics feature nonlinearity,
whereby change in disease risk is not always proportional
to the change in exposure, and feedback, where disease
can modulate exposure just as exposure can modulate dis-
ease. These dynamics are not often explored in social epi-
demiology, although they may have profound implications
for population health. For ins tance, a central observation
in social epidemiology is that low social status predi cts
poor health [27]. However, poor health can also predict
low social status [28]. Therefore, mutually reinforcing in a
positive feedback loop, low social status and poor health
may ultimately converge, with reinforcing implications for
a third social illinequality (which itself plausibly f eeds
back on low s ocial status and poor health) [29,30]. As is
characteristic of positive feedback loops, the relationships
between social status, health, and inequality are likely to
feature nonlinear, accelerating behavior because of amplifi-
cation at each turn of the loop. As an illustration of the
inability of the current epidemiologic paradigm and toolset
to negotiate these dynamics , consider the use of dir ected
acyclic graphs (DAGs) in traditional epidemiologic ana-
lyses. DAGs are mental models used to specify and forma-
lize the causal relationships between exposures and
outcomes. However, like the regression models they edu-
cate, these mental models, by definition, forbid cyclical
relationships between exposure and outcome, and t here-
fore the feedback and reciprocity that likely characterize
the true relationships between them.
Third, the c ounterfactual conceptual framework that
underpins epidemiologic inquiry falls short when consid-
ering both fundamental social causes and macrosocial
causes of disease. Our etiologic understanding of the
social determinants of disease rests on th e counterfactual
exercise of contrasting outcome occurrence probabilities
corresponding to two or more m utually-exclusive expo-
sures [11,12]. However, social factors, of fundamental
importance in social epidemiology, such as race, ethnicity
and gender, are attributes of individuals, rather than
exposures. Because these attributes are fundamental to
identity, authors have argued that the counterfactual
approach is theoretically implausible [30-33]. Similarly,
understanding macrosocial causes requires the assump-
tion that a counterfactual universe could be unchanged
barring a lar ge-scale social cause. However, causes across
levels are inevitably i nterlinked, suggesting that an alter-
nate universe comparable to the present universe , save
changes in a macrosocial influence, may also be
These three challenges may be limiting the progress of
social epidemiology at this stage in its evolution [34], and
have resulted in calls to adopt newer methods that can
overcome them [32,34-36]. Several authors have suggested
the adoption of systems approaches in social epidemiology
as a way past these challenges [37-39]. Systems thinking
suggests that complex dynamic systems, such as popula-
tions, which feature multiple interdependent components
whose interactions may include feedback, non-linearity,
and lack of centralized control [40], are best understood
holistically [41]. This epistemological approach is best
contrasted to reductionism, which suggests that systems
are best understood by aggregating information gathered
via the independent study of their components. By con-
trast, a systems approach implies that the dynamics and
behavior of a system are different, qualitatively, from those
of the sum of its parts. A systems approach, therefore,
emphasizes the dynamics of relationships between compo-
nents of a system, rather than th e characteristics of those
components themselves [41,42].
Two systems approaches that may be part icularly use-
ful in social epidemiology include social network analy-
sis and agent-based modeling. With respect to social
epidemiology, the first involves the characterizatio n of
the structures of social networks or subsets of these net-
works to understand their influence on health behaviors
and outcomes. The second involves the use of stochastic
computer simulations of si mulated individuals, in simu-
lated space, over simulated time to understand how
macro-level health and disease distribution patterns may
emerge from explicitly programme d, micro-level health
behaviors, social interactions, and movement of these
individuals in their environments.
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 2 of 9
Page 2
A developing body of work has begun to apply these
approaches to address social epidemiologic research ques-
tions. For example investigators have used these
approaches to better understand the social etiology of
complex conditions, [21-24,37,43-46] such as obesity
[24,43]. Particularly compelling is a recent high-profile
study by Christakis and Fowler [24], which used social net-
work analysis to demonst rate the spread of obesity via
social relatio nships in a social network. Another study
used stochastic networks nested within agent-based mod-
els (ABMs) to assess strategies for population-level obesity
prevention [43]. Two more recent studies used agent-
based models to understand t he m ech anisms underlying
socioeconomic disparities in diet quality [47], and to assess
how resource allocation may influence socioeconomic dis-
parities in walking behavior [48].
While these papers are good early examples of the adop-
tion of complex systems approaches to social epidemiolo-
gic inquiry, the field remain s young. Here, we aim to
synthesize the extant literature that has called for applica-
tions of these approaches in population health. We will
begin with an examination of each method and its
approach, and then examine each methods strengths and
applicability with regard to social epidemiologic research,
suggesting particular avenues where each may be appro-
priate. Finally, we will discuss limitations to the application
of each method in the field.
Social network analysis
A summary of social network analytic approaches
In the context of social epidemiology, social network ana-
lysis involves the characterization of the structures of
social networks or subsets of networks so as to understand
the flow of health-relevant facto rs (i.e., disease, informa-
tion, social support, etc.) between network nodes. Network
approaches emphasize the structural characteristics of net-
works rather than the characteristics of their nodesthis
implies that the social ties that bind actors have important
consequences for their behavior [49]. Social network ana-
lysis is particularly interested in the patterns and implica-
tions o f relationships between social actors [50], and is
therefore most valuable in characterizing population-level
outcomes when there are relational characteristics
involved in the behavior of networked individuals [51].
Social network analysis has three main branches: (a) net-
work visualization and (b) network characterization, each
principally descriptive, and (c) emerging methods around
stochastic and longitudinal network analysis [50-52]. Net-
work visualization is the process of diagramming network
connections in two an d three-dimensional space so as to
visualize netwo rk structure and relationships [51]. Net-
work characterization involves analyses directed toward
understanding the roles played by individual actors,
subgroups of actors, or overall network structures in char-
acterizing the flow of factors of interest within networks.
For example, at the level of the individual, analyses might
address the number of connections a particular actor has
within a network, the degr ee to which the actor bridges
between other actors in the network, the social distance
(measured in relationships) between an actor and other
actors, or the degree of connectedness of an actor relative
to others [50-52]. By contrast, analyses of network sub-
structures or full networks may address degrees of connec-
tivity between actors, the degree of centralization or
hierarchy in a network, or lengths of paths between parti-
cular actors of interest [50-52].
A third wing, emerging from network characterization,
aims to develop methods for inferential analysis and
hypothesis testing regarding network influences. These
methods, still in development, include stochastic, and
longitudinal network analytic techniques [51,52]. Longitu-
dinal network analysis allows investigators to study tem-
poral changes in networks, their characteristics and
dynamics, and/or t he characteristics of their constituent
parts, while stochastic analy tic techniques allow for the
construction of network models for use in simulations
Applicability of social network approaches in social
epidemiologic research
Social epidemiology is interested in the influences of social
factors on health and disease d istribution in populations.
Because of its focus on social interaction as a potential dri-
ver of indi vidual and collective characteristics and beha-
viors, social network analysis has the potential to yield
valuable insight about the social production of health and
disease in three areas: understanding the social contagion
of non- infectious exposures and outco mes [21-24,44-46],
understanding the role of social network str ucture as a
determinant of population health disparities, and under-
standing the influences of modes of social interaction,
such as social support, social cohesion, and social capital
on population health [51].
1. Understanding the social contagion of non-infectious
exposures and outcomes
Social network analysis is particularly useful for studying
how social phenomena spread through social networks
and influence health in this manner. Using social network
analysis, sever al studies have demonstrated the spread of
non-infectious conditions through social networks, includ-
ing obesity [24], smoking [20,44], alcohol u se [45], back
pain [46], teen substance use [21,22], and general well-
being [23]. These findings suggest that other complex
exposures, conditions, and diseases of population health
interest also may be, in part, communicable.
However, the current literature that has employed
social network approaches to understand the commu-
nicability of non-infectious health outcomes has only
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 3 of 9
Page 3
scratched the surface of the applicability of these
approaches. For example, while it h as been suggested
that obesity may be communicable via network ties [24],
little is known about heterogeneity thereinwhy might
some contacts of o bese individuals become obese while
others may not? Moreover, does the answer to the pre-
vious question lie in characteristics of the obese con-
tacts, the characteristics of those exposed, or in
characteristics of the relationships they share? Finally,
while these approaches have yielded insight about the
social etiology of these outcomes, we know little about
the potential for interventions to exploit network
2. Assessing the role of social network structure as a
determinant of population health disparities
Differences in the density and character of social networks
connecting individuals within different social strata could
partially explain population-level disparities between those
strata. In that sense, these approaches may be used to
study the mediation of health disparities by characteristics
of social networks. For example, ethnic minority groups in
their white counterparts [53]. Moreover, social networks
have been shown to be et hnically and racial ly segregated,
and those among ethnic and racial minority grou ps ha ve
been shown to be stronger and more highly-intercon-
nected than those among their White counterparts [54]. If
obesity has a communicable etiology, as findings from
Christakis and Fowler suggest [24], then it is plausible that
obesity may spread faster and more complet ely in more
dense social networks, and therefore faster and more com-
pletely among minority groups than among whites. In this
manner, differences in social network structure among
ethnic majo rity and minority groups may influence the
spread of obesity among them, helping to explain dispari-
ties in obesity. Therefore, differences in social network
characteristics across strata at several levels, including eth-
nicity, socioeconomic status, and educ ation at the indivi-
dual-level, as well as deprivation or ethnic density at the
area-level, may mediate differences in health between
these groups. Network approaches, used in this way, may
yield insight into the mechanisms underpinning popula-
tion health disparities.
3. Social network analysis and the influences of social
interactional constructs on population health
With the potential to distinguish between the effects of the
nature and volume of social interactions, the environmen-
tal contexts in which they exist, and the characteristics of
the actors involved, social network approaches provide the
correct framework within which to conceptualize and
operationalize modes of social interaction, like social sup-
port and social capital [51], which have been shown to
influence population health [6,55,56].
This is an important methodological development, as
standard regre ssion-based approa ches are not suited to
analyze these constructs because they are fundamentally
relational, and therefore violate the assumption of inde-
pendence of observations. At best, some investigators have
attempted to use multilevel regression techniques to
incorporate a measure of social support, cohesion or capi-
tal at the area-level [57,58]. Aside from the issue of con-
founding by other area-level factors correlated with these
interaction constructs, these approaches cannot accurately
capture the role of these exposures in heterogeneous
populations, where social support is not evenly distributed.
Rather, because social network analysis a llows investiga-
tors to more accurately represent and analyze these con-
structs, this methodology is well situated for studies of the
relationship between social interaction and population
Limitations to the application of social network approaches
in social epidemiology
Social network analysis is not without limitations. T wo
principal limitations are the implicit trade-off between the
use of network analytic techniques and generalizability in
net work data, and the problem of confounding by either
homophily and/or by shared environments in studies
about social induction of exposures or outcomes through
With regard to the first limi tation, b eyond data about
the char acteristics of individuals in networ ks (traditional
data collected in health surveys), social network analysis
requires data about the relationships between individuals.
Traditional sampling techniques designed to im prove
study generalizability by randomly sampling across envir-
onments are not conducive to using network approaches,
as the data yielded about relationships from these techni-
ques is not of sufficient completeness or quality to support
them. For example, a national survey may collect data
about exposures and health outcomes among a random
sample of the population, as well as about the number of
social contacts each respondent has, but it does not collect
data about exposures and outcomes among those respon-
dents. Conversely, studies attempting to maxim ize the
quality of network data may have limited generalizability.
Therefore, because of cost and feasibility constraints,
investigators interested in applying network methodologies
in their work may be forced to balance tradeoffs between
the analytic benefits of social network approaches and the
importance of generalizability when planning epidemiolo-
gic studies.
Studies about the networked spread of exposures or
disease are primarily interested in social induction, or
the causal influence of social interaction on behavior or
health outcomes. The second limitation to the use of
social network analysis in social epidemiologic research
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 4 of 9
Page 4
is the d ifficulty of adjusting for either homophily or
environmental effects, two potential confounders in ana-
lyses about the causal influence of induction [59-61].
Homophily is the tendency for agents with similar a
priori likelihoods of developing an outcome to preferen-
tially form social relationships [24,54]. Along with
homophily, shared spatial e nviro nments between indivi-
duals in networks can also confound studies of social
induction. A series of recent studies have drawn atten-
tion to the difficulty of disentangling social inductio n
from confounding by homophily and environmental
effects (see work by Cohen-Cole and colleagues [60,61]).
Given these challenges, the development of methods to
differentiate induction from its potential c onfounders
remains an active area of research [59].
Agent-based models
A summary of agent-based modeling
ABMs are stochastic computer simulations of simulated
age nts, or individuals, in simulated space, over simu-
lated time. These models allow macro-level behavioral
patterns to emerge from explicitly describ ed, micro-level
behaviors, interactions, and movements of agents in their
environments. Because model conceptualization and
parameterization take place from the bottom, up these
models are ideal for assessing emergence, or macro-
level patterns that arise from micro-level behavior [62].
Emergence, an idea that is central in systems sciences,
may allow for tremendous new insight into important
questions in the social and natural sciences [63].
Agent-based approaches are particularly appropriate
when: 1) individual agent behavior is complex, featuring
learning and adaptation, feedback loops, and/or recipro-
city; 2) when heterogeneous environments can influence
agent behavior and interaction, and agents are not fixed
in space or time, and 3) when inter-agent interactions are
complex, non-linear, and influence agent behavior
[39,62]. In sum, agen t-based approaches are ideal when
agent behavior is a complex function of agent attributes
and characteristics, environments, and inter-agent inter-
action over time.
Agent-based modeling requires the investigator to expli-
citly describe and program agent characteristics and updat-
ing rules during implementation. This includes specifying
agent characteristics and behaviors, as well as changes to
them with time (e.g., learning and adaptation). Agents can
be nested within social networks that influence the degree
and character of inter-agent interaction, and social interac-
tion can be programmed to influence future behavior.
Moreover, investigators can explicitly define the space
within which agents are situated through time, as well as
the influence of that space on agent behavior with time.
ABMs are particularly well-suited for research that is
concerned wi th understanding soc ial processes because
they maintain the centrality of the individual agent and
its attributes, characteristics, and behaviors in the pro-
duction of population-level phenomena. This is con-
trasted with other methods, such as regression models
or differential equations (e.g., laws that determine
dynamics of predato rs and prey), which focus on aggre-
gated data [63]. For this reason, agent-based approaches
have beco me increasingly common throughout the
social sciences, with applications in economics [64,65],
sociology [66], and political science [67,68].
Applications of agent-based modeling in social
epidemiologic research
ABMs place a focus on the individual and the individuals
characteristics and interactions in time and space. They
also allow investigators to run multiple simulations under
various model conditions, thereby isolating the effects of
particular conditions on outcomes of int erest. Therefore,
this approach has the potential to move social epidemiol-
ogy forward in four important ways. First, ABMs move
beyond the limitations of reductionist approaches that
have centered social epidemiology around measurement of
decontextualized risk factors that do not account for
interrelatedness and reciprocity between social exposures.
Second, ABMs may play a useful role in helping us under-
stand causal inference in social epidemiology. Third, this
approach may allow investigators to articulate and explore
mechanisms that underlie our understanding of the social
production of health. Fourth, agent-based approaches may
provide a more robust means to forecast the outcome of
policy interventions.
1. Moving away from independent effects in social
Individual attributes and behaviors, social interaction, and
environmental factors are largely interdependent, interact-
ing to shape health and disease distribution. However the
dominant re ductionist models tha t are used in the field
are limited in their ability to analyze interactions, feedback,
and reciprocity between exposures with on e another and
with outcomes [34], For these reasons, several authors
have challenged the assumption that elucidating risk fac-
tors or independent effects should be the object of epide-
miology [32,34,36,37].
ABMs present a departure from this approach. They
allow investigators an opportunity to model the influences
of individual, inter-individual, and environmental factors
in a m echanistically cogent manner. Unlike reductionist
approaches, ABMs are uniquely suited for the assessment
of the simultaneous etiologic effects of heterogeneous
population characteristics, social intera ction, and the
environment. They also allow for the exploration of
mechanis tic interactions, feedback loops, and reciprocity
between exposures and outcomes operating at multiple
levels in the etiology of complex conditions. For these rea-
sons, these methods may allow us to move beyond the
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 5 of 9
Page 5
reductionist measurement of independent risk factors,to
a more realisti c and nuanced unders tanding of the com-
plex production of health and disease.
2. Agent-based models and causal inference in social
ABMs may also lend themselves to the question of cau-
sal inferenc e in social epidemiology. Social epidemiolo-
gists are centrally concerned with understanding the
social production of disease [12]. This involves the
counterfactual exercise of contrasting outcome occur-
rence probabilities corresponding t o two or more
mutually-exclusive exposures [11,12].
In agent-based modeling, inve stigators program initial
conditions and update rules that specify the characteris-
tics of agents, embedded in networks, and placed within
spatial contexts, all of which influence update rules that
specify stochastically-applied changes in agent behavior
and characteristics each time-step, or simulated unit of
real time. Assuming no changes in initial conditions and
update rules, each simulation is identical within the
bounds of stochasticity. Calling this initial simulation the
control, an investigator can then simulate other models
where only one aspec t of the initial conditions or the
update rules is changed in each simulation these being
experimental simulations. Because both experimental
and control simulations are applied to the same popula-
tion of agents, this approach may bypass Hollands Fun-
damental Problem of Causal Inference, which exists
when it is impossible to observe the effects of multiple
exposures on the same unit in epidemiologic analysis
[31]. Therefore, by comparing any number of experi-
mental simulations to the control simulation, the
investigator can s imulate population-level experiments
using the ABM [63]. This experimental framework is par-
ticular powerful in social epidemiology, where there are
considerable logical limitations to causal inference from
observational models [11,12,69], and where randomized
social interventions are fraught with logistical, ethical,
and logical limitations of their own [12].
Therefore, ABMs allow investigators to use agent-
based counterfactual simulations (ABCs ) in social epid e-
miologic inquiry. In this way, they may allow for more
robust analyses of the effects of counterfactuals that chal-
lenge reductionist approaches. These include social fac-
tors that range from characteristics as fundamental in
social epidemiology as race or gender, to broader social
phenomena, such as macrosocial exposures, which are
difficult to operationalize using real world data. ABCs
allow investigators to identify the etiologic effects of
these factors by positing implausible counterfactuals that
deductively clarify their influences on population health.
For example, an investigator interested in the effects of
segregation on inequalities in obesity in a re al population
could counterfactually simulate the effects of complete
desegregation, comparing the degree of inequality in this
simulation to that of a segregated simulation. Anot her,
interested in the effects of income inequality o n suicide,
could simulate the effects of perfect income parity rela-
tive to varying levels of income inequality on suicide
overall and by subgroup. This method of analysis pose s
considerable strengths over curren t analytic paradigms,
largely relegated to cross-sectional ecologic analyses of
observation al data between vas tly different societies, lim-
ited by comparability and the inability to properly adjust
for confounding.
There is another potential benefit of this counterfactual
appr oach in social epidemiology: a more unifi ed concep-
tualization of disease etiology. K aufman and colleague s
argue that it is most appropriate to articulate exposures of
interest as defined int erventions that would eliminate
them [12]. Conceptually similar t o r andomized social
interventions in their experimental ontology, ABCs also
force us to articulate exposures as defined interventions
the prog rammed difference between the experimental
simulation and the control. In this s ense, investigators
using ABCs must frame ex posures of interest in terms of
the counterfactual in which the exposure in question does
not exist, an exercise that serves to clarify and define that
exposure more accurately.
3. Agent-based counterfactual simulations and the
exploration of etiologic pathways
The opportunity for ex perimentation via A BCs allows us
to explore mechanisms that underlie the social production
of disease. The population health implications of perturba-
tions on a hypothetical control model are driven both by
initial conditions, and by mec hanisms underlying update
rules that relate exposures with one another or with out-
comes. ABCs that attempt to replicate empiric o bserva-
tions can be run to clarify the mechanistic relationships
necessary in a model to yield real-world outcomes. In this
way, ABCs can allow investigators to study mechanisms
underlying the production of health and disease.
4. Modeling the effects of social policy interventions on
A potentially important contribution of ABMs is the
opportunity to test the outcomes of policy interventions.
This represents a key departure from current methods, as
traditional regression-based approaches are limited with
regard to yielding policy-relevant inference in several
ways. First, inference of policy implications from regres-
sion models is limited to interpreting the appropriate tar-
gets for interventions from independent effect measures.
This is limiting, as these effect measures have no direct
relevance to potential intervention schemes. Second,
because of limitations to regression models with regard to
the appropriate representation and operationalization of
macrosocial phenomena, macrosocial interventions are
difficult to study in this manner. Third, because the
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 6 of 9
Page 6
outcomes of regression models are specific to the data
analyzed, policy relevant analyses using these approaches
should be perf ormed on data from the population withi n
which a proposed policy intervention is to be enacted,
necessitating often costly and time-consuming data collec-
tion. Fourth, as disc ussed above, regression-based
approaches are not equipped to a ccount for reciprocity,
feedback, or non-linearity in relations between exposures
and outcomes, which may be important in understanding
the effects of policy interventions on outcomes (intended
and unintended).
By contrast, the counterfactual experimental approach
may allow for the simulated development and testing of
proposed health policy interventions [37,39]. Investigators
can tailor initial conditions to populations within which a
proposed policy intervention is to be enacted. The policy
intervention itself can then be operationalized as a coun-
terfactu al simulation and proposed interventional strate-
gies can be compared head-to-head or to baseline.
Limitations to agent-based approaches in social
Agent-based approaches require the investigator to bal-
ance the importance of mechanistic rigor (i.e. the inclusion
of relevant factors) and model parsimony (i.e. overcompli-
cating the model) [39,62]. In this sense, the process of
model implementation sh ould be tailored to questions of
interest to avoid unnecessary complexity [62]. However,
model tailoring may be logically problematic, as purposive
model tailoring implies the a priori exclusion of factors
that should have no apparent influence on the outcomes
of interest. Howev er, a central argument for agent-based
approaches is the ability of these models to yield emergent
phenomena, which r ely o n t he aggregation of complex
micro-level processes to yield macro-level insight. The
very notion of emergence suggests that the aggregation of
micro-level factors is likely too complex to allow for a
priori exclusion of any factor that could potentially influ-
ence, however indirectly, the outcome of interestor that
it is impossible to know which factors may have influences
on outcomes in the first place. Alternatively, because these
models are stochastically implemented, added computa-
tional factors will generally (although not in all cases)
increase the degree of uncertainty in simulation outcomes,
imposing a pragmatic limit to the number of factors that
can be included in a given model.
Along that line, another limitation is that ABMs, as
computer simulated models, produce quantitative output,
tempting investigators to quantit atively int erpret their
findings. However, ABMs can include any number of fac-
tors, each parameterized from any number of sources
and introducing their own biases and assumptions into
model output. Therefore, quantitative interpretation of
ABM output may not be appropriate [39,62]. This is par-
upon multiple sources for parameterization. In that
sense, ABMs may not be particularly useful when
attempting to forecast absolute population health indica-
tors, such as population prevalence or incidence of a
given outcome. Rather, these models are better su ited for
etiologic inquiry reliant upon qualitative comparisons of
output between counterfactual simulations. One tool that
may allow for more objective assessments of qualitative
differences in output is the us e of Monte Carlo simula-
tions [70]. Averaging model output over several Monte
Carlo simulations can provide confidence intervals on
mod el estimates that can then be used to elicit measures
of signifi cance on differences in output between counter-
factual simulations.
Another set of limitations arises when considering vali-
dation. ABMs, by their nature as systems models, are diffi-
cult (and sometimes impossible) to validate completely.
Generally, there are two strategies an investigator might
use to validate a model by situating it in reality. The inves-
tigator can 1) use real data to parameterize the model, or
2) work backwards by building a model from conceptual
relationships, the findings of which can then be compared
to real world observations. In this way, either the relation-
ships between factors in the model or the outcomes of the
model can be validated, but rarely both.
Both of these approaches are limited. In the first case,
the validity of counterfactual simulations depends on the
valid operationalization of factors in the model, w hich i s
nearly impossible to verify, even when those factors are
operationalized from real data. In the second case, where a
model parameterized based on conceptual relationships
between factors in the model is then validated using real
world observations, there is the threat of affirming the
consequen t. In this case, while a particular configuration
of factors in the model might produce outcomes that pre-
dict the obse rved data, it is possible that there are many
other plausible configurations that would also predict the
observed data, and no way to affirm t hat the investigator
has isolated t he one that operates in reality. However,
model building is not idiosyncratic and model construc-
tion should be educated by current knowledge about the
phenomena in question increasing the likelihood that the
specified conceptual model is accurate.
More generally , a central premise underlying the sys-
tems approach, as discussed above, is that the whole
may be different from the sum of its parts. Therefore, it is
plausible that outcomes that emerge from systems models,
such as ABMs, may contradict findings from reductionist
approaches [62,63]. There is a conceptual limitation to
validating emergent findings using models that cannot, by
definition, capture this phenomenon. Thus, it may be diffi-
cult to differentiate new insight yielded b y models which
capture emergence from spurious findings, because, by
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 7 of 9
Page 7
definition, another model capable of capturing emergence
Another important limitation to the use of agent-based
approaches is that they may be beholden to regression-
based parameters in their construction. Therefore, they
may incorporate into their findings m any o f the same
biases and limitations that arise from reductionist regres-
sion-based approaches.
Finally, agent-based approaches can be computation-
ally in tensive. Therefore, these a pproaches can require
considerable computing resources for efficient use [62].
The systems approaches discussed here, social network
analysis and agent-based modeling, have the potential to
reframe social epidemiology. Chief ly, these too ls allow
investigators to move social epidemiology beyond the
independent effects paradigm that some describe as
conceptually inappropriate [32,35] and reframe o ur pur-
suit of the complex social causes of health and disease in
a holistic framework. In addition, these approaches allow
investigators to understand the etiologic implications of
heterogeneity within the population, social interaction,
and environmental influencesimultaneously,andto
explore mechanistic interactions, feedback loops, and
reciprocity between exposures and outcomes. M oreover,
they can better articulate and provide a framework for
analyzing the health effects of social interaction. Finally,
the counterfactual approach made possible by a gent-
based modeling may promote causal thinking in social
epidemiology and improve our mechanistic understand-
ing and conceptual articulation of exposures when con-
sidering the social production of health.
However, severa l limitations need to be addressed as
these approaches become more prevalent in soci al epide-
miologic research. Considerable methodological develop-
ment is needed in the area of longitudi nal social network
approaches to improve causal inference from social net-
work analysis. Furthermo re, s ocial epidemiologist s inter-
ested in agent-based approach es to etiologic inquiry need
to develop best practices with regard ABM design, para-
meteriz ation, interpretation, and validation in population
health research.
List of Abbreviations
ABM(s) is an abbreviation for agent-based models (or modeling). ABCs is an
abbreviation for agent-based counterfactual simulations.
This study was funded in part by the Rhodes Trust (AME), the British Heart
Foundation (PS), and NSF award #0840889 (LS)
Author details
Department of Public Health, University of Oxford, Oxford, UK.
of Epidemiology, Columbia University, New York, NY, USA.
College of
Physicians and Surgeons, Columbia University, New York, NY, USA.
Department of Physics, University of Houston, Houston, TX, USA.
Authors contributions
AME conceived of the study and drafted the manuscript. PS advised on the
manuscript design and critically revised the manuscript for intellectual
content. LS helped with background research and critically revised the
manuscript for intellectual content. SG advised on the manuscript design
and critically revised the manuscript for intellectual content. All authors read
and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Received: 1 June 2011 Accepted: 1 February 2012
Published: 1 February 2012
1. Berkman LF, Kawachi I: Social epidemiology New York: Oxford University
Press; 2000.
2. Bhopal R, Hayes L, White M, Unwin N, Harland J, Ayis S, Alberti G: Ethnic and
socio-economic inequalities in coronary heart disease, diabetes and risk
factors in Europeans and South Asians. J Public Health 2002, 24(2):95-105.
3. Kaplan GA, Keil JE: Socioeconomic factors and cardiovascular disease: a
review of the literature. Circulation 1993, 88(4):1973-1998.
4. Bermudez OI, Falcon LM, Tucker KL: Intake and food sources of
macronutrients among older Hispanic adults: association with ethnicity
acculturation, and length of residence in the United States. J Am Diet
Assoc 2000, 100(6):665-673.
5. Saelens BE, Sallis JF, Black JB, Chen D: Neighborhood-based differences in
physical activity: an environment scale evaluation. Am J Public Health
2003, 93(9):1552-1558.
6. Cohen SE, Syme SL: Social Support and Health San Diego: Academic Press;
7. Sundquist J, Malmstrom M, Johansson SE: Cardiovascular risk factors and
the neighbourhood environment: a multilevel analysis. Int J Epidemiol
1999, 28(5):841-845.
8. Laveist TA: Racial segregation and longevity among African Americans:
an individual-level analysis. Health Serv Res 2003, 38(6):1719-1733.
9. Lynch JW, Smith GD, Kaplan GA, House JS: Income inequality and
mortality: importance to health of individual income, psychosocial
environment, or material conditions. BMJ 2000, 320(7243):1200-1204.
10. Link BG, Phelan JC: Understanding sociodemographic differences in
healththe role of fundamental social causes. Am J Public Health 1996,
11. Greenland S, Robins JM, Pearl J: Confounding and collapsibility in causal
inference. Stat Sci 1999, 14(1):29-46.
12. Kaufman JS, Kaufman S, Poole C: Causal inference from randomized trials
in social epidemiology. Soc Sci Med 2003, 57(12):2397-2409.
13. El-Sayed AM, Galea S: Community context, acculturation, and low birth
weight risk among Arab-Americans: Evidence from the Arab-American
Birth Outcomes Study. J Epidemiol Community Health 2010, 64(20):155-160.
14. Peak C, Weeks JR: Does Community Context Influence Reproductive
Outcomes of Mexican Origin Women in San Diego, California? J Immigr
Minor Health 2002, 4(3):125-136.
15. Cagney KA, Browning CR, Wallace DM: The Latino Paradox in
Neighborhood Context: The Case of Asthma and Other Respiratory
Conditions. Am
J Public Health 2007, 97(5):919-926.
16. Ostir GV, Eschbach K, Markides KS, Goodwin JS: Neighbourhood
composition and depressive symptoms among older Mexican
Americans. J Epidemiol Community Health 2003, 57(12):987-982.
17. Neeleman J, Wessely S: Ethnic minority suicide: a small area geographical
study in south London. Psychol Med 1999, 29(2):429-436.
18. Neeleman J, Wilson-Jones C, Wessely S: Ethnic density and deliberate self
harm a small area study in south east London. J Epidemiol Community
Health 2001, 55(2):85-90.
19. El-Sayed AM, Tracy MS, Scarborough PS, Galea S: Suicide among Arab-
Americans. PLoS One 2011, 6(2):e14704.
20. Eschbach K, Ostir GV, Patel KV, Markides KS, Goodwin JS: Neighborhood
context and mortality among older Mexican Americans: is there a barrio
advantage? Am J Public Health 2004, 94(10):1807-1812.
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 8 of 9
Page 8
21. Andrews JA, Tildesley E, Hops H, Li F: The Influence of Peers on Young
Adult Substance Use. Health psychology 2002, 21(4):349-357.
22. Pearson M: Drifting Smoke Rings. Connections 2003, 25(2):59-76.
23. Fowler JH, Christakis NA: Dynamic spread of happiness in a large social
network: longitudinal analysis over 20 years in the Framingham Heart
Study. BMJ 2008, 337(2):a2338.
24. Christakis NA, Fowler JH: The spread of obesity in a large social network
over 32 years. N Engl J Med 2007, 357(4):370-379.
25. Kawachi I, Kennedy BP, Glass R: Social capital and self-rated health: a
contextual analysis. Am J Public Health 1999, 89(8):1187-1193.
26. Hawe P, Shiell A: Social capital and health promotion: a review. Soc Sci
Med 2000, 51(6):871-885.
27. Phelan BG, Link J: Social conditions as fundamental causes of disease. J
Health Soc Behav 1995, , Spec No: 80-95.
28. Goldman N: Social inequalities in health. Ann N Y Acad Sci 2006,
29. Marmot MG, Wilkinson RG: Psychosocial and material pathways in the
relation between income and health: a response to Lynch et al. BMJ
2001, 322(7296):1233-1236.
30. Pickett K, Wilkinson R: The Spirit Level: Why greater equality makes societies
stronger New York: Bloomsbury; 2009.
31. Holland PW: Statistics and causal inference. J Amer Stat Assoc 1986,
32. Kaufman JS, Cooper RS: Seeking causal explanations in social
epidemiology. Am J Epidemiol 1999, 150(2):113-120.
33. Kaufman JS: Epidemiologic analysis of racial/ethnic disparities: some
fundamental issues and a cautionary example. Soc Sci Med 2008,
34. Kaplan GA: Whats wrong with social epidemiology, and how can we
make it better? Epidemiol Rev 2004, 26:124-135.
35. Diez Roux AV: Towards a realistic and relevant public health: the
challenges of useful simplification. J Public Health 2008, 30(3)
36. Diez Roux AV: Integrating social and biologic factors in health research: a
systems view. Ann Epidemiol 2007, 17(7):569-574.
37. Galea S, Riddle M, Kaplan GA: Causal thinking and complex system
approaches in epidemiology. Int J Epidemiol 2010, 39(1):97-106.
38. Galea S, Hall C, Kaplan GA: Social epidemiology and complex system
dynamic modelling as applied to health behaviour and drug use
research. Int J Drug Policy 2009, 20(3):209-216.
39. Auchincloss AH, Diez Roux AV: A new tool for epidemiology: the
usefulness of dynamic-agent models in understanding place effects on
health. Am J Epidemiol 2008, 168(1):1-8.
40. Ladyman J, Lambert J, Wiesner K: What is a complex system? Bristol:
University of Bristol 2009, Available at:
41. Meadows DH: Thinking in Systems: A Primer White River Junction: Chelsea
Green; 2008.
42. Skyttner L: General Systems Theory: Problems, Perspectives, Practice Singapore:
World Scientific Publishing; 2005.
43. Bahr DB, Browning RC, Wyatt HR, Hill JO: Exploiting social networks to
mitigate the obesity epidemic. Obesity 2009, 17(4):723-728.
44. Christakis NA, Fowler JH: The collective dynamics of smoking in a large
social network. N Engl J Med 2008, 358(21):2249-2258.
45. Rosenquist JN, Murabito J, Fowler JH, Christakis NA: The spread of alcohol
consumption behavior in a large social network. Ann Intern Med 2010,
152(7):426-33, W141.
46. Raspe H, Hueppe A, Neuhauser H: Back pain, a communicable disease? Int
J Epidemiol 2008, 37(1):69-74.
47. Yang Y, Diez Roux AV, Auchincloss AH, Rodriguez DA, Brown DG: A spatial
agent-based model for the simulation of adults daily walking within a
city. Am J Prev Med 2011, 40(30):353-361.
48. Auchincloss AH, Riolo RL, Brown DG, Cook J, Diez Roux AV: An agent-
based model of income inequalities in diet in the context of residential
segregation. Am J Prev Med 2011, 40(3):303-311.
49. Freeman LC: The Development of Social Network Analysis Vancouver:
Empirical Press; 2004.
50. Wasserman S: Social Network Analysis: Methods and Applications Cambridge:
Cambridge University Press; 1994.
51. Luke DA, Harris JK: Network analysis in public health: history, methods,
and applications.
Annu Rev Public Health 2007, 28(1):69-93.
52. Carrington PJ: Models and Methods in Social Network Analysis Cambridge:
Cambridge University Press; 2005.
53. Wang Y, Beydoun MA: The obesity epidemic in the United States
gender, age, socioeconomic, racial/ethnic, and geographic
characteristics: a systematic review and meta-regression analysis.
Epidemiol Rev 2007, 29(1):6-28.
54. McPherson M, Smith-Lovin L, Cook JM: Birds of a feather: Homophily in
social networks. Annu Rev Sociol 2001, 27(1):415-444.
55. Cobb S: Presidential Address-1976. Social support as a moderator of life
stress. Psychosom Med 1976, 38(5):300-314.
56. Berkman LF: Assessing the physical health effects of social networks and
social support. Annu Rev Public Health 1984, 5(1):413-432.
57. Kim D, Subramanian SV, Kawachi I: Bonding versus bridging social capital
and their associations with self rated health: a multilevel analysis of 40
US communities. J Epidemiol Community Health 2006, 60(2):116-122.
58. Subramanian SV: Social trust and self-rated health in US communities: a
multilevel analysis. J Urban Health 2002, 79(9):s21-s34.
59. Anagnostopoulos A, Kumar R, Mahdian M: Influence and correlation in
social networks. Proceedings of the 14th ACM SIGKDD International
Conference on Knowledge discovery and data mining 2008.
60. Cohen-Cole E, Fletcher JM: Is obesity contagious? Social networks vs.
environmental factors in the obesity epidemic. J Health Econ 2008,
61. Cohen-Cole E, Fletcher JM: Detecting implausible social network effects
in acne, height, and headaches: longitudinal analysis. BMJ 2008, 337(2):
62. Bonabeau E: Agent-based modeling: Methods and techniques for
simulating human systems. Proc Natl Acad Sci USA 2002,
63. Bankes SC: Agent-based modeling: A revolution? Proc Natl Acad Sci USA
2002, 99(90003):7199-7200.
64. Hommes CH: Modeling the stylized facts in finance through simple
nonlinear adaptive systems. Proc Natl Acad Sci USA 2002,
65. LeBaron B: Short-memory traders and their impact on group learning in
financial markets. Proc Natl Acad Sci USA 2002, 99(90003):7201-7206.
66. Macy MW, Willer R: From Factors to Actors: Computational Sociology and
Agent-Based Modeling. Annu Rev Sociol 2002, 28(1):143-166.
67. Cederman LE: Endogenizing geopolitical boundaries with agent-based
modeling. Proc Natl Acad Sci USA 2002, 99(90003):7296-7303.
68. Lustick IS, Miodownik D, Eidelson RJ: Secessionism in multicultural states:
Does sharing power prevent or encourage it? Amer Polit Sci Rev 2004,
69. Shafer G: The situation of causality. Found Sci 1995, 1(4):543-563.
70. Gilli M: A global optimization heuristic for estimating agent based
models. Comput Stat Data An 2003, 42(3):299-312.
Cite this article as: El- Sayed et al.: Social network analysis and agent-
based modeling in social epidemiology. Epidemiologic Perspectives &
Innovations 2012 9:1.
Submit your next manuscript to BioMed Central
and take full advantage of:
Convenient online submission
Thorough peer review
No space constraints or color figure charges
Immediate publication on acceptance
Inclusion in PubMed, CAS, Scopus and Google Scholar
Research which is freely available for redistribution
Submit your manuscript at
El-Sayed et al. Epidemiologic Perspectives & Innovations 2012, 9:1
Page 9 of 9
Page 9
  • Source
    • "Computational modelling and simulation can be used at various stages in a research endeavor and for multiple purposes. Basic research can be conducted, such as testing counterfactual simulations and exploring conceptual models [21,35]. Simulation experiments can also aid in identifying critical high-leverage points and optimal interventions or policy decisions by running multiple scenarios and examining outcomes [36]. "
    Preview · Article · Jan 2016
    • "Probabilistic computational models have been to used to analyze complex phenomena in areas that include the study of global ecology, forced migration, the spread of infectious diseases and threats to international security arising from local and regional conflicts [48,828384. While the use of mathematical models for computer simulations in systems biology is not new, recent trends show a marked qualitative change in the nature of the models, with the explosive growth in the use of agentbased models because of their natural ability to represent multi-scale biological systems. "
    [Show abstract] [Hide abstract] ABSTRACT: Background: Probabilistic models have gained widespread acceptance in the systems biology community as a useful way to represent complex biological systems. Such models are developed using existing knowledge of the structure and dynamics of the system, experimental observations, and inferences drawn from statistical analysis of empirical data. A key bottleneck in building such models is that some system variables cannot be measured experimentally. These variables are incorporated into the model as numerical parameters. Determining values of these parameters that justify existing experiments and provide reliable predictions when model simulations are performed is a key research problem. Domain experts usually estimate the values of these parameters by fitting the model to experimental data. Model fitting is usually expressed as an optimization problem that requires minimizing a cost-function which measures some notion of distance between the model and the data. This optimization problem is often solved by combining local and global search methods that tend to perform well for the specific application domain. When some prior information about parameters is available, methods such as Bayesian inference are commonly used for parameter learning. Choosing the appropriate parameter search technique requires detailed domain knowledge and insight into the underlying system. Results: Using an agent-based model of the dynamics of acute inflammation, we demonstrate a novel parameter estimation algorithm by discovering the amount and schedule of doses of bacterial lipopolysaccharide that guarantee a set of observed clinical outcomes with high probability. We synthesized values of twenty-eight unknown parameters such that the parameterized model instantiated with these parameter values satisfies four specifications describing the dynamic behavior of the model. Conclusions: We have developed a new algorithmic technique for discovering parameters in complex stochastic models of biological systems given behavioral specifications written in a formal mathematical logic. Our algorithm uses Bayesian model checking, sequential hypothesis testing, and stochastic optimization to automatically synthesize parameters of probabilistic biological models.
    No preview · Article · Dec 2015 · BMC Bioinformatics
    • "However, narrowly focused scientific methods are not well-suited to identifying potential unintended consequences (Levins, 2010). Our approach used an agent-based approach in order to explicitly incorporate complex interactions between simulated individuals in a counterfactual framework (El-Sayed et al., 2012). Of course, inference from agent-based simulations is not assumption-free; simulations merely move the realm of assumption from the data-generating process to the validity of model mechanism and parameterization. "
    [Show abstract] [Hide abstract] ABSTRACT: Background: Obesity and depression are comorbid more often than chance predicts. However, depression among the obese is more common in settings where obesity is less common. This suggests that body habitus norms and social stigmatization may play a role in the etiology of depression among the obese. Methods: We developed an agent-based social network model to explore mechanisms by which deviance from normative body habitus may contribute to social isolation in the obese. At each of 240 simulated months (20 years), each agent updated its body mass index based on environmental, peer influence, and stochastic factors. At each month, each agent was subject to social ostracization and consequent depression if its body mass index deviated from that of its peers and the network-wide mean. We compared risk of depression as a function of obesity and obesity norms through simulations of a high-obesity context simulating the US state of Mississippi and a low-obesity context simulating the US state of Colorado, then explored the relationship between global obesogenic forces and agent-specific resistance to the forces. Results: Over 1000 simulations in each context, 25 percent of obese agents in simulated Colorado were ever-depressed as compared to 21 percent in simulated Mississippi, although 10 percent overall were ever-depressed in both settings. High and low levels of resistance to obesogeneity prevented the most depression, whereas medium resistance levels were more depressogenic. Conclusions: Social stigma and ostracization that occur as a consequence of deviance from body habitus norms may be a plausible mechanism by which weight stigma may influence depression in the obese. Public health interventions targeting individuals rather than obesogenic environments may modify body habitus norms with the unintended consequence of increasing stigma-based social isolation among those who remain obese.
    No preview · Article · Nov 2015 · Social Science [?] Medicine
Show more