articles
nature publishing group
Epidemiology
Are Obesity and Physical Activity Clustered?
A Spatial Analysis Linked to Residential Density
Nadine Schuurman1, Paul A. Peters2 and Lisa N. Oliver2
The aim of this study was to examine spatial clustering of obesity and/or moderate physical activity and their
relationship to a neighborhood’s built environment. Data on levels of obesity and moderate physical activity were
derived from the results of a telephone survey conducted in 2006, with 1,863 survey respondents in the study sample.
This sample was spread across eight suburban neighborhoods in Metro Vancouver. These areas were selected to
contrast residential density and income and do not constitute a random sample, but within each area, respondents were
selected randomly. Obesity and moderate physical activity were mapped to determine levels of global and local spatial
autocorrelation within the neighborhoods. Clustering was measured using Moran’s I at the global level, Anselin’s Local
Moran’s I at the local level, and geographically weighted regression (GWR). The global-level spatial analysis reveals
no significant clustering for the attributes of obesity or moderate physical activity. Within individual neighborhoods,
there is moderate clustering of obesity and/or physical activity but these clusters do not achieve statistical significance.
In some neighborhoods, local clustering is restricted to a single pair of respondents with moderate physical activity. In
other neighborhoods, any moderate local clustering is offset by negative local spatial autocorrelation. Importantly, there
is no evidence of significant clustering for the attribute of obesity at either the global or local level of analysis. The GWR
analysis fails to improve significantly upon the global model—thus reinforcing the negative results. Overall, the study
indicates that the relationship between the urban environment and obesity is not direct.
Obesity (2009) doi:10.1038/oby.2009.119
Introduction
Obesity levels in Canada have risen over the past several
­decades and these increases are troubling because of their
associated comorbidities such as diabetes, cancer, and heart
disease (1). In an effort to understand this trend, researchers
are now focusing on the possible links between the physical
environment and rising rates of obesity (2). In Canada,
individuals who engage in walking and other forms of physical
activity have been found to be less likely to be overweight or
obese (1). Recent studies have broadened the scope of this
exploration to probe the associations among obesity, the
likelihood of engaging in physical activities, and the nature
of the built urban environment (3–9). The built environment
comprises all the elements constructed or modified by people—
ranging from buildings, to transportation systems, to public
spaces, and parks (10). These studies have shown that the built
environment can play both a positive and a negative role in
facilitating physical activities like walking (3,11–15).
Although walkability, urban form, density, and access to
green spaces (3–9) have been identified as components of the
built environment that may influence obesity and levels of physical activity, studies have revealed conflicting results about the
extent of this influence (13,16–19). This finding underscores
the complexity of these relationships. For example, although
some studies indicate a positive association between access to
green space and physical activity (9), the results of other studies
showed no significant relationship (17,18). A possible reason
for this inconsistency is that multiple variables must be considered. It becomes important to understand, for example, the
extent to which the physical environment actually influences
one’s choice to walk—and how this balances with the role of
other determinants of walking, like one’s ­ultimate destination (20). As this field of study matures, there is an increasing
need for precision in parsing these associations and it is, therefore, critical that we employ reliable objective instruments to
­measure and confirm the actual existence and extent of these
associations (21). Spatial methods of analysis are uniquely able
to contribute to this level of evidence.
Most studies examining the influence of the built environment
on physical activity use nonspatial methods such as multilevel
modeling (22) or regression (12,21,23). Studies using spatial
methods to examine relations between the built environment
and physical activity or obesity are rare. Such studies, however,
may be useful because they can identify clusters of individuals
1
Department of Geography, Simon Fraser University, Burnaby, British Columbia, Canada; 2Health Information and Research Division, Statistics Canada, Ottawa,
Ontario, Canada. Correspondence: Nadine Schuurman (nadine@sfu.ca)
Received 21 May 2008; accepted 21 March 2009; advance online publication 23 April 2009. doi:10.1038/oby.2009.119
obesity
1
articles
Epidemiology
exhibiting similar health behaviors or patterns, which may
be linked to the built environment. Traditionally, standard
regression has been used to reveal such patterns. The argument
presented in this paper asserts that spatial tests of local
clustering are less susceptible to measurement bias than their
aspatial global counterparts, as they account for the possible
spatial autocorrelation of observations. Aspatial regression
methods are generally predicated on the assumption of
spatial independence of observations. An exception to this is
the geographically weighted regression (GWR) method that
employs a locally weighted approach to regression analysis and
allows for parameter estimates to vary across the study space.
The Centers for Disease Control defines a cluster as “an
unusual aggregation, real or perceived, of health events that
are grouped together in time and space and that are reported
to a health agency” (24). Many health-related behaviors are
clustered with links to the built environment. Chaix et al., for
example, found a distinct spatial distribution for two classes
of mental disorder at the neighborhood and local level (25).
Clusters of health-related behaviors associated with weight gain
and physical activity have also been linked to factors within
the built environment. One study found youth distributed into
nonoverlapping activity/inactivity clusters that could be used
to predict overweight (26). In another study, obesity, diabetes,
and hypertension were found to be clustered based on census
tract variables (27). Mobley et al. investigated the presence of
spatial clustering in high BMI, low BMI, and smoking to find
significant differences in neighborhood characteristics linked
to smoking (28). They concluded that cluster analysis indicates
the value of spatial methods for policy application and evaluation purposes. These initial uses of spatial cluster analysis is for
determining patterns of correlation in the built environment
inform this study.
Spatial clustering techniques are commonly employed in
conjunction with geographic information systems to explore
whether patterns of distribution for a variable have significance.
The search for positive—or negative—spatial autocorrelation
among variables is based on Tobler’s first law of geography
which states that “everything is related to everything else, but
near things are more related than distant things” (29). Spatial
analysis allows researchers to determine whether observations
within a study area are random or exhibit a significant deviation
from a pattern that would likely arise from random underlying
factors. If a set of observations is not random, then it becomes
important to measure the degree and the nature of the spatial
distribution in order to make relevant hypotheses on its underlying cause. It is important to rely on quantitative metrics for
objective measures of spatial patterns, because even when data
are mapped, individuals often have difficulty recognizing or
identifying random distributions; rather they often “see” clusters when none truly exist (30). Spatial analysis instruments in
this study include Moran’s I, Anselin’s Local Moran’s I—a local
indicator of spatial autocorrelation (LISA)—and GWR.
The purpose of this study is to determine whether significant
patterns of clustering of obesity and/or moderate physical
activity can be revealed. This hypothesis is tested via spatial
cluster analysis within eight suburban neighborhoods.
Neighborhoods are selected for contrasting levels of income
and residential density and are not selected at random. Multiscale analysis is performed to allow identification of clusters
within and between neighborhoods. This method also allows
comparisons to be made regarding the frequency and density
of clustering within the two residential density categories. This
paper further examines the association between the degree of
clustering of obesity and physical activity relative to specific
combinations of built environment. The central hypothesis of
this paper is that the physical built environment has a direct
influence on clustering of obesity and physical activity.
Methods And Procedures
The neighborhoods
The study area (Figure 1) consists of eight neighborhoods with ­different
combinations of built environments, further qualified by income level.
Each neighborhood area comprises 3–4 census tracts with resulting
populations varying between 11,000 and 17,000. Residential density is
Locations of study neighborhoods
Newport
High income
High density
Maillardville
Low income
Low density
Sapperton
High income
High density
Fraser Heights
High income
Low density
Hammond
High income
Low density
Edmonds
Low income
High density
Greater Vancouver
Whalley
Low income
Low density
Langley
Low income
High density
Area of detail
Figure 1 The eight suburban neighborhoods used in this study and their location in the context of Metro Vancouver.
2
www.obesityjournal.org
articles
Epidemiology
defined as population per hectare of residential land, as designated by
the Greater Vancouver Land Use Data. A description of these data is
available elsewhere (21). Areas with the highest and lowest residential
densities are excluded from the study as densities are needed to reflect
a suburban profile. Of the qualifying areas, four neighborhoods represent a higher residential density while the other four represent a lower
residential density.
The study areas are also categorized by median family income, calculated using census data from the 2001 Census of Canada. Areas that fall
within the highest and lowest income deciles are excluded so as to select
typical neighborhoods rather than areas of extreme poverty of affluence.
Income categories for this study are classified as higher median family
income (CDN $53,000–77,000) and lower median family income (CDN
$32,000–44,000). Neighborhood selection avoided extremes of income
and density in favor of areas with relatively high automobile use and
largely middle class households. These categories are then combined
with the residential density categories to render four groupings, each
represented by two neighborhoods: lower density/higher income; higher
density/lower income; higher density/higher income; and higher density/
lower income.
Survey data
Data on individuals were collected through a telephone survey of
adults ≥aged 19, residing in the eight designated neighborhoods. The
­average length of residence for these respondents was 8.5 years. A
sampling frame of households in each neighborhood was constructed
from the regional telephone provider and numbers were selected by
Random Digit Dialing. A minimum baseline of five call-backs for each
selected number was used to reduce bias due to nonresponse. The
­survey was conducted by trained interviewers using Computer Assisted
Telephone Interviewing. The survey achieved a response rate of 29%
providing 1,935 respondents. For this study, we use data from the 1,863
­respondents with valid postal codes. An analytic sample of 1,789 is used
for weight and an analytic sample of 1,796 is used for moderate ­physical
activity, due to the exclusion of 74 and 67 respondents, respectively,
with missing variables. The survey was piloted in January 2006 and the
full survey was conducted over 2 weeks in February 2006.
Physical activity and obesity measures
Respondents were asked a series of five survey questions that probed
walking and physical activity (Table 1). Each survey question
offered a range of 5–6 possible responses. These responses were then
dichotomized by designating the two least active options as “nonactive”
and the remaining responses as “active.”
The Active responses were then tallied to form a broad index of
overall physical activity. An individual with a score of ≥3 on this index
(an Active response in at least three of the five questions) was designated
as ­moderately active. Moderate physical activity, therefore, was defined
in two possible combinations: a minimum of 3–3.75 h of physical activity
per week combined with >2 h of sedentary leisure per day or, a minimum
of 2–2.75 h of physical activity per week combined with <2 h of sedentary
leisure per day. The benchmark of 3–3.75 h was selected because it closely
reflects the 30 min per day of physical activity recommended by Canada’s
Physical Activity Guide (31). We believe that for individuals with <2 h of
sedentary leisure per day, 2–2.75 h of physical activity per week should
put them in a similar standing because their reduced hours of sedentary
leisure likely imply the presence of some degree of other moderate physical
activity unaccounted for within the framework of the questions asked.
The classification of respondents into the category of obese was done
using self-reports of height and weight. With these data, standard BMI
was calculated and respondents were classified so that an individual with
a BMI ≥30 was classified as obese.
Spatial analysis
The locations of survey respondents are geocoded from their postal
codes provided in the survey. The 2005 Unique Enhanced Postal Code
product from Desktop Mapping Technologies Inc.’s CanMap Postal
obesity
Geography v2005.3 (Markham, Ontario, Canada) is used to determine
precise geographic coordinates corresponding to each postal code (32).
The postal code information provides a means of locating each
respondent so that spatial correlation of particular attributes can be
assessed among all respondents within each suburban neighborhood.
In urban areas, postal codes cover ~15 households and as such provide
a relatively accurate estimate of an individual’s residential location (33).
In some cases respondents have the same postal code, likely due to
residing in the same apartment building, and thus are given the same
point location, resulting in the spatial duplication of some respondents.
These respondents are not excluded as the methods employed can
account for the increased local density of point locations.
The first spatial measure used, the Moran’s I, is a global measure
of ­spatial autocorrelation that quantifies the degree to which data are
clustered or uniformly distributed overall (34,35). Recent studies
have used it to evaluate the presence of spatial autocorrelation and
spatial clusters in a range of attributes such as hemorrhagic fever with
renal syndrome (36), neonatal characteristics and socioeconomic
conditions (37), influenza (38,39), pneumonia (38), human brucellosis
(40), and breast, lung, and colorectal cancer (41). Spatial proximity and
differences in attribute values are evaluated in concert for all pairs of
observations. Values for Moran’s I vary from +1 (strong positive spatial
autocorrelation with similar values located near one another) to −1
(strong negative spatial autocorrelation with high values located nearest to
low values and vice versa). A value near 0 implies a lack of spatial pattern
(i.e., random distribution). The null hypothesis of no spatial pattern can
be evaluated using a Z-score created from the mean and variance of I (42).
Although the value of the Moran’s I itself can be a useful comparison, care
must be taken in interpretation when a low Z-score is obtained.
In this study, the Moran’s I test statistic is used to test for global
spatial autocorrelation of physical activity and obesity across eight
neighborhoods. Inverse distance squared weighting is used to measure
the spatial relationships among respondents for the two attributes,
obesity and moderate physical activity. This measure determines the
relative influence that should be given to all other respondents when
local spatial autocorrelation is calculated for one particular respondent.
With inverse distance squared weighting, the influence of each location
on the overall calculation decreases exponentially with distance. That
is, when calculating the local spatial autocorrelation for a particular
Table 1 The dichotomization of survey responses to
questions regarding physical activity and sedentary leisure
into “active” and “nonactive” responses
Active
Nonactive
Q18: “In a typical week in the past 3 months
how many hours did you usually spend
walking to and from work or school?”
Question
>1 h
<1 h
Q19: “In a typical week in the past few
months how many hours did you spend
walking from home to grocery stores banks
or to do other errands?”
>1 h
<1 h
Q21: “In a typical week in the past 3 months
how many days did you do at least 30
minutes of moderate physical activity such
as brisk walking running swimming or team
sports?”
≥2 days
≤1 day
Q22: “On a typical day in the past 3 months
how much time did you spend walking for
leisure?”
≥16 min
≤15 min
Q23: “On a typical day in the past 3 months
how much time did you spend sitting while
doing leisure activities such as: reading,
watching T.V., on the computer, or playing
video games?”
<2 h
>2 h
3
articles
Epidemiology
respondent, the greatest weight is given to the most proximate
individuals. The Moran’s I test generates an index value and associated
Z-score for each variable for each neighborhood. The Z-scores are used to
identify the attributes that exhibited a significant degree of global spatial
autocorrelation within a particular neighborhood.
Individual clusters or smaller regions of clustering that are not
­evident within the global pattern may appear during analysis at the
local level (30). To evaluate the existence of local clusters, it is necessary to use local statistics. The local Moran’s I decomposes the global
measure into the contributions for each location, detecting similarities
or dissimilarities in values around a given observation and thus measuring the local spatial autocorrelation at each observation point (43).
These statistics are also known as Local Indicators of Spatial Association
(LISAs), which can be indicators of local spatial clustering or used as
diagnostics for outliers in global patterns. The sum of local Moran’s I
values is proportional to global Moran’s I.
Anselin’s Local Moran statistic is subsequently used to test for local
spatial autocorrelation among respondents within each neighborhood
for each of the same two variables. Inverse distance weighting squared is
again used to conceptualize the spatial relationships. Respondents with
missing data for a particular variable are excluded from any further
analysis of that variable.
An index value and Z-score are calculated for each respondent in each
data layer. The Z-scores are used to identify respondents that exhibit
a significant degree of local spatial autocorrelation for a particular
variable within a particular neighborhood. Results of these analyses
are mapped, highlighting respondents showing instances of significant
positive and/or significant negative local spatial autocorrelation. If
a perfectly homogenous cluster exists in the underlying population,
it is expected that a grouping of respondents with significant positive
local spatial autocorrelation would be visually evident. As is common
practice for local analysis of spatial autocorrelation, maps are visually
inspected for the presence of areas with a concentration of respondents
with significant positive local spatial autocorrelation within close
proximity of each other. For the purposes of this study, clear definitions
of what constitutes evidence of local clustering are developed. First,
groupings of four or five significant positive results in close proximity
are considered to be moderate ­evidence suggesting the possibility of
underlying clustering in the population. ­Second, groupings of six or
more significant positive results in close proximity are considered to be
stronger evidence supporting the existence of such underlying clustering.
Inversely, evidence of significant negative local spatial autocorrelation
amongst any of these groups is considered to weaken the evidence of the
existence of an underlying cluster.
The third method used to test the hypothesis for a relationship
between obesity, physical activity, and the built environment is GWR.
The essence of GWR is that a traditional regression framework is respecified using a locally weighted subsample of the population, to allow
for the estimation of local parameters for each observation as opposed
to a single parameter set for all observations. The power of the GWR
approach is that for each observation point i, it is possible to produce a
complete set of parameter estimates, local standard errors, measures of
significance, influence statistics, and importantly, local r2 values. Additionally, it is possible to specify different methods for the local area
sample, allowing for hypotheses on the spatial level of interaction to be
tested. The analysis in this paper employs a binary logistic model with
obesity as the response variable and moderate physical activity as an
independent variable. Dummy variables for neighborhood density and
median income are included.
Results
Table 2 summarizes the findings for obesity, overweight, and
moderate physical activity rates for the eight neighborhoods.
Table 3 identifies levels of physical activity over 3 h a week for
the same neighborhoods.
The global Moran’s I test statistic is evaluated for both
­obesity and moderate physical activity within each neighborhood (Table 3). Global dispersion means that individuals who
are obese/nonobese (physically activity/not physically active)
are not found in close proximity but distributed regularly
throughout all the study areas. None of the neighborhoods
exhibit statistically significant evidence of global clustering.
This suggests that the built urban environment has little or no
effect on the likelihood of finding this variable. None of the
remaining variables exhibit a statistically significant degree of
global clustering.
Anselin’s Local Moran’s I is used to evaluate the existence
of local clustering not evident within the global neighborhood ­patterns. Any region with multiple cases of significant
local spatial autocorrelation provides evidence in support of
the existence of an underlying cluster in the population. The
mapped results of these analyses of local spatial autocorrelation
in obesity are presented in Figure 2. The results for ­physical
activity are not shown here for space considerations. Cases of
both significant positive and significant negative local ­spatial
autocorrelation are indicated.
Table 2 Survey respondents by neighborhood, summarizing
their respective rates of overweight, obesity, and a moderate
level of activity
Respondents
Whalley
233
Overweight
(%)
Obese
(%)
Moderately
active (%)
53.6
17.1
42.9
Edmonds
229
48.1
18.1
56.8
Maillardville
247
46.6
14.0
47.5
Langley
231
55.9
13.5
53.6
Sapperton
230
53.6
19.8
46.4
Hammond
238
61.6
18.1
51.1
Newport
252
42.3
14.2
53.5
Fraser
Heights
203
41.5
13.5
45.9
1,863
50.5
16.0
49.8
Total
Table 3 Global Moran’s I test results for obesity and moderate physical activity by neighborhood
Edmonds
Fraser
Heights
Hammond
Langley
Maillardville
Newport
Sapperton
Whalley
−0.20
−0.01
0.01
0.00
−0.01
0.01
−0.01
0.00
Obesity
(BMI >30)
Moran’s I
Z-score
−1.29
−0.71
1.07
0.15
−0.72
0.10
−0.45
0.01
Physical activity
(activity index 3+)
Moran’s I
−0.02
−0.01
−0.01
−0.02
−0.01
−0.02
0.00
0.00
Z-score
−1.12
−0.75
−0.10
−1.12
−0.63
−1.07
0.52
0.01
4
www.obesityjournal.org
articles
Epidemiology
High density neighborhoods
Low density neighborhoods
Newport
Hammond
Fraser Heights
High income
Sapperton
Langley
Whalley
Maillardville
Low income
Edmonds
Local Moran’s I
<−2.0
−2.0 to −1.0
−1.0 to 1.0
1.0 to 2.0
>2.0
Figure 2 Moderate evidence of local clustering of obesity. There are several cases of strong positive spatial autocorrelation located within close proximity
in several neighborhoods. However, these cases still do not meet the criteria of six cases required to conclude that there is significant local clustering.
In Figure 2, there is moderate evidence of local clustering of
obesity (several cases of strong positive spatial autocorrelation
located within close proximity) within central and southwestern Sapperton. There is stronger evidence of local clustering
in south-central Langley, Fraser Heights, and southeastern
Hammond. The only neighborhood with potentially significant local clustering is in southern Langley, where there is a
cluster of six higher than expected values. All neighborhoods
contain respondents showing significant local spatial autocorrelation but widely separated individuals or pairs of such cases
(e.g., Edmonds, Whalley) do not meet the criteria required to
conclude there is significant clustering within the population.
In other words, sporadic local clustering of obesity does not
constitute significant clustering.
The LISA analysis of physical activity (not shown here)
demonstrates moderate evidence of local clustering of moderate
physical activity within Maillardville and strong evidence for
such clustering within Whalley. Similar to obesity, each of the
study neighborhoods contains cases of significant local positive
spatial autocorrelation but as these cases are dispersed, they
do not provide evidence of local clustering in the underlying
population. The Sapperton area provides an excellent example
of a neighborhood with many individual cases of significant
positive local spatial autocorrelation for respondents with
moderate physical activity. However, there is no coherent
evidence of any clustering in this attribute within the population
given the dispersion of individual cases across the neighborhood.
Neither obesity nor physical activities are affected by higher vs.
lower density or different income levels. This is an unexpected
result that begs the question of whether or not there are clear
links between the built environment and obesity and/or ­physical
activity. It certainly invites further investigation.
obesity
The results of the GWR analysis (Figure 3) do not provide
significantly different results than those found by either the
global-level Moran’s I statistic or the LISA analysis. The local
regression model is highly simplified and developed to test the
relationship between obesity and physical activity within each
of the study neighborhoods. If local variation was apparent
there would be visual differences in both the local r2 values
and in the parameter estimates. As is apparent from Figure 3,
while there are differences between the local r2 values and the
parameter estimates of activity, they are very small and do not
correspond to any statistical significance via Monte Carlo or
Leung significance tests. Thus, from the testing done here, there
is no significant difference between these values, suggesting
that the relationship between obesity and physical activity was
the same for all of the neighborhoods.
Additionally, the GWR model is used to calculate the Aikaike
information criterion for both the global and local models. The
global model has a corrected Aikaike information criterion of
1,509 with the local model reporting a corrected Aikaike information criterion of 1,511. This suggests that not only is there
no significant difference at the local level, the local model is a
slightly poorer fit than at the global level. This confirms our
results: spatial clustering of obesity and physical activity is not
directly influenced by the urban built environment.
Discussion
There is a growing body of literature exploring the influence of
the built environment on both obesity and physical activity. This
literature is mixed in its conclusions with respect to the effect of
the environment on obesity. Whilst studies generally agree that
walkability influences levels of physical activity (7,12,23), there
are a number of other known risk factors including childhood
5
articles
Epidemiology
Local r 2
0.015 − 0.017
0.018
0.019 − 0.021
Activity parameter
−0.261 to −0.253
−0.252 to −0.220
−0.219 to −0.203 0
2.5
5
km
10
N
Figure 3 Schematic representation illustrates that we do not have
significantly different results, using geographically weighted regression,
than those found by either the global-level Moran’s I statistic or the
LISA analysis. None of the local areas indicates a significantly different
relationship between the dependent and independent variables.
obesity, educational status, and related health risks (44–47).
In this study, the assumption is that if the built environment
affects behavior and health, physical activity and obesity will be
clustered. If there is indeed a positive association between the
physical context of one’s life and one’s health-related behaviors
and outcomes, this association may reveal itself in obesity and
physical activity clusters within a neighborhood. The built
environment in any neighborhood determines the ease and
access to commercial and retail amenities, the numbers of and
proximity to parks and recreational facilities, and the extent
to which an area is pedestrian friendly. Physical activity and
obesity, therefore, may show distinct patterns relative to the
underlying built environment, particularly at the local level.
Utilizing the Moran’s I, the Anselin’s Local Moran’s I, and
GWR instruments, this study maps data on obesity and moderate physical activity across eight suburban neighborhoods.
6
A number of neighborhoods are selected in order to provide
diversity between densities of built environment assuming that
different blends of neighborhood amenities and recreational
areas might have revealed different patterns. The neighborhoods, therefore, represent two levels of residential density
and two levels of income in four different combinations: high
density/high income; high density/low income; low density/
high income; and low density/low income. Data are gathered
from adults residing within those neighborhoods through a
telephone survey.
Geographic information systems and spatial mapping techniques are uniquely suited to identifying and analyzing these
physical patterns of distribution. They provide the means to
effectively and efficiently ascertain whether patterns of significance do indeed exist. Once patterns like clustering are
geographically located and verified, further investigation into
ground-level influences can proceed. There is, for instance,
some evidence of moderate local clustering of physical activity
in Whalley. This is an area where the local light rapid transit
system, the SkyTrain, converges with local parks and commercial areas. Certainly moderate local clusters could provide the
basis for more intensive qualitative examinations of possible
correlation between the built urban environment and physical
activity and/or obesity.
The global analysis of possible clustering of obesity and/
or moderate physical activity across eight suburban neighborhoods yielded only one result of statistical significance.
In the Edmonds neighborhood, a high density/low income
area, there was evidence of uniform spatial distribution for
obesity, i.e., dispersion rather than any clustering. We do not
know of any underlying mechanism that would explain this
result, nor have we encountered anything in the literature
to suggest that this might ever occur. We believe that it is
a random result or a product of random factors. Second,
for our purposes of identifying potential clustering it consolidates our conclusion that there is no global clustering in
Edmonds. The remaining neighborhoods showed no significant patterns of positive or negative distribution of clusters
for moderate physical activity. None of the neighborhoods
revealed any significant clustering for obesity. At the local
level, on the other hand, some of the study areas did reveal
patterns suggesting moderate to strong evidence for the
existence of underlying areas of local clustering. This may
suggest underlying spatial nonstationarity, where the structure of spatial dependency for obesity is heterogeneous. In
short, our findings at the global level did not reveal a close
association between obesity and/or moderate physical activity and the built urban environment.
Although many of the studies examining the phenomenon
of clustering employ spatial analysis at the global level only,
this study took a multi-scale approach and “zoomed in” to
­examine patterns of distribution at the local level as well. This
more granular level of analysis allows patterns of local ­spatial
distribution to be revealed that may otherwise have been
masked or obscured by the presence or absence of the more
remote global patterns, as was the case with these data. In this
www.obesityjournal.org
articles
Epidemiology
case, lack of clustering at the global level indicates both lack of
a clear causal relationship and possible obfuscation of a scaledependent pattern.
Although the overall results of this study indicate that
­obesity and physical activity are not closely linked to the built
urban environment, some of the study areas did reveal ­patterns
­suggesting moderate to strong evidence for the existence of
underlying local clustering. Local analysis does, therefore,
offer the basis for a closer examination of possible links that
are only discernable at a large scale (small area) of analysis.
This study did have certain limitations. Although eight areas
were examined, with different density and income profiles, a
greater number of area samples as well as a more granular
level of distinction between each composite area’s factors may
have rendered more detailed information about patterns relative to variables of income, access to retail, greenways, etc.
The nature of the neighborhoods themselves, being suburban,
may have contributed to the low level of findings. However,
our study considered a diverse mix of land-uses as well as
many variations of green space within walking ­distance of
individuals (21).
Deriving data from self-reports of physical activity and
height and weight is often subject to bias. Women and men
tend to under-report weight and over-report height, making
BMI measurements derived from surveys subject to a potential
skew downward (48). Accuracy in recall for levels of physical
activity is less likely to be an issue, as these are current daily
or weekly activities, although demand characteristic bias may
have influenced some respondents to adjust their responses
in the hopes of providing what they felt might be the “right”
answer. The survey was conducted in February, which is typically rainy and cold; results may have differed if the survey was
conducted during a warmer month.
A further limitation to this approach exists with regard to
the attribute of obesity. Although patterns of significance were
not found in this study, pattern analysis would have to take
into consideration the onset period inherent in weight gain
and attaining the BMI values associated with obesity. Should
an influence exist between the built environment and obesity,
it would need to be ascertained whether the condition of
­obesity was primarily influenced by the current neighborhood
or by previous physical contexts. Moreover, the study design
does not permit analysis of the possible influence of social
networks on obesity and levels of physical activity (49).
Social networking is becoming ever more ubiquitous with the
exponential growth of the web. Web 2.0 broadly refers to a
new generation of internet services and technology (50–52).
Web 2.0 heralds a transition from local pockets of isolation to
a global interconnectedness. Should social network become
the pervasive form of interpersonal communication, this
may have important health considerations (53). Certainly
social networking is an emerging model that may partially
explain patterns of obesity not directly related to the urban
environment. Future studies may employ more complex
questioning about social networks to determine possible
influence. Social networks present an example of nonlinear,
obesity
nonspatial, and easily obfuscated influences on incidence of
obesity and physical activity.
Geographic information systems and spatial analysis provide
an important means of broadening the exploration of the
relationship between the built environment and personal levels
of physical activity, interaction with the physical environment,
and health outcomes. They provide a level of evidence much
needed by researchers seeking to determine whether or not
associations between the physical environment and behaviors
are indeed present. Using spatial analysis, this study was
designed to detect the presence or absence of evidence
supporting spatial clustering of the attributes of obesity and
moderate physical activity within eight study neighborhoods.
Although no patterns of significance for either obesity or
moderate physical activity were determined at the global level,
by employing a multi-scale approach pockets of local clustering
of moderate physical activity were identified within a variety
of neighborhoods—pockets that had not been apparent at the
global level. The evidence of physical activity clustering at the
local level, although inconsistent, may yet indicate that there
could be specific area features within a neighborhood that may
warrant further investigation. There were, however, no similar
local findings for the attribute of obesity. Although this does not
rule out the possibility of association between ­obesity and the
build environment, it does indicate that the relationship is more
spatially complex than has been ­previously acknowledged.
Acknowledgments
We thank the Canadian Institutes of Health Research (CIHR) for their funding
through grant #116338, as well as support from the Canadian Institute
for Health Information and the Canadian Population Health Initiative. We
also thank Alex Hall for his substantial technical assistance with the initial
analysis. Ellen Randall thoroughly read two drafts of this paper and offered
valuable editing suggestions. In addition, N.S. acknowledges the Michael
Smith Foundation for Health Research and CIHR for their continued
research support.
Disclosure
The authors declared no conflict of interest.
© 2009 The Obesity Society
References
1. Tjepkema M. Adult obesity. Health Rep 2006;17:9–25.
2. Hill JO, Peters JC. Environmental contributions to the obesity epidemic.
Science 1998;280:1371–1374.
3. Frank LD, Schmid TL, Sallis JF, Chapman J, Saelens BE. Linking objectively
measured physical activity with objectively measured urban form: findings
from SMARTRAQ. Am J Prev Med 2005;28(Suppl 2):117–125.
4. Li F, Fisher KJ, Brownson RC, Bosworth M. Multilevel modelling of built
environment characteristics related to neighbourhood walking activity in
older adults. J Epidemiol Community Health 2005;59:558–564.
5. Boer R, Zheng Y, Overton A, Ridgeway GK, Cohen DA. Neighborhood
design and walking trips in ten U.S. metropolitan areas. Am J Prev Med
2007;32:298–304.
6. Berke EM, Koepsell TD, Moudon AV, Hoskins RE, Larson EB. Association
of the built environment with physical activity and obesity in older persons.
Am J Public Health 2007;97:486–492.
7. Michael Y, Beard T, Choi D, Farquhar S, Carlson N. Measuring the influence
of built neighborhood environments on walking in older adults. J Aging Phys
Act 2006;14:302–312.
8. Craig CL, Brownson RC, Cragg SE, Dunn AL. Exploring the effect of the
environment on physical activity: a study examining walking to work. Am J
Prev Med 2002;23(2 Suppl):36–43.
7
articles
Epidemiology
9. Neuvonen M, Sievänen T, Tönnes S, Koskela T. Access to green areas and
the frequency of visits—a case study in Helsinki. Urban For Urban Green
2007;6:235–247.
10. National Institute of Environmental Health Sciences—National Institutes of
Health. Obesity & the Built Environment <http://www.niehs.nih.gov/news/
events/pastmtg/2004/built/> (2004).
11. Gordon-Larsen P, Nelson MC, Page P, Popkin BM. Inequality in the built
environment underlies key health disparities in physical activity and obesity.
Pediatrics 2006;117:417–424.
12. Frank LD, Andresen MA, Schmid TL. Obesity relationships with community
design, physical activity, and time spent in cars. Am J Prev Med
2004;27:87–96.
13. Wendel-Vos W, Droomers M, Kremers S, Brug J, van Lenthe F. Potential
environmental determinants of physical activity in adults: a systematic
review. Obes Rev 2007;8:425–440.
14. Lake A, Townshend T. Obesogenic environments: exploring the built and
food environments. J R Soc Promot Health 2006;126:262–267.
15. Lopez RP, Hynes HP. Obesity, physical activity, and the urban environment:
public health research needs. Environ Health 2006;5:25.
16. Forsyth A, Oakes J, Schmitz K, Hearst M. Does residential density increase
walking and other physical activity? Urban Stud 2007;44:679–697.
17. McCormack GR, Giles-Corti B, Bulsara M. The relationship between
destination proximity, destination mix and physical activity behaviors.
Prev Med 2008;46:33–40.
18. Hillsdon M, Panter J, Foster C, Jones A. The relationship between access
and quality of urban green space with population physical activity. Public
Health 2006;120:1127–1132.
19. Handy SL, Boarnet MG, Ewing R, Killingsworth RE. How the built
environment affects physical activity: views from urban planning. Am J Prev
Med 2002;23:64–73.
20. Owen N, Humpel N, Leslie E, Bauman A, Sallis JF. Understanding
environmental influences on walking; Review and research agenda.
Am J Prev Med 2004;27:67–76.
21. Oliver LN, Schuurman N, Hall AW. Comparing circular and network buffers
to examine the influence of land use on walking for leisure and errands.
Int J Health Geogr 2007;6:41.
22. Ross NA, Tremblay S, Khan S et al. Body mass index in urban Canada:
neighborhood and metropolitan area effects. Am J Public Health
2007;97:500–508.
23. Frank LD, Engelke P. The built environment and human activity patterns:
exploring the impacts of urban form on public health. J Plann Lit
2001;16:202–218.
24. Guidelines for investigating clusters of health events. MMWR Recomm Rep
1990;39:1–23.
25. Chaix B, Leyland AH, Sabel CE et al. Spatial clustering of mental disorders
and associated characteristics of the neighbourhood context in Malmö,
Sweden, in 2001. J Epidemiol Community Health 2006;60:427–435.
26. Monda KL, Popkin BM. Cluster analysis methods help to clarify the activityBMI relationship of Chinese youth. Obes Res 2005;13:1042–1051.
27. Schlundt DG, Hargreaves MK, McClellan L. Geographic clustering of obesity,
diabetes, and hypertension in Nashville, Tennessee. J Ambul Care Manage
2006;29:125–132.
28. Mobley LR, Finkelstein EA, Khavjou OA, Will JC. Spatial analysis of body
mass index and smoking behavior among WISEWOMAN participants.
J Women’s Health (Larchmt) 2004;13:519–528.
8
29. Tobler WR. A computer model simulation of urban growth in the Detroit
region. Econ Geogr 1970;46:234–240.
30. Rogerson PA. Statistical methods for the detection of spatial clustering in
case-control data. Stat Med 2006;25:811–823.
31. Public Health Agency of Canada (PHAC). Canada’s Physical Activity Guide
to Healthy Living for Older Adults. Public Health Agency of Canada: Ottawa,
Canada, 1998.
32. DMTI Spatial Inc. Unique Enhanced Postal Code. Version 2005.3. CanMap
Postal Geography 2005.
33. Statistics Canada. 2001 Census Dictionary. Report no.: 92-378-XIE (2003).
34. Moran PA. The interpretation of statistical maps. J R Stat Soc Series B Stat
Methodol 1948;10:243–251.
35. Moran PA. Notes on continuous stochastic phenomena. Biometrika
1950;37:17–23.
36. Fang L, Yan L, Liang S et al. Spatial analysis of hemorrhagic fever with renal
syndrome in China. BMC Infect Dis 2006;6:77.
37. d’Orsi E, Carvalho MS, Cruz OG. Similarity between neonatal profile
and socioeconomic index: a spatial approach. Cad Saude Publica
2005;21:786–794.
38. Crighton EJ, Elliott SJ, Moineddin R, Kanaroglou P, Upshur RE.
An exploratory spatial analysis of pneumonia and influenza hospitalizations
in Ontario by age and gender. Epidemiol Infect 2007;135:253–261.
39. Greene SK, Ionides EL, Wilson ML. Patterns of influenza-associated
mortality among US elderly by geographic region and virus subtype,
1968–1998. Am J Epidemiol 2006;163:316–326.
40. Fosgate GT, Carpenter TE, Chomel BB et al. Time-space clustering
of human brucellosis, California, 1973–1992. Emerging Infect Dis
2002;8:672–678.
41. Jacquez GM, Greiling DA. Local clustering in breast, lung and colorectal
cancer in Long Island, New York. Int J Health Geogr 2003;2: 3.
42. Cliff AD, Ord JK. Spatial Processes: Models and Applications. Pion:
London, 1981.
43. Anselin L. Local indicators of spatial association—LISA. Geogr Anal
1995;27:93–115.
44. Martínez-González MA, Martínez JA, Hu FB, Gibney MJ, Kearney J.
Physical inactivity, sedentary lifestyle and obesity in the European Union.
Int J Obes Relat Metab Disord 1999;23:1192–1201.
45. Mozaffari H, Nabaei B. Obesity and related risk factors. Indian J Pediatr
2007;74:265–267.
46. Reilly JJ, Kelly L, Montgomery C et al. Physical activity to prevent
obesity in young children: cluster randomised controlled trial. BMJ
2006;333:1041.
47. Robinson TN, Sirard JR. Preventing childhood obesity: a solution-oriented
research paradigm. Am J Prev Med 2005;28(2 Suppl 2):194–201.
48. Gorber SC, Tremblay M, Moher D, Gorber B. A comparison of direct vs.
self-report measures for assessing height, weight and body mass index: a
systematic review. Obes Rev 2007;8:307–326.
49. Christakis NA, Fowler JH. The spread of obesity in a large social network
over 32 years. N Engl J Med 2007;357:370–379.
50. Solomon G, Schrum L. Web 2.0: New Tools, New Schools. ISTE:
Washington, DC, 2007.
51. Greaves M, Mika P. Semantic Web and Web 2.0. Web Semantics: Science,
Services and Agents on the World Wide Web 2008;6:1–3.
52. Deshpande A, Jadad AR. Web 2.0: could it help move the health system
into the 21st century? J Mens Health Gend 2006;3:332–336.
www.obesityjournal.org