Journal of Transport Geography 62 (2017) 136147

Contents lists available at ScienceDirect

Journal of Transport Geography

journal homepage: www.elsevier.com/locate/jtrangeo

A geographically weighted regression approach to investigating the spatially MARK

varied built-environment eects on community opportunity
Chih-Hao Wanga,, Na Chenb
a
Department of Geography and City and Regional Planning, California State University, Fresno, Fresno, CA 93740, USA
b
City and Regional Planning, The Ohio State University, Columbus, OH 43210, USA

A R T I C L E I N F O A B S T R A C T

Keywords: Geographically weighted regression (GWR) has been increasingly used to better understand the spatially varied
Spatial statistics relationships between socioeconomic outcomes and policy investments. In this study, a community opportunity
GWR index (COI) is computed for Columbus, Ohio, using a set of socioeconomic factors. An ordinary regression and
Land-use and transportation planning GWR model are estimated to account for the global and local eects of land uses and capital investments re-
Community opportunity
spectively, while controlling for socio-demographical characteristics. The global results indicate that the com-
Social equity
munity opportunity increases as the distance from the city center increases, while The Ohio State University
(OSU) has higher positive spillover-eect on near communities than on distant ones. However, the local results
appear a local spatially inverse relationship in the areas adjacent to the international airport (CMH), indicating
the existence of negative externalities. With the advantage of visualizing spatial variations, the GWR results
suggest that the most eective location for allocating future developments is in eastern Columbus, where shows a
clustering of higher COI premiums of a percentage change in residential and commercial uses. A variety of
spatial variations is found among dierent capital-investment eects. Therefore, local characteristics require
consideration when allocating additional public facilities. Finally, the GWR results reveal the existence of spatial
mismatch that socially disadvantaged groups (e.g. black population, other minorities, single-parent families, and
zero-vehicle households) tend to reside in vulnerable communities. These local results provide a new perspective
on land-use and transportation planning to help shape a fair community opportunity framework for the future.

1. Introduction variations of the built-environment eects (Du and Mulley, 2012;

Geographical mapping of community opportunity can help examine
social inequity, using a spatially aggregated index comprised of a set of
socioeconomic factors. This idea is based on the concept of neigh-
borhood eects that where people live inuences their socioeconomic
outcomes (Acevedo-Garcia et al., 2004; Jencks and Mayer, 1990). Using
GIS, a community opportunity index can be computed by combining a
set of un-weighted standardized neighborhood factors. The term
community opportunity in the present study refers to the physical and
socioeconomic outcomes of a residential community. This approach has
been applied for visualizing the spatial opportunity and deprivation in
many America's cities (Powell, 2007; Reece and Gambhir, 2008).
However, the geographical mapping approach does not explicate how
the community opportunity is inuenced by social and institutional
mechanisms. Particularly, the physical setting of land uses and capital
investments (i.e. the built environment) was often overlooked in the
past research (Sampson et al., 2002). In addition, it is necessary to
apply a spatial statistical approach to better understand the spatial
Mulley, 2014). Geographically weighted regression (GWR) has in-
creasingly attracted attention on studying such spatially varied re-
lationships over a geographical area (Brunsdon et al., 1996; Paez, 2006;
Tu, 2011). Through understanding the local built-environment eects,
land-use and transportation planning can help shape a fair community
development framework to remedy spatial disparities.
In this study, a community opportunity index (COI) was computed
and mapped for Columbus, Ohio. Census tracts were selected as the
geographic units. An ordinary regression model was estimated to ac-
count for the global built-environment eects, while controlling for
socio-demographical factors. A GWR model was also estimated to in-
vestigate the spatial variations of the built-environment eects. The
explanatory variables were characterized into four groups, including
metropolitan-location eects, land uses, capital investments, and socio-
demographical characteristics. This study will add to the existing lit-
erature by examining the local built-environment eects on community
opportunity. In the future, the spatially varied relationships can be used
to develop an optimization-modeling framework that would facilitate

Corresponding author at: Department of Geography and City and Regional Planning, California State University, Fresno, 2555 E San Ramon M/S SB69, Fresno, CA 93740, USA.
E-mail addresses: cwang@csufresno.edu (C.-H. Wang), chen.2572@osu.edu (N. Chen).

http://dx.doi.org/10.1016/j.jtrangeo.2017.05.011
Received 4 December 2016; Received in revised form 5 May 2017; Accepted 28 May 2017
Available online 14 June 2017
0966-6923/ 2017 Elsevier Ltd. All rights reserved.
C.-H. Wang, N. Chen
the allocation of future land uses and capital investments, either to
maximize the total COI scores, or to minimize COI dierences among
the 284 census tracts.
The remainder of the paper is organized as follows. Section 2 con-
sists in a literature review. The modeling methodology is presented in
Section 3. The data are described in Section 4. The modeling results and
their analysis are presented in Section 5. Section 6 discusses how GWR
can be used to provide local information for planners. Section 7 pre-
sents conclusions and outlines areas for future research.
2. Background
2.1. Community opportunity
The discussion of community opportunity is based on the concept of
neighborhood eects, which originates from a longstanding argument
in sociology that where an individual lives matters (Acevedo-Garcia
et al., 2004; Jencks and Mayer, 1990). Early applications of neighbor-
hood eects in urban planning studies focused on comparing spatial
variations among America's cities. For instance, Hill et al. (1998) de-
ned the term central city as the image of America's cities that large
municipalities are disproportionately poor and distressed, in terms of
their socioeconomic outcomes. In their study, central city refers to the
discussions on the functions of America's large cities, including the
functions done within these cities and the disproportionately con-
centrated outcomes, such as a central core with poor neighborhoods
surrounded by wealthier suburbs. Using cluster and discriminant ana-
lyses, they reported that there was wide variation among large cities in
the country.
Sampson et al. (2002) examined social-interactional and institu-
tional mechanisms to account for such spatial variation within a city. By
reviewing over 40 relevant studies, they concluded that there are four
valid neighborhood mechanisms, including social ties, collective e-
cacy, institutional resources, and routine activities. Social ties, driven
by social capital, refer to social relations and interactions between
neighbors (Sampson et al., 2002). The concept of social capital, a re-
source through social relationships, reects the connections between
people and organizations, including internal and external networks,
local mobilization of resources, and willingness to consider alternative
ways of reaching goals (Emery and Flora, 2006; Flora et al., 1997).
Collective ecacy can be seen as the mutual trust and shared will-
ingness to intervene for the public good (Sampson et al., 1997). An
individual is unlikely to intervene in a neighborhood context if the rules
are not clear and neighbors fear one another (Sampson et al., 1997).
Institutional resources refer to the quality, quantity, and diversity of
capital investments in the community, such as libraries, schools, and
hospitals (Sampson et al., 2002). Routine activities measure the types of
land use at the neighborhood level, such as commercial, industrial, and
residential units (Peterson et al., 2000; Sampson et al., 2002; Scribner
et al., 1998; Smith et al., 2000).
Additionally, the term neighborhood eects is a broad concept
with dierent meanings in disciplines other than urban planning. In
social epidemiology and health, for instance, a neighborhood eect is
dened as the independent causal eect of residential community on
any number of health and/or social outcomes (Arcaya et al., 2016;
Oakes, 2004; Perchoux et al., 2016). Many studies in health were based
on the assumption that individual health is inuenced by both in-
dividual characteristics and the contexts to which individuals belong
(Arcaya et al., 2016; Dundas et al., 2014; Kestens et al., 2016; Perchoux
et al., 2016). Among these studies, multilevel modeling appeared to be
a widely-used approach to address neighborhood eects at dierent
spatial scales on individual health (Arcaya et al., 2016; Ellaway et al.,
2012; Oakes, 2004; Oakes et al., 2015). However, Oakes et al. (2015)
reported inconsistences among the ndings of such health studies,
137
Journal of Transport Geography 62 (2017) 136147
particularly using data from experimental designs. More importantly,
they also asserted that there could be a bidirectional interaction be-
tween individuals and their neighborhood.
The focus of this review is on the modeling interface between
community opportunity and built environment. Readers interested in
social aspects of neighborhood eects might usefully consult Hill
et al. (1998), Sampson et al. (2002), Oakes (2004), and Oakes et al.
(2015). Given the previous discussion, a possible research framework
for this study, which examines how land-use allocation and capital
investments inuence community opportunity, is supported. The basic
assumption is that built environment would inuence individuals' be-
havior and therefore facilitate or deter their social and economic ac-
tivities (Gilliland et al., 2006). In addition, the physical setting of land
use and public facilities shape the built-environment of a community,
through a series of planning decisions based on local conventions,
economic eciency and social equity considerations, and lobbying
(Sampson et al., 2002; Witten et al., 2003). Thus, the ndings would
allow planners to better understand the relationships between built
environment and community opportunity, for promoting a fair com-
munity development for the future. The review of past studies on
building community opportunity index is presented in Section 2.2,
while spatial statistical modeling is discussed in Section 2.3 to support
the proposed research framework.
2.2. Geographical mapping of community opportunity
Conceptually, community opportunity can be measured by a broad
set of economic and social outcomes (Stimson et al., 2001). Hill et al.
(1998) developed a method to test the null hypothesis of homogeneity
among large America's cities, using a set of socioeconomic variables.
Stimson et al. (2001) applied the same method to identify the com-
munity opportunity and vulnerability for large cities in Australia.
Variables used in their studies included household income, unemploy-
ment rate, income growth, poverty, housing, and population change.
Sampson et al. (2002) also pointed out that there is a need to in-
corporate physical conditions (e.g. vacant houses and neighborhood
quality) in such analyses. In addition to the comparisons among cities,
some studies focused on the spatial variation within a city. In parti-
cular, geographical mapping of the socioeconomic outcomes of a
community was widely used to visualize the neighborhood context,
including sustainable employment, high-quality education, healthy and
safe neighborhood, aordable housing, and equal accessibility. The
mapping results can help address how the neighborhood environment
inuences individuals' future opportunity (Galster and Killen, 1995;
Rosenbaum et al., 2002) and provide a possibility for planners to better
allocate public facilities and services (Powell, 2007; Reece and
Schultheis, 2008).
To fulll this purpose, a GIS-based approach was used to illustrate
community opportunity (Pearce et al., 2006; Wridt, 2010). Particularly
in public health, various GIS-based methods were used to measure
community resource accessibility, such as the access to local libraries,
schools, parks, grocery stores, and pharmacies (Gilliland et al., 2006;
Law et al., 2011; Pez et al., 2012; Pearce et al., 2006; Witten et al.,
2011). Moreover, variables used for composing community opportunity
in recent studies extend from typical socioeconomic factors to other
aspects, such as education, child-wellbeing, and mobility and accessi-
bility (Pearce et al., 2006; Reece and Gambhir, 2008; Robert, 1999;
Witten et al., 2003). Using GIS, a sum of a set of un-weighted stan-
dardized factors (i.e., Z-scores) has been applied to visualize commu-
nity opportunity for many cities, such as Austin, Chicago, Baltimore,
Cleveland, New York City, Detroit, Houston, Atlanta, Los Angeles, and
Washington DC (Powell, 2007; Reece and Gambhir, 2008). The results
were also used in policy advocacy, community organizing, coalition
building, and service delivering (Reece and Schultheis, 2008).
C.-H. Wang, N. Chen
The advantage of geographical mapping is the illustration of com-
munity opportunity variation. However, it does not provide a com-
prehensive relationship between community opportunity and the built-
environment eects. This could be problematic because a vulnerable
community might result from local capital-disinvestments and un-
healthy land uses (Wridt, 2010). Furthermore, spatial autocorrelation
might exist among these neighboring vulnerable communities due to
spillover eects. Therefore, an approach that includes spatial statistical
modeling is recommended to account for such spatial dependence.
2.3. Spatial modeling for community opportunity
Spatial statistical approaches have been extensively used in recent
studies to better understand the nature of spatial dependence on various
social issues. Wang and Chen (2015) used spatial autoregressive (SAR)
models to examine the intensities of spatial autocorrelations (SA) in
three transport-based job accessibilities for Columbus, Ohio. They also
calculated marginal eects to understand the spillover eects of a unit
change in transportation investments. Similar research approaches used
on studying such social-network eects in a transportation context can
also be found in Goetzke (2008); Goetzke and Andrade (2010); Wang
et al. (2015). Moreover, spatial statistical models can be used to only
control for the SA in an ordinary regression model. For example, Akar
et al. (2016) used spatial error models (SEM) to examined the built-
environment and socioeconomic factors that impacted the trip distances
of individuals, while controlling for the SA in the regression residuals.
All of these spatial statistical models, however, assume that the esti-
mated coecients for the explanatory variables remain constant across
the region, even though the SAs are captured and controlled by a spatial
lag term in the model.
Another spatial statistical approach, geographically weighted re-
gression (GWR), can be considered, when the spatially varied re-
lationships between socioeconomic outcomes and policy investments
are examined. GWR models allow for the relationships in a regression
model to vary in a geographical area (Brunsdon et al., 1996; Paez,
2006; Tu, 2011). For instance, Comber et al. (2011) conducted a spatial
analysis of variations in health access, relating to socioeconomic status,
car ownerships, and geographical distance. Mulley (2014) reported that
land value uplift is partially contributed by accessibility by car and
accessibility to employment along the transit route. In his study, the
GWR model showed spatial variations in accessibility eects on land
values over the city of Sydney in Australia. Wang et al. (2011) com-
bined logit specications with GWR techniques to anticipate land-use
changes for Austin, Texas, while controlling for physical, locational,
and socio-demographical factors. GWR can also be used to forecast
travel demand, transit ridership, and car ownership (Blainey, 2010;
Cardozo et al., 2012; Clark, 2007; Pirdavani et al., 2014; Selby and
Kockelman, 2013). Although GWR has been widely used in various
social issues, there has been little research that examines the spatial
variation in the built-environment eects on community opportunity.
The use of GWR modeling would help better understand how the built-
environment attributes locally support or hinder opportunities for in-
dividuals' socioeconomic outcomes. The spatially varied relationships
could also help mitigate or oset social inequality by distributing
compensatory public facilities and services (Witten et al., 2003; Witten
et al., 2011).
3. Methodology
An ordinary regression model is rst specied to account for the
global built-environment eects on community opportunity, while
controlling for the socio-demographical characteristics, written as:
138
Journal of Transport Geography 62 (2017) 136147
COIi = 0 + k k x ik + i (1)
where COIi represents the community opportunity index for tract i, Xik
is the explanatory variables k for tract i, and 0 and k are parameters to
be estimated. The calculation of the COI is discussed in Section 4.1.
Next, a geographically weighted regression (GWR) is estimated to
capture the spatially varied built-environment eects on the COI. GWR
extends the ordinary regression framework by allowing local para-
meters to be estimated (Fotheringham et al., 2003; Wheeler and Pez,
2010), so that the GWR model for the COI is specied as:
COIi = 0 (ui , vi ) + k k (ui , vi ) x ik + i (2)
where COIi and Xik are the same expressions as in Eq. (1), and 0 (ui, vi)
and k (ui, vi) are parameters to be estimated. Note that (ui, vi) denotes
the coordinates of the centroid of tract i, and therefore, the estimated
parameters, 0 (ui, vi) and k (ui, vi), are allowed to vary over the region.
In Eq. (2), a search window, constructed by a bandwidth to identify
the neighbors around a certain location, is moving from one location to
the next across the region. A GWR model is then tted for the subset of
the data. Therefore, the results can be mapped and compared to ex-
amine the spatial variations. Fixed and adaptive kernel functions are
two general ways to construct a bandwidth, where an adaptive function
adjusts the window catchments for the density of data locations and a
xed function does not (Fotheringham et al., 2003; Wheeler and Pez,
2010). The adaptive kernel function is selected here due to the irregular
sizes of census tracts. Once the bandwidth for a location is identied,
data based on the dened catchment will be weighted according to
their nearest order to location i, and the vector of estimated coecients
at location i can be written as (Fotheringham et al., 2003):
(ui , vi ) = (X T W (ui , vi ) X )1X T W (ui , vi ) COI
(3)
where the bold type denotes a matrix, represents an estimate of ,
and W(ui, vi) is an n by n matrix whose diagonal elements denote the
weights for location i and o-diagonal elements are zero.
There are various kernel functions to dene the weighting scheme,
W(ui, vi), such as global, Gaussian, exponential, box-car, Bi-square, and
Tri-cube model (Fotheringham et al., 2003; Gollini et al., 2015). The
Gaussian model is selected here because it species the weights, wij, in
the weighting scheme as a continuous function of dij, written as
(Fotheringham et al., 2003):
2
1 dij
wij = exp
2 ,
b (4)
where dij denotes the distance between location i and j, and b refers to
the selected bandwidth. The weighting of data decreases from unity to a
very small value, according to a Gaussian curve as the distance between
i and j increases. For data away from location i, there will still be some
fairly large weights after the inection point on the Gaussian curve,
given by the bandwidth. This could be helpful in a small dataset, be-
cause the subset of the data theoretically covers all the samples over the
region, which helps ensure sucient local information for calibrating a
given local regression model.
Finally, the estimated parameters in a GWR model depend on the
selected kernel function, and therefore the selection of an appropriate
bandwidth is extremely important (Fotheringham et al., 2003). Cross-
validation (CV) can help estimate the optimal bandwidth by minimizing
all the prediction errors (Fotheringham et al., 2003; Wheeler and Pez,
2010). Alternatively, the Akaike Information Criterion (AIC) is to derive
the bandwidth by minimizing the AIC, providing a trade-o between
goodness of t and degrees of freedom (Fotheringham et al., 2003). In
this study, both methods were tested to nd the optimal bandwidth for
the estimations of the 284 local regression models.
C.-H. Wang, N. Chen Journal of Transport Geography 62 (2017) 136147

Table 1 2005 to 2010, the average number of jobs only decreased by 0.4%,
Denition and descriptive statistics of community opportunity composite variables. indicating a stable economic climate. Using the Network Analyst in
ArcGIS, accessible jobs by transit were calculated based on the Gen-
Variable Denitions Mean S.D.
eral Transit Feed Specication (GTFS) database. Accessible jobs by
Opportunity index car were also calculated using the Ohio street network dataset within a
COI Community opportunity index (Z-score) 0.00000 7.78 15-min driving buer. The average number of the accessible jobs by car
Economic opportunity (CACJ) (328,606) is much larger than that of jobs accessed by transit
INCO Median household income ($) 50,014 2,6385 (TACJ) (16,892).
UEMP Unemployment rate (%) 10.10 9.36 From the ACS estimates, the average number of people having at
PBAS Households with public assistance income 44.18 43.23
JCHA Percentage change of jobs from 2005 to 2010 0.40 39.90
least a bachelor's degree (AEDE) is 1600 per tract. On average, the
(%) dropout rate (DROP) is 6%, and 25% of students (from kindergarten to
TACJ Jobs reached within a 30-min transit 16,892 29,578 grade 12) are below the poverty line (SPOV). The following variables
commuting buer were assembled from Ohio School Report Cards and then converted into
CACJ Jobs reached within a 15-min driving buer 328,606 117,903
the census tract level, using the areal apportionment method. The
Education & child-wellbeing average percentage of student prociency in reading (PSRT) and math
AEDE Population with at least a bachelor's degree 1596 1151
(PSMT) is 75% and 69%, respectively. The average high school gra-
SPOV Student poverty rate (%) 25.16 24.16
DROP Dropout rate (%) 6.05 10.15 duation rate (GRAD) is 85%. In addition, the expenditure per pupil
EXPE Expenditure per pupil ($) 12,855.11 1603.23 (EXPE) has an average of $12,855, and the average student teacher
STRA Student teacher ratio (%) 15.60 0.98 ratio (STRA) is 15.6.
TQUA Full-time teachers 33.35 41.02 Median housing value (HVAL) has an average of $152,537, while
GRAD High school graduation rate (%) 85.22 7.87
the average home-ownership rate is 49% (HOWN). In a tract, about
Neighborhood & housing quality 10% of housing units are vacant (VACA), based on the 2010 Census
HVAL Median value of owner-occupied housing units 152,537 82,712
Summary File 1. Crime rate (CRIM), which was collected from 2008
($)
VACA Housing vacancy rate (%) 10.31 7.92 Justice Atlas of Sentencing and Corrections, reported that there was
CRIM Number of adult residents admitted to prison 3.76 3.99 average of 3.76 adult residents, per 1000 people, that were admitted to
per 1000 people prison. The population of Columbus increased at an average rate of
HOWN Home ownership rate (%) 48.77 23.99 4.55% from 2005 to 2010, suggesting that Columbus is a growing city.
PCHA Change rate of estimated population from 2005 4.55 14.28
to 2010 (%)
According to Census Transportation Planning Products (CTPP), the
average travel time to work (CTIM) is around 27 min. Using GIS, the
Mobility & accessibility
average distance to the nearest hospital from each census tract centroid
CTIM Mean travel time to work (minute) 26.59 9.87
DHOS Distance to nearest hospital (mile) 2.56 1.63 is 2.56 miles, while the average distances to the other facilities are
DSCH Distance to nearest school (mile) 0.47 0.38 around a half-mile.
DROD Distance to nearest primary and secondary road 0.55 0.43 It is worth noting that these COI variables measure the socio-
(mile)
economic outcomes of a community, while the explanatory variables
DBUS Distance to nearest bus stop (mile) 0.42 0.57
(presented in Section 4.2) describe the built-environment conditions.
Number of observations: 284 For instance, the nearest distance to a school (DSCH) is one of the ty-
pical methods to measure accessibility. From the perspective of plan-
4. Data ning, variable DSCH is a physical result of capital investments (i.e. the
allocation of schools), and not the other way around. We applied this
4.1. Dependent variable: community opportunity index and the composite rule for the other COI components to avoid the problem of endogeneity.
variables There are two steps to construct the COI. First, a set of Z-scores are
calculated for each tract by standardizing the factors reported above.
In this study, community opportunity is composed of a set of phy- Secondly, these Z-scores are combined in equal weight to be the COI
sical and socioeconomic outcomes, and the built-environment eects (see Table 1). Note that Z-Scores for negative factors were multiplied by
refer to the physical results of land-use planning and capital invest- 1 to create composite scores that represent low-to-high opportunity,
ments. To create a community opportunity index (COI) for this study, and therefore the signs of these COI components, in terms of their
physical and socioeconomic data were collected to evaluate the phy- meaning, point in the same direction. Thus, a tract with a higher score
sical and socioeconomic outcomes for the residential community in tends to have a better opportunity, and vice versa.
Columbus, Ohio. Census tract data were used as the community proxies. A similar approach has been widely used in evaluating urban
The selection of factors for a COI has been intensively discussed and sprawl. The scoring of the urban sprawl index and the composite factors
applied in the community-opportunity literature (Powell, 2007; Powell are explicitly explained in Ewing et al. (2003) and Ewing and Hamidi
and Reece, 2009; Reece and Gambhir, 2008; Reece and Schultheis, (2014). In their studies, the approach to compute the urban sprawl
2008). index includes two steps. They rst used PCA (principal component
Table 1 shows the descriptive statistics of the composite variables in analysis) to extract the rst principal component (PC1) as the re-
four categories, including economic opportunity, education and child- presentative of a composite dimension. PCA can eectively reduce the
wellbeing, housing and neighborhood quality, and mobility and ac- number of variables, and therefore was selected in their studies because
cessibility. From the 5-year estimates in 20062010 American Com- these variables are measures of the same dimension. There are four
munity Survey (ACS), the average median household income (INCO) of dimensions considered for constructing the urban sprawl index. The
$50,014 is above the state level of $47,333 (US Census Bureau, 2010). authors used an un-weighted approach to combine these four dimen-
The average unemployment rate (UEMP) is 10.1%. On average, 44 sions into one as the urban sprawl index because there is no theory to
households per tract are receiving public assistance income (PBAS), and support that one dimension is more important than one another.
20.63% of total population is below the poverty line (POVE). From Readers interested in the applications of urban sprawl index can consult
James et al. (2013), Ewing et al. (2014), Ewing et al. (2006), Berrigan

139
C.-H. Wang, N. Chen
et al. (2014), and Lee et al. (2009).
When computing the COI, we tried using PCA to avoid the problem
of correlations among the COI variables. However, the results failed to
identify some communities with better opportunities. Instead, we
computed the correlation coecient of any pair of the COI variables.
Variables PSRT, PSMT, and POVE were dropped because they are
highly correlated to some others. No signicant correlation problems
remained in the other COI variables. We observed that around 17% of
all the correlation coecients (325) were larger than 0.5, while only
2.8% of those were larger than 0.7. The spatial distribution of the
computed COI, together with main roads, hospitals, schools, and li-
braries, is illustrated in Fig. 1. Several spatial clusters with high COI
scores are located in the north, such as Upper Arlington, Worthington,
Dublin, and New Albany. In the meantime, vulnerable communities
mostly cluster around the city center and in the south.
140
Journal of Transport Geography 62 (2017) 136147
Fig. 1. Community opportunity index in Columbus, Ohio.
4.2. Explanatory variables: built-environment and socio-demographical
factors
As discussed in Section 2, the explanatory variables (i.e. the built-
environment eects) in a regression model include a set of physical and
socio-demographical factors that fall into the categories of regional
features, land uses, capital investments, and socio-demographical
characteristics. Table 2 shows their descriptive statistics. Regional
factors are the measures of varied metropolitan location-eects. Based
on the GIS calculations for each tract, the average Euclidean distance is
6.54 miles to the city center (DCBD) and 6.33 miles to The Ohio State
University (DOSU). Land-use data are available at the parcel level from
Franklin County Auditor. Residential use (RES), with an average share
of 37%, is the most intensive use in the region. Commercial use (COM)
follows with an average share of 16.4%. The average shares of the other
uses (industrial, utility, and open space/available land) are all less than
7%.
C.-H. Wang, N. Chen Journal of Transport Geography 62 (2017) 136147

Table 2 spatial variations.

Denition and descriptive statistics of explanatory variables.
5.1. Global OLS results
Variable Denitions Mean S.D.

Regional factors The estimates of the OLS model for the calculated COI are presented
DCBD Distances to the city center (mile) 6.54 3.21 in Table 3. The OLS model gives adjusted-R2 value of 0.621, explaining
DOSU Distance to The Ohio State University (mile) 6.33 3.11
about 62% of the variation in the COI scores. Most of the selected ex-
DAIR Distance to Port Columbus International Airport 7.86 3.66
planatory variables are signicant at the 0.05 level. The variable ASIA
Land uses (share of Asian population) and ZEVH (share of housing units with zero
RESI Share of residential use (%) 36.89 18.67
COMM Share of commercial use (%) 16.36 12.57
vehicle) are signicant at the 0.1 level. Nevertheless, there are some
INDU Share of industrial use (%) 4.69 9.21 insignicant variables, including UTIL (utility share), INDU (industrial-
UTIL Share of utility use (%) 1.32 2.83 use share), RAMP (highway ramp), BUDN (bus stop density), LIBA (li-
AGRE Share of agricultural use (%) 3.85 10.26 brary) and HOSP (hospital).
OPEN Share of open space (%) 6.14 7.44
For each tract, distances from the city center (DCBD) and The Ohio
Capital investments (transportation infrastructure and public facilities) State University (DOSU) are the two indicators of metropolitan loca-
RAMP Presence of highway ramp (Dummy) 0.39 0.49
tion-eects. The variable DCBD has a positive sign, suggesting that a
STDN Street intersection density (# per square mile) 191.98 90.93
BUDN Bus stop density (# per square mile) 17.42 17.73 tract farer away from the city center is likely to have a higher COI score
SCHO Number of schools 1.45 1.38 (i.e. suburbanization). The negative sign of the variable DOSU implies
HOSP Number of hospitals 0.06 0.28 that OSU tends to have more positive spillover-eects in near tracts
LIBA Number of libraries 0.13 0.35 than distant ones.
Socio-demographic characteristics An area with a higher share of residential uses usually has fewer
WHIT Share of white population (%) 65.89 26.11 negative externalities from other uses, while an area with intensive
BLAK Share of black population (%) 25.03 26.01
commercial uses tends to locate numerous employment opportunities
ASIA Share of Asian population (%) 3.32 4.03
OTER Share of minorities (%) 5.76 3.74 (Wang and Chen, 2015). Therefore, a tract with a higher share of re-
ZVEH Percentage of housing units with zero vehicle (%) 10.04 10.25 sidential or commercial uses is expected to have a higher COI score.
SPFR Percentage of single-parent families (%) 20.80 12.31 These two variables (RESI and COMM) both have an expected sign. In
PODN Population density (# per square mile) 4512 3486 addition, open space usually represents environmental amenity and
future development opportunities (Geoghegan, 2002; Irwin and
Number of observations: 284
Bockstael, 2001). The variable OPEN is signicantly positive, implying
that the higher the open space share in a tract, the higher the COI score
Capital investment is another important component for measuring
becomes. These spatial relationships can be used to provide planning
the built-environment eects. Columbus has been long regarded as a
information for future land-use allocations. The analysis of marginal
car-oriented city (Wang and Chen, 2015). In the present study, the
eects can help better understand the impacts (i.e. COI premiums) of a
presence of highway ramp (RAMP) is a dummy variable, calculated
unit change in a given land use. A 1% increase in residential or com-
from Ohio Department of Transportation's TIMS (Transportation In-
mercial uses will increase the COI score by 0.11, while a 1% increase in
formation Management System). About 39% of the total 284 tracts have
open space will increase the score by 0.10.
a highway ramp. From 2010 Census TIGER/LINE Shapeles and Central
The variable STDN has a positive sign, indicating that intensive
Ohio Transit Authority (COTA)'s GTFS database, there are, on average,
street-intersections could improve COI scores, through the improved
192 street intersections (STDN) and 17 bus stops (BUDN) per square
street connectivity and accessibility. Investments in schools (SCHO)
mile in a tract. From Mid-Ohio Regional Planning Commission, there
could also give rise to COI scores. The analysis of their COI premiums
are 412 schools, 17 hospitals, and 37 libraries scattering over the region
provides planners an insight into the problems of allocating further
(see Fig. 1).
capital investments. COI has a score change of 0.01 for a unit change in
Several socio-demographical variables, collected from 2010 Census
street-intersection density, and 0.47 for one in schools.
Summary File 1, are used to control for other social-demographical
Black population (BLAK) is treated as the benchmark in the OLS
eects in the regression model. The region is dominated by white po-
model. Black population, together with the minorities (OTER), is con-
pulation (WHIT), with an average share of 66%, followed by black
centrated around the city center, while white (WHIT) population
population (BLAK), with an average share of 25%. On average, 10% of
clusters toward the outskirt areas. Asian (ASIA) population mostly re-
total housing units do not have access to private vehicles (ZVEH). The
sides in the northern suburbs adjacent to Honda of America
average percentage of single-parent families is 21%, which shares a
Manufacturing. The positive sign of the variable WHIT and ASIA in-
similar spatial pattern with black population. Population density
dicates that a tract with a higher share of white or Asian population has
(PODN) is also considered, having an average of 4512 people per square
a better chance of having a higher COI score. The variable OTER has a
mile.
negative sign, implying that a tract with a higher share of minorities is
likely to be at a disadvantage of community opportunity.
5. Results The negative sign of the variable SPAT implies the existence of
spatial mismatch between single-parent families and vulnerable com-
The focus of using GWR in this study is on investigating the spatial munities. Similar results are also found for another socially dis-
variability of the built-environment eects on the calculated commu- advantage group, zero-vehicle households. Population density is used to
nity opportunity index (COI). Some insignicant explanatory variables control for other socio-demographical eects. The negative sign in-
in the OLS (global) model may be locally signicant in the GWR (local) dicates that the higher the population density, the higher the compe-
model. More interestingly, some global parameter estimates may locally tition becomes among residents.
change their signs at some locations. In addition, Mulley (2014) argues
that the best OLS model does not necessarily lead to the best GWR 5.2. GWR diagnostic information
model. Therefore, this study applies the approach suggested in Wheeler
and Tiefelsdorf (2005) that the local estimated coecients should be Moran's I test is used to investigate the existence of spatial auto-
interpreted in the context of the global estimated results. The results are correlation (SA) in the residuals from the OLS model, and the results are
shown below in three subsections, together with the maps that show the presented in Table 4. When running Moran's I tests, the assumptions of

141
C.-H. Wang, N. Chen Journal of Transport Geography 62 (2017) 136147

Table 3
Estimations of community opportunity models.

Models GWR OLS

Min. 1st Qu. Median 3rd Qu. Max. Global VIF

Variables
Intercept 13.06 8.22 4.60 0.26 4.33 4.53 (1.5)
Regional factors
DCBD 0.77 0.13 0.29 0.92 1.63 0.86 (3.6) 7.2
DOSU 2.07 1.26 0.71 0.30 0.49 1.14 (5.1) 6.1
Land use features
RESI 0.02 0.04 0.06 0.12 0.18 0.11 (4.1) 3.1
COMM 0.06 0.07 0.09 0.10 0.12 0.11 (3.2) 2.2
OPEN 0.19 0.01 0.07 0.14 0.20 0.10 (2.0) 1.7
UTIL 0.07 0.15 0.18 0.23 0.27 0.16 (1.4) 1.3
INDU 0.10 0.02 0.00 0.06 0.12 0.005 (0.1) 1.7
Capital investments
RAMP 3.22 1.43 0.66 0.01 1.01 0.65 (0.9) 0.5
STDN 0.00 0.01 0.01 0.01 0.02 0.01 (2.3) 2.1
BUDN 0.06 0.03 0.00 0.02 0.06 0.02 (0.7) 3.1
SCHO 0.08 0.23 0.52 0.93 1.51 0.47 (2.0) 1.2
LIBA 0.42 0.53 1.27 1.72 2.64 1.30 (1.4) 1.2
HOSP 1.60 0.84 0.12 1.62 4.29 0.93 (0.8) 1.2
Races (BLAC as benchmark)
ASIA 0.23 0.00 0.10 0.26 0.77 0.16 (1.7) 1.9
WHIT 0.02 0.06 0.09 0.13 0.16 0.06 (3.1) 2.7
OTER 0.70 0.52 0.38 0.29 0.18 0.39 (3.9) 1.7
Social characteristics
SPAT 0.29 0.20 0.13 0.09 0.04 0.19 (3.7) 4.8
ZVEH 0.16 0.11 0.08 0.05 0.00 0.08 (1.64) 2.9
POPD 0.0005 0.0004 0.0003 0.0001 0.00004 0.0003 (2.2) 2.1
Adaptive bandwidth quantile (%)
Adjusted R-square (%) 79.7 62.1
AICc/AIC 1663.9 1717.0

1. Signicance: 0.05 in bold.

2. Absolute t-value in parenthesis.

neighborhood structures (e.g., graph-based, n-nearest, and distance- coecient is increased by local collinearity (Wheeler and Tiefelsdorf,
based) and spatial weights (e.g., standardized, binary, and inverse 2005; Wheeler and Pez, 2010). In the last column of Table 3, the VIFs
distance weighting (IDW)) are rarely discussed in other studies. The n- values of the explanatory variables are all less than the critical value of
nearest neighborhood structure is selected here, because of the variable 10 (Cardozo et al., 2012; Mason, 1989; Neter et al., 1989), suggesting
size of census tracts. Several n-nearest neighborhood structures that there are no serious multicollinearity problems. The existence of
(n = 10, 20, 40, 80, 160) are used to capture the SA, and the IDW SA and insignicant multicollinearity both resort to an approach of
matrix is used to account for the eects deceasing with distance. The using GWR.
average radii of the ve adaptive catchments range from 2.58 to The results of the GWR estimations are presented in Table 3, using a
9.6 km. The dierence between the observed Moran's I and the ex- Gaussian kernel weighting function with an adaptive bandwidth. The
pectation decreases as n increases. Nevertheless, all the Moran's' I tests cross-validation (CV) and Akaike Information Criterion (AIC) approach
have very small p-values, indicating the existence of SA. This result is are both tested for selecting the optimal bandwidth (Fotheringham
also supported by a spatial non-stationary pattern of the COI in Fig. 1, et al., 2003; Wheeler and Pez, 2010). The AIC approach is preferred
implying that the modelled relationship might vary across the region. because it produces a larger adaptive bandwidth with a quantile of
Multicollinearity is a potential problem in all regression-based 14.4%. The subset of data for each local regression would extend from
models, leading to inated standard errors (Mulley, 2014; Wheeler and the 40 samples captured by the adaptive bandwidth, the inection
Tiefelsdorf, 2005). When estimating a GWR model, Wheeler and Pez point on the Gaussian curve, to more samples with fairly large weights.
(2010) strongly suggest using diagnostic tools to evaluate local colli- In Table 3, a larger adjusted-R2 (0.797 over 0.621) and a lower AIC/
nearity, such as scatter plots of estimated coecients, maps of ap- AICc (1663.9 over 1717.0) both suggest that the GWR model is a better
proximate local coecient correlations, and local variance ination t than the OLS model. This also conrms that this dataset contains
factors (VIFs). VIFs can measure how much the estimated variance of a sucient samples for the local regression to identify the existence of
spatial dependence (Mulley, 2014).
Table 4
Moran's I tests for the residuals of the OLS model.
5.3. Local GWR results
n-Nearest neighborhood n = 10 n = 20 n = 40 n = 80 n = 160
structure GWR modeling allows for visualizing the spatial variations of the
Min. distance (km) 0.47 0.47 0.47 0.47 0.47
explanatory variables. The local estimated coecients, together with
Mean distance (km) 2.58 3.47 4.76 6.69 9.60 the calculated t-values, are mapped for all the explanatory variables in
Max. distance (km) 11.56 14.29 17.99 22.93 30.04 gures below. Similar to the analyses from the OLS model, these local
Observed Moran's I 0.55 0.47 0.38 0.30 0.02 parameter estimates can be seen as the COI premiums of a unit change
Expectation 0.004 0.004 0.004 0.004 0.004
in a given explanatory variable. In each gure, positive local coecient
Standard deviation 20.53 24.44 27.59 30.64 31.50
p-Value 0.00 0.00 0.00 0.00 0.00 values (i.e. positive premiums) are colored in green, while negative
values (i.e. negative premiums) are colored in red. Also, two dierent

142
C.-H. Wang, N. Chen
shading areas indicate where the local parameters estimates are sig-
nicant at the 0.05 and 0.1 level respectively. These gures show
various patterns of the spatial variations in local parameter estimates.
For the variable DCBD, Fig. 2a shows the tracts of positive pre-
miums in the outskirt areas toward the southwest. Most tracts are sta-
tistically signicant at the 0.05 level. However, there is a small clus-
tering of negative premiums (see the red points in Fig. 2a), which is
adjacent to the Port Columbus International Airport. This implies that
the negative externality of noise from the airport might has stronger
impacts than the metropolitan location-eects do. The variable DOSU
has a totally opposite spatial pattern to the variable DCBD. Most tracts
143
Journal of Transport Geography 62 (2017) 136147
Fig. 2. Local coecients and t-values of metropolitan-lo-
cation variables.
in the outskirt areas show negative premiums of a unit change in dis-
tance to OSU (see the green points in Fig. 2b). A few tracts around the
campus show positive premiums, although they are not signicant at
the 0.1 level.
As found in the OLS results, most tracts over the region show sta-
tistical signicance with positive premiums of up to 0.18 COI score per
1% change in residential use (RESI) and 0.12 per 1% change in com-
mercial use (COMM) (see Fig. 3a and b). Tracts having higher positive
premiums cluster toward the south for the variable COMM and toward
the east for RESI. More interestingly, there is a small clustering of in-
signicant local RESI and COMM coecients located in the north (i.e.
Fig. 3. Local coecients and t-values of land-use variables.
C.-H. Wang, N. Chen
Worthington), suggesting that additional residential or commercial uses
would not add to COI scores in long-developed and better-o areas. The
local coecients of the variable OPEN are locally signicant in the east,
with positive premiums of up to 0.20 COI score (see Fig. 3c). Eastern
Columbus is a relatively newly developed area, and the results suggest
that open space tends to contribute to COI scores particularly in new
developments. Finally, the variables UTIL and INDU, although not
signicant in the OLS model, are locally signicant in the GWR model.
A few tracts in the north show positive premiums for the variable INDU
(see Fig. 3d), and two small clusters on the two sides of the city center
show the same results for the variable UTIL. From the perspective of
land-use planning, these ndings point out that eastern Columbus
might be the most eective location for future developments when
maximizing total COI scores. It is worth noting that the residential or
commercial increments would compete with the preservation of open
space in some tracts, because these three variables all create positive
COI premiums.
Transportation infrastructure and location may have eects on re-
sidential property values and the accessibility to reach employment
opportunities (Haider and Miller, 2000; Kain, 1968; Preston and
144
Journal of Transport Geography 62 (2017) 136147
Fig. 4. Local coecients and t-values of capital-investment
variables.
McLaerty, 1999). The dummy variable RAMP is not signicant in the
OLS model. However, the GWR results show a large concentration
around the city center, having signicant negative premiums (see
Fig. 4a). It is sensible that the presence of highway ramp in the
downtown area would lower COI scores, due to the increased trac.
Except some tracts in the north, the variable STDN is signicant over
the region with positive premiums (see Fig. 4b). The setting of bus stops
is an important planning tool to improve the mobility and accessibility
for social disadvantaged groups. The variable BUDN, not signicant in
the OLS model, does show a small cluster of signicant positive pre-
miums in the northwest (i.e. Dublin) (see Fig. 4c). In Fig. 4d and f, the
variable SCHO and HOSP show signicant positive premiums (up to
1.51 and 4.29 respectively) for the tracts in northeastern Columbus,
where relatively lacks for schools and hospitals (see Fig. 1). The nding
suggests that investments in schools and hospitals would add to the COI
score to communities in this area, but might not do so for other long-
developed areas. Finally, the variable LIBA shows a large concentration
particularly in the mid-west (i.e. Upper Arlington) (see Fig. 4e), having
signicant positive premiums of up to 2.64 COI scores. These maps
clearly show considerable spatially variations, suggesting that local
C.-H. Wang, N. Chen
characteristics of COI premiums require consideration, when planning
for the allocation of future capital investments.
The tracts in northern Columbus generally have a higher COI score
(see Fig. 1). The local results from the GWR model can be used to better
understand spatial mismatch problems. For instance, Fig. 5a shows the
tracts of having larger positive premiums for white population (WHIT)
in the north, where also shows larger negative premiums for the
minorities (OTER) (see Fig. 5c). This nding indicates that the other
minorities, a socially disadvantaged group, usually reside in vulnerable
communities. It is sensible that most Asian (ASIA) population resides
northern suburbs (see Fig. 5b), where shows higher COI scores. The COI
premium mapping for these socio-demographical variables can help
local governments and community organizations develop a fair com-
munity opportunity framework. For instance, aordable housing pro-
grams should target at the single-parent families (SFAT) particularly
clustering in the southwest, where clustering of larger negative pre-
miums (see Fig. 5d).
145
Journal of Transport Geography 62 (2017) 136147
Fig. 5. Local coecients and t-values of socio-demographic
variables.
6. Discussion
The local estimated coecients (i.e. the COI premiums) were
mapped for visualizing the spatial variations. The OLS results show a
global trend that a tract would have a higher COI score if it is either far
away from the city center or close to The Ohio State University (OSU).
However, the GWR results show a local inverse relationship that a tract
would tend to have a lower score if it is too close to the international
airport, due to the externalities of noise and disturbance. The GWR
model also oers an opportunity to see if the insignicant global
parameters would turn out to be locally signicant. In the GWR model,
two land-use variables (i.e. utility and industrial share) both become of
having signicant positive premiums in some tracts. More importantly,
residential and commercial increments would add to COI scores, par-
ticularly in eastern Columbus. These new developments would also
compete with the preservation of open space, which also gives high
positive premiums here.
Several globally insignicant capital-investment variables are also
found of having some local signicant positive premiums, including
RAMP (the presence of highway ramp), BUDN (bus-stop density), LIBA
C.-H. Wang, N. Chen
(library), and HOSP (hospital). A considerable variety in local COI
premiums is found among the capital-investment variables. For in-
stance, the results show that improvements in schools and hospitals
would create higher positive premiums in northeastern Columbus,
where relatively lacks for educational and medical resources. Libraries
do add to COI scores particularly in the mid-west (e.g. Upper
Arlington), while the street-intersection density does so in most tracts
over the region.
Finally, the GWR results reveal the spatial mismatch problems in
Columbus. Several socially disadvantaged groups (e.g. black popula-
tion, minorities, single-parent families, and zero-vehicle households)
tend to reside in certain vulnerable communities. This suggests that a
uniform opportunity-based development policy may exacerbate the gap
between winners and losers, and therefore need to be carefully in-
vestigated, depending upon the local socioeconomic characteristics (Du
and Mulley, 2012; Mulley, 2014).
7. Conclusion and areas for future research
A spatial analytical approach has been developed and applied to
analyze community opportunity, using a combination of geographical
mapping and geographically weighted regression (GWR) modeling. The
geographical mapping of the community opportunity index (COI) was
rst presented, using the aggregated scores of a set of socioeconomic
factors. The estimations of the OLS and GWR model followed to ex-
amine the global and local built-environment eects on the COI, while
controlling for the socio-demographical characteristics. The analysis of
marginal eects was also used to understand the COI premiums of a
unit change in a given explanatory variable. In particular, the results
reveal the most eective locations for allocating future land uses and
capital investments when maximizing total COI scores. To the best of
our knowledge, this paper is the rst to have applied the GWR modeling
in answering the questions of how these planning outcomes (i.e. the
built environment) are locally associated with the socioeconomic out-
comes of a community.
The GWR model presented in this study can be extended into an
optimization-modeling framework that would help solve social-equity
problems. The outputs of the present GWR model, including land uses,
capital investments, and their local estimated coecients (i.e. the COI
premiums), could serve as inputs to this new framework. For instance,
an optimization model could be developed to allocate future land-use
and capital-investment increments while maximizing total COI scores.
Alternatively, an optimization model could be formulated to minimize
total COI dierences among the 284 tracts by an optimal built-en-
vironment allocation. Implementing these possible extensions is cer-
tainly beyond the scope of this study, and therefore remains as an area
for future research.
References
Acevedo-Garcia, D., Osypuk, T.L., Werbel, R.E., Meara, E.R., Cutler, D.M., Berkman, L.F.,
2004. Does housing mobility policy improve health? Hous. Pol. Debate 15, 4998.
Akar, G., Chen, N., Gordon, S.I., 2016. Inuence of neighborhood types on trip distances:
spatial error models for Central Ohio. Int. J. Sust. Transp. 10, 284293. http://dx.doi.
org/10.1080/15568318.2014.903447.
Arcaya, M.C., Tucker-Seeley, R.D., Kim, R., Schnake-Mahl, A., So, M., Subramanian, S.,
2016. Research on neighborhood eects on health in the United States: a systematic
review of study characteristics. Soc. Sci. Med. 168, 1629.
Berrigan, D., Tatalovich, Z., Pickle, L.W., Ewing, R., Ballard-Barbash, R., 2014. Urban
sprawl, obesity, and cancer mortality in the United States: cross-sectional analysis
and methodological challenges. Int. J. Health Geogr. 13, 3. http://dx.doi.org/10.
1186/1476-072x-13-3.
Blainey, S., 2010. Trip end models of local rail demand in England and Wales. J. Transp.
Geogr. 18, 153165.
Brunsdon, C., Fotheringham, A.S., Charlton, M.E., 1996. Geographically weighted re-
gression: a method for exploring spatial nonstationarity. Geogr. Anal. 28, 281298.
Cardozo, O.D., Garca-Palomares, J.C., Gutirrez, J., 2012. Application of geographically
weighted regression to the direct forecasting of transit ridership at station-level. Appl.
Geogr. 34, 548558.
Clark, S.D., 2007. Estimating local car ownership models. J. Transp. Geogr. 15, 184197.
146
Journal of Transport Geography 62 (2017) 136147
Comber, A.J., Brunsdon, C., Radburn, R., 2011. A spatial analysis of variations in health
access: linking geography, socio-economic status and access perceptions. Int. J.
Health Geogr. 10, 44. http://dx.doi.org/10.1186/1476-072x-10-44.
Du, H., Mulley, C., 2012. Understanding spatial variations in the impact of accessibility
on land value using geographically weighted regression. J. Transp. Land Use 5,
4659. http://dx.doi.org/10.5198/jtlu.v5i2.225.
Dundas, R., Leyland, A.H., Macintyre, S., 2014. Early-life school, neighborhood, and fa-
mily inuences on adult health: a multilevel cross-classied analysis of the Aberdeen
children of the 1950s study. Am. J. Epidemiol. 180, 197207. http://dx.doi.org/10.
1093/aje/kwu110.
Ellaway, A., Benzeval, M., Green, M., Leyland, A., Macintyre, S., 2012. Getting sicker
quicker: does living in a more deprived neighbourhood mean your health deterio-
rates faster? Health & Place 18, 132137. http://dx.doi.org/10.1016/j.healthplace.
2011.08.005.
Emery, M., Flora, C., 2006. Spiraling-up: mapping community transformation with
community capitals framework. Community Dev. 37, 1935.
Ewing, R., Hamidi, S., 2014. Measuring Sprawl 2014. Smart Growth America. https://
www.smartgrowthamerica.org/app/legacy/documents/measuring-sprawl-2014.pdf
(02/03/2016).
Ewing, R., Schmid, T., Killingsworth, R., Zlot, A., Raudenbush, S., 2003. Relationship
between urban sprawl and physical activity, obesity, and morbidity. Am. J. Health
Promot. 18, 4757.
Ewing, R., Brownson, R.C., Berrigan, D., 2006. Relationship between urban sprawl and
weight of United States youth. Am. J. Prev. Med. 31, 464474.
Ewing, R., Meakins, G., Hamidi, S., Nelson, A.C., 2014. Relationship between urban
sprawl and physical activity, obesity, and morbidity update and renement.
Health & Place 26, 118126.
Flora, J.L., Sharp, J., Flora, C., Newlon, B., 1997. Entrepreneurial social infrastructure and
locally initiated economic development in the nonmetropolitan United States. Sociol.
Q. 38, 623645.
Fotheringham, A.S., Brunsdon, C., Charlton, M., 2003. Geographically Weighted
Regression: The Analysis of Spatially Varying Relationships. John Wiley & Sons.
Galster, G.C., Killen, S.P., 1995. The geography of metropolitan opportunity: a re-
connaissance and conceptual framework. Hous. Pol. Debate 6, 743.
Geoghegan, J., 2002. The value of open spaces in residential land use. Land Use Policy 19,
9198.
Gilliland, J., Holmes, M., Irwin, J.D., Tucker, P., 2006. Environmental equity is child's
play: mapping public provision of recreation opportunities in urban neighbourhoods.
In: Vulnerable Children and Youth Studies. 1. pp. 256268.
Goetzke, F., 2008. Network eects in public transit use: evidence from a spatially auto-
regressive mode choice model for New York. Urban Stud. 45, 407417. http://dx.doi.
org/10.1177/0042098007085970.
Goetzke, F., Andrade, P.M., 2010. Walkability as a summary measure in a spatially au-
toregressive mode choice model: an instrumental variable approach. In: Pez, A.,
Gallo, J., Buliung, N.R., Dall'erba, S. (Eds.), Progress in Spatial Analysis: Methods and
Applications. Springer, Berlin Heidelberg, Berlin, Heidelberg, pp. 217229.
Gollini, I., Lu, B., Charlton, M., Brunsdon, C., Harris, P., 2015. GWmodel: an R package for
exploring spatial heterogeneity using geographically weighted models. J. Stat. Softw.
63, 50. http://dx.doi.org/10.18637/jss.v063.i17.
Haider, M., Miller, E., 2000. Eects of transportation infrastructure and location on re-
sidential real estate values: application of spatial autoregressive techniques. Transp.
Res. Rec. 18.
Hill, E.W., Brennan, J.F., Wolman, H.L., 1998. What is a central city in the United States?
Applying a statistical technique for developing taxonomies. Urban Stud. 35,
19351969.
Irwin, E.G., Bockstael, N.E., 2001. The problem of identifying land use spillovers: mea-
suring the eects of open space on residential property values. Am. J. Agric. Econ. 83,
698704.
James, P., Troped, P.J., Hart, J.E., Joshu, C.E., Colditz, G.A., Brownson, R.C., Ewing, R.,
Laden, F., 2013. Urban sprawl, physical activity, and body mass index: nurses' health
study and nurses' health study II. Am. J. Public Health 103, 369375.
Jencks, C., Mayer, S.E., 1990. The social consequences of growing up in a poor neigh-
borhood. In: Inner-city Poverty in the United States. 111. pp. 186.
Kain, J.F., 1968. Housing segregation, negro employment, and metropolitan decen-
tralization. Q. J. Econ. 82, 175197.
Kestens, Y., Chaix, B., Shareck, M., Valle, J., 2016. Comments on Melis et al. the eects
of the urban built environment on mental health: a cohort study in a large northern
Italian city. Int. J. Environ. Res. Public Health, 2015, 12, 1489814915. Int. J.
Environ. Res. Public Health 13, 250. http://dx.doi.org/10.3390/ijerph13030250.
Law, M.R., Dijkstra, A., Douillard, J.A., Morgan, S.G., 2011. Geographic accessibility of
community pharmacies in Ontario. Healthc. Pol. 6, 3646.
Lee, I.-M., Ewing, R., Sesso, H.D., 2009. The built environment and physical activity le-
vels: the Harvard Alumni Health Study. Am. J. Prev. Med. 37, 293298.
Mason, R.L., 1989. In: Gunst, R.F., Hess, J.L. (Eds.), Statistical Design and Analysis of
Experiments: With Applications to Engineering and Science. Wiley, New York.
Mulley, C., 2014. Accessibility and residential land value uplift: identifying spatial var-
iations in the accessibility impacts of a bus transitway. Urban Stud. 51, 17071724.
Neter, J., Wasserman, W., Kutner, M.H., 1989. Applied Linear Regression Models. Irwin.
Oakes, J.M., 2004. The (mis) estimation of neighborhood eects: causal inference for a
practicable social epidemiology. Soc. Sci. Med. 58, 19291952.
Oakes, J.M., Andrade, K.E., Biyoow, I.M., Cowan, L.T., 2015. Twenty years of neigh-
borhood eect research: an assessment. Curr. Epidemiol. Rep. 2, 8087. http://dx.
doi.org/10.1007/s40471-015-0035-7.
Paez, A., 2006. Exploring contextual variations in land use and transport analysis using a
probit model with geographical weights. J. Transp. Geogr. 14, 167176.
Pez, A., Scott, D.M., Morency, C., 2012. Measuring accessibility: positive and normative
C.-H. Wang, N. Chen
implementations of various accessibility indicators. J. Transp. Geogr. 25, 141153.
Pearce, J., Witten, K., Bartie, P., 2006. Neighbourhoods and health: a GIS approach to
measuring community resource accessibility. J. Epidemiol. Community Health 60,
389395.
Perchoux, C., Chaix, B., Brondeel, R., Kestens, Y., 2016. Residential buer, perceived
neighborhood, and individual activity space: new renements in the denition of
exposure areasthe RECORD cohort study. Health & Place 40, 116122.
Peterson, R.D., Krivo, L.J., Harris, M.A., 2000. Disadvantage and neighborhood violent
crime: do local institutions matter? J. Res. Crime Delinq. 37, 3163.
Pirdavani, A., Bellemans, T., Brijs, T., Kochan, B., Wets, G., 2014. Assessing the road
safety impacts of a teleworking policy by means of geographically weighted regres-
sion method. J. Transp. Geogr. 39, 96110.
Powell, J., 2007. Communies of Opportunity: A Framework for a More Equitable and
Sustainable Future for All. Ohio State University: Kirwin Institute for the Study of
Race and Ethnicity.
Powell, J.A., Reece, J., 2009. The Future of Fair Housing and Fair Credit: From Crisis to
Opportunity, Symposium: New Strategies in Fair Housing.
Preston, V., McLaerty, S., 1999. Spatial mismatch research in the 1990s: progress and
potential. Pap. Reg. Sci. 78, 387402.
Reece, J., Gambhir, S., 2008. The geography of opportunity: review of opportunity
mapping research initiatives. In: Kirwan Institute for the Study of Race and Ethnicity.
Ohio State University, Columbus.
Reece, J., Schultheis, E., 2008. Poverty's place: the use of geographic information systems
in poverty advocacy. Clear. Rev. 42, 430.
Robert, S.A., 1999. Socioeconomic position and health: the independent contribution of
community socioeconomic context. Annu. Rev. Sociol. 489516.
Rosenbaum, J.E., Reynolds, L., DeLuca, S., 2002. How do places matter? The geography
of opportunity, self-ecacy and a look inside the black box of residential mobility.
Hous. Stud. 17, 7182.
Sampson, R.J., Raudenbush, S.W., Earls, F., 1997. Neighborhoods and violent crime: a
multilevel study of collective ecacy. Science 277, 918924. http://dx.doi.org/10.
1126/science.277.5328.918.
Sampson, R.J., Moreno, J.D., Gannon-Rowley, T., 2002. Assessing neighborhood ef-
fects: social processes and new directions in research. Annu. Rev. Sociol. 28,
443478.
Scribner, R.A., Cohen, D.A., Farley, T.A., 1998. A geographic relation between alcohol
147
Journal of Transport Geography 62 (2017) 136147
availability and gonorrhea rates. Sex. Transm. Dis. 25, 544548.
Selby, B., Kockelman, K.M., 2013. Spatial prediction of trac levels in unmeasured lo-
cations: applications of universal kriging and geographically weighted regression. J.
Transp. Geogr. 29, 2432.
Smith, W.R., Frazee, S.G., Davison, E.L., 2000. Furthering the integration of routine ac-
tivity and social disorganization theories: small units of analysis and the study of
street robbery as a diusion process. Criminology 38, 489524.
Stimson, R., Baum, S., Mullins, P., O'Connor, K., 2001. A typology of community op-
portunity and vulnerability in metropolitan Australia. Pap. Reg. Sci. 80, 4566.
Tu, J., 2011. Spatially varying relationships between land use and water quality across an
urbanization gradient explored by geographically weighted regression. Appl. Geogr.
31, 376392.
US Census Bureau, 2010. State & County QuickFacts: Ohio. http://quickfacts.census.gov/
qfd/states/39000.html (02/07/2016).
Wang, C.-H., Chen, N., 2015. A GIS-based spatial statistical approach to modeling job
accessibility by transportation mode: case study of Columbus, Ohio. J. Transp. Geogr.
45, 111. http://dx.doi.org/10.1016/j.jtrangeo.2015.03.015.
Wang, Y., Kockelman, K., Wang, X., 2011. Anticipation of land use change through use of
geographically weighted regression models for discrete response. Transp. Res. Rec.
111123.
Wang, C.-H., Akar, G., Guldmann, J.-M., 2015. Do your neighbors aect your bicycling
choice? A spatial probit model for bicycling to the Ohio State University. J. Transp.
Geogr. 42, 122130.
Wheeler, D.C., Pez, A., 2010. Geographically weighted regression. In: Fischer, M., Getis,
A. (Eds.), Handbook of Applied Spatial Analysis. Springer.
Wheeler, D., Tiefelsdorf, M., 2005. Multicollinearity and correlation among local re-
gression coecients in geographically weighted regression. J. Geogr. Syst. 7,
161187.
Witten, K., Exeter, D., Field, A., 2003. The quality of urban environments: mapping
variation in access to community resources. Urban Stud. 40, 161177.
Witten, K., Pearce, J., Day, P., 2011. Neighbourhood Destination Accessibility Index: a
GIS tool for measuring infrastructure support for neighbourhood physical activity.
Environ. Plan. A 43, 205223.
Wridt, P., 2010. A qualitative GIS approach to mapping urban neighborhoods with
children to promote physical activity and child-friendly community planning.
Environ. Plan. B Plan. Des. 37, 129147.