Journal of Transport Geography 31 (2013) 267–277

Contents lists available at SciVerse ScienceDirect

Journal of Transport Geography

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

Urban sprawl and social interaction potential: an empirical analysis

of large metropolitan regions in the United States
Steven Farber ⇑, Xiao Li
Department of Geography, University of Utah, 260 South Central Campus Drive, Rm. 270, Salt Lake City, UT 84112-9155, United States

a r t i c l e i n f o a b s t r a c t

Keywords: This paper investigates the impact of urban spatial structure on the opportunities for people to partici-
Social interaction potential pate in face-to-face activities. We make use of the Social Interaction Potential (SIP) metric, a tool to mea-
Urban spatial structure sure the average level of opportunity for people to engage in face-to-face activities given some basic
Time geography constraints on their daily mobility patterns. Generally, this opportunity is a reflection of whether the
Urban sprawl
urban spatial structure – a term that broadly applies to the spatial arrangement of land-uses and the
Spatial interaction
interactions between them – constrains or permits the ability for potential activity partners to be at
the same place and time. In this paper, the SIP metric is applied to 42 metropolitan regions in the United
States with populations over 1,000,000 people. These measurements are regressed against a set of indi-
cators of urban sprawl to expose the relationship between spatial structure and SIP. The indicators are
generated by a factor analysis of a large set of variables describing the scale, centrality and dispersion
of land-uses in addition to several other structural and infrastructure-related variables. Cluster analysis
is also used to organize the regions into similar groupings with respect to their structural characteristics
and the level of SIP they provide. The findings indicate that social interaction potential is hampered by
decentralization, fragmentation, and longer commutes in the largest metropolitan regions in the country.
Interestingly, the negative effect of decentralization on SIP Efficiency is found to be nearly ten times
stronger than that of fragmentation and nearly 20 times stronger than the effect of mean commuting
duration.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction are regressed against characteristics of their spatial structure,

Social interaction is a fundamental human need (Maslow,
1943). At the individual level, social contact plays an important
role in psychological development and well-being (House et al.,
1988). At the community level it is responsible for cohesion and
social capital development (Coleman, 1988; Forrest and Kearns,
2001). And economically, it is the basis for innovation and
creativity (Storper and Venables, 2004). If the city is an arena in
which social interaction takes place, then urban spatial structure
can either be an obstacle that constrains Social Interaction Poten-
tial (SIP), or an asset that supports it. In this paper, we quantify
empirically how SIP is related to regional-scale spatial structure,
with a particular focus on the impact of urban sprawl on the ability
for people to engage in face-to-face activities.
This research uses the previously developed social interaction
potential metric to measure the average quantity of joint time–
space accessibility between all commuters in 42 metropolitan re-
gions across the continental United States (Farber et al., in press;
Neutens et al., 2008). The calculated SIP metrics for these regions
⇑ Corresponding author. Tel.: +1 801 585 8218; fax: +1 801 581 8219.
E-mail address: steven.farber@geog.utah.edu (S. Farber).
revealing the sign, magnitude, and explanatory strength of each
factor.
The total set of social contact opportunities is composed of
many elements: interactions within the household, interactions
between friends, interactions between people who work, interac-
tions between people who don’t work, impromptu interactions,
planned interactions, etc. This paper focusses on one particular
element of this set, after-work interaction opportunities amongst
people who commute. Of course, this is a large component repre-
senting a high percentage of the population in the United States
(e.g. 54.9% of Americans over 16 commute to a workplace outside
of their home according to the 2010 American Community Survey
5 year estimates). Restricting this even further, the SIP metric is
particularly suited to capture the ability for people to conduct
planned face-to-face social contact activities, as opposed to casual,
impromptu contact with neighbors or fellow bus-riders, or contact
via electronic media. Also, it only looks at one representative,
weekday evening time-period. But for working days, the evening
time-period is an important time many workers have to perform
discretionary social activities. So, while we contend that there is
more to understand before having a complete picture of social
interaction opportunities in a city, we may begin by adding to

0966-6923/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jtrangeo.2013.03.002
268 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277
our understanding of how the urban spatial structure impacts this
rather pervasive component of daily social life – planned evening
interactions amongst workers.
The SIP metric attempts to balance the social-contact benefits of
living or working in dense neighborhoods with the inherent time
constraints associated with the regimen of the daily commute
activity. The contention put forward by the SIP body of work is that
the effects of marginal density distributions (residential and
employment densities) on social contact opportunities are dimin-
ished by the burdens of lengthy commutes. Thus, a method that
takes both the spatial and temporal implications of urban spatial
structure into consideration is required to more accurately repre-
sent how the opportunities for social interaction are impacted by
metropolitan-scale urban spatial structure.
The remainder of the paper is structured as follows. First we
provide an overview of the relevant literature. Following that, the
SIP metric is defined, and the data used to measure spatial struc-
ture are described. The results of the analysis are put forward in
the next section, followed by a discussion and conclusions.
2. Literature review
The literature relevant to this research is split between work
motivating the empirical question, that is, the role of spatial struc-
ture in social interaction potential, and the work that underpins
the methodological approach.
2.1. Urban form and social interaction
The degree to which urban form hinders or fosters social inter-
action is most commonly investigated at the local scale. Much of
the existing research emanates from the proponents of New
Urbanism, who propose that dense, mixed-use development cou-
pled with architectural and civic design strategies can be used at
the local scale to induce impromptu neighborly interaction (Talen,
1999). The findings of this body of research are mixed, partly due
to the complexity of heterogeneous tastes, perceptions, and behav-
iors (Sander, 2002). The research, however, is consistent in finding
that automobile use is among the significant factors predicting re-
duced levels of social participation. Placing the emphasis on travel
behavior rather than land use, Freeman (2001) found that land use
density in a neighborhood was not related to the number of social
ties between neighbors, upon controlling for the strong negative
impact of automobile dominance in the neighborhood. Leyden
(2003) concluded that households rating their neighborhoods as
more walkable were also ‘‘more likely to know their neighbors,
participate politically, trust others, and be socially engaged’’ (p.
1546). Similarly, Lund (2003) found the perception of walkability
to be positively related with participating in social activities with
neighbors, but concluded that personal characteristics and atti-
tudes played significant roles as well. Finally, over the last 15 years,
Grannis (1998, 2005, 2009) has shown overwhelming evidence
that local walkability supports the development of strong local
communities.
The above studies suggest that residential walkability leads to
higher incidences of social connectedness at the neighborhood
level, but they are all cross-sectional in design, suggesting that re-
verse causality and self-selection may confound the interpretation
of their results (Cao et al., 2006). For example, it is important to
recognize that these claims echo significant changes in people’s
housing preferences. In fact, a recent Brookings paper reports on
growing evidence that the supply of automobile oriented neigh-
borhoods far exceeds the ever diminishing demand, citing chang-
ing preferences amongst home-buyers as a significant causal
factor (Leinberger, 2011). This echoes earlier, pre-recession
findings, of the existence of latent demand for more walkable,
cohesive neighborhoods (Levine et al., 2005). It is plausible that
the set of preferences that make people want to relocate in more
walkable neighborhoods also make these individuals more likely
to form strong social ties within the neighborhood. This raises
the question of whether or not a more walkable local neighbor-
hood ‘‘causes’’ people to become more socially active.
Furthermore, while these studies identify urban sprawl – or
rather those local-scale urban form elements typical of sprawling
cities – as a culprit in reducing community-level social interaction,
they do not consider the potential metropolitan-scale effects of
sprawl and automobility on time–space trajectory dispersion. Urry
(2002) and Sheller and Urry (2000) hypothesize that the automo-
bile has simultaneously individualized, fragmented, and dispersed
time–space trajectories, thus making synchronization of social
interaction activities difficult to accomplish for many people.
Moreover, Wellman (2001) identifies automobile oriented subur-
banization as a causal factor in the spatial thinning of social
networks. These two sociological arguments speak to a growing
difficulty for people to engage in planned social activities with
members of their social network; cities have expanded, social net-
works have become more spatially dispersed, and individuals are
carrying out more individualistic and less routine daily space–time
patterns than ever before.
These sociological hypotheses were examined empirically by
Farber and Páez (2009, 2011a,b) in a series of time-use studies of
social activity participation. The findings contradict time-geogra-
phy’s traditional conceptualization that the extra velocity provided
by the automobile should enable increased access to opportunities
and therefore more activity participation. Rather, the studies were
in support of the sociological theories of Urry and Welman. They
found that increased automobile use was associated with lower
rates of social activity participation even after controlling for
demographic, household, and socio-economic factors. Also, the in-
crease in overall duration of mandatory automobile travel from
1992 to 2005 was associated with a significant decline in social
activity participation amongst Canadians in the same time period.
They logically reasoned that metropolitan-scale activity dispersion
limits the opportunity for social interaction in urban areas, but this
could not be directly examined by those authors using time-use
data alone.
2.2. Contact fields and activity spaces
A contact field is a geometric estimation of the interaction po-
tential for an individual at a fixed location. The interaction poten-
tial is predicated upon an individual’s trip-distance preference
function and an areal distribution of population density. In the
early part of geography’s quantitative revolution, contact fields
were used to investigate processes that result from interaction be-
tween populations with different spatial distributions (Dacey,
1971; Moore, 1970; Moore and Brown, 1970). They helped
researchers understand and predict spatial behavior within cities
through geometric reasoning and analysis (Golledge et al., 1972).
The intellectual prominence of contact fields diminished once dis-
aggregate travel behavior models derived from random utility the-
ory broadly supplanted the quantitative geographical approaches
(Domencich and McFadden, 1975). Still, they are the starting point
in a decades-long quest to understand how the urban spatial struc-
ture impacts opportunities for spatial interaction (Beckmann et al.,
1983a,b). Indeed, the relationship between spatial (and network)
structures and access to interaction opportunities remains a rele-
vant topic, with the research area receiving attention in a recent
special issue of Networks and Spatial Economics, a journal dedi-
cated to understanding the impacts of human infrastructure on
economic outcomes (Reggiani and Martín, 2011).
 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277
Recently, the concept of the contact field has re-emerged in
transportation and activity behavior research. The present reincar-
nation, more commonly referred to as an activity space, describes
the region in which people regularly travel to conduct activities
within a specified time period (Schonfelder and Axhausen, 2003).
In terms of its use as a contact field, Wong and Shaw (2011) use de-
tailed activity space data for a limited number of respondents to
measure the exposure dimension of segregation. These person-
based exposure measures were then assigned back to their home
neighborhood so that aggregate measures could be constructed
at the zonal level. By shifting to this empirical approach, what is
gained in terms of accurate representation of time–space con-
straints tends to be lost in terms of generality; the geographies
of everyday life are highly heterogeneous and are only captured
for a small sub-sample of the population who are surveyed. It
seems misleading to use such small-sample estimates to approxi-
mate zonal averages. To address this, Farber and Páez (2012) used
the spatial expansion method (Casetti, 1972) to estimate contact
spaces from a highly-detailed travel survey, thereby bridging the
gap between theoretical geometrical work and data-driven
empirical representations. However, their modeled contact-field
approach is isotropic and it does not take into consideration the
true patterns of daily travel largely associated with the daily
commute. It has been shown that access to opportunities is more
accurately estimated when taking into consideration the two main
anchors of daily life, home and work (Salze et al., 2011).
2.3. Social interaction potential
The SIP metric is a time geographic implementation of contact
fields. It uses a rigorous theoretical model of joint-accessibility first
implemented and tested with hypothetical data by Neutens et al.
(2010, 2008). The joint-accessibility work has shown the extent
that such factors as the configuration of transport networks and
land uses, time constraints, and power differentials between group
members affect the opportunities for conducting social activities.
However, while their model is valuable for assessing meeting
opportunities at a microscopic level, due to enormous data require-
ments, its usefulness becomes compromised when insights into
broader spatial trends of social interaction potential at a metropol-
itan scale are desired. Farber et al. (in press) use the concept of
joint-accessibility in the derivation of a metric for metropolitan-
scale social interaction potential. Using journey-to-work data from
the US Census (2000), this approach allows us to construct general-
ized, after-work contact fields for the entire working population of
any city, while explicitly incorporating the two most definitional
daily constraints on activity, the home and workplace.
Prior work using this metric has been focused on internal vali-
dation (Farber et al., in press). In that paper, a basic SIP model was
applied to test-cases based on synthetic land-use/transportation
scenarios. SIP was calculated for eighty scenarios, synthetic cities
with five levels of commuter dispersion and 16 combinations of
employment and residential population distributions. In addition,
SIP and the four zone-based sub-metrics of SIP were computed
for the Salt Lake City region using commuter flow data and net-
work-based travel times. A meta-regression of the test results
clearly indicated that SIP increases significantly with monocentric-
ity of land-uses and decreases significantly with longer commutes.
However it remains to be seen how the metric behaves with
respect to differences between spatial structures of real metropol-
itan regions.
In another paper, the SIP metric was calculated for Flanders,
Belgium, a mega-region home to Brussels, Antwerp, Gent, and
Brugges (Neutens et al., in press). The study showed a strong rela-
tionship between SIP and population scale and density. However,
many cities within the region obtained higher or lower SIP levels
269
than would be expected by population size alone, returning us
again to the unanswered question of how spatial structure relates
to this unexplained variation after controlling for marginal density
distributions.
3. Research method
3.1. The social interaction potential metric
We now describe the proposed model for measuring SIP on a
metropolitan scale. For simplicity, the development of the model
in this application focuses on a social activity within a single,
after-work time budget. In future research, the model can be
extended to other parts of the day and multiple time budgets with
the appropriate modifications. The metric proposed in this section
provides a method for measuring the social interaction potential of
an entire metropolitan region.
Let the scalar value, b > 0, be the after-work time budget for a
discretionary activity under analysis in a region comprising N
administrative zones, Z1, Z2, . . ., ZN. Let tab be the amount of time re-
quired to travel from Za to Zb. Then,

ðb  t ki  t jk Þ if b P t jk þ t ki
Akij ¼ ð1Þ
0 otherwise
is the amount of time available for an after-work discretionary
activity at Zk for an individual, denoted by Iij, that lives in Zi and
works in Zj. The condition in (1) ensures the non-negativity of Akij
in situations when the total required travel time from Zj to Zi
through Zk is longer than the after-work discretionary time budget.
Suppose Iij is attempting to participate in a joint activity at
location Zk with an individual Iqr. In this case, the amount of time
available to them for a joint activity at Zk is:
(
maxð0; minðb  t ki ; b  tkq Þ  maxðtjk ; trk ÞÞ if Akij ; Akqr P 0
Akijqr ¼
0 otherwise
ð2Þ
where max(tjk, trk) is the beginning of the overlapping time period,
min(b  tki, b  tkq) is the ending of the overlapping time period,
and the condition in (2) ensures that both individuals have non-
zero time availability at Zk. The max(0, ) function will replace a
negative number with 0 in the event that Akij and Akqr are temporally
non-overlapping.
It follows that the total social interaction potential for Iij and Iqr
across the N potential zones of interaction is:
X
N
Aijqr ¼ Akijqr : ð3Þ
k¼1
The value of Aijqr can be thought of as an approximation to the
volume of the intersection between the time–space prisms belong-
ing to Iij and Iqr. The total SIP in a region is the average intersection
volume for all pairs of individuals in the region.
In the case that data are aggregated into zones, commute flow
probabilities are used in place of an exhaustive computation. For
P
example, let R ¼ i;j r ij , where rij is the number of workers living
r
in Zi and working in Zj. Then Pij ¼ Rij is the percentage of workers
in the region that travel from zone i to zone j.
It follows that the metric:
XXXXX
SIP ¼ Akijqr P ij Pqr ð4Þ
i j q r k
is the expected social interaction potential for all pairs of
generalized (i.e. aggregated) workers in the city. When the region
produces large prism intersections more often than smaller ones,
the value of the SIP metric goes up. Alternatively, when large prism
270 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277
intersections are few and far between, the SIP metric will be low.
Thus, cities with shorter travel times between activity locations or
more spatially concentrated commuter flow structures will, in
general, produce both larger prism intersections more often, and
more opportunity for social interaction on average.
The metric is bounded by the range [0, Nb]. SIP attains its
minimum value when all pairs of individuals have non-intersecting
time–space prisms, and its maximum value when all pairs of time–
space prisms have at least b minutes of overlap at each location in
the city. In reality, both of these cases are unlikely, but the ratio
SIP/Nb is used as a measure of efficiency; it is the ratio of the
achieved level of SIP divided by the total possible level of SIP. This
efficiency score is useful when comparing SIP between different
cities so we use this as the dependent variable in the regression
modeling that follows. The efficiency score also has a direct
interpretation as the average amount of time available for social
interaction between any two workers at any one location in the
city as a percentage of the total time budget, b.
3.2. Study area
The 42 metropolitan regions with populations greater than 1
million were selected for study in this research. Regions, in this
case, are defined by the boundaries of Metropolitan Planning Orga-
nizations (MPOs) since these are the agencies responsible for
setting long-term transportation policy at a regional, super-muni-
cipal scale. The sample size is somewhat limited, mostly due to the
computational burden of computing the SIP metric, and the time
required to download and compute the explanatory variables used
in the model.
3.3. Data
Data used in this research can be split between those products
incorporated into the metric computations, and those used to
describe the spatial structure of the regions. Despite its computa-
tional complexity, the SIP metric is derived using only two data
inputs: a zone-to-zone origin/destination matrix of morning
commuters, and an inter-zonal matrix of travel times. The origin/
destination matrix for each MPO was obtained from Part 3 of the
Census Transportation Planning Products (http://www.fhwa.dot.-
gov/planning/census_issues/ctpp/) for the year 2000 at the census
tract level of geography. Part 3 of the CTPP is freely available at the
Traffic Analysis Zone (TAZ) level of aggregation; however the Cen-
sus Tract was selected as the unit of analysis for this study since
TAZ definitions vary between MPOs, and it is yet unclear how
the SIP metric is affected by the Modifiable Areal Unit Problem
(Fotheringham and Wong, 1991; Zhang and Kukadia, 2005). While
it would be desirable to use a more current journey-to-work
dataset, at the time this goes to press, small-area aggregations of
American Community Survey (ACS) have not been released, nor
are they expected to have the same coverage afforded by the
large-sample Census 2000 dataset. The travel time matrix is based
on free-flow centroid-to-centroid travel times across the Street-
Map USA street network. The matrix was calculated for each region
using ESRI’s ArcGIS Network Analyst extension.
The variables used to describe urban spatial structure are based
heavily on the sprawl indicators developed by Ewing et al. (2003).
A similar approach to Ewing’s factor analysis is taken in our paper;
however, variables related to observed transportation behavior that
were left out of Ewing’s method as dependent variables are now
included in our factor analysis. The variables used to describe spatial
structure appear in Table 1. They can be grouped into 4 categories:
Scale, Distribution, Centrality, and Other. The Scale variables pertain
to the overall area and population of the region. The Distribution
variables describe the aspatial variation and spatial clustering of
the distributions of origins and destinations in the study area. The
Centrality variables describe the strength of the monocentric core
(or lack thereof), and the rate at which origins and destinations
decline with distance from the center of the region. The center of
each region was determined a priori through an investigation of land
use densities and confirmed by a visual inspection of maps. The
centrality variables are all measured with respect to this central
location for each region. The existence of at least one center is
assumed for each region. However, the centrality variables are used
precisely to tell us whether or not this central location is an impor-
tant one, the extent to which other sub-centers exist in the region,
and how much the population is concentrated around this center
relative to other locations in the region.
Finally, the Other variables include two transportation mea-
sures, highway density and mean commute time, and the fractal
dimension of the urbanized area, a measure of the scattered and
‘‘leap-frogged’’ nature of sprawling land use patterns (Batty and
Longley, 1994; Frenkel and Ashkenazi, 2008; Shen, 2002).
3.4. Factor and regression analysis
The goal of this study is to explain the computed SIP Efficiency
scores as a function of the spatial structure of the regions. The
sample of 42 regions is small compared to the total number of
explanatory variables, and a correlation analysis revealed that
many of the variables are strongly correlated. For these reasons,
we perform a principal component factor analysis with oblique
promax rotation to create 5 latent explanatory factors, only slightly
correlated with each other. Introducing some correlation between
the factors resulted in more easily interpretable factor loadings.
Following the factor analysis, regression analysis of the SIP Effi-
ciency score is carried out using the reduced set of right-hand side
variables. Diagnostic inspection for linearity revealed that several
of the factors are related to SIP Efficiency non-linearly. Optimal
transformation functions were estimated using Multivariable Frac-
tional Polynomial models (MFP) in Stata (Royston and Altman,
1994; Sauerbrei and Royston, 1999). The linear predictor for a frac-
tional polynomial model of order M on covariates X is:
X
K X
M
b0 þ bkm X Pkm
m
k¼1 m¼1
where each power Pm is chosen from the restricted set
{2, 1, 0.5, 0, 0.5, 1, 2, 3} and X0 is taken as ln(X). Using the MFP
algorithm, all pairwise combinations of powers on all variables
are fitted, and the best fitting model, according to the Akaike Infor-
mation Criterion, is selected. The resulting model is composed of a
linear combination of polynomials of the original exogenous factors.
Interpretation of the coefficients is somewhat more tedious since
elasticities will vary with respect to X. Be that as it may, producing
point-estimates of the elasticities is still fairly straightforward by
taking partial derivatives and evaluating at particular values of X.
3.5. Cluster analysis
Cluster analysis is used to group observations into similar clas-
ses called clusters. We employ k-median clustering to find groups
of regions with similar factor scores. Group means for SIP and SIP
Efficiency are then compared to group means for the factor scores.
This analysis enriches our understanding of the relative impor-
tance of the various factors, and how the SIP and SIP Efficiency
measures are differentially impacted by the factors.
 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277 271

Table 1
Variables used to describe spatial structure.a

Variable Description
Scale variables
CT The number of census tracts in the MPO
Area MPO area in square miles
Population Number of people who work and live in the MPO
Distribution variables
Population density Number of people who work and live in the MPO per square mile
MeanO Mean value of census tract residential density
VarianceO Variance of residential density
CVO Coefficient of variation of residential density
MeanD Mean value of census tract employment density
VarianceD Variance of employment density
CVD Coefficient of variation of employment density
MeanET Mean population/employment census tract entropy
VarET Variance of population/employment entropy
CVET Coefficient of variation of population/employment entropy
MoranO Moran’s I of origin count based on queen contiguity
MoranD Moran’s I of destination count based on queen contiguity
MoranOD Moran’s I of origin density based on queen contiguity
MoranDD Moran’s I of destination density based on queen contiguity
Centrality variables
GradientO Slope of residential density distance decay model
lnAO Intercept of residential density distance decay model
AdjR2O Adjusted R-squared of residential density distance decay model
GradientD Slope of employment density distance decay model
lnAD Intercept of employment density distance decay model
AdjR2D Adjusted R-squared of employment density distance decay model
O5, O25, O50, O75, O95 Time travelled from CBD to reach 5%, 25%, 50%, 75% and 95% of origins
D5, D25, D50, D75, D95 Time travelled from CBD to reach 5%, 25%, 50%, 75% and 95% of destinations
Other variables
FD Fractal dimension of urbanized area
Highway Density Highway density (miles/square miles)
mtt Average 1-way commute duration for the region
a
Further explanations for some variables appear in the Appendix A.

4. Results 4.1.2. Origin and destination distributions

Here we put forward the results of the empirical analysis in four
sections. First, we describe the distributions of the variables.
Second we report on the results of the factor analysis. Third, the re-
sults of the regression analysis are displayed and the coefficients
interpreted. Finally, the cluster analysis results are provided.
4.1. Descriptive analysis
Basic descriptive statistics are recorded for each variable in
Table 2. What follows is a summary of the variable classes (Scale,
Distribution, Centrality, and Other) and how the metropolitan re-
gions rank on the various dimensions.
4.1.1. City scale
The cities range quite widely in terms of overall area, the num-
ber of census tracts contained within, and the total number of
workers. Jacksonville, FL is the smallest region in the study com-
prised of only 150 census tracts, a working population of 347,500
and an area of approximately 918 square miles. If the city were a
perfect circle, such an area would be contained within a radius of
only 17 miles. Fort Lauderdale, Providence, and Buffalo are simi-
larly sized regions according to these scale measures. On the other
side of the spectrum is Los Angeles, with an enormous surface area
of approximately 39,992 square miles (radius of 112 miles). Phoe-
nix, Washington DC, Seattle and New York also have very large
areas, while in terms of working population the ranking is New
York (9 million), Los Angeles (6.5 million), Chicago (4.5 million)
and Philadelphia (4.3 million).
Once again, the regions vary quite a lot in terms of census tract
densities and the variation of densities about their means. In terms
of overall density, Phoenix, Salt Lake City and Las Vegas, are
amongst the least dense, while Newark, New York and Boston
top the list. According to the coefficients of variation, the larger
metropolises tend to have more evenly distributed origin and dis-
tribution tract densities. Providence, Ft. Lauderdale, New York and
West Palm Beach are the regions with the highest degree of O/D
evenness within census tracts based on the entropy measures,
whereas Las Vegas, San Francisco, San Diego and Baltimore on
average have tracts which are more polarized. Quite interestingly,
the spatial autocorrelation statistics reveal that spatial distribu-
tions of origins and destinations counts vary from completely
random (San Diego, Las Vegas and Denver for origins, and Virginia
Beach and Houston for destinations), to highly clustered
(Milwaukee, Chicago, Cleveland for origins, and New York, Wash-
ington/Baltimore and Chicago for destinations). On the other hand,
all cities show signs of positive spatial autocorrelation for origin
and destination densities. For densities (as opposed to raw counts),
New York is the most spatially clustered, both in terms of origins
and destinations, while Ft. Lauderdale, Orlando, Houston and Mia-
mi tend to be amongst the least spatially clustered regions in terms
of O&D densities. In general, regions rank similarly in terms of both
origin and destination density clustering with the exceptions of
Pittsburgh, St. Louis, Twin Cities, Sacramento, and Buffalo, which
are far more clustered in terms of origin densities than they are
for destinations.
4.1.3. Centrality
The centrality measures are largely determined by estimated
density gradient functions for origins and destinations (see
272 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277

Table 2 the scale of density at each region’s core. Finally, the adjusted r2
Descriptive statistics. measures the goodness-of-fit of the negative exponential regres-
Variable Mean Std. dev. Min. Max. sion model, with higher levels suggesting a greater feasibility of
Dependent variable a monocentric assumption over the regional spatial structure. In
SIP Efficiency (%) 19.8 12.1 2.6 44.2 terms of residential distributions, Portland (r2 = 57%), Las Vegas
Scale variables (r2 = 41%), Boston (r2 = 57%), and Indianapolis (r2 = 50%) have the
CT 899.1 974.3 150 5237 sharpest rates of decline away from the center, while Cincinnati
Area 7663.2 7078.6 918.8 39992.8 (r2 = 12%), Philadelphia (r2 = 16%), San Francisco (r2 = 16%) and
Population 1,699,260 1,766,855 347,476 8,966,356 Los Angeles (r2 = 25%) have the flattest. The reader will notice that
Distribution variables regions with the flattest gradients also tend to be poorly repre-
Population Density 267.0 161.2 65.5 832.3 sented by the monocentric model, as indicated by their r2’s. For
MeanO 2179.0 1418.5 795.8 9306.0
VarianceO 9882785 2.92E+07 417186.2 1.87E+08
destination distributions, Portland (r2 = 59%), Orlando (r2 = 47%),
CVO 1.084 0.244 0.681 1.814 San Antonio (r2 = 50%), and Jacksonville (r2 = 40%) have the steep-
MeanD 2300.0 1485.2 1060.8 10242.1 est gradients, while Philadelphia (r2 = 13%), Cincinnati (r2 = 12%),
VarianceD 1.18E+08 3.50E+08 1.84E+06 2.26E+09 Cleveland (r2 = 21%) and Los Angeles (r2 = 25%) have the flattest.
CVD 0.365 0.133 0.186 0.782
MeanET 0.560 0.021 0.487 0.617
VarET 0.018 0.003 0.009 0.028
CVET 0.032 0.007 0.014 0.057 4.1.4. Other variables
MoranO 0.315 0.127 0.040 0.579 Mean travel times for each region are based on the CTPP travel
MoranD 0.246 0.110 0.009 0.605 flow matrices and calculated free-flow travel speeds. In other
MoranOD 0.527 0.106 0.229 0.747 words, it is an average of all O/D pair travel times weighted by
MoranDD 0.386 0.116 0.215 0.794
the observed census flow for each pair. Atlanta, LA, Houston, Phoe-
Centrality variables nix and Washington top the list with 21–23 min mean 1-way com-
GradientO 0.053 0.019 0.092 0.019
lnAO 8.212 0.461 7.154 9.323
mute times. Keeping in mind that the calculations were based on
AdjR2O 0.369 0.150 0.096 0.712 free-flow travel-speeds, these commutes can be considered a
GradientD 0.064 0.024 0.113 0.021 hypothetical ‘‘best-case’’ duration, with real commute times being
lnAD 7.983 0.480 7.034 9.856 larger due to congestion and/or use of non-automobile modes. Fort
AdjR2D 0.372 0.130 0.116 0.603
Lauderdale, Buffalo and Providence have the shortest mean travel
O5, O25, O50, O75, O95 – – – –
D5, D25, D50, D75, D95 – – – – times of 13–14.5 min. Since mean travel time is somewhat relative
to the overall size of the region, we can investigate MTT/Area to see
Other variables
FD 1.866 0.049 1.708 1.960 which regions have longer travel times than otherwise would be
Highway Density 0.420 0.226 0.124 0.912 expected based on the region size. From this we observe that Los
mtt 17.8 2.3 13.7 22.8 Angeles, Phoenix, Seattle and New York perform poorly, while Jack-
sonville, Fort Lauderdale, Boston and Providence perform better
than expected.
Appendix A for full description). In both cases, higher magnitude For highway density, the range is quite large, with Boston, Bal-
(more negative) gradients indicate steeper declines of density timore and Newark having nearly 1 mile of highway for every
away from the core, and a higher level of centrality. The lnAO square mile of land area, and Phoenix, Salt Lake City, and Seattle
and lnAD terms are the estimated model constants, and represent with only 0.12–0.14 miles per square mile. It should be cautioned

Table 3
Rotated factor loadings for |loading| > 0.5.

Decentralization Big City Fragmented Low mixing Long travel

Variable Loading Variable Loading Variable Loading Variable Loading Variable Loading
GradientO 1.0125 Population 0.6189 CVO 0.6601 MeanET 0.9641 Area 0.7185
AdjR2O 0.6868 Pop. Density 0.6537 CVD 0.7827 VarET 0.9609 Pop. Density 0.5876
GradientD 0.9811 MeanO 1.0145 MoranO 0.6485 CVET 0.9884 mtt 0.6057
AdjR2D 0.8799 VarianceO 0.9182 MoranOD 0.548 Highway Density 0.7963
O5 0.6109 MeanD 0.9507 AdjR2O 0.7357
O25 0.8523 VarianceD 0.8729 FD 0.6694
O50 0.8877 MoranD 0.7148
O75 0.8831 MoranDD 0.5534
O95 0.7237 lnAO 0.8341
D5 0.6082 lnAD 0.9041
D25 0.8945
D50 0.8919
D75 0.9196
D95 0.7451
High scores: High scores: New High scores: High scores: Las High scores: Los
Philadelphia, Los York, Boston, Chicago, Atlanta, Vegas, San Angeles, Phoenix,
Angeles, Chicago, San New York, Twin Francisco, Kansas, Denver,
Washington, Francisco, Los Cities, Washington San Diego, Detroit Minneapolis,
Baltimore, San Angeles Sacramento
Francisco
Low scores: Low scores: Low scores: Fort Low scores: Low scores:
Portland, Las Nashville, Lauderdale, West Providence, New Milwaukee,
Vegas, Austin Cincinnati, West Palm Beach, Miami, York, Ft. Jacksonville,
New Orleans, Palm Beach, Virginia San Diego, Lauderdale, West Boston,
Boston Beach, Jacksonville Jacksonville Palm Beach, Boston Providence,
Cleveland
 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277 273

Table 4
Results of multivariate fractional polynomial regression.

.4
Partial predictor+residual of SIPE
Variablea Coefficient Std. err. z P P |z|
F11 0.1799 0.0175 10.27 0.000
F12 0.0220 0.0045 4.89 0.000

.3
F2 0.0117 0.0077 1.52 0.129
F3 0.1246 0.0151 8.27 0.000
F4 0.0090 0.0047 1.90 0.057

.2
F5 0.0134 0.0054 2.48 0.013
Constant 0.1715 0.0062 27.66 0.000

Model summary:

.1
Number of obs. = 42
Replications = 10,000
Wald chi2(6) = 561.52
Prob > chi2 = 0.00

0
R-squared = 0.9563
-2 -1 0 1 2 3
Adj R-squared = 0.9488
Root MSE = 0.0275 Decentralization
a
Transformed factors are described in text. Fig. 3. Fractional polynomial plot of Decentralization versus predicted SIP Efficiency.

.5
SIP Efficienty - Predicted vs. Actual
.5

Partial predictor+residual of SIPE

Predicted SIP Efficiency
.4

.4
.3

.3
.2
.1

.2
0

0 .1 .2 .3 .4
.1

SIP Efficiency
-3 -2 -1 0 1 2
95% CI
Fragmentation

Fig. 1. Strength and pattern of linear prediction. Fig. 4. Fractional polynomial plot of Fragmentation versus predicted SIP Efficiency.
.3
.05

Partial predictor+residual of SIPE

.25
Residuals

.2
0

.15
.1
-.05

-2 0 2 4 6
0 .1 .2 .3 .4 .5 Big City
Fitted values
Fig. 5. Fractional polynomial plot of Big City versus predicted SIP Efficiency with
Fig. 2. Residuals versus predicted values. New York appearing as a potential leverage point.
274 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277

that Salt Lake City and Seattle both contain large portions of wil-
derness in their regional boundaries, potentially biasing the true,
or relevant, measurement of highway density.
Finally, the fractal dimension variable, a measure of land use
fragmentation, indicates that Chicago, Baltimore, Houston, Atlanta
and Las Vegas are the most fragmented, while Ft. Lauderdale, Mia-
mi, San Diego and Orlando have the most regular patterns of
urbanized land use.
4.2. Factor analysis results
Factor analysis is used to identify the underlying structure of
the 35 variables describing urban spatial structure. Using a scree
plot and a table of factor loadings, we identify 5 rotated factors that
explain a large proportion of the underlying variance (80% before
rotation). They are easily interpreted as: Decentralization, Big City,
Fragmentation, Low Mixing, and Long Travel. The factor loadings
and high/low ranking cities for each factor appear in Table 3. The
factors are not orthogonal since an oblique rotation was used. This
eased the interpretation of the structure while maintaining a fairly
low level of correlation between factors. The maximum Pearson
correlation is between Big City and Fragmentation at r = 0.37, while
all other pairwise correlations are below 0.3 in absolute value.
4.3. Regression results
Interpretation of the coefficients are fairly straightforward,
eased by the fact that the dependent variable, SIP Efficiency, is a
percentage value, and that each of the (untransformed) covariates
are distributed N(0, 1). For the untransformed covariates, F2, F4
and F5, a change in 1 standard deviation in the independent factor,
is predicted to correspond with a change of b percentage points
where b is the estimated regression coefficient. Thus, the model
predicts SIP Efficiency to decrease significantly with respect to
the Long Travel factor, increase slightly insignificantly with the
Big City factor, and increase ever so slightly with respect to the
Low Mixing factor.
Since F1 and F3 have been entered as non-linear polynomials,
interpretation from the regression equation can be aided with
the fractional polynomial partial predictor plots in Figs. 3 and 4.
In both cases, SIP Efficiency is predicted to decline fairly linearly
with respect to the independent factor, with the effect of Decentral-
ization flattening in the higher range of the factor, and the effect of
Fragmentation being slightly steeper in the lower range of the fac-
tor values. In both cases, the strength of prediction is incredibly
strong, as indicated by their very high z-scores in Table 4 and the
95% confidence interval around the partial prediction plots.
Not all of the variables have such strong levels of prediction in
comparison to the Decentralization and Fragmentation variables. For
example, the Big City factor (p = 0.129) is slightly insignificant at
the 90% level of confidence and a plot reveals that New York may

According to our common-sense hypotheses regarding urban

spatial structure and SIP, we expect that Decentralization, Fragmen-
.25

tation, Low Mixing and Long Travel will be negatively associated

Partial predictor+residual of SIPE

with SIP Efficiency, while Big City will be positively associated. To

test these hypotheses we perform a Multivariate Factional Polyno-
mial (MFP) regression, so that non-linear relationships can be
.2

modeled appropriately. The following regression equation with

six coefficients plus a constant is estimated:
SIPE ¼ 0:172 0:180F 11 þ0:022F 12 þ0:012F 2 0:125F 3 þ0:009F 4 0:013F 5
ð0:006Þ ð0:018Þ ð0:005Þ ð0:008Þ ð0:015Þ ð0:005Þ ð0:005Þ
.15

where F11 = (Decentralization  1.64); F12 = (Decentraliza-

tion  1.64)2  2.68; F2 = BigCity; F3 = (Fragmentation + 3.21)0.5
 1.79; F4 = Polarization; and F5 = LongTravel. The MFP algorithm
centers each covariate about its mean and finds the best fitting
.1

combination of polynomial functions to apply. In this case, since

each of the factors are standard normal distributed by design, only -4 -2 0 2 4
the covariates requiring polynomial expansions (Decentralization Low Mixing
and Fragmentation) need to be centered. Bootstrapped standard er-
rors using 10,000 resampling replications with replacement appear Fig. 6. Fractional polynomial plot of Low Mixing versus predicted SIP Efficiency.

in brackets below each coefficient, and a full table of regression re-

sults is found in Table 4.
.25

The reader will observe an extremely high level of fit indicated

by the Adjusted r2 of 94.5%. This exceptional level of explained
Partial predictor+residual of SIPE

variance suggests that the factors selected for regression nearly

completely determine the efficiency of social interaction potential
in a region. In addition to this, the Root Mean Square Error of
.2

0.0275 suggests that the model predicted values of SIP Efficiency

are very close to the true values which are bound in the (0, 1) inter-
val. This high level of linear prediction fit and strength appears
clearly in Fig. 1. Moreover, a plot of residuals versus predicted
.15

values shows no evidence of nonlinearity or heteroskedasticity

(Fig. 2). Further graphical inspection of the residuals indicates nor-
mality in their distribution. This is confirmed by tests for skewness
(p = 0.91), kurtosis (p = 0.62), and Kolmogorov–Smirnov distance
(p = 0.873). Finally a Bruesch-Pagan/Cook-Weisberg test failed to
.1

find heteroskedasticy in the residuals (p = 0.86). In sum, diagnos- -2 -1 0 1 2 3

tics of the regression model suggest that that the estimators are Long Travel
Best Linear Unbiased Estimators (BLUEs) and that inference based
on the model results is valid. Fig. 7. Fractional polynomial plot of Long Travel versus predicted SIP Efficiency.
 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277 275

be responsible for some bias in the slope estimate (Fig. 5). While
bootstrapping was used to control coefficient stability in general,
when New York is removed from the estimation sample, the
coefficient for the Big City factor increases slightly and the standard
error of the coefficient nearly doubles thus making this estimate
somewhat unreliable. For now, it suffices to say that there is no ef-
fect of city size on SIP Efficiency, once controlling for the other
covariates. Perhaps with a larger sample size this regression factor
can be better understood.
The Low Mixing coefficient is also somewhat troublesome. Not
only does the direction of effect contradict our original hypothesis
that land-use mixing should increase SIP Efficiency, the plot in
Fig. 6 suggests that this relationship is reliant on the tails of the
factor distribution rather than the majority of points near the
mean. Here again, we are less confident about this effect despite
the coefficient being significant at the 90% confidence level. Since
the sign of the coefficient contradicts our expectation, the rela-
tively low z-score for this factor is somewhat reassuring.
Finally, the partial plot of the Long Travel factor can be found in
Fig. 7. While not as tightly distributed as the Fragmentation and
Decentralization plots, we see no cause for concern in this plot as
the values of SIP Efficiency smoothly decline from low to high val-
ues of the factor score. The highest scoring cities on this factor,
Phoenix and Los Angeles, appear to have a slightly uplifting effect
on the regression coefficient, but since their inclusion in the
regression only potentially serves to dampen the effect of longer
commutes, it is clear that the true effect is at least as negative as
found by the regression model.
4.4. Cluster analysis results
Cluster analysis is used to group observations into similar clas-
ses based on the values of the five factor scores used above. When
using k-median clustering, differences between group median val-
ues are maximized while differences within groups are minimized.
The grouping structure is then used to see how mean factor values
for each cluster correspond to mean SIP and SIP Efficiency levels.
The cluster means appear in Table 5. According to a series of 1-
way ANOVAs, the cluster grouping explains a significant portion
of the variance in all the variables in Table 5 except raw SIP
(a = 0.1). In terms of SIP Efficiency, we see a clear division between
clusters 1 and 3, with very low mean efficiencies around 6.5%, and
the remainder with mean efficiencies between 18% and 31%. Clus-
ters 1 and 3 both have high mean values for the Decentralization
factor, and slightly higher means for the Big City factor, indicating
that both of these factors are related to lower levels of SIP Effi-
ciency. Interestingly, when comparing values between Clusters 1
and 3, we can observe that Cluster 3 has a mean value of raw SIP
nearly twice that of Cluster 1. Cluster 3 also has far larger mean
values for Big City, Fragmentation, and Long Commutes, indicating

Table 5
Cluster centers and member regions.

Cluster SIP SIP Efficiency (%) Decentralization Big City Fragmentation Low mixing Long travel Members
1 7526 6.5 1.43 0.11 0.26 0.43 0.86 Baltimore
Cincinnati
Cleveland
Newark
San Francisco
Philadelphia
2 10,094 18.7 0.22 0.25 0.14 0.26 0.83 Arlington
Austin
Denver
Detroit
Houston
Kansas City
Phoenix
Sacramento
San Antonio
St. Louis
Twin Cities
3 14,709 6.7 1.05 1.23 1.12 0.27 0.86 Seattle
Atlanta
Chicago
Los Angeles
New York
Washington
4 8800 31.3 0.53 0.40 1.43 0.36 0.34 Virginia Beach
Fort Lauderdale
Jacksonville
Las Vegas
Miami
Orlando
Salt Lake City
San Diego
West Palm Beach
5 9208 26.3 0.78 0.17 0.31 0.71 0.61 Boston
Buffalo
Columbus
Indianapolis
Nashville
New Orleans
Pittsburgh
Portland
Providence
Milwaukee
276 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277

that larger cities tend to have high absolute values of SIP, but tend a region. This research quantifies the relationship between sprawl
to be less efficient at providing SIP due to higher levels of fragmen- and social contact opportunity, but we don’t yet have a sufficient
tation and commute times. Thus, physical separation between measure of regional scale cohesion or social capital to empirically
inhabitants of a city tends to have a very profound effect on SIP study the relationship between SIP and these higher order
Efficiency. This is echoed by the high efficiency score of Cluster 4 outcomes.
which has a low Fragmentation mean, and the low efficiency score The second direction for empirical work is to delve into the pat-
of Cluster 2 which despite being less decentralized and not very terns of social interaction opportunities existing within the bounds
fragmented, is typified by longer commute times. of a single region. We would like to determine not only how local
transportation and land-use patterns impact social interaction po-
tential differentially throughout a region, but also how this could
5. Discussion and conclusion affect social equity, social segregation, and political polarization.
To achieve this goal we still need to answer some fundamental
The goal of this paper was to determine which aspects of urban questions like ‘‘Who within a city has more or less social interac-
spatial structure support or undermine regions’ abilities to achieve tion potential?’’, ‘‘With whom can these people easily interact?’’,
higher levels of face-to-face contact potential. This objective was and ‘‘Does the urban spatial structure reinforce or break down
achieved by computing a metric for social interaction potential regressive socio-spatial interaction opportunity patterns?’’.
and comparing this to a dataset of spatial structure characteristics
for the 42 most populous metropolitan regions in the United Acknowledgement
States. The findings are quite clear, latent factors describing urban
spatial structure nearly completely determine the levels of SIP Effi- The authors would like to acknowledge The Center for High Per-
ciency achieved in each region. In particular, Decentralization, Frag- formance Computing at the University of Utah for their technical
mentation, and Long Travel, three structural characteristics most support and allocation of processing units on their cluster
identifiable with urban sprawl, significantly reduce SIP Efficiency. computers.
But, by calculating the estimated rate of change inSIP Efficiency
at the mean value of each regressor (in this case X ¼ 0 for all X
since X  N(0, 1)), we find that a 1 standard deviation change in Appendix A. Variable descriptions
Decentralization, Fragmentation, and LongTravel causes a decline in
SIP Efficiency of 0.252, 0.035, and 0.013 percentages points respec- Some of the variables require further explanations. It will be
tively. Thus, the effect of Decentralization is more than seven times convenient to first explain that most of the variables used to
stronger than that of Fragmentation, and nearly 20 times that of describe spatial structure are derived from the origin/destination
Long Travel. This suggests that land-use policies targeting residen- matrix used in the computation of SIP. Many of the quantities
tial and employment intensification, especially at the center of the are based on the marginal row and column sums, the number of
region, will have the most positive effect on SIP, even if this causes origins and destinations for each tract. These are denoted as Oi
longer travel times due to congested roadways. and Di respectively, and let the area of an individual tract be Ai.
Previous theoretical and empirical time-use research in this
X
N
field suggests that urban sprawl and automobility are related to Area ¼ Ai
reduced levels of social contact. The empirical work in this paper i
P
demonstrates that the urban spatial structures associated with Oi
automobile oriented sprawl constrain opportunities for social con- Population Density ¼
P Area
tact. In this case study, the focus is on after-work social contact ðOi =Ai Þ
opportunities. For the sake of computational simplification, we MeanO ¼
N
have made several assumptions. We used a uniform 90 min VarianceO ¼ VARðOi =Ai Þ
time-window between 5 pm and 6:30 pm in which all workers
VarianceO
were assumed available for a discretionary activity. Of course, this CVO ¼
MeanO
P
introduces some error into our calculations provided that workers
ðDi =Ai Þ
have heterogeneous work schedules and other constraints on their MeanD ¼
time. We also assumed free-flow travel-times over the street net-
N
work, not taking into consideration mode-use or congested link
VarianceD ¼ VARðDi =Ai Þ
speeds. These assumptions lead to an overestimation of social VarianceD
CVD ¼
contact potential. However, at the regional scale of analysis, the MeanD
X
findings are generalizable since cities that have structures placing MeanET ¼ Ei =N;
more time-use constraints on individuals will consist of residents
that are more time-pressured, holding all else constant. And, as where
previous research suggests, social interaction is one of the first
types of activities to drop out of people’s daily schedules when
Ei ¼ pO lnðpO Þ  pD lnðpD Þ
faced with increased time-pressures (Farber and Páez, 2011b). Fol- Oi
pO ¼
lowing this chain of reasoning, this research provides evidence in Oi þ Di
support of the claim that sprawling metropolitan regions may Di
pD ¼
one day suffer from unforeseen social consequences as their inhab- Oi þ Di
itants become less able to participate in capital building social VarET ¼ VARðEi Þ
activities. VarET
There are two clear directions for future empirical work that CVET ¼
MeanET
arise from this study. The first is to determine whether we can
associate observed socioeconomic outcomes with social interac- Origin Density Decay Variables: Let the origin density decay func-
tion potential in a region. For example, it has been argued that tion be: ðODÞi ¼ ðODÞ0 egO T i where (OD)i is the density of origins in
sprawl reduces the overall level of social capital and cohesion in tract i, (OD)0 is the density of origins at the CBD, Ti is the travel time
 S. Farber, X. Li / Journal of Transport Geography 31 (2013) 267–277
from the CBD to tract i, and gO is the density gradient. The CBD is the
tract with the maximum number of destinations in the region.
Then:
GradientO = gO, the regression coefficient,
lnAO = ln(OD)0, the regression constant,
AdjR2O = the adjusted R2 of the regression model.
Destination Density Decay Variables: Let the destination
density decay function be: ðDDÞi ¼ ðDDÞ0 egD T i where (DD)i is the
density of destinations in tract i, (DD)0 is the density of destina-
tions at the CBD, Ti is the travel time from the CBD to tract i, and
gD is the density gradient. The CBD is the tract with the maximum
number of destinations in the region.
Then:
GradientD = gD, the regression coefficient,
lnAD = ln(DD)0, the regression constant,
AdjR2D = the adjusted R2 of the regression model.
Fractal Dimension: The urbanized area polygon data that comes
from the US census 2000 TIGER files (the detailed description of the
definition could be found at http://www.census.gov/geo/www/ua/
ua_2k.html) was clipped by the Metropolitan Planning Organiza-
tion (MPO) boundary and then transformed into raster data to
calculate the box-counting fractal dimension of the region.
The box-counting fractal dimension (FD) of the MPO is the coef-
ficient obtained by regressing the dependent variable, log(N(s))
against the independent variable, logf()(1/s), as follows:
 
1 Towards a measure of joint accessibility. Computers Environment and Urban
log NðsÞ ¼ log C þ FD log þe
s Systems 32, 331–342.
A square mesh of various sizes s = 2k(k = 0, 1, 2, . . . , n) is laid over the
city image to cover the census urbanized area shape by using N(s)
boxes. C is the total urbanized area in the shape.
References
Batty, M., Longley, P., 1994. Fractal Cities: A Geometry of Form and Function.
Academic Press, San Diego.
Beckmann, M.J., Golob, T.F., Zahavi, Y., 1983a. Travel probability fields and urban
spatial structure: 1. Theory. Environment and Planning A 15, 593–606.
Beckmann, M.J., Golob, T.F., Zahavi, Y., 1983b. Travel probability fields and urban
spatial structure: 2. Empirical tests. Environment and Planning A 15, 727–738.
Cao, X., Handy, S., Mokhtarian, P., 2006. The influences of the built environment and
residential self-selection on pedestrian behavior: evidence from Austin, TX.
Transportation 33, 1–20.
Casetti, E., 1972. Generating models by the expansion method: applications to
geographical research. Geographical Analysis 4, 81–91.
Coleman, J.S., 1988. Social capital in the creation of human capital. The American
Journal of Sociology 94, 95–120.
Dacey, M.F., 1971. Two-dimensional urban contact fields. Geographical Analysis 3,
109–120.
Domencich, T.A., McFadden, D.P., 1975. Urban Travel Demand: A Behavioral
Analysis. North Holland Publishing Co., Amsterdam.
Ewing, R., Pendall, R., Chen, D., 2003. Measuring sprawl and its transportation
impacts. Transportation Research Record 1831, 175–183.
Farber, S., Páez, A., 2009. My car, my friends, and me: a preliminary analysis of
automobility and social activity participation. Journal of Transport Geography
17, 216–225.
Farber, S., Páez, A., 2011a. Mobility without accessibility: the case of car use and
discretionary activities. In: Lucas, K., Blumenberg, E., Weinberger, R. (Eds.), Auto
Motives: Understanding Car Use Behaviours. Bingley, Emerald.
Farber, S., Páez, A., 2011b. Running to stay in place: the time-use implications of
automobile oriented land-use and travel. Journal of Transport Geography 19,
782–793.
Farber, S., Páez, A., 2012. Activity spaces and the measurement of clustering: a case
study of cross-linguistic exposure in Montreal. Environment and Planning A 44,
315–332.
277
Farber, S., Neutens, T., Miller, H.J., Li, X., in press. The social interaction potential of
metropolitan regions: a time-geographic measurement approach using joint
accessibility. Annals of the Association of American Geographers.
Forrest, R., Kearns, A., 2001. Social cohesion, social capital and the neighbourhood.
Urban Studies 38, 2125–2143.
Fotheringham, A.S., Wong, D.W.S., 1991. The modifiable areal unit problem in
multivariate statistical analysis. Environment and Planning A 23, 1025–1044.
Freeman, L., 2001. The effects of sprawl on neighborhood social ties: an explanatory
analysis. Journal of the American Planning Association 67, 69–77.
Frenkel, A., Ashkenazi, M., 2008. Measuring urban sprawl: how can we deal with it?
Environment and Planning B: Planning and Design 35, 56–79.
Golledge, R.G., Brown, L.A., Williamson, F., 1972. Behavioural approaches in
geography: an overview. Australian Geographer 12, 59–79.
Grannis, R., 1998. The importance of trivial streets: residential streets and
residential segregation. American Journal of Sociology 103, 1530–1564.
Grannis, R., 2005. T-communities: pedestrian street networks and residential
segregation in Chicago, Los Angeles, and New York. City and Community 4, 295–
321.
Grannis, R., 2009. From the Ground up: Translating Geography into Community
through Neighbor Networks. Princeton University Press, Princeton.
House, J.S., Landis, K.R., Umberson, D., 1988. Social relationships and health. Science
241, 540–545.
Leinberger, C.B., 2011. The Death of the Fringe Suburb. The New York Times, New
York ed., New York.
Levine, J., Inam, A., Torng, G.-W., 2005. A choice-based rationale for land use and
transportation alternatives. Journal of Planning Education and Research 24,
317–330.
Leyden, K.M., 2003. Social capital and the built environment: the importance of
walkable neighborhoods. American Journal of Public Health 93, 1546–1551.
Lund, H., 2003. Testing the claims of new urbanism – local access, pedestrian travel,
and neighboring behaviors. Journal of the American Planning Association 69,
414–429.
Maslow, A.H., 1943. A theory of human motivation. Psychological Review 50, 370–
396.
Moore, E.G., 1970. Some spatial properties of urban contact fields. Geographical
Analysis 2, 376–386.
Moore, E.G., Brown, L.A., 1970. Urban acquaintance fields: an evaluation of a spatial
model. Environment and Planning 2, 443–454.
Neutens, T., Schwanen, T., Witlox, F., De Maeyer, P., 2008. My space or your space?
Neutens, T., Schwanen, T., Miller, H.J., 2010. Dealing with timing and
synchronization in opportunities for joint activity participation. Geographical
Analysis 42, 245–266.
Neutens, T., Farber, S., Delafontaine, M., Boussauw, K., in press. Spatial variation in
the potential for social interaction: a case study in Flanders (Belgium).
Computers Environment and Urban Systems.
Reggiani, A., Martín, J., 2011. Guest editorial: new frontiers in accessibility
modelling: an introduction. Networks and Spatial Economics 11, 577–580.
Royston, P., Altman, D.G., 1994. Regression using fractional polynomials of
continuous covariates: parsimonious parametric modelling. Journal of the
Royal Statistical Society: Series C (Applied Statistics) 43, 429–467.
Salze, P., Banos, A., Oppert, J.M., Charreire, H., Casey, R., Simon, C., Chaix, B.,
Badariotti, D., Weber, C., 2011. Estimating spatial accessibility to facilities on
the regional scale: an extended commuting-based interaction potential model.
International Journal of Health Geographics 10.
Sander, T.H., 2002. Social capital and new urbanism: leading a civic horse to water?
National Civic Review 91, 213–234.
Sauerbrei, W., Royston, P., 1999. Building multivariable prognostic and diagnostic
models: transformation of the predictors by using fractional polynomials.
Journal of the Royal Statistical Society: Series A: Statistics in Society 162, 71–94.
Schonfelder, S., Axhausen, K.W., 2003. Activity spaces: measures of social
exclusion? Transport Policy 10, 273–286.
Sheller, M., Urry, J., 2000. The city and the car. International Journal of Urban and
Regional Research 24, 737–757.
Shen, G., 2002. Fractal dimension and fractal growth of urbanized areas.
International Journal of Geographical Information Science 16, 419–437.
Storper, M., Venables, A.J., 2004. Buzz: face-to-face contact and the urban economy.
Journal of Economic Geography 4, 351–370.
Talen, E., 1999. Sense of community and neighbourhood form: an assessment of the
social doctrine of New Urbanism. Urban Studies 36, 1361–1379.
Urry, J., 2002. Mobility and proximity. Sociology – The Journal of the British
Sociological Association 36, 255–274.
Wellman, B., 2001. Physical place and cyberplace: the rise of personalized
networking. International Journal of Urban and Regional Research 25, 227–252.
Wong, D., Shaw, S.L., 2011. Measuring segregation: an activity space approach.
Journal of Geographical Systems 13, 127–145.
Zhang, M., Kukadia, N., 2005. Metrics of urban form and the modifiable areal unit
problem. Transportation Research Record 1902, 71–79.