This article was downloaded by: [UZH Hauptbibliothek / Zentralbibliothek Zürich]
On: 18 March 2015, At: 01:09
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK
Click for updates
Cartography and Geographic Information Science
Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcag20
The impact of using social media data in crime rate
calculations: shifting hot spots and changing spatial
patterns
Nick Malleson
a
& Martin A. Andresen
b
a
School of Geography, University of Leeds, West Yorkshire, LS2 9JT, United Kingdom
b
School of Criminology, Institute for Canadian Urban Research Studies, Simon Fraser
University, 8888 University Drive, Burnaby, BC V5A 1S6 Canada
Published online: 10 Apr 2014.
To cite this article: Nick Malleson & Martin A. Andresen (2015) The impact of using social media data in crime rate
calculations: shifting hot spots and changing spatial patterns, Cartography and Geographic Information Science, 42:2,
112-121, DOI: 10.1080/15230406.2014.905756
To link to this article: http://dx.doi.org/10.1080/15230406.2014.905756
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
The impact of using social media data in crime rate calculations: shifting hot spots and changing
spatial patterns
Nick Mallesona
and Martin A. Andresenb
*
a
School of Geography, University of Leeds, West Yorkshire, LS2 9JT, United Kingdom; b
School of Criminology, Institute for Canadian
Urban Research Studies, Simon Fraser University, 8888 University Drive, Burnaby, BC V5A 1S6 Canada
(Received 28 November 2013; accepted 13 March 2014)
Crime rate is a statistic used to summarize the risk of criminal events. However, research has shown that choosing the
appropriate denominator is non-trivial. Different crime types exhibit different spatial opportunities and so does the
population at risk. The residential population is the most commonly used population at risk, but is unlikely to be suitable
for crimes that involve mobile populations. In this article, we use “crowd-sourced” data in Leeds, England, to measure the
population at risk, considering violent crime. These new data sources have the potential to represent mobile populations at
higher spatial and temporal resolutions than other available data. Through the use of two local spatial statistics (Getis-Ord
GI* and the Geographical Analysis Machine) and visualization, we show that when the volume of social media messages,
as opposed to the residential population, is used as a proxy for the population at risk, criminal event hot spots shift spatially.
Specifically, the results indicate a significant shift in the city center, eliminating its hot spot. Consequently, if crime
reduction/prevention efforts are based on resident population based crime rates, such efforts may not only be ineffective in
reducing criminal event risk, but be a waste of public resources.
Keywords: violent crime; spatial crime analysis; twitter; population at risk
Introduction
The spatially referenced crime rate is a statistic often used
to represent the risk of criminal events. Spatially referenced
crime rates help to reveal clusters of crime in space
and/or time based on an underlying population at risk.
However, the choice of an appropriate population at risk
is non-trivial. Different crime types have different spatial
opportunity sets that necessitate the separate analyses.
Similarly, the population at risk varies for different crime
rates and should be given the same consideration. As
stated by Boggs, “a valid rate … should form a probability
statement, and therefore should be based on the risk or
target group appropriate for each specific crime category”
(Boggs 1965, 900). Despite this importance, most research
uses the residential (census) population as the population
at risk, primarily because of data availability and constraints
in terms of time and money. Although it has
been claimed that it matters little which poulation at risk
is used in the analysis (Cohen, Kaufman, and Gottfredson
1985), recent research suggests that the residential population
is unsuitable as a measure of population at risk for
crimes that involve mobile victims such as assaults
(Boivin 2013); robbery (Zhang, Suresh, and Qiu 2012);
and automotive theft, burglary, and violent crime
(Andresen 2006, 2011).
In an attempt to address these limitations, our article
utilizes “crowd-sourced” data to measure the ambient
population. Specifically, we use messages from mobile
devices (such as smart phones) that are posted to Twitter.
These data have the potential to represent the ambient
population at much higher spatial and temporal resolutions
than previous research in spatial crime analysis, although
there are also considerble difficulties associated with the
data that must be overcome before they can be used by
crime analysts in earnest. The research questions are:
(1) Are crime hot spots stable under the application of
different population-at-risk measures?
(2) Which areas have the highest crime rates when
using both residential (census) and mobile (social
media) population-at-risk data?
Related work
The population at risk in crime analysis
Although a number of studies have made attempts
(Andresen 2006, 2011; Zhang, Suresh, and Qiu 2012;
Boivin 2013), it is needless to say that there is no consensus
on the appropriate way to measure the population
at risk in the scientific community (Andresen and Jenion
2010). This is partially because there are so few available
data sets at a spatial resolution that can be useful to
researchers, particularly in the context of spatial crime
analysis. Boggs (1965) is the earliest known example to
*Corresponding author. Email: andresen@sfu.ca
Cartography and Geographic Information Science, 2015
Vol. 42, No. 2, 112–121, http://dx.doi.org/10.1080/15230406.2014.905756
© 2014 Cartography and Geographic Information Society
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
systematically show the impact of using different populations-at-risk
measures in crime rate calculations. She considered
the business/residential land use ratio for business
crime, parking space availability for vehicle theft, and
sidewalk area (as a proxy for pedestrians) for street robbery.
In her subsequent analysis, Boggs (1965) found that
her alternative populations at risk mattered a lot for some
crime types and very little for other crime types. More
recently, Andresen and colleagues have used the
LandScan Global Population Database as the population
at risk (Andresen 2006, 2011; Andresen and Jenion
2010; Andresen, Jenion, and Reid 2012). The LandScan
data provide an estimate of the ambient population, on a
global scale, at a spatial resolution of approximately 1 km2
– this area varies with the distance from the equator.
Though largely instructive, there are limitations with
these data: (1) the spatial resolution is relatively poor for
spatial crime analysis (approximately the size of a census
tract) because recent research has shown that analyzing
crime at scales greater than the street segment may hide
important lower-level patterns (Andresen and Malleson
2011); and (2) the ambient population estimate is a yearly
average, such that no account is taken for seasonal variations
or the differences in population counts at different
times of day. In an attempt to allieviate some of these
problems, this article will use data contributed by individuals
to social media services to estimate ambient population
at risks.
Social media data for mobile populations
In recent years, the emergence of vast new administrative
and commercial data sources, coupled with warnings
about a “crisis” in an empirical sociology that continued
to rely entirely on traditional small studies (Savage and
Burrows 2007), has spurred some research to engage with
new forms of “crowd-sourced” data to gain insight into
social processes. These data, commonly contributed informally
by citizens rather than being obtained from a formal
survey, are becoming ubiquitous and will undoubtedly
have a dramatic impact on future social science research.
With respect to population dynamics in particular, traditional
large-volume social science data lack information
regarding where people are throughout the day, and
instead represent the night time distribution of the population.
A benefit of new forms of crowd-sourced data, and
social media in particular, is that new technologies enable
researchers to capture large volumes of information
regarding peoples’ daily behavior. This may prove to be
instructive for understanding urban dynamics and developing
more accurate population-at-risk estimates. And in
the context of this article, such data may prove to be
useful for spatial crime analysis.
The number of sources for such data is increasing,
with the more widely used being Twitter, mobile device
data from service providers, public transport usage,
Foursquare, Flickr, and Facebook. Research in the
United States has found that two-thirds of online adults
(66%) use social media platforms (Smith 2011) and that
26% of American Internet users aged 18–29 have been
found to use Twitter (Smith and Brenner 2012). Data from
these sources are also voluminous. For example, there
were supposedly over 100 million active Twitter accounts
in 2011 (Twitter 2011) and 270,000 tweets per minute
produced worldwide in 2012 (TechCrunch 2012).
Social media data have recently been used for a wide
variety of different purposes – a full review of applications
would be an extensive undertaking (and one that would be
outdated before it is published). However, examples of the
application of social media data to the study of social
phenomena are more limitted. Examples include research
into the fear of missing out (Przybylskia et al. 2013), wellbeing
(Hong et al. 2012), and happiness over time (Bliss
et al. 2012). Others make some limitted use of the data,
but still resort to traditional sampling methods (see, for
example, Fischer and Reuber 2011; Wohn et al. 2013).
The most relevant research for this project are those that
have started to make use of the geographical locations of
social media messages, although given the novelty of
utilizing these data sources, examples are still rare.
Relevant research includes: the mathematical analysis of
human mobility patterns (Cheng et al. 2011); the development
of neighborhood boundaries based on the characteristics
of those who commonly frequent them (Cranshaw
et al. 2012); the identification of events such as earthquakes
(Crooks et al. 2013) and other geographical patterns
(Stefanidis, Crooks, and Radzikowski 2013) in social
media data; and the use of Google search trends to estimate
the locations of new outbreaks of influenza
(Ginsberg et al. 2009). However, we are unaware of any
research that uses social media data to better understand
the risk of criminal victimization.
Despite their relatively widespread (and increasing)
use, these data sources do have limitations. Such data
are inherently “messy” in the sense that they are not
gathered using a systematic and statistically guided methodology
such as a census. As a result, data structures may
be poorly defined, missing data are commonplace, and
there are no systematic “corrections” for these issues
because these data are still so new to research.
Additionally, because of these issues, we must also be
concerned with generalizability. For example, Li,
Goodchild, and Xu (2013), as part of a special issue on
mapping cyberspace and social media, found that higher
socioeconomic status groups are overrepresented in
Twitter and Flickr. This is not inherently problematic,
particularly in the current context of measuring populations
at risk, because these higher socioeconomic groups
may be representative of the underlying population distribution,
on average. The main difficulty arises in testing
Cartography and Geographic Information Science 113
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
such a hypothesis. However, even if only a portion of the
actual population is being captured, the bias inherent in
residential populations for measuring the population at
risk may be reduced.
Study area and data overview
Leeds and the census data
Our study area is Leeds, United Kingdom (UK). The
Leeds local authority district is the third largest in the
UK (behind London and Birmingham) with a residential
population estimated at 757,655 in 2012 (Office for
National Statistics 2012). Leeds has a central business
and retailing district with a high concentration of shops,
businesses, and entertainment facilities. This district
attracts large volumes of people from within Leeds,
Bradford, Manchester, and a number of smaller towns/
villages on the outskirts of the city. Such areas have long
been known to have high levels of crime because they
attract large volumes of people (Schmid 1960a, 1960b)
and the center of Leeds is no exception; the district has
high volumes of violent crime relative to surrounding
areas. Related to the alternative population-at-risk literature
in spatial crime analysis, relatively few people live in
the city center, upwardly biasing any representations of
criminal event risk using the resident population.
In order to measure the residential population, we have
used the number of people residing in each Output Area
(OA) at the time of the 2011 UK census. The OA geography
is the smallest area for which census statistics are
released. Each OA has a recommended size of 125 households,
but can vary based on natural boundaries and the
presence (or absence) or high-density housing.
Crime data
The criminal event data used in the analyses below include
all individual occurrences of violent crime in 2011 within
the Leeds Local Authority District (N = 10,625) that were
reported to the police. These data were obtained from the
police.uk service (http://www.police.uk); all policerecorded
criminal events in England and Wales have
been available to the public since December 2011,
although only 44% of violent crimes were made known
to the police (Flatley 2013a). “Violent crime” includes a
variety of crime types ranging from minor assaults to
serious incidents of wounding and murder (Flatley
2013b). A drawback with these data is that it is not
possible to disaggregate the crime type further (for example,
it might be advantageous to analyze robbery and
assault separately as research has shown that the spatial
patterns of spcific crime types can be rather different
(Andresen and Linning 2012)). For privacy reasons, the
police.uk service aggregates individual crime points to the
nearest “anonymous map point” that can be the center of a
street segment, a public place, or a commercial building.
These points are defined with catchment areas that have at
least eight unique postal addresses, approximately the size
of a city block. Although such an aggregation process
inevitably induces some spatial inaccuracy, the impact is
unlikely to influence any results because the direction in
which the criminal event points are moved is random in
the aggregate. Also, because Leeds is a rather densely
populated city, it is unlikely that any individual criminal
event points will be displaced far from their actual location.
Additionally, we could disaggregate the data temporally,
which is an obvious application of social media data
because of the availability of the time when messages are
posted. We do not undertake such an analysis, and leave it
for future research, because the first comparison in the
spatial crime analysis context is with how crime data are
mapped in the majority of research, an aggregated year.
Social media data
The data used in the current article are messages posted to
the Twitter service from within the Leeds local authority
district, 22 June 2011 to 14 April 2013. Although there are
other social-media services that provide publicly available
access to user contributed data (such as Flickr and
Foursquare), data for this study originate solely from
Twitter. Future work will explore the possibility of including
a variety of sources (e.g., Stefanidis, Crooks, and
Radzikowski 2013); currently Twitter is by far the most
widely used service and it is not clear that the incorporation
of additional services is necessary in this application.
Because we are interested in the spatial dimension of
criminal victimization risk, only messages with associated
GPS coordinates have been included. Such data are commonly
generated using mobile devices by users who have
explicitly opted to publish their present location. A manual
inspection of the data revealed that many high-volume
accounts were not representing individuals (examples
include weather forecasts, car advertisements, etc.). After
deleting these data, the number of messages in our sample
was almost 2 million, N = 1,955,655. In addition to the
location, each individual message contains information
regarding the user account, the text itself, and the time
of the message. These additional fields allow for the creation
of a temporally dynamic population at risk or an
exploration of the characteristics of the individuals who
make up the general population. Both of these factors
could lead to even more accurate risk estimates, although
this is not under investigation here and is a direction for
future research.
The density of the messages overlaid with violent
crime hot spots is shown in Figure 1.1
As would be
expected, message densities are greatest in urban areas
and particularly in the city center. This is precisely what
114 N. Malleson and M.A. Andresen
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
would be expected, based on what we know regarding the
ambient population. Consequently, as hypothesized above,
despite not having a representative sample of individuals
based on socioeconomic status, based on local knowledge
of the study area, these data may appear to be representative
of where people actually are. And, of great interest for
the current article, the largest densities of messages appear
to coincide with the greatest densities of violent crime –
this is not the case with the resident population.
Methods and results
The aim of this article is to highlight the areas that suffer
high rates of crime, using both residential and mobile
population-at-risk estimates. To answer this question, the
research will apply two complementary statistics that can
be used to identify clusters in spatial data. Both search for
clusters of crime by comparing volumes in individual
areas to their surrounding neighbors and to global
averages. They are known as Local Indicators of Spatial
Association (LISA) and offer the advantage of testing for
statistical significance of apparent clusters – see Anselin
(1995) for a discussion of LISA statistics. Both statistics
will be used to search for statistically significant crime hot
spots using census data and social media data as the
populations at risk.
Statistic 1: Getis–Ord GI*
The first statistic to be applied is the Getis–Ord GI* (Getis
and Ord 1992; Ord and Getis 1995). This is used here
because its definition closely matches that of a “hot spot”
– local area averages that are significantly greater than
global averages (Chainey and Ratcliffe 2005) – and has
hence become popular within spatial criminological
research. We use first-order queen’s contiguity in the analyses
below.
Figure 2 maps the GI* indices for the two violent
crime rates. Output areas with insignificant p values
(0.05 < p < 0.95) are not shown, regardless of their Z
value. The distribution of significant GI* scores proves to
be instructive. When considering the residential violent
crime rate, there is a statistically significant cluster in the
city center as well as in some of the surrounding neighborhoods.
The violent crime cluster in the city center is
expected, particularly because of the low residential population
and large volume of criminal events. The surrounding
neighborhoods that exhibit clusters of violent crime
Figure 1. Kernel density of social media messages and violent crime contours. The contours depict the areas with the largest volume of
violent crime (densities of 600 and 1400 crimes per km2
, respectively, obtained using Kernel Density Estimation).
Cartography and Geographic Information Science 115
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
largely consist of industrial estates that also have a low
population density. The most notable exceptions are violent
crime clusters surrounding a large hospital (St.
James’s University Hospital) to the north-east and two
small areas in neighborhoods to the south-west. It should
be noted, however, that the violent crime cluster
Figure 2. GI* Z values for crime rates (using ambient and residential population denominators) in Leeds.
116 N. Malleson and M.A. Andresen
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
surrounding the university hospital may simply be a
reporting issue: violent criminal events are coded to
occur at this location because this is where they are
reported.
A number of violent clusters emerge when using the
ambient violent crime rate (see Figure 2 insets A–D).
Curiously, none of the violent crime clusters includes the
city center area, suggesting the violent crime rate there is
not significant when using the ambient population to measure
the population at risk. Rather, the violent crime clusters
are in diverse neighborhoods with no obvious single
explanation for their existence. Each of these neighborhoods
may have high violent crime rates given the size of
the population at risk. This is clearly a direction for further
research.
A drawback with the GI* statistic is that it requires the
spatial aggregation of point data into areas (output areas in
this case). Therefore, it is susceptible to the modifiable
areal unit problem (Openshaw 1984). Hence a second
statistic is also used that avoids aggregation to the output
area geography in order to further assess the differences in
the two violent crime rate calculations.
Statistic 2: the Geographical Analysis Machine
The Geographical Analysis Machine (GAM) (Openshaw
1987) is an algorithm originally developed during research
investigating child leukemia cases near a nuclear reactor
(Openshaw, Charlton, and Craft 1988). However, GAM
has also been applied to research areas such as food
poverty (Farrow et al. 2005) and the analysis of crime
clusters (Corcoran, Wilson, and Ware 2003). The clustering
algorithm operates by iterating over a set of distinct
search points that form a regular grid and then calculating
the concentrations of events within a given radius of each
point. For all search points, i, the algorithm calculates the
number of expected events, ei, standardizing against the
underlying background population:
ei ¼
∑0
∑p
!
pi; (1)
where ∑0 is the total number of observations (crimes), ∑p
is the size of the base population (number of residents or
number of messages), and pi is the size of the base
population within search circle i. Then the difference, di,
is calculated using the actual number of observations, ai,
and the expected number of observations, ei, that occur
within circle i:
di ¼ ai À ei: (2)
If a larger number of cases are found than would be
expected (di > 0), a Poisson test for statistical significance
is performed. The test calculates the probability that the
number of observed events is the same as the number of
expected events (d = 0). If this probability is lower than a
set threshold – in this case the threshold is 0.0099 – then
the null hypothesis is rejected and the difference is statistically
significant at the specified threshold. In these cases
the search circle is stored as a potential cluster. The GAM
output is a list of search points and the difference between
the expected and actual number events (di) when di is
statistically significant.
This algorithm has been chosen to comlement the GI*
analysis because, importantly, it minimizes the impact of
the modifiable areal unit problem by defining arbitrary
search locations on a regular grid and also by varying
the search radius for each search point. In this manner,
clusters that appear at one resolution can be discarded if
they disappear at others. A further advantage of the GAM
algorithm is that it will process raw point data directly –
spatial aggregation is not a prerequisite.
In the following, multiple analyses were run with the
search radii being increased in 100 m increments from
200 m to 1 km. All significant search points at all radii
were used to generate a single density map. The difference
between the expected and actual numbers of crimes at
each search point (i.e., the output of the algorithm) was
used to calculate the density. In this manner, the most
dense areas will be those that have a large difference at
multiple resolutions. Clusters that are only significant at a
small number of search radii will add marginally to the
density of their area. The results are mapped in Figure 3.
The first notable result is that the GAM outputs are
largely in agreement with those of the GI* analysis. Both
techniques reveal broadly similar cluster locations regardless
of the population at risk used. Considering the number
of social media messages, the large volume of violent
crime in the city center is only marginally higher than
would be expected given the ambient population. In
other words, the risk of violent criminal victimization is
not particularly high at the city center. However, the algorithms
both identify violent crime clusters in neighborhoods
to the north- and south-east regardless of the
population at risk used. The consistency with which
these areas have been identified as crime hot spots suggests
that they are indicitative of an exceptionally high
volume of crime, whereas the city center hot spot is more
likely to be an artifact of the size of the ambient
population.
Discussion and conclusions
In this analysis, we have shown that different spatial
patterns of crime rates emerge when using two different
population-at-risk measures: the residential population
(measured by the 2011 UK census) and the ambient population
(measured by counting the number of messages
Cartography and Geographic Information Science 117
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
Figure 3. Clusters of violent crime calculated using the Geographical Analysis Machine with ambient and residential population at risk.
“Cluster strength” is the sum of all significant search circles at all radii from 200 m to 1 km.
118 N. Malleson and M.A. Andresen
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
posted to the Twitter social media service). One may say
that such a conclusion is an obvious one, but it is important
to recognize that the use of an ambient population
measure is justified by theory as well as previous empirical
research despite the widespread use of the residential
population in geography of crime literature. Perhaps most
striking are the results from the Leeds city center. Though
this area has a large volume of violent criminal events, it
does not exhibit a statistically significant rate when the
ambient population is used to measure the population at
risk. Consequently, despite the high volume of violent
criminal events, there is not a statistically significant elevation
in risk of violent criminal victimizaton when considering
a theoretically informed population at risk. No
such conclusion would have been reached with the residential
population.
Additionally, there are a small number of neighborhoods
very close to the city center that exhibit significantly
high violent crime rates when considering both
populations at risk, regardless of the clustering method.
There is no obvious reason for such high rates of violent
crime. These neighborhoods score rather high on the
deprivation scale, with two of the neiborhorhoods scoring
114 and 128, highest in England out of a total of 32,482
neighborhoods. Given that deprivation is a highly complex
phenomenon, considering a multitude of social factors,
it may be the case that this plays some role through (a
lack of) oppportunity in terms of legitimate activities for
residents social tension that leads to violence. This is
clearly an area of future research interest as well.
Though we have had some interesting, and theoretically
expected, results, our analysis is not without its
limitations. Most specifically, we must be cautious with
the use of Twitter data and making generalizations about
general population movements. How well do the spatial
locations of social media messages reflect the actual spatial
locations of the ambient population, in general? We
know that some socioeconomic groups are overrepresented
in these data, but is this necessarily a problem?
Also, to what extent does multiple-counting (users of
Twitter who frequently tweet) bias the spatial distribution
of the population at risk? These users may simply tweet in
locations where there are more people anyway, not causing
any spatial bias, or they may make it appear as though
more people are present than actually are present.
Additionally, despite the user rates of social media are
increasing, the percentage of messages that include accurate
geogrpahic information are as low as 1–2% (Leetaru
et al. 2013; Gelernter and Mushegian 2011). Finally, there
is the potential for participation inequality stemming from
the differences in the prevalence of social media useage
across different social groups. A body of work has
explored the impacts of the “digital divide” (e.g., Yu
2006; Fuchs 2008) and it is possible that the higher
crime rates identified in the north-east and south-west
neighborhoods are an artifact of lower Twitter usage in
these relatively deprived communities. However, it is not
clear how well general trends in digital access are reflected
in Twitter usage – further research is required to establish
whether or not the ambient population in these neighborhoods
is poorly represented by Twitter data. The persistance
of the hot spots regardless of the population at risk
used here does, however, add strength to the results.
In general, there are potential problems that must be
investigated for the appropriate use of crowd-sourced data.
However, if they can be resolved, there is great potential,
particularly for spatial crime analysis. For example,
Twitter data, or social media data more generally, could
be used to estimate particular sub-populations at risk of
particular crime types such as young people who visit bars
during the evening. Therefore, the population at risk could
be tailored according the the most likely victims of a
particular crime category to answer the call made by
Boggs (1965) almost 50 years ago: “the risk or target
group appropriate for each specific crime category”
(Boggs 1965, 900).
As discussed by Savage and Burrows (2007), the
social sciences (spatial or not) must embrace these new
forms of data that, although messy, biased and noisy, have
the potential to describe social phenomena better than
well-organized small surveys or even national censuses.
Mayer-Schonberger and Cukier (2013) share this view:
One of the areas that is being most dramatically shaken
up by N = all is the social sciences. They have lost their
monopoly of making sense of empirical social data, as
big data analysis replaces the highly skilled survey specialists
of the past… When data are collected passively
while people do what they normally do anyway, the old
biases associated with sampling and questionnaires
disappear. (30)
We are confident that the messy, biased and noisy aspects
of big data will soon be reduced for confident use in the
social sciences. Though they may not disappear or be at
the same low level as with more formal data gathering
techniques, these limitations may simply become outweighed
by the sheer volume of crowd-sourced data and
the ways in which it can be utilized. We were able to
obtain nearly 2 million individual datum with a minimal
setup time and negligible financial cost. Also, with
increased use and demand for such data, the providers of
social media may very well enhance the quality of their
data and metadata because they will realize the value of
their commodity. We have argued above that its utility is
significant for spatial crime analysis.
Future research in the area of spatial crime analysis
has a number of obvious directions. The most obvious is
to disentangle these data by day/night, weekday/weekend
(or simply day of week), and so on. For this to be successful,
a more nuanced definition of the crime type than that
Cartography and Geographic Information Science 119
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
provided by the police.uk data will be necessary – individual
police forces do capture these data and might make it
available for research purposes. This would allow for the
identification of theoretically informed crime rates to be
used for clusters in space and time. Additionally, with the
possibility of linking social media users back to their
home census geography unit, we could generate a profile
from those census data of populations at risk in different
locations. But of course, such research necessarily
involves a new set of ethical implications that have yet
to be properly addressed. However, if these ethical issues
can be overcome and the public can see the social benefits
that may emerge from this research, we may be able to
significantly advance our knowledge of the spatial patterns
of crime.
Note
1. The density per unit area is used in order to facilitate
subsequent comparisons in the paper. Violent crime contours
are present in order to show the overlap of violent
crime with the messages.
References
Andresen, M. A. 2006. “Crime Measures and the Spatial
Analysis of Criminal Activity.” British Journal of
Criminology 46 (2): 258–285. doi:10.1093/bjc/azi054.
Andresen, M. A. 2011. “The Ambient Population and Crime
Analysis.” The Professional Geographer 63 (2): 193–212.
doi:10.1080/00330124.2010.547151.
Andresen, M. A., and G. W. Jenion. 2010. “Ambient Populations
and the Calculation of Crime Rates and Risk.” Security
Journal 23 (2): 114–133. doi:10.1057/sj.2008.1.
Andresen, M. A., G. W. Jenion, and A. A. Reid. 2012. “An
Evaluation of Ambient Population Estimates for Use in
Crime Analysis.” Crime Mapping: A Journal of Research
and Practice 4 (1): 7–30.
Andresen, M. A., and S. J. Linning. 2012. “The (In)
Appropriateness of Aggregating across Crime Types.”
Applied Geography 35 (1–2): 275–282. doi:10.1016/j.
apgeog.2012.07.007.
Andresen, M. A., and N. Malleson. 2011. “Testing the Stability
of Crime Patterns: Implications for Theory and Policy.”
Journal of Research in Crime and Delinquency 48 (1):
58–82. doi:10.1177/0022427810384136.
Anselin, L. 1995. “Local Indicators of Spatial Association –
LISA.” Geographical Analysis 27 (2): 93–115.
doi:10.1111/j.1538-4632.1995.tb00338.x.
Bliss, C. A., I. M. Kloumann, K. D. Harris, C. M. Danforth, and
P. S. Dodds. 2012. “Twitter Reciprocal Reply Networks
Exhibit Assortativity with Respect to Happiness.” Journal
of Computational Science 3 (5): 388–397. doi:10.1016/j.
jocs.2012.05.001.
Boggs, S. L. 1965. “Urban Crime Patterns.” American
Sociological Review 30 (6): 899–908. doi:10.2307/2090968.
Boivin, R. 2013. “On the Use of Crime Rates.” Canadian
Journal of Criminology and Criminal Justice/La Revue
Canadienne De Criminologie Et De Justice Pénale 55 (2):
263–277. doi:10.3138/cjccj.2012-E-06.
Chainey, S., and J. H. Ratcliffe. 2005. GIS and Crime Mapping.
Chichester: John Wiley and Sons.
Cheng, Z., J. Caverlee, K. Lee, and D. Z. Sui. 2011. “Exploring
Millions of Footprints in Location Sharing Services.” In
Proceedings of the Fifth International AAAI Conference on
Weblogs and Social Media (ICWSM), Barcelona, July, 81–
88. Menlo Park, CA: AAAI press.
Cohen, L. E., R. L. Kaufman, and M. R. Gottfredson. 1985.
“Risk-Based Crime Statistics: A Forecasting Comparison for
Burglary and Auto Theft.” Journal of Criminal Justice 13
(5): 445–457. doi:10.1016/0047-2352(85)90044-3.
Corcoran, J. J., I. D. Wilson, and J. Ware. 2003. “Predicting the
Geo-Temporal Variations of Crime and Disorder.”
International Journal of Forecasting 19 (4): 623–634.
doi:10.1016/S0169-2070(03)00095-5.
Cranshaw, J., R. Schwartz, J. Hong, and N. Sadeh. 2012. “The
Livehoods Project: Utilizing Social Media to Understand the
Dynamics of A City.” In Proceedings of the Sixth
International AAAI Conference on Weblogs and Social
Media (ICWSM), Dublin, May, 58–65. Menlo Park, CA:
AAAI Press.
Crooks, A., A. Croitoru, A. Stefanidis, and J. Radzikowski.
2013. “#Earthquake: Twitter As A Distributed Sensor
System.” Transactions in GIS 17 (1): 124–147.
doi:10.1111/j.1467-9671.2012.01359.x.
Farrow, A., C. Larrea, G. Hyman, and G. Lema. 2005.
“Exploring the Spatial Variation of Food Poverty in
Ecuador.” Food Policy 30 (5–6): 510–531. doi:10.1016/j.
foodpol.2005.09.005.
Fischer, E., and A. R. Reuber. 2011. “Social Interaction Via
New Social Media: (How) Can Interactions on Twitter
Affect Effectual Thinking and Behavior?” Journal of
Business Venturing 26 (1): 1–18. doi:10.1016/j.
jbusvent.2010.09.002.
Flatley, J. 2013a. Focus On: Violent Crime and Sexual Offences,
2011/12. London: Office for National Statistics.
Flatley, J. 2013b. Crime in England and Wales, Year Ending
September 2012. London: Office for National Statistics.
Fuchs, C. 2008. “The Role of Income Inequality in A
Multivariate Cross-National Analysis of the Digital
Divide.” Social Science Computer Review 27: 41–58.
doi:10.1177/0894439308321628.
Gelernter, J., and N. Mushegian. 2011. “Geo-Parsing Messages
from Microtext.” Transactions in GIS 15 (6): 753–773.
Getis, A., and J. K. Ord. 1992. “The Analysis of Spatial
Association by Use of Distance Statistics.” Geographical
Analysis 24 (3): 189–206. doi:10.1111/j.1538-4632.1992.
tb00261.x.
Ginsberg, J., M. H. Mohebbi, R. S. Patel, L. Brammer, M. S.
Smolinski1, and L. Brilliant. 2009. “Detecting Influenza
Epidemics Using Search Engine Query Data.” Nature 457:
1012–1014. doi:10.1038/nature07634.
Hong, L., A. Ahmed, S. Gurumurthy, A. Smola, and T. Kostas.
2012. “Discovering Geographical Topics in the Twitter
Stream.” Proceedings of the 21st International Conference
on World Wide Web, Lyon, 769–778.
Leetaru, K., S. Wang, A. Padmanabhan, and E. Shook. 2013.
“Mapping the Global Twitter Heartbeat: the Geography
of Twitter.” First Monday 18 (5). doi:10.5210/fm.
v18i5.4366.
Li, L., M. F. Goodchild, and B. Xu. 2013. “Spatial, Temporal,
and Socioeconomic Patterns in the Use of Twitter and
Flickr.” Cartography and Geographic Information Science
40 (2): 61–77. doi:10.1080/15230406.2013.777139.
120 N. Malleson and M.A. Andresen
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015
Mayer-Schonberger, V., and K. Cukier. 2013. Big Data: A
Revolution That Will Transform How We Live, Work and
Think. London: John Murray.
Office for National Statistics. 2012. Mid-2012 Population
Estimates. Accessed November 28, 2013. http://www.ons.
gov.uk/ons/rel/pop-estimate/population-estimates-for-england-
and-wales/mid-2012/mid-2012-population-estimates-for-eng
land-and-wales.html
Openshaw, S. 1984. The Modifiable Areal Unit Problem.
Concepts and Techniques in Modern Geography
(CATMOG). Vol. 38. Norwich: Geo Books.
Openshaw, S. 1987. “An Automated Geographical Analysis
System.” Environment and Planning A 19 (4): 431–436.
Openshaw, S., M. Charlton, and A. Craft. 1988. “Searching for
Leukaemia Clusters Using A Geographical Analysis
Machine.” Papers in Regional Science 64 (1): 95–106.
doi:10.1111/j.1435-5597.1988.tb01117.x.
Ord, J. K., and A. Getis. 1995. “Local Spatial Autocorrelation
Statistics: Distributional Issues and An Application.”
Geographical Analysis 27 (4): 286–306. doi:10.1111/
j.1538-4632.1995.tb00912.x.
Przybylskia, A. K., K. Murayamab, C. R. DeHaanc, and V.
Gladwelld. 2013. “Motivational, Emotional, and Behavioral
Correlates of Fear of Missing Out.” Computers in Human
Behavior 29 (4): 1841–1848. doi:10.1016/j.chb.2013.02.014.
Savage, M., and R. Burrows. 2007. “The Coming Crisis of
Empirical Sociology.” Sociology 41 (5): 885–899.
doi:10.1177/0038038507080443.
Schmid, C. F. 1960a. “Urban Crime Areas: Part I.” American
Sociological Review 25 (4): 527–542. doi:10.2307/2092937.
Schmid, C. F. 1960b. “Urban Crime Areas: Part II.” American
Sociological Review 25 (5): 655–678. doi:10.2307/2090139.
Smith, A. 2011. Why Americans use social media. Technical
report, Pew Research Centre. Accessed November 28,
2013. http://www.pewinternet.org/Reports/2011/Why-
Americans-Use-Social-Media.aspx
Smith, A., and J. Brenner. 2012. Twitter Use 2012. Technical
report, Pew Research Center. Accessed November 28,
2013. http://pewinternet.org/Reports/2012/Twitter-Use-
2012.aspx
Stefanidis, A., A. Crooks, and J. Radzikowski. 2013.
“Harvesting Ambient Geospatial Information from Social
Media Feeds.” Geojournal 78: 319–338. doi:10.1007/
s10708-011-9438-2.
TechCrunch. 2012. “Analyst: Twitter Passed 500M Users In June
2012.” Accessed January 19, 2013. http://techcrunch.com/
2012/07/30/analyst-twitter-passed-500m-users-in-june-2012-
140m-of-them-in-us-jakarta-biggest-tweeting-city/
Twitter. 2011. “One Hundred Million Voices.” Twitter Blog.
Accessed January 2014. https://blog.twitter.com/2011/one-
hundred-million-voices
Wohn, D. Y., N. Ellison, M. L. Khan, R. Fewins-Bliss, and R.
Gray. 2013. “The Role of Social Media in Shaping FirstGeneration
High School Students’ College Aspirations: A
Social Capital Lens.” Computers & Education 63: 424–436.
doi:10.1016/j.compedu.2013.01.004.
Yu, L. 2006. “Understanding Information Inequality: Making
Sense of the Literature of the Information and Digital
Divides.” Journal of Librarianship and Information
Science 38: 229–252. doi:10.1177/0961000606070600.
Zhang, H., G. Suresh, and Y. Qiu. 2012. “Issues in the
Aggregation and Spatial Analysis of Neighborhood Crime.”
Annals of GIS 18 (3): 173–183. doi:10.1080/
19475683.2012.691901.
Cartography and Geographic Information Science 121
Downloadedby[UZHHauptbibliothek/ZentralbibliothekZürich]at01:0918March2015