ArticlePDF Available

Abstract and Figures

Case linkage uses crime scene behaviours to identify series of crimes committed by the same offender. This paper tests the underlying assumptions of case linkage (behavioural consistency and behavioural distinctiveness) by comparing the behavioural similarity of linked pairs of offences (i.e. two offences committed by the same offender) with the behavioural similarity of unlinked pairs of offences (i.e. two offences committed by different offenders). It is hypothesised that linked pairs will be more behaviourally similar than unlinked pairs thereby providing evidence for the two assumptions. The current research uses logistic regression and receiver operating characteristic analyses to explore which behaviours can be used to reliably link personal robbery offences using a sample of 166 solved offences committed by 83 offenders. The method of generating unlinked pairs is then refined to reflect how the police work at a local level, and the success of predictive factors re‐tested. Both phases of the research provide evidence of behavioural consistency and behavioural distinctiveness with linked pairs displaying more similarity than unlinked pairs across a range of behavioural domains. Inter‐crime distance and target selection emerge as the most useful linkage factors with promising results also found for temporal proximity and control. No evidence was found to indicate that the property stolen is useful for linkage. Copyright © 2012 John Wiley & Sons, Ltd.
Content may be subject to copyright.
Linking Personal Robbery Offences Using Offender Behaviour
AMY BURRELL
1,
*, RAY BULL
1
and JOHN BOND
2
1
School of Psychology, University of Leicester, Henry Wellcome Building, Lancaster Road, Leicester
LE1 9HN, UK
2
Department of Chemistry, University of Leicester, Leicester LE1 7RH, UK
Abstract
Case linkage uses crime scene behaviours to identify series of crimes committed by the same
offender. This paper tests the underlying assumptions of case linkage (behavioural consistency
and behavioural distinctiveness) by comparing the behavioural similarity of linked pairs of
offences (i.e. two offences committed by the same offender) with the behavioural similarity of
unlinked pairs of offences (i.e. two offences committed by different offenders). It is hypothesised
that linked pairs will be more behaviourally similar than unlinked pairs thereby providing
evidence for the two assumptions. The current research uses logistic regression and receiver
operating characteristic analyses to explore which behaviours can be used to reliably link
personal robbery offences using a sample of 166 solved offences committed by 83 offenders.
The method of generating unlinked pairs is then renedtoreect how the police work at a local
level, and the success of predictive factors re-tested. Both phases of the research provide
evidence of behavioural consistency and behavioural distinctiveness with linked pairs display-
ing more similarity than unlinked pairs across a range of behavioural domains. Inter-crime
distance and target selection emerge as the most useful linkage factors with promising results
also found for temporal proximity and control. No evidence was found to indicate that the
property stolen is useful for linkage. Copyright © 2012 John Wiley & Sons, Ltd.
Key words: case linkage; behavioural analysis; robbery; mugging; street crime
INTRODUCTION
Case linkage
Identifying serial crime and linking these offences to a single offender is a crucial part of police
work (Bennell, Jones, & Melnyk, 2009; Santtila, Fritzon, & Tamelander, 2004; Sorochinski &
Salfati, 2010; Yokota & Watanabe, 2002). Establishing that a number of offences are attribut-
able to the same person supports the implementation of efcient and productive investigative
*Correspondence to: Amy Burrell, School of Psychology, University of Leicester, Henry Wellcome Building,
Lancaster Road, Leicester, LE1 9HN, UK.
E-mail: amb58@le.ac.uk
Journal of Investigative Psychology and Offender Proling
J. Investig. Psych. Offender Prol. (2012)
Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/jip.1365
Copyright © 2012 John Wiley & Sons, Ltd.
strategies (Labuschagne, 2012; Santtila, Junkkila, & Sandnabba, 2005), for example, pooling
information from all the crime scenes (Bennell et al., 2009), which can lead to faster identi-
cation and apprehension of the offender and/or strengthening the evidential case for court
(Woodhams, Bull, & Hollin, 2007a). Linking offences can be relatively simple if forensic
and/or physical evidence is found at the scene (Grubin, Kelly, & Brunsdon, 2001). However,
forensic evidence is often lacking (Ewart, Oatley, & Burn, 2005; Hazelwood & Warren, 2003).
When this is the case, behavioural analysis may be used to identify a linked series of offences
(Bennell & Jones, 2005; Grubin et al., 2001; Hazelwood & Warren, 2003; Woodhams et al.,
2007a; Woodhams & Toye, 2007). This is known as case linkage.
The central assumptions of case linkage are behavioural consistency and behavioural
distinctiveness. The offender consistency hypothesis (Canter, 1995) postulates that offenders
will behave consistently across their crimes, and behavioural distinctiveness assumes that the
way an offender behaves is heterogeneous from the way in which other offenders commit
crime (Goodwill & Alison, 2006; Salfati & Bateman, 2005). Taken together, these assump-
tions allow individual offendersoffences to be (1) linked together and (2) distinguished from
offences committed by other offenders. Evidence of behavioural consistency has been found
for a range of offence types including sexual assault (Grubin et al., 2001), homicide (Salfati &
Bateman, 2005), arson (Santtila et al., 2004), burglary (Bennell & Canter, 2002; Bennell &
Jones, 2005; Markson, Woodhams, & Bond, 2010), vehicle crime (Tonkin, Grant, & Bond,
2008), and commercial robbery (Woodhams & Toye, 2007). Evidence has also been found
in support of behavioural distinctiveness (Bennell & Canter, 2002; Bennell & Jones, 2005;
Bennell, Gauthier, Gauthier, Melnyk, & Musolino, 2010; Grubin et al., 2001; Santtila,
et al., 2004; Santtila, et al., 2005; Woodhams & Toye, 2007).
Robbery
Robbery is the theft of property using force or the threat of force. There are two categories of
robbery in the UK: commercial and personal. Commercial robberies are committed against
businesses (e.g. a bank robbery with the violence directed at bank employees), whereas
personal offences are committed against individuals (e.g. mugging)(HomeOfce, 2011).
Although there may be some parallels between commercial and personal robberiesfor
example, both typically involve groups of offenders (Gill, 2000; Smith, 2003) and motivations
for committing offences overlap (e.g. money, addiction, and excitement)there are also
differences. For example, commercial robberies are more probably planned; the nancial
rewards are usually larger; and in the UK, the weapon ofchoice is a rearm (Gill, 2000) rather
than a knife. It is not surprising therefore that the two categories are separated for research
purposes. In fact, Matthews (2002) goes so far as to say that research that does not differentiate
between the two categories should be interpreted with caution.
Personal robbery is commonly referred to as muggingor street crime(e.g. Tilley, Smith,
Finer, Erol, Charles, & Dobby, 2004), and small, portable, high-value items (such as mobile
telephones) are commonly targeted (Monk, Heinonen, & Eck, 2010; Smith, 2003). Weapons
are used/displayed in approximately one-third of personal robberies in the UK (Flatley,
Kershaw, Smith, Chaplin, & Moon, 2010), and knives are commonly associated with this type
of offence (Barker, Geraghty, Webb, & Key, 1993; Flatley et al., 2010). With this level of
weapon use, it is not surprising that 40% of robbery victims receive some kind of injury
(Smith, 2003).
It is important now to explore the potential to conduct case linkage on personal robbery for a
number of reasons. First, personal robbery typically accounts for around 2% of recorded crime
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
in England and Wales per annum. Second, personal robbery has a signicant impact on
victims (Barker et al., 1993; Dolan, Loomes, Peasgood, & Tsuchiya, 2005; Monk et al.,
2010), including loss of goods, injury, fear (Monk et al., 2010), and in some cases long-term
psychological trauma (Barker et al., 1993). There is also evidence that robbers commonly
commit series of robbery offences (Wright & Decker, 1997), so if accurate methods of case
linkage could be developed, it is possible that robbery offences could fall substantially with
the apprehension of relatively few offenders. Third, in the professional experience of the third
author, forensic evidence is often absent in personal robberies because of a lack of physical
contact between the offender(s) and the victim, thus limiting the scope to link robberies using
such evidence.
Personal robbery in the UK is typically committed by small groups of young men (aged
younger than 20 years) against other men (Smith, 2003) although some offenders do
operate alone. The impact of group dynamics on behavioural consistency is largely
unknown. In the US, research has found that although co-offending does not seem to have
a signicant impact on victim selection, it often does increase planning (Alarid, Burton, &
Hochstetler, 2009). This means that the robberies an offender commits with a group might
differ from those they commit alone, potentially making their crimes more difcult to link.
However, research in the UK on behavioural coherence (in rape) has demonstrated the
existence of thematic similarities between offenders committing multiple crimes with the
same co-offenders (Porter & Alison, 2004). Porter and Alison (2006) went on to examine
behavioural coherence in robbery; the results of which suggested that offenders within the
same group might indeed behave in a homogenous fashion. Furthermore, Alarid et al.
(2009) reported that if offenders commit robberies in a short span of time, they probably
select co-offenders from the same group of associates. This suggests that co-offending
might not negatively impact on behavioural consistency (and therefore the ability to link
offences) so long as the offences are committed in relatively quick succession by the same
group of offenders.
Numerous UK studies have linked personal robbery to street culture (e.g. Deakin,
Smithson, Spencer, & Medina-Ariza, 2007; Wright, Brookman, & Bennett, 2006). Motiv-
ation ranges from material gain (Alarid et al., 2009; Monk et al., 2010) to alleviating boredom
(Tilley et al., 2004) or for the buzz(Deakin et al., 2007). Personal robberies are often
described as spontaneous and unplanned (Alarid et al., 2009; Woodhams & Toye, 2007)
and consequently situation dependent. If offences are very situation dependent,any factor that
affects the similarity/dissimilarity of the situation potentially impacts on the ability to identify
linked offences. However, rational choice theory (Cornish & Clarke, 1986)which presents
offenders as decision makerswould suggest that, even where a spontaneous offence occurs
in response to a presented opportunity, the offender(s) will still make rational decisions about
whether the potential benets are worth the risk before embarking on a robbery attempt. As
previous experience of success and failure is a key factor in decision making (Juliusson,
Karlsson, & Gärling, 2005) it is probable that the offender will seek similar robbery opportun-
ities to those that have resulted in successful robberies in the past, thus increasing the feasibil-
ity of the linkage task which is grounded in behavioural consistency.
Offender learning and offender adaptation could prove challenging to case linkage. Modus
operandi (MO) has been found to evolve over time (Yokota & Watanabe, 2002) as some
offenders learn what is effective (Keppel, 1995) and gain condence (Douglas & Munn,
1992). Furthermore, offenders can adapt their MO in response to crime prevention measures
(Tilley et al., 2004) or as new opportunities for crime arise. However, research has found that
many robbery offenders develop a consistent method of committing their offences (Deakin
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
et al., 2007), suggesting that offender learning and adaptation might not impact crime scene
behaviours as much as initially thought.
CURRENT RESEARCH
The new research presented here tested the underlying assumptions of case linkage by compar-
ing the behavioural similarity of linked pairs of personal robbery offences (i.e. two offences
committed by the same offender) with the behavioural similarity of unlinked pairs of personal
robbery offences (i.e. two offences committed by different offenders). It was hypothesised that
linked pairs would be more behaviourally similar than unlinked pairs, thus providing evidence
for both of the theoretical assumptions. This approach has been successfully used by a number
of studies (e.g. Bennell & Canter, 2002; Bennell & Jones, 2005; Tonkin, et al., 2008;
Woodhams & Toye, 2007) and across a range of crime types (e.g. burglary, car theft, commer-
cial robbery). It was therefore anticipated that the approach would provide evidence that
offender behaviour can be used to distinguish between linked and unlinked pairs of personal
robbery.
The current research was novel in two ways. First, the research focused on personal
robbery, an offence type not previously explored by case linkage researchers. Personal
robbery is quite different to commercial robbery, particularly in relation to planning (as
discussed in the introduction), and the literature even warns that the two offence types should
not be analysed together because of their diverse nature (Matthews, 2002). As such, the
existing work on commercial robbery (Woodhams & Toye, 2007) is not necessarily applicable
for analysts and police ofcers working to link series of personal robbery offences. This
research, therefore, aimed to address this gap in the literature.
The research also considered the implications of the size of a study area (e.g. police force
area) on the linkage task. Previous research has utilised the whole study area when generating
unlinked pairs of offences. This is problematic for several reasons. First, offenders tend to
operate in a relatively small geographical area (Santtila, Laukkanen, & Zappalà, 2007), and
so, selecting random pairs of offences from anywhere in the police force area to act as unlinked
pairs is not reective of known patterns of offending, thus potentially biasing the results.
Second, police analysts conduct case linkage at a local level (i.e. borough/policing district)
as well as a force-wide level (Burrell & Bull, 2011); and so, considering how the distance
between unlinked pairs might impact on linkage accuracy is very relevant to the practical
application of case linkage. Third, there is evidence that offenders learn from each other
(Clarke & Eck, 2005); and so, offenders working in the same local area might adopt similar
methods of operation, thus making the linkage task more challenging. Therefore, identifying
which behaviours can be used to reliably link crimes at a local level will help the analysts
working with a local remit.
The issues described earlier could impact on the similarity of unlinked pairs, and so, it is
possible that the case linkage performance of some behavioural domains would deteriorate
if the unlinked offences were geographically closer together. The current research was there-
fore conducted in two phases. Phase 1 compared the similarity of linked pairs of offences with
unlinked pairs of offences committed within the same police force area (which is quite a large
geographical area), allowing comparison of the results to previous research. Phase 2 reduced
the geographical area that the unlinked pairs were sourced from by ensuring that both offences
within the unlinked pair occurred in the same borough. This overcame the limitation of
generating unlinked pairs using a large geographical area, allowing the research to test whether
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
behaviour can be used to distinguish between linked and unlinked pairs on a more local level
as well as on a force-wide basis (as tested by phase 1). Previous research regarding burglary
has shown that although there are differences between policing districts, the performance of
individual domains as useful predictors can be demonstrated across multiple districts (Bennell
& Jones, 2005). It was therefore anticipated that some behavioural domains might be
identied as useful predictors by both phases of the research.
METHOD
Sample
The data sample was extracted from police records for solved personal robbery offences
recorded by Northamptonshire Police between 1st January 2005 and 31st December 2007.
The sample contained 166 offences committed by 83 offenders. The offenders were aged
between 10 and 44 years with an average age of 18years at the time of their offence.
Seventy-seven offenders were male, and ve were female (the gender was recorded as
unknown for one offender). Over 70% (n=58) of the offenders were recorded as being White
people (including four of the women), 13 were Black people, and 12 (including 1 woman)
were of mixed heritage.
Procedure
It is common practice to include a constant number of offences (usually two) per offender in
case linkage analysis to remove the bias that might be presented by prolic offending (Bennell
& Canter, 2002; Bennell & Jones, 2005; Woodhams & Toye, 2007). The current research
replicates previous research (e.g. Bennell & Canter, 2002; Woodhams & Toye, 2007)
conducting the analysis using data from two offences per offender. Furthermore, a recent
publication from the Federal Bureau of Investigation in the US identies serial murder as
the unlawful killing of two or more victims by the same offender(s) in separate events
(Morton, 2008, p. 9), and so, there is precedent for dening a series as two or more offences.
Selecting linked pairs
This research uses the two most recent offences for each offender to create a linked offence
pair mirroring the approach used by other researchers (e.g. Woodhams & Toye, 2007). A total
of 135 serial personal robbers (i.e. those who had committed more than one offence in the
timeframe) were identied from the raw data. However, the Home Ofce Counting Rules
(Home Ofce, 2011) state that a separate crime should be recorded by the police for each
victim rather than each incident; and so, a single incident can result in multiple offences if
there is more than one victim. There were cases where the date, time, and location of offences
were identical and the MO information suggested that the two most recent offences were
actually part of the same incident. To include such pairs in the analysis would falsely inate
the level of similarity in linked pairs. Nineteen offenders were excluded from the analysis
for this reason. In a similar vein, a further 21 offenders were omitted as one or both of their
offences already appeared in the dataset as part of the crime series of another offender (i.e.
their co-offender); and so,the inclusion of the pair would againcompromise the independence
of the linked pairs sample. A further 12 offenders were excluded as a result of missing data
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
about their offences. The two most recent offences (that were not part of the same incident)
were selected for each of the remaining 83 offenders, forming the linked sample for analysis.
Selecting unlinked pairs for phase 1
The current research mirrors previous case linkage research utilising an unlinked sample with
the same number of pairs as the linked sample (e.g. Markson et al., 2010, Tonkin, Woodhams,
Bull, Bond, & Palmer, 2011b; Tonkin, et al., 2008). The unlinked pairs were generated using
the =RAND() function in Microsoft Excel to randomly re-order the rows in the linked sample.
The unlinked pairs were created by using rows 1 and 2 as pair1, rows 3 and 4 as pair 2, and so
on. The data were then checked manually to ensure that all the unlinked pairs were in fact
unlinked as the random re-ordering of rows could result in linked offences being matched
together as an unlinked pair. A total of 83 unlinked pairs were created on the basis of the
166 offences contained within the linked sample. This dataset is labelled unlinked1through-
out this paper.
Selecting unlinked pairs for phase 2
A second unlinked sample (labelled unlinked2throughout the paper) was created for further
comparison. The random nature of allocating offences to an unlinked pair in phase 1 meant
that a single offence could be matched with an unrelated crime located anywhere in the police
force area.The police force area is geographically large (913 square miles), and so, there was a
high likelihood of unlinked pairs being located far apart. The police force area is broken down
into six boroughs, and further examination of the data revealed that the two offences for the
majority of linked pairs (82%) occurred in the same borough, whereas the two offences within
unlinked1 pairs typically occurred in different boroughs (75%). This difference between the
samples of linked and unlinked1 pairs could introduce bias into the analysis, potentially inat-
ing the predictive ability of some of the behavioural domains.
The unlinked2 pairs were generated by randomly re-ordering therows in the linked sample
to create new pairs but this time controlling for borough to ensure the offences in each
unlinked2 pair occurred in the same local area. The borough for each offence was identied
using the crime reference number, which includes a reference to the borough. There were
slightly fewer unlinked pairs in the unlinked2 sample (n= 81) as two boroughs had an odd
number of offences associated with it, meaning that there was a single offence left overafter
pairs had been created. In addition, one borough only had two offences associated with it
these were both committed by the same offender and therefore could not be included as an
unlinked pair. This second unlinked sample (unlinked2) was then combined with the linked
sample from phase 1 and the statistical analyses re-run to determine if the same behaviours
emerged as useful linking factors.
Data coding
A description of the way in which each offence was committed (i.e. the MO) was included in
the police records (i.e. the dataset used in this study). Content analysis of these descriptions
was conducted as part of a wider study and a checklist of dichotomously coded behaviour
variables created. Binary codingi.e. 1 denoting the presence of a behaviour, and 0 the
absence of a behaviourwas used because previous research has indicated that more complex
coding methods are difcult to apply to police data in a reliable way (Canter & Heritage,
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
1990). A total of 67 crime commission behaviour variables were identied; however, a large
proportion of these variables were not included in the current analysis. This was because they
were deemedto be more indicative of victim or bystander behaviour than offender behaviour,
they occurred in less than 1% of cases, and/or the variable had a poor inter-coder reliability
score. In addition, variables in relation to the approach used to commit the robbery were
excluded as preliminary analysis revealed there was insufcient information about the
approach used in offences available for analysis. Furthermore, where approach was possible
to identify, the coders did not always agree on how that approach should be categorised
(e.g. blitz, condence trick, distraction).
Overall, 15 MO behavioural variables, which had a very good overall inter-coder reliability
score (k=0.95), were selected for inclusion in the current study. These variables were
combined with other variables extracted from the recorded crime data (e.g. time of day, day
of week, property stolen) to form a nal behaviourchecklist of 48 behaviours (Table 1).
Individual offence behaviours can be arranged into clusters, each thought to serve a differ-
ent purpose in the offence (Tonkin et al., 2008). For example, weapon use and threatening
language are both examples of how to control victims during an offence. Thus, the behaviours
were grouped into behavioural clusters or domains for analysis. Domains were originally
modelled on those used by Woodhams and Toye (2007) in their analysis of commercial
robbery, namely planning, target selection, control, and property. However, adjustments were
made because of available data and the differing nature of commercial and personal robbery.
First, it was not possible to create a domain for planning because of a lack of relevant beha-
viours recorded in relation to the personal robbery offences. The target selection domain
encompasses some behaviours used by Woodhams and Toye (2007), that is, day of week
and time of day. However, the time of day variables were structured differently to be more
representative of patternsin personal robberyand were based on the timebands used by Smith
(2003) in his research on personal robbery. Variables relating to whether the offender was
known to the victim, and whether the victim was at a cashpoint at the time of the attack were
also included in the target selection domain. The control domain included six variables in
relation to weapon use, variables relating to violent actions, offender commands, and whether
the victim and/or offender were alone or in a group when the offence occurred (n=15
behaviours in total). The property domain contained 14 types of property plus whether any
property was returned to the victim by the robber(s) during or following the offence. Temporal
proximityi.e. the number of days between offencesand inter-crime distance (calculated
using Pythagorastheorem to determine the number of metres between the grid references
[xand yco-ordinates] for the two crimes in each pair) were also included in the analysis. These
nal two behaviours were included as they have been found to be useful predictors of linkage
by previous research (e.g. Tonkin, Santtila, & Bull, 2011a).
Measuring similarity
The similarity of pairs across each behavioural domain was measured using Jaccardscoef-
cient. Jaccardscoefcient does not take joint non-occurrences into account (Real & Vargas,
1996), so using this coefcient means that the level of similarity does not increase if the behav-
iour is not reported to have occurred within an offence pair (Woodhams, Grant, & Price,
2007b). This is an important issue when working with police data as the absence of abehaviour
does not necessarily mean that this behaviour did not occur, rather that it was not reported or
was not recorded (Tonkin et al., 2008). Some research ndings have indicated that taxonomic
similarity might be a more appropriate measure of similarity than Jaccardscoefcient
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
(Woodhams et al., 2007b); however, recent research failed to replicate these ndings
particularly when large sample sizes were used (Melnyk, Bennell, Gauthier, & Gauthier,
2011). Furthermore, research has also demonstrated that Jaccardscoefcient outperforms
Table 1. Behavioural domains
Behavioural domain Offence behaviours N(out of 166 offences) %
Target selection Monday 39 23.5
Tuesday 22 13.3
Wednesday 22 13.3
Thursday 23 13.9
Friday 24 14.5
Saturday 25 15.1
Sunday 11 6.6
22:00 to 01:59 34 20.5
02:00 to 05:59 5 3.0
06:00 to 09:59 5 3.0
10:00 to 13:59 22 13.3
14:00 to 17:59 46 27.7
18:00 to 21:59 54 32.5
Known offender 41 24.7
Unknown offender 64 38.6
Victim at cashpoint/bank 2 1.2
Control Weapon used 60 36.1
Knife 21 12.7
Firearm 4 2.4
Weapon other 11 6.6
Group of offenders versus group of victims 31 18.7
Group of offenders versus lone victim 72 43.4
Lone offender versus group of victims 9 5.4
Lone offender versus lone victim 47 28.3
Offender(s) searches victim/victims property 24 14.5
Violencephysical assault 55 33.1
Weapon threatened 36 21.7
Weapon shown/seen 29 17.5
Offender requests property 32 19.3
Offender demands property 54 32.5
Victim resistsmet with threat 9 5.4
Property Cash 39 23.5
Mobile phone 51 30.7
Cards 8 4.8
Jewellery/watch 5 3.0
Wallet/ purse 11 6.6
Keys 6 3.6
Documents 8 4.8
Luggage 6 3.6
MP3 player 7 4.2
Clothing/ footwear 7 4.2
Food 3 1.8
Cigarettes 4 2.4
Pedal cycle 14 8.4
Miscellaneous 13 7.8
Property returned 3 1.8
Inter-crime distance Inter-crime distance (m)
Temporal proximity Temporal proximity (days)
Total variables 48
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
the taxonomic similarity measure across a range of conditions (Bennell, et al., 2010). Hence,
the decision was made to use the Jaccards similarity measure in this study.
Jaccards coefcients are expressed as a value of between 0 and 1, with 0 indicative of no
similarity and 1 denoting perfect similarity. Thus, higher Jaccardscoefcients for linked pairs
compared with unlinked pairs would provide support for behavioural consistency and distinct-
iveness. Jaccards coefcients were calculated using the Statistical Package for the Social
Sciences (SPSS) version 18.0 © (IBM Corporation, Armonk, NY). SPSS calculates the similar-
ity of pairs of offences on the basis of the behaviours input into the analysis producing a matrix
containing the Jaccardscoefcients for all possible combinations of offences. Coefcient
matrices were produced for each behavioural domain (i.e. target selection, control, and
property). In line with other research (e.g. Tonkin et al., 2011a), Jaccardscoefcients were
also calculated for a combined domain. This contains all behaviours from the target selection,
control, and property domains.
The relevant Jaccardscoefcients for each domain were extracted from each matrix for
each pair in the linked, unlinked1, and unlinked2 samples (i.e. the Jaccards coefcient for
target selection for linked pair 1, the Jaccards coefcient for target selection for linked pair
2, etc.). All other coefcients were excluded from the analysis. The coefcients for each
domain plus the variables temporal proximity and inter-crime distance formed the dataset
for the next stage of the analysis.
Comparing similarity of domains
The KolmogorovSmirnov (D) test of normality revealed that the distribution of Jaccards
scores, inter-crime distances, and temporal proximities were signicantly different from
normal. This means that the median rather than the mean scores should be reported to compare
similarity(Field, 2005) and that a non-parametric test of signicance should be used to assess
whether there is a statistically signicant difference between similarity scores for linked and
unlinked pairs. Previous research (e.g. Markson et al., 2010; Tonkin et al., 2008; Woodhams
& Toye, 2007) has used Wilcoxon matched-pairs signed rank tests to test differences because
of concerns about the independence of the data. The present authors contend that the current
data can be considered to be independent. This is for two reasons: rst, although the linked
and unlinked samples utilisethe same crime data, individual scores (i.e. Jaccards coefcients,
inter-crime distances, and temporal proximities) are generated using data points from two
separate offences. As such, no individual score impacts on the value of another. Second,
unlike some previous research (e.g. Bennell et al., 2009), the current research only compares
linked pairs with one possible combination of unlinked pairs at a time. Therefore, the under-
lying crime data are only used twice within a single set of analyses rather than to a large extent.
AMannWhitney U-test was therefore selected to determine if there was a statistically signi-
cant difference between linked and unlinked pairs for each domain in each phase of analysis.
Reporting effect size is good practice in statistics (Field, 2005) as this provides a measure of
the size of the difference between two populations (in this case linked versus unlinked pairs).
Effect sizes were calculated by converting z-scores from the MannWhitney Uanalysis into
the effect size estimate rusing the equation cited in Field (2005, p. 532).
Identifying predictive factors of linkage
A split-half validation method was introduced at this stage by dividing the data into experi-
mental samples (to build the predictive models) and test samples (to test the predictive
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
models). This mirrors the approach used by other researchers (e.g. Bennell & Canter, 2002;
Bennell & Jones, 2005; Tonkin et al., 2011a). The data were randomly split in half with 42
linked pairs, and 42 unlinked1 pairs forming the experimental sample, and 41 linked pairs
and 41 unlinked1 pairs forming the test sample for phase 1 of the research. The unlinked2
sample was also split into an experimental dataset (composing of 41 unlinked pairs of
offences) and a test sample (made up of 40 pairs). These were combined with the experimental
and test datasets for the linked sample to create the dataset for phase 2 of the analysis.
Single-factor logistic regression models, exploring whether any of the behavioural domains
could be used as accurate predictors of linkage, were calculated using the experimental data-
sets. Regression models consisting of multiple factors were also tested for each phase. This
was achieved through standard logistic regression with multiple factors or utilising stepwise
logistic regression. This determined whether the single factors could be combined to generate
optimal models with improved predictive performance.
The constant (a) and logit (b) values from the regression models were used to calculate the
estimated probabilities for each pair in the test samples using the process outlined in Bennell
and Canter (2002). To clarify, the aand bvalues from the experimental sample for phase 1
were used to calculate the probabilities for the test sample for phase 1, and the aand bvalues
from the experimental sample for phase 2 were used to calculate probabilities for the test
sample for phase 2.
The probabilities were used to perform receiver operating characteristic (ROC) analyses.
ROC analysis is becoming standard practice in case linkage research. ROC analysis is a meas-
ure of predictive accuracy and uses the area under the curve (AUC) to assess the linkage
accuracy of the data that gives rise to the ROC curve (Bennell et al., 2009). An AUC of 0.5
indicates chance level, and an AUC of 1.0 indicates perfect discrimination, meaning the larger
the AUC, the higher the predictive accuracy (Woodhams et al., 2007a). AUCs of between 0.5
and 0.7 are indicative of low levels of accuracy, 0.7 to 0.9 indicate moderate levels of
accuracy, and 0.9 to 1.0 high levels (Bennell & Jones, 2005; Swets, 1988). The approach
has many advantages and has been used to overcome concerns about statistical
independence and to set decision thresholds (e.g. Bennell & Jones, 2005; Bennell et al.,
2009). ROC analysis is also a useful method of calibrating the validity of linkage features
identied by regression models, and this is what it was used for in the current study.
The ROC analysis was conducted for each domain using SPSS version 18.0 © using the
probabilities as test variables and linkage status as the state variable. It was hypothesised that
the ROC analyses would mirror the trends found in the regression analyses, thus providing
validation for the regression models developed with the experimental sample.
RESULTS
Test of difference
Table 2 shows the median Jaccards scores, inter-crime distances, and temporal proximities for
each sample. The results of the MannWhitney U-tests and effect sizes are also listed.
Unlinked1 and unlinked2 pairs display lower Jaccards similarity scores across the target
selection, control, and combined behavioural domains than linked pairs. Both sets of unlinked
pairs have larger inter-crime distances and more days between offences than linked pairs.
Furthermore, the MannWhitney U-test indicates that these differences between linked and
unlinked samples are statistically signicant for all observations except control in phase 1
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
although this is close to signicance (p= 0.063). Interestingly, the effect sizes for the target
selection, control, and combined behavioural domains increase between phase 1 and phase 2,
indicating that these domains might be more useful for linking crimes within borough
compared with force-wide.
Conversely, inter-crime distance and temporal proximity are more useful on a force-wide
basis than on a local level with lower effect sizes reported in phase 2 than in phase 1.The
results for inter-crime distance are not surprising given the methodology for creating
unlinked2 pairs, which reduced the median distance between an unlinked pair from
12,990 to 2,314 m. It is promising, however, that linked pairs are still demonstratively
closer together on average than unlinked pairs and that this nding is statistically signi-
cant. This indicates that inter-crime distance remains a useful linkage tool at a local level;
in fact, its effect size suggests it remains better than the other behavioural domains
examined. The reduction in the average number of days between offences and the effect
size from phase 1 to phase 2 is not as easily explained for temporal proximity. However,
a re-examination of the raw data (i.e. the 166 offences) revealed that offences within each
borough tended to be weighted towards either the start or the end of the timeframe
examined. In fact, no borough had offences from all three years represented within their
sample; thus, this anomalous nding is attributed to the distribution of date of offence
across the data.
There are no differences in the median Jaccards coefcients for the property domain.
Furthermore, the effect size is small, indicating that this domain is unhelpful for linkage
purposes. Overall, these analyses provide support for the assumptions, albeit not across
all behavioural domains.
Table 2. Comparison of similarity scores
Domain Median
MannWhitney U(z)* Effect size (r)
Phase 1 Phase 2 Phase 1 Phase 2
Target selection Linked = 0.250 2,380.000 1,975.000 0.28 0.38
Unlinked1 = 0.000 (3.590)* (4.840)*
Unlinked2 = 0.000
Control Linked = 0.250 2,884.000 2,670.500 0.14 0.18
Unlinked1 = 0.167 (1.859) (2.350)*
Unlinked2 = 0.125
Property Linked = 0.000 3,429.000 3,355.000 0.01 0.00
Unlinked1 = 0.000 (0.071) (0.031)
Unlinked2 = 0.000
Combined Linked = 0.200 2360.500 1,975.000 0.27 0.36
Unlinked1 = 0.143 (3.508)* (4.572)*
Unlinked2 = 0.091
Inter-crime
distance (metres)
Linked = 803.6 659.500 1,735.000 0.69 0.41
Unlinked1 = 12,989.8 (8.889)* (5.176)*
Unlinked2 = 2,313.5
Temporal
proximity (days)
Linked = 36 1,491.500 2,164.000 0.49 0.31
Unlinked1 = 292 (6.313)* (3.942)*
Unlinked2 = 144
Field (2005) indicates that r= 0.10 is a small effect size (explaining 1% of variance), r= 0.30 is a medium effect
size (explaining 9% of the variance), and r= 0.50 is a large effect size (explaining 25% of the variance).
*p<0.05.
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
Regression
The MannWhitney U-tests are a useful starting point for determining which behavioural
domains might be the most useful for differentiating between linked and unlinked samples.
However, additional analysis is needed to identify the predictive value of each domain.
Tables 3 and 4 outline the results of the regression analyses.
The single-factor model for target selection performed fairly well, explaining 12% of the
variance and improving predictive accuracy by 16% in phase 1 (Table 4). The positive result
was replicated in phase 2, this time explaining even more of the variance (20%) and improving
predictive accuracy by 17% beyond chance. This replicates the trends highlighted by the
MannWhitney U-test.
The results for the control domain are contrary to what would be expected given the Mann
Whitney Uresults with phase 1 outperforming phase 2 in relation to R
2
and predictive accur-
acy. Furthermore, the chi-square was signicant for phase 1 but not for phase 2. Overall, the
predictiveaccuracy of the models for phase 1 and phase 2 was low compared with chance, and
neither model explained much of the variance. Combined with low Wald statistics, this
analysis suggests control is of limited value to linkage analysis in cases of personal robbery.
Property was identied as a poor predictive factor for linkage in both phases of the research,
as non-signicant chi-squares indicated that the models did not t the data well, and the non-
signicant Wald statistics indicated that this individual predictor should be removed from the
regression model. Furthermore, the single-factor model did not explain much of the variance,
and predictive accuracy did not improve much beyond chance. The results were not
unexpected given the ndings of the MannWhitney U-tests and add weight to the argument
that the property stolen during a robbery is not particularly useful when predicting whether
two offences are linked or not.
The combined domain (which is composed of target selection, control, and property)
performed favourably compared with the single-factor models in phase 1. Although the
predictive accuracy of the target selection regression model was slightly better (16% compared
with 14%), the combined domain explained more of the variance (18% compared with 12%).
The combined regression model for phase 2 performed on par with the combined model for
phase 1, suggesting that this domain has some value for predicting whether offences are linked
at both a local and a force-wide level.
As expected from the MannWhitney U-tests, the regression model for inter-crime distance
in phase 1 performed very well, explaining 63% of the variance and improving predictive
accuracy by over 30%. Furthermore, the model in phase 2 was much weaker (explaining
17% of the variance and only improving predictive accuracy by 7%), replicating the trend
highlighted by the MannWhitney U-test. Similarly, the trends for temporal proximity repli-
cated those highlighted by the MannWhitney U-tests with a deterioration in how useful the
behaviour was between phase 1 and phase 2. The very poor performance of the phase 2 model
for temporal proximity (explaining less than 1% of the variance and not improving predictive
accuracy much beyond chance) is probably due to the distribution of date of offence in the data
(as stated previously). Furthermore, temporal proximity did explain 14% of the variance and
improved predictive accuracy by 27% in phase 1, indicating that it might still be useful in
certain circumstances.
A forward stepwise logistic regression produced optimal models consisting of target
selection and inter-crime distance for both phases of the research. Although there was some
deterioration in the predictive ability of the models between phase 1 and phase 2, the optimal
models performed the best in both conditions. The optimal model explained 69% of the
variance and improved predictive accuracy by 33% in phase 1 and explained 31% of the
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
Table 3. Regression models
Model
Constant a(SE) Logit b(SE) Wald (df) Model w
2
Nagelkerke R
2
Phase 1 Phase 2 Phase 1 Phase 2 Phase 1 Phase 2 Phase 1 Phase 2 Phase 1 Phase 2
Target selection 0.570 (0.307) 0.675 (0.305) 2.519 (0.969) 3.563 (1.121) 6.757 (1)** 10.097 (1)** 7.929** 13.440** 0.12 0.20
Control 0.387 (0.298) 0.112 (0.295) 1.832 (0.974) 0.553 (0.799) 3.535 (1) 0.480 (1) 3.899* 0.485 0.06 0.008
Property 0.103 (0.240) 0.098 (0.241) 0.806 (.788) 0.623 (0.848) 1.047 (1) 0.540 (1) 1.097 0.554 0.02 0.009
Combined 1.225 (0.465) 0.946 (0.404) 7.789 (2.319) 5.410 (0.946) 9.569 (1)** 7.717 (1)** 12.006** 10.148** 0.18 0.15
Inter-crime
distance
1.831 (0.431) 0.720 (0.328) 0.000 (0.000) 0.000 (0.000) 12.132 (1)** 6.426 (1)* 52.737** 11.313** 0.63 0.17
Temporal
proximity
0.736 (0.336) 0.117 (0.304) 0.003 (0.001) 0.001 (0.001) 8.161 (1)** 0.195 (1) 9.632** 0.196 0.14 0.003
Optimal target
selection
inter-crime
distance
1.044 (0.469) 0.005 (0.399) 4.647 (2.036)
0.000 (0.000)
3.451 (1.218)
0.000 (0.000)
5.209 (1)*
12.283 (1)**
8.032 (1)**
4.855 (1)*
60.267** 22.038** 0.69 0.31
SE, standard error.
*p<0.05,
**p<0.01.
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
variance and improved predictive accuracy by 18% in phase 2. Furthermore, the chi-square
values were signicant to p<0.05 for both optimal models, indicating that they tthe
data well.
Receiver operating characteristic analyses
The results of the ROC analyses (Table 5) were largely consistent with the logistic regression
analyses, with the optimal model (which combined target selection and inter-crime distance)
and the single-factor inter-crime distance model emerging as the best predictors of linkage in
phase 1, with AUCs of 0.90 and 0.92, respectively. AUCs of between 0.90 and 1.00 represent
a high measure of discrimination accuracy for the linkage feature(s) that gave rise to the curve
(Swets, 1988), indicating that the optimal model and, more particularly, inter-crime distance
were very useful for discriminating between linked and unlinked pairs of personal robbery.
Temporal proximity, target selection, and combined were not far behind with moderate
AUCs of 0.83, 0.64, and 0.64, respectively.
The value of inter-crime distance declined substantially in phase 2 (from 0.92 to 0.75). This
difference is statistically signicant as the condence intervals do not overlap (Melnyk et al.,
2011). The AUC for temporal proximity also declines in phase 2; however, this result is not
statistically signicant. As with the regression analysis, the value of target selection improved
in phase 2 (the AUC increases to 0.69); however, as the condence intervals overlap, this
difference is not statistically signicant. Interestingly, the AUC for the combined domain
improves to 0.70 in phase 2 compared with 0.64 in phase 1. This is contrary to the regression
ndings but is in line with the MannWhitney Uresults. Similarly, the AUC for control
improves from 0.56 to 0.66 when moving from phase 1 to phase 2, which is in line with the
Table 5. Receiver operating characteristic analysis
Model
AUC (SE) 95% condence interval
Phase 1 Phase 2 Phase 1 Phase 2
Target selection 0.640 (0.061)* 0.691 (0.059)* 0.5200.760 0.5750.806
Control 0.563 (0.064) 0.657 (0.061)* 0.4370.689 0.5380.776
Property 0.448 (0.064) 0.451 (0.064) 0.3230.573 0.3250.577
Combined 0.640 (0.062)* 0.703 (0.059)* 0.5190.761 0.5880.818
Inter-crime distance (m) 0.918 (0.028)* 0.750 (0.055)* 0.8620.974 0.6430.857
Temporal proximity (days) 0.829 (0.045)* 0.717 (0.059)* 0.7400.917 0.6010.832
Optimal 0.904 (0.033)* 0.782 (0.050)* 0.8400.969 0.6840.881
Note: AUC, area under the curve; SE, standard error. An AUC value of 0.5 is non-informative, a value of 0.500.70 is
low, 0.700.90 is moderate, and 0.901.00 is high (Swets, 1988).
*p<0.05.
Table 4. Predictive accuracy of models
Target
selection Control Property Combined
Inter-crime
distance
Temporal
proximity Optimal
Phase 1 Random 50.0 50.0 50.0 50.0 50.6 50.0 50.0
Model 65.5 56.0 56.0 64.3 81.9 66.7 83.1
Phase 2 Random 50.6 50.6 50.6 50.6 50.6 50.6 50.6
Model 67.5 54.2 55.4 63.9 57.8 51.8 68.7
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
MannWhitney U-test (and its associated effect size) but not the regression analyses.
Comparison of the condence intervals indicates that the difference in AUCs between phase
1 and phase 2 is not signicant for either the combined or control domains.
As would be expected on the basis of the MannWhitney Uresults and the regression
analyses, the property domain did not achieve a signicant AUC in either phase. In fact, with
AUCs of less than 0.50, the domain is non-informative (Swets, 1988), performing at below the
threshold set by chance.
DISCUSSION
This research found that linked pairs were more similar than unlinked pairs across a range of
domains (i.e. behavioural consistency and distinctiveness was observed). Linked pairs had
larger similarity scores for target selection, control, and combined domains plus smaller
inter-crime distances and fewer days between offences than unlinked1 and unlinked2 pairs.
The current ndings, although for the newcrime of personal robbery, are somewhat simi-
lar to those of previous case linkage research that has consistently found inter-crime distance to
be one of the most useful single-factor models for linkage (Bennell & Canter, 2002; Bennell &
Jones, 2005; Markson et al., 2010; Tonkin et al., 2008; Tonkin et al., 2011a; Woodhams &
Toye, 2007). Furthermore, the AUCs attained are comparable with those found in the litera-
ture, as Bennell and Jones (2005) reported a range of 0.76 to 0.91 for inter-crime distance in
their research on burglary, with other researchersAUCs for inter-crime distance falling within
this range. The inter-crime distance model also performed well in terms of predictive accuracy
when applied at the force level (31% improvement over the random model). However, the pre-
dictive accuracy of the regression model and the AUCs for inter-crime distance deteriorated
between phase 1 and phase 2. This suggests that caution needs to be applied when linking local
crimes using inter-crime distance alone as it can no longer be considered a strong predictor of
linkage. However, inter-crime distance did still achieve a moderate AUC (and in fact the high-
est AUC for a single-factor model) in phase 2, indicating it may still have some value when
working at a local level. The key message from this research would be that although inter-
crime distance remains valuable when working at a local level, it should be treated with more
caution than if working at a force-wide level. It is argued therefore thatit should not be used in
isolation to link crimes. This is particularly important as analysts may have successfully used
inter-crime distance to link other offence types locally and assumed that this could simply be
extended to personal robbery. Research exploring the thresholds for deciding whether crimes
are linked on the basis of inter-crime distance in different-sized geographical areas (i.e. force-
wide and borough) needs to be identied to assist analysts to make informed linkage
decisions.
There are a number of reasons why inter-crime distance might emerge as a useful linkage
factor. First, research consistently demonstrates that offenders tend to operate within a limited
geographical area or comfort zone. For example, Santtila et al. (2007) found the median
distance for committing a rape was 2.44 km from the offenders home; this fell to 0.85 km
for homicide. Furthermore, as rational decision makers, offenders have a tendency to act on
the rst or closest opportunity to commit crime (the least effort principle) (Rossmo &
Rombouts, 2008). As such, once offenders have found a good location to commit a robbery,
there is no immediate reason for them to travel far to commit the next robbery. This effect
would be more pronounced in rural areas such as Northamptonshire (which is 90% rural;
Ofce for National Statistics, 2004), where opportunities to commit crime are limited and/or
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
clustered geographically (e.g. robberies will cluster in the more urbanised parts of rural areas,
such as small market towns with surrounding villages, which are often located far from one
another). The clustering of targets maytherefore help the linkage task as the analystonly needs
to search a limited geographical area for other crimes in the series. However, it may also have
an adverse effect as multiple offenders will operate within any cluster of potential targets.
Thus, the frequency of offending in these areas may make it difcult to distinguish between
offenders (Bennell & Jones, 2005).
The target selection domain did not perform as well as in previous research. The AUCs
achieved (0.64 to 0.69) were lower than the AUC reported by Woodhams and Toye in their
2007 study oncommercial robbery (0.79). Furthermore, the predictive accuracy of the regres-
sion models were somewhat lower than that found by Woodhams and Toye (2007), an
increase of between 16% and 17% over chance compared with over 20% reported by
Woodhams and Toye (2007). However, the literature indicates that the performance of the
target selection domain can vary considerably, suggesting that this domain might be more
useful in some offences than others. For example, target selection behaviours have performed
better with samples of commercial robbery (Woodhams & Toye, 2007) and commercial
burglary (Bennell & Canter, 2002) compared with car theft (Tonkin et al., 2008). Also, its
performance seems to vary by country, for example, target selection performed well in the
study of Tonkin et al. (2011a) of residential burglary in Finland; however, it performed less
well in the study of Markson et al. (2010) of residential burglary in the UK. This could occur
for a number of reasons including which behaviours are included (or not included) in the tar-
get selection domain, the quality of data recording and coding, and/or because differing social
structures may present different opportunities to commit crime. Further research is needed to
explore the reasons for the differences and to identify the optimal combination of target
selection behaviours to use for linkage purposes.
The current study also found that the target selection domain performed slightly better at a
local level than on a force-wide basis (although the difference was not statistically signicant).
This is perhaps unexpected because different areas present different opportunities (or targets)
for robbers, so some homology of targets might be anticipated when multiple offenders are
operating in the same area. Therefore, as active decision makers and risk assessors (Clarke
& Cornish, 1986), robbers operating in the same area would be expected to identify the same
or similar people to target, therefore making it difcult to distinguish between offenders. How-
ever, many offenders operate within a patch(Deakin et al., 2007); and if these patches do not
overlap, combined with the evidence that offenders do not travel far to commit their offences
(Santtila et al., 2007), this might be a reason why an individual offenderscrimes might be
easier to link using target selection at the more locallevel. Furthermore, it is probable that
there will be fewer active robbers operating in any single local area than force-wide, increas-
ing the chances of distinguishing between different offence series by using target selection.
Control has not been included as a behavioural domain in many studies, possibly because
this can be difcult to code or is not relevant in some crime types (e.g. burglary). Woodhams
and Toye (2007) found control to be the best predictor of linkage in their study on commercial
robbery, even performing better than inter-crime distance. The current research failed to
replicate this nding for personal robbery. This may be due to the differing variables included
in the domain. For example, Woodhams and Toye (2007) included information about the
manner in which the offence was committed (i.e. calm/condent, anxious/agitated, or loud/
aggressive), whereas the current study did not as it was not possible to code this behaviour
from the MO information available for the research. It is possible that replicating the present
study with a different set of variables within the control domain might increase predictive
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
accuracy. Despite the disappointing results, it is argued that there is potential for control to be a
useful predictor of robbery linkage, although more work is needed to identify which
behaviours are the most useful to include in the domain. More in-depth information about
the interaction between the victim and the offender(s) would be benecial in terms of identi-
fying and coding control behaviours in a way that could be utilised for caselinkage. This could
be achieved through access to original victim and witness statements as these probably contain
more detailed information than the MO data accessed for this study.
The property domain performed poorly. The inability of the property domain to distinguish
between linked and unlinked pairs of robbery can perhaps be at least partially explained by
situation dependence as property stolen is largely determined by what that victim is carrying
at the time of the attack. Another reason why the propertydomain is unhelpful to link offences
is probably that many robbers typically target the same specic items such as mobile phones,
cash, and jewellery (Monk et al., 2010; Smith, 2003), that is, those items that can be easily
carried. Therefore, the type of property stolen is probably a characteristic of robbery generally,
thus not distinguishing between offenders. Overall, the results for property are not unexpected,
particularly when the fact that some researchers have reported low levels of predictive
accuracy and AUCs compared with other behavioural domains (e.g. Tonkin et al., 2011a) is
taken into account.
The combined domain (which is composed of target selection, control, and property)
performed favourably compared with the single-factor models in phase 1. Overall, the effect
size reported for the MannWhitney U-test and the AUCs in the ROC analysis increased
between phases 1 and 2 of the research. However, the difference in AUCs was not statistically
signicant. Furthermore, the combined regression model for phase 2 performed on par with the
combined regression model for phase 1. Taken together, this suggests that the combined
domain has some value for predicting whether offences are linked at both local and force-wide
levels. The target selection regression model outperformed the combined model for phase 2.
Although this was not statistically signicant, taking the performance of control and property
into account, this suggests that target selection behaviours are driving the success of the
combined domain. Further research, testing different combinations of behaviours in the com-
bined domain, would be useful to determine whether there is any evidence for this proposition.
The signicant results from phase 1 suggest that temporal proximity might be a useful
linking factor in certain circumstances. However, given the methodology used to create the
linked pairs (i.e. selecting the two most recent offences), it is perhaps surprising that temporal
proximity did not perform better in phase 1, particularly when the poor results for phase 2 are
probably inuenced by the distribution of data for date of offence. This highlights the impact
that the distribution of data can have on the outcome of statistical tests and emphasises the
importance of taking this into account when interpreting results. It is, however, important to
continue to explore the value of temporal proximity in case linkage as temporal behaviour
(combined with spatial behaviour) has been highlighted as a useful method of concentrating
investigative efforts in serial cases (Rossmo & Rombouts, 2008). It is compulsory for temporal
information to be recorded by the police for all offences in the UK, meaning that this resource
is readily available to analysts for case linkage work. Thus, it is important to establish whether
temporal proximity canbe used to link offences; and if so, how temporal data might be used in
the most effective way.
The optimal models are composed of target selection and inter-crime distance in both
phases, and these performed better than the single-factor models. However, in phase 1, these
differences were not large compared with inter-crime distance alone, accounting for a similar
proportion of the variance and comparable improvements to predictive accuracy recorded.
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
However, the optimal model was favourable compared with the inter-crime distance model in
phase 2, accounting for 31% of the variance compared with 17% and improving predictive
accuracy by 18% compared with 7%. This suggests that, although inter-crime distance is
the most useful linking factor if working at force level, it may be useful to combine this with
target selection if working locally. This is somewhat supported by the ROC analysis as inter-
crime distance recorded the highest AUC in phase 1 but the optimal model performed best in
phase 2, and although the difference between the AUCs did not quite reach signicance, the
condence intervals did not overlap by much (0.8620.974 compared with 0.6840.881).
Overall, the results provide some support for the theoretical assumptions of case linkage as
there were statistical differences between linked and unlinked pairs across a number of
behavioural domains. However, the results also suggest that the predictive ability of some
behavioural domains may be sensitive to whether the unlinked pairs have meaningful
constraints put on their inter-crime distance (i.e. localversus force-widepairings).
Limitations and directions for future research
Much of the case linkage research has been conducted using police recorded crime data, and
this research is no exception. The limitations of working with police data are clearly outlined in
the case linkage literature, including the challenges presented by missing data (Tonkin et al.,
2008) and the inability to assess the reliability of data coding within police data systems
(Bennell & Canter, 2002). However, using police recorded crime data remains one of the most
ecologically valid methods of conducting linkage research for the people who perform case
linkage inan applied setting, typically police analysts (Woodhams & Toye, 2007). Given that,
overall, the research strongly suggests that case linkage with police data is indeed feasible,
then even better data recording and better victim interviewing (Milne & Bull, 1999) by the
police seem warranted. The data gaps and limitations highlighted within this (and other)
research provide a useful starting point for improving data quality. Access to more in-depth
data already held by the police (e.g. original victim and witness statements) would also be
benecial.
Concerns have also been raised about the use of solved offences as the basis for the link-
age task (Bennell & Canter, 2002); not only is this not reective of case linkage in an
applied setting, but it is also possible that one of the reasons cases are solved is that they
are behaviourally similar and/or geographically and temporally proximal (Bennell & Jones,
2005). Thus, using solved offences could inate the similarity scores or articially reduce
the geographical and temporal distances of linked offences compared with unsolved serial
crime. Future research needs to address this issue. This could be achieved by assessing
the behavioural consistency and distinctiveness of unsolved series of offences that have
been linked using another means (e.g. DNA or ngerprints) (Woodhams et al., 2007a) or
by comparing the similarity of linked pairs rst identied through MO to offences
rst linked through DNA (as Woodhams & Labuschagne [2011] have just explored with
interesting results).
This research compares a linked sample with unlinked samples of a comparable size. In an
applied setting, the analyst is looking for series of offences from within all recorded crime.
Thus, limiting the sample of unlinked pairs in this way is not reective of the linkage task,
and this might have inated or depressed the value of behavioural domains for linkage. This
limitation could be overcome by comparing the linked sample with all possible combinations
of unlinked pairs.
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
It is important to consider whether ndings can be generalised to other areas (e.g. other
police force areas or countries) (Bennell & Jones, 2005; Tonkin et al., 2008). For example,
the ndings of this research might not be generalisable to other police force areas in the UK
because of the way different population densities, the level of urbanisation, and/or geograph-
ical layout might impact on crimeopportunities. It is therefore necessary to replicate the work
in a different police force area to determine whether the current ndings hold true elsewhere.
Such a study (using the same coding dictionary and domains) is already underway using data
from a large, UK urban police force. Replicating the research using data from a comparable
police force in another country (such as New Zealand or Norway, which both have large rural
areas) would also be valuable to determine whether the same behavioural domains emerge as
strong linkage factors.
Finally, there is a clear gap in the literature in relation to personal robbery. There have been
a number of studies of commercial robbery (e.g. Gill, 2000), but not many studies that focus
on personal robbery. In some cases where studies of personal robbery are identied, it
becomes apparent that these are focused on street crime, which encompasses other crimes
(such as snatch thefts, low level violence, and sometimes gang activity), rather than personal
robbery specically. Furthermore, denitions of personal robbery differ (particularly in differ-
ent countries), making comparisons between studies challenging. Future research on the scale
and nature of personal robbery would therefore be benecial. It would also be useful to
conduct some research with robbery offenders to explore the factors that affect decision
making to add context to the existing literature.
CONCLUSION
This study found that the predictive accuracy of domains is subject to change if geographical
constraints for selecting the unlinked pairs are imposed, which more closely reect how a
crime analyst might work at a local level. For some behavioural domains, predictive accuracy
improved with these constraints, most notably for target selection behaviours. Even where
performance deteriorated (as measured by the predictive accuracy of regression models, and
the AUC of ROC curves), statistical differences were still found between linked and unlinked
pairs of offences for some domains, most notably inter-crime distance, indicating that for such
domains, the assumptions of case linkage were still supported whether operating at a force-
wide or more local level.
REFERENCES
Alarid, L. F., Burton, V. S., & Hochstetler, A. L. (2009). Group and solo robberies: Do accomplices
shape criminal form? Journal of Criminal Justice,37,19. DOI: 10.1016/j.jcrimjus.2008.12.001
Barker, M., Geraghty, J., Webb, B., & Key, T. (1993). The prevention of street robbery. Police
research group crime prevention unit series paper no. 44. London: Home Ofce. Retrieved March
3, 2011, from http://rds.homeofce.gov.uk/rds/prgpdfs/fcpu44.pdf
Bennell, C., & Canter, D. V. (2002). Linking commercial burglaries by modus operandi: Tests using
regression and ROC analysis. Science & Justice,42, 153164. DOI: 10.1016/S1355-0306(02)
71820-0
Bennell, C., Gauthier, D., Gauthier, D., Melnyk, T., & Musolino, E. (2010). The impact of data
degradation and sample size on the performance of two similarity coefcients used in behavioural
linkage analysis. Forensic Science International,199,8592. DOI: 10.1016/j.forsciint.2010.03.017
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
Bennell, C., & Jones, N. J. (2005). Between a ROC and a hard place: A method for linking serial
burglaries by modus operandi. Journal of Investigative Psychology and Offender Proling,2,
2341. DOI: 10.1002/jip.21
Bennell, C., Jones, N. J., & Melnyk, T. (2009). Addressing problems with traditional crime linking
methods using receiver operating characteristic analysis. Legal and Criminological Psychology,
14, 293310. DOI: 10.1348/135532508X349336
Burrell, A., & Bull, R. (2011). A preliminary examination of crime analystsviews and experiences
of comparative case analysis. International Journal of Police Science and Management,13,215.
DOI: 10.1350/ijps.2011.13.1.212
Canter, D. (1995). Psychology of offender proling. In R. Bull & D. Carson (Eds.), Handbook of
psychology in legal contexts (pp. 343355). Chichester: John Wiley & Sons.
Canter, D. & Heritage, R. (1990). A multivariate model of sexual offences behaviour: Developments
in offender proling.Journal of Forensic Psychiatry,1, 185212. DOI: 10.1080/
09585189008408469
Clarke, R.V., & Eck, J. E. (2005). Crime analysis for problem solvers in 60 small steps. Retrieved
September 8, 2011, from http://www.popcenter.org/library/reading/PDFs/60steps.pdf
Cornish, D. B. & Clarke, R. (1986). The reasoning criminal: Rational choice perspectives on offending.
New York, NY: Springer-Verlag.
Deakin, J., Smithson, H., Spencer, J., & Medina- Ariza, J. (2007). Taxing on the streets: Understanding
the methods and process of street robbery. Crime Prevention and Community Safety,9,5276.
DOI:10.1057/palgrave.cpcs.8150033
Dolan, P., Loomes, G., Peasgood, T., & Tsuchiya, A. (2005). Estimating the intangible victim costs
of violent crime. British Journal of Criminology,45, 958976. DOI: 10.1093/bjc/azi029
Douglas, J. E., & Munn, C. (1992). Violent crime scene analysis: Modus operandi, signature, and
staging. FBI Law Enforcement Bulletin. Retrieved December 16, 2010, from http://www.criminalpro
ling.ch/violent-crime.html
Ewart, B. W., Oatley, G. C., & Burn, K. (2005). Matching crime using burglarsmodus operandi:
A test of three models. International Journal of Police Science & Management,7, 160174.
DOI: 10.1350/ijps.2005.7.3.160
Field, A. (2005). Discovering statistics using SPSS (2nd ed.). London: Sage.
Flatley, J., Kershaw, C., Smith, K., Chaplin, R., & Moon, D. (2010, July). Crime in England
and Wales 2009/10. Findings from the British Crime Survey and police recorded crime (2
nd
Ed).
Home Ofce Statistical Bulletin 12/10. London: Home Ofce. Retrieved December 16, 2010, from
http://rds.homeofce.gov.uk/rds/pdfs10/hosb1210.pdf
Gill, M. (2000). Commercial robbery: Offendersperspectives on security and crime prevention.
London: Blackstone Press Ltd.
Goodwill, A. M., & Alison, L. J. (2006). The development of a lter model for prioritizing suspects in
burglary offences. Psychology, Crime, & Law,12, 395416. DOI: 10.1080/10683160500056945
Grubin, D., Kelly, P., & Brunsdon, C. (2001). Linking serious sexual assaults through behaviour
(HORS 215). London: Home Ofce.
Hazelwood, R. R., & Warren, J. I. (2003). Linkage analysis: Modus operandi, ritual, and signature in
serial sexual crime. Aggression and Violent Behavior,8, 587598. DOI: 10.1016/j.avb.2004.02.002
Home Ofce. (2011). Home ofce counting rules for recorded crime: Robbery. Retrieved August 4,
2011, from http://www.homeofce.gov.uk/publications/science-research-statistics/research-statistics/
crime-research/counting-rules/count-robbery?view=Binary
Juliusson, E. Á., Karlsson, N., & Gärling, T. (2005). Weighing the past and future in decision making.
European Journal of Cognitive Psychology,17, 561575. DOI: 10.1080/09541440440000159
Keppel, R. D. (1995). Signature murders: A report of several related cases. Journal of Forensic Science,
40,670674. DOI: 10.1002/jip.22
Labuschagne, G.. (2012). The use of a linkage analysis as an investigative tool and evidential
material in serial offenses. In K. Borgeson & K. Kuehnle (Eds.) Serial offenders: Theory and
practice (pp.187215). London: Jones & Bartlett Learning, LLC.
Markson, L., Woodhams, J., & Bond, J. (2010). Linking serial residential burglary: Comparing the
utility of modus operandi behaviours, geographical proximity, and temporal proximity. Journal
of Investigative Psychology & Offender Proling,7,91107. DOI: 10.1002/jip.120
Matthews, R. (2002). Armed robbery. Cullompton, Devon: Willan Publishing.
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
Melnyk, T., Bennell, C., Gauthier, D., & Gauthier, D. (2011). Another look at across-crime similarity
coefcients for use in behavioural linkage analysis: An attempt to replicate Woodhams, Grant, and
Price (2007). Psychology, Crime, and Law,7, 359380. DOI: 10.1080/10683160903273188
Milne, R., & Bull, R. (1999). Investigative interviewing: Psychology and practice. Chichester: Wiley.
Monk, K. M., Heinonen, J. A., & Eck, J. E. (2010). Street robbery. Problem-oriented guides for police
problem-specic guides series no. 59. Retrieved September 8, 2011, from http://www.popcenter.
org/problems/pdfs/street_robbery.pdf
Morton, R. J. (Ed.) (2008). Serial murder: Multi-disciplinary perspectives for investigators.
US Department of Justice: Federal Bureau of Investigation. Retrieved January 27, 2012, from
http://www.fbi.gov/stats-services/publications/serial-murder/
Ofce for National Statistics (2004). Rural & urban area classication 2004 [data le]. Retrieved January
23, 2012, from http://www.ons.gov.uk/ons/publications/re-reference-tables.html?edition=tcm%
3A77-211129
Porter, L. E., & Alison, L. J. (2004). Behavioural coherence in violent group activity: An interpersonal
model of sexually violent gang behaviour. Aggressive Behavior,30,449468. DOI: 10.1002/
ab.20047
Porter, L. E., & Alison, L. J. (2006). Behavioural coherence in group robbery: A circumplex model of
offender and victim interactions. Aggressive Behavior,32, 330342. DOI: 10.1002/ab.20132
Real, R., & Vargas, J. M. (1996). The probabilistic basis of Jaccards index of similarity. Systematic
Biology,45, 380385.
Rossmo, K., & Rombouts, S. (2008). Geographic proling. In R. Wortley, & L. Mazerolle (Eds.)
Environmental criminology and crime analysis (pp. 136149). Devon: Willan Publishing.
Salfati, C. G., & Bateman, A. L. (2005). Serial homicide: An investigation of behavioural consistency.
Journal of Investigative Psychology and Offender Proling,2,121144. DOI: 10.1002/jip.27
Santtila, P., Fritzon, K., & Tamelander, A. (2004). Linking arson incidents on the basis of crime
scene behavior. Journal of Police and Criminal Psychology,19,116. DOI: 10.1007/BF02802570
Santtila, P., Junkkila, J., & Sandnabba, N. K. (2005). Behavioural linking of stranger rapes. Journal
of Investigative Psychology and Offender Proling,2,87103. DOI: 10.1002/jip.26
Santtila, P., Laukkanen, M., & Zappalà, A. (2007). Crime behaviours and distance travelled in homicides
and rapes. Journal of Investigative Psychology and Offender Proling,4,115. DOI: 10.1002/jip.56
Smith, J. (2003). The nature of personal robbery (home ofce research phase 254). London: Home
Ofce. Retrieved August 4, 2011, from http://library.npia.police.uk/docs/hors/hors254.pdf
Sorochinski, M., & Salfati, C. G. (2010). The consistency of inconsistency in serial homicide: Patterns
of behavioural change across series. Journal of Investigative Psychology & Offender Proling,7,
109136. DOI: 10.1002/jip.118
Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science,240, 12851293.
Retrieved January 24, 2012, from: http://www.ncbi.nlm.nih.gov/pubmed/3287615
Tilley, N., Smith, J., Finer, S., Erol, R., Charles, C., & Dobby, J. (2004). Problem solving street
crime: Practical lessons from the street crime initiative. London: Home Ofce.
Tonkin, M., Grant, T., & Bond, J. W. (2008). To link or not to link: A test of the case linkage principles
using serial car theft data. Journal of Investigative Psychology and Offender Proling,5,5977.
DOI: 10.1002/jip.74
Tonkin, M., Santtila, P., & Bull, R. (2011a). The linking of burglary crimes using offender behaviour:
Testing research cross-nationally and exploring methodology. Legal and Criminological Psychology.
Advance online publication. DOI: 10.1111/j.2044-8333.2010.02007.x
Tonkin, M., Woodhams, J., Bull, R., Bond, J., & Palmer, E. (2011b). Linking different types of crime
using geographical and temporal proximity. Criminal Justice and Behavior,38, 10691088.
DOI: 10.1177/0093854811418599
Woodhams, J., Bull, R., & Hollin, C. R. (2007a). Case linkage: Identifying crimes committed by
the same offender. In R. N. Kocsis (Ed.) Criminal proling: International theory, research, and
practice (pp. 117133). Totowa, NJ: Humana Press.
Woodhams, J., Grant, T. D., & Price, A. R. G. (2007b). From marine ecology to crime analysis:
Improving the detection of serial sexual offences using a taxonomic similarity measure. Journal
of Investigative Psychology and Offender Proling,4,1727. DOI: 10.1002/jip.55
Woodhams, J., & Labuschagne, G. (2011). A test of case linkage principles with solved and unsolved
serial rapes. Journal of Police and Criminal Psychology. Advanced online publication.
DOI:10.1007/s11898-011-9091-1
Linking personal robbery offences using offender behaviour
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
Woodhams, J. & Toye, K. (2007). An empirical test of the assumptions of case linkage and offender
proling with serial commercial robberies. Psychology, Public Policy, and Law,13,5985.
DOI: 10.1037/1076-8971.13.1.59
Wright, R., Brookman, F., & Bennett, T. (2006). The foreground dynamics of street robbery in
Britain. British Journal of Criminology,46,115. DOI: 10.1093/bjc/azi055.
Wright, R. T. & Decker, S. H. (1997). Armed robbers in action: Stick ups and street culture. Boston:
Northeastern University Press.
Yokota, K., & Watanabe, S. (2002). Computer-based retrieval of suspects using similarity of modus
operandi. International Journal of Police Science & Management,4,515.
A. Burrell et al.
Copyright © 2012 John Wiley & Sons, Ltd. J. Investig. Psych. Offender Prol. (2012)
DOI: 10.1002/jip
... This research has tested different behaviors such as similarity of modus operandi (MO) (methods of the break-in, targets of crime and stolen items) as well as geographical proximity (distance between offense locations) and temporal proximity (time interval between offenses) between cases. Notably, geographical proximity has attracted attention from researchers and shown its effectiveness in crime linkage (Burrell et al., 2012;Bennell and Canter, 2002;Bennell and Jones, 2005;Davies et al., 2012;Markson et al., 2010;Tonkin et al., 2008;Tonkin et al., 2012aTonkin et al., , 2012b. Previous research have generally shown the importance of geographic proximity in crime linkage compared to M.O. and temporal proximity (Burrell et al., 2012;Davies et al., 2012;Markson et al., 2010;Tonkin et al., 2012a). ...
... Notably, geographical proximity has attracted attention from researchers and shown its effectiveness in crime linkage (Burrell et al., 2012;Bennell and Canter, 2002;Bennell and Jones, 2005;Davies et al., 2012;Markson et al., 2010;Tonkin et al., 2008;Tonkin et al., 2012aTonkin et al., , 2012b. Previous research have generally shown the importance of geographic proximity in crime linkage compared to M.O. and temporal proximity (Burrell et al., 2012;Davies et al., 2012;Markson et al., 2010;Tonkin et al., 2012a). Furthermore, geographical proximity has been shown to be effective together with temporal proximity in linking serial offenses across different types of crimes (Haginoya, 2016b;Tonkin et al., 2011) and in the mixed condition of serial and non-serial crimes (Haginoya, 2016a;Tonkin et al., 2012a). ...
... 95% CI = confidence interval of AUC; sensitivity and specificity were calculated based on the Youden index specificity). These accuracies are equivalent to or better than in previous studies that examined crime linkage using geographical proximity for burglaries (Bennell and Canter, 2002;Bennell and Jones, 2005;Markson et al., 2010;Tonkin et al., 2012aTonkin et al., , 2012bHaginoya, 2014a), robberies (Burrell et al., 2012;Woodhams and Toye, 2007) and car theft or engine key theft (Davies et al., 2012;Tonkin et al., 2008;Tonkin et al., 2012b). Therefore, the present study replicated the robust usefulness of geographical proximity and showed the effectiveness of all three types of distance measures in linking crimes. ...
Article
Purpose The purpose of this paper was to compare the accuracy of linking crimes using geographical proximity between three distance measures: Euclidean (distance measured by the length of a straight line between two locations), Manhattan (distance obtained by summing north-south distance and east-west distance) and the shortest route distances. Design/methodology/approach A total of 194 cases committed by 97 serial residential burglars in Aomori Prefecture in Japan between 2004 and 2015 were used in the present study. The Mann–Whitney U test was used to compare linked (two offenses committed by the same offender) and unlinked (two offenses committed by different offenders) pairs for each distance measure. Discrimination accuracy between linked and unlinked crime pairs was evaluated using area under the receiver operating characteristic curve (AUC). Findings The Mann–Whitney U test showed that the distances of the linked pairs were significantly shorter than those of the unlinked pairs for all distance measures. Comparison of the AUCs showed that the shortest route distance achieved significantly higher accuracy compared with the Euclidean distance, whereas there was no significant difference between the Euclidean and the Manhattan distance or between the Manhattan and the shortest route distance. These findings give partial support to the idea that distance measures taking the impact of environmental factors into consideration might be able to identify a crime series more accurately than Euclidean distances. Research limitations/implications Although the results suggested a difference between the Euclidean and the shortest route distance, it was small, and all distance measures resulted in outstanding AUC values, probably because of the ceiling effects. Further investigation that makes the same comparison in a narrower area is needed to avoid this potential inflation of discrimination accuracy. Practical implications The shortest route distance might contribute to improving the accuracy of crime linkage based on geographical proximity. However, further investigation is needed to recommend using the shortest route distance in practice. Given that the targeted area in the present study was relatively large, the findings may contribute especially to improve the accuracy of proactive comparative case analysis for estimating the whole picture of the distribution of serial crimes in the region by selecting more effective distance measure. Social implications Implications to improve the accuracy in linking crimes may contribute to assisting crime investigations and the earlier arrest of offenders. Originality/value The results of the present study provide an initial indication of the efficacy of using distance measures taking environmental factors into account.
... There is also a well-established body of research evidence that supports the use of series linking (e.g. Bennell & Jones, 2005;Burrell et al., 2012;Ellingwood et al., 2013;Santtila et al., 2008;Woodhams et al., 2019). This research has predominantly focused on testing whether offenders repeat (at least some of) their MO from one crime to the next (referred to as behavioural consistency) and whether it is possible to distinguish the MO of one offender from that of another offender (referred to as behavioural distinctiveness). ...
Article
The memory of eyewitnesses is a valuable form of evidence within criminal investigations. However, both investigators and eyewitnesses are not exempt from making mistakes that may impede upon the accuracy of witness statements. One such mistake is known as memory blindness. Memory blindness describes the process by which - unintentional or intentional - alterations to witness statements made by investigators can lead witnesses to not notice (i.e., be blind to) such changes, which in turn may distort witnesses’ memories for these altered details. Memory blindness holds serious implications for both eyewitness recall memory, and may therefore compromise the criminal investigation. The current paper will discuss recent developments in the research on memory blindness and describe the factors that have so far been identified as impacting upon memory blindness in witnesses. Recommendations as to best practices for avoiding distorted memory during interviews resulting from memory blindness are also provided in light of the current research base.
... what happens to linking accuracy when one-offs are added to the mix?), and also to counter the problem (outlined above) of assumed perfect knowledge of series membership in the data. A method developed by Craig Bennell (2002) (and subsequently used in several crime linking studies, Bennell and Canter, 2002;Burrell et al., 2012;Tonkin et al., 2017;Woodhams and Labuschagne, 2012) was used to measure the similarity, in terms of offender crime scene behavior, between all of the individual homicides. This was done by calculating Jaccard's coefficient of similarity for each pair of crimes; J = a/ (a þ b þ c), where a is the number of behaviors present in both crimes in the pair, b the number of behaviors present in crime one but not in crime two, and c the number of behaviors absent in crime one but present in crime two. ...
Article
Full-text available
Purpose – Crime linkage analysis (CLA) can be applied in the police investigation-phase to sift through a database to find behaviorally similar cases to the one under investigation and in the trial-phase to try to prove that the perpetrator of two or more offences is the same, by showing similarity and distinctiveness in the offences. Lately, research has moved toward more naturalistic settings, analyzing data sets that are as similar to actual crime databases as possible. One such step has been to include one-off offences in the data sets, but this has not yet been done with homicide. The purpose of this paper is to investigate how linking accuracy of serial homicide is affected as a function of added hard-to-solve one-off offences. Design/methodology/approach – A sample (N = 117–1160) of Italian serial homicides (n = 116) and hard-to-solve one-off homicides (n = 1–1044, simulated from 45 cases) was analyzed using a Bayesian approach to identify series membership, and a case by case comparison of similarity using Jaccard’s coefficient. Linking accuracy was evaluated using receiver operating characteristics and by examining the sensitivity and specificity of the model. Findings – After an initial dip in linking accuracy (as measured by the AUC), the accuracy increased as more one-offs were added to the data. While adding one-offs made it easier to identify correct series (increased sensitivity), there was an increase in false positives (decreased specificity) in the linkage decisions. When rank ordering cases according to similarity, linkage accuracy was affected negatively as a function of added non-serial cases. Practical implications – While using a more natural data set, in terms of adding a significant portion of non-serial homicides into the mix, does introduce error into the linkage decision, the authors conclude that taken overall, the findings still support the validity of CLA in practice. Originality/value – This is the first crime linkage study on homicide to investigate how linking accuracy is affected as a function of non-serial cases being introduced into the data.
... Crime linkage(CL) refers to the identification of crimes committed by the same suspect based on behavioral similarity [3], also called linkage analysis, case linkage, crime linkage [7]. It has been applied to many types of crime, such as serial sexual offenses [8,9], serial robbery [10,11], serial burglary [12,13] and so on. In crime linkage, the similarity of criminal behaviors is generally characterized by Modus operandi (M.O.). ...
Article
Crime linkage is a challenging task in crime analysis, which is to find serial crimes committed by the same offenders. It can be regarded as a binary classification task detecting serial case pairs. However, most case pairs in the real world are nonserial, so there is a serious class imbalance in the crime linkage. In this paper, we propose a novel random forest based on the information granule. The approach doesn’t resample the minority class or the majority class but concentrates on indistinguishable case pairs at the classification boundary. The information granule is used to identify case pairs that are difficult to distinguish in the dataset and constructs a nearly balanced dataset in the uncertainty region to deal with the imbalanced problem. In the proposed approach, random trees come from the original dataset and the above mentioned nearly balanced dataset. A real-world robbery dataset and some public imbalanced datasets are employed to measure the performance of the approach. The results show that the proposed approach is effective in dealing with class imbalances, and it can be extended to combine with other methods solving class imbalances.
... The M.O. is the main basis for criminal behavior. Researchers primarily extract attributes from the M.O. of robbery cases according to the following domains: the action of criminals [16], weapons of criminals [17], ways of threat [18], ways of harm [19], and ways of disguise [20]. ...
Article
Detecting serial crimes is one of the most challenging tasks in crime analysis. Linking crimes committed by the same criminal can improve the work efficiency of police offices and maintain public safety. Previous crime linkage studies have focused on the crime features of modus operandi (M.O.) but did not address the crime process. In this paper, we proposed an approach for detecting serial robbery crimes based on understanding offender M.O. by integrating crime process information. According to the crime narrative text, a natural language processing method is used to extract the action and object characteristics of the crime process, a dynamic time warping method was introduced in the similarity measurement of these characteristics, and an information entropy method was used to weight the similarity of the action and object characteristics to obtain the comprehensive similarity of criminals’ crime process. A real-world robbery dataset is employed to measure the performance of finding serial crimes after adding the crime process information. According to the results, information about the crime process obtained from the case narrative text has significant separability and can better characterize better the offender’s M.O. Five machine learning algorithms are used to classify the case pairs and identify serial cases and nonserial cases. Based on the crime features, the results show that the addition of crime process information can substantially improve the effect of detecting serial crimes.
... Theoretical research has reflected the broad application of crime linkage, finding support for the principles of behavioural consistency and distinctiveness in burglary (Bennell & Canter, 2002;Markson, Woodhams, & Bond, 2010), personal and commercial robbery (Burrell, Bull, & Bond, 2012;Woodhams & Toye, 2007), sexual assault (Harbers, Deslauriers-Varin, Beauregard, & van der Kemp, 2012;Santtila, Junkkila, & Sandnabba, 2005;Woodhams & Labuschagne, 2012), car theft (Davies, Tonkin, Bull, & Bond, 2012;Tonkin, Grant, & Bond, 2008), arson (Santtila, Fritzon, & Tamelander, 2004), and homicide (Salfati & Bateman, 2005;Santtila et al., 2008). Support for these principles has also been found using samples containing several crime types , both unsolved and solved offences (Woodhams & Labuschagne, 2012), one-off and a series of offences (Tonkin, Santtila, & Bull, 2012), and using different methodologies and data from different countries (Ellingwood, Mugford, Bennell, Melnyk, & Fritzon, 2013;. ...
Article
Crime linkage has been the subject of increasing attention in academic research. Research has found support for the principles of behavioural consistency and distinctiveness, which underpin crime linkage, but this does not provide direct evidence as to whether crime linkage is useful in practice. This literature review draws together documentation that refers to the practice of crime linkage, from assessing analysts' efficacy, to discussing the usage of computerised tools to assist with the linkage process, to providing a comprehensive outline of the process itself. The implications of the amount and type of information currently available are discussed, including the variations in practice and terminology that were explored. Avenues for future investigation and the manner in which future research could be conducted are set out in a research agenda.
Article
La presente investigación tiene como objetivo analizar el perfil de los delincuentes que cometen delitos leves de hurto en la ciudad de Barcelona, en función del hurto cometido, ya sea comercial o personal. La muestra fue extraída de diligencias policiales de la Guardia Urbana de Barcelona. Se han estudiado 56 delitos de hurto leve, con un total de 71 delincuentes. Los resultados revelan que los delincuentes comúnmente son hombres, adultos jóvenes, extranjeros, que actúan de manera individual (o grupal en el caso de los más jóvenes), con domicilio en Barcelona y antecedentes penales de reincidencia en el hurto, además de versatilidad criminal para cometer también otro tipo de delitos. Se hallan diferencias en función de si el hurto es comercial o personal. Se concluye que los hurtadores comerciales son más planificados y premeditados, mientras que los personales son más oportunistas y dependen en mayor medida del azar. Finalmente, este estudio facilita la toma de decisiones en el trabajo policial para la prevención del delito, considerando las características y el comportamiento de los delincuentes según el tipo de hurto.
Article
The current study investigated characteristics of robbery offences with a focus on lone versus multiple perpetrator crimes. The main aim was to outline differences between lone, duo, and 3+ perpetrator crimes. Police recorded crime data from West Midlands Police relating to 1574 personal robberies (including attempts) committed in Birmingham city centre (UK) were used in the analysis. Robbery offences were compared on victim and offender demographics and offence characteristics. The findings showed significant differences between the three group sizes (lone, duo, 3+ offenders) in victim and offender age; approach styles; victim control variables; weapon use; alcohol and violence; injuries; and carjacking. The findings of the current research can feed into developing policing strategies which take into account differences in group size.
Article
Full-text available
This study compared the ability of seven statistical models to distinguish between linked and unlinked crimes. The seven models utilised geographical, temporal, and modus operandi information relating to residential burglaries (n = 180), commercial robberies, (n = 118), and car thefts (n = 376). Model performance was assessed using receiver operating characteristic analysis and by examining the success with which the seven models could successfully prioritise linked over unlinked crimes. The regression‐based and probabilistic models achieved comparable accuracy and were generally more accurate than the tree‐based models tested in this study. The Logistic algorithm achieved the highest area under the curve (AUC) for residential burglary (AUC = 0.903) and commercial robbery (AUC = 0.830) and the SimpleLogistic algorithm achieving the highest for car theft (AUC = 0.820). The findings also indicated that discrimination accuracy is maximised (in some situations) if behavioural domains are utilised rather than individual crime scene behaviours and that the AUC should not be used as the sole measure of accuracy in behavioural crime linkage research.
Chapter
Full-text available
This chapter begins by explaining the purposes of linking crimes committed by the same offender and what case linkage can add to a police investigation and prosecution. The various steps involved in the process of case linkage are explained. The assumptions of behavioral consistency and inter-individual behavioral variation, which case linkage rests on, are outlined, and the research that has begun to test these assumptions is reported. The effect of poor-quality data on the case linkage process and on empirical research is examined. Current methods and future developments for overcoming this difficulty are described. The obstacles to identifying linked crimes across police boundaries are discussed. Case linkage research and practice are compared with various criteria for expert evidence with promising results. The chapter closes by considering future avenues for research and practice in case linkage.
Article
Full-text available
Comparative Case Analysis (CCA), typically conducted by crime analysts, uses crime scene behaviours to try to identify series of crimes committed by the same offender. Accurate identification of series of offences allows the police to pool resources and evidence, thereby boosting the potential to identify and apprehend the offender. This paper discusses the results of a survey of a sample of crime analysts working in two UK police forces about their views and experiences of CCA. The results focus on how CCA is conducted, what evidence and information is considered, and how useful CCA is to criminal investigation. Suggestions for how CCA might be developed further are also included.
Article
Full-text available
Street robbery offences can be perpetrated in a variety of ways. Offenders adopt a particular method of conducting the offence that appears to be closely related to the underlying purpose of the offence and the type of victim that is targeted. This paper reviews existing research and presents new findings relating to the "decisions" involved in the commission of street robbery from the perspective of the offenders. Twenty face-to-face interviews were undertaken with offenders convicted of street robbery. Findings specifically focus on the modus operandi employed by offenders based on their knowledge of the risks, struggles and advantages of different ways of working.
Article
Full-text available
In the absence of forensic evidence (such as DNA or fingerprints), offender behavior can be used to identify crimes that have been committed by the same person (referred to as behavioral case linkage). The current study presents the first empirical test of whether it is possible to link different types of crime using simple aspects of offender behavior. The discrimination accuracy of the kilometer distance between offense locations (the intercrime distance) and the number of days between offenses (temporal proximity) was examined across a range of crimes, including violent, sexual, and property-related offenses. Both the intercrime distance and temporal proximity were able to achieve statistically significant levels of discrimination accuracy that were comparable across and within crime types and categories. The theoretical and practical implications of these findings are discussed and recommendations made for future research.
Book
The technique commonly referred to as 'criminal profiling' is one in which crime behaviours are evaluated to predict the likely attributes of an offender. In practice, the technique has enjoyed steady growth and interest over the past three decades. The present volume draws together research from scholars and practitioners from across the globe that examines from a range of multidisciplinary perspectives the use, application and value of this form of profiling. The material canvassed throughout this text not only examines research and theoretical principles for the assessment of crimes, but extends to exploring the range of issues facing the theoretical development of criminal profiling and its practical, legal and professional practice. Topics include the examination of homicidal syndromes, the association of crime scene behavior patterns with offender characteristics, criminal propensity and opportunity, the identification of crimes committed by the same offender, the legal admissibility of profiling, its accuracy, and the problems confronting the contemporary practice of criminal profiling. This volume is likely to become an essential reference for forensic practitioners and students alike operating in the fields of criminology, law and policing and anyone with a genuine interest in the investigation and resolution of violent crime.
Article
This report presents the development and evaluation of the suspect retrieval system based on modus operandi developed by the National Research Institute of Police Science in Japan. The database used in the system stores a large number of records consisting of the modus operandi of prior offenders. A score is assigned to each record, where each score represents the similarity of the modus operandi between each record and a crime under investigation. The similarity is statistically calculated based on the choice probability of each modus operandi. Suspects in the database are rank ordered according to scores. The higher a rank is, the more likely a suspect is expected to have committed the crime under investigation. The validity of the system is evaluated with data about Japanese burglars (n = 12,468) and some factors influencing the accuracy of the retrieval are discussed.