ArticlePDF Available

The linking of burglary crimes using offender behaviour: Testing research cross‐nationally and exploring methodology

Authors:

Abstract and Figures

Purpose. The current study tests whether existing behavioural case linkage findings from the United Kingdom (UK) will replicate abroad with a sample of residential burglaries committed in Finland. In addition, a previously discussed methodological issue is empirically explored. Methods. Seven measures of behavioural similarity, geographical proximity, and temporal proximity are calculated for pairs of burglary crimes committed by 117 serial burglars in Finland. The ability of these seven measures to distinguish between pairs of crimes committed by the same offender (linked pairs) and different offenders (unlinked pairs) is tested using logistic regression and receiver operating characteristic (ROC) analysis. Two methodologies for forming the unlinked pairs are compared; one representing the ‘traditional’ approach used by research and, the other, a new approach that represents a potentially more realistic and statistically sound approach to testing case linkage. Results. A wider range of offender behaviours were able to distinguish between linked and unlinked crime pairs in the current Finnish sample than in previous UK-based research. The most successful features were the kilometre-distance between crimes (the intercrime distance), the number of days separating offences (temporal proximity), and a combination of target, entry, internal, and property behaviours (the combined domain). There were no statistically significant differences between the two methodological approaches. Conclusions. The current findings demonstrate that a wider range of offender behaviours can be used to discriminate between linked and unlinked residential burglary crimes committed in Finland than in the UK. The use of a more realistic and statistically sound methodology does not lead to substantial changes in case linkage findings.
Content may be subject to copyright.
1
Legal and Criminological Psychology (2011)
C2011 The British Psychological Society
The
British
Psychological
Society
www.wileyonlinelibrary.com
The linking of burglary crimes using offender
behaviour: Testing research cross-nationally
and exploring methodology
M. Tonkin1,P.Santtila
2andR.Bull
1
1School of Psychology, Forensic Section, University of Leicester, UK
2Department of Psychology and Logopedics, ˚
Abo Akademi University, Turku,
Finland
Purpose. The current study tests whether existing behavioural case linkage findings
from the United Kingdom (UK) will replicate abroad with a sample of residential
burglaries committed in Finland. In addition, a previously discussed methodological
issue is empirically explored.
Methods. Seven measures of behavioural similarity, geographical proximity, and
temporal proximity are calculated for pairs of burglary crimes committed by 117
serial burglars in Finland. The ability of these seven measures to distinguish between
pairs of crimes committed by the same offender (linked pairs) and different offenders
(unlinked pairs) is tested using logistic regression and receiver operating characteristic
(ROC) analysis. Two methodologies for forming the unlinked pairs are compared; one
representing the ‘traditional’ approach used by research and, the other, a new approach
that represents a potentially more realistic and statistically sound approach to testing
case linkage.
Results. A wider range of offender behaviours were able to distinguish between
linked and unlinked crime pairs in the current Finnish sample than in previous UK-
based research. The most successful features were the kilometre-distance between
crimes (the intercrime distance), the number of days separating offences (temporal
proximity), and a combination of target, entry, internal, and property behaviours (the
combined domain). There were no statistically significant differences between the two
methodological approaches.
Conclusions. The current findings demonstrate that a wider range of offender
behaviours can be used to discriminate between linked and unlinked residential burglary
crimes committed in Finland than in the UK. The use of a more realistic and statistically
sound methodology does not lead to substantial changes in case linkage findings.
The investigation of crime can be enhanced if the police are able to identify crimes that
have been committed by the same person or the same group of people (i.e., link crimes
Correspondence should be addressed to Mr M. Tonkin, School of Psychology, Henry Wellcome Building, University of Leicester,
Lancaster Road, Leicester LE1 9HN, UK (e-mail: mjt25@le.ac.uk).
DOI:10.1111/j.2044-8333.2010.02007.x
2M. Tonkin et al.
as an offence series). The primary benefit of this process is that it allows the evidence
from multiple investigations to be pooled, thus potentially increasing the quantity and
quality of evidence that is available to prosecute offenders (Grubin, Kelly, & Brunsdon,
2001).
One approach that the police use to link crimes is through offender behaviour,
whereby crimes that are committed in a similar way behaviourally are judged to have been
committed by the same offender (linked crimes), whereas crimes that are committed
in different ways are judged to be the work of separate offenders (unlinked crimes)
(Woodhams, Hollin, & Bull, 2007). Behavioural case linkage relies on two theoretical
assumptions about offender behaviour: offenders must demonstrate a degree of similarity
in the way they behave from one crime to the next (behavioural consistency) and their
behaviour must be different from the way in which other offenders behave (behavioural
distinctiveness) (Bennell, Jones, & Melnyk, 2009; Woodhams, Hollin et al. (2007).
Over the last decade, a number of empirical studies have begun to test the validity of
behavioural case linkage with a range of crime types (e.g., Bennell & Canter, 2002; Salfati
& Bateman, 2005; Santtila, Fritzon, & Tamelander, 2004; Tonkin, Grant, & Bond, 2008;
Woodhams & Toye, 2007). This research has focused on a range of issues, including
the exploration of methodology (e.g., Bennell, Gauthier, Gauthier, Melnyk, & Musolino,
2010; Woodhams, Grant et al., 2007), extending the evidence for case linkage to new
crime types (e.g., Tonkin et al., 2008; Woodhams & Toye, 2007), and attempting to
replicate existing findings in new geographical locations and time periods (Bennell &
Jones, 2005; Markson, Woodhams, & Bond, 2010).
But, despite this growing body of research, significant work is still required. For
example, although the case linkage findings relating to residential burglary have been
replicated across the United Kingdom (UK), there has been no attempt to replicate these
findings abroad. Also, the range of research into methodological issues is somewhat
limited. The aims of the current study are, therefore, twofold: (1) to examine the
replicability of existing case linkage findings abroad using a sample of residential
burglaries committed in Finland and (2) to explore a methodological issue that has
been discussed in the literature but is yet to receive substantial attention (Bennell &
Canter, 2002; Woodhams, 2008).
Research into behavioural case linkage with burglary
There now exist a number of studies that have examined behavioural case linkage using
burglary data (e.g., Bennell & Canter, 2002; Bennell & Jones, 2005; Ewart, Oatley, &
Burn, 2005; Goodwill & Alison, 2006; Markson et al., 2010).
This research has found that certain offender behaviours demonstrate sufficient
consistency and distinctiveness to allow linked crimes to be reliably distinguished from
unlinked crimes. The kilometre-distance between offence locations (the intercrime
distance) has been particularly successful in this task, with the intercrime distance
outperforming target characteristics, entry behaviours, internal behaviours (such as
offender search behaviour), and property stolen when differentiating between linked
and unlinked burglaries. These findings have been shown to replicate in various locations
around the UK (Bennell, 2002; Bennell & Jones, 2005; Markson et al., 2010).
The number of days separating burglaries (temporal proximity) has also been shown
to reliably differentiate between linked and unlinked crimes (Markson et al., 2010). In
Markson et al.’s study, the temporal proximity was second only to the intercrime distance
in terms of discrimination accuracy (it outperformed target, entry, internal, and property
Behavioural linking of burglaries 3
behaviours), and the combination of intercrime distance and temporal proximity was
able to facilitate the greatest level of linking success. These findings are corroborated by
studies that have utilized different methodologies (Ewart et al., 2005; Goodwill & Alison,
2006).
The case linkage literature on burglary can be criticized, however, because it has only
ever focused on samples from the UK. It is, therefore, unclear whether the same findings
apply in other countries. Indeed, different countries can be expected to vary in terms of
physical and social geography, the availability, type and distribution of potential targets,
and approaches to policing and data recording/storage, which might lead to variation
in case linkage performance. We briefly review some of these differences now, with
particular focus on differences between the UK and Finland that are relevant to burglary
crime.
Cross-national differences between the UK and Finland
The UK and Finland differ in terms of population density, with Finland averaging
approximately 16 persons per square kilometre compared to 254 persons per square
kilometre in the UK.1Housing is, therefore, much more dispersed in Finland than in
the UK, which would be expected to impact considerably on offender spatial behaviour
such as journey-to-crime and the intercrime distance.
There are also differences between the UK and Finland in terms of housing. The
predominant type of residential accommodation in the UK is a house or bungalow (82% of
households in 2006), whereas in Finland the majority of housing is split across two types
(44% are flats and 40% detached housing in 2003). The slightly wider variation in housing
that is evident in Finland might impact on offender consistency and distinctiveness
because there would be more scope for offenders to target different types of house
and they may need to employ a wider range of entry behaviours. Thus, it might be
hypothesized that case linkage performance for target and entry behaviours would be
enhanced in a Finnish compared to a UK sample.
Also, it is not unreasonable to suggest that police forces in Finland will differ from
those in the UK in terms of how they record information about burglary crime. There
may be additional behaviours recorded in Finland that are not recorded in the UK and/or
the same behaviours may be recorded in different ways. Differences such as these have
the potential to impact on case linkage performance.
Methodology in behavioural case linkage research
Much of the research currently conducted into behavioural case linkage has followed
a methodology originally proposed by Bennell (2002). This involves creating pairs of
linked and unlinked crimes and coding each crime for the presence or absence of a
range of offender behaviours. The ability of these behaviours to distinguish between
linked and unlinked offences is then tested by logistic regression and receiver operating
characteristic (ROC) analysis. The hypothesis is that linked crimes will display greater
behavioural similarity than unlinked crimes, which leads to statistically significant
1All statistics included in this section of the manuscript were obtained from http://en.wikipedia.org/wiki/Main_
Page, http://www.ymparisto.fi/default.asp?node=10078&lan=en, http://www.communities.gov.uk/documents/corporate/
pdf/971061.pdf
4M. Tonkin et al.
discrimination accuracy as measured by these two types of analysis (e.g., Bennell &
Canter, 2002).
Researchers adopting this approach have recently begun to explore how this method-
ology might be improved, including whether more appropriate statistical measures of
consistency and distinctiveness exist (e.g., Bennell et al., 2010) and whether alternative
approaches to forming the linked and unlinked pairs might be utilized (Woodhams,
2008). In terms of the latter, Woodhams (2008) has argued that the current practice
of forming the unlinked pairs from the same sample of crimes that are used to form
the linked pairs may be problematic for several reasons. We will focus on the statistical
criticisms she presents. Woodhams argues that such an approach leads to the assumption
of statistical independence becoming violated during logistic regression. The impact of
violating this assumption is that the confidence interval is spuriously inflated and the
subsequent p-value of any statistical test diminished (Hopkins, 2001). Consequently, the
statistical significance of certain offender behaviours may have been underestimated in
previous research, thus leading to them being inappropriately rejected as valuable case
linkage features.
Another issue that arises from forming the unlinked pairs from the same sample as
the linked pairs is that it leads to a sample consisting solely of serial offences (Bennell
& Canter, 2002; Woodhams, 2008). This is problematic because police crime analysts
who are charged with conducting behavioural case linkage in real life must distinguish
linked crimes from a backdrop of not just serial crimes but non-serial crimes as well. The
current methodology for testing case linkage is not, therefore, a true reflection of the
reality in which case linkage is expected to perform. This is problematic for an area of
applied research such as this that aims to be relevant to police practice.
The current study
The current study, therefore, aimed to explore whether existing case linkage findings for
residential burglary in the UK would replicate cross-nationally with a sample of Finnish
burglaries. Based on previous research, it was hypothesized that the intercrime distance
and temporal proximity would achieve the highest levels of discrimination accuracy.
However, it was also predicted that target and entry behaviours would perform more
successfully than they have in previous UK-based research. The analyses were initially
conducted using Bennell’s (2002) original methodology to ensure comparability with
previous research, but they were also conducted using a methodology whereby the
unlinked pairs were formed from a statistically independent sample of serial and non-
serial crimes. The outcomes from these two methodological approaches were then
compared to determine whether the findings altered when using an approach that is
potentially more ecologically valid and statistically sound. Given the paucity of research
in this area, it was not possible to make specific predictions regarding these comparisons.
Method
The data
To facilitate the replication aspect of this study, 234 solved residential burglary crimes
committed by 117 serial burglars in the Greater Helsinki region of Finland2(between
2The greater Helsinki region of Finland covers an area of approximately 815 km2that contains the capital of Finland,
Helsinki, and the neighbouring cities of Espoo and Vantaa.
Behavioural linking of burglaries 5
1990 and 2001) were extracted from a dataset that had been established during a previous
project (Laukkanen, Santtila, Jern, & Sandnabba, 2008; Santtila, Ritvanen, & Mokros,
2004). The 234 crimes represent a random selection of solved residential burglary
crimes committed during this period. These data were originally collected to facilitate
an investigation of offender and geographical profiling in Finland. Two offences were
randomly selected from the series of each offender, which was necessary to prevent
highly prolific offenders with unusually consistent or inconsistent offence behaviour
having an undue influence on the findings (Bennell, 2002). This dataset is referred to as
dataset one.
To facilitate the analysis of methodology, 508 serial and non-serial burglaries were
extracted from the same dataset described above (Laukkanen et al., 2008; Santtila,
Ritvanen et al., 2004). None of these 508 crimes were included in the 234 burglaries
that were extracted for the purposes of replication, so these two datasets can be
considered statistically independent. Due to the possibility that serial burglaries are
more common in any jurisdiction than non-serial burglaries (Bennell & Canter, 2002;
Goodwill & Alison, 2006), the 508 crimes contained a disproportionate number of serial
to non-serial burglaries. In the absence of any published literature to suggest exactly
how disproportionate serial and non-serial burglaries are, a ratio of approximately 3:1
was used. It was hoped that this approach would contribute to creating a more realistic
research environment in which to test behavioural case linkage. This dataset is referred
to as dataset two.
For each of the crimes in these two datasets a range of behavioural data existed,
including the location of the crime (x,ycoordinates indicating the offence location
to the nearest metre), the date the crime was committed (in many cases this was the
mid-point between an ‘earliest crime date’ and a ‘latest crime date’ because the exact
offence time was unknown, which is not unusual for burglary crime; Ratcliffe, 2002),
the type of property burgled, the method of entry, the search behaviour, and the type
and cost of property stolen (see Appendix for a full list of behavioural data included in
this study).
Apart from the location and temporal information, the data were stored in a binary
format (1 =present in the crime; 0 =absent). The use of binary data is consistent with
previous literature on behavioural case linkage and is justified by findings suggesting
that more complex coding schemes are unreliable with police data (Canter & Heritage,
1990). Satisfactory inter-rater reliability has been reported for the larger dataset from
which the current data were selected (Mdn [median] case-by-case =0.78 and Mdn
variable-by-variable =0.88; Santtila, Ritvanen et al., 2004).
Procedure
The offence behaviours were first grouped into behavioural domains that contained
behaviours that either served a similar function during the offence (e.g., they facilitated
entry into the property), or that occurred at a similar stage of the offence (e.g., they
occurred at the start of the offence when a burglar was selecting the target), or that
represented one ‘type’ of offender behaviour (e.g., spatial behaviour) (see Appendix for
a full listing of which behaviours comprised each domain). Seven behavioural domains
were created: (1) target characteristics (e.g., the type of property burgled); (2) entry
behaviours (e.g., the point and method of entry); (3) internal behaviours (e.g., search
behaviour); (4) property stolen (e.g., cash, keys etc.); (5) the intercrime distance; (6)
temporal proximity; (7) a combined behavioural domain, which included all behaviours
6M. Tonkin et al.
in the target, entry, internal, and property domains. These domains were derived from
previous case linkage studies on burglary and the behaviours were placed into domains
according to their placement in previous research (Bennell, 2002; Bennell & Canter,
2002; Markson et al., 2010).
Pairs of crimes were then created from the two burglary datasets. Initially, 117
linked pairs of crimes were created (one for each offender) from dataset one. Each pair
contained two crimes committed by the same offender that were taken randomly from
each offender’s series. One-hundred-and-seventeen pairs of unlinked crimes were then
created from dataset one, which contained two crimes committed by different offenders.
Finally, a further set of 117 unlinked pairs were created from dataset two.
Having created these crime pairs, each group of pairs (linked dataset one; unlinked
dataset one; and unlinked dataset two) was split into two halves to form a development
sample and a test sample. So, 58 of the 117 linked pairs and 58 of the 117 unlinked
pairs from dataset one were used to create a development sample and the remaining 59
linked and 59 unlinked pairs from dataset one were used to create a test sample. These
development and test samples were used to examine discrimination accuracy using
Bennell’s (2002) original methodology. Also, 58 of the 117 unlinked pairs from dataset
two were used to form a further development sample and the remaining 59 pairs formed
a further test sample. These development and test samples were used in combination
with the samples formed from the linked pairs in dataset one to test discrimination
accuracy using a new methodology that addresses statistical and practical issues with
Bennell’s (2002) existing approach. It should be noted that the larger number of pairs in
the test samples is due to there being an uneven number of offenders.
The procedure of splitting data into development and test samples is known as split-
half validation (Efron, 1982; Gong, 1986), and is discussed by Bennell and colleagues as
a way of reducing the potential bias that might arise from developing and testing linkage
models on the same sample (e.g., Bennell & Jones, 2005).
Data analysis
The first stage of analysis was to calculate the degree of behavioural similarity between the
linked and unlinked crime pairs. To achieve this, a measure called Jaccard’s coefficient
was used, which ranges from 0 (indicating no behavioural similarity) to 1 (indicating
complete behavioural similarity). Jaccard’s coefficient has been used extensively in
previous studies of behavioural case linkage (e.g., Bennell & Canter, 2002; Markson
et al., 2010). Jaccard’s coefficients were calculated for each crime pair in terms of target,
entry, internal, and property behaviours separately, as well as for the combination of
these behaviours. In addition to this, the kilometre-distance and number of days between
the two crimes in each pair were calculated.
The potential value of these seven measures of offender behaviour for distinguishing
between linked and unlinked crimes was then assessed using logistic regression and
ROC analysis (e.g., Bennell, 2002). In order to test the cross-national replicability of
case linkage research, seven direct logistic regression analyses were conducted on the
development sample from dataset one (one regression for each of the seven linkage
features) with linkage status (linked versus unlinked) as the dependent variable and
the linkage features as the independent variables. These analyses allowed the case
linkage performance of each linkage feature to be judged independently from the others
(Woodhams & Toye, 2007). A forward stepwise logistic regression was then conducted,
where all linkage features were entered into the model simultaneously, thus allowing
Behavioural linking of burglaries 7
the optimal combination of linkage features to be identified (Bennell & Canter, 2002).
However, the combined domain was not included in these analyses because this domain
was comprised of a combination of behaviours from the target, entry, internal, and
property domains. Consequently, the inclusion of this variable in the same regression
model as the other domains would risk violating the assumption of multicollinearity,
which can lead to reduced p-values, incorrect regression coefficients and, ultimately, to
incorrect conclusions (Field, 2005). Furthermore, the decision to exclude the combined
domain was consistent with previous research (e.g., Bennell, 2002), which is important
given that one of the primary aims of the current study was to replicate previous work.
Having developed regression models on the development sample, these same models
were used to produce predicted probabilities (ranging from 0 to 1) for each crime
pair in the test samples from dataset one (Bennell & Canter, 2002). These predicted
probabilities were then used as the test variables and linkage status (linked, unlinked) as
the state variable to produce ROC curves for each of the seven single-feature behavioural
domains and for the optimal combination of domains. The Statistical Package for the
Social Sciences (SPSS), version 17.0 (c) (IBM Corporation, NY United States), was used
to produce these ROC curves.
The associated ROC statistic is the area under the curve (AUC), which gives a measure
of discrimination accuracy for the linkage feature/s used to construct the curve (i.e., how
well the offender behaviour discriminated between linked and unlinked crimes). The
AUC can range from 0 (indicating perfect inaccuracy) to 1 (indicating perfect accuracy),
with a value of 0.5 indicating a chance level of accuracy (Bennell & Jones, 2005). AUC
values of 0.50–0.70 are described as a low level of discriminative accuracy, values of
0.70–0.90 are moderate, and values of 0.90–1.00 are high (Swets, 1988). The regression
statistics and AUC values were compared visually with those obtained in previous UK-
based burglary studies to allow the cross-national replicability of case linkage findings to
be examined.
To explore the impact of using a statistically independent sample of serial and non-
serial burglaries to form the unlinked pairs, the same regression and ROC analyses were
run using the linked pairs from dataset one and the unlinked pairs formed from dataset
two. The regression statistics obtained from the first set of analyses were then compared
with these analyses visually and the AUC statistics compared statistically using ROCKIT
1.1B2 C(University of Chicago, IL, United States). SPSS version 17.0 was used to produce
the ROC curves, but it is currently not possible to compare ROC curves statistically
using SPSS, so ROCKIT was used for this purpose. Any differences would suggest that
the choice of methodology impacts on case linkage findings.
Results
A cross-national replication of UK burglary case linkage research
The results from seven direct logistic regression analyses using linked and unlinked pairs
from dataset one are summarized in Table 1. A degree of success was evident for all
seven models of offence behaviour, although some models clearly outperformed others.
The most successful single-feature models were intercrime distance, followed by the
combined domain, then temporal proximity. These models all had highly significant
model 2values and Wald statistics ( p<.001), with between 24% and 57% of the
variability in linkage status explained individually by each of these three behavioural
domains (Brace, Kemp, & Snelgar, 2003; Kinnear & Gray, 2000). Furthermore, all three
models offered an improvement in predictive accuracy above the level one would
8M. Tonkin et al.
Ta b l e 1 . Nine logistic regression models for a sample of Finnish burglars: Bennell’s (2002) methodology
R2(Cox and
Model Constant (SE)Logit(SE)2(df )Wald(df ) Snell-Nagelkerke)
Combined 2.98 (0.65) 9.57 (2.03) 35.68 (1)∗∗∗ 22.13 (1)∗∗∗ 0.27–0.35
Ta r g e t 0.94 (0.29) 2.96 (0.71) 23.99 (1)∗∗∗ 17.61 (1)∗∗∗ 0.19–0.25
Entry 1.17 (0.33) 3.14 (0.74) 22.37 (1)∗∗∗ 17.80 (1)∗∗∗ 0.18–0.23
Internal 1.14 (0.39) 2.86 (0.87) 12.97 (1)∗∗∗ 10.93 (1)∗∗ 0.11–0.14
Property 0.76 (0.39) 2.89 (1.27) 5.48 (1)5.14 (1)0.05–0.06
Intercrime
distance
2.19 (0.51) 0.32 (0.06) 54.77 (1)∗∗∗ 25.95 (1)∗∗∗ 0.43–0.57
Temporal
proximity
1.01 (0.28) 0.00 (0.00) 31.96 (1)∗∗∗ 20.59 (1)∗∗∗ 0.24–0.32
Optimal 1 1.80 (0.69) 80.41 (3)∗∗∗ 0.56–0.75
Intercrime 0.24 (0.07) 12.68 (1)∗∗∗
Temporal 0.00 (0.00) 9.31 (1)∗∗
Target 4.65 (1.60) 8.44 (1)∗∗
Optimal 2 2.77 (0.60) 67.62 (2)∗∗∗ 0.50–0.67
Intercrime 0.26 (0.06) 16.51 (1)∗∗∗
Temporal 0.00 (0.00) 8.39 (1)∗∗
p.05; ∗∗p.01; ∗∗∗p.001.
expect through chance, with each model offering an approximate 23–30% improvement
(see Table 2). In contrast to these models, however, the target, entry, and internal
models performed less favourably. These models explained between 11% and 25% of the
variability in linkage status and offered an approximate 15% improvement in predictive
accuracy above chance. The poorest performance was for the property domain, with just
5–6% of the variability accounted for by this model and a 6% improvement in predictive
accuracy.
The signs of the logit coefficients indicated that linked crimes were characterized by
greater behavioural similarity in terms of combined, target, entry, internal, and property
behaviours and shorter intercrime distance and temporal proximity values than unlinked
crimes.
To determine whether these individual domains could be combined to produce supe-
rior discriminative performance, a forward stepwise logistic regression was conducted.
The stepwise regression proceeded through three steps before it converged on a final
model. The final model (referred to as optimal 1 in Tables 1 and 2 below) contained three
domains (intercrime distance, temporal proximity, and target characteristics), which
accounted for between 56% and 75% of the variance in linkage status and facilitated
an improvement in predictive accuracy of almost 30% above chance (see Tables 1 and
2). These results indicate that optimal model 1 was superior to any of the single-feature
regression models in terms of discriminative performance.
Ta b l e 2 . Predictive accuracy of the models (%): Bennell’s (2002) methodology
Intercrime Temporal Optimal Optimal
Combined Target Entry Internal Property distance proximity 1 2
Random 50.00 50.00 50.00 50.00 50.00 58.60 50.00 58.60 58.60
Model 75.90 65.50 66.40 65.50 56.00 79.80 73.30 86.90 85.90
Behavioural linking of burglaries 9
Ta b l e 3 . Summary of the receiver operating characteristic (ROC) analyses: Bennell’s (2002)
methodology
95% Confidence Classification
Model AUC (SE) interval category
Combined 0.72 (0.05)∗∗∗ 0.63–0.81 Moderate
Target 0.73 (0.05)∗∗∗ 0.64–0.82 Moderate
Entry 0.66 (0.05)∗∗ 0.56–0.76 Low
Internal 0.66 (0.05)∗∗ 0.56–0.76 Low
Property 0.58 (0.05) 0.48–0.69 Low
Intercrime distance 0.84 (0.04)∗∗∗ 0.75–0.93 Moderate
Temporal proximity 0.82 (0.04)∗∗∗ 0.74–0.90 Moderate
Optimal 1 0.86 (0.04)∗∗∗ 0.78–0.93 Moderate
Optimal 2 0.86 (0.04)∗∗∗ 0.79–0.94 Moderate
Note. AUC =area under the curve. Classification categories are according to Swets (1988), where an
AUC value of 0.50 is non-informative, a value of 0.50–0.70 is low, 0.70–0.90 is moderate, and 0.90–1.00
is high. ∗∗p.01; ∗∗∗p.001.
It should be noted that a model combining the intercrime distance and temporal
proximity was able to perform at a similar level to optimal 1 in terms of predictive
accuracy and the percentage of variance explained (see optimal 2 in Tables 1 and 2
below).
To facilitate further comparisons, nine empirical ROC curves were produced (one
for each of the seven single-feature regression models and two for the optimal regression
models). The results are summarized in Table 3.
The ROC results are largely consistent with those obtained from the logistic regression
analyses, with the combined, intercrime distance, and temporal proximity domains
performing well in comparison to the entry, internal, and property domains. Also, the
optimal models performed marginally better than any of the single-feature models. The
fact that similar findings emerged from both the development and test samples suggests
that the current findings are robust and potentially have wider applicability to other
burglary crimes committed in this region of Finland.
However, there were several areas of slight inconsistency between the regression and
ROC findings that are worthy of comment. First, the target domain performed slightly
better in the ROC analyses than would have been expected from the regression analyses.
This suggests that it might be premature to discount the potential independent value
of target characteristics in the linkage task. The second area of slight inconsistency was
the equivalent performance of the two optimal models in terms of the AUC. Based on
the stepwise regression results, one would predict that performance would be superior
in optimal model 1. But, it is important to remember that the difference between these
two optimal models was small in the regression analyses, so it is perhaps unsurprising
that they achieved a similar level of discrimination accuracy in the ROC analyses.
The impact of methodological variation in case linkage research
The findings from seven direct logistic regression analyses using linked pairs from dataset
one and unlinked pairs from dataset two are summarized in Tables 4 and 5. When
these findings are compared with those obtained previously, we see that there is a
trend towards reduced discrimination accuracy in the current set of analyses, with less
10 M. Tonkin et al.
Ta b l e 4 . Nine logistic regression models for a sample of Finnish burglars: new methodology
R2(Cox and
Model Constant (SE)Logit(SE)2(df )Wald(df ) Snell-Nagelkerke)
Combined 2.52 (0.60) 7.80 (1.77) 27.88 (1)∗∗∗ 19.43 (1)∗∗∗ 0.21–0.29
Ta r g e t 0.98 (0.29) 3.18 (0.73) 26.79 (1)∗∗∗ 19.02 (1)∗∗∗ 0.21–0.28
Entry 1.02 (0.33) 2.60 (0.69) 16.83 (1)∗∗∗ 14.33 (1)∗∗∗ 0.14–0.18
Internal 1.09 (0.39) 2.69 (0.85) 11.55 (1)∗∗ 9.91 (1)∗∗ 0.10–0.13
Property 0.55 (0.38) 1.99 (1.22) 2.73 (1) 2.64 (1) 0.02–0.03
Intercrime
distance
1.73 (0.45) 0.25 (0.06) 36.23 (1)∗∗∗ 19.16 (1)∗∗∗ 0.34–0.46
Temporal
proximity
0.93 (0.27) 0.00 (0.00) 28.97 (1)∗∗∗ 17.71 (1)∗∗∗ 0.22–0.30
Optimal 1 2.13 (0.75) 68.02 (3)∗∗∗ 0.55–0.73
Intercrime 0.24 (0.07) 14.00 (1)∗∗∗
Temporal 0.00 (0.00) 9.08 (1)∗∗
Target 3.74 (1.36) 7.57 (1)∗∗
Optimal 2 2.93 (0.65) 58.16 (2)∗∗∗ 0.49–0.66
Intercrime 0.23 (0.06) 14.56 (1)∗∗∗
Temporal 0.00 (0.00) 11.16 (1)∗∗
∗∗p.01; ∗∗∗p.001.
substantial model 2, Wald, and R2statistics for all domains except the target domain.
But, the magnitude of these differences is small. Indeed, when the predictive accuracies
from these two analyses are compared (Tables 2 and 5), none of the domains differ by
more than 5.20%.
A forward stepwise logistic regression was conducted to facilitate further compar-
isons. The stepwise regression proceeded through the same previous three steps before
converging on a final model, which contained the same domains (target, intercrime
distance, and temporal proximity). The only slight difference was in terms of the
performance of the optimal models, whereby a reduced performance was observed
in the current set of analyses (as indicated by the model 2, Wald, and R2statistics).
However, it should be noted that the predictive accuracies in Tables 2 and 5 indicate
an improved rather than a reduced performance. The reason for this contradiction is
probably due to the way in which these measures of model performance are calculated
(Field, 2005).
The analyses thus far indicate that minor differences exist as a function of how the
unlinked pairs are formed. To further examine this issue, nine empirical ROC curves were
created as before. The results are summarized in Table 6 and compared statistically with
the previous ROC results in Table 7. There were no statistically significant differences
in terms of the AUC statistics produced in the two sets of analysis.
Ta b l e 5 . Predictive accuracy of the models (%): new methodology
Intercrime Temporal Optimal Optimal
Combined Target Entry Internal Property distance proximity 1 2
Random 50.00 50.00 50.00 50.00 50.00 52.30 50.00 52.30 52.30
Model 70.70 67.20 66.40 61.20 56.90 76.70 74.10 87.20 87.20
Behavioural linking of burglaries 11
Ta b l e 6 . Summary of the receiver operating characteristic (ROC) analyses: new methodology
95% Confidence Classification
Model AUC (SE) interval category
Combined 0.73 (0.05)∗∗∗ 0.64–0.82 Moderate
Target 0.71 (0.05)∗∗∗ 0.61–0.80 Moderate
Entry 0.66 (0.05)∗∗ 0.56–0.76 Low
Internal 0.72 (0.05)∗∗∗ 0.63–0.81 Moderate
Property 0.55 (0.05) 0.44–0.66 Low
Intercrime distance 0.85 (0.04)∗∗∗ 0.76–0.93 Moderate
Temporal proximity 0.82 (0.04)∗∗∗ 0.74–0.90 Moderate
Optimal 1 0.88 (0.04)∗∗∗ 0.81–0.95 Moderate
Optimal 2 0.89 (0.03)∗∗∗ 0.82–0.96 Moderate
Note. AUC =area under the curve. Classification categories are according to Swets (1988), where an
AUC value of 0.50 is non-informative, a value of 0.50–0.70 is low, 0.70–0.90 is moderate, and 0.90–1.00
is high. ∗∗p.01; ∗∗∗p.001.
Discussion
In the current study, the cross-national replicability of case linkage findings relating to
residential burglary was examined using a sample of burglaries committed in Finland.
There was evidence to suggest that a range of offender behaviours can be used to
distinguish between linked and unlinked crimes. The most successful behaviours were
the intercrime distance, temporal proximity, and the combined domain, which is
somewhat consistent with previous research that has shown the value of the intercrime
distance and temporal proximity for linking burglary crimes committed in the UK (e.g.,
Bennell, 2002; Bennell & Canter, 2002; Ewart et al., 2005; Goodwill & Alison, 2006;
Markson et al., 2010).
However, the magnitude of discrimination accuracy in the current study was larger
for the combined, target, entry, and internal domains than in previous UK-based work.
Most notably, the combined and target domains both achieved AUC values in excess of
0.70, which indicates a moderate degree of discrimination accuracy (Swets, 1988). In
previous work, the AUC values obtained in over 10 UK police jurisdictions have never
exceeded 0.69 for these domains (mean combined AUC =0.65; mean target AUC =0.60;
Ta b l e 7 . A comparison of the area under the curve (AUC) statistics produced using Bennell’s (2002)
methodology and a new methodology
Model Bennell approach AUC New approach AUC
Combined 0.72 0.73
Target 0.73 0.71
Entry 0.66 0.66
Internal 0.66 0.72
Property 0.58 0.55
Intercrime distance 0.84 0.85
Temporal proximity 0.82 0.82
Optimal 1 0.86 0.88
Optimal 2 0.86 0.89
Note. All statistical comparisons were non-significant (p.05).
12 M. Tonkin et al.
Bennell, 2002; Bennell & Canter, 2002; Bennell & Jones, 2005; Markson et al., 2010).
Likewise, the entry and internal domains (AUCs =0.66) compare favourably to previous
research on UK data (mean entry AUC =0.58; mean internal AUC =0.51). These findings
suggest that a wider range of offender behaviours in Finland than in the UK demonstrate
the relative consistency and distinctiveness required to facilitate successful behavioural
case linkage.
There are several potential explanations for these differences. First, it is possible
that Finnish burglars are more consistent and distinctive in their offence behaviour
than burglars from the UK. This might be due to individual differences (such as the
presence of particularly rigid and unique behavioural scripts for offending) and/or due
to environmental differences (such as the availability and diversity of potential targets
with which to offend against). In terms of the former, there is no theoretical basis to
suggest that Finnish and English burglars possess different characteristics that would
be expected to impact on consistency and distinctiveness. In terms of the latter, one
potential environmental factor was discussed earlier that might partially account for the
observed differences. There is a wider variety of housing in Finland than in the UK, which
could allow for between-offender differences in burglary behaviour to emerge more
readily among Finnish than English offenders. A comparison between the current data
and those from Markson et al. (2010) support this suggestion. In the current sample, 65%
of the crimes in dataset one targeted detached housing or second floor apartments (two
separate categories) and the remaining 35% were split across the other four categories
of housing. This compares with 84% of crimes in Markson et al.’s sample that fell under
one category of housing. The wider variation in types of housing targeted by Finnish
burglars might partially account for the superior discrimination accuracy observed in
the current study for target characteristics.
Another explanation is that the UK and Finnish police may differ in terms of
their data recording and storage practices. A comparison between the data available
for the current and previous studies suggests that there may be some value in this
explanation. The internal domain in this study, for example, included the number of
offenders responsible for the crime and how the offender/s exited the crime scene.
These variables were not included in previous research on UK data (Bennell, 2002). It
is plausible that additional behaviours such as these led to the improved case linkage
performance observed in the current study. It is also plausible that the behaviours were
operationalized in a more appropriate way in the current study compared with past
research. For example, the target domain in this study included several variables related
to the owner’s occupancy, whereas in previous research occupancy has been defined
simply in terms of one variable. Differences such as these may also have contributed to
the improved discrimination accuracy in the current study. If these explanations are valid,
then it suggests that the UK police may be able to enhance case linkage performance by
altering the types and nature of information that is recorded on police databases.
However, it is difficult to tease out and test these potential explanations using the
current set of data, which means that it is not possible to draw any definitive conclusions.
Despite the difficulty in definitively explaining these findings, the potential implica-
tions are clear. From a practical perspective, they suggest that the Finnish police can
use a range of offender behaviours to identify linked residential burglary crime series.
But, it seems that they should prioritize the use of the intercrime distance and temporal
proximity in this process because these two features offer the greatest level of predictive
accuracy with a minimum of cost in terms of the time necessary to use these features in
practice. However, in the absence of data relating to these features, it seems that there
Behavioural linking of burglaries 13
is scope for the Finnish police to rely on target, entry, and internal behaviours to link
burglaries.
It is worth highlighting at this point, though, that the AUC values obtained in this
study were not perfect (1.00), so a degree of error can be expected when linking
burglary crimes in practice using these behaviours. Indeed, given the low base rate of
crimes that are linked in real life, high AUC values may not necessarily translate into
error-free prediction (Bennell & Jones, 2005; Szmukler, 2001). This is an important point
for practitioners who are involved in linking crimes to consider.
From a theoretical point of view, the consistency and distinctiveness of intercrime
distance in this study provides support for several seminal theories of offender spatial
behaviour, which suggest that offenders seek to minimize the efforts and risks involved
in offending (e.g., by returning to the same places that are familiar to them and by not
travelling great distances to offend) (Brantingham & Brantingham, 1981, 1984; Clarke &
Felson, 1993; Cohen & Felson, 1979; Cornish & Clarke, 1986; Felson, 1986, 1994).
Furthermore, these findings support previous case linkage research in the UK that has
shown burglary, robbery, and car theft offenders tend to offend in somewhat distinct,
non-overlapping geographical areas (e.g., Bennell & Canter, 2002; Bennell & Jones, 2005;
Tonkin et al., 2008; Woodhams & Toye, 2007). That is, the geographical locations that
one offender chooses to offend in are somewhat different from the areas that a different
offender may choose. A potential explanation for this finding comes from previous
research using this dataset (Laukkanen et al., 2008), which has shown that Finnish
burglars do not travel far from home to offend (a median of 3.88 km). Thus, it may be
that the current sample chose to offend close to home and, by virtue of the fact that the
offenders live in different areas, somewhat distinct, non-overlapping patterns of offender
spatial behaviour emerged.
However, as noted above, the AUC values observed in the current study were
not perfect (1.00), so the offending ‘territories’ of the burglars in this sample were
not completely non-overlapping. Furthermore, the geographical area studied here was
relatively large (815 km2), so it is unclear whether these findings would be replicated at
a smaller geographical scale. But, we might be cautiously optimistic given that similar
findings have been observed with burglars in the UK using study areas that are much
smaller in size (112–230 km2; Bennell & Jones, 2005).
Another main aim of the current study was to explore the impact of using different
methodological approaches to forming the unlinked crime pairs. The regression and
ROC analyses revealed minor differences in discriminative accuracy (none of which
reached statistical significance) when the ‘traditional’ methodology (Bennell, 2002)
was compared with a new methodology that offers a potentially more realistic and
statistically sound approach to testing case linkage. This is reassuring because it suggests
that previous research is valid and that the current set of findings is robust across different
methodological approaches.
Based on these findings and the significant amount of additional work that is required
to utilize an independent sample of crimes, it might be argued that researchers can
continue with Bennell’s (2002) original methodology at present. However, it is important
to conduct further comparisons between the two methodologies using different datasets.
In terms of limitations, the current study shares many of the limitations associated
with previous case linkage research, including the fact that these findings are only
applicable to the geographical and temporal period studied. However, the use of split-
half validation suggests that these findings may be applicable to other similar areas in
Finland. Another limitation is that the sample was restricted to solved crime. Researchers
14 M. Tonkin et al.
have discussed the fact that this may artificially inflate discrimination accuracy and that
this approach moves research away from the reality in which case linkage is expected to
perform (i.e., with unsolved crimes) (Bennell, 2002). Future research should address this
issue by testing case linkage with samples of unsolved crime. Nonetheless, the current
study has contributed to the growing body of case linkage literature by extending the
evidence to a new country and by exploring new methodological issues in a systematic
way.
The challenge now, however, is to test findings such as these in a prospective way by
applying them in practice to ongoing police investigations to determine whether they
can truly yield improved behavioural case linkage performance. Indeed, this would seem
particularly pertinent given recent evidence that statistically significant AUC statistics
do not necessarily translate into substantial practical accuracy in discriminative tasks
(Szmukler, 2001).
Acknowledgement
The authors would like to acknowledge the kind assistance of Dr Jessica Woodhams who
commented on an earlier version of this manuscript.
References
Bennell, C. (2002). Behavioural consistency and discrimination in serial burglary. Unpublished
doctoral dissertation. Liverpool, UK: University of Liverpool.
Bennell, C., & Canter, D. V. (2002). Linking commercial burglaries by modus operandi: Tests
using regression and ROC analysis. Science and Justice,42, 153–164. 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 coeffi-
cients used in behavioural linkage analysis. Forensic Science International,199, 85–92.
doi:10.1016/j.forsciint.2010.03.017
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 Profiling,
2, 23–41. 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, 293–310. doi:10.1348/135532508X349336
Brace, N., Kemp, R., & Snelgar, R. (2003). SPSS for psychologists: A guide to data analysis using
SPSS for Windows (2nd ed.). Basingstoke: Macmillan.
Brantingham, P. J., & Brantingham, P. L. (1981). Environmental criminology. Beverly Hills: Sage.
Brantingham, P. J., & Brantingham, P. L. (1984). Patterns in crime. New York: Macmillan.
Canter, D., & Heritage, R. (1990). A multivariate model of sexual offences behaviour:
Developments in ‘offender profiling’. Journal of Forensic Psychiatry,1, 185–212.
doi:10.1080/09585189008408469
Clarke, R. V., & Felson, M. (1993). Routine activity and rational choice. New Brunswick, NJ:
Transaction.
Cohen, L. E., & Felson, M. (1979). Social change and crime-rate trends: A routine activity approach.
American Sociological Review,44, 588–608. Retrieved from: http://www.jstor.org/stable/
2094589
Cornish, D. B., & Clarke, R. V. (1986). The reasoning criminal: Rational choice perspectives on
offending. New York: Springer-Verlag.
Efron, B. (1982). The jackknife, the bootstrap and other re-sampling plans. Philadelphia, PA:
Society for Industrial and Applied Mathematics.
Behavioural linking of burglaries 15
Ewart, B. W., Oatley, G. C., & Burn, K. (2005). Matching crimes using burglars’ modus operandi: A
test of three models. International Journal of Police Science and Management,7, 160–174.
Retrieved from: http://www.vathek.com/ijpsm/contents.php?vi=7.3
Felson, M. (1986). Linking criminals’ choices, routine activities, informal control, and criminal
outcomes. In D. Cornish & R. V. Clarke (Eds.), The reasoning criminal: Rational choice
perspectives on offending (pp. 119–128). New York: Springer-Verlag.
Felson, M. (1994). Crime and everyday life: Insights and implications for society. Thousand
Oaks, CA: Pine Forge Press.
Field, A. (2005). Discovering statistics using SPSS (2nd ed.). London: Sage.
Gong, G. (1986). Cross-validation, the jackknife, and the bootstrap: Excess error estimation in
forward logistic regression. Journal of the American Statistical Association,81, 108–113.
Retrieved from: http://www.jstor.org/stable/2287975
Goodwill, A. M., & Alison, L. J. (2006). The development of a filter model for prioritizing suspects in
burglary offences. Psychology, Crime & Law,12, 395–416. doi:10.1080/10683160500056945
Grubin, D., Kelly, P., & Brunsdon, C. (2001). Linking serious sexual assaults through behaviour
(Home Office Research Study 215). London: Home Office Research, Development and Statistics
Directorate.
Hopkins, W. G. (2001). A new view of statistics. Retrieved from: http://www.sportsci.org/
resource/stats/modelsdetail.html
Kinnear, P. R., & Gray, C. D. (2000). SPSS for Windows made simple. Hove: Psychology Press.
Laukkanen, M., Santtila, P., Jern, P., & Sandnabba, K. (2008). Predicting offender home
location in urban burglary series. Forensic Science International,176, 224–235.
doi:10.1016/j.forsciint.2007.09.011
Markson, L., Woodhams, J., & Bond, J. W. (2010). Linking serial residential burglary: Comparing
the utility of modus operandi behaviours, geographical proximity, and temporal proxim-
ity. Journal of Investigative Psychology and Offender Profiling,7, 91–107. doi:10.1002/
jip.120
Ratcliffe, J. H. (2002). Aoristic signatures and the spatio-temporal analysis of high volume crime
patterns. Journal of Quantitative Criminology,18, 23–43. doi:10.1023/A:1013240828824
Salfati, C. G., & Bateman, A. L. (2005). Serial homicide: An investigation of behavioural consistency.
Journal of Investigative Psychology and Offender Profiling,2, 121–144. doi:10.1002/
jip.27
Santtila, P., Fritzon, K., & Tamelander, A. L. (2004). Linking serial arson incidents on the
basis of crime scene behavior. Journal of Police and Criminal Psychology,19, 1–16.
doi:10.1007/BF02802570
Santtila, P., Ritvanen, A., & Mokros, A. (2004). Predicting burglar characteristics from crime scene
behaviour. International Journal of Police Science & Management,6, 136–154. Retrieved
from: http://www.vathek.com/ijpsm/contents.php?vi=6.3
Swets, J. A. (1988). Measuring the accuracy of diagnostic systems. Science,240, 1285–1293.
Retrieved from: http://www.ncbi.nlm.nih.gov/pubmed/3287615
Szmukler, G. (2001). Violence risk prediction in practice. The British Journal of Psychiatry,178,
84–85. Retrieved from: http://bjp.rcpsych.org/cgi/reprint/178/1/84
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 Profiling,5, 59–
77. doi:10.1002/jip.74
Woodhams, J. (2008). Understanding juvenile sexual offending: An investigative perspective.
Unpublished doctoral dissertation. Leicester, UK: University of Leicester.
Woodhams, J., Grant, T. D., & Price, A. R. G. (2007). From marine ecology to crime analysis:
Improving the detection of serial sexual offences using a taxonomic similarity measure. Journal
of Investigative Psychology and Offender Profiling,4, 17–27. doi:10.1002/jip.55
Woodhams, J., Hollin, C. R., & Bull, R. (2007). The psychology of linking crimes:
A review of the evidence. Legal and Criminological Psychology,12, 233–249.
doi:10.1348/135532506X118631
16 M. Tonkin et al.
Woodhams, J., & Toye, K. (2007). An empirical test of the assumptions of case linkage and offender
profiling with serial commercial robberies. Psychology, Public Policy, and Law,13, 59–85.
doi:10.1037/1076-8971.13.1.59
Received 5 October 2010; revised version received 29 November 2010
Appendix
Content dictionary of offence behaviours and behavioural domains
Target characteristics
A detached house
1st floor of a multi-storey building
2nd floor or above of a multi-storey building
A studio flat
A terraced or semi-detached house
Other target
Owners present
Owners temporarily away (i.e., less than 24 hours)
Owners away 1–3 days
Owners away 3 +days
Target has safeguards present (alarm, security light, dog etc.)
Target in an urban city area
Entry behaviours
Door
Back door
Balcony door
Window
Mailbox
Open or unlocked door
Manual force
Breaking glass
Climbing (above street level)
Lock
Key
Tool
Crowbar
Hook
Sharp weapon
Garden tool
Screwdriver or spike
Brick or stone
Tool brought to the scene
Tool used from the scene or the immediate vicinity
Internal behaviours
Interrupted
Interrupted by a guard or the owner
Behavioural linking of burglaries 17
Fingerprints, footprints, or DNA left at the scene
Tools used in the burglary left at the scene
1 offender
2 offenders
3+offenders
Tidy search
Untidy search
Only first room entered was searched
Whole target searched
Drawers/cabinets opened and searched
Drawers pulled out and contents possibly thrown on floor
Inner doors opened using force
Property piled up to be carried away
Stolen items hidden close by
Stolen items abandoned
Used facilities (consumed food/drink, used toilet/shower, defecated/urinated)
Exit by car
Exit on foot
Property stolen
Cash
Credit or bank cards, cheques, bank book, shares
Firearms, ammunition, explosives
Sharp weapons (not cutlery)
Watches, wristwatches
Small-size consumer electrical items
Large electrical equipment, musical instruments that need to be carried with both hands
Tapes, CD’s LP’s, videotapes
Jewellery
Fake jewellery (costume)
Prescription medication
Tobacco products/smoking tools
Cosmetic, hygiene products
Alcohol
Plates, cups, cutlery, and other utensils
Food
Clothes
Purses, hand bags, suitcases, backpacks
Wallet
Keys (home, car)
Identity documents (e.g., passport, driving licence, library card etc.)
Spectacles, sunglasses, or other optical items
Antique or art objects
Construction tools or materials
Porcelain, crystal glass, silverware
Games or sports equipment
Vehicle
Items stolen could be carried by one person
18 M. Tonkin et al.
Stolen items worth less than 170 Euros
Stolen items worth up to 1700 Euros
Stolen items worth more than 8400 Euros
Combined
Contains all behaviours listed above under target characteristics, entry behaviours,
internal behaviours, and property stolen.
Intercrime distance (in kilometres)
The distance in kilometres between two crime sites.
Temporal proximity (in days)
The number of days between two crime dates.
... The behaviour of foraging burglary offenders has not previously been researched in isolation in respect of its ability to predict linkage between either foraging or traditional burglary offences. Previous studies have categorised these behaviours with that of entry behaviour to define what is commonly referred to as a modus operandi or MO Grant et al, 2008 andDavies andTonkin et al, 2012). This study disconnects the physical entry characteristics from other behaviours displayed by the offenders and as such provides new insights. ...
... However, this result was not entirely unexpected as multiple other studies Grant et al, 2008 andDavies andTonkin et al, 2012) have concluded that traditional MO characteristics, which the offender behaviours are drawn from, provide the lowest levels of prediction accuracy. One possible explanation for this result is that all but one of the offender behaviours in this study are reliant on being identified by either eye witness evidence i.e. multiple offenders and use of a motor vehicle or being identified through a thorough forensic examination of the scene i.e. the wearing of gloves to mask the presence of finger prints. ...
... Previous research Grant et al, 2008 andDavies andTonkin et al, 2012;Tonkin and Santtila, 2011) has been conducted into the crime linkage behaviors of domestic burglary offenders. Similarly, to hotspot research, this study is the first to extend this method of examination to crimes committed by foraging burglary offenders and in doing so enables increased understanding of linkage indicators. ...
Thesis
Full-text available
Drawn from ecology, the optimal forager predictive policing methodology has been identified as the primary tasking tool used by police services to tackle domestic burglary. Built upon established findings that the target selection behaviour of foraging domestic burglary offenders can be predicted, this thesis examines the physical offending and geographical characteristics of foraging offenders in greater detail. This study evolves established research evidence by drawing upon criminological methods that have potential to increase the approaches effectiveness before testing their applicability in respect of foraging criminals. Ecological research evidence relating to assumptions of foraging behaviour are used to devise theoretical manifestations within criminal behaviour which are subsequently tested for and used to build a theoretical model to combat them. The study achieves all of this through a number of key research chapters, these include (1) identifying predictive thresholds for linking burglary offences committed by foraging criminals (2) drawing on existing assumptions within ecology the study then seeks to identify their presence within foraging criminals, including the presence of significant crime displacement, and (3) geographical profiling is identified and tested as a potential solution to combat the evasive behaviour of foraging offenders as a response to the increased police presence that the optimal forager model is designed to co-ordinate. Underpinning the study throughout is an examination of the enablers and blockers present that impact upon the effectiveness of such transitions of theory into practice. Overall, the thesis provides new theoretical material by creating a framework of foraging offender typologies. The key practical implications for policing include a model for tackling the identified theoretical foraging typologies to increase the crime prevention and reduction efforts in respect of domestic burglary.
... However, the link in respect of predicted foraging offenders was slightly less (AUC = 0.89) than that that identified in previous studies (Bennell and Jones 2005) of linked dwelling burglaries (AUC = 0.90). In respect of target selection, previous studies (Bennell and Jones 2005;and Tonkin et al. 2011) examining randomly selected linked burglary dwellings have identified wide varying degrees of prediction accuracy in respect of target selection, with AUCs of 0.58 and 0.73, respectively; however, both were below the results identified within this study (AUC = 0.76). In an effort to optimise the predictive accuracy the two optimal characteristics of target selection and inter-crime distance were combined. ...
... This study disconnects the physical entry characteristics from other behaviours displayed by the offenders while committing the crime and as such provides new insights on behaviour committed by burglars during the offence. Only one other study has attempted to do this by examining the internal search behaviour of offenders (Tonkin et al. 2011) in which a strong discriminatory prediction accuracy was identified (AUC = 0.66). This study however identified a very low predictive accuracy from analysing the personal behaviour of the foraging offenders. ...
Article
Full-text available
Crime linkage is a systematic way of assessing behavioural or physical characteristics of crimes and considering the likelihood they are linked to the same offender. This study builds on research in this area by replicating existing studies with a new type of burglar known as optimal foragers , who are offenders whose target selection is conducted in a similar fashion to foraging animals . Using crimes identified by police analysts as being committed by foragers this study examines their crime scene behaviour to assess the level of predictive accuracy for linking crimes based on their offending characteristics. Results support previous studies on randomly selected burglary offence data by identifying inter-crime distance as the highest linking indicator, followed by target selection, entry behaviour, property stolen and offender crime scene behaviour. Results discuss distinctions between this study and previous research findings, outlining the potential that foraging domestic burglary offenders display distinct behaviours to other forms of offender (random/marauder/commuter).
... Whereas CCA is concerned with structured queries within large databases such as ViCLAS to find similar cases [1], CLA is a process where a behavioural investigative advisor (BIA) examines a smaller number of two or more offences with regard to their linkage potential. The empirical crime linkage literature has rather focussed on CCA scenarios and amassed a substantial body of research (Burrell et al., 2012;Davies et al., 2012;Ellingwood et al., 2013;Tonkin et al., 2011a;Tonkin et al., 2011b;Salo et al., 2012;Santtila et al., 2005;Santtila et al., 2008;Woodhams and Labuschagne, 2011;Woodhams and Toye, 2007) that has demonstrated two main findings: First, crime linkage is a worthwhile endeavour as large scale studies such as Woodhams et al. (2019b) have proved that the theoretical assumptions of consistency and distinctiveness do hold up in international samples of serial sexual assaults. Second, various statistical algorithms do perform objectively well in linking crimes committed by the same offender. ...
... As in other studies, a separate optimal threshold for maximizing the amount of correct links and maximizing the number of correct rejections (i.e. Youden's index) was calculated for each ROC curve (Bennell and Jones, 2005;Markson et al., 2010;Tonkin, et al., 2011a;Tonkin et al., 2011b;Woodhams and Toye, 2007). All ROC analyses were conducted with MedCalc for Windows, version 12.0. ...
Article
Purpose Traditional crime linkage studies on serial sexual assaults have relied predominantly on a binary crime linkage approach that has yielded successful results in terms of linkage accuracy. Such an approach is a coarse reflection of reality by focussing mainly on the outcome of an offence, neglecting the forceful differences due to the intricate offender-victim interaction. Only few researchers have examined sexual assaults through the lens of a sequence analysis framework. This paper aims to present the first empirical test of offence sequence-based crime linkage, moving beyond exploratory analyses. Design/methodology/approach Offence accounts from 90 serial sexual assault and rape victims from the UK were analysed and sequentially coded. Sequence analysis allowed to compare all offences combinations regarding their underlying sequence of events. The resulting comparison was transformed and plotted in two-dimensional space by multidimensional scaling analysis for a visual inspection of linkage potential. The transformed proximities of all offences were used as predictors in a receiver operating characteristic analysis to actually test their discriminatory accuracy for crime linkage purpose. Findings Sequence analysis shows significant discriminatory accuracy for crime linkage purpose. However, the method does perform less well than previous binary crime linkage studies. Research limitations/implications Several limitations due to the nature of the data will be discussed. Practical implications The practical limitations are as follows: the study is a potential practical value for crime analysts; it is a complimentary methodology for statistical crime linkage packages; it requires automated coding to be useful; and it is very dependent on crime recoding standards. Originality/value The exploratory part of this study has been published in a book chapter in 2015. However, to the best of the authors’ knowledge, the succinct test of crime linkage accuracy is the first of its kind.
... Since Bennell et al.'s (2010) review, researchers have also taken steps to improve the ecological validity of their research studies. For example, behavioural consistency and distinctiveness have been found across samples including both solved and unsolved offences, and one-off and serial offences (Slater et al., 2015;Winter et al., 2013;Woodhams et al., 2019), as well as using a wider range of statistical analyses and comparing samples from different countries Tonkin et al., 2012). ...
Article
Purpose The Violent Crime Linkage Analysis System (ViCLAS) is a computerised database which is used by law enforcement in several countries to find potential links between serial violent crimes. In 2012, Bennell, Snook, MacDonald, House and Taylor identified a number of assumptions that must be valid for these computerised systems to be effective. Design/methodology/approach This paper revisits and expands on these assumptions with specific reference to the use of ViCLAS, looking at research that has been conducted since this 2012 review and outlining where research is still outstanding. Findings The importance of evaluating ViCLAS is highlighted in this paper. Practical implications Particularly, the research agenda highlights how the practice of comparative case analysis when using ViCLAS could be improved. Originality/value To the best of the authors’ knowledge, this is the first review of the research dedicated specifically to the evaluation of ViCLAS.
... There is also evidence that offenders might be differentiated by their crime and routine activity locations. Nearby crimes are more likely to have been committed by the same offender than different offenders [38][39][40][41][42]. That all people exhibit highly distinctive spatiotemporal routine activity patterns [43], suggests that among offenders, activity spaces may also be distinctive. ...
Article
Full-text available
It is well established that offenders' routine activity locations (nodes) shape their crime locations, but research examining the geography of offenders' routine activity spaces has to date largely been limited to a few core nodes such as homes and prior offense locations, and to small study areas. This paper explores the utility of police data to provide novel insights into the spatial extent of, and overlap between, individual offenders' activity spaces. It includes a wider set of activity nodes (including relatives' homes, schools, and non-crime incidents) and broadens the geographical scale to a national level, by comparison to previous studies. Using a police dataset including n=60,229 burglary, robbery, and extra-familial sex offenders in New Zealand, a wide range of activity nodes were present for most burglary and robbery offenders, but fewer for sex offenders, reflecting sparser histories of police contact. In a novel test of the criminal profiling assumptions of homology and differentiation in a spatial context, we find that those who offend in nearby locations tend to share more activity space than those who offend further apart. However, in finding many offenders' activity spaces span wide geographic distances, we highlight challenges for crime location choice research and geographic profiling practice.
... Combining domestic and other violence is likely to have masked differences that our model predicts, due to their typical settings (home/public places). 7 Here the discrete choice research converges with literature on the near repeat phenomenon (e.g., Bernasco, 2008;Johnson et al., 2009) and crime linkage (e.g., Tonkin et al., 2011Tonkin et al., , 2012, which confirms that offences that are close in space and time are more likely to have been committed by the same offender. This pattern is often explained in terms of a 'boost' mechanism, whereby successful crime commission causes a follow-up crime by motivating the offender to return to the same location, and a 'flag' mechanism, whereby repeated offending at the same location (possibly but not necessarily by the same offenders) is merely a symptom indicating that the location continuously provides criminal opportunities (Johnson et al., 2009;Lantz & Ruback, 2017;Pease, 1998). ...
Article
Full-text available
This paper extends Crime Pattern Theory, proposing a theoretical framework which aims to explain how offenders' previous routine activity locations influence their future offence locations. The framework draws on studies of individual level crime location choice and location choice in non-criminal contexts, to identify attributes of prior activities associated with the selection of the location for future crime. We group these attributes into two proposed mechanisms: reliability and relevance. Offenders are more likely to commit crime where they have reliable knowledge that is relevant to the particular crime. The perceived reliability of offenders' knowledge about a potential crime location is affected by the frequency, recency and duration of their prior activities in that location. Relevance reflects knowledge of a potential crime location's crime opportunities and is affected by the type of behaviour, type of location and timing of prior activities in that location. We apply the framework to generate testable hypotheses to guide future studies of crime location choice and suggest directions for further theoretical and empirical work. Understanding crime location choice using this framework could also help inform policing investigations and crime prevention strategies.
Article
In many countries, data collection on sexual violence incidents is not integrated into the healthcare system, which makes it difficult to establish the nature of sexual offences in this country. This contributes to widespread societal denial about the realities of sexual violence cases and the collective oppression of survivors and their families. Capturing detailed information about incidents (e.g., characteristics of perpetrators, where it happened, victims, and the offence) can dispel myths about sexual violence and aid in crime prevention and interventions. This article examines how information about sexual violence incidents—in particular, offences committed against children in Kenya—is gathered from two different data sources: the Violence Against Children Survey (VACS) and data collected by the Wangu Kanja Foundation (WKF), a survivor-led Kenyan NGO that assists sexual violence survivors in attaining vital services and justice. These two surveys provide the most comprehensive information about sexual and gender-based violence. The analysis indicates that, while the VACS provides information about the prevalence of sexual violence, it provides less detailed information about the nature of violence (e.g., characteristics of perpetrators, victims, and the offence) compared with the WKF dataset. We critically reflect on how validity and informativeness can be maximised in future surveys to better understand the nature of sexual violence, as well as other forms of gender-based violence, and aid in prevention and response interventions/programming.
Thesis
Full-text available
Behavioural crime linking refers to the practice of trying to tie two or more offences to the same offender using behaviour observable at the crime scene. It rests on the assumptions that offenders behave consistently enough from one offence to another, and distinctively enough from other offenders allowing offences to be successfully linked together. Conceptualised in the 70s, and developed methodologically with increased scientific rigour from the 90s, the last decade has seen a sharp rise in published studies on behavioural crime linking. From empirical validation of the underlying assumptions to mapping out practice and more ecologically valid tests of linkage accuracy, the field has developed considerably. Considering that investigating homicide is resource intensive, not to mention serial homicide, reliable and valid behavioural crime linking has the potential to aid and prioritise investigative avenues and help solve serial homicide. Most studies on serial homicide have been carried out on North American samples. While some research has studied the consistency and distinctiveness of serial homicide offenders, few have empirically tested models of behavioural crime linking and linkage accuracy with serial homicide. Another shortcoming in behavioural crime linking research is the use of mostly serial cases to model crime linking, even though real crime databases include both serial and one-off offences. Some studies have tested the effect of added one-offs on the linkage accuracy of burglary and rape, but none so far the effect this would have on homicide. Additionally, while some studies have compared serial homicide offences to one-off homicides, none have tested whether it would be possible to predict whether a homicide belongs to a series or is a singular offence. Cognitive bias, especially confirmation bias or the expectancy effect, has been shown to have a considerable effect on crime investigation. No studies to date have explored the effect of such bias in behavioural crime linking. The general aim of the thesis was to increase ecological validity of behavioural crime linking research, especially with regard to sampling choices and analyses that strive to answer questions relevant for homicide investigation. The main sample consisted of 116 Italian serial homicides, committed in 23 separate series of homicide. Additionally, information about 45 cases of hard-to-solve one-off homicide was gathered, coded, and added to the sample. Study I found seven behavioural dimensions of offending (e.g., sexually motivated homicides and aspects of control-behaviour) in line with previous research. Notably, also other motives than sexual were found in the killings. A majority of offences (63%) were correctly classified to their actual series in the predictive part of the study. Study II was an experiment that investigated whether knowledge of series membership increased perceived (coded) behavioural similarity in homicides committed by the same offender. While no support was found for a strong expectancy effect, the experimental task may have lacked in sufficient complexity, and replication is thus needed. Study III found several key differences between serial and singular homicides and was able to successfully use these differences to predict with good accuracy whether an offence was part of a series. Study IV combined all the advances in the methodology thus far and showed that behavioural crime linking was still viable even with a large proportion (10:1) of one-off homicides added into the sample. As a function of added one-off homicides, the specificity of the model worsened (more false positives), as did the proportion of offences belonging to a series found near the top of a ranked listing from more behaviourally similar to less behaviourally similar. Overall model accuracy remained good, though, further validating the practice of behavioural crime linking with more ecologically valid data. The studies of the present thesis contribute to the methodology of behavioural crime linking research. Replication on local crime databases is needed to maximise the practical usefulness of the models in different jurisdictions. Going forward, a close-knit collaboration between researchers and practitioners is called for, to keep the research relevant for practice and to develop evidence-based practice. As we gain a clearer picture of the accuracy and error rate of behavioural crime linking models, their usefulness increase in both the criminal investigative phase and in the trial phase with behavioural crime linking being presented as expert evidence.
Chapter
Crime linkage can be a useful tool in the investigation of sexual offenses when other, physical evidence is unavailable or too costly to process. It involves identifying behavior that is both consistent and distinctive, and thus forms an identifiable pattern through which a series of offenses committed by the same offender can be distinguished. While there is a substantial body of research to support the principles of crime linkage, samples often contain only one type of sexual offense, and further research is needed into offenses such as voyeurism and exhibitionism. In practice, there are a number of ways in which crime linkage can be conducted, and a variety of terms are used to describe these different processes. While writings from practitioners provide insight into how crime linkage is conducted, research now needs to focus more on systematically mapping its practice and documenting procedural differences. There are also a number of additional considerations that require further research attention where the practice of crime linkage is concerned, such as the utility of computerised databases designed to assist with the process, the human decision-making element of linking and how bias can affect this, and the effects of expertise and training on linkage efficacy.
Article
Purpose The purpose of this paper is to understand (i) how crime linkage is currently performed with residential burglaries in New Zealand, (ii) the factors that promote/hinder accurate crime linkage and (iii)whether computerised decision-support tools might assist crime linkage practice. Design/methodology/approach A total of 39 New Zealand Police staff completed a questionnaire/interview/focus group relating to the process, challenges, products and uses of crime linkage with residential burglary in New Zealand. These data (alongside four redacted crime linkage reports) were subjected to thematic analysis. Findings The data clearly indicated wide variation in crime linkage process, methods and products (Theme 1). Furthermore, a number of factors were identified that impacted on crime linkage practice (Theme 2). Research limitations/implications Future research should develop computerised crime linkage decision-support tools and evaluate their ability to enhance crime linkage practice. Also, researchers should explore the use of crime linkage in court proceedings. Practical implications To overcome the barriers identified in the current study, greater training in and understanding of crime linkage is needed. Moreover, efforts to enhance the quality of crime data recorded by the police will only serve to enhance crime linkage practice. Social implications By enhancing crime linkage practice, opportunities to reduce crime, protect the public and deliver justice for victims will be maximised. Originality/value The practice of crime linkage is under-researched, which makes it difficult to determine if/how existing empirical research can be used to support ongoing police investigations. The current project fills that gap by providing a national overview of crime linkage practice in New Zealand, a country where crime linkage is regularly conducted by the police, but no published linkage research exists.
Article
Full-text available
The aim of the present study was to predict offender characteristics on the basis of crime scene behaviour in urban burglaries. The police files of 633 burglaries in the Finnish Metropolitan Area between 1990 and 2001 were content analysed using a predetermined list of variables. The crime scene behaviour variables were subjected to a principal component analysis. Fourteen factors indicative of different types of burglaries were identified and used to predict the characteristics of the 244 offenders using regression models. Statistically significant predictors of almost all offender characteristics were identified. From a practical point of view, the predictive models could be used in police investigations to narrow down the number of suspects.
Article
This paper discusses the development of a filter model for prioritizing possible links in dwelling burglary. The filters utilize the central aspects of crime scene information that is available and accessible to investigators in burglary, namely geo-spatial, temporal, behavioural, and dwelling information. The proposed filters were analysed using a sample of 215 dwelling burglaries committed by 43 serial burglars (i.e. 5 offences each) in order to determine the sequence in which the filters should be considered in prioritizing possible linked offences. The results indicated that the following order (i.e. better performance to worse performance) was most effective at linking offences, utilizing: (1) geo-spatial information, (2) temporal aspects, (3) behavioural information and, lastly, (4) dwelling characteristics. Specifically, the results indicated that offences in close proximity to one another should be given priority. Further, any offence occurring within a 28-day span before or after the index offence should be given priority. The paper argues that behavioural and dwelling characteristics are less effective for linking than geo-spatial and temporal information because the former two aspects are influenced significantly by situational and contextual cues on offender decision-making.
Article
’Hard’ forensic evidence (eg DNA) may be the best means of linking crimes, but it is often absent at burglary crime scenes. Modus operandi information is always present to some degree, but little is known of its significance in matching burglaries. This paper evaluates the ability of three algorithms to match a target crime to the actual offender within a database of 966 offences. The first (RCPA) uses only MO information, the second (RPAL) only temporal and geographic data and a third (COMBIN) is a combination of the two. A score of one indicates a perfect match between the target crime and the case selected by the algorithm. The lowest possible rank is 965 showing that 965 cases were selected before the target offence. The RPAL and COMBIN each achieve a perfect match for 24 per cent of the crimes and succeed in matching over half of the crimes at a score of 10 or less. For prolific offenders, using MO information alone is better than temporal and geographic data, although the best performance is achieved when in combination. Behavioural, spatial and temporal information is collected by many Police Services. The value and means of utilising such data in linking crimes is clearly demonstrated.