Space-Time Modelling to support
local policing
Robert Haining
Department of Geography,
University of Cambridge, England.
AAG; New York; Feb 2012
1
INTRODUCTION
Recent trends in policing in the UK (Crime and Disorder
Act 1998):
• greater use of intelligence-led, proactive, interagency, crime prevention measures;
• recognition of the importance of people’s fear of
crime and of “re-assurance”.
The importance of the neighbourhood for delivering
crime reduction strategies and tackling the fear of crime.
• targeted strategies (no cold calling zones; crime
hotspot responses);
• tailored strategies (situational crime prevention
measures; reassurance policing).
2
Together with the availability of geo-coded offence, offender, victim data =>
- opportunities to explore the importance of place and space in
crime and criminality;
- provide inputs into policing practice at different scales:
• risk mapping and modelling;
• crime hotspot detection;
• policy assessment.
Bayesian Hierarchical Models:
• exploiting spatial and temporal autocorrelation;
• exploring spatial and temporal heterogeneity.
3
4
PROJECT 1:
Policy evaluation: ‘no-cold calling’
zones in Peterborough, England.
Guangquan Li, Sylvia Richardson and Nicky Best
Imperial College, London
5
Background to project
• Cold calling is a visit or a telephone call to a consumer
by a trader, whether or not the trader supplies goods
or services, which takes place without the consumer
expressly requesting the contact.
• Not illegal but often associated with “rogue trading”;
and doorstep cold calling associated with burglary.
• Creation of “no cold calling” (NCC) zones.
6
• Setting
up an NCC zone combines:
•“re-assurance policing”
•Reducing fear of crime by prioritising a crime
and disorder threat that concerns neighbourhood
residents.
• situational crime prevention
• reducing crime by altering the “environment” so
as to (i) reduce the opportunity to offend and (ii)
increase the risk of getting caught if the motivated
offender chooses to offend.
• has roots in Rational Choice Theory (Cornish
and Clarke, 1986); Routine Activity Theory
(Cohen and Felson, 1979)
7
• advice to residents on how to deal with cold callers;
• police presence through street and household signage;
• Trading Standards approve legitimate cold calling;
• Target hardening;
• Higher levels of surveillance; information sharing
But do these schemes reduce burglaries in the targeted NCC
zones?
8
NCC areas in Peterborough
9
Data on “No Cold Calling”
10
Raw data: aggregated temporal profile
4
6
8
Overall (without NCC)
Individual NCC
Aggregated NCC
2
Positive
impact of
policy?
0
Annual DwellBurglary risk per 100 dwellings
2005/2006 NCC groups (10 COA)
2001
2002
2003
2004
2005
year
2006
2007
2008
11
12
Constructing control groups
• To form a control group, areas are selected on the basis of
having similar local characteristics (e.g., burglary rates or
deprivation scores) to those in the NCC-targeted group.
– Lower Super Output Areas (LSOAs) are the basic units.
ID
Matching criterion
No. of LSOAs
1
All LSOAs in Peterborough
88
2
±10% burglary rate of the NCC group in 2005
9
3
±20% burglary rate of the NCC group in 2005
20
4
±30% burglary rate of the NCC group in both 2004 and 2005
8
5
LSOAs containing the NCC-targeted COAs (but excluding the
NCC-targeted COAs)
9 (one LSOA is outside
Peterborough)
6
LSOAs that had “similar” multiple deprivation scores (MDS)
to those for the NCC LSOAs in 2004
46
13
14
15
16
17
18
19
Some wider questions arising from the evaluation:
•
Data limitations
• burglary count as the measure of success
• short length of time series
•
Impacts before implementation (publicity effects?)
•
Displacement effects? (diffusion of benefits?; net effect?)
•
: threshold or dilution?
20
• How generalizable are these findings?:
Urban schemes
Rural schemes
12UDGQ0021 (0.54)
12UEHH0013 (0.68)
12UCGE0010 (0.036)
00JAPA0012 (0.77)
12UEGR0010 (0.072)
00JANK0015 (0.134)
12UCGA0013 (0.78)
12UEHL0012 (0.164)
00JAND0001 (0.83)
00JANB0005 (0.223)
12UDGS0014 (0.81)
12UEHP0017 (0.215)
12UEHS0014 (0.227)
12UDGQ0006 (0.83)
12UEGN0016 (0.236)
12UCGD0002 (0.86)
12UEGP0011 (0.251)
00JANG0024 (0.87)
12UEGQ0016 (0.306)
00JANH0009 (0.281)
12UCGA0003 (0.87)
00JANB0001 (0.309)
12UDGQ0024 (0.87)
00JANB0003 (0.398)
NCC2007
00JANG0009 (0.92)
12UCGF0002 (0.397)
00JANB0009 (0.471)
00JAPB0010 (0.72)
NCC2006
00JANG0025 (0.96)
12UGHX0020 (0.491)
12UEHR0020 (0.516)
12UEGQ0003 (0.544)
00JANY0010 (0.45)
12UCGF0013 (0.527)
00JANC0016 (0.47)
12UEHL0011 (0.544)
00JANE0006 (0.79)
12UCFW0016 (0.566)
00JANK0014 (0.590)
00JANG0013 (0.81)
12UCGF0001 (0.608)
00JANE0010 (0.84)
NCC2007
12UCGM0003 (0.653)
00JANT0027 (0.90)
12UEHR0012 (0.299)
00JANQ0023 (0.91)
NCC2005
12UEHR0014 (0.520)
00JANH0003 (0.649)
12UEHR0017 (0.70)
Overall (0.95)
NCC2006
Overall (0.29)
−100
−50
0
50
100
−200
Percentage change in burglary rate compared to controls
−100
0
100
200
Percentage change in burglary rate compared to controls
21
Not all neighbourhoods respond well to targeted strategies (Gillham 1992):
• persistently high rates of particular crimes;
• residents want to “fight back”;
• high levels of social cohesion.
In some cases schemes can cause an increase in crime (Pease 1999).
22
PROJECT 2:
Stable and Time Varying
Components of Variation in
Burglary Rates in Peterborough,
England.
Guangquan Li, Sylvia Richardson and Nicky Best
Imperial College, London
23
Domestic burglaries cluster in space and time
reflecting:
-Concentrations of events (crime hotspots – more
deprived neighbourhoods)
- repeat victimization (same or nearby houses –
more affluent neighbourhoods)
(Bowers 2004)
Less well studied is how stable these patterns are
over time.
24
25
Summary of raw data
Pairwise
correlation plots
show positive
correlations in the
annual burglary
rates.
26
• The basic idea behind the SCM is
a) to extract the spatial pattern that is shared across
time periods (linked to stable attributes of areas
e.g. housing type) and;
b) to capture spatial risk patterns that are specific to
each time period (linked to changeable attributes
e.g. population changes).
27
The shared component modelling (SCM)
framework
28
The estimated shared component of the model
2005
2007
2006
2008
2005
2007
2006
2008
The figures on the right hand side show which of the local
burglary rates are consistently above/below the Peterborough
29
average.
The estimated time-specific patterns
2005
2007
2006
2008
2005
2007
2006
2008
The time-specific maps show which of the local burglary rates
are above/below the Peterborough average in each year.
30
• shared = “predictable” or “stable” pattern
– help with longer term targeting of policing
resources (strategic planning).
– different from “hot-spot analysis” (or cluster
analysis) which tends to be useful for immediate
action (tactical deployment).
• specific = “unpredictable” or “unstable”
pattern
– May be difficult to interpret without expert
knowledge; specific circumstances at the time
31
SCM with covariates
32
Posterior probabilities: shared pattern
• Maps in the first row highlight LSOAs with consistently and
“significantly” higher risk than the Peterborough average.
• Maps in the second row highlight LSOAs with consistently
33
“significantly” higher risk after adjusting for IMD and JSA.
Posterior probabilities of the shared pattern
scaled to 2006 (adjusted for IMD and JSA)
Area 11
(p=0.72)
Area 6
(p=0.95)
34
Shared component modelling can also be applied
to multivariate space-time data (e.g. two or
more crimes over several time periods).
35
Benefits of multilevel/hierarchical models
• Multilevel/hierarchical models offer a natural framework
to combine information and hence to strengthen
estimation.
• The sparsity issue often encountered in analysing georeferenced and/or time-series data can be addressed by
“smoothing” over space or time or both.
• Idea is to “borrow information” from neighbouring
areas or time periods to produce better (more stable,
less noisy) estimates in each area.
• Modelling spatial or temporal structure achieved by
appropriate choice of random effects distribution.
36
Benefits of the Bayesian framework
• Since all parameters are treated as random variables
with associated posterior distributions, probability
statements about the parameters can be easily made,
– the probability of success
– the probability that the burglary rate of an area is
above the Peterborough average
• Uncertainty can be naturally accounted for and/or
propagated if necessary:
– e.g., uncertainty associated with the reference trend
estimates can be propagated into measuring the NCC
policy's impact.
37
Implementation and other issues
• Estimation of BHM requires computationally
intensive simulation methods (McMC)
– Implemented in free WinBUGS and GeoBUGS
software: www.mrc-bsu.cam.ac.uk/bugs
– Need to ensure the McMC chains have converged to
the target distributions.
– Free software INLA (Rue et al, 2008) implements fast
approximation: www.r-inla.org
• Assessing sensitivity of the results to different
prior specifications.
38