Lancaster-Warwick-Stirling node

As a contribution to this, researchers from Lancaster-Warwick-Stirling node of NCRM have conducted a descriptive analysis of the spatio-temporal distribution of crime in Lancashire. This note summarises the statistical model and the results of the analysis. The data used cover the period April 2003 to March 2009 inclusive. Information on each recorded crime consists of its location, time and type of crime, classified as criminal damage (51% of crimes), serious acquisitive crime (30%) and other wounding crime (19%). To preserve confidentiality, location is aggregated into lower super-output areas (LSOA’s). In a mixed urban-rural county such as Lancashire, LSOA’s are highly variable in geographical size to achieve relatively constant population size (average 1,500 per LSOA). The population size and relevant covariates for each LSOA were extracted from the 2001 national census. The time of reporting of each crime is recorded to the nearest minute, but the time of occurrence is not recorded. For this reason, and to eliminate time-of-day effects, our analysis uses a time-resolution of one day. Our methodological approach is to derive the properties of the resulting spatiotemporal count data from an underlying spatially and temporally continuous point process model. To estimate the “normal” pattern of incident crimes we consider a multiplicative decomposition of the local (in space and time) intensity of crimes into three components. The first two correspond to spatially averaged temporal variation and temporally averaged spatial variation, and are modelled by loglinear regressions. The third component corresponds to residual spatio-temporal variation and is modelled as a latent stochastic process, whose covariance structure describes any spatio-temporal variation not captured by the available covariates.

 

The regression components of the model reveal strong and statistically significant day-of-week effects for crime-types criminal damage and other wounding. For criminal damage the effects are highest for Mondays, followed by Sundays and Saturdays. On other week-days, crime-rates are between 20% and 30% lower than on Sundays. For other wounding, the difference between weekdays and weekends is greater. For serious acquisitive crime, day-of-week effects and seasonal effects are smaller and only marginally statistically significant. All three crime types show seasonal variation in incidence and a decreasing trend over the study-period that we model as logquadratic. Seasonal effects are strongest for criminal damage, whilst the decreasing trend is most pronounced for serious acquisitive crime.

 

The available spatial covariates are the density of licensed premises, the deprivation rates for income and employment, and domains of the index of multiple deprivation corresponding to health and disability, education skills and training, barriers to housing and services, and living environment. The effect of density of licensed premises is highly significant for all three types of crime, whereas the significance of the effects of the individual domains of deprivation are not consistent over the three categories of crime. Crime rates are generally higher in and near urban areas. Rates in the other wounding category show a strong peak at each city centre. For the criminal damage category, the gradation in incidence from rural to urban areas is less strong, and for the serious acquisitive crime category even less so. For serious acquisitive crime, incidence also appears to be generally higher in the southern part of the county than in the northern part. Direct examination of animations of the data indicates that the urban areas have consistently high rates, whereas departures from zero in the rural areas can be characterised more as sporadic events. The pattern of crime rates varies considerably over the 14 districts in Lancashire. We consider the results relating to criminal damage in three representative districts: Blackpool, Preston and Lancaster. These districts vary in size, overall crime rates, decreasing trends and seasonal pattern. All three show a highly significant association between local crime rates and density of licensed premises. Examining associations between deprivation effects and crime rates we conclude that health, barriers to housing and services, and living environment deprivation effects are significant for Blackpool; income, employment, education and health effects are significant for Preston; income, employment, health, and barriers to housing and services effects are significant for Lancaster.

 

The results relating to the stochastic model for residual spatio-temporal variation also suggest that the covariance functions for each of the three crime categories are separable, with marginal correlations following exponential decay with increasing separation in time and space. The residual spatio-temporal correlation is strongest for serious acquisitive crime and weakest for other wounding. We conjecture that this might be explainable by the spatial pattern of land-use types affecting the rates of acquisitive crimes, whereas the other wounding category includes crimes that are the result of violent interactions between individuals, irrespective of their location.

 

In future work, we intend to use the descriptive model described in this note as the basis for developing a real-time surveillance system to detect emergent clustering of criminal activity and inform decisions on short-term tactical responses to these emergent clusters.

 

For further details please see http://www.lancs.ac.uk/staff/diggle/MADE