Burglary in London: Insights from Statistical Heterogeneous Spatial Point Processes

Jan Povala, Seppo Virtanen, Mark GirolamiJournal of the Royal Statistical Society Series C (Applied Statistics) 2020 · DOI

TL;DR

This paper introduces a novel Bayesian mixture model for analyzing spatial point patterns, specifically applied to burglary data in London. By extending mixture models with spatial dependencies and allowing for localized covariate effects, the authors provide a more interpretable and accurate approach than standard methods for understanding crime patterns.

What Problem Does This Paper Address?

Understanding and predicting criminal activity, particularly burglary, is crucial for effective policing and policy-making. Traditional statistical models often struggle to capture the complex spatial nature of crime, which exhibits both spatial dependence (nearby events are more related) and spatial heterogeneity (crime patterns vary across different areas, like urban centers vs. residential neighborhoods). Existing methods like Log-Gaussian Cox Processes (LGCPs) can model spatial dependence but often fail to adequately account for this localized heterogeneity, leading to less interpretable results. This paper aims to bridge this gap by developing a model that explicitly addresses both spatial dependence and heterogeneity, offering deeper insights into the factors driving burglary.

Methods

The core contribution is a Spatially-Aware Mixture of Generalized Linear Models (SAM-GLM). This Bayesian model is a finite mixture of Poisson generalized linear models. The key innovations are:

  • Mixture Components for Heterogeneity: Each location is probabilistically assigned to one of several "groups" (mixture components). Each group has its own set of regression coefficients, allowing for localized effects of covariates (e.g., the impact of "number of households" might differ significantly between a residential area component and a city-center component).
  • Block-Based Spatial Dependence: To model spatial dependence, the study region is divided into blocks (e.g., census tracts). The mixture weights (the probability of a location belonging to a specific group) are modeled using Gaussian Processes that operate at the block level, allowing for spatial correlation between nearby blocks. This is more computationally tractable than modeling dependence at the individual cell level.
  • Bayesian Inference: The model is estimated using Markov Chain Monte Carlo (MCMC) methods, which allows for quantifying uncertainty in the estimates.

The authors also explore a simplified version where blocks are treated independently, offering a trade-off between computational complexity and model flexibility.

Datasets

The primary dataset used is monthly spatial point pattern data of residential and non-residential burglaries in London, published by UK police forces. For analysis, this data was aggregated into a regular grid.

  • Training Data:
    • One-year dataset (2015)
    • Three-year dataset (2013-2015)
  • Test Data:
    • 2016 (for one-year training)
    • 2016-2018 (for three-year training)

Additional covariates used include:

  • Land use data
  • Socioeconomic census data (2011)
  • Points of Interest (POI) data (2018)

Key Insights and Ideas

  • Burglary is Spatially Heterogeneous: The model successfully identifies distinct "types" of burglary patterns across London, suggesting that different factors drive crime in different areas. This moves beyond a one-size-fits-all approach.
  • Interpretable Components: The mixture components offer a way to interpret localized effects. For instance, one component might highlight areas where "number of households" and "house price" are key drivers (suggesting residential burglary driven by opportunity and reward), while another might focus on "points of interest" and "accessibility" (suggesting commercial or public space crime).
  • Covariate Importance Varies by Component: The importance of covariates like "number of households" or "points of interest" is not uniform across London; their influence is captured by the specific mixture component active in a given area.
  • Spatial Dependence is Important, but Block-Level is Sufficient: While spatial dependence is crucial, modeling it at the block level (e.g., MSOAs) proved sufficient and more computationally efficient than modeling it at a finer grid level.

Key Results

  • Improved Predictive Performance: SAM-GLM consistently outperformed the standard Log-Gaussian Cox Process (LGCP) in terms of out-of-sample prediction (measured by held-out log-likelihood and RMSE), especially for the one-year dataset. For the three-year dataset, SAM-GLM matched LGCP performance with fewer components.
  • Better Hotspot Prediction: The model showed competitive or superior performance in predicting crime hotspots (measured by PAI and PEI) compared to LGCP.
  • Reduced Overdispersion and Excess Zeros: The mixture approach effectively handled issues of overdispersion and excess zeros present in standard Poisson models, leading to a better model fit (e.g., an 81% decrease in Pearson $\chi^2$ statistic).
  • Computational Tractability: While complex, the proposed Bayesian MCMC approach was found to be more computationally tractable than LGCP for similar performance levels, particularly due to the block-based approach.
  • Identified Key Drivers: The model identified that "number of households" and "points of interest" are generally strong predictors of burglary intensity, but their relative importance varies significantly across the identified spatial components.

Limitations and Future Work

  • Identifiability: The authors acknowledge the common issue of label-switching in mixture models, where the order of components can be arbitrary. They address this by inspecting traceplots to ensure consistent interpretation.
  • Computational Complexity: While more tractable than LGCP in some aspects, the MCMC inference for the full SAM-GLM can still be computationally intensive, especially with a large number of blocks and components. The authors suggest exploring more efficient inference schemes like variational methods or alternative Gaussian Process factorizations.
  • Covariate Specification: Future work could explore forcing certain covariates to share coefficients across components or investigating spatially varying coefficient processes.
  • Data Granularity: The analysis was performed on aggregated grid data. Future work could explore finer-resolution data if available and anonymization allows.
  • Temporal Dynamics: The current model focuses on spatial patterns over aggregated time periods. Extending it to explicitly model temporal dynamics could be a valuable future direction.

Summary updated

Abstract

Summary To obtain operational insights regarding the crime of burglary in London, we consider the estimation of the effects of covariates on the intensity of spatial point patterns. Inspired by localized properties of criminal behaviour, we propose a spatial extension to mixtures of generalized linear models from the mixture modelling literature. The Bayesian model proposed is a finite mixture of Poisson generalized linear models such that each location is probabilistically assigned to one of the groups. Each group is characterized by the regression coefficients, which we subsequently use to interpret the localized effects of the covariates. By using a blocks structure of the study region, our approach enables specifying spatial dependence between nearby locations. We estimate the proposed model by using Markov chain Monte Carlo methods and we provide a Python implementation.
BibTeX
@article{povala2020burglary,
  author = {Jan Povala and Seppo Virtanen and Mark Girolami},
  title = {{Burglary in London: Insights from Statistical Heterogeneous Spatial Point Processes}},
  journal = {Journal of the Royal Statistical Society Series C (Applied Statistics)},
  year = {2020},
  doi = {10.1111/rssc.12431},
}
Figure