TL;DR
This study quantifies the impact of COVID-19 lockdowns on crime patterns in London using Bayesian Structural Time Series (BSTS) and Geographically Weighted Logistic Regression (GWLR). It reveals significant shifts in crime, particularly a decrease in acquisitive crimes and an increase in expressive crimes, with distinct regional variations. The findings offer data-driven insights for optimizing policing strategies and addressing socio-economic factors contributing to crime.
What Problem Does This Paper Address?
The COVID-19 pandemic and subsequent lockdowns drastically altered daily life, leading to predictable changes in human behavior and, consequently, crime patterns. While previous research has explored these changes, many studies offered a pan-city analysis, overlooked regional disparities, or treated the entire lockdown period as a single event. This paper addresses these gaps by:
- Quantifying and analyzing the impact of COVID-19 on London's crimes at a finer, regional scale (MSOAs).
- Differentiating the effects of London's three distinct lockdown periods.
- Identifying the underlying socio-economic and demographic factors driving these crime changes.
Understanding these nuanced impacts is crucial for developing effective, data-driven crime prevention and deduction strategies in the post-pandemic era.
Methods
The study employs a multi-stage methodological approach:
- Exploratory Data Analysis (EDA): Initial analysis of crime data to identify overall trends and seasonality using Seasonal and Trend decomposition using LOESS (STL). Spatial analysis using Spatial Point Pattern Test (SPPT) and local Moran's I was used to assess changes in crime distribution and identify hotspots.
- Quantifying Lockdown Impacts (BSTS): The core of the temporal analysis uses a Bayesian Structural Time Series (BSTS) model. This model is novel in this context for its ability to:
- Model counterfactual crime trends (what crime would have looked like without lockdowns).
- Incorporate covariates like temperature, holidays, and crucially, a restriction-level time series to capture the varying intensity of lockdowns.
- Allow for separate models for each of the three distinct lockdown periods, enabling a comparative analysis of their unique impacts.
- Identifying Driving Factors (LR & GWLR): Based on the BSTS results (specifically, identifying MSOAs with significant crime increases), Logistic Regression (LR) and Geographically Weighted Logistic Regression (GWLR) models were used.
- GWLR is a key innovation here, as it accounts for spatial autocorrelation, meaning it models how relationships between factors and crime can vary across different geographic regions (MSOAs), offering a more localized understanding than traditional LR.
Datasets
- Crime Data: Monthly crime counts for eleven crime types (three acquisitive: burglary, robbery, theft; two expressive: drug offences, violence against the person) across 982 London MSOAs, spanning from March 2011 to July 2021. This data was provided by the Metropolitan Police and sourced from data.police.uk.
- COVID-19 Restriction Data: A time series of restriction levels provided by the Greater London Authority, documenting measures like school closures, shop closures, and stay-at-home orders.
- Temperature and Holiday Data: Monthly average temperatures and counts of UK bank holidays from 2011 to 2021.
- Demographic and Socio-Economic Data: Primarily derived from the 2021 UK Census at the MSOA level, including variables like age, gender, population density, unemployment rates, and poverty indicators.
- Mobility Data: Google Mobility data to assess the reduction in mobility across London boroughs during the pandemic.
- Stop and Search Data: Data on stop and search instances, also from data.police.uk, to evaluate its potential role.
Key Insights and Ideas
- Lockdowns weren't monolithic: The study highlights that the three distinct lockdown periods had varying impacts, with the first and third lockdowns showing more pronounced effects on acquisitive crimes.
- Acquisitive vs. Expressive Crime Divergence: A clear pattern emerged: acquisitive crimes (burglary, theft, robbery) generally decreased during lockdowns, while expressive crimes (drug offences, violence) saw a significant increase. This aligns with routine activity theory (fewer opportunities for acquisitive crime) and general strain theory (increased stress leading to expressive crime).
- Spatial Heterogeneity is Key: Crime changes were not uniform across London. GWLR revealed that the factors influencing crime increases varied significantly by MSOA, underscoring the need for localized policing strategies.
- Unemployment as a Major Driver: Prolonged unemployment emerged as a significant predictor for increased acquisitive crimes (especially robbery) and expressive crimes (drug offences), validating the general strain theory's emphasis on economic and psychological pressures.
- Age and Density Matter: Younger populations were associated with increased burglary and theft, while population density played a role in violent crime increases. Conversely, older populations seemed to have a protective effect against burglary and drug offenses.
- Stop-and-Search's Nuanced Role: While stop-and-search showed a consistent role in curbing drug offenses, its impact on theft was less clear, suggesting a need for careful resource allocation.
Key Results
- Temporal Impact: Lockdowns 1 and 3 had the most significant impact on acquisitive crimes, with reductions seen in 34% and 28% of MSOAs, respectively. Lockdown 2 had a more mixed effect. Expressive crimes consistently escalated during lockdowns.
- Spatial Impact: Significant crime reductions were concentrated in central London areas like Westminster, while increases were observed in peripheral areas, often coinciding with new hotspots for expressive crimes.
- Factor Contributions:
- Burglary: Reduced risk associated with older populations and increased risk with higher proportions of minors and population density.
- Robbery: Primarily driven by long-term unemployment, with drug offenses also playing a role in lockdown 2.
- Theft: Reduced risk associated with older populations and those with higher employment rates.
- Drug Offences: Significantly increased with higher proportions of youth and unemployed individuals. Stop-and-search consistently reduced the risk.
- Violence Against the Person: Increased risk associated with higher proportions of males and population density.
Limitations and Future Work
- Data Granularity: While MSOA-level data was used, finer-grained data (e.g., street-level) could offer even more localized insights.
- Causality: While the study identifies correlations and uses BSTS for inferring causal impact, establishing definitive causality remains complex.
- Generalizability: Findings are specific to London; similar studies in other cities would be needed to assess broader applicability.
- Future Work: The authors suggest further investigation into the role of non-residential land use and residents in the clustering of expressive crimes, and a deeper exploration of the interrelationships between different crime types. They also propose exploring the effectiveness of interventions beyond policing, such as enhanced unemployment support.
Summary updated