Empirical evidence of the impact of mobility on property crimes during the first two waves of the COVID-19 pandemic

The study examines how government stay-at-home (SAH) orders during COVID-19 altered mobility and, in turn, affected registered property offences (robbery, burglary, theft) in Tamil Nadu (TN), India, and its capital city Chennai across two waves (2020, 2021). Wave two was markedly more severe in infections and fatalities. The research question is whether and how changes in mobility induced by varying lockdown intensities (complete and partial) causally impacted property crime levels, and whether effects differed across urban (Chennai) versus state-wide (TN) contexts. The purpose is to provide empirical, counterfactual evidence using official daily FIR counts over a decade and assess implications for policing when restrictions are tightened or lifted. The study is important due to sparse peer-reviewed, primary-data analyses from India and because it informs resource prioritization and preventive strategies during and after large-scale restrictions.

Global evidence indicates pronounced declines in acquisitive/property crimes following stringent SAH orders, with heterogeneous effects by offence type and geography. Studies across the US reported declines in residential burglary and larceny, mixed burglary trends, and increases in auto theft and homicide in some contexts. In England and Wales, property crimes like shoplifting and robbery dropped sharply, while public order and violent crimes rose; second lockdown impacts were less pronounced than the first. Ireland saw declines during lockdowns with rebounds post-lockdown and persistent elevations in cybercrime. Middle-income contexts (Mexico City, Rio de Janeiro, Buenos Aires) showed reductions in several crimes tied to reduced mobility and government presence, though reporting behavior can confound trends. India saw sharp drops in visible property crimes during lockdowns. Theoretical framing draws on Routine Activities Theory (RAT) and crime opportunity theory: crime requires offender-target convergence without capable guardianship and rational choices based on reward-risk tradeoffs. Lockdowns reduced mobility and social interactions, altering opportunity structures, particularly affecting instrumental/property crimes. Prior work (e.g., in China, New York, Los Angeles, London, Sydney) found differential impacts across acquisitive crimes linked to target distribution and guardianship, supporting RAT and opportunity perspectives.

Design: Counterfactual time-series analysis using daily FIR counts for robbery, burglary, and theft from 1 Jan 2010 to 31 Dec 2021. Two geographies: Tamil Nadu (state-wide; 1356 police stations) and Chennai (urban; 115 stations). Lockdown phases: Wave 1 (2020): CL-2020 (23 Mar–30 Apr), PL-2020 (1 May–8 Jun), gradual relaxations, Post-L-2020 (1–30 Sep). Wave 2 (2021): PL-One-2021 (10 Apr–5 May), CL-2021 (6 May–7 Jun), PL-Two-2021 (8 Jun–6 Jul), Post-L-2021 (1–30 Sep). Mobility measurement: Google Community Mobility Reports (six domains: retail & recreation, parks, grocery & pharmacy, workplaces, transit stations, residential) expressed as percentage change relative to 5-week baseline (3 Jan–6 Feb 2020) matched by weekday. Data limitations acknowledged: relative (not absolute) measures, device/user opt-in bias, short baseline, population shifts, changing place-learning. Crime reporting limitations (under-reporting) discussed; assumed relatively constant pre/post for causal inference. Modeling: Forecasting via Auto-Regressive Recurrent Neural Network (ARNN) using the GluonTS DeepAR model with RNN architecture (LSTM/GRU) and negative binomial likelihood for count data. Training: 1 Jan 2010–31 Dec 2019. Validation (unseen): 1 Jan 2020–22 Mar 2020. Prediction/evaluation: 23 Mar 2020–31 Dec 2021. No covariates used in the DeepAR forecasts; mobility is analyzed in association and in effect-size adjustments post hoc. Forecast generation: At each step, sample from output distribution; median of samples fed to next step; repeated n times to form probabilistic forecasts and confidence intervals. Accuracy metric: Weighted Mean Absolute Percentage Error (WMAPE) on validation period to avoid division-by-zero issues of MAPE. Causal impact estimation: Interrupted time-series approach. The counterfactual series (DeepAR forecast) represents expected crime absent pandemic restrictions. Causal effect computed as difference between actual and predicted counts over each phase. Statistical testing: Normality via Shapiro test. If violated, Wilcoxon signed-rank test for significance. Effect size: Cliff’s Delta (non-parametric), with interpretations (negligible, small, medium, large). Additional comparative mobility analyses across phases with effect sizes to characterize differences in mobility between wave 1 and wave 2, and between PL and CL phases. Focus offences: Theft, burglary, robbery (dacoity and murder for gain excluded due to sparse counts).

- Pandemic severity: Mean daily infections/fatalities in wave 2 were ~5–6× those in wave 1 (infections: 12,265 vs 2,403; fatalities: 153 vs 34), overwhelming health infrastructure. - Mobility: Sharp declines from baseline in retail & recreation, workplaces, and transit stations during CL/PL, with larger reductions in wave 1 than wave 2; residential mobility increased. Chennai generally showed larger mobility reductions (and larger residential increases) than TN. Effect-size comparisons indicated mobility during CL-2020 < CL-2021 (i.e., larger reductions in 2020), and PL-2020 < PLs-2021, often with medium to large Cliff’s Delta. - Model performance: ARNN (DeepAR) achieved lowest validation WMAPE in 2/3 series and was competitive in robbery. Validation WMAPE: Robbery 0.358 (ARNN), Burglary 0.288 (best), Theft 0.141 (best), outperforming ARIMA, Holt-Winters, BSTS, GAM in most cases. - Tamil Nadu crime changes vs counterfactual (percent change; Cliff’s Delta): • Robbery: CL-2020 −83% (−0.983, large); PL-2020 −25% (−0.456, medium); Post-L-2020 +30% (0.389, medium). PL-One-2021 −6% (−0.21); CL-2021 −67% (−0.867, large); PL-Two-2021 −14% (−0.30); Post-L-2021 +56% (0.476, medium-large). • Burglary: CL-2020 −69% (−0.99, large); PL-2020 −47% (−0.78, large); Post-L-2020 +23% (0.49, medium-large). PL-One-2021 −19% (−0.49, medium-large); CL-2021 −59% (−0.98, large); PL-Two-2021 −27% (−0.48, medium); Post-L-2021 +12.6% (0.26, small-medium). • Theft: CL-2020 −80% (−1, very large); PL-2020 −54% (−0.98, large); Post-L-2020 ~0% (−0.09). PL-One-2021 −24% (−1, very large); CL-2021 −64% (−0.76, medium-large); PL-Two-2021 −35% (−0.76, medium-large); Post-L-2021 −10% (−0.25, small-medium). - Chennai crime changes vs counterfactual (percent change; Cliff’s Delta): • Robbery: CL-2020 −75.8% (−1, very large); PL-2020 −47.5% (−0.944, large); Post-L-2020 +7.2% (0.591). PL-One-2021 −37.1% (−0.376); CL-2021 −78.4% (−0.879, large); PL-Two-2021 −32.7% (−0.341); Post-L-2021 +23.4% (0.25). • Burglary: CL-2020 −56.5% (−1, very large); PL-2020 −37.6% (−0.889, large); Post-L-2020 +17.2% (0.102). PL-One-2021 −3.5% (−0.09); CL-2021 −44.1% (−0.889, large); PL-Two-2021 −13.3% (−0.375); Post-L-2021 −18.5% (0.065). • Theft: CL-2020 −84.0% (−1, very large); PL-2020 −61.9% (−1, very large); Post-L-2020 +17.9% (0.714). PL-One-2021 −34.0% (−0.622); CL-2021 −66.3% (−1, very large); PL-Two-2021 −43.7% (−0.686); Post-L-2021 +15.6% (0.75). - Patterns: • Across both geographies, CL phases caused drastic declines in property crimes; PL phases caused moderate declines; Post-L phases often saw increases, especially robberies in 2021 (TN +56%; Chennai +23%). • Theft exhibited the largest magnitude declines across phases (often Cliff’s Delta near −1). Burglary declines were generally smaller than robbery/theft. During the lull between waves (late 2020–early 2021), actual and predicted theft/robbery converged; burglary remained moderately above predictions. - Mobility–crime linkage: Reductions in retail & recreation, workplaces, and transit stations were synchronous with declines in robbery and theft (opportunity reduction). Burglary trends inversely tracked residential mobility increases (enhanced guardianship). - Urban–rural differences: Chennai showed larger mobility changes and generally steeper crime declines during restrictions than TN, consistent with stricter enforcement and urban context, but overall patterns were consistent across both.

Findings support Routine Activities and crime opportunity theories: lockdown-induced mobility reductions diminished offender–target convergence and increased guardianship, especially in residential settings, leading to pronounced declines in robbery and theft and moderate declines in burglary. The inverse association between residential mobility and burglary underscores the role of capable guardians. When restrictions were lifted (Post-L), robberies increased above counterfactuals even after adjusting for mobility, suggesting additional criminogenic pressures (e.g., economic strain, unemployment) beyond mere movement volume. The study demonstrates that mobility functions as a spatio-temporal proxy for opportunity structures: lower mobility in public spaces corresponds with reduced opportunities for acquisitive crimes; higher residential presence reduces burglary risk. Urban–rural comparisons show similar directional effects, with Chennai’s larger mobility shifts aligning with sharper crime declines under restrictions, likely reflecting enforcement capacity and urban infrastructure. The results have practical implications: as restrictions ease, agencies should anticipate and proactively counter surges in robberies, reassessing protocols and resource allocation that had shifted during the pandemic.

- The two-year analysis with a decade of training data establishes a clear, direct relationship between mobility and property offences: strict SAH orders produced very large effect-size reductions; milder restrictions yielded medium-to-large reductions. - Post-lockdown 2021 saw robberies rise above counterfactuals in TN (≈+56%) and Chennai (≈+23%) even after adjusting for mobility, indicating pandemic-related economic and social consequences; theft and burglary showed mixed post-lockdown trends. - Urban–rural differences were limited in directionality; Chennai’s stronger enforcement and larger mobility changes corresponded with relatively greater reductions during restrictions. - Methodologically, ARNN (DeepAR) delivered accurate, robust forecasts for crime time series with multi-seasonality and volatility, requiring minimal manual specification. - Policy implication: Police should strengthen preventive work and rapidly reset protocols when restrictions are lifted to mitigate robbery surges. - Future research: Incorporate qualitative studies to better understand behavioral and compliance drivers of mobility differences across waves and to quantify under-reporting dynamics.

- Crime data limitations: Under-reporting and non-reporting of property offences are well-documented; during the pandemic, fear of infection, constrained access to police/health services, and altered law-enforcement protocols may have further reduced registrations beyond mobility effects; these factors likely amplify observed declines but were not quantified. - Mobility data limitations: Google mobility reflects relative changes, not absolute counts; relies on opt-in location histories and smartphone users; short 5-week baseline (cannot capture seasonality), potential population shifts (relocation, remote work), and evolving place definitions. - Model covariates: DeepAR forecasts were univariate (no covariates); mobility was analyzed associationally rather than integrated into the forecasting model. - Wave differences: Higher mobility during wave two despite greater severity may reflect compliance fatigue, reduced fear, vaccinations, and improved treatments; reasons were not directly measured; absence of qualitative components limits causal attribution for mobility compliance changes.