The COVID-19 pandemic necessitated unprecedented stay-at-home orders (lockdowns) globally to curb the spread of the virus. These lockdowns drastically altered lifestyles, causing a sharp decline in human mobility. Tamil Nadu, India, experienced two waves of the pandemic in 2020 and 2021, providing a unique opportunity to study the impact of lockdowns on crime. While existing literature suggests a global decline in property crime during lockdowns, there's a scarcity of research using primary data from India. This study aims to address this gap by investigating the impact of varying degrees of lockdown restrictions on property offenses (robbery, burglary, theft) in Tamil Nadu and its capital city, Chennai, considering the complexities of crime reporting within the Indian context. The research questions revolve around the impact of mobility changes, as measured by Google Mobility Reports, on property crime rates during different lockdown phases, and whether there are significant urban-rural differences in this impact. The study's importance lies in its potential to inform law enforcement strategies and resource allocation during similar public health emergencies or unexpected events that significantly alter mobility patterns. The study utilizes a robust Auto-Regressive Recurrent Neural Network (ARNN) model for time-series forecasting, chosen for its superior performance compared to traditional methods in handling fluctuating data with multiple seasonality factors. The study's time frame covers two waves of the pandemic, enabling analysis across various phases of lockdown with differing restriction levels.

Existing literature on the impact of COVID-19 lockdowns on crime reveals a global trend of decreased property crime. Studies across different continents and income levels show reductions in burglary, robbery, and theft during strict lockdown periods (Nivette et al., 2021). However, the effects varied geographically and across crime types, with some studies noting increases in violent crime or specific offenses (Mohler et al., 2020; Ashby, 2020; Scott and Gross, 2021; Meyer et al., 2022; Hou et al., 2022). Research from England and Wales (Agrawal et al., 2022; Halford et al., 2020; Farrell and Dixon 2021), Ireland (Buil-Gil et al., 2021), Mexico City (Estévez-Soto, 2021), Rio De Janeiro (Bullock et al., 2021), and Buenos Aires (Perez-Vincent et al., 2021) show mixed results, with variations dependent on the severity and duration of lockdowns and socio-economic factors. Studies in India reported sharp drops in visible crimes (Som et al., 2020), while others in Nigeria (Akanmu et al., 2021) found increases due to factors such as poverty and poor governance. Much of the literature employs routine activities theory (RAT) and crime opportunity theory to explain these variations. RAT suggests crime occurs when an offender, a suitable target, and lack of guardianship converge (Cohen and Felson, 1979), while crime opportunity theory highlights rational choices based on reward vs. risk (Agnew, 1992). Studies utilizing machine learning and these theories successfully account for crime pattern shifts (Li et al., 2021; Chen et al., 2021; Koppel et al., 2022; Campedelli et al., 2020). This existing research laid the groundwork for a more in-depth analysis of the Indian context, particularly focusing on the differential impacts of lockdowns and subsequent mobility changes on various property crimes.

This study leverages daily crime data from 1356 police stations across Tamil Nadu and 115 in Chennai, covering the period from January 1, 2010, to December 31, 2021. The data includes counts of robberies, burglaries, and thefts. The Auto-Regressive Recurrent Neural Network (ARNN) model, specifically the GluonTS DeepAR model (Alexandrov et al., 2020; Salinas et al., 2020), was employed to forecast daily crime counts in the absence of the pandemic. This model, which outperforms traditional methods like ARIMA, Holt-Winters, Bayesian Structural Time Series Model, and Generalized Additive Model (Table 4), uses an RNN architecture (LSTM or GRU) and a negative binomial likelihood function to generate probabilistic forecasts. The model was trained on data from 2010-2019, validated on data from January 1, 2020, to March 22, 2020, and used to predict daily crime counts from March 23, 2020, to December 31, 2021. The difference between the actual and predicted crime counts represents the causal impact of the lockdown. Google Community Mobility Reports provided mobility data, categorized into six domains: retail & recreation, parks, groceries & pharmacies, workplaces, transit stations, and residential areas. The percentage change in mobility from a pre-pandemic baseline was used as a measure of mobility changes during different lockdown phases. The study employs interrupted time series analysis, using Cohen’s d or Cliff’s Delta (Hess and Kromrey, 2004) to measure effect sizes (ES) of lockdowns on crime rates, accounting for potential non-normality through the Wilcoxon signed-rank test. The analysis assesses the impact of complete lockdowns (CL) and partial lockdowns (PL) across both pandemic waves, comparing actual and predicted crime rates during these periods and in the post-lockdown (Post-L) phases. The study includes both state-level (Tamil Nadu) and city-level (Chennai) analyses to identify urban-rural differences.

The analysis revealed a strong correlation between mobility changes and property crime rates. During complete lockdowns (CL) in both waves, there was a significant decline in all three property crime types (robberies, burglaries, and thefts) in both Tamil Nadu and Chennai. The decline was more pronounced in Chennai, a predominantly urban area, than in Tamil Nadu, which has a larger rural population. Table 5 and Table 6 show the percentage change from the predicted values during the various lockdown phases. Effect sizes (ES) were largely negative and significant during CL and PL phases (Table 3, Table 5, and Table 6). The magnitude of decrease was highest for thefts, followed by burglaries, and then robberies. The sharpest decline in mobility occurred in retail and recreation, workplaces, and transit stations during CL-2020, which corresponds with the largest decrease in robberies and thefts. However, when restrictions were lifted during the Post-L phase, a significant increase in robberies was observed in both Tamil Nadu (56% increase; ES 0.476) and Chennai (23.4% increase; ES 0.25), even after adjusting for mobility (Figure 4, Table 5, Table 6). The increase in robberies suggests a potential relationship between the pandemic's economic fallout (increased unemployment) and the rise in property crimes. Burglary trends showed a more complex pattern, with a lower magnitude of decrease than robberies and thefts during lockdowns and a mixed trend during the post-lockdown period. This suggests that while overall mobility impacted robberies and thefts, the impact of lockdowns on burglaries was less strongly linked to changes in overall mobility, perhaps due to changes in guardianship. There was a substantial fall in the magnitude of decrease in all three property offences during the various phases of lockdown in the second wave when compared to the first wave. This can be attributed to changes in public adherence to lockdown orders, reduction in the fear of COVID-19 due to increased knowledge and vaccinations, and other factors. The ARNN model demonstrated superior accuracy compared to other forecasting methods, exhibiting low WMAPE values (Table 4). The figures (Figure 3, Figure 4, Figure 5, and Figure 6) further illustrate the relationship between mobility and crime trends.

The findings strongly support the hypothesis that mobility plays a significant role in influencing property crime rates. The observed sharp declines in property crime during strict lockdowns, particularly in retail areas, workplaces, and transit stations, align with routine activity theory and crime opportunity theory. Reduced mobility minimized the convergence of offenders, suitable targets, and absence of guardianship, leading to a decrease in crime opportunities. The post-lockdown surge in robberies, despite reduced overall mobility compared to the pre-pandemic period, highlights the complex interplay of factors. The rise in robberies is likely linked to pandemic-induced economic hardship, resulting in increased desperation and criminal activity. This suggests that mobility is a critical component, but not the sole determinant, of crime rates. The relatively higher decrease in property crimes in Chennai, compared to Tamil Nadu, could be attributed to several factors including higher literacy rates, better healthcare infrastructure, potentially leading to reduced mobility, and stricter enforcement of lockdown measures. The ARNN model proved effective in analyzing the impact of lockdowns on crime, demonstrating the advantages of deep learning approaches for time-series crime forecasting. The study also acknowledged the limitation of crime reporting in the Indian context, where under-reporting could potentially influence the results. Future research could explore qualitative data to further understand the reasons behind variations in mobility during different phases of the pandemic and to gather more detailed information on the victims' experience.

This study provides strong empirical evidence for the link between mobility and property crime rates during the COVID-19 pandemic. The significant decline in crime during lockdowns and the subsequent rise in robberies after restrictions were eased highlight the importance of mobility in crime opportunity. The ARNN model showcased its usefulness as a forecasting tool. Future research could integrate qualitative data, further explore the economic impacts of the pandemic on crime, and investigate the effectiveness of different policing strategies in managing crime during similar public health crises. The findings emphasize the need for proactive crime prevention measures, particularly for robberies, during periods of changing mobility patterns.

The study acknowledges potential underreporting of crime, particularly during the pandemic. Situational factors such as fear of infection, limited access to police services, and the police’s own changed priorities during the pandemic could have affected crime registration. While the under-reporting would likely amplify the findings, it was not quantified. The use of Google Mobility Reports, while widely used, has limitations. The data is not a measure of absolute numbers but relative changes from the baseline and may not capture all mobility patterns. Furthermore, the study does not delve into the qualitative aspects behind the variations in mobility between the two waves of the pandemic, which would enhance the study's comprehensiveness.