Counterfactual analysis of the impact of the first two waves of the COVID-19 pandemic on the reporting and registration of missing people in India

The study investigates how pandemic-driven mobility restrictions affected the reporting and registration of missing person cases in Tamil Nadu, India. Missing persons cases are distinct from other crimes in both incidence and reporting processes, often involving prolonged investigations and significant psychological and socio-economic impacts on families. Despite technological advances and wider awareness, India recorded rising missing person reports pre-pandemic. The COVID-19 lockdowns created a natural experiment to examine latent factors that enable or impede reporting—particularly mobility and access to police. The research question centers on quantifying the impact of mobility changes across lockdown phases on registered missing person cases using counterfactual time-series forecasts, thereby revealing how situational factors influence access to justice for families of missing persons.

Pre-pandemic, reasons for missing persons in India span unintentional absence (disorientation, separation, kidnapping) and intentional absence (elopement, academic failure, fleeing abuse, bankruptcy, suicide), with socio-cultural pressures (e.g., inter-caste/religion unions) and gender-based violence implicated. Internationally, conflict zones (e.g., Syria, Iraq), authoritarian contexts (e.g., China’s dissidents, Uighurs), and developed regions (USA, Europe) also report large numbers of missing persons. The pandemic literature highlights indirect impacts on crime reporting, mobility, and the strain on law enforcement, with increased domestic violence and mental health stressors affecting missing persons. Methodologically, crime forecasting traditionally relied on ARIMA, Holt-Winters, and exponential smoothing, but these have limits under volatile, rapidly changing conditions. Recent advances in deep learning (e.g., RNN-based models like DeepAR) offer improved probabilistic forecasting for count data and can handle seasonality, holidays, and limited data. Prior studies for India’s context are sparse, especially spanning both COVID waves; findings from high-income countries may not transfer due to differing socio-political and cultural contexts.

Design: A natural experiment leveraging India’s COVID-19 lockdown phases to study the impact of mobility on the registration of missing person cases in Tamil Nadu (population ~80 million). Daily FIR-based counts of missing persons (and unidentified dead bodies, UIDB) were compiled for 2010–2021. Lockdown phases defined eight windows: Wave 1 complete lockdown (CL, Mar 23–Apr 30, 2020), Wave 1 partial lockdown (PL, May 1–Jun 8, 2020), post-lockdown 2020 (from Sep 1, 2020), Wave 2 PL-One (Apr 10–May 5, 2021), Wave 2 CL (May 6–Jun 7, 2021), Wave 2 PL-Two (Jun 8–Jul 6, 2021), and post-lockdown 2021 (Sep 1–Sep 30, 2021), etc. Reporting is via FIR per Indian CrPC Section 154; missing persons receive FIRs despite not being crimes per se. Forecasting model: DeepAR (an autoregressive recurrent neural network) implemented via GluonTS in Python. The model provides probabilistic forecasts suitable for positive count data using a negative binomial likelihood. Advantages include: minimal structural assumptions, probabilistic outputs with confidence intervals, handling of nonstationarity/seasonality/holiday effects, and feature creation for trends/seasonality. Model development followed standard train/validation/test splits; performance was evaluated with Weighted Mean Absolute Percentage Error (WMAPE) and compared against ARIMA, GAM, BSTS, and Holt-Winters, with DeepAR showing the lowest error. Counterfactual inference: DeepAR was trained on the pre-pandemic period and used to predict the counterfactual (no-lockdown) daily counts for 2020–2021. Forecast uncertainty was summarized with 95% intervals (0.025 and 0.975 quantiles from n sampled traces). The causal impact was estimated as the difference between actual counts and the counterfactual prediction during each lockdown window. Distributional assumptions were checked using the Shapiro–Wilk test; t-tests were applied when normality held, otherwise Wilcoxon signed-rank tests. Effect sizes were quantified using Cliff’s delta, robust under non-normality. Mobility measurement: Mobility changes were obtained from Google Community Mobility Reports (CMR) across six domains: retail and recreation, grocery and pharmacy, parks, transit stations, workplaces, and residential. Baseline: Jan 3–Feb 6, 2020, day-of-week adjusted. Mobility was analyzed for Tamil Nadu overall and Chennai specifically, across all eight windows. Normality was assessed (Shapiro–Wilk), differences tested with Wilcoxon, and effect sizes computed (Cliff’s delta). Analyses examined associations between mobility shifts and deviations of actual from counterfactual missing person registrations.

- Descriptive trends: The daily mean of registered missing person cases increased from 32.44 (SD 14.42) pre-pandemic (Jan 1, 2010–Mar 22, 2020) to 54.89 (SD 20.89) during the pandemic period (Mar 23, 2020–Dec 31, 2021). Trends show long-run increases, necessitating counterfactual adjustment. - Mobility: Severe declines from baseline occurred in non-residential domains during complete lockdowns, with partial recovery in PL phases; residential mobility increased markedly. Reductions were larger in Wave 1 than Wave 2 and more pronounced in Chennai than statewide. Cliff’s delta comparisons across phases indicated large effect sizes in many domains (e.g., TN retail/recreation PL-2020 vs CL-2020: 0.548 large; Chennai workplaces CL-2020 vs CL-2021: -0.43 large). - Missing persons (Tamil Nadu): • CL-2020: Actual vs counterfactual showed a sharp decline of -73.87% (Cliff’s delta -0.981). • CL-2021: Decline of -35.29% (Cliff’s delta -0.742). • PL-2020: Decline of -39.66% (Cliff’s delta -0.839). • PL-One 2021: Increase of +9.14% (Cliff’s delta 0.33). • PL-Two 2021: Decline of -14.63% (Cliff’s delta -0.264). • Post-L-2020: Increase of +35.40% (Cliff’s delta 0.931). • Post-L-2021: Increase of +39.98% (Cliff’s delta 0.784). - UIDB (Tamil Nadu): • CL-2020: -26.77% (Cliff’s delta -0.68); CL-2021: -6.12% (Cliff’s delta -0.218). • PL-2020: +12.59% (0.204); PL-One 2021: +42.32% (0.65); PL-Two 2021: -18.60% (-0.513). • Post-L-2020: +26.41% (0.666); Post-L-2021: +17.93% (0.393). - Chennai patterns mirrored statewide results with magnified magnitudes; e.g., CL-2020 missing persons -45.39% (Cliff’s delta -0.94), Post-L-2020 +247% (Cliff’s delta 1.0), CL-2021 -41.4% (Cliff’s delta -1.0), Post-L-2021 +136% (Cliff’s delta 1.0). - Interpretation: Restrictive phases corresponded to significant shortfalls relative to counterfactuals, indicating impeded reporting/registration mechanisms; relaxation phases saw rebounds or overshoots, consistent with restored access to reporting channels and delayed complaint lodgment being cleared.

Findings demonstrate a strong linkage between mobility and the registration of missing person cases. During complete and partial lockdowns, movement restrictions, reduced public transport, and fear of infection created practical barriers to reaching police stations or relief centers, yielding large negative deviations from counterfactual predictions. When mobility curbs eased, registrations rose markedly above counterfactuals, consistent with backlogged or delayed reporting being enabled once access improved. These effects were more pronounced in Chennai than in Tamil Nadu overall, reflecting greater urban mobility suppression. Similar, though smaller, patterns appeared for unidentified dead bodies, aligning with the conceptual relationship between missing persons and UIDB case trends. The results underscore that variations in mobility translate into changes in access to justice and that registered case counts during crises reflect both true incidence and the functionality of reporting channels. Ensuring resilient, accessible reporting mechanisms—independent of mobility constraints—is crucial, particularly for time-sensitive missing person investigations.

The pandemic offered a natural experimental context to uncover the role of mobility in the reporting and registration of missing person cases. Counterfactual analysis via DeepAR revealed sharp declines relative to predicted levels during complete lockdowns (−74% in 2020; −36% in 2021) and substantial increases during post-lockdown periods (+35% in 2020; +40% in 2021). Results highlight that mobility is a pivotal situational determinant of registration and that inhibited access during crises can delay reporting, potentially reducing the likelihood of successful tracing. Policy implications include strengthening accessible, resilient reporting channels (including technological solutions and multi-agency coordination), enhancing transport/access provisions during emergencies, and focusing resources for rapid response within the critical early period after disappearance. Future research could extend analyses across other Indian states, integrate richer covariates (e.g., socio-economic indicators), and evaluate specific interventions designed to maintain reporting access under mobility constraints.

- The analysis is based on registered daily FIR counts from Tamil Nadu (and Chennai), which may not capture unreported cases. - Data supporting the findings are held by the State Crime Records Bureau of the Tamil Nadu Police Department and are not publicly shareable; access requires permission, limiting external reproducibility. - Mobility is measured via Google Community Mobility Reports, which provide aggregated, device-based proxies rather than direct measures of individual access. - Findings pertain to Tamil Nadu/Chennai during 2010–2021 and may not generalize without caution to other regions or contexts.