Estimating the effect of social inequalities on the mitigation of COVID-19 across communities in Santiago de Chile

Chile experienced one of the most severe COVID-19 epidemics globally by September 1, 2020, with more than 400,000 cases and 590 deaths per million, and Santiago Metropolitan Region accounting for roughly 70% of national cases. Initial NPIs (school closures, bans on public gatherings, self-isolation for high-risk travelers) were introduced in mid-March, but rapid case growth led to a full lockdown across the Metropolitan Region on May 15, 2020. This study investigates the spatio-temporal spread of SARS-CoV-2 across 37 comunas in the Santiago urban area (6.4 million people), aiming to quantify how NPIs altered mobility and contacts and how these changes affected epidemic dynamics. Using anonymized mobile phone data to measure mobility and infer contact reduction, integrated into a spatially and age-structured mechanistic epidemic model calibrated to surveillance data, the study also examines how socioeconomic disparities, proxied by the Human Development Index (HDI), shaped the capacity to reduce mobility and the resulting heterogeneous disease burden across communities.

Prior work has modeled COVID-19 spread and NPIs at urban scales, including studies for Boston, Wanzhou, New York City, and London. NPIs and mitigation strategies have varied widely across and within countries, necessitating context-specific modeling. Mobile device data have been used to evaluate intervention impacts and inform large-scale epidemic models. The basic reproduction number R0 for SARS-CoV-2 has been estimated between 2 and 3 in multiple countries. Studies have highlighted socioeconomic gradients in mobility reduction and capacity for social distancing in France, Italy, the United States, Colombia, Mexico, and Indonesia, and linked socioeconomic inequalities to worse pandemic outcomes in the United States, Singapore, and the UK. Seroprevalence research in the US, Spain, Italy, Brazil, and Iran indicates that infections typically exceed reported cases by factors of 4 to 20, suggesting substantial under-ascertainment in many settings.

Data: Anonymized mobile phone eXtended Detail Records (XDR) from Telefónica Movistar (market share ~24.6%) covering 1.4 million devices (≈22% of the study population) from February 27 to June 1, 2020. Device location is approximated by serving antenna coordinates; home antenna is defined by nighttime activity. Antennas are assigned to comunas. Mobility measurement: For each day t, a mobility rate matrix M(t) is built, where Mij(t) is the fraction of devices residing in comuna i that visited comuna j. Daily matrices are averaged over three phases: baseline (pre–March 16), partial lockdown (March 16–May 15), and full lockdown (post–May 15). Mobility includes all purposes (work, errands, recreation); no POI classification is used. Contacts reduction proxy: Direct contacts are not inferred. Instead, for each antenna a in comuna j, with resident users Na and day-t visitors va(t) from the same comuna, the maximum possible pairwise contact count ca(t) is approximated (homogeneous mixing). Contacts reduction during partial and full lockdowns is computed as the ratio of median ca(t) across antennas before versus after each intervention, aggregated at comuna level. Socioeconomic indicator: HDI at comuna level is computed from Chilean census data (2013–2015) following UNDP guidelines, capturing life expectancy, education, and income components. Epidemic model: A stochastic metapopulation SLIR model inspired by GLEAM represents 37 comunas as subpopulations, age-structured into K=16 five-year brackets using comuna demographics and country-specific contact matrices. Transmission occurs within and between comunas via an effective coupling derived from the empirical mobility network using a time-scale separation approximation, justified by trip durations (mean 4.5 h; 85% ≤8 h; <3% >1 day). Disease natural history: Latent period 4 days; infectious period 2.5 days; generation time ~6.5 days. Deaths are simulated by applying an infection fatality rate and a reporting delay (Δ) after removal, accounting for hospitalization-to-report lags (>2 weeks). NPIs in model: On March 16 and May 15, mobility matrices are switched to reflect observed inter-comuna mobility changes; comuna-level contact matrices are scaled by estimated contact-reduction parameters for partial and full lockdowns. School closures are implicitly captured via observed changes. Stochastic simulation: Chain binomial transitions between compartments. Initial seeding: projected active cases in the Metropolitan area on March 1, 2020, allocated to comunas proportional to population. Calibration: Approximate Bayesian Computation (ABC) Rejection using weekly deaths (confirmed + suspected) from the Chilean Ministry of Health. Priors: R0 ∈ [2,4] (step 0.02), death delay Δ ∈ [14,21] days (step 1). Distance metric: median absolute percentage error with 20% tolerance. 140,000 iterations (~200 realizations per parameter set). Posterior sampling yields ensembles (n=5000) for medians and 95% CIs. Sensitivity analyses include alternative compartmental structures (with pre-symptomatic and asymptomatic transmission), varying generation times, and alternative IFR assumptions.

• Mobility and contacts: The first NPIs on March 16 reduced inter-comuna travel by ~48%; the May 15 full lockdown produced an additional ~17% reduction, for an overall ~65% drop vs baseline. Contacts decreased by 36% after March 16 and by an additional 11% under full lockdown. • Detection and prevalence: As of August 1, 2020, estimated detection rate was 102 cases per 1000 infections (90% CI: 95–112 per 1000). Median projected prevalence was 38.7% (95% CI: 35.1–41.6%). Using a higher IFR yielded 28.3% (95% CI: 23.5–32.3%) with a worse fit. • Model fit and transmissibility: Calibrated R0 median 2.66 (95% CI: 2.58–2.72). Simulated vs reported weekly deaths showed high agreement (Pearson r=0.99, p<0.001), MAPE 12%. Simulated vs reported cumulative cases across comunas correlated strongly (r=0.84, p<0.001). Days to reach 200 infections correlated between model and surveillance (Kendall τ=0.61, p<0.001). Effective Rt dropped below 1 only after the full lockdown; simulated and reported Rt time series correlated (r=0.78, p<0.001). • Inequalities and mobility: Average mobility reduction after March 16 correlated positively with comuna HDI (Pearson r≈0.80, p<0.001): higher-HDI comunas reduced mobility more. • Inequalities and burden: Attack rates vs HDI showed strong negative correlations: reported r=−0.74 (p<0.001); simulated r≈−0.69 (p<0.001). Deaths per 1000 vs HDI were negatively correlated: simulated r≈−0.50 (p<0.002); reported r≈−0.40 (p<0.02). Wealthier comunas experienced lower incidence and mortality. • Impact of NPIs (counterfactuals): Without the full lockdown, the incidence peak would have been 21.6% higher (95% CI: 7.5–41.3%) and deaths 34.7% higher (95% CI: 27.2–44.1%). A 1-week delay in the full lockdown would increase peak intensity by 18.1% (95% CI: 6.0–34.0%) and deaths by 7.7% (95% CI: 1.3–13.7%); a 2-week delay yields a 21.6% (95% CI: 7.4–41.1%) higher peak. Uniformly applying the mobility/contact reductions observed in the highest-HDI quartile to all comunas would have reduced cases by 83.8% (95% CI: 77.6–88.6%) and deaths by 70.5% (95% CI: 55.0–80.9%) by May 15.

The study demonstrates that while early NPIs in Santiago reduced mobility and contacts, they were insufficient to control transmission until a full lockdown was implemented, which pushed Rt below 1 and served as a tipping point for epidemic control. The effects of NPIs on mobility were strongly stratified by socioeconomic development: higher-HDI comunas decreased mobility more and suffered lower attack rates and mortality, indicating that systemic social and economic inequalities modulated both behavioral responses and health outcomes. These findings underscore the need for policies that consider differential capacities to comply with mobility restrictions and the potential of stringent, timely NPIs to shorten outbreaks and mitigate unequal burdens. Comparisons with simpler modeling approaches highlighted the importance of incorporating spatial coupling and data-driven mobility/contact changes to capture heterogeneous dynamics across comunas. The observed late divergence between simulated and reported deaths suggests possible lockdown fatigue, warranting further investigation as more data become available.

By integrating large-scale mobile phone mobility data with a spatially and age-structured stochastic epidemic model, the study quantifies how NPIs and social inequalities shaped COVID-19 transmission across Santiago’s comunas. The full lockdown, despite only modest additional mobility and contact reductions relative to earlier measures, was decisive in bringing Rt below 1 and curbing the outbreak. Higher-HDI comunas achieved larger mobility reductions and experienced lower disease burden, evidencing inequality-driven heterogeneity in both behavior and outcomes. Counterfactual analyses emphasize the substantial mortality reductions achievable through uniform, stronger, and earlier interventions. Future work includes testing hypotheses about compliance changes and lockdown fatigue as additional data become available.

• Geographic scope: Analysis is restricted to the Santiago Metropolitan Area; national and international importations after March 1, 2020, are not modeled explicitly. • Disease model simplicity: A SLIR structure is used; although sensitivity analyses suggest robustness, it is simpler than models including pre-symptomatic and asymptomatic compartments. • Mobility/contact proxies: Mobility lacks POI categorization, and not all mobility entails equal transmission risk. Contacts are inferred via co-location at antennas under homogeneous mixing, not measured directly. • Sample representativeness and infrastructure: Mobile phone users may not perfectly represent the population; data depend on antenna distribution. Aggregation at comuna level mitigates some biases. • Data coverage: Mobile data end June 1, 2020; deviations in late July/early August may reflect unobserved compliance changes. • Seroprevalence: Lack of contemporaneous serosurveys in Santiago limits direct validation of infection prevalence estimates.