Reduction in mobility and COVID-19 transmission

The study addresses how reductions in human mobility, implemented via social distancing, isolation, and shielding policies, relate to SARS-CoV-2 transmission as measured by the effective reproduction number (R). Early in the pandemic, many countries instituted strict mobility-restricting measures to suppress transmission and prevent overwhelming healthcare capacity. Mobility datasets (Apple and Google) emerged as timely proxies for changes in contact rates. With countries experiencing multiple waves and varying degrees of relaxation and re-imposition of measures, the research question is whether mobility data reliably reflect changes in transmission over time, whether this relationship changes with policy shifts, and how it can inform control thresholds required to keep R below 1. The purpose is to quantify and compare the mobility–transmission relationship across 52 countries and to evaluate how much of the variation in transmissibility can be explained by mobility, including before and after relaxation of strict measures.

Prior work showed that early mobility reductions correlated with decreases in COVID-19 incidence and supported the use of non-pharmaceutical interventions (NPIs) such as social distancing, isolation, and shielding. Observational and modeling studies (e.g., analyses in China, Hong Kong, Europe, and the US) documented that NPIs reduced transmission and mobility proxies tracked these changes. The paper references established methods for estimating time-varying reproduction numbers from case or death data and notes robust associations between mobility patterns and transmission in multiple settings. This study builds on that literature by systematically quantifying the mobility–transmission relationship across a wide set of countries using Apple and Google mobility streams and by examining changes in this relationship over time as policies were relaxed.

Study design: Comparative analysis across 52 countries with sustained SARS-CoV-2 transmission, linking mobility data to transmission inferred from reported deaths. Data: Daily COVID-19 death counts from WHO and ECDC. Mobility data from Apple (driving, transit, walking) and Google (grocery and pharmacy, parks, residential, retail and recreation, transit stations, workplaces). Data processing: Mobility streams smoothed using a 7-day rolling average and rescaled relative to pre-pandemic country-specific maximums. Composite mobility indices were created: combined Apple, combined Google (excluding parks and residential), and a combined Apple-Google measure used for main results; analyses also conducted for each individual stream. Inclusion: 52 countries meeting active transmission thresholds; 36 had both Apple and Google data; 16 had Google only. Transmission modeling: - Parametric model linking time-varying effective reproduction number R_t to mobility m_t. Deaths are related to a delayed reproduction number reflecting infection-to-death delays. Negative binomial likelihood used to account for overdispersion in reported deaths. - Non-parametric estimate of the delayed reproduction number derived from deaths alone using a rolling (weekly) window approach (EpiEstim-like framework), providing data-driven Rt trajectories without imposing a specific functional link to mobility. Delay distributions: Infection-to-death interval modeled as gamma distributed (mean 18.3 days, SD 8.64). Serial interval for transmission modeled as gamma with mean 6.48 days and variance 13.3 days. Inference and comparison: Bayesian inference via MCMC estimated country-specific parameters, including potential change points in the mobility–transmission relationship (a post-relaxation dampening). Model fit and predictive ability were assessed by comparing parametric Rt (mobility-linked) with non-parametric Rt from deaths and by adjusted R-squared metrics. Model comparisons across mobility streams used Deviance Information Criterion (DIC). Mobility thresholds: For countries where transmission decreased significantly with mobility reductions, mobility thresholds were estimated: the reduction in mobility required to bring R below 1, estimated both pre- and post-change in the mobility–transmission relationship. Robustness checks: Sensitivity to serial interval assumptions; robustness to excluding very early epidemic periods (before 100 cumulative deaths). Software and code: Analyses conducted in R (v3.5.1); code available on GitHub (https://github.com/powered/Mobility20).

• Mobility trends: Across 52 countries, mobility dropped sharply early in the pandemic (median 63% reduction from baseline at March 11, 2020; IQR 51–67%), then partially recovered to a 14% reduction by October 25, 2020 (IQR for late-October/November reductions ~17–22%). - Variation by country: Smallest mobility reductions (37–51%) were observed in Moldova, Afghanistan, Switzerland, Ecuador, Paraguay, Sweden, Ukraine, Panama, Dominican Republic, Denmark. Largest reductions (72–83%) in Honduras, Poland, Costa Rica, Italy, Guatemala, Peru, Philippines, Argentina, France, Bolivia. - Mobility–transmission link: Transmission significantly decreased with initial mobility reductions in 73% of countries. Evidence of decoupling (a changed/dampened relationship following relaxation of strict measures) occurred in 80% of countries. - Explanatory power: Mobility explained a substantial portion of transmission variability across countries; median adjusted R-squared 48% (IQR 27–77%). Where a change in relationship was inferred, predictive ability dropped from a median adjusted R-squared of 74% (IQR 49–91%) pre-relaxation to 30% (IQR 12–48%) post-relaxation. - Mobility thresholds (R<1): Median threshold reduction across countries decreased from 51% pre-change to 16% post-change, indicating that after dampening, smaller mobility reductions were sufficient for control, likely due to additional protective behaviors and measures. - UK example (Apple-Google stream): Pre-change mobility reduction needed to achieve R<1 was 43% (95% CrI 41–46%). Post-change threshold was 18% (95% CrI 14–21%). On Oct 25, 2020, UK mobility reduction was ~15%, implying likely R>1 with uncertainty (within post-change CrI). Estimated Rt for new infections on Oct 25 was 1.07 (95% CrI 0.99–1.17); delayed Rt from deaths was 1.21 (95% CrI 1.10–1.33). - Country heterogeneity in thresholds: In Germany, Hungary, and Turkey, contemporaneous mobility reductions were below the lower 95% CrI of thresholds, suggesting ongoing epidemics. In 12 countries (Belgium, Chile, Denmark, France, Ireland, Israel, Italy, Netherlands, Spain, Switzerland, UK, USA), latest mobility overlapped threshold CrIs, indicating uncertainty in control status. - Model and data-stream performance: Combined Apple-Google stream generally outperformed individual streams. In 8 countries, individual streams (notably Apple Transit in Brazil, France, Mexico, Philippines, UK, USA) provided substantially better fit (DIC differences >10). - Robustness: Mobility thresholds were robust to serial interval assumptions; thresholds were not correlated with basic reproduction number estimates; results robust to excluding very early epidemic periods.

The findings demonstrate a strong and temporally coherent association between reductions in human mobility and decreases in SARS-CoV-2 transmissibility in most of the 52 countries analyzed, affirming mobility as a useful, near-real-time proxy for changes in contact rates and transmission. The observed dampening of the mobility–transmission relationship after relaxation of strict measures suggests that other sustained preventive behaviors and interventions (mask use, distancing, hygiene, targeted restrictions, improved testing and tracing) increasingly contributed to lowering transmission, allowing control at smaller measured mobility reductions. Nonetheless, as mobility increased, many countries experienced resurgent transmission, emphasizing that relaxation must be accompanied by equally effective alternative control strategies. The analysis highlights policy-relevant mobility thresholds for achieving R<1 and underscores heterogeneity across countries. Real-time mobility data, updated with short delays (2–4 days for Apple; 7–10 for Google), can support ongoing surveillance, situational awareness, and timely policy adjustments. However, post-relaxation decoupling reduces predictive ability, indicating that reliance on mobility alone is insufficient and must be complemented by direct epidemiological indicators and contextual information on NPIs and behaviors.

Across 52 countries, automated mobility measures were strongly linked to SARS-CoV-2 transmission, especially during initial stringent control periods. After relaxation, the relationship often dampened, suggesting that sustained protective behaviors and interventions allowed epidemic control at lower levels of measured mobility. The study quantifies how much transmission variability mobility can explain, identifies mobility thresholds for achieving R<1, and compares the utility of different mobility streams, with combined Apple-Google generally performing best. Policymakers can use mobility indicators as part of a broader toolkit to gauge and adjust control measures. Future work should integrate mobility with detailed, time-varying data on specific NPIs, behavioral adherence, contact patterns by setting, vaccination coverage, and variant-specific dynamics, and refine models to account for decoupling and heterogeneity across populations.

• Mobility as a proxy: Mobility metrics approximate contact rates but may not fully capture risk-relevant contacts (e.g., masking, distancing, ventilation), contributing to post-relaxation decoupling and reduced predictive power. - Data quality and delays: Deaths-based inference involves substantial lags (infection-to-death mean ~18.3 days) and potential under-reporting or reporting delays; although estimates are relatively robust to under-reporting, variability in reporting and response strategies can affect inference. - Heterogeneity across countries: Differences in demographics, health systems, NPIs, adherence, and testing can confound the mobility–transmission relationship; thresholds could not be identified for some countries and some upper 95% CI limits were unidentifiable. - Model assumptions: Assumptions about serial interval and infection-to-death distributions, and the parametric functional form linking mobility to Rt, may not hold uniformly; despite sensitivity analyses, residual model misspecification is possible. - Data gaps: Mobility measures had missing data in places and varying update delays across providers; not all countries had both Apple and Google data. - Post-relaxation predictive decline: The adjusted R-squared substantially decreased after inferred changes in relationship, limiting the utility of mobility-only models for forecasting in later epidemic phases.