Human mobility and infection from Covid-19 in the Osaka metropolitan area

The study addresses the question: Where should governments control human mobility to reduce COVID-19 cases during the medium-term of the pandemic? While early pandemic periods showed short-term sustainability benefits and broad mobility reductions, medium-term effects and policies (e.g., repeated emergency declarations in Japan) led to complex behavioral changes, including shifts from public transport to private cars with adverse implications for CO2 emissions. Human mobility is a policy-leveraged factor and has been shown to affect transmission more strongly than many other determinants (e.g., population density, temperature, vaccination). Given that vaccination coverage varies and novel pathogens may lack vaccines initially, nonpharmaceutical interventions that target mobility remain fundamental. The Osaka metropolitan area (Osaka, Kyoto, Hyogo) provides a relevant case due to interconnected urban centers and coordinated prefectural policies, including multiple states of emergency and non-compulsory restrictions (“soft lockdown”). The study focuses on identifying which mobility types most strongly relate to subsequent infection levels to inform targeted, sustainable policy measures.

Prior research using mobile phone and mobility data has identified high-risk venues such as restaurants, shops, theaters, cafes, bars, and hotels, and highlighted nightlife areas as particularly influential in Japan. Policymakers often referenced transit-station mobility when devising countermeasures. Early analyses (e.g., March–April 2020) using Google Community Mobility Reports and machine learning (including random forests) found retail/recreation, groceries/pharmacies, and transit to be important correlates of case trends. International studies (e.g., Germany, Italy, Portugal) reported that increased groceries/pharmacies mobility and decreased workplace and retail/recreation mobility correlated with case dynamics, and workplace closures were nearly as effective as stay-at-home orders. Japan experienced among the largest global declines in mobility, with notable increases in park use worldwide during the pandemic. However, medium-term relationships can be nonlinear due to policy changes, new variants, and behavioral adaptations, raising risks of overfitting for some machine learning approaches and underscoring the need for robust methods like random forests.

Study area and period: Osaka metropolitan area (Osaka, Kyoto, Hyogo Prefectures), March 1, 2020 to September 30, 2021. Data: (1) Human mobility from Google Community Mobility Reports across six categories—retail/recreation, groceries/pharmacies, parks, transit stations, workplaces, residential—reported as percentage change from a baseline (median Jan 3–Feb 6, 2020). (2) Daily newly confirmed COVID-19 cases by prefecture from Japan’s Ministry of Health, Labour and Welfare. Ethics: Approved by the Research Ethics Committee of the Graduate School of Life Science, Osaka City University (No. 21-58); use followed guidelines for de-identified location data. Analysis: Random forest regression was used to model the relationship between daily mobility (predictors) and the total number of COVID-19 cases after a two-week lag (response), reflecting the approximate incubation and reporting delay. Models were implemented in JMP Pro 16 with 10,000 trees, bootstrap sampling, and random feature selection at splits. Variable importance was assessed using dependent resampled inputs constructed via a k-nearest neighbors approach. Performance was evaluated using R². The analysis emphasized mobility behavior and relative importance to identify mobility categories most associated with subsequent case counts.

- Mobility patterns: After March 2020, all mobility types decreased except residential, which increased—indicating people stayed home more even without a formal stay-at-home order. Workplaces showed declines during holiday periods; groceries/pharmacies generally remained around 0% with slight changes during emergencies; parks initially increased during the first emergency then declined. - Epidemic waves and policy: Five waves occurred (1st: Apr–May 2020; 2nd: Jul–Sep 2020; 3rd: Dec 2020–Feb 2021; 4th: Mar–Jun 2021; 5th: Jul–Sep 2021). States of emergency (1st, 3rd, 4th, 5th waves) effectively reduced cases. - Model performance: Random forest models for each prefecture achieved R² > 0.7; Osaka Prefecture R² = 0.777. - Variable importance/effects: In Osaka, total effects were highest for groceries/pharmacies (0.437), parks (0.368), and workplaces (0.253), indicating these categories most strongly influenced case counts after two weeks. - Policy-relevant thresholds: Results indicate it is essential to keep groceries/pharmacies mobility within −5% to +5% of baseline and parks mobility above −20% (i.e., avoid large reductions in park use). - Transit finding: Transit stations mobility was not a significant source of infection in this context. With appropriate hygiene processes (e.g., maintained daytime train frequency, incentives to ride at less-crowded times), encouraging a return to transit use appears consistent with infection control and supports sustainability goals (e.g., CO2 reduction). - Parks: Increased park mobility is associated with fewer infections, suggesting parks can be actively promoted during restrictions, potentially substituting for higher-risk indoor activities.

The findings directly address where to control mobility to reduce infections. Constraining groceries/pharmacies mobility to a narrow band around baseline likely minimizes crowding and transmission in essential indoor venues. Encouraging or at least not overly restricting park use is associated with lower case counts, offering a safer outlet for activity during restrictions. Contrary to common policy practice that closely tracked transit-station mobility, the analysis did not identify transit as a key infection source under the hygiene and operational measures applied in Osaka (e.g., maintained service frequency, staggering demand). This implies potential to lift stringent transit-related restrictions, aiding mode shift back to public transport and advancing urban sustainability (lower emissions). The results also highlight heterogeneity across prefectures, implying that local characteristics and policy implementation nuances matter for tailoring mobility guidance. Collaboration between governments and private sector data holders could further refine understanding by linking de-identified mobility and infection data while preserving privacy.

This study identifies the mobility categories most associated with subsequent COVID-19 case counts in the Osaka metropolitan area during the medium-term pandemic. Using random forests on Google mobility data with a two-week lag, the analysis shows that maintaining groceries/pharmacies mobility near baseline (−5% to +5%) and avoiding large reductions in parks mobility (keep above −20%) are associated with reduced infections, while transit mobility was not implicated as a transmission source under Osaka’s hygiene protocols. Policymakers can prioritize managing essential indoor spaces and promote safe outdoor alternatives, while restoring confidence in public transit to support sustainability goals. Future work should integrate richer, privacy-preserving mobility datasets, explore finer spatial-temporal scales, and assess context-specific heterogeneity and evolving factors such as variants and vaccination.

- Medium-term dynamics likely involve nonlinear relationships influenced by shifting policies, behavioral adaptation, variants, and vaccination, which can challenge model stability and generalizability. - Aggregated, prefecture-level Google mobility data may mask within-prefecture heterogeneity and context (e.g., specific venues, time-of-day patterns). - Privacy constraints limit the ability to match individual-level mobility and infection data, constraining causal inference. - Findings are context-specific to the Osaka metropolitan area and to periods with particular hygiene measures in transit; applicability elsewhere depends on local practices and compliance. - While random forests mitigate overfitting relative to some methods, model interpretability and causal attribution remain limited in observational analyses.