Public transit mobility as a leading indicator of COVID-19 transmission in 40 cities during the first wave of the pandemic

Study design, setting, and study period

We conducted a longitudinal ecological study to estimate the association between public transit mobility reductions and population-level COVID-19 transmission 2 or 3 weeks later. We included 40 cities covered by Citymapper (excluding only Monaco, which did not record 100 cumulative COVID-19 cases by the end of the study period) (Fig. S1). We considered COVID-19 transmission during the 4 weeks between March 16, 2020 and April 12, 2020.

Public transit mobility

Citymapper is an app that provides users in select major cities with real-time information on public transportation options, alongside other modes of transport. The Citymapper Mobility Index (CMI) was released on March 19, 2020 (Citymapper, 2020b) and included data on 41 cities and metropolitan areas across 23 countries and five continents (Citymapper, 2020c). Historical measurements were available back to March 2 for most cities. CMI measures the relative frequency of trips planned within the app compared to a “recent typical usage period” of January 6, 2020–February 2, 2020 (exceptions: February 3, 2020–March 1, 2020 for Paris and December 2, 2019–December 22, 2019 for Hong Kong and Singapore) (Citymapper, 2020a). A value of 100% indicates a typical volume of trips planned with the Citymapper app, whereas values above or below 100% indicate increased or decreased trip volumes, respectively.

To compare the evolution of mobility as measured by CMI to the announcement of formal government restrictions, we plotted the value of CMI before and after the first announcement of a major national or sub-national physical distancing intervention, namely gathering restrictions and/or mandatory closures (Table S1).

COVID-19 transmission

We measured COVID-19 transmission using two different quantities: the weekly case ratio and the effective reproduction number (R_t).

To calculate these quantities, we obtained time series of COVID-19 case counts for cities represented in the Citymapper dataset (Table S2). For most cities, it was possible to obtain case data at the local level (e.g., cities, counties, metropolitan areas) corresponding to the area served by the Citymapper app. For a small number of cities where local data were not available, we used national or sub-national data. These data should nonetheless serve as good approximations for the global cities represented in our dataset, as initial transmission of the disease was concentrated in these dense, internationally connected hubs.

We defined the weekly case ratio for a given week as the ratio between cases reported in that week and cases reported in the week prior. A value of 1 indicates the number of cases reported in the present week is unchanged from the prior week.

The effective reproduction number represents the average number of secondary infections a person infected at time t would be expected to generate given that conditions remain unchanged (Fraser, 2007), with a value of 1 marking the threshold between the exponential growth (R_t > 1) and exponential decay (when R_t < 1). We estimated R_t via the parametric serial interval method (Cori et al., 2013) implemented in the EpiEstim (version 2.2-4) (Cori, 2019) package in R (version 4.1.3) (R Core Team, 2021). We used estimates of the serial interval of the original wild-type virus by Du et al. (2020) (mean = 3.96 days, SD = 4.75 days).

For each location, we calculated the weekly case ratio and R_t for the 4 weeks comprising the study period (March 16, 2020 and April 12, 2020). The weekly case ratio could not be calculated during the week beginning March 16 for the cities of Amsterdam, Lyon, and Paris due to insufficient data. Istanbul and Montreal were excluded from the week of March 16 for both the weekly case ratio and R_t because the calculated weekly case ratios exceeded 30. There was a change in how tests were confirmed during this period in the province of Quebec, where Montreal is situated (Rocha, 2020).

Statistical analysis

We used multilevel linear regression models to estimate the association of public transit mobility reductions with COVID-19 transmission during the initial wave of the pandemic. We fit two sets of models using different outcomes: the logarithm of the weekly case ratio and the effective reproduction number. The primary exposure variable was the weekly mean CMI 2 weeks prior. We chose a 2-week lag to reflect the delay between infection and symptom onset (the incubation period) and the delay between symptom onset and when the results of confirmatory testing are reported (the reporting lag). A meta-analysis using data from the first few months of the pandemic estimated the incubation period of COVID-19 at 6.5 days (Alene et al., 2021). We estimate reporting lag to be between 5–15 days early on in the pandemic, since testing was scarce and rapid diagnostics were unavailable. Some estimates of reporting delays include 6.4 days in Singapore (up to March 17) (Tariq et al., 2020), 4–10 days in Bavaria, Germany (up to April 8) (Günther et al., 2021), and 7–9 days in Japan (up to May 13) (Miyama et al., 2022).

We fit models using four pairs of weeks: comparing average weekly CMI for the weeks beginning March 2, March 9, March 16, and March 23 to the case ratio and reproductive number 2 weeks later (the weeks beginning March 16, March 23, March 30, and April 6, respectively). Most countries implemented their first major physical distancing policies throughout the second week of March (March 9–March 15), so this 4-week period covers time both before and after official government interventions. For most countries in the dataset, this period was a time of growth and/or an early peak in the first wave of the pandemic (Dong, Du & Gardner, 2020).

The multilevel linear regression models used random intercepts to account for the clustering of outcomes by city, nested within countries. Models were fit using glmmTMB (version 1.1.8) (Brooks et al., 2017). To adjust for epidemic timing, we fit adjusted models for both outcomes that included a new covariate: days since the 100^th case in the country. For the first week of case data (March 16), this covariate was negative for some cities. Unadjusted and adjusted models were compared with a likelihood ratio test. A complete mathematical description of the models used in this analysis is available in the Supplemental Methods.

Sensitivity analyses

To account for uncertainty in the reporting delay, we fit an alternative set of models using a 3-week lag between CMI and outcomes instead of a 2-week lag. In these models, we could use only 3 weeks of outcome data (March 23, March 30, April 6), since the CMI dataset begins on March 2 for most cities.

Data on imported versus locally acquired cases were not available for most cities, so we were unable to adjust for this factor during the analysis. Since many countries implemented international travel restrictions during the third week of March (Hale et al., 2021), cases reported at least 2 weeks later should be much less likely to represent imported cases than cases reported before this time. As a sensitivity analysis, we re-ran models excluding outcome data from the third and fourth weeks of March (beginning March 16 and March 23).

To assess whether the association of CMI with outcomes differed across the 4 weeks included in the 2-week lag models (both unadjusted and adjusted for epidemic timing), we re-fit these models with an additional interaction term between CMI and week and compared them to the original models using a likelihood ratio test.

Ethics

This study was exempt from review because it used exclusively publicly available data.

Public transit mobility

Nearly all cities experienced substantial reductions in public transit mobility during the month of March (March 2: median = 100%, first quartile (Q₁)–third quartile (Q₃) = 94–108%; March 29: median = 10%, Q₁–Q₃ = 6–15%) (Fig. 1). Decreases were less pronounced in Hong Kong, Seoul, Tokyo, and Milan, where public transit mobility was already substantially reduced at the beginning of the month. Cities in Europe, Australia, and the Americas showed similar patterns in public transit mobility reductions generally coinciding with the dates of national or sub-national physical distancing measures, including restrictions on public gatherings or business closures (Table S1).

Weekly case ratio

The weekly value of the mobility index was associated with the logarithm of the weekly case ratio 2 weeks later (Fig. 2, Fig. S2). A 10% lower mean mobility index was associated with a 12.3% lower weekly case ratio 2 weeks later (exp(β) = 0.877; 95% confidence interval (CI): [0.859–0.896]) (Table 1). This model explained a portion of the variance in weekly case ratios (conditional R² = 0.595). When the model was adjusted for days since the 100^th case in the country (a measure of epidemic timing), the association was attenuated to 7.8% (exp(β) = 0.922; 95% CI: [0.885–0.961]) and variance explained increased slightly (conditional R² = 0.641). Being an additional 10 days since the 100^th case in the country was associated with a 20.8% lower weekly case ratio (exp(β) = 0.792; 95% CI: [0.668–0.938]).

Since reporting delay is unknown and variable across geographies, we also ran a model using a 3-week lag for CMI, which produced slightly weaker associations when compared to the full model (Table 1, Fig. S3). The sensitivity analysis excluding the first 2 weeks of outcome data (which have a higher probability of representing imported cases, rather than local transmission) produced comparable results to the full model (Table S3). In the sensitivity analysis of adding an interaction between CMI and week, the likelihood ratio tests did not favor the interaction model in either the unadjusted or adjusted analyses (unadjusted p = 0.400, adjusted p = 0.911).

Effective reproduction number

The weekly value of the mobility index was associated with the effective reproduction number 2 weeks later (Fig. 3, Fig. S4). A 10% lower mean mobility index was associated with a decrease in the effective reproduction number 2 weeks later (β = −0.058; 95% CI: [−0.068 to −0.048]) (Table 1). This model explained a portion of the variance in the effective reproduction number (conditional R² = 0.587). When the model was adjusted for days since the 100^th case in the country, the association was attenuated (β = −0.032; 95% CI: [−0.052 to −0.013]) and variance explained increased slightly (conditional R² = 0.647). Being an additional 10 days since the 100^th case in the country was associated with a decrease in the effective reproduction number (β = −0.121; 95% CI: [−0.202 to −0.040]).

Using a 3-week lag for CMI produced a slightly weaker strength of association (Table 1, Fig. S5). The sensitivity analysis excluding the first 2 weeks of outcome data (which have a higher probability of representing imported cases, rather than local transmission) produced roughly comparable results to the primary analysis, based on 2-week lag (Table S3). In the sensitivity analysis of adding an interaction between CMI and week, the likelihood ratio tests did not favor the interaction model in either the unadjusted or adjusted analyses (unadjusted p = 0.539, adjusted p = 0.958).