Infection data were collected during a 34-week period from March 1 to November 14, 2020. The study period was determined to observe the initial confirmed case reports in Tokyo, followed by the implementation of self-restraint–based contact reduction across all areas of Tokyo and social measures in restaurants and bars by the Tokyo Metropolitan Government. The number of new COVID-19 case reports and symptom onsets per day was made publicly available at websites by the Tokyo Metropolitan Government [16]. The number of symptom onsets was sequentially added over 3-week periods of prompt reports, as there were several time lags between symptom onset, testing, diagnosis, and inclusion in the daily report of new cases. We thus obtained data on the number of symptom onsets as of December 5, 2020. In this study, the weekly numbers of symptom onsets and confirmed case reports were used for analysis.

The effective reproduction number (R_t), the average number of secondary cases generated by a single primary case, was estimated as a function of the estimated calendar date of infection. To achieve this, the date of infection was statistically inferred using confirmed cases that were divided into two groups: cases with (1) known date of illness onset and (2) known date of confirmation and without known date of illness onset. For the latter group, we back-projected the suggested date of illness onset nonparametrically, and then we summed these cases with group (1). Subsequently, we implemented nonparametric back-projection of the date of infection using the incubation period distribution with a mean of 5.2 days. Using the assumed median generation time of 4.6 days [17], the renewal equation was employed to estimate everyday R_t. The Bayesian Markov chain Monte Carlo method was implemented using rstan, version 2.19.3; the estimation code is available in the GitHub repository [18]. To control for local heterogeneities that are exhibited as daily fluctuations, we smoothed the time series by computing the 7-day rolling averages, and we adopted the number on the fourth day of each week for analysis.

Population volumes at 10 PM to midnight in seven Tokyo metropolitan areas were estimated using location data from smartphones created by LocationMind xPop. LocationMind xPop uses aggregated people flow data originally collected by NTT Docomo, Inc, through their application service “Docomo Map Navi” using only cell phone location data collected with user consent to the auto GPS function of the service, and the data are then statistically processed by NTT Docomo in their entirety before being provided to LocationMind Inc [19]. The original location data are GPS data (latitude and longitude) sent at a frequency of every 5 minutes at the shortest interval, and these data do not include information that specifies individuals [20]. NTT Docomo is the largest mobile phone operator in Japan, and it accounts for approximately 44% of total mobile phone subscribers [21]. The population per hour was estimated using the location data and calibration with the national census.

The users in this study agreed to provide their location information. The data are anonymized so that individuals cannot be specified, and personal information such as gender, age, and occupation are unknown. The population in this study represents the number of mobile devices that appear in a 500 m² grid at a specific time. Mobile phone trajectories were used to distinguish and extract on-site dining behavior from stay-at-work and stay-at-home behaviors. A stay point was assumed at which the GPS data of a device are concentrated within the area with positional errors for approximately 15 minutes or longer. A user’s place of work and home were estimated using the information on the stay points, including period of time and length of stay per user. Stay points that were spatially distinct from the estimated place of work and home were accounted for with stay episodes other than stay-at-work and stay-at-home. The following seven Tokyo metropolitan areas were selected to represent districts with restaurants and bars: Kabuki-cho (mesh code 533945361), Ginza-Corridor-gai (mesh code 533946002), Shibuya-Center-gai (mesh code 533935952), Ueno-Nakamachi-dori (mesh code 533946512), Shinjuku-Ni-chome (mesh code 533945264), Ikebukuro (mesh code 533945764), and Roppongi (mesh code 533935984). The selection was based on designated areas in Tokyo for monitoring of people flow data by the Cabinet Office [22].

Because using a single data source at a specific time may not represent the population distribution in the whole period, the average population concentration was used to estimate the nighttime population volumes in seven Tokyo metropolitan areas between 10 PM and midnight. A scaling factor was calculated using the estimated place of home for each user in combination with the national census, which gave a value of the total population represented by the user. The number of people staying in a certain area per hour was calculated by summing the scaling factors of the users. The estimated number of people per hour was calculated on a weekly basis to meet a sufficient sample size, as areas that contain few persons should be removed to prevent the identification of individuals based on their location.

During the first wave of COVID-19, self-restraint–based contact reduction was implemented for the entire population of Tokyo for a 7-week period from April 7 to May 24, 2020. First-wave countermeasures included declaration of a state of emergency, call for self-restraint in travel and going out, restrictions of the maximum capacity of events and gatherings, and restrictions of use of facilities where clusters were occurring and the “3Cs” (closed space, crowded space, and closed contact setting) were observed [23]. Restaurants and bars were asked to cooperate by shutting down their services at 8 PM during this period.

During the second wave, social measures for restaurants and bars in Tokyo were in place for a 6-week period from August 3 to September 15, 2020. Restaurants and bars serving alcohol were asked to shorten their operating hours to 10 PM for this 6-week period.

Neither the first nor the second wave interventions were legally binding; for example, owners of restaurants and bars were asked to cooperate with the shutdown of services, and instead, the owner received a subsidy from Tokyo. In the analysis, the implementation of public health and social measures was presented as a dichotomous variable (implemented=1, absent=0).

The initial and peak weeks were identified by visual inspection of the observed time series in the night-time population, symptom onsets, confirmed case reports, and R_t.

A vector autoregression (VAR) model was employed to predict endogenous outcomes (ie, nighttime population volume, symptom onsets, confirmed case reports, R_t) while investigating how they are influenced by each other in terms of Granger causality [24]. VAR is a system of equations with outcomes that depend on other outcome variables. Suppose we have a vector of time series data Wt, Xt, Yt, and Zt (four endogenous variables); then, the VAR model with one exogenous variable and p lags can be expressed as below:

where ν is the intercept, ρ is the coefficient of exogenous variable C, α is the coefficient of the lagged terms of W, β is the coefficient of the lagged terms of X, γ is the coefficient of the lagged terms of Y, δ is the coefficient of the lagged terms of Z, and υ is assumed to be white noise. In the W equation, β, γ, and δ infer how W is influenced by other outcomes (X, Y, and Z) in terms of Grander causality.

The number of lags can be statistically evaluated by information criteria such as the Akaike information criterion (AIC). To measure the effect of restrictions on growth in the nighttime population and infections, the first difference in the natural log was used for a respective time series to stabilize the variances with removing trends, as was applied in a published study [25]. Regarding R_t, the first difference was used for analysis. First-wave countermeasures and social measures implemented by restaurants and bars were included in the VAR model as exogenous covariates. Of interest to our research was the effect of the implementation of restrictions on restaurants and bars on the nighttime population volume and the lagged coefficient of the nighttime population volume in symptom onsets and R_t. Granger causality tests were performed for each equation to assess whether any pairs among nighttime population volume, symptom onsets, confirmed case reports, and R_t had a causal relationship.

A sensitivity analysis was conducted using the vector autoregression model in which the symptom onsets and confirmed case reports were excluded from the endogenous variables.

All statistical analyses were performed using Stata, version 16.1 (StataCorp LLC). The significance level in the two-tailed tests was set to .05.

The increase in the nighttime population in the seven examined districts in Tokyo was followed by increased symptom onsets and increased R_t thereafter. Even during the intervention period, the nighttime population increased significantly following the decrease in confirmed case reports. The implementation of social measures by restaurants and bars was not significantly associated with the nighttime population during the second wave in Tokyo.

The results of our study imply that people adjusted their level of mobility according to the number of confirmed case reports and, thus, their perceived risk of their own infection. They may have reduced on-site dining behavior voluntarily after an increasing number of new confirmed cases was reported and then relaxed their mobility restrictions after a decreasing number of cases was reported. Although people dynamically adjusted their behavior in response to information and policies [26] and recognized that dining inside restaurants should be avoided [27], public attitudes related to behavioral restrictions changed as the COVID-19 epidemic changed. It is notable that there were considerable time lags between behavior changes, symptom onsets, and R_t. A time lag between major mobility reduction and the peak of confirmed cases was also observed in South Korea [28]. There could have been a delay in the voluntary reduction in on-site dining from the time that suppression of the surge of COVID-19 should have begun. This latency was also observed in the reaction to policy measures to restrict mobility in northern Italy [29]. Therefore, policy measures to mitigate COVID-19 should be planned at an earlier stage informed by social behavior dynamics, rather than based on the number of confirmed case reports. Early implementation of strategies is crucial to successful COVID-19 suppression and to avoid national lockdowns [30]. A tracking system of the nighttime population using mobile phone location data would be helpful for policy decision makers to monitor these dynamics in a real-time manner. Space-time dispersions in transmission of COVID-19 have occurred from metropolitan areas towards the countryside [31]. Thus, an automated information system to support strategic policy decision making is of particular importance in metropolitan areas with high population density and mobility, where there is an elevated risk of COVID-19 transmission.

The implementation of first-wave interventions in early 2020 was significantly associated with an increase in the nighttime population volume in Tokyo. Less restrictive measures, requesting that restaurants and bars shorten their operating hours, were not significantly associated with the nighttime population volume. The nighttime population volume started to decrease one month before the implementation of first-wave countermeasures and social measures to restaurants and bars, and it reached the minimum in the early phase of the interventions. People may have reduced on-site dining behavior earlier than the implementation of public health and social measures in response to preceding policy arguments and announcement of the implementations along with the increase in confirmed cases. It should be noted that the decreasing trend appeared to be weaker against the second wave than in March 2020. Although we could not confirm this from our data, many people may have experienced caution fatigue by repeated COVID-19 surges and the highest number of new cases, and they may have become tired of physical distancing. Another possible explanation is that the public health and social measures were suboptimal to suppress the recovery of on-site dining behavior based on the level of the restrictions and the timing of the implementation. However, we could not determine the optimal level of restrictions and length of interventions with regard to COVID-19 mitigation or the economic effects of the interventions. Closure of restaurants and bars may result in decreased consumption expenditure and economic decline. Further investigation is needed to estimate the effect of reductions in the nighttime population volume on the economy and on COVID-19 transmission. This estimation can also inform the government as to how financial compensation should be provided to entities that comply with requests to close at night. In Japan, the national government launched a Go To Eat campaign to encourage people to dine at restaurants starting in October 2020 [32]. This campaign aimed to support the recovery of the industry from economic decline caused by first-wave countermeasures and other behavior restrictions; however, there are concerns that it increased on-site dining and that it resulted in the recent large resurgence of COVID-19 [33]. Strategies to support businesses need to be organized in conjunction with a level of behavior restrictions to prevent COVID-19 resurgence.

This is the first study to relate public health and social measures to prevent COVID-19 transmission to the nighttime population volume in Tokyo. A strength of this study is the estimation of probable on-site dining behavior as opposed to stay-at-work and stay-at-home behavior using mobile phone location data. However, our study has some limitations that must be noted. The suspected mode of transmission in each case was unavailable due to the nature of open data, potentially leading to underestimation of the association between the nighttime population volume and the onset of symptoms. Lags may have been shorter if symptom onset were stratified by transmission via on-site dining. The risk of COVID-19 transmission in on-site dining could also be mitigated over time as restaurants and bars increasingly adopted physical distancing guidelines [34]. The number of reported cases in the first wave might have been suppressed because of limited testing capacity in early 2020. The effectiveness of the measures to reduce on-site dining may vary across regions and countries that have different response levels to policies among the general population. The level of effectiveness would also differ according to the stage of pandemic (first, second, or later), as public health and social measures against later waves can be subject to cumulative caution fatigue against COVID-19. Social behavioral dynamics could also be caused by the media’s reporting on COVID-19 [35]. Future examination on the association between on-site dining and media reporting will be beneficial to provide guidelines to the media that support effective mobility restrictions.

Regardless of restrictions on restaurants and bars in Tokyo metropolitan areas, people reduced their mobility restrictions after decreases in new confirmed cases were reported. An increase in the nighttime population volume may result in increased symptom onset and R_t. A tracking system of the nighttime population using mobile phone location data would be helpful for policy decision makers to monitor social behavior dynamics and suppress COVID-19 transmission at an earlier stage.