Introduction
Confirmed cases of coronavirus disease 2019 (COVID-19) caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) were first reported in Wuhan, China in December 2019 [1]. The World Health Organization (WHO) declared the epidemic a Public Health Emergency of International Concern (PHEIC) on 30 January 2020 and a pandemic on 11 March 2020 [1]. COVID-19 is transmitted by inhalation of infected aerosols and contact with infected droplets [1]. Complicating its control, secondary transmission can occur from both asymptomatic and pre-symptomatic infected individuals [1]. Crowded places, close-contact indoor environments, and confined spaces with poor ventilation, often referred to as the ‘three C’s’, allow the virus to spread easily [1].
Human mobility data have already contributed to understanding the temporal patterns of COVID-19 [Reference Zhang2–Reference Jung6]. In the absence of effective vaccines or therapeutics during the early phase of the pandemic, the implementation of public health and social measures (PHSMs), including movement restrictions, physical distancing, mask-wearing, school and business closures, and travel restrictions, played a key role in suppressing the incidence at least temporarily and buying time for the development of pharmaceutical countermeasures. Published studies from many countries have attempted to estimate the effectiveness of PHSMs that aimed to reduce human-to-human physical contact in suppressing the incidence of COVID-19. Studies have found substantial reductions in the incidence during the early months of 2020 [Reference Vokó and Pitter7, Reference Nakajo and Nishiura8]. Many studies have also used modelling approaches to simulate counterfactual epidemic scenarios and evaluate the impact of movement restrictions on COVID-19 transmission dynamics [Reference Delport9–Reference Flaxman11].
During the early years of the COVID-19 pandemic in Japan, a ‘hammer and dance’ approach was adopted to suppress the impact of the pandemic on the healthcare system. Intensive PHSMs, including the declaration of states of emergency, were implemented during periods of rapid case escalation to slow down transmission, that is, the hammer phase and to prevent excessive case-load demand on healthcare system. Once the incidence declined to a manageable level by the healthcare system, these interventions were gradually relaxed and the low incidence level was maintained for as long as possible, that is, the dance phase [Reference Suzuki12]. This approach was also used in other countries, but the situation in Japan was special for two reasons. First, movement restrictions were ordered via the declaration of a state of emergency (SoE), but were purely requested and not strictly legally binding (i.e., initially there was no legal penalty for not adhering to the request). Second, a stay-home approach was taken during the first wave, but interventions were then gradually focused on situations with a high risk of transmission (e.g., eating and drinking facilities with alcohol, and mass gatherings). The second wave was brought under control without declaring a SoE or calling for people to stay at home. The controls were therefore not compulsory for the whole population, although it is likely that people tightly and strictly abided by the government’s suggestions, successfully suppressing the incidence (Figure 1) [Reference Suzuki12–14]. To understand the population impact of these countermeasures for future options against pandemics, it is interesting to assess the impact of movement restrictions on the transmission dynamics of COVID-19 in Japan.
Weekly confirmed cases during the first three COVID-19 epidemic waves in Osaka (a) and Tokyo (b). Orange- and red-shaded areas represent restriction measures and states of emergency, respectively.
One major source of data to assess the de-facto population size is spatial location data for mobile phones. In Japan, different sources of mobile phone location data have been used to assess the relationship between SARS-CoV-2 transmission and human mobility [Reference Kurita5, Reference Jung6, Reference Nagata15–Reference Okada17]. Nightlife districts with large numbers of eating and drinking establishments have been closely linked to COVID-19 transmission dynamics [Reference Nagata15, Reference Imamura18]. Two studies in particular found a positive relationship between the effective reproduction number of COVID-19 and the nighttime population, implying the significance of PHSMs in addressing human mobility during the evening and night [Reference Nakanishi16, Reference Okada17]. Because the nighttime de-facto population size captured the core temporal pattern of transmission, it provides a critical opportunity to retrospectively assess the effectiveness of interventions. The purpose of this study was to epidemiologically estimate the impact of PHSM in Japan in suppressing the incidence of COVID-19 by comparing the observed mortality and the counterfactual scenario in which no restrictions were in place.
Methods
Epidemiological dataset for COVID-19
The Infectious Disease Control Law requires physicians to notify any laboratory-confirmed cases of COVID-19 to Japan’s Ministry of Health, Labour and Welfare (MHLW) via local health centres. From 2020 to 2021, confirmation of COVID-19 was carried out through real-time reverse transcriptase polymerase chain reaction testing. We obtained the publicly available dataset of age-stratified confirmed cases and confirmed deaths from the Japanese government and from the metropolitan governments of Osaka and Tokyo [13, 14]. The scarcity of cases and the weekly reporting format for the de-facto population meant that we compiled the reported epidemiological data into a weekly dataset and every week started with Sunday. For Osaka, the study periods covered for the first, second, and third COVID-19 waves were 23 February–13 June 2020, 14 June–10 October 2020, and 11 October 2020–2027 February 2021, which was consistent with previously defined incidence-based timing of separate epidemic waves [Reference Kurahara19]. For Tokyo, the study periods covered for the first, second, and third COVID-19 waves were 9 February–23 May 2020, 24 May–19 September 2020, and 20 September 2020–2027 February 2021 [Reference Imamura18] (Supplementary Table S1). The prevalent SARS-CoV-2 lineages during the first three epidemic waves were B.1.1, B.1.1.284, and B.1.1.214, respectively [Reference Matsumura20]. The study period preceded the dominance of B.1.1.7 (so-called ‘Alpha’ variant), one of variants of concern (VOC), featured by increased transmissibility compared to the prior non-VOC strains [Reference Matsumura20]. It also preceded the introduction of the nationwide vaccination program against COVID-19 since February 2021 [Reference Kayano21].
Mobile phone location data
In this study, ‘LocationMind xPop’ data on the hourly population volume estimates in selected areas in the Tokyo and Osaka metropolitan areas were provided by LocationMind Inc. [22] ‘LocationMind xPop’ are data that NTT DOCOMO has processed statistically and comprehensively from cell phone location data sent with permission from users of the Zenrin Map Navi service provided by ZENRIN DataCom CO., LTD. The location information is GPS data (latitude and longitude information) taken at a minimum of every 5 min. They do not contain information that identifies individuals.
For each metropolitan area, mobile phone trajectories were used to selectively collect population volume that did not involve stay-at-work and stay-at-home behaviours. The areas of interest in this study were selected on the basis of designated areas for monitoring of people flow data by the Cabinet Office [23]. The ‘LocationMind xPop’ data showed within-week variations, and were therefore aggregated by epidemiological weeks. The periods of interest for the de-facto population size exactly corresponded to the first three COVID-19 waves described in Supplementary Table S1. We used the nighttime population data in the same format from 2019 as the counterfactual de-facto nighttime population size for each city.
Night time population model
The nighttime population data can be used to explain the socially heterogeneous nature of the COVID-19 transmission during the early epidemic waves in Japan. Contact tracing of early cluster outbreaks showed concentration of cases among people in high-risk and high-contact settings like drinking and entertainment establishments in nightlife districts [Reference Nagata15, Reference Nakanishi16, Reference Imamura18]. Limitations using such data include bias towards younger age groups, especially among those aged 20s and 30s [Reference Imamura18], exclusion of other transmission settings like workplace, hospitals, nurseries, and households, and inability to take into account the changing public behaviour (e.g., young people drinking at home). However, there was a long-term positive association of the nighttime population with the COVID-19 transmission throughout the pandemic in Japan [Reference Okada17].
Our study used mobile phone GPS-based nighttime population size data to establish the effective reproduction number R t, the average number of newly infected cases caused by an infected individual during an epidemic in the presence of changing susceptibility and various countermeasures. The response variable in our model was therefore the effective reproduction number during reporting week k and epidemic wave i, R k,i, and the explanatory variables were the natural logarithmic forms of the nighttime population and its relative change from the previous reporting week.
where n k is the weekly nighttime population during reporting week k. Of course, this is the same as
and, therefore, we have
The time lag was taken to be 1 week, as the closest value to the 8 days estimated by another study [Reference Okada17]. The coefficients α i, β i, and γ i varied with the corresponding epidemic wave i. The R k,i estimated by our model was applied to the next-generation matrix component in the multi-type renewal equation model [Reference Green, Ferguson and Cori24] and a heterogeneous contact pattern among four different age groups, that is, those aged 0–19, 20–39, 40–59, and ≥ 60 years [Reference Jung25], was considered. We then estimated the unknown parameters using the maximum likelihood method (see Supplementary Material). We also derived the 95% confidence intervals (CIs) of these parameters using the non-parametric bootstrap method. A comparison between the nighttime population model with wave-specific parameter coefficients, Eq. (1), and the model with fixed parameter coefficients across all three epidemic waves combined was made to assess the stability of the relationship between the nighttime population and the effective reproduction number (see Supplementary Material). Additionally, a simpler nighttime population model with only α i, and β i coefficients was compared with Eq. (1) (see Supplementary Material).
Counterfactual scenarios
To construct the counterfactual epidemic waves, we replaced the weekly observed nighttime population data with the corresponding weekly pre-pandemic 2019 data in each of Osaka and Tokyo prefectures. The counterfactual is therefore the absence of restriction, and assumed that in 2020 without the pandemic restrictions, everyone would have behaved as they did in 2019. Counting from the first week in which a primary case was identified, each counterfactual simulation was carried out for 52 weeks. During the counterfactual scenario analysis, depletion of susceptibles from the total population was explicitly taken into account in computing R k (see Supplementary Material).
We assumed there would be no overlapping of epidemic waves, no re-infection within the same epidemic wave (i.e., the susceptible group is monotonously depleted), no immunity from infection during a previous epidemic wave, and that confirmed cases represent 25% of all infected individuals [Reference Sanada26]. The median final cumulative epidemic sizes for each of the four age groups during each epidemic wave were estimated along with the 95% CIs using the non-parametric bootstrap method.
Empirically estimated age-dependent confirmed case fatality risk (cCFR) during each epidemic wave was used to compute the mortality (Supplementary Figure S1). The cCFR was calculated by dividing observed age-specific deaths by observed age-specific confirmed cases throughout the corresponding epidemic wave. We assumed that each age-specific cCFR remained a constant over the course of the counterfactual epidemic.
A sensitivity analysis was carried out using different mean generation times (3.0 days, 4.6 days, and 5.9 days) [Reference Nishiura, Linton and Akhmetzhanov27] to estimate variations in the R k, and counterfactual age-dependent cases and deaths during the first epidemic wave in Osaka (see Supplementary Material). A second sensitivity analysis was carried out, assuming an increased cCFR among the elderly population in the counterfactual scenarios for all epidemic waves in both Osaka and Tokyo (see Supplementary Material).
Summary of assumptions
The following key assumptions were made in our study model. Assumption 1: the SoE and other PHSMs influenced the effective reproduction number, and the change in the effective reproduction number was regulated by the change in the nighttime population. Assumption 2: no SoE or mobility-restricting PHSMs were implemented in counterfactual scenario and this scenario was captured by using the pre-pandemic 2019 data. Assumption 3: the generation time was a constant 3.5 days for all three epidemic waves. Assumption 4: the observed cCFR during each wave remained unchanged in counterfactual scenarios (i.e., not elevated even when healthcare capacity was overwhelmed). Assumption 5: Counterfactual epidemic waves were considered distinct with no overlapping; there was no re-infection within each wave, no immunity from infection in previous waves, and only 25% of infected cases were reported.
Statistical analysis
All statistical and numerical calculations used JMP® Pro, Version 17.0.0. and JMP® Student Edition, Version 18.2.0. SAS Institute Inc., Cary, NC, USA, 1989–2023.
Results
Model fit
Figure 2 and Supplementary Figure S2 visually assess our model fit, comparing observed and modelled age-specific incidence in Osaka and Tokyo, respectively. The model contained the majority of observed data points within the predicted uncertainty bounds. R k using our regression model Eq. (1) and the 95% CIs were also compared with R k estimated from the epidemiological data (Figure 3a–c, Supplementary Figures S3a, S3b, and S3c), where both predicted uncertainty bounds overlapped for most of the weeks.
Comparison of the weekly observed and predicted cases by age group during the first three COVID-19 epidemic waves in Osaka. (a), (b), and (c) show the weekly observed (blue dots) and predicted (red lines) COVID-19 cases among different age groups during the first, second, and third COVID-19 epidemic waves, respectively. These waves started from week 23–29 February 2020, 14–20 June 2020, and 11–17 October 2020, respectively. For the first two waves, predicted cases were obtained using the nighttime population model from week 5 onwards. For the third wave, predicted cases were obtained using the nighttime population model from week 3 onwards except week 13. For all three waves, estimates from weeks 1 and 2 were not calculated using our method. Estimates from weeks 3 and 4 during both the first and second waves, and estimates from week 13 during the third wave were obtained using the effective reproduction numbers calculated from the epidemiological data. The shaded areas represent the 95% confidence intervals (CI) of the weekly predicted cases computed using the nonparametric bootstrap method.
Comparison of the weekly reproduction numbers estimated from the epidemiological data and predicted by the nighttime population model during the first three COVID-19 epidemic waves in Osaka. (a), (b), and (c) compare the weekly reproduction numbers estimated from the epidemiological data (blue dotted lines) and predicted by the nighttime population model (red lines) during the first, second, and third COVID-19 waves, respectively. (d), (e) and (f) compare the weekly reproduction numbers predicted by the nighttime population model (red lines) observed during the first, second, and third COVID-19 waves, respectively, and the counterfactual version from 2019 data (green dashed lines) for the corresponding epidemiological weeks. The waves started from week of 23–29 February 2020, 14–20 June 2020, and 11–17 October 2020, respectively. As an exception, for the nighttime population model (red lines), the reproduction numbers in weeks 3 and 4 during both the first and second waves, and week 13 during the third wave were estimated from the epidemiological data. The blue whiskers represent the normal-based 95% confidence intervals (CIs) of the blue dotted lines, and the red shaded areas represent the profile-likelihood 95% CI of the red lines. The green dashed lines were obtained by replacing the 2020 nighttime population data with those from 2019, the red lines. The green whiskers represent the 95% CIs of predicted effective reproduction numbers of the counterfactual scenario computed by the nonparametric bootstrap method. There is a decrease in the reproduction number over time because both the mean estimate and the CIs of the green lines were multiplied by the susceptibility depletion factor with the assumed ascertainment rate at 25%. The gray horizontal lines in all panels represent the threshold value of reproduction number at 1, below which the epidemic will die out. The first state of emergency was in effect from 7 April 2020 to 21 May 2020 and the second from 13 January 2021 to 1 March 2021.
During the comparison of our model with wave-specific parameter coefficients and the model with fixed parameter coefficients for all three waves combined, we found a better model fitness to the incidence data with our model in Eq. (1). The latter significantly underestimated or overestimated the incidence data during the peaks of the second and third waves across all age groups. This was more pronounced in the 20–39 years group (Supplementary Figure S4). A comparison between our model in Eq. (1) and a simpler model with only two parameter coefficients showed that the former had lower Akaike information criterion (AIC) values for all epidemic waves in Osaka and Tokyo except the third wave in Osaka (Supplementary Tables S2 and S3). A similar model comparison was performed for all three epidemic waves combined and the model with three parameter coefficients had lower AIC values for both Osaka and Tokyo (Supplementary Tables S4 and S5).
Counterfactual simulations
Figure 3d–f and Supplementary Figures S3d, S3e, and S3f compare estimated R k against the counterfactual R k based on the nighttime population data for the same epidemiological week in 2019. Using the counterfactual R k, the epidemic scenario without restrictions was derived, allowing the computation of the causal impact of PHSMs. The estimated counterfactual epidemic sizes substantially surpassed the observed values for the three epidemic waves in both Osaka and Tokyo (Figure 4 and Supplementary Figure S5). However, during the second wave in Tokyo, the 95% CIs of the cumulative counterfactual cases included the observed cumulative counts for all age groups.
Comparison of the observed and counterfactual cumulative number of cases during the first three COVID-19 epidemic waves in Osaka. Panels (a), (b), and (c) present the observed (blue lines) and counterfactual (red lines) median cumulative cases during the first, second, and third COVID-19 epidemic waves, respectively. Within each panel, sub-panels (i), (ii), (iii), and (iv) show the corresponding results by age group: 0–19 years, 20–39 years, 40–59 years, and ≥ 60 years, respectively. The unit of measurement of cases was Log10(case+1) people. The waves started from week of 23–29 February 2020, 14–20 June 2020, and 11–17 October 2020, respectively. The counterfactual waves started from the corresponding epidemiological weeks in 2019, and the simulation was run for 52 weeks. The shaded areas represent the 95% confidence intervals (CI) of the counterfactual cumulative number of cases computed using the nonparametric bootstrap method.
In Osaka, with a population of about 8.5 million in 2020, the estimated counterfactual cumulative reported cases (and deaths) were 1.47 million (71525), 0.92 million (13589), and 0.75 million (18814) for the first three waves, respectively (Table 1). These estimated cases were up to 814, 99, and 21 times the observed values for the first three waves, respectively. In Tokyo, with 13.8 million people, the estimated counterfactual cumulative reported cases (and deaths) were 2.15 million (57018), 0.84 million (4361), and 1.21 million (17100) for the first three waves, respectively (Table 1). The case estimates were 420, 42, and 14 times the observed values, respectively.
Cumulative numbers of cases and deaths of first three epidemic waves of COVID-19 in Osaka and Tokyo, Japan
Table 1. Long description
Seven survey respondents are labelled from A to G and their answers to five statements are shown, describing how far they agree or disagree. The statements were, I learnt something new, the tutor was engaging, I would recommend the course, I enjoyed the course, and I can apply what I learnt in my role. Four respondents answered positively to all the statements, selecting either Agree or Strongly Agree, one answered with either positive or neutral answers, and two answered all the statements with negative answers, selecting either Disagree or Strongly Disagree, plus one Neutral.
Abbreviation: CI: confidence interval.
Note: Observed and counterfactual data are compared. Impact estimates were calculated by taking the ratio of median counterfactual estimate to the observed value.
Figure 5a shows the result of the sensitivity analysis using different mean generation times for the first wave in Osaka. R k was not very sensitive to different mean generation times, while the estimated cumulative deaths and cases during the counterfactual first wave in Osaka would be larger if the mean generation time was shorter than the 3.5 days assumed in our main analysis (Figure 5b and Supplementary Figure S6). Additionally, a sensitivity analysis on the cCFR with increased elderly mortality when the healthcare system was overwhelmed was conducted in the counterfactual simulations across all epidemic waves in Osaka and Tokyo (Supplementary Table S6). Since the majority of the observed mortality came from the elderly population aged 60 years and over, the impact of increased elderly mortality significantly contributed to the overall increase in counterfactual deaths.
Sensitivity analysis on mean generation times in predicting weekly reproduction numbers and counterfactual mortality for the first COVID-19 epidemic wave in Osaka from 23 February 2020 to 13 June 2020. (a) The weekly reproduction numbers, R, using different assumptions of mean generation time (3 days, 3.5 days, 4.6 days, and 5.9 days) starting from week 3, 8–14 March 2020. R values before week 5 were estimated from the epidemiological data and from week 5 onwards were predicted by the nighttime population model. The gray horizontal line represents the threshold value of R at 1, below which the epidemic will die out. (b) Using different assumptions of mean generation time (3 days, 3.5 days, 4.6 days, and 5.9 days), counterfactual cumulative mortality by age group was predicted over a period of 52 weeks. The whiskers represent the 95% confidence intervals computed using the nonparametric bootstrap method.
Discussion
Using the nighttime population to establish the reproduction number, this study attempted to retrospectively evaluate the effectiveness of PHSMs across the first three COVID-19 epidemic waves in Osaka and Tokyo. After reconstructing the observed epidemic waves, the counterfactual scenarios were drawn by swapping the nighttime population data for 2020 with those for 2019, allowing the computation of hypothetical situations in which SoE and/or other mobility-restricting PHSMs were not in place. The most drastic difference was seen during the first wave for both Osaka and Tokyo, where the counterfactual reported cases and deaths were estimated to be over 800 and 400 times the observed reported cases and deaths, respectively, highlighting the impact of PHSMs in suppressing the epidemic.
The key take-away message from our study is that SoE and other mobility-restricting PHSMs had substantial effectiveness in reducing the epidemic size and mortality during the first three COVID-19 waves in Osaka and Tokyo. Although accompanied by huge uncertainties, the median counterfactual estimate of cases was more than 16 times the observed count, implying that the indirect population impact of PHSMs was substantial. Our study underscores the importance of substantially suppressing the early incidence of a pandemic via PHSMs to buy time for pharmaceutical control options to be developed.
To our knowledge, this study is the first to have used the pre-pandemic nighttime population in 2019 to derive the counterfactual scenario for PHSMs, enabling causal modelling analysis that has also been applied in a vaccine effectiveness study [Reference Kayano21]. An essential underlying assumption for the counterfactual analysis is that people in 2020 would have behaved similarly to 2019 in the absence of PHSMs. The COVID-19 pandemic acted as an innovative opportunity to predict the effective reproduction number using mobility, temperature, population density, and other predictive variables, as seen in other studies [Reference Nouvellet3–Reference Jung6, Reference Nagata15–Reference Okada17, Reference Bokányi28–Reference Vanni30]. In line with this, our regression model of R k linearly used nighttime population data. However, its extension to the counterfactual scenario had to assume that the predictive performance of the regression model remained unchanged even when the epidemic was in progress and transmission was widespread in the community in the absence of PHSMs.
We analysed three distinct epidemic waves separately as the comparison between the model with fixed parameter coefficients and our model with wave-specific parameter coefficients showed that the latter could capture the incidence data better. This is partly because public health interventions progressed as a function of time, but also because contact behaviours and societal responses evolved over the course of pandemic. The first wave in 2020 was tackled by declaring the SoE and people were advised to stay at home. The announcement of the first SoE on 7 April 2020 in both Osaka and Tokyo [Reference Nakajo and Nishiura8, 31] was followed by a huge reduction in human mobility including at night. When the SoE was lifted, the risk awareness among the population remained but self-restraint declined over time. During the second wave, PHSMs in Japan evolved and challenged epidemic control by targeting only the high-risk section of the population [Reference Nakajo and Nishiura8, Reference Jung32]. The epidemic was brought under control without having to declare a SoE. The effective result, and the smaller population impact of PHSMs during the second wave, especially in Tokyo, implies that population-wide contact was different from before the first wave, and people continued to use precautionary behaviours during the ‘new normal’ life. This idea is supported by the absence of any statistically significant difference in cumulative cases between the observed and counterfactual scenarios during the second wave. Compared to the first wave, the observed cCFR among the elderly during the second wave was lower most likely due to different age compositions of cases, improved diagnosis, and early treatment and contact tracing, although the change in pathogen virulence was also possible. This might potentially have led to a relatively lower population impact of PHSMs on mortality. The depletion of susceptible population from prior epidemic waves was not considered during wave-specific analyses and, therefore, its possible contribution towards a smaller population impact of PHSMs during the second wave especially in Tokyo could not be explained by our model. The third wave occurred following an interplay between precautionary ‘new normal’ behaviours and the promotion of economic activities including subsidization of travel (e.g., Go To travel campaign). A second SoE was announced for the third wave on 13 January 2021, in Osaka and 7 January 2021, in Tokyo [31] and caused another fall in nighttime population volumes. This led to a significant population impact estimate of PHSM than in the second wave. There are probably more practical explanations for the difference as well, but our study indicates that the population effectiveness of PHSM can vary with the type of intervention, people’s behaviour, and background societal situations.
There are advantages and disadvantages of inferring the effectiveness of states of emergency and other mobility-restricting PHSM using the nighttime population data. First, we did not take into account autonomous psychological and social responses, including those informed by risk awareness. A previous study found that increases and decreases in incidence regulated mobility during the pandemic [Reference Okada17]. The mobility restrictions during the states of emergency were not legally binding, and our estimates must therefore have included societal psychosocial reactions as part of intervention effectiveness. Second, during the SoE, both daytime and nighttime population contact was reduced. The use of a variable covering the high-risk population (nighttime population data) in the model therefore did not capture the contact from or impact of daytime social activities. Our simplistic model was designed to develop causal inferences using a relatively strong predictor variable (nighttime population data) and we did not aim to achieve quantitative forecasting capturing all social heterogeneities. Third, the transmission was probably oversimplified in this study, but our method was innovative in predicting the effective reproduction number during the COVID-19 pandemic using mobility data [Reference Nouvellet3–Reference Jung6, Reference Okada17, Reference Bokányi28–Reference Vanni30]. Our study emphasizes the importance of using strong predictors for causal impact assessment, and the simple nighttime population model provided the best fit to resolve the unanswered question in Japan (i.e., the effectiveness of interventions during the first, second, and third epidemic waves).
Our study also had several limitations. First, three counterfactual epidemic waves were simulated separately and the length of each simulation was set arbitrarily at 52 weeks or one year. In reality, the second and third waves were continuous with no distinct demarcation. Second, we simplified the immune response from natural infection with the wild-type SARS-CoV-2 by assuming that there was no re-infection within the same wave and no acquired immunity from infection during previous waves. Third, this study focused on Tokyo and Osaka, two metropolitan prefectures, and the analyses did not account for spatial interactions (e.g., with other prefectures or countries). This was essential for our computation, but would change the results. Fourth, behaviour was different during certain weeks in the initial periods of the epidemic wave and around the new year holidays, and the nighttime population model could not be used to derive the effective reproduction numbers. In those cases, we used the effective reproduction numbers calculated from the observed epidemiological data, rather than employing the regression model. Fifth, only the 2019 nighttime population data was used in the counterfactual simulations, and we could not assess the consistency of the pre-pandemic nighttime population data over multiple years.
Conclusion
If the 2020 nighttime population had remained unchanged from 2019, we could have seen far more reported cases and deaths due to COVID-19 in both Osaka and Tokyo. Estimating the effective reproduction number using mobility data provides critical insights into the impact of mobility-restricting PHSMs on COVID-19 transmission dynamics.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0950268826102234.
Data availability statement
All data and JMP files necessary to reproduce the results are available as supplementary material.
Acknowledgements
The authors would like to thank Tetsuro Kobayashi and Katsuma Hayashi for assisting data access. They thank the local health centres and the Infectious Disease Surveillance sections of Tokyo Metropolitan Government and Osaka Prefecture for their work in making time-series data publicly available in real time. The authors also thank Melissa Leffler, MBA, from Edanz (https://jp.edanz.com/ac) for editing a draft of this manuscript.
Author contribution
Conceptualization: H.N.; Formal analysis: T.M.S.; Investigation: T.M.S.; Methodology: T.M.S., M.N., Y.O., and H.N.; Supervision: H.N.; Validation: T.M.S.; Visualization: T.M.S.; Writing – original draft: T.M.S.; Writing – review and editing: T.M.S. and H.N.
HN conceived the research idea. All authors contributed to building statistical models. TMS conducted statistical analyses on the datasets and HN evaluated the validity of the analyses. TMS drafted the manuscript including figures and tables. HN edited earlier versions of the manuscript. All authors read and approved the final version of the manuscript.
Funding statement
TMS received financial support as a scholarship from the Ministry of Education, Culture, Sports, Science and Technology (MEXT) of Japan for his doctoral degree program. MN was supported by Japan Science and Technology Agency (JST) SPRING fellowship (JPMJSP2110). YO received funding from the SECOM Science and Technology Foundation. HN received funding from Health and Labour Sciences Research Grants (grant numbers 20CA2024, 21HB1002, 21HA2016, and 23HA2005), the Japan Agency for Medical Research and Development (grant numbers JP23fk0108612, and JP23fk0108685), JSPS KAKENHI (grant numbers 21H03198 and 22 K19670), the Environment Research and Technology Development Fund (grant number JPMEERF20S11804) of the Environmental Restoration and Conservation Agency of Japan, the Daikin GAP Fund of Kyoto University, HU-RIZONT international research excellence program (Rapid-GRIP project: 2024–1.2.3-HU-RIZONT-2024-00034 to HN), Japan Science and Technology Agency SICORP program (grant numbers JPMJSC20U3 and JPMJSC2105), the CREST program (grant number JPMJCR24Q3), and RISTEX program for Science, Technology, and Innovation Policy (grant number JPMJRS22B4). The funders had no role in the study design, data collection and analysis, the preparation of the manuscript, or the decision to publish.
Competing interests
The authors declare that they have no known competing commercial or financial interests that could influence the content of this study.
Ethical standard
We used anonymized GPS location data and publicly available epidemiological data, and ethical approval was not required because of the nature of our study.
Open access
