Pandemic data
In Japan, COVID-19 has been designated a notifiable infectious disease according to the Infectious Disease Control Law [27]; cases diagnosed using a nucleic acid amplification test or antigen test are notified as confirmed COVID-19 cases. Confirmed cases are mandatorily reported to the government within 24 h of diagnosis. During the study period of the present investigation, all suspected patients were requested to quarantine for 14 days [28], and all underwent real-time reverse transcription polymerase chain reaction testing on days 7 and 14. As a result of these control measures, the cumulative incidence of COVID-19 by the end of 2020 was 235,700 cases, which was less than 0.2% of the national population.
In this study, we obtained the COVID-19 incidence reported from September 10 to November 9, 2020 from open data published by the Ministry of Health, Labour and Welfare [29]. The data were based on an online surveillance system, the Health Center Real-time Information-sharing System on COVID-19 [30].
We intentionally restricted our analyses to this time frame, taking into account the 30 days before and after the start of the second travel campaign (October 1, 2020). The time delay in reporting SARS-CoV-2 infection in Japan was approximately 9 days during the study period [31]. To obtain an epidemic curve as a function of the time of infection, we shifted the entire epidemic curve (originally drawn by reporting date) to the left for a fixed time delay of 9 days. The reporting involved weekend bias; therefore, we used a 7-day rolling average, and we counted the number of prefectures that underwent an incidence of three, five, and seven COVID-19 cases per 100,000 population per week. These values were specifically used because the pandemic situation in each prefecture was classified into four discrete stages during late 2020 (Additional file 2: Tables S1 and S2). Detection of 15 cases per 100,000 population per week led prefectural governments to declare a stage III pandemic and consider stringent public health and social measures that could involve restrictions on movement and other personal rights (Additional file 2: Tables S1 and S2) [32]. Stage IV involved 25 cases per 100,000 people. Even if prefectures could control the pandemic below stage III, they were advised to monitor the incidence. Stage II did not involve an explicit threshold value: each prefecture was advised to determine the value based on the local epidemiological situation. The World Health Organization guidance applies thresholds for the level of community transmission of 20, 50, and 150 COVID-19 cases per 100,000 people [33]. Following our exploratory analysis, we found that the number of newly reported cases per population in Japan during the study period was relatively low compared with other nations. Fewer than four prefectures exceeded the threshold of 10 per 100,000 people. Thus, in the present study, we employed slightly lower threshold levels of three, five, and seven cases per 100,000 population per week to capture local epidemic activity. Those lower threshold levels allowed us to determine more clearly how the situation changed with low levels of incidence.
We obtained the population estimates by prefecture from the Statistics Bureau of Japan [34] and the daily average temperature in each prefecture during the corresponding time period from the Japan Meteorological Agency [35]. As the representative value for temperature across all 47 prefectures, we applied the median value of daily average temperature from the data for each prefectural capital. During the study period, the proportion of positive cases among the total number of weekly tests remained below 10%, and there were no major changes (Additional file 2: Figs. S1 and S2). We did not standardize the number of newly infected cases because there were no significant differences in the consistency of reporting among prefectures. In the study period, B.1.1.284 and B.1.1.214 were the dominant SARS-CoV-2 lineages, but they were not determined to be variants of concern (Additional file 2: Fig. S3).
Interrupted time-series model
Given the start date of the second campaign (October 1, 2020) and 9 days’ reporting delay, we set a campaign period in the interrupted time-series model from October 10 to November 9, 2020. The control period was 30 days before the second campaign (intervention) period. Through the corresponding time periods, we investigated the number of prefectures that had COVID-19 incidence in excess of the defined thresholds for weekly incidence (i.e., three, five, and seven cases per 100,000 population), Yt, which can be modeled as model 1:
$${Y}_{t}={\beta }_{0}+{\beta }_{1}T+{\beta }_{2}{X}_{t}+{\beta }_{3}\left(T-{T}_{i}\right){X}_{t},$$
(1)
where T is the time elapsed from the start of observation, Xt is the dichotomous variable representing the campaign state (0, pre-campaign; 1, post-campaign), and Ti is the time when the campaign started. \({\beta }_{0}\) is the parameter for the baseline level of the outcome, \({\beta }_{1}\) for increase in the outcome following the time-unit increase, \({\beta }_{2}\) change in the level of outcome immediately after the campaign, and \({\beta }_{3}\) the rate of increase following the campaign.
It has been reported that temperature may be associated with SARS-CoV-2 transmission [36, 37]; therefore, we developed an extended model (model 2) by incorporating temperature into model 1. Model 2 was as follows:
$${Y}_{t}={\beta }_{0}+{\beta }_{1}T+{\beta }_{2}{X}_{t}+{\beta }_{3}\left(T-{T}_{i}\right){X}_{t}+{\beta }_{4}Z,$$
(2)
where Z is the median value of the daily average temperature for Japan’s 47 prefectures.
We also examined a model without considering the immediate change in outcome after implementing the second campaign (\({\beta }_{2}\)). Model 3 was as follows:
$${Y}_{t}={\beta }_{0}+{\beta }_{1}T+{\beta }_{3}\left(T-{T}_{i}\right){X}_{t}+{\beta }_{4}Z.$$
(3)
Assuming that \({Y}_{t}\) follows a Poisson distribution, we applied the maximum-likelihood method to estimate all parameters; we derived the 95% confidence intervals (CIs) of the estimates using the parametric bootstrap method with 10,000 samples. Lastly, to select the best model among those proposed, we calculated the Akaike information criterion.
To examine the robustness of our results, we conducted sensitivity analyses. Our sensitivity analysis covered different geographic groups, timing of the intervention, holiday periods, the different models, and the study period (additional analysis). To exclude the possibility that an increase in infections was caused simply by geographic bias—especially heterogeneities associated with urbanization—we conducted a subgroup analysis using two discrete prefectural groups (urban and nonurban groups) based on each prefecture’s population density. To account for uncertainty about the time of illness onset to official reporting, we undertook sensitivity analysis using different values for the timing of the intervention. We also conducted interrupted time-series analyses: (1) using holiday periods as a possible explanatory variable; (2) employing another functional model with an exponential function for inference; and (3) extending the length of the study period until termination of the campaign.
Data-sharing statement
The original data analyzed in the present study are available in the Additional file 1.