Abstract
This study investigates how human mobility has changed in the long-term in response to the COVID-19-related information in Japan. We use publicly available data from Google on human mobility in retail & recreation and residential spent time. These variables can be explained using daily data on the number of infected cases, whether the state of the emergency is declared or not, and the cumulated number of the vaccinated person. In the regression analysis, we use the ‘interactive effects model’ to control complicated unobservable factors that vary across time and cross-sectional dimensions. Our regression results find that people feared an unknown virus in the 1st wave, but the habituation trend for human mobility is noticed for the repeated similar infection information. However, from a different kind of information about the spread of new variants, people’s habituation comes to a halt to some extent. Further, the spatial interaction of infection information is observed. We also show that people reacted to infection information even without a state of emergency declaration. Also, vaccination promotion encourages people to go out with security. When implementing policies to control human mobility, it is essential to consider the timing, and the degree of information penetration, carefully.
1 Introduction
Coronavirus disease 2019 (COVID-19) infections have had a significant socio-economic impact for the whole world including Japan [1, 2]. Transmission of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) occurs through direct, indirect, or close contact with an infected person [3]. Thus, the increased human mobility raises opportunities for human-to-human contact and spreads COVID-19 infection [4-8]. Hence, human mobility control called Non-Pharmaceutical Interventions (NPIs), such as lockdown is an effective method for reducing COVID-19 infections [9-14], human mobility [15, 16], or both [17]. Additionally, when the infection spreads, people will act in a risk-averse manner out of caution and fear, and stop going out, thereby decreasing human mobility [18-20]. From March to May 2020, in the U.S., consumption-related travel fell by 60% overall, of which 7% was due to the lockdown, and the rest was due to people’s own choices; they restricted their travel due to fear of infection [19]. From these perspectives, both human mobility and COVID-19 infection are profoundly interrelated with a negative relationship [21].
How human behaviour reacts to various types of information regarding COVID-19 is a significant concern regarding socio-economic impact and control of COVID-19 transmission. In particular, people are exposed to two main pieces of information: first, the increase in COVID-19 infections, and the NPI [22-25]. There are studies of behaviour variations over time against the COVID-19-related information. Using data from 124 countries, there was a gradual decline in adherence to physical distancing actions under the continuance of NPIs from March to December 2020 [26]. Also, there was a diminished ratio of stay-at-home, a behavioural change in response to increased infection in Japan over three waves: from March to May 2020 (first wave of COVID-19), July to September (second wave), and from October to December (third wave) [24]. A study using data through November 2020 shows that stay-at-home reactions to the information about the number of infected people declined after May 2020, when Japan’s first state of emergency declaration was lifted [27]. Regarding the declaration of a state of emergency, an NPI issued between 2020 and 2021 in Japan, the study finds that the extent of curtailment of outgoing behaviour decreased from the first to the fourth declaration of the state of emergency [28, 29].
However, few have conducted long-term analyses of human behaviour reactions to both COVID-19 infections and the NPI. As the COVID pandemic has prolonged for almost two and half years (as of July 2022), people are supposed to have gradually become accustomed to two types of information for alerts. Long-term analysis is crucial for evidence-based policy making (EBPM).
The first COVID-19 case was found on January 16, 2020, in Japan [30]. In response to the expansion of COVID-19 infection, the Japanese government issued a COVID-19 emergency declaration for the first time on April 7, 2020 [31]. Since then, as of July 2022, Japan has been through six waves of COVID-19 infection [32]. In the first wave of COVID-19 infection, people were quite alarmed and afraid of an unknown virus and an unprecedented pandemic [33]. Thenceforth, a COVID-19 vaccine was developed, and Japan’s approval was on February 14, 2021, for Pfizer Inc. and May 21, 2021, for Takeda Pharmaceutical Company Limited / Moderrna Inc. [34]. In addition, therapeutic agents are being developed [35], and much is known about the post-infection situation [36]. Furthermore, people are getting tired of restraint due to the COVID-19, so-called ‘Pandemic Fatigue [37, 38].’ Therefore, although recently, as of July 2022, we have faced the Omicron variant and its subvariant with a much faster infection rate [39, 40], it is also true that we are getting used to COVID-19.
For two years since the start of the pandemic, Japan initiated a rapid vaccination program in the first half of 2021 [41] and witnessed the emergence of the new variants. However, human behavioural changes in response to new information over the past two years under the COVID-19 pandemic are still in question.
This study will first examine whether people have become habituated to information about the spread of infection after six waves of infection over two years by examining the variation of the degree of human mobility in response to the COVID-19 infection information. Habituation [42, 43] refers to the diminishing response to a stimulus over time as the stimulus is repeated [44]. To our knowledge, no studies to date have empirically examined habituation by exposure to COVID-19 infection information over a long period.
There are two phases in a single wave of infection: increasing of infection, and decreasing of infection. The responses of human mobility are considered heterogeneous between the two phases. In the increasing phase, people are less likely to go out, and in the decreasing phase, they gradually begin to go out. However, the more fearful they are, the more inhibited and the more cautious they are in both phases. We validate this by figuring out the start, the end, and the peak of infection of different waves in each prefecture in Japan.
On top of this, we consider spatial interactions between prefectures using cross-prefecture human travel to and from other prefectures. Since changes in human mobility in one prefecture are also influenced by the other prefectures’ information. The importance of considering regional spatial interaction under the COVID-19 pandemic is pointed out in some studies [45, 46]. Therefore, we incorporate a cross term using a spatial weights matrix representing spatial interactions between prefectures. In addition, information about an increase in the cumulative number of people vaccinated by the prefecture would also reduce fear of infection and promote outgoing behaviour [47]. From the strand of literature, this investigation is not accomplished so far. The impact of the multiple emergency declarations issued will also be re-examined. We use the ‘interactive effects model [48]’ in regression analysis to control complicated unobservable factors regarding human mobilities that vary across time and cross-sectional dimensions.
While this study does not directly assess people’s attitudes, it indirectly explores changes in people’s fear, risk awareness, and ‘pandemic fatigue’ by examining people’s mobility responses to COVID-19-related information over time.
2 Data
2.1 Human mobility
Human mobility data is from Google’s COVID-19 Community Mobility Reports daily data [49]. This data consists of retail & recreation, grocery & pharmacy, parks, transit stations, workplaces, and residential. Regarding six types of human mobilities: we select retail & recreation, and residential. When avoiding unnecessary mobility, either to avoid infection or as a response to government’s declaration of a state of emergency, many people reduce outings to retail stores (not grocery and pharmacy) and entertainment places. Similarly, to avoid infection or to comply with a state of emergency declaration, there is a greater tendency to stay-at-home. Residential in Google’s data indicates time spent at home.
Google’s COVID-19 Community Mobility Reports is a percent change compared to a day-of-week baseline calculated from median values by day of the week for the five weeks from January 3 to February 6, 2020, just prior to the global outbreak of the COVID-19 infection. Since the data fluctuates by the days of the week, by taking the difference from the previous week, we consider the variation width from the previous week-day (percentage points) of percentage change from the baseline. Using the difference of percentage change from the previous week, rather than the percentage change from the base date itself, one would capture the behaviour of people who are deciding whether to go out or not in response to new information varying in the short term, such as daily newly infected cases. We denote this dependent variable as Δyit, where i = 1, 2, …, N stands for the prefecture, t = 1, 2, …, T is the index of day, and Δ is the difference from the previous week.
2.2 Infected cases of COVID-19
Data on the daily number of newly infected cases of COVID-19 are obtained from NHK (NIPPON HOSO KYOKAI; Japan Broadcasting Corporation) [50]. Since the number of newly infected cases has very large range it is better to, for example, take the logarithm of the data to mitigate heteroscedasticity. However, this method is not possible since the data contains zero. So instead, an inverse hyperbolic sine (IHS) transformation is used, which transforms 0 to 0 and performs a transformation similar to the log transformation [19, 22, 27]. Let Iit, denote newly infected cases, IHS transformation is shown below,
where 
is IHS of new infections. On top of that, newly infected cases have day-of-week variations. For this reason, IHS of new infections is converted to the difference from the same day of the previous week. Hence, the IHS-difference of daily newly infected cases 
approximates the growth rate from the previous week of newly infected cases.
There is another implication of taking the previous week’s difference. As an example, according to an NHK news article on August 13, 2021 [51], the peak date of infection in Tokyo during the Delta variant was dominant, “A record number of 5,773 cases were confirmed in Tokyo on Friday, August 13. This is the highest number ever recorded. The number of cases has also increased by 1,258 since a week ago Friday, and the rapid spread of infection continues (translated from the Japanese article by author).” Daily news is delivered mainly on the number of new infections of the day and changes in the number of infections compared to the same day the previous week. Simply put, people have been judging the severity of the infection situation based on the change from the same day of the previous week. In this way, approximating the growth rate from the previous week of newly infected cases reasonably illustrates people’s infection information received.
2.3 The declaration of a state of emergency
The declaration of a state of emergency is obtained from the Cabinet Secretariat ‘COVID-19 Information and Resources [52].’ The timing of the declaration of a state of emergency varies between prefectures. We set a dummy variable that takes 1 if a state of emergency is declared to the prefecture and 0 if not, denoting Eit. Fig.S1 in the Supplementary Information displays the periods of the declaration of a state of emergency for each prefecture.
As for NPI in Japan, the lockdown accompanied by legal penalties has not been implemented. Instead, a state of the emergency declaration has requested people to refrain from going out except for essential and urgent affairs. The declaration also has requested stores to shorten their hours of operation or to temporarily close down the business. Japanese policies for controlling the COVID-19 expansion are so-called soft lockdown [53] or voluntary lockdown [27]. Thus, people’s outgoing behaviour depends also on whether a state of emergency is declared or not, and they voluntarily decide their behaviour [54]. In light of the above, in this study, we also consider the declaration of a state of emergency from the Japanese government as COVID-19-related information.
2.4 Vaccination
The number of daily COVID-19 vaccination data by prefecture is from ‘COVID-19 vaccination status’ by Digital Agency [55]. The data is reworked into a cumulative number to produce data on the vaccination rate per million persons. Population data by the prefecture as of October 1, 2020, are obtained from ‘Population Estimates,’ the Statistics Bureau of Japan [56].
Vaccination in Japan began in February 2021 [57], and vaccination was accelerated in 2021 [41]. Due to the announcement at the time by the Ministry of Health, Labour and Welfare that the COVID-19 vaccine provides high preventive efficacy with two doses, the majority of those who received the first dose also received the second dose [58]. Therefore, there is an almost perfect correlation between the first and second vaccination rates (correlation coefficient = 0.99). Additionally, since the number of vaccinations varied with the day of the week. Therefore, the data is converted to the increased range from the previous day of the week, and denoted as ΔVit.
As for COVID-19-related information, the larger the increase in vaccination rate per population compared to last week, the more people have been vaccinated and form antibodies, which we assume means that people are more comfortable going out.
2.5 Time series plots
Time series plots of our concerning variables are in Fig. 1. For visibility, we demonstrate 7-day backward moving average using the geometric mean of human mobility in retail & recreation, the IHS transformation of the 7-day backward moving average number of infected persons, and the IHS transformation of the cumulative number of people vaccinated for each of 1 to 3 doses. Pink-coloured shades are the period of the declaration of a state of emergency.
Time series plots of our concerning variables used in the estimation. The blue-green-coloured line (on the right axis) is the 7-day backward moving average using the geometric mean of human mobility in retail & recreation. The red-purple-coloured line is the IHS transformation of the 7-day backward moving average number of infected persons. The purple-coloured two-dash lines are the IHS transformation of the cumulative number of people vaccinated for each of 1 to 3 doses. The pink-coloured shades are the period of the declaration of a state of emergency.
2.6 Spatial weight
To construct a spatial weighted matrix, ‘Cross-prefecture Travel in Each Prefecture’ is acquired from V-RESAS provided by the Cabinet Secretariat and the Cabinet Office. This data is based on ‘Agoop Corporation’s Current Population Data’ and is the demographic data based on day/night population from GPS data obtained with user consent from specific smartphone applications. The data is made by the average weekly migratory population across prefectures in 2019 as 1 and is of ISO 8601-week number base. There are two types of data, movement from other prefectures to the relevant prefecture and movement from the relevant prefecture to other prefectures. We choose the former one. Furthermore, there are two types of population movement: composition (%) and index. We chose the composition ratios because our focus is on which prefectures are travelled to (j) and from (i) more frequently weekly. Details of the construction of spatial weighted matrix 
are in the Methods section.
During the pandemic, the greater the travel population from the specific prefecture i, the more the COVID-19 trend in that prefecture is expected to substantially affect the human mobility inside the relevant prefecture j. There are mainly two aspects to this. One is that the higher the people’s interaction, the higher the risk of infection transmission across prefectural borders. Another is that, for instance, for those who commute, trends in the prefecture i they travel to are of concern.
Using the constructed spatial weighted matrix 
, we will create the cross-terms, spatially weighted infected cases, 
, and spatially weighted state of emergency declaration, 
. These are to inspect the impact of information on the number of infected people in other prefectures and the impact of information on the announcements of a state of emergency declaration in other prefectures, respectively. In a strand of studies regarding human mobility, few studies have considered these spatial interactions.
As an illustration, using V-RESAS’s Cross-prefecture Travel for the fifth week of 2020, the last week of January 2020 (January 27-February 2, 2020), the spatial weighted matrix is illustrated in Fig. 2. This period is just before the pandemic, and the irregular movements due to the New Year in Japan have already calmed down. Thence it is considered to be the period when normal inter-prefecture travel can be observed. From Fig. 2, the dark-red coloured cells that the more people travel between prefectures, are found in or around large cities such as Tokyo, Aichi, Osaka, and Fukuoka.
Spatial weighted matrix for the 5th week of 2020, the last week of January 2020 (January 27 to February 2, 2020). Constructed using ‘V-RESAS’s Cross-prefecture Travel’ data by the Cabinet Secretariat and the Cabinet Office. The darker the red colour, the more people travel between prefectures.
2.7 Controls
Control variables are also employed. We exploit temperature and precipitation data, taking from Japan Meteorological Agency, which would be relevant to human mobility [59]. Besides this, day-of-the-week dummies and holidays-dummies are prepared for time fixed effects. As for the prefectural fixed effects, population density per square kilometre of inhabitable land area and percentage of the population over 65 years old are extracted from the ‘Regional Statistics Database’ (System of Social and Demographic Statistics), the Statistics Bureau of Japan [60]. Further details of the controls are in the Supplementary Information. The controls are denoted as Cit.
2.8 Setting up wave periods of COVID-19 infection
This study ascertains the human mobility response to each wave of infection, so it needs to identify each wave. Since there is no official announcement from the government, each prefecture’s duration of the wave of COVID-19 infection is determined. Further details are in the Supplementary Information.
3 Results
3.1 Empirical strategies
We use the ‘interactive effects model’ [48] to control for unobservables regarding human mobility, our object of the analysis. For example, suppose a new variant has been identified in some country, and there is concern about a future epidemic in Japan however, the sense of caution differs between prefectures with high population density (the infection is seriously spread) and low population density (not as serious). In such a case, when common factors vary over time and affect each individual (prefectures in our case) heterogeneously, this method is powerful. Details are in the Methods section.
We use long-term panel data with 47 prefectures for the cross-section. The time dimension is 762 days from February 22, 2020, to March 25, 2022. Consequently, the total sample size is 35,814.
Lags are taken from 1 day to 7 days for our concerning variables: growth rate from the previous week of newly infected cases, previous week’s difference in vaccination rate per population, and the spatial weighted of these variables. This is because whether people go out on a day depends not only on information on infection and vaccination status from the previous day, but also from over the past few days (here, we assume one week). In contrast, the declaration of a state of emergency is valid on the day it is issued, so there is no need to take lags. Similarly, lags are not taken for the control variables, day-of-the-week dummies, holidays-dummies, temperature, and precipitation since they are only relevant for that day. The population over 65 years old and the population density are used as fixed effects. Estimation is performed 7 times stepwise from 1 to 7 in daily lag to account for multicollinearity. Only results estimated at lag 1 are presented for variables that are not taken lag. The estimation is also conducted for robustness using the polynomial degree 2 Almon lag model.
3.2 Retail & recreation human mobility response
3.2.1 Infected cases in increasing phase
Inspecting the phase of increasing infection in Fig. 3, the fear of new infectious diseases leads to a high degree of suppression of human mobility in the first wave, and although the second wave is lower than the first wave, it is still high. Excluding insignificant results, a 1% increase in the number of infected cases from the previous week resulted in a maximum of 1.29 percentage points (pp) (lag: 2, estimated coefficient (Est.) = - 1.290, standard error (s.e.) = 0.118, 95% confidence interval (CI) = [-1.521 to -1.059], P < 0.001) week-on-week (WoW) decrease in the percentage deviation from the baseline of human mobility, in the first wave. The suppression of human mobility in the first wave decreased with each passing day and a minimum of 0.65 pp decrease is observed (lag: 7, Est. = -0.649, s.e. = 0.253, 95% CI = [-1.145 to -0.153], P = 0.010). The 6 and 7-day lags have a wide confidence interval, indicating that the behaviour on these lag days after receiving infection information varied widely among prefectures. In the second wave, the mobility responses also decrease with each passing day. We see a maximum of 0.70 (lag: 3, Est. = -0.701, s.e. = 0.059, 95% CI = [-0.817 to -0.585], P < 0.001) and a minimum of 0.31 pp WoW decrease (lag: 7, Est. = -0.307, s.e. = 0.034, 95% CI = [-0.374 to -0.240], P < 0.001), respectively. The third wave shows a more modest response than the first and second waves, indicating a tendency towards habituation, the tendency for people to become accustomed to similar infection information. The response remains flat for each lag day. A maximum is a WoW decrease of 0.29 pp (lag: 4, Est. = -0.288, s.e. = 0.078, 95% CI = [-0.441 to -0.135], P < 0.001).
Retail & recreation human mobility reactions to the COVID-19-related information. Points are estimated coefficients, and bars indicate upper and lower 95% confidence intervals. Infection waves are six, but the state of emergency declaration has been issued only on the first, third, fourth, and fifth waves. Infected cases in increasing phase, infected cases in decreasing phase, spatially weighted cases, and vaccination are taken daily lag from 1 day to 7 days. We conduct the regression analysis seven times, from lag-1-regression to lag-7-regression. The state of emergency declaration, spatially weighted state of emergency declaration, and controls are not taken daily lag; these estimates are from lag-1-regression.
In the fourth wave, the Alpha variant spread, followed by the Delta variant in the fifth wave, and in the sixth wave, the Omicron variant became predominant [61]. Each new variant exacerbated the rapidness of infectivity [62], and people feared each new variant. Accordingly, the human mobility response rose in the fourth wave. The degree of suppression of human mobility increases on lag days 4 and 5, with a maximum WoW decrease of 0.49 pp (lag: 5, Est. = -0.491, s.e. = 0.064, 95% CI = [-0.616 to -0.366], P < 0.001). In the fifth wave, human mobility is in a suppression trend from the 3 to the 7 lag days, showing the same human mobility suppression level as in the fourth wave, although the first and second lags are insignificant. The wave show a maximum of 0.39 pp WoW decrease (lag: 5, Est. = -0.387, s.e. = 0.132, 95% CI = [-0.646 to -0.128], P = 0.003). The sixth wave also shows the same level of suppression as the fourth and fifth waves. The suppression trend gradually strengthened from lag days 1 to 4, and the estimates are significant through lag 7. This wave has a maximum of 0.42 pp WoW decrease (lag: 4, Est. = -0.422, s.e. = 0.155, 95% CI = [-0.726 to -0.118], P = 0.006).
Unlike up to the third wave, this tendency from the fourth to the sixth waves is possible since the number of infected people increased more than expected due to the spread of the variants, and people were more reluctant to go out after a few days of receiving information about the successive days’ information of infections increase. Interestingly, the suppressing effect on human mobility is not as serious as in the first and second waves. Thus, one can argue that there is a tendency for habituation in general.
3.2.2 Infected cases in decreasing phase
Regarding the phase of decreasing infection in Fig. 3, if the negative range of the estimated value is extensive, the WoW increase in the percentage change from the baseline human mobility will be prominent during the recovering phase. Suppose people become less fearful of infection, then the increase in the rate of change (the negative range of the estimate) will be substantial. Also, if people’s mobility decreases greatly during the exacerbating phase, a WoW difference in the percentage change from the baseline will be a positive repercussion during a recovering phase.
As a result, in the first wave, the negative range is relatively large, so it seems to be a repercussion of the increasing phase. However, human mobility remains somewhat suppressed since the magnitudes of the estimates are smaller than the ones in the increasing phase. A 1% decrease in the number of infected cases will result in a maximum of 0.48 pp (lag: 1, Est. = -0.483, s.e. = 0.136, 95%CI = [-0.750 to -0.216], P < 0.001) and a minimum of 0.2 pp (lag: 5, Est. = -0.196, s.e. = 0.068, 95% CI = [-0.328 to -0.064], P = 0.004) WoW increase in human mobility. In the first wave, the fear of new infectious diseases seemed strong. The same was true for the second and third waves, which are smaller than the magnitude of the increasing phase. In the second wave, the maximum is 0.32 pp (lag: 7, Est. = -0.317, s.e. = 0.043, 95% CI = [-0.401 to -0.233], P < 0.001), and the minimum is 0.15 pp (lag: 5, Est. = -0.152, s.e. = 0.059, 95% CI = [-0.268 to -0.036], P = 0.010) on WoW changes. Only lag 5 is significant for the third wave, but the estimated coefficient is small. They suggest that people were cautious about the recovery of human mobility in 2020 or early 2021.
In contrast, the fourth wave in the spring of 2021 shows the similar magnitude as in the increasing phase, a maximum of 0.44 pp (lag: 1, Est. = -0.438, s.e. = 0.057, 95%CI = [-0.550 to -0.326], P < 0.001) and a minimum of 0.23 pp (lag: 6, Est. = -0.232, s.e. = 0.068, 95% CI = [-0.364 to -0.100], P = 0.001), WoW increase in human mobility. This high magnitude may be due to the start of vaccination and the increasing availability of information on infection of COVID-19. In the fifth wave, conversely, the estimates are no longer significant. One reason may be that lag 1 in the increasing phase is insignificant. Another reason may be that the Delta variant is more transmissible than the Alpha variant [63]; thus, people were cautious of the new variant. Finally, in the sixth wave, only lag 1 shows a significant impact of 0.60 pp in lag day 1 (Est. = -0.600, s.e. = 0.132, 95% CI = [-0.859 to -0.341], P < 0.001); however, after lag 2, results are insignificant except for lag 5, which shows a small estimated coefficient. The Omicron variants are more transmissible than previous variants but are less severe [63]. Therefore, in the decreasing phase, people tend to go out on the next day when the number of infected patients decreases.
3.2.3 Spatially weighted infected cases
We do not separate the increasing and decreasing phases for the spatially weighted infected cases. We realize that people’s outgoing behaviour significantly reacts to infection information of the other prefectures in the first wave. The maximum is 1.00 pp (lag: 2, Est. = -1.000, s.e. = 0.149, 95%CI = [-1.292 to -0.708], P < 0.001), and the minimum is 0.62 pp (lag: 4, Est. = -0.622, s.e. = 0.270, 95%CI = [-1.151 to -0.093], P = 0.021). However, the effect of information on infection in other prefectures in the first wave is insignificant after lag 5. In the second through the fourth wave, the results are insignificant for most lags; otherwise, the magnitudes are modest.
On the contrary, in the fifth wave, lag 5 through 7 are significant, with a maximum of 0.31 pp (lag: 7, Est. = -0.305, s.e. = 0.078, 95%CI = [-0.458 to -0.152], P < 0.001). The sixth wave also shows considerable reactions with the largest of 0.34 pp (lag: 3, Est. =-0.339, s.e. = 0.053, 95%CI = [-0.443 to -0.235], P < 0.001), and all the lags are significant. The fifth and sixth waves show epidemics of the more infectious variants, the Delta and Omicron, which have once again affected people’s going out behaviour due to information about infections in other prefectures. Overall, the spatial interaction of the infected cases was substantial for the first, fifth, and sixth waves, when people were afraid of the new infectious disease and its more infectious variants.
3.2.4 The state of emergency declaration and spatially weighted of them
The state of emergency declaration was issued only for the first, third, fourth, and fifth waves. In the first wave, the declaration of a state of emergency significantly reduced human mobility, although the CI is quite large, with a 3.59 pp WoW decrease in the percentage deviation from the baseline of human mobility (Est. = -3.590, s.e. = 1.580, 95%CI = [-6.687 to -0.493], P = 0.023). In the third wave, although significant, the magnitude is low (Est. = -0.318, s.e. =0.107, 95%CI = [-0.528 to -0.108], P = 0.003). This low magnitude probably is the habituation. However, another ample reason for this low magnitude is that when emergency declarations were issued in prefectures, the infection had entered a phase of decline, so human mobility had begun to recover (see Fig. 1). In the fourth wave, the magnitude is significant and large, 1.34 pp WoW decrease (Est. = -1.340, s.e. = 0.459 95%CI = [-2.240 to -0.440], P = 0.003). In the fifth wave, the magnitude is slightly lower but significant, with a 0.66 pp WoW decrease (Est. = -0.657, s.e. = 0.141 95%CI = [-0.933 to -0.381], P < 0.001). Again, a habituation trend could be observed here. The state of emergency declaration is not significant in any of the waves for the spatially weighted variable, possibly because the state of emergency declaration was valid only in one’s prefecture.
3.2.5 Vaccination
Regarding vaccination, positive estimates will be expected because the more the vaccination rate increased from the previous week within their own prefecture, the more secure people felt. As a result, in the case of 1st dose of vaccination, only lag 1 and lag 3 are significant with a subtle impact, while the other lags are insignificant. The effect is more apparent for the 2nd dose than 1st dose, which is significant for all lag days. A 1 pp increase in the previous week’s difference in vaccination rate brings at most about 0.21 pp WoW increase in percentage deviation from the baseline of human mobility, at lag 3 (Est. = 0.210, s.e. = 0.058, 95%CI = [0.096 to 0.324], P < 0.001). The smallest case was an increase of 0.14 pp in lag 7 (Est. = 0.142, s.e. = 0.053, 95%CI = [0.039 to 0.245], P = 0.007). In 2021, when vaccination began, information was brought to the public that two doses of the vaccine were effective [58]. As more people started to receive the second dose, the population felt reassured, which may have led to a recovery in human mobility. However, all are not significant for the 3rd dose.
3.2.6 Controls
For Controls, each day and holiday dummy are significant, and precipitation is negative and significant, so more precipitation reduced human mobility.
3.3 Residential time spent response
The “Residential” category shows a change in time spent. According to Google, given that people are already spending a large portion of their day at their residences, the residential time spent change is not great even on workdays compared to the human mobility response [64]. As in Fig.4, the responses are subtle; however, the trends are clear.
3.3.1 Infected cases in increasing phase
Excluding the results that are not significant, a 1% increase in the number of infected cases compared to the previous week in the first wave is associated with a maximum of 0.31 pp (lag: 3, Est. = 0.312, s.e. = 0.0582, 95% CI = [0.198 to 0.426], P <0.001) and a minimum of 0.20 pp (lag: 4, Est. = 0.199, s.e. = 0.061, 95% CI = [0.080 to 0.318], P = 0.001) WoW increase from the previous week in the percentage change of residential spent time from the baseline. In the second wave, the confidence interval is quite narrower, and the magnitude of residential spend time increases is low, 0.11 pp at maximum (lag: 2, Est. = 0.112, s.e. = 0.016, 95% CI = [0.080 to 0.144], P <0.001). As in the case of retail & recreation, the impact in the first and second waves decreases with each passing day.
In the third and fourth waves, the effect is even more negligible, with a maximum of 0.09 pp increase in the third wave (lag: 3, Est. = 0.090, s.e. =0.015, 95% CI = [0.060 to 0.119], P <0.001); and in the fourth wave is 0.09 pp increase at maximum (lag: 7, Est. = 0.089, s.e. = 0.036, 95% CI = [0.019 to 0.160], P = 0.013). In the fourth wave, the impact increases as the lag period are longer. For the fifth wave, when the Delta variant was prevalent, the impact increase from lag 1 to lag 4. The maximum is 0.14 pp increase (lag: 4, Est. = 0.138, s.e. = 0.042, 95% CI = [0.056 to 0.220], P = 0.001). Similarly, the sixth wave, when the Omicron variant was prevalent, shows an increase in impact from lag 1 to lag 4, with a maximum of 0.10 pp increase (lag: 7, Est. = 0.097, s.e. = 0.026, 95% CI = [0.047 to 0.147], P < 0.001). A habituation trend is also observed for stay-at-home behaviour. Due to the variants’ prevalence, the fourth to sixth waves show an increasing tendency each lag day has passed, as in the retail & recreation case.
3.3.2 Infected cases in decreasing phase
In the decreasing phase, the time spent in the residences is reduced with a decrease in infection when the estimates are positive and significant. In the first wave, the maximum is 0.14 pp, reducing residential stay time (lag: 4, Est. = 0.144, s.e. = 0.025, 95% CI = [0.095 to 0.193], P < 0.001). However, the impact is smaller than in the increasing phase. As in the retail & recreation phase, people’s behaviour did not entirely return to normal due to fear of COVID-19. In the second wave, the impact is more negligible, with a maximum of 0.09 pp reduction (lag: 2, Est. = 0.087, s.e. = 0.021, 95% CI = [0.047 to 0.127], P < 0.001). At the same time, the third wave is insignificant except for lag 3, and the impact of lag 3 is small. In the fourth wave, the magnitude of the estimates increases slightly, with a maximum increase of 0.11 pp (lag: 2, Est. = 0.111, s.e. = 0.016, 95% CI = [0.081 to 0.141], P < 0.001). The fifth wave is insignificant for all lag days, same as for retail & recreation. The sixth wave, again, as in the case of retail & recreation, has a more considerable impact on lag 1, with an increase of 0.13 pp reduction (lag: 1, Est. = 0.132, s.e. = 0.031, 95% CI = [0.071 to 0.193], P < 0.001).
3.3.3 Spatially weighted infected cases
For spatially weighted infected cases, as with retail & recreation, the impact in the first wave is significant. The maximum impact is 0.35 (lag: 1, Est. = 0.353, s.e. = 0.052, 95% CI = [0.251 to 0.455], P < 0.001), and the minimum 0.20 pp change (lag: 5, Est. = 0.196, s.e. = 0.070, 95% CI = [0.058 to 0.334], P = 0.005) in residential spent time (i.e., spent time increases when other prefectures’ infection increases and spent time decreases when other prefectures’ infection decreases). In the second wave, the impact falls, with a maximum of 0.10 pp (lag: 2, Est. = 0.104, s.e. = 0.026, 95% CI = [0.052 to 0.156], P < 0.001). In the third wave, all but lag 3 are insignificant, and the magnitude of lag 3 is relatively small. In the fourth wave, the estimates are slightly higher, with a maximum of 0.15 pp (lag: 5, Est. = 0.151, s.e. = 0.057, 95% CI = [0.040 to 0.262], P = 0.007). In the meantime, all lag days in the fifth wave are insignificant. The sixth wave, which shows rapid infection by the Omicron variant, again shows the influence of infection information from other prefectures; however, with a little effect with a maximum of 0.06 pp (lag: 2, Est. = 0.058, s.e. = 0.009, 95% CI = [0.040 to 0.076], P < 0.001).
In total, from Fig. 3 and 4, although the magnitude differs, the impact of each wave’s infection information on human mobility is just the opposite for retail & recreation, and residential.
Residential spent time reaction to the COVID-19-related information. Points are estimated coefficients, and bars indicate upper and lower 95% confidence intervals. Infection waves are six, but the state of emergency declaration has been issued only on the first, third, fourth, and fifth waves. Infected cases in increasing phase, infected cases in decreasing phase, spatially weighted cases, and vaccination are taken daily lag from 1 day to 7 days. We conduct the regression analysis seven times, from lag-1-regression to lag-7-regression. The state of emergency declaration, spatially weighted state of emergency declaration, and controls are not taken daily lag; these estimates are from lag-1-regression.
3.2.4 The state of emergency declaration and spatially weighted of them
Although the estimates for the state of the emergency declaration are positive, only the third wave is significant (Est. = 0.180, s.e. = 0.020, 95% CI = [0.142 to 0.218], P < 0.001). On another note, the spatially weighted state of emergency declaration is positive and significant for the first and third waves (first wave: Est. = 0.161, s.e. = 0.035, 95% CI = [0.092 to 0.230], P < 0.001; third wave; Est. = 0.099, s.e. = 0.038, 95% CI = [0.024 to 0.175], P = 0.010). The state of emergency declaration has a relatively small impact on residential, compared to the case of retail & recreation.
3.3.5 Vaccination
For Vaccination, a growth in vaccination rates within one’s own prefecture will have a negative effect on stay-at-home (easier to go out). As a result, only lags 6 and 7 for the 2nd dose are significant. Although the magnitude is relatively small, similar to the case of retail & recreation, 2nd dose of vaccine effectively changes human behaviour.
3.3.6 Controls
As for the control variables, only precipitation is positively significant, with more precipitation making it easier to stay-at-home.
3.4 Robustness
For robustness, the results from the Almon lag model are demonstrated in the Supplementary Information Fig. S3 and 4. From the figures, the results show a similar tendency to our main results.
4 Discussion
Initially, both in retail & recreation mobility and residential spent time, people had sensitive reactions with fear to the increase in the number of infections and complied with the government’s declaration of a state of emergency. As the first three waves progressed, people became gradually habituated to the new infections, and their reactions to the increased number of infections declined. However, people were still cautious about COVID-19; the recovery of their behaviour in decreasing phase was moderate.
Contrarily, from wave 4 to wave 6, the Alpha, Delta, and Omicron variants began to dominate, and the response of people, which had declined until the third wave, started to be sensitive again. In the increasing phase, people’s reactions in retail & recreation mobility and residential spent time tended to intensify around lag day 3 to lag day 5, rather than lag day 1. This tendency may be because people were more cautious about the new variants, which show a further increase in infections over several days than anticipated compared to the previous three waves.
In the fourth wave, during the decreasing phase of infection, human mobility reaction was at the same level as in the increasing phase. Namely, human mobility had recovered to the extent that it decreased during the increasing phase of the fourth wave. There are many possible backgrounds. The fourth wave of the decreasing phase was roughly from April to July 2021 (see Fig.S2). This time was just around when the promotion of prompt vaccination was announced [41], and the infection and treatment information of COVID-19 had been gradually available [35, 36]. These various factors may have also contributed to the recovery of human mobility. By comparison, in the fifth wave, the Delta variant exposed a more severe outbreak, so people were cautious even during the decreasing phase. Thus, outgoing behaviour did not recover.
In contrast, in the sixth wave, when the Omicron variant was prevalent, the response in the lag day 1 was prominent in the decreasing phase. They immediately started to go out after receiving information about declining infections. Despite its rapid infection speed, the Omicron variant is relatively mild [63], which may have eased people’s caution during the decreasing phase.
What we can tell through all six waves is that with or without the state of emergency’s declaration, on average, people respond to infection information of COVID-19. When a similar situation persists, people tend to habituate. However, from a different kind of information (the spread of new variants) people’s habituation come to a halt to some extent. Nevertheless, on the whole, people are still more habituated to the infection information than when they were first exposed to it.
Turning to the declaration of a state of emergency, people reacted substantially to the first declaration in the first wave. However, the third wave, in which prefectures with particularly severe infection situations received a second declaration in early January 2021, did not considerably decrease human mobility. The obscure response is probably due to the habituation to the state of emergency declaration. Another may be because of the decreasing trend in infection cases in many prefectures (see Fig. 1). This decrease in infection is due to the fact that human mobility had already considerably reduced during the New Year’s holidays (before the second declaration), where many people cancelled their trips to their parents’ homes in response to the number of infected people rising at the end of 2020 [65]. Nevertheless, people reacted profoundly to the declarations in the fourth wave (third declaration). The reaction was muted for the fifth wave (fourth declaration). This muted reaction is probably due to habituation to the state of emergency declaration, again.
Timing is critical for policy implementation during a crisis, and the degree of information perception should also be taken into account when considering people’s habituation. While not legally restricted, emergency declarations in Japan have considerably constrained economic activity, particularly in restaurants. Examining its effects in more detail will be necessary for future evidence-based policy invocations.
The effect of vaccination is apparent: completing the second dose worked in the direction of recovery of human mobility. We had received information on the efficacy of the 2nd dose [58]. Moreover, at that time, the Japanese government promoted rapid vaccination [41]. Therefore, a repeated strong message is vital for proceeding with some policies. Effective vaccination program reassures people against COVID-19 which will likely continue in the future. The ability of people to feel secure and to go out is vital for maintaining economic activity during a pandemic.
There are some limitations to this study:
The data, which are macro data aggregated by the prefecture, do not address the heterogeneity of people’s behaviour within a prefecture or each person.
It does not consider policy measures other than emergency declarations, such as pre-emergency measures, which is an issue for future study and for EBPM.
Although each variant’s infection speed or severity is essential information for the infection status and human mobility reaction, we cannot incorporate this information into our model because of the complexity of these transmission mechanisms and people’s reactions to the information. Similarly, it is not possible to disentangle human mobility recovery to a) habituation, b) receiving announcements about the promotion of vaccination, c) receiving information about COVID-19 symptoms and treatment options, and d) any other components in each prefecture (however, time-varying factors of the ‘interactive effects model’ assign different effects for each prefecture in our model, which can capture such unobservable information).
Given the limitations above, our results have the following policy implications. In Japan, declaring a state of emergency required restaurants and others to shorten their business hours or refrain from opening. Although not legally binding, many restaurants have complied. Our results show that people react to infection information of their own and other prefectures, and curb their outing behaviour even when a state of emergency is not declared. Vaccination promotion has also been shown to encourage people to go out to a certain degree. Therefore, although, at the moment (as of July 2022), the government is promoting vaccination and advancing the coexistence with COVID-19 [66], when invoking policies to control human mobility, it is essential to consider: 1) timing and 2) information penetration (consideration of habituation) carefully so as not to constrain economic activities more than necessary.
5 Methods
5.1 Regression model
We use Bai’s ‘interactive effects model [48]’ with the following estimation method. Consider a panel dataset in which {(yit, Xit)}, i = 1, 2, …, N, t = 1, 2, …, T ; the interactive effects model is described as
where Yit is a dependent variable, and Xit is a p × 1 vector of explanatory variables, such that p is the number of explanatory variables and β is a p × 1 vector of the parameters to be estimated. While λi is a d × 1 vector of factor loadings that differ among the unit i, Ft is a d × 1 vector of a common factor that varies over time t, vit is an unobservable term, and eit is the error term. The unobservable term vit has a factor structure λ′iFt and a random part eit. The factor structure allows the model to capture unobserved variations and their loadings.
Applying the interactive effects model, our estimation model will be
where Δyit is the previous week’s difference in percent change from the baseline of human mobility or spent time in residence (in pp), and 
is the IHS difference (from the previous week) of newly infected persons. The variable 
approximates the growth rate of newly infected persons from the previous week. While Eit is the state-of-emergency declaration dummy that takes the value of 1 under emergency declaration, and 0 otherwise. Here, 
is an element of the spatial weight matrix; therefore, the cross-terms 
and 
exhibit the impact of information disclosure on the number of infected people in other prefectures and on the state-of-emergency declaration in other prefectures, respectively. The variable ΔVitv is the increased range of vaccination rate per million persons from the previous week. The index v denotes the number of vaccine doses (from 1st dose to 3rd dose) provided. The term Cit is a control variable, and 
denotes spatial weighted fixed effects. Ftd represents the common factors of dimension d; λid represents the factor loadings with dimension d, where d = 1, 2, …, D stands for the dimension of factors, that is, the number of factors. Lastly, εit is an i.i.d. random component. Parameters α, β1, β2, β3, γ1, γ2, δ, and ϕj are to be estimated. Ft, λ′i, and εit are unobservable. The parameters of our interest are β1, β2, β3, γ1, and γ2.
The estimation of the ‘interactive effects model’ is conducted using ‘phtt’ R library with circumstances of R 3.6.3.
5.2 Spatial-weight matrices
Using V-RESAS’s ‘cross-prefecture travel data in each prefecture,’ spatial-weight matrices, Ws, where s is the week index, are created to apprehend the weekly time-series changes in the movement of people between prefectures. This spatial-weight matrix W for the specific week is standardized using Kelejian and Prucha’s method [67],
where wij is an element of the spatial weight matrix with row i (travel from) and column j (travel to), and where j = 1,2, …, M, and the diagonal element wii is 0. This standardization is conducted through all the sample weeks s; then, we have 
. The time dimension of our estimates is in daily units, while V-RESAS’s cross-prefecture travel data in each prefecture are in weekly units. Therefore, in our estimation, we use 
of weeks s, corresponding to the day t of the analysis. The larger the element of 
, the greater the number of people moving from prefecture i to j.
5.3 The lags
We take lags for the IHS difference (from the previous week) of newly infected persons, 
; its spatial weighted variables 
; and the increased range of vaccination rate per million persons from the previous day of the week, ΔVit. For day t, the daily lag k is taken from one day (k = 1) to seven days (k = 7). The elements of the spatial weight matrix, 
, have a weekly index, s; thus, we take a lag l, which corresponds to the day of our estimation. For example, 25 March 2022, corresponds to the 12th week of 2022. If we take k = 1 lag; subsequently, we get 24 March 2022, which also corresponds to the 12th week of 2022. Therefore, we choose the spatial weighted matrix for the 12th week. Nevertheless, if lag k = 5 is taken from March 25, the observed date would be 20 March 2022, which corresponds to the 11th week of 2022. Therefore, the spatial weighted matrix will be the one observed in the 11th week; thus, lag l = 1. We do not create a cross-term for ΔVit with the spatial weighted matrix since the vaccination status of other prefectures does not largely affect the human mobility in that prefecture.
The lag is taken because the infection situation of a day will not be reported until that day’s evening (around 17:00) or later (e.g. announcement time of web news on the number of infected people in Tokyo on 13 August 2021 by NHK: 23:25, Nikkei: 17:00, Bloomberg: 17:19, Jiji Press News: 22:34, TV Asahi: 18:45, and Tokyo Shinbun: 16:56) [68]. Therefore, most people decide whether they should go out based on the information on the infection status recorded on the previous day and before that. Thus, in our estimation, the maximum number of lag days is set to seven. The same is true for vaccination ΔVit, where people undertake decisions about their behaviour based on the information about increased vaccination rates recorded on previous days. Among the variables of our interest, lags are not applied to the state of emergency declaration, Eit. This is because people’s outgoing behaviour on a given day is affected more if a state of emergency is declared on the same day than if it is declared on the days leading up to the previous day.
Estimates are proposed separately for each lag day; thus, the estimation is conducted seven times. For robustness, the model encompassing all lag orders is estimated. To avoid multicollinearity, the polynomial degree 2 Almon lag model is employed. Further details of the Almon lag model estimation are in the Supplementary Information.
5.4 Common factors and loadings
Common factors, denoted by Ftd, are unobservable elements that vary through the time dimension; they are common for all cross-sectional observable elements: prefectures in our study. However, the loadings of the common factors, Ftd differ across cross-sectional observable elements, which are prefectures; thus, λid describes the difference in the loadings. In a sense, the component 
corresponds to the generalization of unit-fixed effects and time-fixed effects in panel data analyses; this component can better describe complicated and correlated unobservable elements through cross sections and time dimensions than ordinary two-way fixed effects.
There exists considerable but unobservable factors which affect human mobility. For the analysis of COVID-19-related information, it is better to control complicated unobservable factors that vary across time and cross-sectional dimensions. For example, suppose a new variant of COVID-19 is spreading in a country other than Japan. In that case, people in urban areas such as Tokyo and Osaka, which have international airports and a large population concentration, will be more alert to the outbreaks than those in rural areas. Alternatively, when alarmed by the severity and speed of the new variant infection outbreaks, people in the urban areas and rural areas can have difference sense of caution towards the information.
The other example is regarding the government’s policies or announcements. For instance, even if there is an announcement that vaccinations are to be promoted expeditiously by the Japanese government, different prefectures have different systems for promoting vaccinations, so each prefecture’s residents (or, more strictly, each municipality, which is the main body promoting vaccinations) will receive the announcement differently. Alternatively, people may welcome the rapid vaccination program more in prefectures where the infection situation is more difficult than in those where it is less so.
In such cases, the information is either unobservable or difficult to incorporate into the model. Additionally, such information simultaneously affects all prefectures; however, the level of sensitivity differs by prefecture. Thus, the interactive effects model is a better method to control these unobservables. In the first example, the common factors, Ftd, capture the fear of a future epidemic of new variants, while the loading λid captures differences among prefectures in their susceptibility to a future epidemic of new variants.
The dimension d of factors is chosen by consistent estimation, which also deals with the underestimation of true variance by Bai and Ng [69]. For the variance-covariance matrix, we use heteroscedasticity- and autocorrelation-consistent estimators, as proposed by Bai [48].
5.5 Spatial weighted fixed effects
We use the elements of a standardized spatial weight matrix instead of dummy variables, as in the least squares dummy variable (LSDV), to control for spatial interactions. Component 
controls spatial interaction with unobservable δj, accompanied by the observable cross-prefecture travel in each prefecture, 
, which varies over the cross-sectional (prefectural) dimension i and time dimension s. In another expression, this component, representing the spatial weighted fixed effects, captures unobservable spatial interactions that arise from the cross-prefecture travel of each prefectural pair, which varies over time.
Data Availability
The datasets generated during and/or analysed during the current study are available from the author on reasonable request.
Funding Statement
This study did not receive any funding
Conflicts of interest
The author declares no competing interests.
Acknowledgments
The author thanks Takeshi Miura, Yudai Higashi, Dung Luong Anh, and Yasuharu Ukai for their helpful comments and suggestions. Any remaining errors are the author’s own.
References
