JHPS and JHPS-COVID19 data

The JHPS was established in 2014 as a result of the integration of the Keio Household Panel Survey (KHPS), a survey that has been implemented since 2004, and the former JHPS that was introduced in 2009. The first wave of the KHPS was conducted in 2004 and collected information from a sample of 4,005 respondents aged 20 to 69 years. To address the sample attrition problem, the KHPS sample was extended through the recruitment of an additional 1,419 individuals in 2007 and 1,012 individuals in 2012. The first wave of the former JHPS was conducted in 2009 and obtained data from 4,022 respondents aged 20 and over. Due to the similarity of these two surveys, the KHPS and the former JHPS were combined in 2014 and named the “Japan Household Panel Survey (JHPS)”, thereby indicating the adoption of a common questionnaire. After integration, the JHPS received a top-up sample of 2,203 respondents aged 20 and over in 2019. Fig 1 shows the total sample sizes of the JHPS for each survey year.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Sample size of the Japan Household Panel Survey.

Source: [4].

https://doi.org/10.1371/journal.pone.0294189.g001

Aiming to provide data that represent the Japanese population, the survey subjects were selected according to a two-stage stratified random sampling method. Japan was stratified into 24 strata following a regional and municipal classification. The number of subjects in each stratum was allocated in proportion to the registered population according to the Basic Resident Register of the previous year. Next, the districts inside each stratum were selected following a systematic random sampling process, and an average of 10 subjects per district were randomly selected until the predefined total number of subjects per stratum was achieved. Reserve subjects were randomly selected (they had characteristics similar to those of the original subjects they were supposed to replace) and used, when necessary, to replace original subjects who could not be contacted or declined to participate in the survey [4].

Respondents were recruited in January and were informed about the anonymity and confidentiality of their answers. The participants received the questionnaires in their home (in February) after agreeing to join the survey [5]. Informed consent was not obtained, given that the data were analyzed anonymously. In addition, the survey was approved by the Institutional Review Board of the Institute for Economic Studies, Keio University (approval number: 22012R).

The JHPS-COVID19 was a special survey conducted with the respondents of the JHPS. The initial waves targeted the respondents from the JHPS in 2020, while subsequent waves targeted the respondents from the previous JHPS-COVID19. The exception was the third wave that targeted the respondents from the second wave and the respondents of the JHPS in 2021. Table 1 details the survey target, the number of total respondents, the response rates for each survey and the number of respondents in this study.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Number of respondents and response rates for each survey.

https://doi.org/10.1371/journal.pone.0294189.t001

Study variables

y_ijt is a dependent variable and took a value of 1 if the respondent adopted preventive measures, including avoiding traveling, avoiding close and crowded spaces and close-contact settings (the 3Cs in Japan), regular handwashing, reducing contact with people by at least 70%, refraining from attending events and dining with people other than relatives, and teleworking. Respondents were asked whether they performed each prevention measure in April (JHPS-COVID19 first and third waves), September (JHPS-COVID19 second and fourth waves), or January (JHPS2021). They were also asked whether they worked from home for at least one day in the last week of the corresponding month.

D_jt(SOE) or D_jt(quasi-SOE) is a dummy variable that took a value of 1 when the SOE or quasi-SOE was adopted at least one day of the month in which the preventive behavior was reported or at least one day of the week in which telecommuting was reported.

m_jt indicates the natural logarithm of the number of deaths due to COVID-19 in the month or week the dependent variable was considered. To account for the case of 0 deaths, the number of deaths plus 1 was logarithmically transformed. We confirmed that the use of an inverse hyperbolic sine transformation, as in previous studies [1–3], did not significantly change our results.

fear_ijt is a dummy variable that took a value of 1 if the respondent had a fear of infection. In each wave of the JHPS-COVID19 and JHPS2021, the participants were asked “How concerned are you about being infected?” and “How concerned are you about someone in your family being infected?” We determined that the respondents feared COVID-19 infection when the answer was “very concerned” or “somewhat concerned” for at least one of the questions.

x_ijt are the following individual attributes:

Sex: a dummy variable that took a value of 1 if the respondent was male.

Marital status: a dummy variable that took a value of 1 if the respondent was married.

Age category: respondents were separated into those aged 20 to 39 years, 40 to 59 years and those aged 60 years or older.

Education status: a dummy variable that took a value of 1 if the respondent completed tertiary education.

Income category: respondents were separated into quartiles according to their income level.

Work status: a dummy variable that took a value of 1 if the respondent worked.

COVID-19 vaccination status: a dummy variable that took a value of 1 if the respondent received at least two doses of the COVID-19 vaccine.

High grit: a dummy variable that took a value of 1 when the respondent’s Grit Scale score was above the average. The Grit Scale measures the extent to which individuals are able to maintain focus and interest and persevere in achieving long-term goals [6]. It is expected that a person with a high grit score would have a higher probability of adopting preventive behaviors since they would be more committed and capable of maintaining the adoption of behaviors to protect themselves and others.

The nature of Japanese SOEs and quasi-SOEs

The Japanese government can declare an SOE and quasi-SOE at the prefecture level after considering three conditions: the infection situation (such as the number of newly reported cases), medical service system (such as hospital beds available), and surveillance system (such as PCR test situation). An SOE and quasi-SOE allow prefecture governments to take stock of emergency measures to curb the spread of COVID-19, such as refraining from unnecessary and nonurgent going out, reducing the business hours of restaurants and bars, and adopting telework to reduce the movement of the population. However, unlike countries such as Australia [7], the UK [8] and the US [9], a specificity of the Japanese SOE and quasi-SOE is the fact that only recommendations and not mandates can be adopted. A clear example of this difference is the fact that Japan did not adopt lockdown policies and recommended only that businesses close and the population stay at home. Interactive maps from the Oxford COVID-19 Government Response Tracker [10] illustrate this difference, revealing that during the pandemic, Japan was one of the few countries that adopted only recommendations.

Measures are more restrictive and stringent in an SOE than a quasi-SOE. For example, during an SOE, the government can recommend the reduction of business hours or business closures, while during a quasi-SOE, the former is allowed, but the latter is not possible [11, 12]. Another important difference is that SOEs are declared for a whole prefecture, but for quasi-SOEs, the prefectural governor can geographically limit this state to specific municipalities if necessary [11–13].

In February 2021, the government revised the law allowing the imposition of fines up to 300,000 yen for violators of reduced bar and restaurant business hours during the SOE and fines up to 200,000 yen during the quasi-SOE [11, 12]. However, there is no evidence that businesses were punished with such fines, despite reports of businesses that violated the law. Instead of relying on fines, the Japanese government adopted positive reinforcement by offering compliant businesses daily rewards of up to 200,000 yen.

According to the Japanese government definition, a quasi-SOE can be declared when the third stage (rapid increase in the number of infected people) in a four-stage classification of the prevalence of infection is achieved. If the number of infections continues to increase, advancing to the fourth and highest stage defined as an “explosive infection spread”, then the government is able to declare an SOE. Although decisions were made on a case-by-case basis, the government defined shortages of hospital beds (hospital bed occupancy rate of 20% and hospital admission rate of 40%) and weekly new infections (15 people per 100 thousand) as criteria to determine if a prefecture entered the third stage. The criteria to determine if a prefecture entered the fourth stage were also shortages of hospital beds (hospital bed occupancy rate of 50% and hospital admission rate of 25%) and weekly new infections (25 people per 100 thousand) [11].

The evolution of the pandemic

Fig 2 illustrates the pandemic situation in Japan during 2020 and 2021. The number of new infections indicated that during the study period, there were five waves of COVID-19. The first cases of infection occurred in April 2020, and the highest peak of infections occurred in September 2021. The number of deaths from COVID-19 also indicated the existence of five waves, with the highest death rate occurring after the peak of the 4^th wave of infections. Coincidentally, the surveys were fielded in periods after infection peaks. The fourth wave was the exception, with the peak of infections occurring during the period the survey was fielded.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Number of new infections and deaths from COVID-19.

Source: Author calculations based on [14].

https://doi.org/10.1371/journal.pone.0294189.g002

Statistical analysis

We estimated the equation as a random-effects or fixed-effects linear probability model. If the coefficients αs (intervention effects) were significantly positive, intervention policies promoted the adoption of preventive behaviors. Similarly, if βs (information effects) were positive, exposure to COVID-19-related information (or the fear induced by the available information) encouraged people to adopt infection prevention actions. We analyzed the data using STATA/SE version 14.2 for Windows (StataCorp, College Station, TX, United States).

Demographic characteristics and prevention measures

The urgency of understanding the effects of the unprecedented pandemic and the measures adopted by governments worldwide resulted in a myriad of studies focusing on topics ranging from socioeconomic and behavioral impacts to physical and mental health issues. One of the main issues of these studies, especially those related to prevention behaviors, is the availability of appropriate data to conduct the investigations, with many depending on cross-sectional data and rapid internet surveys with quota samples or convenience samples (people who did not have access to the internet were not able to participate in the survey). In other words, they depended on evidence based on correlations from cross-sectional data that did not necessarily represent the country population and lacked long-term behavioral follow-up [15]. Aiming to overcome these issues, we employed a large and high-quality longitudinal survey that, derived from probability samples, represented the Japanese adult population.

The characteristics of the sample are summarized in Table 2. The sex and age group distributions of the sample were similar to those of the Japanese population aged 21 years or older in 2020: 48.07% were males, and 51.93% were females; and 23.86% of participants were aged 20 to 39 years, 33.71% were aged 40 to 59 years, and 42.43% were aged 60 years or older according to the 2020 Population Census (Statistical Bureau [16]). Despite the existence of missing values (approximately 2,100 cases), the sample’s average monthly household income was still representative (approximately 491 thousand yen), being slightly higher than the average of the Japanese population. According to the Comprehensive Survey of Living Conditions (Ministry of Health, Labour and Welfare), the average household income was 5,643 thousand yen in 2020 or approximately 470 thousand yen per month [17]. The share of respondents with a tertiary education level was slightly greater than that of the Japanese population, possibly given the existence of missing values for this variable (approximately 1,650 cases). According to the 2020 Population Census, 41.7% of Japanese individuals with some level of education and aged 15 years or older had a tertiary education level [18]. In contrast, the share of respondents from Tokyo and the three neighboring prefectures (Chiba, Kanagawa, and Saitama) was also similar to that of the Japanese population aged 21 years or older, approximately 30% [16].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Demographic characteristics of the respondents.

https://doi.org/10.1371/journal.pone.0294189.t002

Regarding prevention behaviors, in more than 90% of the observations, respondents avoided events and dining out, close and crowded spaces and close-contact settings, and traveling. Approximately 88% of the respondents regularly washed their hands, while 80% reduced their contact with people. The percentage of cases in which telework was adopted on specified dates was approximately 17.5%.

Empirical results

The estimation results based on the random-effects model are summarized in Table 3. The dependent variable in columns (1)-(4) is a dummy that took a value of 1 if all five prevention measures were adopted, while it was a telework dummy in columns (5)-(8). Columns (1)-(4) show that both intervention effects (SOE and quasi-SOE) and information effects (number of deaths and fear of infection) were positive and statistically significant. As the addition of information effect variables resulted in only a small decrease in the coefficients of intervention effect variables, we focused on the full model in column (4). The implementation of an SOE increased the probability of adopting all five preventive measures by 6%, while the implementation of a quasi-SOE increased the probability by 7.1%, implying that the impact of an SOE and quasi-SOE was basically the same (the F test could not reject the null hypothesis that the coefficients of the SOE and quasi-SOE were equal), although a quasi-SOE imposes fewer restrictions on movement. In Japan, if a quasi-SOE was not enough to control the outbreak, an SOE declaration with stronger restrictions on movement was issued. Thus, it is possible to interpret that a quasi-SOE could have served as a threat of a stricter SOE and promoted people’s behavioral changes to take more prevention measures. This may imply that a phased intervention method with quasi-SOEs and SOEs as a policy response to the pandemic is effective.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Estimation of intervention and information effects: Random-effects model.

https://doi.org/10.1371/journal.pone.0294189.t003

Regarding the information effect, the estimated coefficients showed that a 1% increase in the number of deaths increased the percentage of people who adopted all five prevention behaviors by 0.036%. If we calculated the substitution rate of the SOE and deaths, the imposition of the SOE corresponded to an increase in the number of deaths by approximately 167% (6%÷0.036%) or approximately 1.7 times. Additionally, fear of infection led to an increase of 6.4% in the adoption of all preventive behaviors.

Following the previous literature [2], we decomposed the contributions of an SOE and quasi-SOE, number of deaths, fear, and other factors to the adoption rate of all preventive behaviors based on the coefficients in column (4). Fig 3 shows the decomposition results when we used each variable’s average values for Tokyo residents. The SOE contribution was approximately 5%, deaths 18%, fear 5%, and other reasons 46%, which indicates that the information effect was approximately 23%, while the intervention effect was approximately one-fifth (5%). If we suppose that other factors represented the contribution of usual or “new normal” prevention habits (even with no change in SOE status, deaths, or fear), they could be included in the information effect. In this case, the information effect totaled 69%, with the intervention effect becoming less than one tenth of the information effect. These results are similar to previous findings based on prefecture-level data [2, 3] in the sense that information effects have a greater impact on people’s behavior. To check for possible endogeneity bias in the model, we also used lagged data for the SOE, quasi-SOE, and number of deaths. The results were similar to the ones presented, so they were omitted for simplicity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Decomposition of the contribution of intervention and information effects: The case of Tokyo.

Note: The decomposition of the contribution of intervention and information effects was based on the estimations from column 4 of Table 3.

https://doi.org/10.1371/journal.pone.0294189.g003

Given that the Delta variant had high transmissibility, we checked whether the spread of the Delta variant impacted the estimation results. As the Delta variant was first reported in Japan in March 2021 [19], we considered that the influence of the Delta variant was captured in the 3^rd and 4^th JHPS-COVID19 surveys. Therefore, as a robustness check exercise, we estimated the model without the data of the 3^rd and 4^th surveys (an estimation without the data of the 4^th survey, a period when the Delta variant was dominant, was also conducted). The estimation results were similar to those presented in Table 3, indicating no or very small influence of the Delta variant.

For other factors, the estimation result showed that people with the following characteristics tended to adopt preventive behaviors: female sex [20–24], married, 60 years or older [23, 24], people who did not work, and perseverant. In contrast, variables such as income and vaccination status were not statistically significant.

Next, columns (5)-(8) show that the declaration of an SOE increased the probability of adopting telework by 7.6 to 8.4%, while the coefficients of the quasi-SOE were not significant. These results indicate that a decision to conduct telework largely depended on an SOE declaration. Fear was not significant, while the number of deaths was negative and significant. Although it was difficult to interpret why the death coefficient was negative, its magnitude was extremely small (a 1% increase in the number of deaths results in a 0.014% decrease in the telework rate), being close to zero. Moreover, such small information effects are understandable if we suppose that a decision to implement telework depended more on companies than on individual workers.

Individual attributes indicated that the probability of implementing telework was significantly higher among males, individuals with a tertiary education level, high-income earners, and perseverant individuals, while vaccination status was significantly negative, showing that after the vaccination series was completed, there was a tendency to return to commuting.

Even in cases in which teleworking was implemented, its frequency could be different. To account for this, we conducted a robustness check analysis by replacing a telework dummy with the weekly number of days worked from home, which is available in each survey. We found no major changes from the estimation results in Table 3.

Furthermore, the availability of telework could change depending on the infection status and other factors. To account for this possibility, we controlled for the number of deaths due to COVID-19 in Table 3. It is also likely that the availability of telework increased as time passed after the outbreak of the pandemic because firms and workers would be able to prepare the equipment and network necessary for telework. To observe a difference by time, we conducted a robustness check analysis where we estimated the basic model without the 3^rd and 4^th survey data. We found no major changes from the original estimation results in Table 3.

Table 4 presents the estimation results for each preventive behavior based on the fixed-effects model. Columns (1) and (7) present similar results even after controlling for time-invariant region- and individual-specific characteristics. The exception is that fear coefficients were basically not significant. Coefficients from columns (2)-(6) indicate a difference between an SOE and quasi-SOE, with the latter being relatively smaller for avoiding traveling and not statistically significant for the 3Cs and avoiding events and dining with others. In contrast, the quasi-SOE coefficient for reducing contact with people was almost three times larger than that of the SOE, with the F test confirming that their difference was statistically significant. Thus, we can understand that this behavior’s contribution made significantly positive coefficients of the quasi-SOE in column (1) of Tables 3 and 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Estimation of intervention and information effects: Fixed-effects model.

https://doi.org/10.1371/journal.pone.0294189.t004

Interestingly, the SOE had a positive effect even on the likelihood of handwashing. The SOE had a direct effect on traveling, avoiding the 3Cs, reducing contact with people, and refraining from attending public events and dining out, such as prohibiting large-scale events and reducing restaurant and bar business hours. However, it was impossible for the government to directly intervene in the behavior of people’s handwashing. Thus, the positive effect of the SOE in column (4) should have reflected just the indirect announcement effect.

Regarding the information effect, an increase in the number of deaths increased the chances of avoiding the 3Cs, having contact with other people and joining events and dining out, while fearing infection increased the chances of avoiding travel.

We also investigated how the adoption of all preventive behaviors and telework in heterogeneous groups were affected by intervention and information effects. First, Table 5 indicates that the information effect was similar, while the intervention effect was noticeably larger for younger individuals, which is consistent with the previous literature [3]. From a geographical perspective, the intervention effect was positive and significantly larger in the region around the capital (Chiba, Kanagawa, Saitama, and Tokyo), where the population and number of infected people were larger, while the information effect was positive and significant only in the rest of the country. Regarding work status, the intervention effect of the SOE was greater for people who did not work, whereas the information effect of the number of deaths was greater for people who worked. Regarding firm size, the intervention effect was significantly greater for large firms and government agencies than for medium and small firms.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Estimation of intervention and information effects by individual attributes: Fixed-effects model.

https://doi.org/10.1371/journal.pone.0294189.t005

For teleworking, the SOE coefficient was higher among young people, residents of Tokyo and its three neighboring prefectures, regular employees, and those in large firms and government agencies. Although the coefficients for the number of deaths were negative and statistically significant, as mentioned above, the coefficients were extremely small, close to zero.