Variations in Early-Stage Responses to Pandemics: Survey Evidence from the COVID-19 Pandemic in Japan.

During the initial phase of pandemics, swift behavioral responses by individuals, such as social distancing, can temper the speed and magnitude of further infections. However, individual choices in this period are often made in the absence of reliable knowledge and coordinated policy interventions, producing variation in protective behaviors that cannot be easily deduced from that in later periods. Using unique monthly panel survey data, we examine variations in the association between changes in infections and risky behavior, particularly the frequencies of face-to-face conversations and dining out, between January to March 2020. We find that the increase in confirmed cases is negatively associated with the likelihood of these behaviors. However, high school graduates are less responsive than university graduates. We provide evidence that this can be attributed to their lower perception of infection risk, while we cannot fully rule out the roles of income opportunity costs. These results point to the benefits of interventions incorporating nudges to raise individuals' risk perceptions during the initial phase of pandemics. We also discuss the potential efficacy of such interventions in later periods of pandemics. Supplementary Information: The online version contains supplementary material available at 10.1007/s41885-021-00103-5.

Introduction

Infectious diseases are one of the leading causes of deaths in the world. Even in the twenty-first century, novel viruses, such as H1N1, SARS, and COVID-19, have posed a great threat to human lives and society. To mitigate the infection spread, it is essential for individuals to avoid risky behavior and maintain appropriate social or physical distance from one another (Fenichel 2013; Fenichel et al. 2011; Institute of Medicine 2007; Ipsen 1959).1 However, it can be difficult to achieve sufficient levels of distancing, because of attendant economic costs, free-riding behavior, and uncertainties about transmission risk. Earlier studies show that individuals’ behavioral responses vary across their demographic, socio-economic, and psychological characteristics (Bish and Michie 2010). Evidence from the COVID-19 pandemic in 2020 has also demonstrated similar patterns (Barrios et al. 2021; Brodeur et al. 2021; Dasgupta et al. 2020;).

While insightful, most earlier studies have left an important issue unaddressed. Individuals’ behavioral response to the initial phase of a pandemic may differ from that of later periods, due to changes in knowledge about the virus, policy interventions, and socio-economic conditions. However, how they respond in the initial phase of the infection spread is not well understood. This is crucial for two reasons. First, prompt, appropriate initial responses are essential to minimizing later infection spreads. Second, it is even more costly for people to adopt—and justify—protective behavior in early periods, because of poor knowledge about the virus’s risks and effective policy interventions. Therefore, it is relevant to identify obstacles that prevent individuals from taking protective behavior, such as social distancing, in this period.

This study bridges this knowledge gap by examining the case of Japan during the initial phase of the COVID-19 infection spread, prior to the announcement of a state of emergency on April 7th, 2020. Japan is suitable for this exercise, because it was one of the earliest countries to confirm COVID-19 cases outside of China, following Thailand (WHO 2020). Therefore, citizens, as well as the government, suffered from a lack of knowledge about the virus. Furthermore, the Japanese government was less interventionist than other countries, in that it did not restrict residents’ activities or provide financial support. Reverse-transcription polymerase chain reaction (RT-PCR) tests were not made widely available. Rather, the government simply recommended that citizens avoid risky behavior and stay home voluntarily. Exploiting these features, this study analyzes the extent to which increases in infection risk are associated with the prevalence of risky behavior—such as face-to-face conversation and dining outside—between January and March 2020. We also examine differences in the magnitude of association by individuals’ demographic and socio-economic conditions.

Crucially, this study also uncovers obstacles to voluntary compliance with risk-reducing measures, such as economic conditions, poor access to information, low perceptions of transmission risks, and socio-psychological characteristics. Disentangling these obstacles allows us to discuss the interactive roles between individuals’ responses and public policies. For example, if people do not modify their behavior due to the low perception of infection risks, then interventions that elevate risk perceptions should mitigate the spread of COVID-19 effectively, without the need for drastic legal restrictions.

Using original survey data and a fixed effect model, we regress risky behavior on the monthly average of confirmed cases per day in each prefecture—the main unit of subnational government in Japan. Considering the absence of a natural experimental condition, it is difficult to fully rule out the possibilities of reverse causality and sample selection. However, we provide evidence that these biases are unlikely to be severe, and if anything should work against our central hypotheses.

We find that the increase in the number of confirmed cases is associated with decreases in risky behavior, even in the early pandemic period and even without legally enforceable policy interventions. However, the association is weaker among high school graduates than university graduates, implying that exposure to infections may not be equal across individuals. We also provide suggestive evidence that the differences in the perception of infection risk is the most plausible reason for the heterogeneity. These results suggest the importance of interventions that incorporate nudges to heighten perceptions of risk in the early phase of pandemics.

This study is most closely related to Barrios and Hochberg (2020), Machida et al. (2020), and Muto et al. (2020). Using daily panel data at the region level in the U.S., Barrios and Hochberg (2020) find that relative to Republicans, Democrats are more concerned about the infection spread and economic damages and are more likely to avoid risky behavior, given the increase in the confirmed cases. A distinction between this study and theirs is that they do not examine the role of socio-economic status, which is an important predictor of disaster preparedness and resilience. Furthermore, Barrios and Hochberg (2020) analyze risky behavior in the U.S. after the government started to restrict residents’ activities, while we study Japan before the government intervened. The findings of this study are also in line with those of Muto et al. (2020) and Machida et al. (2020), who conducted a survey in Japan as early as or even earlier than this study to examine individuals’ risky behavior. Muto et al. (2020) find a negative correlation between socio-economic status and risky behavior in line with this study, but they do not test the potential reasons for the correlation. Machida et al. (2020) find insignificant association between socio-economic status and behavior. Another distinction is that these studies analyze cross-sectional datasets, while we employ monthly panel data. This enables us to examine individuals’ behavioral changes in response to the infection spread more rigorously.

Our findings are also relevant to understanding individuals’ protective behavior in the absence of policy interventions. Governments enforce social distancing through various interventions, such as closing public transportation and workplaces, making viral or antibody tests widely available, and providing financial subsidies (Hale et al. 2020). Existing studies suggest that these domestic regulations can be an effective tool to control the infection spread (Gatto et al. 2020; Hoeben et al. 2021; Jarvis et al. 2020; Katafuchi et al. 2021).2 However, an obvious concern regarding these legal interventions is their economic consequences, such as increases in the unemployment rate (Acemoglu et al. 2020; Gharehgozli et al. 2020; Inoue and Todo 2020; Mandel and Veetil 2020; Martin et al. 2020). Mandatory social distancing also affects residents’ mental and physical health negatively (Liu et al. 2020; Pfefferbaum and North 2020; Yamamura and Tsustsui 2021) and exacerbates anti-social behavior, including violence and suicide (Dsouza et al. 2020; Mazza et al. 2020). As a result, governments often lift social distancing requirements to restart social and economic activities even before infections are brought under control. These negative side-effects point to the importance of illuminating how governments can cope with infections without relying on costly regulations. One component is uncovering the obstacles to voluntary social-distancing behavior in the absence of legal regulations.

Finally, this study also contributes to the extensive literature on the relationship between risk perception and health behavior. Prior studies have argued that perceptions of health risk play pivotal roles in predicting risky/protective behavior, such as smoking, the purchase of health insurance, and immunization (Brewer et al. 2007; Lin and Sloan 2015; Schaller et al. 2019; Zhou-Richter et al. 2010). The same patterns have been confirmed for protective behavior from infectious diseases (Bennett et al. 2015; Gidengil et al. 2012; Lakdawalla et al. 2006). Since the health impact of these behaviors are scientifically confirmed and widely known, individuals’ risk perception for these behaviors is determined by their knowledge of and trust in scientific research. Hence, not surprisingly, those with higher socio-economic status, particularly with higher educational attainment, are more likely to take protective behavior (Lowcock et al. 2012; Maurer 2009). By contrast, scientific knowledge about COVID-19 was still scarce during the initial phase of the pandemic. Furthermore, unlike other infectious diseases such as SARS and H1N1, COVID-19 has distinctive features, including a high proportion of asymptomatic infections, limited capacities to conduct RT-PCR tests, and frequent mutations of the virus. Consequently, both scientifically confirmed and unconfirmed information about COVID-19 spread on social media (Cato et al. 2021b). These features could cause the perceived risk of COVID-19 to vary even among those with similar educational backgrounds. Therefore, it is informative to confirm that even under these conditions, those with higher educational attainment have both higher risk perceptions and are more likely to avoid risky behavior.

The rest of this study is organized as follows: Section 2 summarizes the infection spread and government responses in Japan. Sections 3 and 4 describe the dataset and identification strategy, respectively. Section 5 presents the results. Section 6 disentangles the obstacles to avoiding risky behavior, and finally Section 7 concludes.

Background

Infection Spread of COVID-19 in Japan (January to March 24th, 2020)

Japan was one of the earliest countries to confirm COVID-19 cases outside of China, following Thailand (WHO 2020). The first case in Japan was confirmed on January 15th, 2020 in Kanagawa, a region in the suburb of Tokyo, and 15 more cases were reported by the end of January (Fig. 1). Most of these cases (13 out of 16) were attributed to visitors and returnees from China. The first report of human-to-human transmission, however, appeared in January 28th in Nara, a tourist site in western Japan.

Fig. 1

Infection spread in Japan

In February, the virus gradually and silently spread in several rural prefectures in addition to large cities. By the end of February, a total of 239 cases were reported. However, more than half of the 47 prefectures had not yet confirmed any cases, and even populous prefectures, such as Miyagi and Osaka, had found only a few cases Fig. 2 and 3.

Fig. 2

Infection spread in Japan

Fig. 3

Infection spread in Japan

Infection spread accelerated in March. More populated prefectures started to find new cases regularly, and over 10 prefectures announced their first cases in the first week of March. While about 30 cases were found nationwide each day until the 9th, a big jump occurred on the 10th, when 70 cases were reported. Around the same time, fatalities from COVID-19 started being reported regularly.

Government Response and Economic Consequences

Despite the confirmation of infected citizens earlier than in most countries, the Japanese government’s response was comparatively passive. It gradually tightened immigration controls for visitors from Hubei Province, China, and also asked Japanese residents in Wuhan to return to Japan in the beginning of February. However, in stark contrast to other countries that closed public transportation and workplaces, there was no legal regulation of residents’ activities in Japan. In fact, as late as early April, the prime minister emphasized that there was no need to declare a state of emergency and only requested self-restraint (Jishuku Yosei) in hosting or attending large-scale public events. The one exception was on February 27th, when the national government requested the closures of all elementary, junior, and senior high schools until the beginning of the new academic year in April. However, the final decision was left to the governor of each prefecture, and some prefectures did not close their schools. No restrictions were placed on economic activities .

While the national government was cautious about declaring a state of emergency, several local governments initiated measures of their own. That said, these were also limited in the scope and time frame of regulated activities and, more importantly, lacked legal enforcement. On February 28th, the Governor of Hokkaido announced a state of emergency, although it had no legal force, and requested that residents avoid leaving their homes for three weeks. The Governor of Osaka also asked for the refrainment of movement to and from Hyogo, the neighboring prefecture, between March 20th and 22nd.

The low number of RT-PCR tests in Japan is also striking.3 There were two paths for Japanese residents to be tested as of March 2020. First, those who had “close contact” with an infected person were requested to visit a designated medical facility.4 Second, those who did not have close contact but suffered from severe symptoms could consult with their family doctor or local public health call center, who would then refer the patient to a designated facility, if considered necessary. Only those persons whom the facility suspected were infected could take a RT-PCR test, which was administered at public health centers or local public health institutions. Therefore, there was no way to detect asymptomatic infection except for those who had “close contact”. The accuracy of detecting infected people also depended on the screening ability of home doctors, call centers, and designated medical facilities.

Because of these passive policy interventions, economic conditions in Japan did not decline as much as in other countries during the first quarter of 2020. Although the number of bankruptcies increased from 651 cases in February to 740 in March, as shown in Fig. A1, only 12 cases were related to COVID-19 (Tokyo Shoko Research 2020). The unemployment rate was also stable between January and March, in contrast to other countries experiencing a rapid increase in infections, such as the U.S. and Ireland (Fig. A2).

Data

This study employs two datasets. First, to approximate the risk of COVID-19 infection, we construct prefecture-level monthly panel data on the average number of daily confirmed cases between December 2019 and March 2020 (4 periods × 47 prefectures). We use this information as the main independent variable. Because the number of newly confirmed cases is reported daily by the government and mass media, it is the most easily accessible information for people regarding the infection spread.

Second, this study uses data from an original, nationwide online panel survey. We discuss the survey design in detail in Online Appendix A1. Our survey targeted those in their 30s and 40s, given that working-age individuals account for a high proportion of confirmed cases compared to the elderly and teenagers. While the behavior of the elderly, who are susceptible to COVID-19, is undeniably important, it is difficult to collect a representative sample of older generations due to disparities in internet access and low likelihood of owning smartphones (Ministry of Internal Affair and Communication 2018 p156).

The first round of the survey was conducted between March 25th and 27th, 2020. We conducted quota sampling with regard to gender (two categories), age group (four 5-year categories), and location of residence (10 categories) among those who registered with Rakuten Insight, a survey company in Japan, so that the distribution of these characteristics was comparable to that of the Japanese population. Table A1 presents the summary statistics of prefecture and respondent characteristics. The distribution of age, gender, and occupation is comparable with the population, supporting the national representativeness of our data. However, our dataset may oversample those with higher socio-economic status (Online Appendix A1).

The first-round survey data contain two behavioral variables related to risky behavior, our outcome of interest. First, we asked respondents to retrospectively report their frequency of face-to-face conversation on a typical day in each month between December 2019 and March 2020 (4 periods). Second, we also asked for the frequency of dining out for dinner in a typical week in each month between January and March 2020 (3 periods), based on recall. To mitigate potential concerns about recall bias, we asked respondents to choose an answer from a list of frequency intervals which included the option, “do not want to answer.” Specifically, for conversations, we asked the following question for each month: “On a typical day, with how many people do you have face-to-face conversation in your daily life and job?” The answer options included: (1) Rarely, (2) 1 to 2, (3) 3 to 5, (4) 5 to 10, (5) 11 or more, (6) do not want to answer. For eating out, we asked: “On a typical week, how often do you dine out for dinner per week?” The answer options included: (1) Rarely, (2) 1 to 3, (3) 4 to 6, (4) everyday, (5) do not want to answer. After dropping the sample of Hokkaido prefecture, 2624 respondents answered these questions. From this information, a monthly panel dataset was compiled.5

On April 27th to May 7th, we re-surveyed the same respondents to collect further information on their social and psychological traits, such as civic attitudes and social capital, although we use these only in Section 6. A total of 2293 individuals participated in both surveys, but the sample size in Section 6 becomes even smaller because of missing values.

Table 1 presents the proportion of those who had face-to-face conversations with five people or more (roughly around the median) on a typical day, and the proportion of those who had dinner outside at least once a typical week. It shows reductions of risky behavior over time, particularly in March, but the changes are small, likely due to two reasons. First, there was no legal regulation of residents’ activities in Japan. Therefore, many Japanese firms did not take actions to encourage social distancing among employees, such as remote work, at that time (Okubo and NIRA 2020). Second, people were not yet aware of the severity of virus, given the scarcity of scientifically confirmed information.

Table 1

Changes in risky behavior

	Obs.	Mean
Face-to-face conversation (1 if talked with five people or more on a typical day)
December 2019	2609	0.535
January 2020	2609	0.537
February 2020	2609	0.515
March 2020	2612	0.486
Dining out (1 if had dinner outside at least once a typical week)
January 2020	2620	0.484
February 2020	2616	0.461
March 2020	2619	0.429

Identification Strategy

This study estimates the following fixed effect model:where, R denotes the binary indicators of risky behavior of individual i in prefecture p in month t, and is the mean of R over the periods. For face-to-face conversations, R takes unity if the individual talks with five people or more per day, and zero otherwise. For dining outside, it takes unity if the individual undertakes the activity at least once a week. X includes predetermined respondent characteristics, such as age, gender, and educational attainment, which we use as interaction terms. Inf denotes the monthly average of newly confirmed cases per day in the prefecture in which the respondent resides. AdjInf denotes the summation of Inf over the adjacent prefectures, to account for high levels of cross-prefectural movement in urban areas in particular. Damage denotes proxies for the economic damages from the infection spread, such as the number of bankruptcies and the active job-openings-to-applicants ratio. Finally, T denotes monthly fixed effects. The use of fixed effect model controls for those characteristics invariant between January and March 2020, including socio-economic conditions at the prefecture and individual levels. Monthly fixed effects capture the impact of country-level shocks, such as announcement of school closure, news about the infection spread in other countries, and restrictions on overseas travel. In this model, β is the coefficient of interest.

Our identification strategy relies on four assumptions: no reverse causality, parallel trend assumption, limited impact of economic damage and government intervention, and limited spillover effect. We examine their plausibility in Appendix.

Results

Benchmark Results

Before showing the main results of Eq. (1), Table 2 presents the results of models that do not incorporate any interaction terms. It shows that an increase in confirmed cases per prefecture is negatively associated with risky behavior. Furthermore, compared to the naïve models (Columns (1) and (5)), the association becomes even larger after controlling for economic conditions (Columns (2) and (6)). The results are also robust to the additional control for confirmed cases in adjacent prefectures (Columns (3) and (7)) and the exclusion of respondents with a schooling-age child (Columns (4) and (8)). Hence, changes in economic conditions or government interventions cannot explain the significantly negative coefficients of confirmed cases.6 Looking at respondents with no children, Columns (4) and (8) show that a one standard deviation increase in COVID-19 cases (S.D. = 1.9 as of March) is associated with a decrease in the likelihood of talking with more than five people per day and dining out at least once a week by 1.5 and 1.0 percentage points, respectively. We present the robustness of the results in Online Appendix A2.1.

Table 2

The association between infection spread and behavior

	Conversation
Sample:	All	All	All	No child
	(1)	(2)	(3)	(4)
Confirmed cases	−0.007***	−0.008***	−0.009***	−0.008***
	(0.002)	(0.002)	(0.002)	(0.003)
Confirmed cases			0.001	0.002
in adjacent prefectures			(0.002)	(0.002)
Bankruptcy cases		0.371**	0.370**	0.220
		(0.178)	(0.179)	(0.281)
Job-openings- to-applicants ratio		−0.164***	−0.173***	−0.123
		(0.059)	(0.057)	(0.087)
Monthly FE	Yes	Yes	Yes	Yes
Individual FE	Yes	Yes	Yes	Yes
Mean Dep. Var.	0.518	0.518	0.518	0.486
Observations	10,439	10,439	10,439	7299
Obs. at the month-prefecture level	184	184	184	184
R-squared	0.0174	0.0185	0.0186	0.0183
Number of respondents	2613	2613	2613	1827
	Dining
Sample:	All	All	All	No child
	(5)	(6)	(7)	(8)
Confirmed cases	−0.006**	−0.007***	−0.005**	−0.005**
	(0.002)	(0.002)	(0.003)	(0.002)
Confirmed cases			−0.002	−0.003
in adjacent prefectures			(0.002)	(0.002)
Bankruptcy cases		0.457	0.464	0.630*
		(0.391)	(0.372)	(0.334)
Job-openings- to-applicants ratio		−0.009	0.006	−0.032
		(0.084)	(0.087)	(0.095)
Monthly FE	Yes	Yes	Yes	Yes
Individual FE	Yes	Yes	Yes	Yes
Mean Dep. Var.	0.458	0.458	0.458	0.464
Observations	7855	7855	7855	5494
Obs. at the month-prefecture level	138	138	138	138
R-squared	0.0199	0.0202	0.0204	0.0140
Number of respondents	2624	2624	2624	1835

The OLS coefficients are reported. Standard errors clustered at the prefecture level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1

One may be concerned that the estimated coefficients are small in magnitude, but it should be emphasized that we examine behavioral changes in the initial phase of the pandemic when people were not aware of the severity of infection risks. In addition, there was no government intervention to encourage social distancing, and therefore these behavioral changes are fully attributed to individuals’ voluntary decisions. Finally, there was growing social awareness that the number of confirmed cases was not a good proxy for the extent of infection spread, because of asymptomatic transmissions. It is, therefore, valuable to still find significant behavioral changes despite these situations. The small point estimates also suggest the importance of looking further at heterogeneities in sensitivity to the infection spread across respondents.

Main Results: Variations in the Responses to Infection Spread

Does behavioral sensitivity to infection risk vary across individuals according to demographic and socioeconomic characteristics? We address this question by estimating Eq. (1). Table 3 demonstrates significant differences by educational attainment, particularly for the frequency of conversations. Columns (1) and (2) suggest that the impact of a one standard deviation increase in confirmed cases for university graduates is larger by 1.7 and 2.5 percentage points, respectively, than for high school graduates. The results are robust to controls for month-prefecture fixed effects (Columns (3) and (4)) and the exclusion of interaction terms with characteristics other than education levels (Columns (5) and (6)). Therefore, our result is unlikely to be driven by unobserved heterogeneity at the individual and prefectural levels. The results for dining out are qualitatively the same, while the coefficients become statistically insignificant in the even-numbered columns, where we exclude respondents with children. We address the possibility of ceiling effects in Online Appendix A2.2.

Table 3

Heterogeneous effect

	Conversation
Sample:	All	No child	All	No child	All	No child
	(1)	(2)	(3)	(4)	(5)	(6)
Confirmed cases	0.005	0.006
	(0.016)	(0.017)
Confirmed cases	−0.009*	−0.013***	−0.009*	−0.013***	−0.009	−0.013***
x University	(0.005)	(0.003)	(0.005)	(0.003)	(0.005)	(0.004)
Confirmed cases	−0.009*	−0.015**	−0.009*	−0.015**	−0.011***	−0.017***
x Vocational	(0.005)	(0.007)	(0.005)	(0.007)	(0.004)	(0.006)
Confirmed cases	−0.004	−0.003	−0.003	0.002
x Age	(0.040)	(0.042)	(0.040)	(0.041)
Confirmed cases	−0.007*	−0.005	−0.007*	−0.005
x Female	(0.004)	(0.005)	(0.004)	(0.005)
Confirmed cases	−0.005		−0.004
x Live with schooling-age child	(0.005)		(0.005)
Monthly Fixed Effect	Yes	Yes	No	No	No	No
Month-Prefecture Fixed Effect	No	No	Yes	Yes	Yes	Yes
Individual Fixed Effect	Yes	Yes	Yes	Yes	Yes	Yes
Other prefecture characteristics	Yes	Yes	No	No	No	No
Observations	10,192	7203	10,192	7203	10,339	7231
Obs. at the month-prefecture level	184	184	184	184	184	184
R-squared	0.0198	0.0194	0.0331	0.0412	0.0322	0.0415
Number of respondents	2551	1803	2551	1803	2588	1810
	Dining
Sample:	All	No child	All	No child	All	No child
	(7)	(8)	(9)	(10)	(11)	(12)
Confirmed cases	0.004	0.007
	(0.014)	(0.024)
Confirmed cases	−0.011**	−0.003	−0.011**	−0.003	−0.010**	−0.003
x University	(0.005)	(0.005)	(0.005)	(0.005)	(0.005)	(0.004)
Confirmed cases	−0.008**	0.001	−0.007**	0.001	−0.012***	−0.002
x Vocational	(0.003)	(0.005)	(0.003)	(0.005)	(0.003)	(0.003)
Confirmed cases	0.027	−0.012	0.023	−0.013
x Age	(0.029)	(0.055)	(0.029)	(0.055)
Confirmed cases	−0.016***	−0.010*	−0.017***	−0.011*
x Female	(0.004)	(0.005)	(0.004)	(0.006)
Confirmed cases	−0.015*		−0.015*
x Live with schooling-age child	(0.008)		(0.007)
Monthly Fixed Effect	Yes	Yes	No	No	No	No
Month-Prefecture Fixed Effect	No	No	Yes	Yes	Yes	Yes
Individual Fixed Effect	Yes	Yes	Yes	Yes	Yes	Yes
Other prefecture characteristics	Yes	Yes	No	No	No	No
Observations	7665	5422	7665	5422	7777	5443
Obs. at the month-prefecture level	138	138	138	138	138	138
R-squared	0.0272	0.0168	0.0427	0.0406	0.0380	0.0391
Number of respondents	2559	1810	2559	1810	2597	1817

The OLS coefficients are reported. Standard errors clustered at the prefecture level are in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.1

Regarding other characteristics, first, we find a difference in the frequency of dining out by gender. Second, the coefficient of interaction with respondents’ age is statistically insignificant for most columns and small in magnitude. Finally, those with a schooling-age child are less likely to eat out, given the increase in infection risk.

Suggestive Evidence on the Mechanisms of Heterogeneous Impact

Why are less educated individuals less responsive to the infection spread? This section tests eight potential mechanisms relating to differences in their economic circumstances, knowledge and perception about the transmission risk, socio-psychological characteristics, and other factors. Our identification strategy relies on two analyses. First, we test the association between education levels and potential mediating variables in Subsections 6.1 to 6.4. Second, we test the association between the mediating variables and risky behavior in Subsection 6.5. In the main text, we discuss the results of analyses using the sub-sample of respondents who have no children, as child-rearing responsibilities may have complicated interactive effects with the mediating variables. The estimation results based on the full sample are reported in the Appendix.

Economic Conditions

Individuals’ educational attainment may be associated with their occupation and economic status, which may be drivers of heterogeneous effects. We test these channels in this subsection.

Occupational Suitability for Teleworking

High school graduates may engage in a job that is not suitable for teleworking or remote work, such as in retail or the restaurant business. To test this channel we construct an industry-level proxy using the survey results of Okubo and NIRA (2020). Based on an online survey in Japan, Okubo and NIRA (2020) show the proportion of respondents working at home by industry as of March 2020. We combine these proportions and our respondents’ occupation to approximate the suitability of their jobs for teleworking. We then regress this proxy on respondent characteristics to examine whether high school graduates actually engage in jobs unsuitable for telework.

Column (1) of Table 4, however, shows that the coefficient for university graduates is negative among the no-child sub-sample, counter to the hypothesis. The observed patterns do not change in the full sample estimation (Table A11). Since the suitability of working at home may vary even within an industry, our proxy may include measurement errors. However, the measurement errors alone are unlikely to explain the negative correlation between the education level and suitability for remote work.

Table 4

The relationship between education and socio-economic indices (sample with no schooling-age child)

	Suitability of job for teleworking	Economic status	Information access	Risk perception	Risk preference	Social capital	Alternative protective measures
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
University	−0.101***	0.642***	0.194***	0.199***	0.026	0.500***	0.170***
	(0.024)	(0.054)	(0.061)	(0.058)	(0.041)	(0.064)	(0.048)
Vocational	−0.047	0.258***	0.147***	0.117*	0.038	0.390***	0.214***
	(0.035)	(0.066)	(0.053)	(0.060)	(0.070)	(0.081)	(0.056)
Age	−0.001	−0.005**	0.023***	0.004	−0.002	−0.001	−0.012***
	(0.002)	(0.002)	(0.004)	(0.004)	(0.004)	(0.006)	(0.003)
Female	0.157***	−0.427***	−0.124***	−0.073	−0.213***	0.262***	0.387***
	(0.021)	(0.048)	(0.044)	(0.047)	(0.041)	(0.058)	(0.046)
Prefecture FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes
R-squared	0.1069	0.1954	0.0556	0.0362	0.0506	0.0629	0.0769
Observations	1465	1586	1798	1785	1790	1451	1787

The OLS coefficients are reported. Standard errors clustered at the prefecture level are in parentheses. The sample sizes of Columns (1) and (6) are smaller than the others, because the data on respondents’ occupation and social capital were collected in the second-wave survey. Column (2) also has a small sample size due to missing values in the annual income data. *** p < 0.01, ** p < 0.05, * p < 0.1

Economic Status

If the economic status of high school graduates is lower, they may suffer from credit constraints that make the disutility from the income loss caused by staying home larger than for the wealthy. We conduct a polychoric principal component analysis to construct a composite index of economic status from two variables (Kolenikov and Angeles 2004): annual income, and a binary indicator that takes unity for self-employment, executive, or regular employment.7 We examine the correlation between this index and education level in Column (2) of Table 4. It confirms that the economic status of university graduates is significantly higher than that for high school graduates, in line with our hypothesis.

Knowledge and Perception of Transmission Risk

Poor information access and low risk perception are major causes of risky behaviors (Kenkel 1991), suggesting that the maladapted behavioral response of high school graduates may be due to these issues.

Information Access

To test the channel through poor information access, we construct a composite index from three variables: the frequencies of reading paper newspapers, reading newspaper websites, and watching television news. Then, we estimate the association between this index and education level in Column (3) of Table 4. The results are consistent with the hypothesis: university graduates follow mass media more frequently than do high school graduates.

Risk Perception

The Protection Motivation Theory in psychology proposes that a high risk perception—which is attributed to subjective factors such as expectations of infection probability and the severity of symptoms—is essential if individuals are to take protective actions (Rogers and Prentice-Dunn 1997). Risk perception is formed through exposure to information from the media and from their peers, the cognitive ability to process the (numeric) information, and engagement in risky behavior (Ferrer and Klein 2015). When reliable information is scarce and cognitive ability is limited, people suffer from cognitive overload. This causes various cognitive biases in decision-making, including the normalcy bias: the optimistic underestimation of the probability and severity of negative events (Kahneman and Tversky 1972).

There are reasons to think that high school graduates have lower risk perceptions about COVID-19 infections. First, because the actual number of infected individuals is unobservable, people infer the infection probability from the information available, but news related to COVID-19 frequently includes professional, foreign language terms such as “RT-PCR tests”. Processing such information may cause them to suffer from cognitive overload, exacerbating the normalcy bias. Second, while mass media reported the severity of the infection spread, a relatively small number of people were actually confirmed to be infected as of March 2020. Therefore, if high school graduates do not rely on or collect information about COVID-19 from the mass media as carefully as do university graduates, they may assess risks based primarily on their peers’ experiences with infection. This generates a gap in risk perception based on educational attainment.

To test this channel, we construct a composite index of risk perception using the following two questions: how many infected people that respondents think there actually are in Japan as of the survey period; the extent to which COVID-19 will cause serious problems for themselves. The regression result in Column (4) of Table 4 shows that university graduates are more likely to take the infection risk seriously than are high school graduates, supporting our hypothesis.

Socio-Psychological Characteristics

The heterogeneous impact of educational attainment may also reflect variations in individuals’ risk preference and social capital. This subsection tests these possibilities.

Risk Preference

High school graduates may be less likely to take precautionary actions, because they are less risk averse (Anderson and Mellor 2008). Given the difficulty in conducting an economic experiment to elicit the risk preference of respondents in our online survey, we test this channel through two proxy variables. First, we asked the following question: which of the following two sayings characterizes you better, “nothing ventured, nothing gained” or “a wise man never courts danger”? The answer options are in Likert-scale. Second, we also asked the following question: at which precipitation probability do you bring an umbrella when going out? A lower score to these answers indicates greater risk aversion. These questions are frequently used in the literature (Ikeda et al. 2016 p142; Iida 2016) and draws from earlier work in the United States. In Column (5) of Table 4, we estimate the relationship between the composite index of these variables and respondent characteristics. We find that education level is uncorrelated with risk preference.

Social Capital

Social distancing during the COVID-19 pandemic is a public good, and therefore, people have an incentive to freeride (Brodeur et al. 2021; Cato et al. 2020; Cato et al. 2021a). This suggests a channel that university graduates may possess more social capital, and so may care more about their reputation or disapproval from neighbors, causing them to follow societal norms of social distancing. The second wave of our survey asks about respondents’ social capital through six questions on general trust, pure altruism, and social norms. More detail about each question is reported in Table A1. We use these answers to construct a composite index. Column (6) of Table 4 demonstrates that social capital is higher for university graduates than for high school graduates, supporting the hypothesis.

Alternative Interpretations

Alternative Protective Measures

High school graduates may take alternative actions to protect themselves, such as wearing masks and washing their hands with disinfectant. Although our survey does not include items on the use of facemasks or disinfectant soap, it does ask respondents whether they wished to buy them more than usual. We regress the composite index of these variables in Column (7) of Table 4. The result shows that university graduates are more likely to answer affirmatively than high school graduates, counter to the hypothesis.

Less Confidence in Confirmed Cases as a Proxy for Infection Risk

High school graduates may recognize that the number of confirmed cases underestimates the actual infection risk, and therefore, they may be more sensitive to other types of information, such as the ratio of positive RT-PCR tests. However, this hypothesis assumes that those with lower education have more knowledge about COVID-19 than educated respondents. This assumption contradicts our findings that high school graduates spend less time collecting information on COVID-19 than university graduates (Table 4, Column (3)).

Association between Mediating Variables and Risky Behavior

The results so far show that respondents’ education levels are associated with economic status, information access, risk perception, and social capital. To further test whether they are also associated with risky behavior, we additionally control for the interaction terms between these seven indices and the number of confirmed cases, based on the specifications in Table 3.

Table 5 presents robust evidence that in prefectures with more confirmed cases, those with high risk perception are more likely to reduce the frequency of risky behavior. The table also reports False Discovery Rate q-values (Anderson and Mellor 2008) to adjust the p-values of the 14 coefficients of each outcome, confirming a robust association between risk perception and frequency of dining out. Among the other three likely mechanisms, the coefficient for economic status is significantly associated with the frequency of conversations, but it does not predict the frequency of dining out. For robustness, we re-estimate the model by controlling for only the interaction term between confirmed cases and risk perception, in addition to the terms included in Table 3. Table A13 shows that the coefficient of risk perception is still statistically significant and comparable in magnitude with that of Table 5. In Online Appendix A3 we test the validity of this model more carefully, particularly the potential issue of endogeneity of risk perception and multicollinearity.

Table 5

The relationship between socio-economic indices and risky behavior

	Conversation				Dining
Sample:	All		No child		All		No child
	(1)		(2)		(3)		(4)
Confirmed cases	−0.010	[0.828]	−0.017*	[0.319]	−0.010	[0.524]	−0.003	[0.868]
x Suitability of job for teleworking	(0.013)		(0.009)		(0.009)		(0.008)
Confirmed cases	−0.004*	[0.319]	−0.005***	[0.183]	−0.006	[0.364]	−0.003	[0.364]
x Economic status	(0.002)		(0.002)		(0.004)		(0.002)
Confirmed cases	−0.000	[1.000]	−0.002	[0.664]	−0.003	[0.596]	−0.004	[0.596]
x Information access	(0.002)		(0.002)		(0.003)		(0.005)
Confirmed cases	−0.005**	[0.319]	−0.006*	[0.319]	−0.009***	[0.001]	−0.007**	[0.069]
x Risk perception	(0.002)		(0.003)		(0.002)		(0.003)
Confirmed cases	−0.002	[0.828]	−0.001	[1.000]	−0.006***	[0.007]	−0.006**	[0.090]
x Risk preference	(0.002)		(0.002)		(0.002)		(0.002)
Confirmed cases	−0.002	[0.828]	−0.002	[0.828]	−0.001	[0.868]	−0.002	[0.524]
x Social capital	(0.002)		(0.003)		(0.002)		(0.002)
Confirmed cases	−0.001	[1.000]	0.001	[1.000]	−0.000	[0.928]	0.000	[0.928]
x Alternative protective measures	(0.004)		(0.005)		(0.004)		(0.004)
Confirmed cases	−0.009		−0.012***		−0.001		0.006
x University	(0.006)		(0.003)		(0.004)		(0.004)
Confirmed cases	−0.015**		−0.020**		−0.003		0.006
x Vocational	(0.007)		(0.009)		(0.005)		(0.006)
Confirmed cases	−0.000		−0.000		0.000		−0.000
x Age	(0.001)		(0.001)		(0.001)		(0.001)
Confirmed cases	−0.007		−0.003		−0.024***		−0.014*
x Female	(0.007)		(0.006)		(0.005)		(0.008)
Confirmed cases	−0.011**				−0.017
x Live with schooling-age child	(0.005)				(0.011)
Month-Prefecture Fixed Effect	Yes		Yes		Yes		Yes
Individual Fixed Effect	Yes		Yes		Yes		Yes
Observations	6918		4901		5197		3685
Obs. at the month-prefecture level	184		184		138		138
R-squared	0.0491		0.0526		0.0646		0.0657
Number of respondents	1738		1230		1740		1233

The OLS coefficients are reported. Standard errors clustered at the prefecture level are in parentheses. Anderson’s (2008) q-values that adjust the p-values of 14 coefficients in each outcome are in brackets. The sample size is smaller than Table 3, because the data on respondents’ occupation and social capital were collected in the second-wave survey. *** p < 0.01, ** p < 0.05, * p < 0.1

Given these arguments in this section, differences in risk perception are the most likely driver of heterogeneity by education level, although we cannot fully rule out the potential role of income opportunity costs.

Conclusion

Using unique survey data collected in the initial phase of the COVID-19 pandemic in Japan, we find that an increase in the number of confirmed cases is negatively associated with the frequency of face-to-face conversation and dining out. However, high school graduates do not respond as much as do university graduates. We provide suggestive evidence that this heterogeneity is driven primarily by the former’s lower perception for infection risk, although we cannot fully rule out the role of income opportunity costs.

The following policy implications can be derived for the initial phase of pandemics. Our findings suggest that socio-economically vulnerable individuals are exposed to higher infection risk in this period, and thus can become the primary vectors of the virus. This is consistent with the argument of Ahmed et al. (2020). It is, therefore, incumbent upon the government to implement a prompt, targeted intervention for this subpopulation. One approach is for governments to provide information on the risks of infection transmission in an easily accessible and understandable manner to mitigate cognitive overload and normalcy biases. Another promising approach is interventions that incorporate nudges to elevate risk perceptions, as suggested by Van Bavel et al. (2020).

Our findings also have implications for later pandemic periods. Infection risks change with the community diffusion of the virus, as well as mutations of the virus itself. Various information shocks may also influence individuals’ risk perception. For example, governments often lift legal regulations before the infection spread is brought under control in order to restart economic activities. This may provide a wrong message to some citizens and cause them to lower their risk perception excessively. Individuals must interpret such information and update their risk perception in a measured way, but this may be difficult given cognitive overload and risk biases. The importance of voluntary social distancing remains high until herd immunity can be achieved through large-scale vaccinations. However, access to vaccines is still limited, particularly in developing countries with poor national/local governance (Aida and Shoji 2021). Vaccine hesitancy is another issue that can reduce vaccination coverage among the population (Kawata and Nakabayashi 2021; Miyachi et al. 2020). Therefore, the policy interventions suggested above should be effective even in later periods of pandemics.

Finally, potential limitations of this study may include the usage of retrospective data (errors in recall), the relatively low explanatory power of estimation models, and low factor loadings of composite indices, although these are not uncommon in survey-based research. In addition, we should note that our data does not cover those aged over 50 or under 30. Given that their behavioral patterns could differ from our respondents (Shoji et al. 2021), we should be careful in generalizing our findings to other generations.