Global Behaviors, Perceptions, and the Emergence of Social Norms at the Onset of the COVID-19 Pandemic

Abstract

We conducted a large-scale survey covering 58 countries and over 100,000 respondents between late March and early April 2020 to study beliefs and attitudes towards citizens’ and governments’ responses at the onset of the COVID-19 pandemic. Most respondents reported holding normative beliefs in support of COVID-19 containment measures, as well as high rates of adherence to these measures. They also believed that their government and their country’s citizens were not doing enough and underestimated the degree to which others in their country supported strong behavioral and policy responses to the pandemic. Normative beliefs were strongly associated with adherence, as well as beliefs about others’ and the government’s response. Lockdowns were associated with greater optimism about others’ and the government’s response, and improvements in measures of perceived mental well-being; these effects tended to be larger for those with stronger normative beliefs. Our findings highlight how social norms can arise quickly and effectively to support cooperation at a global scale.

1. Introduction

At the onset of the COVID-19 pandemic, individuals were called upon to adopt a variety of individually costly behaviors intended to help reduce the spread of COVID-19 and thereby benefit others and society more broadly. These included washing hands more frequently, wearing face masks, avoiding social gatherings, closing businesses, caring for their children as schools close, and staying at home (Dowd et al., 2020, Hollingsworth et al., 2011). Although media attention often focused on failures to adhere to COVID-19 abatement measures, and some measures were controversial (Webster et al., 2020), a variety of sources suggest that people largely adhered to the measures requested of them (Benítez et al., 2020, Chan, Moon, Savage, Skali, Torgler, Whyte, 2020, Clark et al., 2020, Collis et al., 2021, Hoeben et al., 2021, Pedersen and Favero, 2020), and that their combined efforts meaningfully slowed the spread of the virus (Chernozhukov et al., 2021, Courtemanche et al., 2020, Dave et al., 2021). The early stages of the COVID-19 pandemic therefore provide a striking example of rapid cooperation to overcome a public goods dilemma at a global scale.

In this paper, we investigate whether social norms played a primary role in motivating this widespread and rapid adherence to COVID-19 containment measures. That is, we explore whether this behavioral adherence was accompanied by beliefs that one should adhere, and that adherence was expected by others (Bicchieri, 2005, Bicchieri, 2016, Fehr and Fischbacher, 2004). To do so, we use data from a large-scale international survey of COVID-19 behaviors, beliefs, and attitudes that we developed and distributed at the onset of the pandemic. Participants were recruited globally through online snowball sampling, beginning on March 20, 2020. By April 7, 2020, the last date included in our analyses, the survey garnered over 110,000 participants in 175 countries. It was-to our knowledge-the first large-scale survey of its kind, and offers unique insights into people’s self-reported behaviors, motivations, and experiences at the early and accelerating phase of the COVID-19 pandemic.

Consistent with others’ findings (Barari et al., 2020, Clark et al., 2020, Collis et al., 2021, Pedersen and Favero, 2020), respondents in our survey overwhelmingly reported adhering to a variety of COVID-19 containment measures, including avoiding social gatherings, washing hands more frequently, and staying home from work. Social norms appear to have played an important role in motivating this high level of adherence. In our survey, we asked individuals two sets of questions related to social norms. The first assessed their first-order injunctive beliefs–what they believed people should do–regarding eschewing social gatherings, avoiding handshakes, closing businesses, and stay-at-home orders. These questions were phrased, e.g., “What do you think: should people in your country cancel their participation at social gatherings because of the coronavirus right now?” The second set of questions assessed their second-order injunctive beliefs–what they believed others believed people should do. These questions were phrased e.g., “How many of 100 people in your country do you think believe that participation at social gatherings should be cancelled because of the coronavirus right now?”

We find that participants report having very high first-order injunctive beliefs. For instance, over three-quarters reported believing that businesses should be closed. For other behaviors, injunctive beliefs were even higher. People’s second-order injunctive beliefs were, while lower than first-order beliefs, still high–always over 50%.

If social norms indeed played a role in motivating adherence to COVID-19 containment measures, then we would make the following two predictions (Bicchieri, 2005, Bicchieri, 2016, Fehr and Fischbacher, 2004). First, first-order and second-order beliefs should be correlated with each other. Indeed, we find that they are correlated, $r=.2$. Second, these beliefs should be correlated with behavior. Again, this is what we find. Both sets of beliefs were associated with adherence to containment measures, even in regressions that controlled for both at once. A one standard deviation increase in our first-order and second-order injunctive beliefs scales is associated with a 0.270 and 0.144 standard deviation increase in the likelihood of reporting compliance with at least one COVID-19 containment measure, respectively. These results are robust to controlling for a battery of covariates, including perceptions of one’s own risk from COVID-19. Moreover, the magnitude of the correlations with containment measures was as large or larger for injunctive beliefs as it was for other covariates.

Although respondents report high rates of adherence with containment measures themselves, nearly 60% report being dissatisfied with others’ adherence. They express similar dissatisfaction with their governments’ response to the pandemic: only 9% of respondents report that their government’s response was too extreme. We find that those who had internalized social norms were most dissatisfied. A one standard deviation increase in our first-order injunctive beliefs scale was associated with a 0.078 and 0.101 standard deviation increase in reporting dissatisfaction with others’ response and the government, respectively. Once governments do respond with lockdowns, we find that both first-order and second-order injunctive beliefs increase, though our estimates are noisy and not statistically distinguishable from zero. We also find that dissatisfaction with others’ response and the government response falls, and that it falls most for those with higher first-order injunctive beliefs.

Finally, we investigate the relationship between social norms and government responses on individuals’ reports of their mental well-being. Our survey included both a widely-used depression scale known as the PHQ-8 (Kroenke et al., 2001), and a scale we developed to capture anxieties and worries specific to the COVID-19 pandemic, with items such as, “I am nervous when I think about current circumstances” and “I am worried about my health”. We correlate these scales with dissatisfaction of others’ responses, dissatisfaction with the government response, as well as measures of trust in government, and find that both measures of mental well-being are negatively correlated with dissatisfaction with others’ response, dissatisfaction with government response, and distrust of government. Changes in perceptions of others’ and the government’s response were strongest for those with higher first-order injunctive beliefs.

We expand on this analysis with a detailed case study of the effect of the U.K. government’s lockdown on March 23, 2020, finding that in the days following the lockdown, individuals’ perceptions of others’ response and the government’s response improved, as did their well-being. (Note, however, that in this case study, we do not find evidence of moderation via first-order injunctive beliefs.)

The current research makes three primary contributions. First, we provide evidence that social norms did, in a relatively short span of time, arise all over the world, to motivate what ultimately proved to be a global effort to mitigate the pandemic. Although, as we will shortly discuss, there are many well-documented cases demonstrating the role of norm enforcement in motivating contributions to public goods, none to our knowledge show that such social norms can arise so quickly and at such a large scale.

Second, we contribute to the ongoing debate on the impact of lockdowns on mental well-being. While some have found negative impacts of lockdowns on mental well-being (Brodeur et al., 2021, Giuntella et al., Möhring et al., 2021, Sibley et al., 2020), our paper finds positive effects. This perhaps highlights the importance of considering a counterfactual in which, in the absence of lockdowns, individuals would have been more concerned about COVID-19, at least early in the pandemic.

Our third contribution is the data themselves, which constitute a large-scale survey of attitudes, beliefs, and behaviors, with substantial coverage early in the pandemic. We have made our survey data publicly available at https://osf.io/zgfjc/ and hope that it serves as a public good for other researchers.

The paper proceeds as follows. In Section 2, we discuss the potential role of social norms for encouraging adherence to COVID-19 containment measures. In Section 3, we discuss the methods by which we developed and distributed our survey. In Section 4 we discuss each of the above results in turn. Finally, in Section 5, we discuss the contribution of our research in more detail, as well as key limitations to our findings.

2. Social Norms and Adherence to COVID-19 Containment Measures

In a pandemic, individuals are called upon to take actions like washing hands more frequently, avoiding social gatherings, and adhering to school closures, store closures, and stay-at-home orders (Dowd et al., 2020, Hollingsworth et al., 2011). Such actions benefit the individual somewhat by mitigating the risk from the disease, but, for many individuals–e.g., those at lower risk from the disease, or whose livelihood is at stake–these actions may, on net, be costly. However, such actions have major benefits for society as whole: They help to reduce the infection rate, keep hospitals from being overwhelmed, and reduce mortality rates (Hollingsworth et al., 2011). This situation mirrors that of other large-scale public goods dilemmas which can be combated via costly personal actions. For instance, when it comes to combating climate change, individuals can take individually costly actions like flying less or paying more for electric vehicles and green energy. The reduced carbon emissions from these actions benefit all of us.

Humans are unique in their ability to use social norms to ameliorate public good dilemmas like pandemics and climate change (Akerlof, 1976, Bicchieri, 2005, Bicchieri, 2016, Boyd, 2019, Boyd and Richerson, 2021, Henrich and Muthukrishna, 2021, Ostrom, 1990), using them when, for example, encouraging hygiene (Gauri et al., Munger and Harris, 1989, Nalbone et al., 2005), constructing and maintaining communal infrastructure (Boyd and Richerson, 2021, Ostrom, 1990), encouraging peaceful coexistence (Ellickson, 1986, Leeson, 2007), conserving common resources (Acheson, 1988), going to war (Boyd and Richerson, 2021, Henrich and Muthukrishna, 2021, Leeson, 2007, Mathew and Boyd, 2011), and protecting individual property rights (Ellickson, 1986, Leeson, 2007).

Social norms are diverse, both in theory and in practice (Bicchieri, 2016, Gelfand, 2012, Gelfand et al., 2011, Gelfand et al., 2017, Harrington and Gelfand, 2014, Henrich and Muthukrishna, 2021, Nowak, 2006). At their core, social norms have in common a desired but individually costly behavior like wearing a mask or staying at home, which is expected of individuals, and incentivized via sanctions or rewards (Bicchieri, 2005, Bicchieri, 2016, Boyd, 2019, Fehr and Fischbacher, 2004). Such sanctions or rewards are often meted out by ‘third parties,’ so called because even individuals who are not directly impacted by a norm violation will typically sanction the violation. When punishing is itself costly, this requires that it too is incentivized, for instance via ‘higher-order’ punishment, which is when those who fail to punish when expected to are themselves punished, e.g., via ostracizing (Boyd and Richerson, 2021, Henrich and Muthukrishna, 2021). Norms are sustained in equilibrium so long as interactions amongst individuals involved are repeated and reputations conveyed across interactions with sufficiently high likelihood (see, for instance, Kandori, 1992, Nowak, 2006, Panchanathan and Boyd, 2004), and the behavior is sufficiently observable (Bicchieri et al., 2020, Rand et al., 2014, Yoeli et al., 2013). There are also some additional criteria that help norms be sustained in equilibrium, e.g., due to their coordinated nature, they should be categorical (Hoffman et al., 2020) and unambiguous (Yoeli and Rand, 2020).

Social norms have deeply shaped our psychology (Bicchieri, 2005, Boyd, 2019, Boyd and Richerson, 2021, Henrich and Muthukrishna, 2021). Humans intuitively engage in third-party punishment even at a very young age (McAuliffe et al., 2015), carrying over these intuitions into anonymous laboratory environments (Fehr and Fischbacher, 2004, Henrich and Muthukrishna, 2021). Often, people are not aware of the role that social norms play, and for example report getting “a warm glow,” believing they are “doing the right thing” when engaging in the desired behavior, or feeling “it is wrong” when someone fails to do so. That is, the very preferences or beliefs that lead people to engage in normative behaviors are learned or ‘culturally evolved’ in response to the social rewards and punishments in place due to social norms (Henrich and Muthukrishna, 2021).

Although individuals may not be aware of the role social norms play in motivating their behavior, social norms will often leave tell-tale signs of their role in shaping preferences and beliefs. Psychologists and economists employ a multitude of experimental and survey measures–incentivized and not–for uncovering their role (Fehr and Fischbacher, 2004, Gelfand et al., 2011, Henrich and Muthukrishna, 2021, Jordan et al., 2016, Krupka and Weber, 2013, Paluck and Shepherd, 2012). In this study, we employ relatively straightforward measures of first-order injunctive beliefs and second-order injunctive beliefs. The former is one way of measuring the degree to which individuals have themselves internalized the social norm. The latter gets at the degree to which they view the behavior as expected, and can be a powerful predictor of norm compliance (Jachimowicz et al., 2018).

Because of the importance of social norms in motivating contributions to public goods, many behavioral interventions, or ‘nudges,’ used to promote contributions to the public good can be understood as engaging some aspect of people’s psychology related to social norms (Bicchieri et al., 2020, Prentice and Paluck, 2020, Rand et al., 2014, Williamson and Thulin, 2021). For instance, communicating descriptive norms is a common and often-effective nudge used to encourage behaviors ranging from resource conservation to reduced alcohol consumption (Bicchieri and Dimant, 2019, Borsari and Carey, 2003, Cialdini, 1984, Gelfand and Harrington, 2015, Goldstein et al., 2008, Sheeran et al., 1999), that likely works by communicating that the desired behavior is indeed expected (Elster, 1989, Muldoon et al., 2014, Rand et al., 2014). Of course, social norms are not the only tool available to humans when combating public good dilemmas. We also have more formal institutions–laws, police, courts, etc.–that can be called upon to enforce desired behaviors. As we will later discuss, these institution also play an important role in shaping people’s attitudes, beliefs, and behaviors during the COVID-19 pandemic. These institutions, however, are not entirely divorced from norms. Indeed, they often depend, at some level, on norms (Acemoglu and Robinson, 2012). For instance, if a U.S. President violates the Emoluments Clause of the U.S. Constitution, we depend on members of the U.S. Senate to enforce the clause, and on other voters to vote out members of the Senate who do not do so. No formal institution exists to punish voters who fail to do so. Instead, we must rely on social norms.

There are many reasons to think that social norms would play an important role in motivating COVID-19 containment measures like avoiding social gatherings, washing hands more often, maintaining two meters (six feet) of distance from others, or staying at home. As already stated, these are individually costly actions whose benefits accrue to the group–precisely the kinds of behaviors humans often use social norms to encourage. Most of the required behaviors, such as avoiding social gatherings, maintaining sufficient distance from others, or staying at home, are relatively observable. (The key exception is probably handwashing, but even this is somewhat observable in some contexts.) Thus, it is conceivable that these behaviors would be sustained as social norms. Moreover, early in the pandemic, other more formal institutions had not yet had a chance to respond, and there did not yet exist fines or laws requiring containment measures. This left a void for social norms to fill. That said, social norms can take a long time to change (for a case study, see Tankard and Paluck, 2017), so it is not necessarily the case that they could be harnessed sufficiently quickly to fill this void and motivate the kind of rapid behavior change necessary at the beginning of a pandemic.

If social norms did play a role in motivating people to adopt new behaviors to help contain COVID-19, then we might expect to see their imprint in a number of ways. We make the following key predictions following from our perspective on social norms and the COVID-19 pandemic, most of which will be borne out in our upcoming analyses:

• 1.Measures of the strength of social norms–e.g., first order and second order injunctive beliefs–should be positively correlated with each other, and with COVID-19 containment behaviors.
• 2.Social norms would also be expected to increase people’s expectations of others, meaning they will be more disappointed in others’ behavior if they have more strongly internalized the social norm. The most direct way to provide evidence for this is to test whether our measure of first-order beliefs moderates people’s disappointment in others’ compliance.
• 3.People’s expectations of others might extend to governments as well. That is, they might also be more disappointed in the government’s response to COVID-19 if they hold higher first-order beliefs. Again, we can test for this by estimating whether our measure of first-order beliefs moderates people’s disappointment in their government.
• 4.Finally, we also explore whether people who internalized the norm (i.e., those with higher first-order beliefs) will react most positively to government lockdowns. Specifically, we consider the effect that government lockdowns had on respondents’ perceptions of others’ response, the government’s response, and also their own mental well-being–how concerned they were about COVID-19 and how negative their experiences were. In addition to the main effect of government lockdowns on these survey measures, we will explore whether first-order injunctive beliefs moderate the effect of lockdowns on these measures. We expect lockdowns to have a larger effect on perceptions of others’ and the government’s reaction for those with high first-order injunctive beliefs. However, the effect of social norms on people’s proximate sense of well-being–the feelings and thoughts that they consciously experience–is unclear. Even their effect on baseline (pre-lockdown) well-being is unclear. Internalizing norms might have made people more anxious about COVID-19. In this case we would expect those with high first-order injunctive beliefs to be more worried and have a more negative experience of the pandemic. At the same time, compliance with norms can sometimes be associated with positive emotions such as “warm glow” (Andreoni, 1990).

In the next section, we outline the methods used to test these predictions in the context of the early and accelerating phases of the COVID-19 pandemic.

3. Methods

We conducted a global online survey at the onset of the COVID-19 pandemic. The survey was conducted through the Qualtrics survey platform and contained 58 total items, including 3 scales (23 total items), and 8 demographic measures. The complete survey instrument is available at https://osf.io/zgfjc. Our primary analyses employ items measuring self-reported adherence with COVID-19 containment measures (5 items), first-order injunctive beliefs (4 items), second-order injunctive beliefs (4 items), perceptions of sufficient government response (3 items), perceptions of sufficient public response (1 item), and two mental health scales (13 total items). We use other items such as demographics as control measures, as described below. We obtained exempt IRB review (#E-2065) from MIT.

We posted our first call for participants via social media on March 20, 2020, which ultimately produced nearly 1.4 million impressions on Twitter and accounted for 31,211 clicks on the landing page (www.covid19-survey.org). We amplified this first call via our personal social media accounts, and also mobilized a broad range of individuals engaged in traditional media (journalists, TV hosts or anchors from national and international media outlets) along with social media influencers (ranging from entrepreneurs, to comedians and accounts of online cat communities), broader engagements from a host of international and national NGOs (including various United Nations initiatives such as the UN-s Sustainable Development Solutions Network and the World Values Survey Association network) along with university and alumni Networks. We provide a list of the handles of accounts for which we were able to obtain the retweet information on https://bit.ly/3gAZgO9.

The link to our survey was shared widely, due to our own outreach efforts and organic multiplication through a variety of channels. We identified a further 1,105 original tweets promoting the survey to a combined follower count of 5,601,915 followers. There were at least a further 14,404 retweets of these calls to participate from other users. The survey also appeared in a large number of popular media outlets, including El País (Spain), Frankfurter Allgemeine Zeitung (Germany), Tribune de Geneve, Le Matin, 20Min and 24heures (Switzerland), and the Jakarta Globe (Indonesia), among many others.

Ultimately, between March 20th and April 7th 2020, the landing page was accessed by 391,476 different users, of whom more than 110,000 individuals from 175 countries completed the survey. Our original call accounts for just 7.9% of the direct clicks to the landing page; the bulk of traffic was due to broader dissemination. Indeed, a referrer analysis based on Google Analytics data suggests that 27.5% of referrals came through Twitter and a further 22.8% of landing page clicks were referrals via Facebook. This constitutes only an incomplete list: there are at least 350 other referrers, and a further 37.4% of visits are not identifiable (e.g., because of browser privacy settings), a challenge that is common to data collected through online platforms (e.g., Awad et al., 2018).

We note that our resulting sample should not be viewed as representative, and that it is possible that individuals who were more concerned about the COVID-19 pandemic were disproportionately more likely to take or share this survey. To account for this possibility, we provide further analysis of our dissemination in Appendix B and discuss the limitations of our sampling strategy toward the end of the manuscript.

In this article, we report results from the 58 countries in which at least 200 people participated, corresponding to a sample of 108,075 individuals (see Table S1 for breakdown of participants by country; see Figs. S8 and S9 for replications of our main analyses in which we include all countries in which at least 100 people participated). At the launch of the survey, on March 20th, there had been 240,000 confirmed COVID-19 cases and 9,900 people deaths attributed to COVID-19 (Dong et al., 2020). Two and a half weeks later, on April 7th, confirmed cases and deaths had increased four and five-fold, respectively. The 58 countries included in our sample accounted for 92% of all known COVID-19 cases globally, and 93% of deaths. Within our sample and study period, 13 countries switched from no stay-at-home requirements to nationwide lockdowns (Hale et al., 2021). Our data thus capture global public attitudes in the early and accelerating phases of the COVID-19 pandemic, before, during, and after many governments made challenging and consequential policy decisions.

In the analyses presented below, we re-weight observations to improve their representativeness at the country level, based on respondents’ gender, age, income, and education; however, our results are qualitatively similar without weights (see Table S2 and S3). All underlying data and code to reproduce our findings are available at the following repository: https://osf.io/3sn2k.

4. Results

We next present key results from this study. We begin by discussing respondents’ first-order and second-order injunctive beliefs. We then discuss behaviors (i.e., people’s adherence to COVID-19 abatement measures), satisfaction with others’ adherence and government response, and mental well-being; for each we also discuss the relationship with respondents’ first-order injunctive beliefs, and, when appropriate, with their second-order injunctive beliefs. (For the motivation behind these analyses, please see our predictions at the end of Section 2.)

4.1. Normative Beliefs

A large majority of respondents hold first-order injunctive beliefs that it is important for others to engage in protective behaviors (Fig. 1 , Panel B, dark gray bars): 97% believe that people in their country should cancel their participation in social gatherings because of COVID-19; 92% say people should not shake each other’s hands; 77% think that all shops other than particularly important ones, such as supermarkets, pharmacies, post offices, and gas stations, should be closed; and 81% support a general curfew that only excepts grocery shopping, necessary family trips, and the commute to work. Moreover, 70% of respondents think that risky behaviors should be financially punished. In Panel A of Fig. S1, we present mean first-order beliefs across all questions, by country. The mean first-order belief is above 70% in nearly every country included in our analyses.

Fig. 1.

Behaviors and Beliefs at the Onset of the COVID-19 Pandemic Notes: This figure shows descriptive statistics of personal and perceptions of societal reactions to COVID-19. Panel A presents self-reported engagement in abatement behaviors. Panel B contrasts respondents’ attitudes and perceived attitudes of compatriots about abatement behaviors and policies. Panel C shows the share of respondents who think that the government action has been excessive by country. Respondents from countries with at least 200 responses are included (see Figs. S8 and S9 in the appendix for results when the threshold is 100 observations). Responses are weighted to be representative at the country level in terms of age, gender, income, and education. Panel A and B are further weighted by country population to account for different country sizes. Panel C displays weighted country averages.

Respondents’ second-order injunctive beliefs (their beliefs about what others believe should be done) are somewhat lower, but still quite high. Specifically, respondents estimate that 67% of their fellow citizens believe that social gatherings should be cancelled, 74% of people in their country support avoiding handshakes, 63% of people believe stores should be closed, and 55% are in favor of curfews. In Panel B of Fig. S1, we present the mean second-order belief across all questions, by country. Second-order beliefs are more spread out, but above 60% for roughly three-quarters of the countries included in our analyses.

For both first-order and second-order beliefs, we form scales to use in our upcoming analyses. For first-order beliefs, we form our scale by summing subjects’ affirmative responses to our questions. That is, a participant who responded in the affirmative that they believed people should cancel participation in social events and that risky behaviors should be financially punished, but not for the remaining questions would receive a first-order injunctive belief score of 2. For second-order beliefs, we form our scale by averaging subjects’ responses across all second-order injunctive belief questions. For both scales, we normalize the scale so that the mean and standard deviation are 0 and 1, respectively.

We first test the hypothesis that respondents internalized the social norm more strongly (i.e., their first-order injunctive beliefs should be higher) when they believed others expect compliance (i.e., their second-order injunctive beliefs are higher) (Bicchieri, 2005, Bicchieri, 2016, Fehr and Fischbacher, 2004). Consistent with this hypothesis, we find that first-order and second-order injunctive beliefs are correlated. In particular, the unconditional correlation of our measures of first and second-order injunctive beliefs is 0.20, and this correlation is 0.12 ($p<0.001$) in weighted regressions with country and date fixed-effects effects and controls for the severity of COVID-19.

It is noteworthy that respondents’ second-order injunctive beliefs are substantially lower than their first-order injunctive beliefs–between 14 percentage points and 30 percentage points lower. What should we make of this pattern? One possible interpretation is that respondents underestimate others’ injunctive beliefs. Such underestimates might be the result of, for example, media attention, which tended to focus on those who disagreed with containment measures. If this interpretation is true, then this gap between first-order and second-order injunctive beliefs might have led individuals to reduce their own adherence (Bursztyn et al., 2018, Jachimowicz et al., 2018). However, there is another possibility, namely that subjects who participated in our survey were more concerned about COVID-19 than the average person in their country, because they were more motivated to attend to information pertaining to the survey, and were more interested to provide their opinion about the then-rapidly spreading pandemic.

4.2. Behavior

We document broad adherence to COVID-19 containment measures at the onset of the COVID-19 pandemic (see Fig. 1, Panel A). Globally, 87% of respondents reported that they did not attend any social gatherings in the past week at the time they were surveyed; 82% washed their hands more frequently than a month earlier; 90% said that they would immediately inform people around them if they experienced COVID-19 symptoms; 67% reported keeping a distance of at least 2 meters (6 feet) to other people; and 76% said that they stayed home in the past week (SI Figure 4 presents the country-level averages). People also planned to maintain these behaviors. For example, while 42% of respondents reported that they would leave their home in the next 5 days to buy food, only 19% said that they would go to work, and 45% of respondents indicated that they would not leave their home for any reason in the subsequent 5 days. In panel C of Fig. S1, we present the mean of self-reported adherence to all containment measures, by country. It is above 70% in nearly every country included in our analyses. These results are in line with prior work revealing the broad adherence to COVID-19 protective measures for Italy (Barari et al., 2020), and other countries around the world (Allcott et al., 2020, Andersen, 2021, Kushner Gadarian et al., 2021).

We use responses to these questions to create a self-reported prevention behaviors index, summing respondents’ responses, then normalizing the result so that its mean and standard deviation is 0 and 1, respectively.

We next tested two predictions related to social norms: That people should be more likely to adhere to the norm when they internalize this norm themselves (their first-order injunctive beliefs are high), and that they should be more likely to adhere when they believe that others expect compliance (their second-order injunctive beliefs are high). To test this hypothesis we estimate the following equation:

$$beh\_inde{x}_{ict}={\alpha }_0+{\alpha }_{1}Fo{B}_{ict}+{\alpha }_{2}So{B}_{ict}+{\alpha }_3{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $beh\_inde{x}_{ict}$ is the index combining four different measures of adherence to preventative behavior. $Fo{B}_{ict}$ and $So{B}_{ict}$ are our measures of first-order and second-order injunctive beliefs, respectively. $X_{ict}$ are time-varying control variables (including individual level characteristics and controls for country-level changes and levels in COVID-19 cases and deaths), ${\delta }_c$ and ${\theta }_t$ are country and day fixed effects, respectively.

As predicted, both sets of injunctive beliefs are associated with adherence to containment measures, even in regressions that controlled for both at once. A one standard deviation item increase in our first-order injunctive beliefs scale is associated with a 0.270 (SE $=0.027$, $p<0.001$) increase in the self-reported prevention behaviors index, and and a one standard deviation increase in our second-order belief scale is associated with a 0.144 (SE $=0.025$, $p<0.001$) increase in the self-reported prevention behaviors index. These results are robust to controlling for the battery of covariates described above. Moreover, the magnitude of the correlations is as large or larger than for other covariates, including, importantly, those related to one’s own risk from COVID-19 (see Table 1 ).

Table 1.

Relationship Between Self-Reported Prevention Behaviors and First- and Second-Order Beliefs

	Self-reported behavior index
	(1)	(2)	(3)	(4)
First-order belief index	0.301***	0.270***	0.270***	0.272***
	(0.031)	(0.027)	(0.027)	(0.027)
Second-order belief index	0.260***	0.144***	0.140***	0.136***
	(0.032)	(0.025)	(0.022)	(0.023)
Age			-0.415	-0.421
			(0.240)	(0.238)
Income bracket			0.022*	0.023*
			(0.011)	(0.010)
Education bracket			0.042**	0.043**
			(0.013)	(0.014)
Own health			0.027	0.027
			(0.022)	(0.022)
Confirmed COV-19 cases per capita				0.085**
				(0.031)
Lagged confirmed COV-19 cases per capita				0.005
				(0.026)
Confirmed COV-19 deaths per capita				-0.080***
				(0.014)
Lagged confirmed COV-19 deaths per capita				0.115
				(0.067)
Constant	-0.099	-0.093***	0.098	0.098
	(0.052)	(0.001)	(0.085)	(0.084)
Country-age-gender FE	No	Yes	Yes	Yes
Observations	108075	107863	102163	102163

Notes. This table shows four different regressions of our index of self-reported adherence to COVID-19 prevention behaviors on first-order and second-order injunctive beliefs. The table presents standardized beta coefficients. Standard errors are clustered by country and show in parentheses. * $p<0.05$,** $p<0.01$,*** $p<0.001$.

4.3. Satisfaction with Others’ Adherence and Government Response

Fifty-eight percent of respondents perceived that the reaction of their country’s public to the COVID-19 outbreak had been insufficient. Respondents held similarly pessimistic beliefs about their governments. Globally, just 9% of respondents believed that the response of their country’s government had been too extreme (see Fig. 1, Panel C). This pattern holds robustly across different social groups: When splitting the sample based on country, gender, median income, and median age, there is not a single socio-economic stratum in which a majority of individuals thought that the government reaction had been “somewhat extreme” or “too extreme” (see Figure S4). Indeed, many respondents believed that their government’s actions were not sufficient, with 42% of respondents across the 58 countries holding such beliefs. In addition, a substantial proportion of respondents (36%) indicated that they did not trust their government’s handling of COVID-19, and 34% even stated that they believed their government had not been truthful about COVID-19. Taken together, these results highlight that at the onset of the COVID-19 pandemic, a substantial number of respondents believed that their citizens’ and government’s response was insufficient.

We also tested the relationship between first-order injunctive beliefs and dissatisfaction with others and the government’s response, by estimating the following equation:

$$y_{ict}={\alpha }_0+{\alpha }_{1}Fo{B}_{ict}+{\alpha }_{2}So{B}_{ict}+{\alpha }_3{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $y_{ict}$ indicates respondents’ satisfaction in others’ response or the government’s response. $Fo{B}_{ict}$ and $So{B}_{ict}$ are our measures of first order and second order injunctive beliefs, respectively. $X_{ict}$ are time-varying control variables (including individual level characteristics and controls for country-level changes and levels in COVID-19 cases and deaths), ${\delta }_c$ and ${\theta }_t$ are country and day fixed effects, respectively.

Unsurprisingly, those who had most strongly internalized social norms (and thus had relatively high first-order injunctive beliefs) were more dissatisfied both with others’ response and the government’s response: a one standard deviation-item increase in our first-order injunctive beliefs scale is associated with a 0.288 (SE $=0.017$, $p<0.001$) and 0.273 (SE $=0.017$, $p<0.001$) standard deviation increase in dissatisfaction with others’ response and the government’s response, respectively (see Table 2 ).

Table 2.

Relationship Between Judgments’ of Others’ and the Government’s Response and First-order Injunctive Beliefs

	(1)	(2)
	Public reaction insufficient	Gov’t reaction insufficient
First-order belief index	0.288***	0.273***
	(0.017)	(0.017)
Age	-0.094	0.000
	(0.140)	(0.062)
Income bracket	0.025*	0.011
	(0.011)	(0.008)
Education bracket	0.022	0.021
	(0.013)	(0.017)
Own health	-0.030	-0.069***
	(0.016)	(0.013)
Confirmed COV-19 cases per capita	-0.054	-0.046
	(0.042)	(0.037)
Lagged confirmed COV-19 cases per capita	0.004	-0.022
	(0.042)	(0.028)
Confirmed COV-19 deaths per capita	-0.070	0.001
	(0.052)	(0.022)
Lagged confirmed COV-19 deaths per capita	-0.054	-0.029
	(0.081)	(0.082)
Constant	0.028	-0.014
	(0.050)	(0.025)
Country-age-gender FE	Yes	Yes
Observations	102163	102157

Notes. This table shows two different regressions of the perception of an insufficient public or government response on first-order injunctive beliefs. The table presents standardized beta coefficients. Standard errors are clustered by country and show in parentheses. * $p<0.05$,** $p<0.01$,*** $p<0.001$.

4.4. Mental Well-being

We next test the effects on mental well-being, which we assessed using two measures: one, the PHQ-8, is a widely-used depression index (see Kroenke et al., 2001; $\alpha =.86$),2 while the second contained items we developed to capture anxieties and worries specific to the COVID-19 pandemic, with items including, “I am nervous when I think about current circumstances,” and “I am worried about my health” ($\alpha =.58$); we subsequently refer to this scale as the COVID-Brief Anxiety Inventory “CBAI” or just the “worries index”.3 Further details on the contents and construction of the scale appear in Appendix C.9.

We regress these two measures of mental well-being on our first-order injunctive beliefs scale, as well as four variables that capture respondents’ views of their fellow citizens and government during the COVID-19 pandemic: (a) the perception of an insufficient public response; (b) the perception of an insufficient government response; (c) a lack of trust in their government; and (d) the perception that their government is untruthful. We use the following regression equation to estimate the effects:

$$m{h}_{ict}={\beta }_0+{\beta }_{1}per{c}_{ict}+{\beta }_2{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $m{h}_{ict}$ is the first-order belief or mental health outcome of interest for individual $i$ on day $t$ in country $c$, $per{c}_{ict}$ is the standardized value of the perception of others. We report standardized coefficients to enable comparisons with prior research. This means that the coefficient can be interpreted as the change in the dependent variable, in standard deviations of this variable, associated with an increase in the independent variable by one standard deviation.

We find that respondents who had internalized norms of adherence (those with higher first-order injunctive beliefs) had higher depression scores ($\beta =0.042$, SE=0.019, $p<.001$; see Fig. 2 , Panel A; and Table S4) and higher worries scores ($\beta =0.336$, SE=0.031, $p<0.001$).

Fig. 2.

Effects of Government Response at Onset of COVID-19 Pandemic Notes: Panel A depicts the individual-level weighted pairwise relationship between the variables indicated in the figure heading and row, controlling for respondents’ age, gender, education, health as well as country and date fixed effects. The regressions in Panel B are estimated using the individual-level weighted data, controlling for country and day fixed effects; the independent variable is an indicator of whether the country implemented a lockdown (“stay at home” policy). Standard errors in Panel A and B are clustered by country. The regressions in Panel C are estimated using the individual-level data from the UK and a set of control group countries. The regressions control for country-by-education-by-gender fixed effects and date fixed effects. The independent variable is an indicator variable taking the value 1 for respondents participating from the UK after the 23 March 2020. Panels D, E and F illustrate the impact of the UK lockdown announcement among respondents from the UK compared to the average time trend among participants from control group countries on perceptions of sufficiency of the government response, the CBAI and the second-order beliefs. Standard errors in Panels C-F are clustered by country-age-gender. All figures show standardized beta coefficients.

We further find that perceptions of an insufficient public response are not significantly related to the PHQ-8 ($\beta =.036$, $SE=.022$, $p=.101$) but are related to higher values on the CBAI ($\beta =.140$, $SE=.020$, $p<.001$). Similarly, perceptions of an insufficient government response are not significantly related to the PHQ-8 ($\beta =.048$, $SE=.036$, $p=.180$) but are related to higher values on the CBAI ($\beta =.183$, $SE=.019$, $p<.001$). In addition, a lack of trust in the government’s response is related to both higher values on the PHQ-8 ($\beta =.090$, $SE=.027$, $p=.001$) and CBAI ($\beta =.093$, $SE=.018$, $p<.001$), similar to perceptions of the government being untruthful (PHQ-8: $\beta =.083$, $SE=.029$, $p=.005$; CBAI: $\beta =.079$, $SE=.017$, $p<.001$).

Respondents’ first-order injunctive beliefs, pessimistic beliefs about their fellow citizens’ and government’s response to COVID-19 are thus all associated with lower mental well-being, particularly in terms of their anxieties specific to the COVID-19 pandemic (as measured by the CBAI). We note that these results are in line with prior research which found that trusting other citizens and the government are important predictors of well-being (Helliwell, 2006, Helliwell and Huang, 2008); indeed, this relationship may be amplified given the particularly high levels of uncertainty likely to occur at the onset of the COVID-19 pandemic (Kay et al., 2009).

We also study the relationship between perceptions of the government and the general public with our measures of mental health using the same specification. We find that perceptions of an insufficient public response are not significantly related to the PHQ-8 ($\beta =.036$, $SE=.022$, $p=.101$) but are related to higher values on the CBAI ($\beta =.140$, $SE=.020$, $p<.001$; see Fig. 2, Panel A; and Table S4). Similarly, perceptions of an insufficient government response are not significantly related to the PHQ-8 ($\beta =.048$, $SE=.036$, $p=.180$) but are related to higher values on the CBAI ($\beta =.183$, $SE=.019$, $p<.001$). In addition, a lack of trust in the government’s response is related to both higher values on the PHQ-8 ($\beta =.090$, $SE=.027$, $p=.001$) and CBAI ($\beta =.093$, $SE=.018$, $p<.001$), similar to perceptions of the government being untruthful (PHQ-8: $\beta =.083$, $SE=.029$, p=.005; CBAI: $\beta =.079$, $SE=.017$, $p<.001$).

4.5. The Effects of Government Lockdowns

Our sampling time period covered the onset of the COVID-19 pandemic in which government policies became increasingly stringent (Cheng et al., 2020, Hale et al., 2021). One particularly salient government response was the institution of nationwide lockdowns (“stay-at-home” orders) in many countries during our survey period. We explore whether these policy changes were related to respondents’ normative beliefs, as well as their views of their fellow citizens’ and government’s response, government beliefs, and whether they improved mental well-being.

In this subsection, we explore this question using daily individual data, exploiting time variation in country-level lockdown announcements (Hale et al., 2021), thus comparing individuals in the same country over time (see Fig. 2, Panel B; and Table S5, Panel A). We consider subjects’ first-order and second-order injunctive beliefs, as well as four variables that capture respondents’ views of their fellow citizens and government during the COVID-19 pandemic: (a) the perception of an insufficient public response; (b) the perception of an insufficient government response; (c) a lack of trust in their government; and (d) the perception that their government is untruthful. We regress these variables on a dummy indicating whether a government lockdown was in place at the time the respondent participated in the survey, using the following equation:

$$y_{ict}={\gamma }_0+{\gamma }_{1}lockdow{n}_{ict}+{\gamma }_2{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $y_{ict}$ for individual $i$ on day $t$ in country $c$ is the outcome of interest and $lockdow{n}_{ict}$ is an indicator variable that is equal to one after a country implemented a lockdown.

We find that lockdowns are associated with a variety of (short-term) positive effects. Their estimated effect on first-order injunctive beliefs is positive ($\beta =.08$, $SE=.047$, $p=.09$), as is the effect on second-order injunctive beliefs ($\beta =.15$, $SE=.11$, $p=.18$), though we note that these coefficients are not statistically significant within commonly accepted thresholds. We also find that as a country announced a nationwide lockdown, respondents were less likely to view their government as not being truthful ($\beta =-.115$, $SE=.056$, $p=.045$), less likely to view the government’s reaction as insufficient ($\beta =-.165$, $SE=.068$, $p=.018$), and reported lower values on the CBAI ($\beta =-.091$, $SE=.044$, $p=.043$). While these effects are statistically significant using commonly used thresholds, we urge caution in their interpretation given that they do not meet more stringent significance thresholds which have been proposed recently (Benjamin et al., 2018) and which may take on particular importance in the context of a pandemic (Enserink and Kupferschmidt, 2020). In addition, we find no statistically significant effect of lockdown announcements around the world on the perceived sufficiency of public response ($\beta =-.034$, $SE=.064$, $p=.594$), trust in government ($\beta =-.090$, $SE=.067$, $p=.183)$, or the PHQ-8 ($\beta =.108$, $SE=.112$, $p=.340$).

We further find significant interactions between first-order injunctive beliefs and a lockdown on a variety of these measures. To assess the importance of these interactions we estimate the following equation:

$$y_{ict}={\gamma }_0^{\prime }+{\gamma }_1^{\prime }lockdow{n}_{ict}+{\gamma }_2^{\prime }lockdow{n}_{ict}\times Fo{B}_{ict}+{\gamma }_3^{\prime }Fo{B}_{ict}+{\gamma }_4^{\prime }{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $FoB$ is the standardized sum of all first-order beliefs. Significant interaction effects (${\gamma }_2^{\prime }$) include perceived insufficient reaction of the government ($\beta =-.078$, $SE=.034$, $p=0.027$) and not trusting the government ($\beta =-.101$, $SE=0.056$, $p=.079$). However, we do not find significant interaction effects for the PHQ-8 ($\beta =-.005$, $SE=.026$, $p=0.827$) and CBAI ($\beta =-.029$, $SE=.050$, $p=.557$). That is, those who most strongly internalized normative beliefs reacted most positively to the lockdowns when it came to satisfaction with their government’s response, and to trust in their government.

We also explored whether, in line with prior theory (Kay et al., 2009), individuals were more likely to rely on their governments to cope with the arising uncertainty if they placed greater trust in them. To assess the importance of trust in governments, we estimate the following equation:

$$y_{ict}={\gamma }_0^{+}+{\gamma }_1^{+}lockdow{n}_{ict}+{\gamma }_2^{+}lockdow{n}_{ict}\times trus{t}_c+{\gamma }_3^{+}trus{t}_c+{\gamma }_4^{+}{X}_{ict}+{\delta }_c+{\theta }_t+{ϵ}_{ict}$$

where $trus{t}_c$ is a standardized country-level measure of trust based on the latest available (pre-pandemic) data from Gallup.4 We find that lockdowns had a stronger effect on respondents’ views of the public and their government in countries where citizens placed a greater amount of confidence in the government before the onset of the COVID-19 pandemic (see Table S8) than in other countries. Respondents in these countries indicated greater trust in the government (interaction effect: $\beta =-.223$, $SE=.066$, $p=.001$) and greater sufficiency of the public response (interaction effect: $\beta =-.301$, $SE=.110$, $p=.008$) after lockdowns were put in place than those in other countries. In terms of respondents’ mental well-being, we find that while the interaction effects are directionally consistent with this view’displaying lower values on the depression ($\beta =-.207$, $SE=.122$, $p=.090$) as well as the CBAI ($\beta =-.129$, $SE=.070$, $p=.069$)–they are not statistically significant within commonly used thresholds. That is, nationwide lockdown announcements had larger effects on respondents’ views of the public and their government in countries where citizens placed greater confidence in the government before the onset of the COVID-19 pandemic than in other countries, while these effects were less consistent in terms of respondents’ mental well-being.

4.6. Case study: Effect of Government Lockdown in the UK

We next provide a more granular view of one specific policy change in Panels C, D, and E of Fig. 2 and Table S5, Panel B. On the evening of March 23, 2020, Prime Minister Boris Johnson announced a nationwide lockdown in the United Kingdom (UK), allowing us to conduct a within-country event study to further illuminate the effects of lockdown announcements on beliefs and mental well-being. Among the countries for which we have dense day-on-day samples needed for this type of analysis–Brazil, Germany, Sweden, the US, and the UK–the UK is the only country that, within our study period, went from virtually no stay-at-home requirements to a complete lockdown overnight. Against this backdrop, we estimate a difference-in-differences model, comparing changes in perceptions in the UK before versus after the lockdown announcement to changes in these other four countries. We implement this by estimating the same regression as for the main sample with the restricted sample.

We note that the estimates of this sudden shift in the UK’s policy may be different from the effects we document across countries in general, particularly given that’as Fig. 1B demonstrates’perceptions about the UK government’s reaction to the COVID-19 pandemic were at the far-left end of the spectrum (i.e., very few people in our sample believed that the government reaction was too extreme). In addition, the UK’s policy response to COVID-19 for a short while involved pursuing a strategy of herd immunity, which may have resulted in higher levels of anxiety. The lockdown announcement brought the UK’s policy response in line with the responses of other peer group countries that had responded earlier and stronger to the crisis.

We find that the nationwide lockdown announcement in the UK was associated with less pessimistic views about the public and government response, and better mental well-being. Specifically, we find that the UK lockdown announcement, relative to other countries, was associated with higher first-order injunctive beliefs ($\beta =.132$, $SE=.118$, $p=.263$) and second-order injunctive beliefs ($\beta =.543$, $SE=.141$, $p<.001$). We also find that the UK lockdown announcement was associated with a lower perceived insufficient response of the public ($\beta =-.408$, $SE=.074$, $p<.001$) and the government ($\beta =-.487$, $SE=.047$, $p<.001$). Respondents also viewed the government as more truthful ($\beta =-.157$, $SE=.046$, $p=.001$), displayed greater trust in the government’s response ($\beta =-.275$, $SE=.045$, $p<.001$) and, crucially, display lower values of depression ($\beta =-.105$, $SE=.041$, $p=.010$) and CBAI ($\beta =-.180$, $SE=.046$, $p<.001$). In this case, we did not find a positive interaction between first-order injunctive beliefs and mental well-being. Indeed, the interaction between first-order injunctive beliefs and the lockdown was negative (CBAI: $\beta =-0.210$, $SE=.085$, $p=.015$; PHQ-8: $\beta =-0.096$, $SE=.095$, $p=.313$; see Figure S3).

In terms of respondents’ views, the increase in trust in the government (.223 standard deviations) and greater sufficiency of the public response (.301 standard deviations) associated with the lockdown announcement in the cross-country analysis is about half the size as the increase in trust in the government (.487 standard deviations) and greater sufficiency of the public response (.408 standard deviations) associated with the lockdown announcement in the UK. In terms of respondents’ mental well-being, the decrease in the CBAI (.091 standard deviations) associated with the lockdown announcement in the cross-country analysis is similarly roughly half the size of the decrease in the CBAI (.180 standard deviations) associated with the lockdown announcement in the UK. In addition, in the UK, we find that the lockdown announcement was also associated with a decrease on the PHQ-8 (.105 standard deviations), though we do not replicate this effect in our cross-country data, which may arise in part because this relationship is dependent on high levels of pre-COVID government trust, as discussed above. In related research, these magnitudes are similar to the effect of direct cash transfers on self-reported worries in a sample of Kenyan individuals (.13 standard deviations; see Haushofer and Shapiro, 2016), and somewhat smaller than the effects of targeted mental health interventions (typically between .20 and .30 standard deviations; see Cuijpers et al., 2013) as well as the cross-sectional association between self-reported health and the worries and depression indices within our sample (.163 standard deviations and .269 standard deviations, respectively; see Table S7).

Taken together, these results suggest that timely and decisive government action–i.e., imposing a nationwide lockdown–was consistently associated in both the cross-country analysis as well as the within-country event study with more favorable views of the public and government response. For mental well-being outcomes, our cross-country analysis provides weak evidence, and our within-country event study stronger evidence, that these outcomes also improved. However, we note that the within-country event study was based on a major policy shift in the UK, which should not be treated as representative of the rest of the world, but rather as a case study of what occurred in one particular country.

4.7. Robustness Checks and Limitations

We conducted several analyses to assess the robustness of our findings. First, we examined the robustness of our results to different permutations over different sets of control variables. To address basic sample composition over our specifications, all contain a particular set of baseline control variables we consider essential (e.g., country fixed effects and date fixed effects, self-reported health, and current daily and lagged COVID-19 cases and deaths per capita in the country of residence). On top of this, we permutated over 128 different combinations of additional control variables. As shown in Figures S5-S7, the majority of specifications have p-values below the .05 threshold, and our main specification (highlighted in red) commonly lies somewhere in the middle of the distribution of coefficients. We further evidence the robustness of our findings by showing that, as we include more control variables and thus explain a larger share of the variation in the dependent variable, the absolute size of our effect grows (see Figure S8).5 These results provide some confidence in our estimates and provide a rationale for choosing a specification with an exhaustive set of control variables.

Second, we address potential concerns arising from the composition of our sample.6 One possibility, as noted above, is that participants in our survey were more concerned about COVID-19 than a representative survey might capture. To evaluate whether this is the case, we conducted a survey sampled to be representative in terms of age, sex, and ethnicity via Prolific (Palan and Schitter, 2018, Peer et al., 2017), with 1,000 participants each in both the US and UK on March 28 and 29, i.e., on dates overlapping with the main sample. We restrict our comparison group in the main survey to respondents in the US and UK who completed the survey on the same dates, and compared responses on self-reported measures assessing preventive behaviors (e.g., whether individuals sheltered-in-place), which we argue approximate individuals’ concerns about COVID-19. We find only very small differences in self-reported preventative behavior across both samples (see Table S8), highlighting that at least for the US and UK, our survey captured a similar level of concern about COVID-19 as a representative survey did.

We also used the Prolific survey as an opportunity to assess the possibility that responses to our survey were driven by social desirability bias. We did this by incorporating a list experiment in this representative survey: a random half of participants (the control group) were presented with a list of four protective measures (e.g., handwashing), and asked how many of them they favored, while the other half of participants (the treatment group) were presented with the same list of four measures, plus one additional item: whether they think there should be a curfew in their country. Respondents in the control group on average agreed with 2.91 statements, while respondents in the treatment group agreed with 3.64 items. The average difference in statements agreed with between the control and treatment group is thus 0.73, which very closely mirrors the share of respondents who, when asked directly, agree that there should be a curfew in their country (0.74 in the representative sample; 0.71 in the main sample). These results suggest that social desirability bias did not play an important role.

Another possibility is that our survey attracted particularly neurotic individuals who may have experienced heightened levels of worries or depression (Jylhä and Isometsä, 2006), and are both more likely to have experienced emotional distress (Kroencke et al., 2020, Modersitzki et al., 2021) and to have sheltered-in-place during the COVID-19 pandemic (Chan, Moon, Savage, Skali, Torgler, Whyte, 2020, Götz et al., 2020). To ascertain the distribution of personality traits in our sample, we accessed and compared the participants in our dataset that reported living in the UK to the BBC Lab Dataset (Ebert et al., 2020, Jokela et al., 2015), which contains Big Five personality data for 588,014 UK residents measured between November 2009 and April 2011. As Figure S7 exhibits, the distributions show considerable overlap for most personality traits; if anything, our sample is somewhat more conscientious, less extroverted, and less neurotic than the sample in the BBC Lab Data. This analysis provides some evidence against the possibility that–at least in the UK–we attracted particularly neurotic people into the sample.

In addition, we explored whether potential selection into our sample is likely to make a difference, that is, we conducted analyses to explore whether the effect of the lockdown varies by demographic characteristics and personality traits. Our analysis reveals no systematic differences across a range of variables, including age, gender, education, income, marital status, and most personality traits (see Table S9). The exception to this pattern is the personality trait openness to new experience, where we find that treatment effects are smaller for people with higher openness scores. These analyses provide some evidence to suggest that even if there were selection on demographic characteristics and personality traits in our sample, they are unlikely to account for our treatment effect estimates.

Despite these reassuring results, we note that our findings should be viewed within the constraint that our sample was not recruited to be representative for each country, which we were unable to do given the tight timeline necessary to field the survey in time to capture the effects of lockdown announcements. As described above, we took steps to diversify the sources of our sample, are able to ascertain bounds around the representativeness in the UK and US, and show the limited impact of potential selection effects. At the same time, it is possible that selection biases continue to occur in our data, for example affecting the representation of political orientation, which has been closely tied to COVID-related attitudes and behaviors in the US (Bursztyn et al., 2020, Gollwitzer et al., 2020, Pennycook et al., 2020). However, while the link between political partisanship and pandemic behavior is particularly prominent in the US, this is not necessarily the case elsewhere, and the global nature of our sample makes it unlikely that our findings would be severely impacted by this potential confound. We encourage future research that seeks to evaluate the effectiveness of governmental interventions around the world during the COVID-19 pandemic to use representative samples to better estimate such sampling biases, which can in turn further inform the interpretation of our results.

5. Discussion

In this paper, we reported results from a large-scale, international survey of COVID-19 attitudes and behaviors at the onset of the pandemic. We find rapid, widespread adoption of normative beliefs associated with COVID-19 containment behaviors: The vast majority of respondents across virtually every country in our sample believed that people in their country should cancel their participation in social gatherings, should not shake each other’s hands, should close shops other than particularly important ones, should impose a strict curfew, and that non-adherence should be financially punished. Most respondents also believe that the majority of others hold these beliefs as well. Moreover, these normative beliefs were strongly associated with adherence to these behaviors.

These findings provide a striking example of how quickly social norms can arise, and motivate meaningful behavior change in the service of solving large-scale public good dilemmas. While there is ever-growing evidence of the importance of social norms for permitting humans to cooperate at a large scale (e.g., Bicchieri, 2005, Bicchieri, 2016, Boyd, 2019, Boyd and Richerson, 2021, Henrich and Muthukrishna, 2021, Ostrom, 1990), our study stands out for two reasons. First, it demonstrates just how quickly social norms can arise. Our survey was administered just a few weeks after the first mentions of COVID-19 in national news sources (Taylor, 2020), and before the peak of the first wave of the COVID-19 pandemic in most countries. Yet, we found that the vast majority of our participants already held high first-order and second-order injunctive beliefs, and were complying with COVID-19 containment measures. These findings suggest that social norms can be effective at promoting contributions to public goods even in rapidly changing environments.

Second, our study demonstrates that social norms can support cooperation at a truly global scale. High first-order injunctive beliefs, second-order injunctive beliefs, and behavioral adherence were not concentrated in a small number of places. Rather, in nearly all countries included in our analyses, participants reported very high first-order beliefs and self-reported adherence. Even second-order beliefs, though more variable, were quite high in most countries. The ability of norms to operate on a global scale is not only of theoretical interest, but also of practical importance, suggesting that social norms can play a role in ameliorating global-scale public good dilemmas like climate change.

Consistent with others’ writings (Acemoglu and Robinson, 2012), social norms in our setting appear to be complements to more formal means of motivating desired behaviors, namely government lockdowns. Those who most strongly internalized social norms were also the most supportive of increased government response, and improved their opinion of government response more when their government instituted a lockdown.

Our study also offers an opportunity to consider not only the influence of social norms on behavior, but also on people’s proximate experiences–the beliefs and feelings they were conscious of, which we aimed to capture using our PHQ-8 and CBAI measures of mental well-being. Theoretically, social norms can have both a positive and negative net effect on people’s mental well-being since they can motivate them to act by causing them to be proud or to experience ‘warm glow’ (and thus have a positive influence on their mental well-being), by motivating people to act by causing them to be anxious about COVID-19 as well as them being seen as non-compliant, or to experience shame if they fail to comply’all of which could have a negative influence on their perceived well-being. We find more evidence of the latter in our data: Those who internalized social norms more strongly also appear to have felt more anxious and depressed. These results are consistent with others’ findings (e.g., DellaVigna et al., 2012), though we emphasize that norms need not always be associated with negative emotions and experiences.

Although our paper focused on the role of social norms in motivating behavior and shaping respondents’ beliefs and experiences, we highlight that our findings on mental well-being are noteworthy in their own right. Whereas others have noted the negative effects of the pandemic and lockdowns on well-being (e.g., Brodeur et al., 2021, Giuntella et al., Möhring et al., 2021, Sibley et al., 2020), our results suggest that lockdowns also had a positive effect. That is, they reduced depression and anxiety on average for our respondents, and particularly strongly for those who had internalized social norms of adherence to COVID-19 containment behaviors. There are, however, a number of ways to contextualize these findings. First, note that our study considers the effects of lockdowns relatively early in the pandemic. It is possible that the longer lockdowns last, the more negative their impact on mental well-being may become (though, for a contrasting perspective, see Lara Aknin and Dunn, 2021), or that some people will react negatively to lockdowns (as Fazio et al., 2021 found for individuals who most strongly supported fining those who failed to adhere to abatement measures). Differences could also arise as a function of how concerned people are about COVID-19, with those who are particularly vested in the pandemic’s effects (such as parents) responding particularly strongly to lockdowns (Möhring et al., 2021). Lockdowns could also have other negative effects, e.g., leading to hostility towards immigrants (Bartos et al., 2020). Finally, we note that, like us, Sibley et al., 2020 found that lockdowns increased trust in government.

We also recognize that social norms were likely only one contributing factor to participants’ pessimism about others’ and the governments’ response, which could also stem from broader worries about labor market consequences. That is, it is possible that participants intuited that others might not be able to comply with containment measures because doing so would make it challenging to make a living. Similarly, their pessimism reflected a concern that if others did not comply, this would have dire consequences for their own livelihoods.

The main limitation of our study is our sampling methodology. As discussed in Section 3, we attempted to make our results as informative of the general population as possible, but we must nonetheless caution against making claims about the general population based on our study. Participants in the study may well have differed from the general population, e.g., in how tech savvy they are, or how concerned they are about COVID-19–factors that may well have skewed their responses in ways that we cannot control for.

We end with a note of gratitude for the rapid, global collaboration behind this very project which mirrors our findings. Dozens of volunteers, listed in the acknowledgements, came together within hours to translate a survey into 70+ languages in order to cover the rapidly unfolding, and often messy, events as they were happening in March 2020. Thousands of people shared our call to participate, and over one hundred thousand participants graciously donated their time so we may learn more about what was going on in their own surroundings, and in their own minds, during this time.

Declaration of Competing Interest

The authors declare no competing interests.

Appendix A. Authors, Affiliations and Contribution Statement

• A. Stefano Caria, Associate Professor, Department of Economics, Warwick University, UK. Survey design, literature and write-up review, and subgroup analysis.

• Thiemo Fetzer, Associate Professor, Department of Economics, Warwick University, UK. Organization, Media outreach, survey design, data preparation, data analysis, research design, write-up, and dissemination analysis.

• Stefano Fiorin, Assistant Professor, Bocconi and IGIER, US. Survey design, translation implementation.

• Friedrich M. Götz, Assistant Professor, Department of Psychology, University of British Columbia, Canada. Media outreach, survey implementation, translation, maintenance, write-up.

• Margarita Gomez, Executive Director of the People in Government Lab, Blavatnik School of Government, University of Oxford, UK. Survey implementation and maintenance, communication, translation, outreach.

• Johannes Haushofer, Assistant Professor, Stockholm University, Sweden. Organization, survey implementation, translation and maintenance, outreach, survey design, data analysis, write-up.

• Lukas Hensel, Postdoctoral Research Fellow, Guanghua School of Management, Peking University, China. Survey implementation, translation and maintenance; data preparation; weight construction; data analysis, write-up.

• Andriy Ivchenko, PhD Student, Social Policy Department, London School of Economics and Political Science, UK. Survey design, implementation, and maintenance, translation, website design, outreach, and dissemination analysis.

• Jon M. Jachimowicz, Assistant Professor, Organizational Behavior Unit, Harvard Business School, US. Project management lead, survey design implementation, translation and maintenance, outreach, survey design, data analysis, write-up.

• Gordon Kraft-Todd, Postdoctoral Research Associate, Boston College, US. Survey implementation, translation, maintenance, and write-up.

• Elena Reutskaja, Associate Professor, IESE Business School, Barcelona, Spain. Survey design, outreach for data collection, media, translations, GDPR compliance, write-up.

• Christopher Roth, Assistant Professor, Department of Economics, Warwick University, UK. Survey design, survey implementation, translation and maintenance, write-up.

• Marc Witte, Research Associate, Institute of Labor Economics (IZA), Bonn, Germany. Survey design, maintenance, data preparation, data analysis, results preparation, write-up.

• Erez Yoeli, Research Scientist, MIT Sloan School of Management, US. IRB Application, survey design, survey implementation, translation and maintenance, analysis, write-up.

Appendix B. Survey Dissemination

The call for participation was initially posted on Twitter via 10 social media accounts of the team members of this project on March 20, 2020. On Twitter, these accounts combine around 22,266 followers with the call being retweeted directly by 2,193 users. The call to participate produced nearly 1.4 million impressions on Twitter and accounted for 31,211 clicks on the landing page www.covid19-survey.org.

In addition to our call for participation, we identified at least a further 1,105 original tweets promoting the survey to a combined follower count of 5,601,915 followers. There were at least a further 14,404 retweets of these calls to participate from other users. Unfortunately, we cannot extract or identify each tweet due to limitations on the Twitter search feature’for example, Twitter only allows us to identify the 100 most recent retweets per tweet. This leaves us with data on just 5,191 or 36% of the retweets we identified. The combined number of followers of accounts behind these retweets stands at an additional 14,460,460 followers.

Among the social media influencers that were mobilized are a broad range of individuals engaged in traditional media (journalists, TV hosts or anchors from national and international media outlets) along with social media influencers (ranging from entrepreneurs, to comedians and accounts of online cat communities), broader engagements from a host of international and national NGOs (including various United Nations initiatives such as the UN’s Sustainable Development Solutions Network and the World Values Survey Association network) along with University- and Alumni Networks. We provide a list of the handles of accounts for which we were able to obtain the retweet information on https://bit.ly/3gAZgO9. We emphasize that this is only a subset of the actual tweets and retweets that we were able to download using various Twitter features ex-post.

While Twitter provides us at least with some data on dissemination, it was only one of many sources for recruitment. Out of the 391,476 different unique landing page visitors, a referrer analysis based on Google Analytics data suggests that only 27.5% of referrals were due to users clicking the link on Twitter. Note that our survey was featured in the media, including the Spanish El Pais and the German Frankfurter Allgemeine Zeitung, along with newspapers Tribune de Geneve, Le Matin, 20Min and 24heures (Switzerland) and the Jakarta Globe (Indonesia), among many others. In addition, our survey was also spread through additional means, including via mailing lists and WhatsApp groups, and was distributed by several NGOs and companies.

Appendix C. Data Dictionary

The survey instrument is provided in the Appendix. This document provides a description of the variables that can be found in the data file on our OSF page (https://osf.io/3sn2k/). We also indicate below which variables were derived measures and how they have been computed. Derived variables are created in the accompanying do-file and not part of the main dataset.

C.1. Meta-information

The survey collection software Qualtrics provides the following meta information:

StartDate – for confidentiality reasons, we do not provide the exact start time but only the date on which a participant took the survey. Dates are coded based on the time of interview recorded in GMT time zone.

ResponseID – Unique ID for an individual survey attempt.

Duration – This variable measures the number of seconds the respondent took to fill out the survey.

UserLanguage – This variable captures the ISO language code for the language in which the survey was taken.

CountryofLiving – This variable indicates the country in which a respondent lives.

iso2c – ISO 3166 alpha-2 country code

year – year that survey was taken

month – month that survey was taken

day – day that survey was taken

region – region

continent – continent

C.2. Past Behaviors

To what extent do the following statements describe your behavior for the past week? [0 = Does not apply at all; 100 = Applies very much]

• •beh_stayhome I stayed at home.
• •beh_socgathering I did not attend social gatherings.
• •beh_distance I kept a distance of at least two meters to other people.
• •beh_tellsymp If I had exhibited symptoms of sickness, I would have immediately informed the people around me.
• •beh_handwash I washed my hands more frequently than the month before.

As a derived variable we generate the beh_index, which is the sum of the above five beh_* dummies pertaining to protective behaviors. This variable is then standardized.

C.3. Future Behaviors

• •leavehome Do you need to leave your home in the next 5 days? [0 = No; 1 = Yes]
• •
  What are the reasons for you to leave your home (check all that apply)? Please try to be as honest as possible. Your answers will be kept confidential. [0 = did not check; 1 = checked]

  • –
    leavehome_reason_work Going to work
  • –
    leavehome_reason_pet Walking a pet
  • –
    leavehome_reason_physical Doing physical activity (e.g. exercising, jogging)
  • –
    leavehome_reason_food Procuring food for yourself or family
  • –
    leavehome_reason_pharmacy Going to the pharmacy
  • –
    leavehome_reason_hospital Going to the hospital / receiving medical treatments
  • –
    leavehome_reason_care Taking care of dependents
  • –
    leavehome_reason_friends Meeting friends or relatives
  • –
    leavehome_reason_tired Getting tired of being inside of the house
  • –
    leavehome_reason_bored Getting bored
  • –
    leavehome_reason_adrenaline Getting some adrenaline (from breaking the law)
  • –
    leavehome_reason_freedom Exercising my freedom
  • –
    leavehome_reason_other Other

As derived variables, we distinguish “good” and “bad” reasons for leaving the house:

• •leavehome_badThis variable is the sum of all “bad” reasons to leave the home, which is leaving the house to: visit friends, because one is bored, for the adrenaline, and to exercise one’s freedom.
• •leavehome_goodThis variable is the sum of all “good” reasons to leave the home, which is leaving the house to: go to work, walk a pet, get exercise, procure food, go to the pharmacy, the hospital, or to care for somebody else.

C.4. Personal attitudes about coronavirus measures

• •fob_social - What do you think: should people in your country cancel their participation at social gatherings because of the coronavirus right now? [No = 0; Yes = 1]
• •fob_handshake - What do you think: should people in your country not shake other people’s hands because of the coronavirus right now? [No = 0; Yes = 1]
• •fob_stores - What do you think: should all shops in your country other than particularly important ones, such as supermarkets, pharmacies, post offices, and gas stations, be closed because of the coronavirus right now? [No = 0; Yes = 1]
• •fob_curfew - What do you think: should there be a general curfew in your country (with the exception of grocery shopping, necessary family trips, and the commute to work) because of the coronavirus right now? [No = 0; Yes = 1]

We refer to these variables as “first-order beliefs”. As a derived variable, we construct the weighted average country-level beliefs about coronavirus attitudes (first-order beliefs), meaning that we generate the weighted average of respondents per country agreeing with each of the beliefs above. We use within-country weights (weight) to account for selection on age, gender, education, and income (see Section 3 below for further details on weight construction).

• •m_fob_social - the country-level average of fob_social [min 0; max 100]
• •m_fob_handshake - the country-level average of fob_handshake [min 0; max 100]
• •m_fob_stores - the country-level average of fob_stores [min 0; max 100]
• •m_fob_curfew - the country-level average of fob_curfew [min 0; max 100]

C.5. Perception of others’ beliefs about coronavirus measures

• •sob_social - How many of 100 people in your country do you think believe that participation at social gatherings should be cancelled because of the coronavirus right now? [slider ranging from 0 to 100 ’ initiated at 0]
• •sob_handshake - How many of 100 people in your country do you think believe that one should not shake other people’s hands because of the coronavirus right now? [slider ranging from 0 to 100 ’ initiated at 0]
• •sob_stores - How many of 100 people in your country do you think believe that all shops in your country other than particularly important ones, such as supermarkets, pharmacies, post offices, and gas stations, should be closed because of the coronavirus right now? [slider ranging from 0 to 100 ’ initiated at 0]
• •sob_curfew - How many of 100 people in your country do you think believe there should be a general curfew in your country (with the exception of grocery shopping, necessary family trips, and the commute to work) because of the coronavirus right now? [slider ranging from 0 to 100 ’ initiated at 0]

We refer to these variables as “second-order beliefs”.

Misperceptions: Differences between first- and second-order attitudes

As derived variables, we construct individual-level misperceptions. We construct our misperception measures by taking the difference between people’s beliefs about others’ attitudes, minus the weighted average of actual attitudes of other respondents ($diff\_ind\_X=\mathtt{m}\_\mathtt{fob}\_\mathtt{X}-\mathtt{m}\_\mathtt{fob}\_\mathtt{X}$). We do this for four dimensions: (i) cancellation of social gatherings, (ii) appropriateness of hand-shakes, (iii) store closures, (iv) a general curfew. The resulting misperception variables are:

• •diff_ind_social - Difference between first- and second order attitudes regarding social gatherings. [min -100; max 100]
• •diff_ind_handshake - Difference between first- and second order attitudes regarding hand shaking. [min -100; max 100]
• •diff_ind_stores - Difference between first- and second order attitudes regarding store closures. [min -100; max 100]
• •diff_ind_curfew - Difference between first- and second order attitudes regarding curfew. [min -100; max 100]

We use the z-score transformation of the sum of all four misperception items to construct our misperception index (misperception_index). Higher values indicate higher levels of misperceptions.

C.6. Financial sanctioning of risky behaviors

• •financialpunishment - What do you think: should risky behaviors, which might enable further spread of the coronavirus, be financially punished? [0 = No; 1 = Yes]
• •
  Which fines should be enforced for the following risky behaviors (amount in your country currency)?

  • •
    financialpunish_1: Participation at social gatherings (amount in country currency) [value as entered by respondent, numerical values only]
  • •
    financialpunish_2: Going out despite exhibiting symptoms of coronavirus [value as entered by respondent, numerical values only]

C.7. Case predictions

• •infect_now - Without looking it up, what is your estimate of the number of people in your country who are currently infected? [value as entered by respondent]
• •infect_onemonth - How many people in your country do you think will be infected 1 month from now? [value as entered by respondent]

We also use the log of case growth as a derived variable:

log_case_growth = log((infect_onemonth-infect_now)/infect_now)

C.8. Perceptions of government/public response & efficacy

• •perceivedreaction - Do you think the reaction of your country’s government to the current coronavirus outbreak is appropriate, too extreme, or not sufficient? [5-point scale; 1 = The reaction is much too extreme; 2 = The reaction is somewhat too extreme; 3 = The reaction is appropriate; 4 = The reaction is somewhat insufficient; 5 = The reaction is not at all sufficient]
• •govtrust - How much do you trust your country’s government to take care of its citizens? [5-point scale; 1 = Strongly distrust; 2 = Somewhat distrust; 3 = Neither trust nor distrust; 4 = Somewhat trust; 5 = Strongly trust]
• •govfact - How factually truthful do you think your country’s government has been about the coronavirus outbreak? [5-point scale; 1 = Very untruthful; 2 = Somewhat untruthfu; 3 = Neither truthful nor untruthful; 4 = Somewhat truthful; 5 = Very truthful]
• •react_pub_appr - Do you think the reaction of your country’s public is appropriate, too extreme, or not sufficient? [5-point scale; 1 = The reaction is much too extreme; 2 = The reaction is somewhat too extreme; 3 = The reaction is appropriate; 4 = The reaction is somewhat insufficient; 5 = The reaction is not at all sufficient]
• •perceivedeffectivnes - What do you think: How effective are social distancing measures (e.g., through a general curfew) to slow down the spread of the coronavirus? [5-point scale; 1 = Not at all effective; 2 = Not effective; 3 = Neither effective nor ineffective; 4 = Effective; 5 = Very effective]

For the analysis, we generate dummies of all measures of the perceptions of government and public response and efficacy described above. Those equal 1 if the individual’s response is above the midpoint of the Likert scale, and 0 if it is below the midpoint. When the data is collapsed at the country-level, it can then be interpreted as the share of respondents who gave a response above the midpoint.

C.9. Worries battery

• •mh_anxiety_1 - I am nervous when I think about current circumstances. [5 point scale; same for all items in this battery; 1 = Does not apply at all; 2 = Somewhat does not apply; 3 = Neither applies nor does not apply; 4 = Somewhat applies; 5 = Strongly applies]
• •mh_anxiety_2 - I am calm and relaxed. [reverse coded]
• •mh_anxiety_3 - I am worried about my health.
• •mh_anxiety_4 - I am worried about the health of my family members.
• •mh_anxiety_5 - I am stressed about leaving my house.

As a derived variable, we generate the “worries index” (mh_index), which is the z-scored sum of the 5 worries questions questions above. Higher values indicate higher levels of worries.

C.10. Depression questionnaire (PHQ9)

The data includes responses eight questions from the commonly used PHQ9 depression questionnaire (Kroenke et al., 2001) (without the suicide question). Higher values indicate higher levels of depression.

How often have you been bothered by the following over the past 2 weeks?

• •PHQ9_1 - Little interest or pleasure in doing things? [4 point scale; same for all items in this battery; 1 = Not at all; 2 = Several days; 3 = More than half the days; 4 = Nearly every day]
• •PHQ9_2 - Feeling down, depressed, or hopeless?
• •PHQ9_3 - Trouble falling or staying asleep, or sleeping too much?
• •PHQ9_4 - Feeling tired or having little energy
• •PHQ9_5 - Poor appetite or overeating?
• •PHQ9_6 - Feeling bad about yourself ’ or that you are a failure or have let yourself or your family down?
• •PHQ9_7 - Trouble concentrating on things, such as reading the newspaper or watching television?
• •PHQ9_8 - Moving or speaking so slowly that other people could have noticed? Or so fidgety or restless that you have been moving a lot more than usual?

Derived variable: The “depression index” (phq9_index) is the z-scored sum of the 8 PHQ9 questions above.

C.11. Personality Battery

The data also includes a ten item version of the Big-Five personality questionnaire (Gosling et al., 2003).

To which extent do the following questions apply to you? I see myself as $\dots$

• •personality_b5_1 - Extroverted, enthusiastic [7 point scale; same for all items in this battery; 1 = Disagree strongly; 2 = Disagree moderately; 3 = Disagree a little; 4 = Neither agree nor disagree; 5 = Agree a little; 6 = Agree moderately; 7 = Agree strongly]
• •personality_b5_2 - Critical, quarrelsome
• •personality_b5_3 - Dependable, self-disciplined
• •personality_b5_4 - Anxious, easily upset
• •personality_b5_5 - Open to new experiences, complex
• •personality_b5_6 - Reserved, quiet
• •personality_b5_7 - Sympathetic, warm
• •personality_b5_8 - Disorganized, careless
• •personality_b5_9 - Calm, emotionally stable
• •personality_b5_10 - Conventional, uncreative

C.12. Personal Information

• •age - Age in years (2020 - year of birth).
• •educ - How many years of education did you complete? [numerical value provided by participants]
• •income - What is your monthly household income, before tax, in your country’s currency? [numerical value provided by participants]
• •marital_status - What is your marital status? [1 = married/co-habitating, 2 = single/divorced]
• •hhmember - How many people live in your household? [numerical value provided by participants]
• •gender - Which gender do you identify with? [1 = Male; 2 = Female; 3 = Other]
• •health - How healthy are you? [1 = poor; 2 = fair; 3 = good; 4= excellent]
• •Comorbidities - How many of the following conditions do you have: cardiovascular diseases, diabetes, hepatitis B, chronic obstructive pulmonary disease, chronic kidney diseases, and cancer? [values ranging from 0 to 5 or more]

Derived variables:

• •age_yr - Age in years (2020- year of birth).
• •age_yr_bin - categorical variables subdividing individuals into five year bin groups.

C.13. Merged external data

C.13.1. John’s Hopkins Data

We merge the JHU COVID19 data introduced in (Dong et al., 2020). The data are matched to each countries ISO2 letter country code and merged to individual respondents based on the date on which a respondent took the survey. This implies that the variables are varying within-country over time.

The variables are:

• •covid_confirmed - confirmed cases in country on date respondent participated
• •covid_death - confirmed COVID19 deaths in country on date respondent participated
• •covid_recovered - patients recovered from COVID19 in a country on date respondent participated
• •l1covid_confirmed - confirmed cases in country on one day prior to date that respondent participated
• •l1covid_death - confirmed COVID19 deaths in country one day prior to date respondent participated
• •l1covid_recovered - patients recovered from COVID19 in a country one day prior to date respondent participated
• •l2covid_confirmed - confirmed cases in country two days prior to the day that respondent participated
• •l2covid_death - confirmed COVID19 deaths in country two days prior to date on which respondent participated
• •l2covid_recovered - patients recovered from COVID19 in a country two days prior to date on which respondent participated

C.13.2. Oxford COVID-19 Policy tracker

We merged the Oxford COVID-19 Policy tracker introduced in (Hale et al., 2021). This data tracks government policy relating to COVID-19 responses. The data is described in detail in (Hale et al., 2021). The data, being a country-by-day panel, is merged based on the date at which a respondent took the survey and thus reflects the policy landscape at the date the person took the interview.

For the purposes of our analysis, we further constructed some derived statistics separately. We leverage the data from (Hale et al., 2021) that is up to date as of April 6th. The data is a country-by-daily data set capturing the different measures countries adopted to constrain the spread of COVID-19. We focus on the main subcomponents S1-S6, capturing government actions grouped as: S1 School closures, S2 Workplace closures, S3 Cancellation of public events, S4 Closure of public transportation, S5 Public information, and S6 Restrictions on internal movement.

The data distinguishes between general country-wide restrictions versus targeted ones. In addition, the data also distinguishes between recommendations versus requirements. We construct restrictions indices that are specific to our sample-countries and time window, and do not use the stringency index that is provided by (Hale et al., 2021). This is because for the sample period under consideration, from March 20 to April 5, 2020, most countries had already adopted quite stringent measures with regard to international travel and public information campaigns. As a result, these and a few other sub-components add little variation.

Instead, we use the (Hale et al., 2021) data to create a set of indicator variables that capture whether a country applied measures in a specific domain S1-S6 that are general, i.e., apply to the country as a whole. Similarly, we also use a robustness measure that constructs an index based on transformed indicator variables capturing whether restrictions in the domain S1-S6 are general and are mandatory. Using these sets of dummy variables, we then construct the first principal component of the data for the set of countries in our estimations and for the time period under considerations.

Table 3 presents the factor loadings for the first three principal components for the measures S1-S6 that are coded as applying countrywide (but are not necessarily mandatory). We observe that the first principal component is positively loaded with little weight being placed on the Public Information component. This is not surprising as most countries had ongoing public information efforts by March 20th, 2020 (the day our data collection began).

Table 3.

Principal components to our construction of a COVID-19 country restriction index: general-country wide (but not necessarily mandatory) measures

	Comp1	Comp2	Comp3	Unexplained
General recommended/mandated School closures	0.407	-0.050	0.889	0.020
General recommended/mandated Workplace closures	0.451	-0.199	-0.388	0.329
General recommended/mandated Public event cancelation	0.425	0.207	-0.126	0.479
General recommended/mandated Public transport closure	0.400	-0.334	-0.159	0.458
General recommended/mandated Public information	0.222	0.892	-0.077	0.094
General recommended/mandated Restrictions internal movement	0.491	-0.086	-0.107	0.357

Table 4 presents the factor loadings for the first three principal components constructed on the dummy variables capturing general and mandated policy changes. Naturally, information campaigns do not have a “mandatory” dimension. As such, this feature, that already added little variation, is dropped. As before, we observe positive loading on all subcomponents.

Table 4.

Principal components to construction COVID-19 country restriction index: general and mandatory restrictions

	Comp1	Comp2	Comp3	Unexplained
General mandated School closures	0.400	0.613	0.681	0.000
General mandated Workplace closures	0.476	-0.010	-0.252	0.393
General mandated Public event cancelation	0.436	0.347	-0.565	0.220
General mandated Public transport closure	0.412	-0.672	0.373	0.130
General mandated Restrictions internal movement	0.503	-0.228	-0.118	0.319

We use these two first principal components to study the impact of country-level policy changes on perceptions at the individual level.

C.14. Further external data

We used two external data sets for our analysis that we cannot include for proprietary reasons.

First, we use data from Gallup to measure confidence in government. This data is collected through the Gallup World Polls. Many universities have institutional subscriptions to the data. For licensing reasons, we are unable to post that data on the OSF. It is available from Gallup and measures the share of respondents in a country that respond ”yes” to the question ”Do you have confidence in national government?” (see https://www.gallup.com/analytics/232838/world-poll.aspx).

Second, we use data collected by the British Broadcast Service too compare personality traits in our sample to other, more general samples. This data can be accessed at https://beta.ukdataservice.ac.uk/datacatalogue/studies/study?id=7656.

Appendix D. Weight Construction

This section describes the construction and use of weights included with the survey data. The included weights correct for differences in income, education, age, and gender between survey respondents and the general population in each country. For countries that lack data on one of the dimensions, the weights correct for the available dimensions. We use data on the population structure from the United Nations statistical agency to construct the weights.7 To weight by income, we use data from the Gallup World Poll.8

Table 5 displays the age bins used for reweighting for each gender to account for the population structure. This definition means that we cannot construct weights for respondents who indicate ‘other’ as gender. To construct income weights, we use country-level income quintiles. Finally, we use three education categories to construct weights: less than 8 years of education, between nine and 14 years of education, and 15 and more years of education.

Table 5.

Age bins used for reweighting

Age bin
18 - 19
20 - 24
25 - 29
30 - 34
35 - 39
40 - 44
45 - 49
50 - 54
55 - 59
60 - 64
65+

These age bins are then used to construct weights based on the frequency of observations in the survey data according to the following formula:

$$j\_\mathit{weigh}{t}_{\mathit{ibc}}=\frac{\mathit{weigh}{t}_{\mathit{jb}}·N_c}{N_{\mathit{jb}}}$$ (1)

where $j\_weigh{t}_{ibc}$ is the weight for individual $i$ in bin $b$, for category $j$ (age-gender, income, education), and country $c$. $weigh{t}_{jb}$ is the fraction of the population in bin $b$ of category $j$. $N_{jb}$ is the number of individuals in our survey in bin $b$ and $N_c$ is the number of observations in country $c$. Intuitively, this formula put more weight on individuals in in bins with few observations and individuals in larger bins.

To construct aggregate individual-level weights, we multiply the weights in different categories.9

$$weigh{t}_{ijbc}={\Pi }_{j\in (educ,inc,ag)}j\_weigh{t}_{ibc}$$ (2)

We also construct weights that account for the differential sample size across countries (weighting all countries equally) by dividing the weights by the number of observations in our sample.

$$weight\_sampl{e}_{ijbc}=weigh{t}_{ijbc}/N_c$$ (3)

D1. Included weight variables

• •weight_sample - weights to reweight individuals within country weighting all countries equally.
• •weight - weight to reweight individuals within country.
• •ag_weight - age-gender weights based on UN population data.
• •no_ag_weights - dummy indicating missing age-gender weights.
• •educ_weight - education weights based on GALLUP World Poll education data.
• •no_educ_weights - dummy indicating missing education weights.
• •inc_weight - income weights based on GALLUP World Poll income data.
• •no_inc_weights - dummy indicating missing income weights.
• •N_country - number of respondents in country of respondent.

Supplementary material

Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.jebo.2021.11.015.

Appendix E. Supplementary materials

Supplementary Data S1

Supplementary Raw Research Data. This is open data under the CC BY license http://creativecommons.org/licenses/by/4.0/