Individual-Level Protective Behavior During the First Wave of the COVID-19 Pandemic

Compliance Without Fear: Individual-Level Protective Behavior During the First Wave of the COVID-19 Pandemic

Frederik Jørgensen1, Alexander Bor1, and Michael Bang Petersen1*

1Department of Political Science, Aarhus University, Denmark *Corresponding author: [email protected]

March 9, 2021

Abstract

Objectives: The outbreak of the COVID-19 pandemic required rapid public compliance with advice from health authorities. Here, we ask who was most likely to do so during the first wave of the pandemic. Design: Quota-sampled cross-sectional and panel data from eight Western democracies (Den- mark, France, Germany, Hungary, Italy, Sweden, UK and USA). Methods: We fielded online public opinion surveys to 26,508 citizens between March 19 and May 16. The surveys included questions about protective behavior, perceptions of the pandemic (threat and self-efficacy), as well as broader attitudes towards society (institutional and interpersonal trust). We employ multilevel and fixed-effects regression models to analyse the relationship between these variables. Results: Consistent with prior research on epidemics, perceptions of threat culturally uniform determinants of both avoidant and preventive forms of protective behavior. On this basis, authorities could foster compliance by appealing to fear of COVID-19, but there may be normative and practical limits to such a strategy. Instead, we find that another major source of compliance is a sense of self-efficacy. Using individual-level panel data, we find evidence that self-efficacy is amendable to change and exerts an effect on protective behavior. Furthermore, the effects of fear are small among those who feel efficacious, creating a path to compliance without fear. In contrast, two other major candidates for facilitating compliance from the social sciences, interpersonal trust and institutional trust, have surprisingly little motivational power during the first wave of the COVID-19 pandemic. Conclusions: To address future waves of the pandemic, health authorities should thus focus on facilitating self-efficacy in the public.

Accepted for publication in British Journal of Health Psychology In most social crises political leaders ask citizens to “keep calm and carry on” while they seek to resolve the crisis themselves. Such an approach is not viable during an epidemic. Be- cause pathogens travel between citizens, the growth of the epidemic is fundamentally tied to citizen behavior (Anderson et al. 2020). Citizens cannot carry on with their lives as usual; rather they need to engage in rapid physical distancing to slow the spread of pathogens (Lew- nard and Lo 2020). Accordingly, during the outbreak of the coronavarius pandemic in 2020, governments across the globe launched interventions to facilitate the public’s compliance with advice regarding avoidant (e.g., physical distancing) and preventive behaviors (e.g., increased hand-washing).

The intensity of these interventions ranged from the use of soft power in the form of information campaigns and leadership to the use of force such as strict curfews and mandatory quarantines for large sections of the population. In democratic societies, however, there are always normative and practical limits to the use of force; repressive interventions interfere with concerns for democratic rights. Furthermore, the public’s compliance with health advice in democratic societies is difficult to effectively police without a massive up-scaling of public surveillance. Thus, in a democracy, compliance with public advice and policy is a function of citizen discretion and, hence, the authorities must rely on persuasion and voluntary compliance (Tyler and Jackson 2014). Influence via communication requires an understanding of the psychology that this communication should tap into (Bonell et al. 2020; Van Bavel et al.

2020; West et al. 2020). For democratic governments’ success in handling both the ongoing coronavirus pandemic and potential future pandemics, it is key to understand the psychological correlates that motivated citizens to comply with public health advice in the initial drastic phase of the current pandemic.

The present manuscript identifies and investigates key individual-level psychological correlates and causes of self-reported protective behavior during the first wave of the coronavirus pandemic across eight Western democracies. This analysis is based on large online surveys re- sembling the populations of interest (N=26,508). Building on a general theoretical framework for understanding protective behavior, protective motivation theory, the study provides insights on the correlates and causes of protective behavior which apply to most Western democracies and, hence, on the factors that efficient health communication from governments, health au-

1 thorities and the media needs to target in order to eﬃciently generate compliance among the public during new waves of the pandemic or in future pandemics.

Protective motivations during a pandemic

The outbreak of the pandemic was followed by the rapid spread of information about the threat of the coronavirus and how to cope with the virus via protective behaviors (Amidon et al.

2021). The framework of protection motivation theory is particularly useful for understanding behavior and motivations in such situations as individual responses to these two components of risk communication constitutes the theory’s core focus. Protection motivation theory stipulates that people’s motivation to comply with risk-relevant recommendations–including in the domain of health–is a function of appraisals of the threat confronting the individual and appraisals of the individual’s ability to cope with threat through compliance with recommendations (Rogers

1975; Maddux and Rogers 1983).

The role accorded to feelings of threat in protection motivation theory is shared with “nu- merous theories in social and health psychology” (Sheeran et al. 2014). Consistent with this, individuals who perceive themselves to be more at risk have been found to be more compliant with protective advice in epidemics more generally (Brug et al. 2009) and measures of fear of the coronavirus have been found to be major determinants of protective behavior compliance in the current pandemic (Brouard et al. 2020; Harper et al. 2020; Kachanoﬀ et al. 2020).

Yet, according to protection motivation theory, individuals need to feel not only threatened but also efficacious for efficient protective behavior. Following Bandura (1982), self-efficacy is understood as a belief in the self’s personal ability to engage in behavior that protects them from a given threat; here, infection with the coronavirus. Self-efficacy therefore requires mastery of both knowledge about the coronavirus and the capability to comply with behavioral recommendations (Rippetoe and Rogers 1987). While knowledge is sometimes assessed separately from assessments of capability, knowledge may be particularly important for self-efficacy when confronted with a novel threat such as a new virus. Consistent with the proposed role of self-efficacy, several strands of research demonstrate that this factor is an important predictor of protective behavior including in epidemics (Bish and Michie 2010; Ruiter et al. 2001; Teasdale et al. 2012).

2 From the perspective of protection motivation theory, an added psychological effect of self- efficacy is to increase the effects of threat appraisals (Maddux and Rogers 1983; Sheeran et al.

2014). Consistent with this, meta-analyses of research on protective behavior against mundane threats (e.g., smoking, sunburn, stress) provides supportive evidence (Sheeran et al. 2014).

However, experimental research using simulations of immediate and large-scale threats such as pandemics and terrorist attacks finds that self-efficacy is more predictive of compliance than feelings of threat, and that self-efficacy beliefs can generate compliance even if personal fear is low (Pearce et al. 2019; Teasdale et al. 2012).Thus, in large-scale crises where the salience of the threat is omnipresent such as a pandemic, self-efficacy may potentially not increase but decrease the relevance of a personal sense of threat. As emphasized by Teasdale et al. (2012) prior to the outbreak of the coronavirus pandemic: “Coping appraisals appear to be an important, and hitherto underresearched, predictor of how people may behave in pandemics”. Given these conflicting theoretical possibilities, it is not only important to assess the separate effects of fear and self-efficacy but also their potential interactive effects.

While appraisals of threat and self-efficacy have a prominent role in the psychological literature, it is also important to note that research on protective behavior in prior epidemics suggests that threat and self-efficacy may not be sufficient for high levels protective behavior (Leung et al.

2005; Brug et al. 2004). Accordingly, we also assess two other psychological predictors that may be important in the context of collectively mobilizing events such as the outbreak of a pandemic

(Johnson et al. 2020): Interpersonal trust in fellow citizens and institutional trust in society’s institutions. These forms of trust have been highlighted as major thereotical determinants of compliance with collective rules and recommendations (Ostrom 1998; Tyler and Jackson 2014) including in relation to the COVID-19 pandemic (Johnson et al. 2020). Interpersonal trust entails the perception that fellow citizens are willing to follow similar norms and institutional trust entails the perception that institutions are willing to enforce the norms (Ostrom 1998).

Some prior work on protective behavior during epidemics has indeed found that compliance is higher among individuals with higher degrees of trust (Rubin et al. 2009) (but see Fong and

Chang (2011)). Also, initial work during the current pandemic found that geographical areas in

United States and Italy with higher levels of trust were more likely to engage in social distancing (Brodeur et al. 2020; Durante et al. 2020). Again, it is relevant to consider both the main

3 eﬀects of interpersonal trust and institutional trust and their interactive eﬀects with appraisals of threat. Thus, individuals who do not feel threatened may still engage in protective behavior as a contribution to the collective project of halting infection spread (Johnson et al. 2020) and, hence, feelings of threat may be less important for compliance among individuals high in trust.

The present study

While some evidence already exists of the predictors of behavior during the coronavirus pandemic, none of these published studies have considered how these alternative non-threat related psychological factors may interact with feelings of threat to shape protective behavior. This is not least important because prior evidence suggests that in extremely salient crises–such as the coronavirus pandemic–trust in the self (i.e., self-efficacy) or others may make feelings of threat less rather than more predictive of compliant behavior. Consequently, the aim of this study is to empirically assess the following two research questions: First, to what extent are feelings of threat, self-efficacy, interpersonal trust, and institutional trust positively associated with protective behaviors during the outbreak of the coronavirus pandemic? Second, to what extent and in what direction, does feelings of threat interact with self-efficacy, interpersonal trust, and institutional trust in predicting protective behaviors during the outbreak of the coronavirus pandemic? In addition to the theoretical contributions underlying these two research questions, the present study also contribute to the existing literature in two important ways: First, the present study seeks to establish generalizable findings by utilizing large-scale representative surveys in eight countries during the first wave of the COVID-19 pandemic (N = 26,508). Second, while most of our results remain correlational, we are able exploit a longitudinal component in our data that allows us to increase causal leverage of key conclusions.

Materials and Methods

Data

We ﬁelded surveys in eight countries during the ﬁrst wave of the COVID-19 pandemic:

Denmark, France, Germany, Hungary, Italy, Sweden, the United Kingdom (UK), and the United

States of America (USA). The sampled set of countries were chosen to represent the diversity of national responses to the COVID-19 pandemic as well as the diversity in the severity of the local

4 epidemic (see further discussion below). In each of the eight countries, the survey ﬁrm Epinion sampled adult respondents using online panels. To increase representativity, survey respondents were quota sampled to match the population margins on age, gender, and geographic location of each of the eight countries in our study. All participants provided informed consent and the study was conducted in accordance with the guidelines of the Danish National Committee of

Health Research Ethics for survey research that do not involve human biological material. All data and required code is publicly available in a repository at the webpage of the Open Science

Framework: osf.io/asczn (link anonymised). We utilize two samples during the ﬁrst wave of the

COVID-19 pandemic. First, a cross-sectional sample collected between March 19 and April 3 including 26,508 respondents overall (the supplementary Table SM1 provides country-speciﬁc starting dates). Second, a panel sample that includes 10,569 respondents who were observed at least twice for a total of 24,720 observations in the period between March 13 and May 16. In this sample, we observe key measures over time and are able to increase causal leverage. (See supplementary section A.1 for further details on the samples and the data collection process).

Measures

Following prior research on behavior during epidemics (Bish and Michie 2010), we assess two types of protective behavior: Avoidant and preventive. Avoidant behaviors include physical distancing, such as avoiding crowds or hugging and kissing people outside your close family.

Preventive behaviors include hygienic precautions (e.g. hand-washing or coughing into ones sleeve).

To assess avoidant behaviors, respondents were asked: (1) did you shake someone’s hand yesterday?, (2) did you hug or kiss someone outside your closest family yesterday?, (3) were you in a room with more than 10 people yesterday?, (4) did you use public transport yesterday?, (5) were you careful yesterday to keep your distance from elderly and chronically ill people?, and

(6) to what degree were you careful yesterday to keep your distance from people outside your closest family?. To assess preventive behaviors, respondents were asked: (1) when you coughed and/or sneezed yesterday, did you do this in your sleeve each time?, and (2) how many times do you estimate that you washed your hands or used hand sanitizer yesterday?.

For each type of behavior, we add together these items to form indexes of avoidant and

5 preventive behavior, respectively. Each index is scaled to vary between 0 and 1 with higher values reﬂecting a higher degree of protective behavior. In addition to these outcomes, we also obtained a measure that directly asked respondents whether they changed their behavior during the COVID-19 pandemic in order to avoid spreading the infection. For the cross-sectional sample, we report analyses with this measure as an alternative outcome in the supplementary materials (see section B.4). All analyses replicate those presented in the main text.

To examine key correlates of protective behavior, we assessed (1) appraisals of threat, (2) self-eﬃcacy, (3) interpersonal trust, and (4) institutional trust.

To measure appraisals of threat, we focus on individual-level worries related to the coronavirus. Speciﬁcally, we create an index that adds together three items that directly measure the extent to which our respondents are concerned about the consequences of the coronavirus:

To what degree are you concerned about the consequences of the Corona virus . . . (1) for your- self?, (2) for your family?, and (3) for your close friends? Respondents answered on 4-points scale from “not at all” to “to a high degree”. Together, the three items form a reliable scale,

α = 0.83, scaled to range from 0 to 1, with higher values indicating higher worry.

To measure self-efficacy, we rely on a five-item scale.1 As noted, self-efficacy requires mastery of both knowledge about how specific measures can protect against COVID-19 and a feeling of being capable of following protective advice (Bandura 1982; Rippetoe and Rogers 1987). While the first four items in our self-efficacy scale tap into the knowledge dimension of mastery, the

ﬁfth item taps into the capability dimension. The ﬁrst four items ask: To what degree do you feel that you know enough about . . . (1) how to avoid being infected and/or infecting others with the Corona virus?, (2) the symptoms of the Corona virus?, (3) what you should do if you fall ill with the Corona virus?, (4) what you as a citizen should do in relation to the

Corona virus? On these questions, respondents answered on a 4-point-scale from “not at all” to “to a high degree”. The ﬁfth item asked: To what extent do you agree or disagree with the following statement: I’m certain I can follow oﬃcial advice to ”distance myself” from others if

I want to. Here, respondents answered on a 5-point scale from “disagree completely” to ”agree completely”. The ﬁve items form a reliable scale as measured by alpha (α = 0.79). For the

ﬁnal index, we add together the ﬁve times and scale it to range from 0 to 1 with higher values

1Due to a programming error, we did not observe self-eﬃcacy in France. Consequently, French respondents are left out of models including this measure.

6 indicating more eﬃcacy. 2

To measure interpersonal and institutional trust, we use standard measures. On interpersonal trust, we ask respondents: “Do you think that most people by and large are to be trusted, or that you cannot be too careful when it comes to other people?”. On institutional trust, we ask: “On a scale from 0 to 10, how much conﬁdence do you personally have in the government?”.

In the analyses, we also include a battery of demographic control variables (see below).

Table SM10 provides descriptive statistics for all variables in our pooled sample, while Tables

SM11-SM18 provide the descriptives by country.

In the panel sample, not all measures were available. As discussed below, the panel sample focuses on the potential causal effect of self-efficacy on protective behavior. To this end, we have one repeated measure of self-efficacy available (specifically, the question: ”To what degree do you feel that you know enough about what you as a citizen should do in relation to the

Corona virus?”) and three repeated measures of protective behavior are available: (1) avoiding crowds, (2) washing hands and (3) the alternative outcome that more broadly assess whether respondents changed their behavior in order to avoid spreading the infection. We summarize these three measures into a modiﬁed index of protective behavior (see supplementary section

B.6 for details).

Statistical analysis

To answer the research questions, we use the cross-sectional sample and regress protective behavior on each psychological predictor (perceived worry, self-eﬃcacy, interpersonal trust, and institutional trust) separately, while controlling for a battery of covariates (age, sex, education, occupation, income, and vote choice). In this benchmark model, data is pooled from all

2Although the scale covers both aspects in the theoretical definition of self-efficacy, we would ideally have observed more items that directly measure the capability dimension. Empirically, however, there is significant evidence that the utilized scale in fact measures a single latent trait of self-assessed mastery. First, the findings are not an artefact of the scale construction. Hence, we rerun our analyses while including only the capability item as a measure of self-efficacy (i.e., the agreement with the statement: I’m certain I can follow official advice to “distance myself” from others if I want to). These analyses show that all results replicate those in the main manuscript (see supplementary section B.3 for details). Note, we are unable to repeat this analysis for the longitudinal results because we, unfortunately, do not measure both dimensions of self-efficacy repeatedly. Second, beyond the evidence provided by the satisfactory alpha-value, we conducted a polychoric PCA, which demonstrates that the five items clearly load on only one dimension. Specifically, the first component in the polychoric PCA has an eigenvalue of 3.20, while the second component’s eigenvalue is 0.80. Third, to assess the consequence that the scale includes more items reflecting knowledge than capability, we also created an alternative scale where the items for each component were equally weighted (i.e., 50 % each). This alternative scale is highly correlated with our preferred scale (r = .91), suggesting again that the used scale does indeed reflect one single trait of self-efficacy.

7 countries and uses multilevel modelling to account for the fact that observations are nested within countries. Speciﬁcally, models with random intercepts for country and time are ﬁtted.

Moreover, the slopes of the psychological correlates are allowed to vary by country (given that likelihood ratio tests show that random slope models ﬁt the data better compared to the random intercepts-only models). The models provide the overall correlations between protective behavior and each psychological predictor across all countries, as well as country-speciﬁc deviations from the overall patterns.3

In addition, we use the panel sample to estimate the causal impact of a change in self-efficacy on protective behavior using individual-level fixed effects linear probability models. Individual- level fixed effects control for stable individual differences (i.e. time invariant confounders) in protective behavior by only using within-individual variation in self-efficacy and protective behavior. Instead of comparing protective behavior among individuals with high versus low levels of self-efficacy, we estimate how protective behavior changes when a respondent reports changes in their self-efficacy compared to an earlier interview. To account for potential time trends in protective behavior we also include time fixed effects. To correct for clustering in self- efficacy, we cluster standard errors by individual and time. The two-way fixed effects estimator gives an unbiased estimate of the causal impact of self-efficacy on protective behavior on the assumption that the protective behavior of individuals had followed parallel trends in the absence of changes in efficacy (Angrist and Pischke 2008). In the supplementary materials, we provide tests that support that this assumption is valid here (see section B.6 for details).

In the analyses, all measures are scaled to range from 0-1. Accordingly, the size of the estimated correlations (in the cross-sectional sample) and eﬀects (in the panel sample) reported below reﬂect the change in the outcome variable when we compare individuals at the minimum and maximum values of each of the independent variables. All reported p-values are from two-sided tests. 3The supplementary materials include results from a full model that includes all psychological predictors as well as all demographics in the same model. All analyses replicate those presented in the main text (see supplementary section B.2).

8 Results

Descriptive levels of protective behavior across countries

Before turning to examining the proposed research questions, we begin by providing descriptive results regarding the pandemic context of the sampled countries as well as the level of protective behavior in each country. To this end, Figure 1 displays information about the pandemic severity, the government response and the levels of observed preventive and avoidant behavior for each country. To represent diﬀerences in government strategies, a dashed line plots a measure of the stringency of government responses to the COVID-19 pandemic over the survey period (Hale et al. 2020). This is a composite measure of the number of non-pharmaceutical interventions taken in a speciﬁc country (e.g., school closings and curfews). The bars display the count of COVID-19 infections per capita as an indicator of the severity of the local epidemic

(also taken from Hale et al. (2020). In addition, Figure 1 plots levels of protective behaviors as a function of time. The solid black lines represent avoidant behavior and the solid grey lines represent preventive behavior.

There is substantial variation in the severity of the local epidemic across countries and, while all countries have implemented some form of stringent measures, there is also substantial variation with Sweden having the less stringent response and Italy and France the most stringent responses. Independently of both stringency and severity, however, the level of protective avoidant behavior is exceptionally high. At the same time, we observe a signiﬁcantly lower, but similarly stable, levels of preventive behavior across the countries.

Main correlations between protective behavior and psychological predictors

To answer our ﬁrst key research question, Figure 2 displays the estimated correlations between each of the protective behavior measures and the psychological correlates.4 Worry correlates positively with both avoidant and preventive behavior such that people who are more worried are more likely to comply with the protective advice on each measure. Whereas we observe a marked correlation between preventive behavior and worry (βpreventive = 0.20, p < 0.0001), the correlation with avoidant behavior is more moderate (βavoidant = 0.06, p < 0.0001). In contrast, self-eﬃcacy is relatively strongly correlated with both types of protective behav-

4for correlates between protective behavior and demographic variables, see supplementary Figure SM1).

9 Figure 1: Selection of Countries for Data Collection

Note: Solid black (avoidant behavior) and grey (preventive) lines are the developments in the present measures of self-reported protective behavior. Dashed lines are the developments in policy stringency (Hale et al. 2020). Red bars display the developments in the COVID-19 case counts per capita.

ior. Compared to the least efficacious, the most efficacious are hence about 11 percentage points more likely to display avoidant behavior (βavoidant = 0.11, p < 0.0001) and about 19 percentage points more likely to display preventive behavior (βpreventive = 0.19, p < 0.0001). Against the theoretical expectation, interpersonal trust is negatively related to both types of behavior (βavoidant = −0.04, p < 0.0471; βpreventive = −0.05, p < 0.0001) such that respondents who trust most other people are less likely to comply with protective advice. Finally, institutional trust has a small but statistically significant positive association with avoidant behavior

(βavoidant = 0.02, p = 0.024), whereas its correlation with preventive behavior is somewhat larger (βpreventive = 0.05, p = 0.024). While the correlations in Figure 2 identify the average correlations with protective behaviors, they potentially mask heterogeneity across countries. To investigate this heterogeneity,

10 Figure 2: Correlations between protective behavior and psychological correlates

Note: Correlations from our benchmark model. Filled black circles (grey triangles) show the estimated country- speciﬁc correlations between each of our psychological variables and avoidant behavior (preventive behavior). Error bars are 95% conﬁdence intervals.

we utilize the benchmark model to extract the country-speciﬁc correlations in Figure 3. Each

filled black circle (grey triangle) shows an estimated country-specific correlation between the respective psychological variables and avoidant behavior (preventive behavior). The black (grey) dashed lines refer to the estimated overall associations between each of the psychological variables and avoidant behavior (preventive behavior) from Figure 2.If the confidence intervals overlap the dashed lines, it means that the country-specific slope is not statistically significantly different from the overall mean slope. On the whole, Figure 3 shows that the estimated correlations are strikingly uniform across the countries.

Interactions between worry and alternative psychological motivations

Our second key research question focuses on the potential interactions between worry, on the one hand, and the other set of psychological correlates, on the other. We examine these potential

11 Figure 3: Country-speciﬁc correlations between protective behavior and psychological correlates

Note: Filled black circles (grey triangles) show the estimated country-specific correlations between each of our psychological variables and avoidant behavior (preventive behavior). Black (grey) dashed lines refer to the estimated overall associations in Figure 2. Error bars are 95 % confidence intervals that show whether the country-specific correlations are statistically significantly different from overall associations.

interactions by re-specifying the benchmark models such that they include a linear interaction between worry and each of the remaining psychological correlates, respectively. Figure 4 shows the results from these moderation analyses. The solid grey lines show predicted compliance over the range of worry at low levels on the respective moderators, the dashed lines show predicted compliance at medium levels, while solid black lines show compliance at high levels.

For self-efficacy (top-left panel), we observe a substantive and statistically significant moderation of the association between worry and avoidant behavior (βavoidant = −0.22, p < 0.0001). Among those with minimal levels of self-efficacy, worry has a substantial effect on avoidant behavior such that the difference in predicted values among those who feel worried and those who do not is 0.24 (p < 0.0001). Among those with maximal levels of self-efficacy, however, this difference is statistically indistinguishable from 0. For preventive behavior (bottom-left panel),

12 Figure 4: Do self-eﬃcacy and trust moderate the correlation between worry and protective behavior?

Note: Solid black lines show predicted values at high levels of each moderator, black dashed lines show predicted values at medium levels, and solid grey lines show predicted values at low levels.

we observe a similar empirical pattern. The difference in preventive behavior between those who are worried and those who are not is largest at low levels of self-efficacy and decreases as self-efficacy increases (βpreventive = −0.23, p < 0.0001). On institutional trust, we similarly find that worry is significantly and substantively moderated by institutional trust. For avoidant behavior (top-right panel), we observe a substantive and statistically significant decrease in the difference between the worried and unworried

(βavoidant = −0.14, p < 0.0001) from 0.14 (p < 0.0001) at low levels of institutional trust to 0 (p = 0.599) at high trust levels. For preventive behavior (bottom-right panel), we similarly observe that the diﬀerence decreases from 0.27 (p < 0.0001) at minimal levels of institutional trust to 0.13 (p < 0.0001) at maximal trust levels.

On interpersonal trust, we observe more modest moderations for both types of behavior

(βavoidant = −0.06, p < 0.0001; βpreventive = −0.08, p < 0.039). On avoidant behavior, the

13 diﬀerence between the worried and unworried decreases from 0.08 (p < 0.0001) at low levels of trust to 0.02 (p = 0.0004) at high level of trust. On preventive behavior, the decrease reﬂects a change from 0.23 (p < 0.0001) at low trust levels to 0.15 (p < 0.0001) at high trust levels.5

Figure 5: Country-speciﬁc moderations

Note: Filled black circles (grey triangles) show the estimated country-specific moderations. Error bars are 95 % confidence intervals that show whether the country-specific moderations are statistically significantly different from overall moderations (see Figure 4).

Figure 5 displays the consistency of these interactions across the countries in our sample.

For each country, the black filled circles (grey triangles) show the estimated country-specific interaction between worry and each of the moderators on avoidant (preventive) behavior. If the confidence intervals overlap the dashed black (grey) lines, the country slopes are not statistically

5Following the advice of (Hainmueller et al. 2019), we test the robustness of these linear interactions in the supplementary section B.5. Figure SM14 shows the binned estimator that provides no evidence for non-linearities in the interactions. Figure SM15 shows the more flexible kernel estimator. Although it should be noted that there is some tendency for a curvilinear interaction between between worry and self-efficacy on preventive behavior, it is crucial that this observed curvilinearity is driven by very few observations. For more than 95 percent of the observations, we observe a linear negative moderation (please see the distribution on the self-efficacy moderator in SM15). Altogether, this corroborates the robustness of the results.

14 distinguishable from overall mean slope on the specific interaction term. Overall, we find the interactions to be relatively consistent across cultures, suggesting that the substantive patterns observed in Figure 4 applies across the different countries in our sample.

Assessing longitudinal eﬀects of self-eﬃcacy on protective behavior

The above analyses highlight the importance of self-efficacy for protective behavior during the outbreak of a pandemic. To increase the causal leverage of any conclusions regarding this key variable, we utilize the panel sample and estimate the effect of self-efficacy on protective behavior using the two-way fixed effects estimator. We observe a substantial effect of efficacy of 0.06 (p < 0.0001) (see Table SM19). Additional analyses shows that this effect is relatively homogeneous across countries and varies between βDenmark = 0.04 (p = 0.065) and βUSA = 0.12 (p = 0.0251) (see the supplementary Figure SM16).

Discussion

In this study, we asked two research questions: First, to what extent does appraisals of threat, self-efficacy, interpersonal trust and institutional trust predict protective behavior during the first wave of the coronavirus pandemic? Second, does self-efficacy and trust moderate the effects of feelings of threat on protective behavior?

We found extremely high levels of avoidant behavior (e.g., physical distancing) in all countries and medium to high levels of preventive behavior (e.g., increased hand-washing) across the countries. Most likely, this reflects that policy measures during the first wave of the pandemic were particularly targeted towards fostering avoidant behavior in the form of distancing. In answering the first research question, we found that individual-level variation in these protective behaviors strongly reflected individual-level differences in threat appraisals (i.e., self-assessed worry) and self-efficacy but not measures of trust. In answering the second research question, we found strong evidence that individual-level differences in self-efficacy decreased the importance of threat as a predictor of protective behavior and also found some evidence that the trust measures similarly negatively moderated the association between threat and protective behavior. Finally, additional analyses using panel data suggested (1) that efficacy is amendable to change and does not just reflect a set of stable traits and (2) that changes in efficacy causally

15 impacts protective behavior.

Several years prior to the coronavirus pandemic, Teasdale et al. (2012) used protection motivation theory to argue that self-efficacy could be “an important, and hitherto underresearched, predictor of how people may behave in pandemics”. Overall, the present findings provide direct evidence for this assertion during the outbreak of an actual pandemic. Furthermore, the findings extend previous studies of protective motivations by suggesting that self-efficacy may be even more important in the face of exceptionally salient threats such as a pandemic. Thus, prior research on protection motivation in the context of mundane threats (e.g., smoking, sunburn and stress) suggest that self-efficacy enhances the effects of threat appraisals (Sheeran et al.

2014). In contrast, we found that high levels of self-eﬃcacy made individual-level feelings of threat almost irrelevant for the protective behaviors promoted by authorities worldwide during the coronavirus pandemic, providing a pathway to compliance without fear.6

This finding might be key for health communication during both the current and future pandemics. Thus, the risk profile of many diseases including COVID-19 is highly asymmetric with some individuals being more at risk (Jordan et al. 2020). As physical distancing and other protective measures are most effective if most parts of society comply (Anderson et al. 2020), such asymmetries imply that health communication that exclusively focuses on personal risk and threat may not motivate sufficiently high levels of protective behavior. Instead, the present

findings suggest that an alternate and more effective focus for health communication is self- efficacy. A focus on promoting self-efficacy rather than fear may also be normatively desirable.

While an increased sense of threat may increase compliance, it can entail mental health costs

(Ornell et al. 2020) and increase the acceptance of undemocratic treatments of other groups

(Marcus et al. 1995).

The fact that individual diﬀerences in trust was not strongly associated with compliance during the outbreak of the coronavirus pandemic also has important theoretical implications.

In particular, it is relevant to note that people high in interpersonal trust were less likely to engage in protective behaviors compared to people low in interpersonal trust across all countries.

While unexpected, this eﬀect is consistent with some prior ﬁndings on the relationship between

6While efficacious individuals may not need to feel personally threatened in order comply, it is possible to ask whether they do need to feel that society as such is threatened by CVOID-19? To assess this, we conducted additional analyses, reported in Section B.7, which replicated the findings using a measure of feelings of societal threat. Hence, during first wave of the COVID-19 pandemic, efficacious individuals did not need to feel that they or society was threatened in order comply.

16 trust and protective behavior during epidemics (Lear 1995; Fong and Chang 2011). While trust may increase other forms of protective behavior, the trusting mindset seems to make it psychologically diﬃcult to treat others as infection threats (see Aarøe et al. (2016). For institutional trust, the inconsistent and weak correlations between protective behavior and institutional trust may be viewed as welcome news. According to the present data, political views and polarization does not necessarily jeopardize compliance in the outbreak of a massive global crisis like a pandemic, although it may naturally inﬂuence compliance at later stages (see

Gollwitzer et al. (2020)).

These conclusions notwithstanding, three important limitations of the present study should be noted. First, results from the cross-sectional analyses are limited in terms of causality, and while the applied two-way ﬁxed eﬀects estimator controls away all time-invariant unobserved heterogeneity, it does not control for time-varying factors that may still bias our estimates.

Future research should seek to corroborate our findings relying on more longitudinal and experimental designs including to assess how the interaction between threat appraisals and self- efficacy may change as a function of the magnitude of the crisis. Second, the study relied on high-quality online panels for recruiting in our participants and not random probability sampling. Accordingly, the samples may be systematically different to the national population, for example in terms of digital literacy. This sampling bias may increase due to non-response. In the present study, response rates for the individual countries vary between 18-38 % and may imply that the samples are biased towards those most interested in or concerned about the pandemic. Third, it is relevant to note that social desirability may bias the present findings.

At the same time, several recent studies have assessed whether or not participants over-report compliance when asked about protective behavior in the context of the COVID-19 pandemic.

Most of these studies ﬁnd no evidence to suggest that self-reports of protective behavior are tainted by social desirability bias (Larsen et al. 2020; Munzert and Selb 2020; Galasso et al.

2020). Furthermore, the exceptional high levels of reported avoidant behavior are consistent with studies from United States using actual behavioral measures (Lee et al. 2020), compliance is exceptionally high when it comes to avoidant behavior. These observations suggest that social desirability may not be a major factor in the context of the present studies.7

7Nonetheless, in the supplementary section B.8, we seek to directly limit any confounding from social desirability bias using the personality trait of agreeableness. The personality trait of agreeableness is a trait that is frequently discussed as a potential target of social desirability bias (Graziano and Tobin 2002). Independently

17 Conclusion

The present findings demonstrate that self-efficacy was both necessary and sufficient for protective behavior during the first wave of the coronavirus pandemic and constitutes a pathway to compliance with pandemic health advice not driven by personal fear. Seemingly, the salience of the emergency created an unprecedented motivation to obtain and act on health advice, while ignoring other common psychological considerations including those related to fear as well as trust in fellow citizens and political institutions. Accordingly, the findings suggest that a major focus during the current and future pandemics should be on providing clear information about protective behavior and formulating guidelines that facilitate a sense of self-efficacy in the public. To facilitate society-wide protective behavior during a massive crisis such as the

ﬁrst wave of COVID-19 pandemic, the establishment of a strong sense of self-eﬃcacy is key.

of the larger research question of how much agreeableness is tainted by social desirability, it is clear that people who report that they lack agreeableness are willing to disclose socially undesirable traits. We make use of this fact and subset our analyses to the bottom half of the agreeableness scale and ﬁnd that results replicate among individuals who are not unwilling to disclose undesirable behaviors and attitudes.

18 References

Aarøe, L., Osmundsen, M., and Petersen, M. B. (2016). Distrust as a disease avoidance strategy:

Individual diﬀerences in disgust sensitivity regulate generalized social trust. Frontiers in

psychology, 7:1038.

Amidon, T. R., Nielsen, A. C., Pﬂugfelder, E. H., Richards, D. P., and Stephens, S. H. (2021).

Visual risk literacy in “ﬂatten the curve” covid-19 visualizations. Journal of Business and

Technical Communication, 35(1):101–109.

Anderson, R. M., Heesterbeek, H., Klinkenberg, D., and Hollingsworth, T. D. (2020). How

will country-based mitigation measures inﬂuence the course of the covid-19 epidemic? The

Lancet, 395(10228):931–934.

Angrist, J. D. and Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist’s

Companion. Princeton University Press.

Bandura, A. (1982). Self-eﬃcacy mechanism in human agency. American psychologist,

37(2):122–147.

Bish, A. and Michie, S. (2010). Demographic and attitudinal determinants of protective be-

haviours during a pandemic: a review. British Journal of Health Psychology, 15(4):797–824.

Bonell, C., Michie, S., Reicher, S., West, R., Bear, L., Yardley, L., Curtis, V., Amlˆot,R.,

and Rubin, G. J. (2020). Harnessing behavioural science in public health campaigns to

maintain ‘social distancing’in response to the covid-19 pandemic: key principles. Journal of

Epidemiology and Community Health.

Brodeur, A., Grigoryeva, I., and Kattan, L. (2020). Stay-at-home orders, social distancing and

trust.

Brouard, S., Vasilopoulos, P., and Becher, M. (2020). Sociodemographic and psychological

correlates of compliance with the covid-19 public health measures in france. Canadian Journal

of Political Science/Revue canadienne de science politique, pages 1–8.

Brug, J., Aro, A. R., Oenema, A., De Zwart, O., Richardus, J. H., and Bishop, G. D. (2004). Sars

19 risk perception, knowledge, precautions, and information sources, the netherlands. Emerging

infectious diseases, 10(8):1486.

Brug, J., Aro, A. R., and Richardus, J. H. (2009). Risk perceptions and behaviour: towards pan-

demic control of emerging infectious diseases. International Journal of Behavioral Medicine,

16(3).

Durante, R., Gulino, G., et al. (2020). Asocial capital: Civic culture and social distancing

during covid-19.

Fong, E. and Chang, L.-y. (2011). Community under stress: Trust, reciprocity, and community

collective eﬃcacy during sars outbreak. Journal of community health, 36(5):797–810.

Galasso, V., Pons, V., Profeta, P., Becher, M., Brouard, S., and Foucault, M. (2020). Gender

diﬀerences in covid-19 attitudes and behavior: Panel evidence from eight countries. Proceed-

ings of the National Academy of Sciences, 117(44):27285–27291.

Gollwitzer, A., Martel, C., Brady, W. J., P¨arnamets,P., Freedman, I. G., Knowles, E. D.,

and Van Bavel, J. J. (2020). Partisan diﬀerences in physical distancing are linked to health

outcomes during the covid-19 pandemic. Nature human behaviour, 4(11):1186–1197.

Graziano, W. G. and Tobin, R. M. (2002). Agreeableness: Dimension of personality or social

desirability artifact? Journal of personality, 70(5):695–728.

Hainmueller, J., Mummolo, J., and Xu, Y. (2019). How much should we trust estimates from

multiplicative interaction models? simple tools to improve empirical practice. Political Anal-

ysis, 27(2):163–192.

Hale, T., Webster, S., Petherick, A., Phillips, T., and Kira, B. (2020). Oxford covid-19 govern-

ment response tracker. Blavatnik School of Government, 25.

Harper, C. A., Satchell, L. P., Fido, D., and Latzman, R. D. (2020). Functional fear predicts

public health compliance in the covid-19 pandemic. International Journal of Mental Health

and Addiction, page 1.

20 Johnson, T., Dawes, C., Fowler, J., and Smirnov, O. (2020). Slowing COVID-19 transmis-

sion as a social dilemma: Lessons for government oﬃcials from interdisciplinary research on

cooperation. Journal of Behavioral Public Administration.

Jordan, R. E., Adab, P., and Cheng, K. K. (2020). Covid-19: risk factors for severe disease and

death. BMJ, 368.

Kachanoﬀ, F., Bigman, Y., Kapsaskis, K., and Gray, K. (2020). Measuring two distinct psy-

chological threats of covid-19 and their unique impacts on wellbeing and adherence to public

health behaviors. Preprint at PsyArXiv.

Larsen, M., Nyrup, J., and Petersen, M. B. (2020). Do survey estimates of the public’s com-

pliance with covid-19 regulations suﬀer from social desirability bias? Journal of Behavioral

Public Administration, 3(2).

Lear, D. (1995). Sexual communication in the age of aids: The construction of risk and trust

among young adults. Social Science & Medicine, 41(9):1311–1323.

Lee, M., Zhao, J., Sun, Q., Pan, Y., Zhou, W., Xiong, C., and Zhang, L. (2020). Human

mobility trends during the early stage of the covid-19 pandemic in the united states. PLoS

One, 15(11):e0241468.

Leung, G. M., Ho, L.-M., Chan, S. K., Ho, S.-Y., Bacon-Shone, J., Choy, R. Y., Hedley, A. J.,

Lam, T.-H., and Fielding, R. (2005). Longitudinal assessment of community psychobehavioral

responses during and after the 2003 outbreak of severe acute respiratory syndrome in hong

kong. Clinical Infectious Diseases, 40(12):1713–1720.

Lewnard, J. A. and Lo, N. C. (2020). Scientiﬁc and ethical basis for social-distancing interven-

tions against covid-19. The Lancet. Infectious diseases.

Maddux, J. E. and Rogers, R. W. (1983). Protection motivation and self-eﬃcacy: A revised the-

ory of fear appeals and attitude change. Journal of Experimental Social Psychology, 19(5):469–

479.

Marcus, G. E., Theiss-Morse, E., Sullivan, J. L., and Wood, S. L. (1995). With malice toward

some: How people make civil liberties judgments. Cambridge University Press.

21 Munzert, S. and Selb, P. (2020). Can we directly survey adherence to non-pharmaceutical

interventions? Survey Research Methods, 14(2):205–209.

Ornell, F., Schuch, J. B., Sordi, A. O., and Kessler, F. H. P. (2020). “pandemic fear” and

covid-19: mental health burden and strategies. Brazilian Journal of Psychiatry.

Ostrom, E. (1998). A behavioral approach to the rational choice theory of collective action:

Presidential address, american political science association, 1997. American Political Science

Review, 92(1):1–22.

Pearce, J. M., Lindekilde, L., Parker, D., and Rogers, M. B. (2019). Communicating with

the public about marauding terrorist ﬁrearms attacks: Results from a survey experiment on

factors inﬂuencing intention to “run, hide, tell” in the united kingdom and denmark. Risk

analysis, 39(8):1675–1694.

Rippetoe, P. A. and Rogers, R. W. (1987). Eﬀects of components of protection-motivation

theory on adaptive and maladaptive coping with a health threat. Journal of Personality and

Social Psychology, 52(3):596.

Rogers, R. W. (1975). A protection motivation theory of fear appeals and attitude change. The

journal of psychology, 91(1):93–114.

Rubin, G. J., Amlˆot,R., Page, L., and Wessely, S. (2009). Public perceptions, anxiety, and

behaviour change in relation to the swine ﬂu outbreak: cross sectional telephone survey. BMJ,

339:b2651.

Ruiter, R. A., Abraham, C., and Kok, G. (2001). Scary warnings and rational precautions: A

review of the psychology of fear appeals. Psychology and health, 16(6):613–630.

Sheeran, P., Harris, P. R., and Epton, T. (2014). Does heightening risk appraisals change

people’s intentions and behavior? a meta-analysis of experimental studies. Psychological

bulletin, 140(2):511.

Teasdale, E., Yardley, L., Schlotz, W., and Michie, S. (2012). The importance of coping appraisal

in behavioural responses to pandemic ﬂu. British journal of health psychology, 17(1):44–59.

22 Tyler, T. R. and Jackson, J. (2014). Popular legitimacy and the exercise of legal authority:

Motivating compliance, cooperation, and engagement. Psychology, Public Policy, and Law,

20(1):78.

Van Bavel, J. J., Baicker, K., Boggio, P. S., Capraro, V., Cichocka, A., Cikara, M., Crockett,

M. J., Crum, A. J., Douglas, K. M., Druckman, J. N., et al. (2020). Using social and

behavioural science to support covid-19 pandemic response. Nature Human Behaviour, pages

1–12.

West, R., Michie, S., Rubin, G. J., and Amlˆot,R. (2020). Applying principles of behaviour

change to reduce sars-cov-2 transmission. Nature Human Behaviour, pages 1–9.

Wolfers, J. (2006). Did unilateral divorce laws raise divorce rates? a reconciliation and new

results. American Economic Review, 96(5):1802–1820.

23 Supplementary Materials

Contents

A Materials and Methods 25

A.1 Data collection ...... 25

A.2 Population and sample characteristics ...... 27

A.3 Descriptive statistics ...... 36

B Supporting Results 45

B.1 Correlations between protective behavior and demographics ...... 45

B.2 Full model ...... 46

B.3 Self-eﬃcacy: Capability dimension ...... 50

B.4 Behavior change as alternative outcome ...... 54

B.5 Robustness of moderations ...... 58

B.6 Individual fixed effects linear regression analyses: Effects of efficacy ...... 60

B.7 Societal threat ...... 64

B.8 Overall correlations, by agreeablenss ...... 65

24 A Materials and Methods

A.1 Data collection

In each of the eight countries in our sample, the Danish survey ﬁrm Epinion sampled adult respondents using online panels from panel suppliers. In Denmark, Norstat supplied the online panels, while CINT supplied the online panels in the remaining seven countries. The panel suppliers, recruit new panelists for their online panels mostly through online channels, but also to a lesser extent, via telephone interviews. Panelists are then invited to participate in surveys. Participation is compensated via lotteries for gift certiﬁcates. After answering a survey, respondents are ’quarantined’ for eight days. This means that respondents, once their embargo period has ended become eligible to be invited again. We exploit this for our panel sample.

The collection of Danish data started March 13, while starting dates varied for the remaining countries (see Table SM1, below). The variables included in our cross-sectional sample were collected between March 19 and April 3. Data included in our panel sampled was collected between March 13 and May 16. For the the cross-sectional sample, we sampled about about

500 Danes every day and about 250 respondents from each of the remaining seven countries.

Our panel sample consists of all respondents who have completed our rolling survey at multiple time points. On the one hand, this means that respondents are randomly selected by the panel providers to be invited again, there are no selection on any observed “wave 1” response. At the same time, respondents from smaller panel pools, and from underrepresented demographic groups are potentially over represented in our panel sample. Indeed, almost half of all respondents in the panel sample come from Denmark, where we interviewed twice as many respondents on a daily basis, relying on a somewhat smaller online panel (given that Denmark is a country with only 5.8 Million citizens).

Survey respondents were quota sampled to match the population margins on age, gender, and geographic location for each of the eight countries in our study (Section A.2 below compares the sample and census characteristics for each country). In our study, the median interview length, across all countries, was 8.75 minutes. Among the panelists invited to take our survey, the response rate (calculated as the fraction of complete responses over invited, eligible participants)

25 across the countries in our sample was between 18% (Hungary) and 38 % (Denmark).8 The survey was conducted in line with the national ethical guidelines for conducting survey-based research involving human subjects. Informed consent was obtained from each participant at the beginning of the survey.

Table SM1 gives an overview of the data collection process, including starting dates and overall sample sizes for each country.

Table SM1: Overview of data collection

Cross-sectional sample Panel sample Country Dates Obs. Dates Obs. Individ. Denmark March 19 - April 3 7,391 March 13 - May 16 11,831 5,136 Sweden March 21 - April 3 3,025 March 21 - May 16 2,032 862 Germany March 24 - April 3 2,236 March 24 - May 16 2,304 949 France March 24 - April 3 2,319 March 24 - May 16 1,965 860 Italy March 21 - April 3 3,156 March 21 - May 16 2,433 1,000 Hungary March 24 - April 3 2,288 March 24 - May 16 2,996 1,215 United Kingdom March 21 - April 3 3,039 March 24 - May 16 2,052 904 United States March 21 - April 3 3,054 March 24 - May 16 1,072 503

8Response rates are calculated for the period between March 13 and April 2.

26 A.2 Population and sample characteristics

In Tables SM2-SM9, we compare the population and sample characteristics of each country.

Similar to most surveys based on Internet panels, our samples from some of the countries included in our study are skewed towards more educated and younger eligible voters compared to the overall population of eligible voters. In contrast, the samples are overall well balanced on sex. In our models in the main text, we address these imbalances by controlling for a battery of covariates that match these imbalances. Note, that we cannot obtain valid census data for the share of potential voters that did not vote in all countries. Therefore, we impute the proportion who did not vote from the proportion in each sample and scale the remaining party choice values accordingly such that the variable sums to 1. In the tables below, we report the scaled proportions in parentheses).

27 Table SM2: Population and sample characteristics, Denmark

Census Sample Sex and Age Male 18-34 years 0.13 0.11 Male 35-55 years 0.19 0.14 Male 56+ years 0.17 0.21 Female 18-34 years 0.13 0.18 Female 35-55 years 0.19 0.18 Female 56+ years 0.19 0.19 Geography Nordjylland 0.10 0.09 Midtjylland 0.22 0.25 Syddanmark 0.21 0.20 Hovedstaden 0.32 0.34 Sjælland 0.15 0.12 Education ISCED Lv0-4 0.59 0.35 ISCED Lv5-8 0.41 0.65 Vote choice Socialdemokratiet 0.26 (0.22) 0.25 Radikale 0.09 (0.07) 0.07 Konservative 0.07 (0.06) 0.06 Nye Borgerlige 0.02 (0.02) 0.02 Socialistisk Folkeparti 0.08 (0.07) 0.08 Liberal Alliance 0.02 (0.02) 0.02 Danske Folkeparty 0.09 (0.07) 0.07 Venstre 0.23 (0.20) 0.16 Enhedslisten 0.07 (0.06) 0.08 Alternativet 0.03 (0.03) 0.02 Other 0.04 (0.04) 0.03 Did not vote NA (0.15) 0.15

28 Table SM3: Population and sample characteristics, Sweden

Census Sample Sex and Age Male 18-34 years 0.14 0.12 Male 35-55 years 0.17 0.18 Male 56+ years 0.18 0.19 Female 18-34 years 0.13 0.17 Female 35-55 years 0.17 0.18 Female 56+ years 0.20 0.16 Geography Ostra¨ Sverige 0.40 0.30 SödraSverige 0.43 0.40 Norra Sverige 0.17 0.30 Education ISCED Lv0-4 0.63 0.62 ISCED Lv5-8 0.37 0.38 Vote choice Centerpartiet 0.09 (0.07) 0.05 Kristendemokraterna 0.06 (0.05) 0.05 Liberalerna 0.05 (0.05) 0.04 Moderaterna 0.20 (0.16) 0.12 Miljöpartiet 0.04 (0.04) 0.03 Socialdemokraterna 0.28 (0.23) 0.24 Sverigedemokraterna 0.18 (0.14) 0.18 Vänsterpartiet 0.08 (0.07) 0.09 Other 0.02 ((0.01) 0.02 Did not vote NA (0.18) 0.18

29 Table SM4: Population and sample characteristics, Germany

Census Sample Sex and Age Male 18-34 years 0.13 0.09 Male 35-55 years 0.18 0.22 Male 56+ years 0.19 0.19 Female 18-34 years 0.12 0.13 Female 35-55 years 0.17 0.19 Female 56+ years 0.22 0.18 Geography Baden-Württemberg 0.13 0.13 Bayern 0.16 0.15 Berlin 0.04 0.03 Brandenburg 0.03 0.03 Bremen 0.01 0.01 Hamburg 0.02 0.03 Hessen 0.08 0.07 Mecklenburg-Vorpommern 0.02 0.02 Niederscahsen 0.10 0.10 Nordrhein-Westfalen 0.22 0.22 Rheinland-Pfalz 0.05 0.05 Sachsen 0.05 0.06 Sachsen-Anhalt 0.03 0.03 Schleswig-Holstein 0.03 0.03 Saarland 0.01 0.01 Thüringen 0.03 0.03 Education ISCED Lv0-4 0.71 0.59 ISCED Lv5-8 0.29 0.41 Vote choice CDU/CSU 0.37 (0.28) 0.19 SPD 0.25 (0.19) 0.13 AfD 0.11 (0.09) 0.10 FDP 0.07 (0.05) 0.06 Die Linke 0.09 (0.07) 0.09 Grüne 0.08 (0.06) 0.15 Other 0.03 (0.02) 0.05 Did not vote NA (0.24) 0.24

30 Table SM5: Population and sample characteristics, France

Census Sample Sex and Age Male 18-34 years 0.13 0.11 Male 35-55 years 0.17 0.19 Male 56+ years 0.18 0.18 Female 18-34 years 0.13 0.16 Female 35-55 years 0.17 0.20 Female 56+ years 0.22 0.16 Geography Auvergne Rhône Alpes 0.12 0.12 Bourgogne France-Comtéand Grand Est 0.12 0.13 Bretagne and Normandie 0.10 0.11 Centre-Val de Loire and Pays de la Loire 0.09 0.10 Hauts-de-Freance 0.09 0.09 Île-de-France 0.18 0.18 Nouvelle-Aquitaine 0.13 0.09 Occitanie 0.09 0.09 Provence-Alpes-Côted’Azu 0.08 0.07 Education ISCED Lv0-4 0.67 0.59 ISCED Lv5-8 0.33 0.41 Vote choice Dupont-Aignan 0.05 (0.03) 0.03 Fillon 0.20 (0.14) 0.08 Hamon 0.06 (0.04) 0.06 LE Pen 0.21 (0.15) 0.17 Macron 0.24 (0.17) 0.21 Melenchon 0.20 (0.13) 0.10 Other 0.04 (0.03) 0.04 Did not vote NA (0.31) 0.31

31 Table SM6: Population and sample characteristics, Italy

Census Sample Sex and Age Male 18-34 years 0.11 0.12 Male 35-55 years 0.19 0.26 Male 56+ years 0.19 0.12 Female 18-34 years 0.10 0.15 Female 35-55 years 0.19 0.25 Female 56+ years 0.23 0.11 Geography Nortwest Italy 0.27 0.26 Norteast Italy 0.19 0.19 Central Italy 0.20 0.20 South Italy 0.23 0.23 Insular Italy 0.11 0.12 Education ISCED Lv0-4 0.83 0.66 ISCED Lv5-8 0.17 0.34 Vote choice Centre-Right 0.37 (0.27) 0.29 Five Star Movement 0.33 (0.24) 0.24 Centre-Left 0.23 (0.16) 0.15 Free and Equal 0.03 (0.02) 0.02 Other 0.04 (0.03) 0.02 Did not vote NA (0.28) 0.28

32 Table SM7: Population and sample characteristics, Hungary

Census Sample Sex and Age Male 18-34 years 0.13 0.14 Male 35-55 years 0.19 0.18 Male 56+ years 0.16 0.15 Female 18-34 years 0.12 0.14 Female 35-55 years 0.19 0.19 Female 56+ years 0.22 0.20 Geography Central Hungary 0.31 0.30 Transdanubia 0.30 0.30 Great Plain and North 0.39 0.40 Education ISCED Lv0-4 0.74 0.61 ISCED Lv5-8 0.26 0.39 Vote choice Fidesz KDNP 0.49 (0.28) 0.26 Jobbik 0.19 (0.11) 0.08 MSZP-PM 0.12 (0.07) 0.03 LMP 0.07 (0.04) 0.02 DK 0.05 (0.03) 0.09 MM 0.03 (0.02) 0.05 Other 0.04 (0.03) 0.04 Did not vote NA (0.42) 0.42

33 Table SM8: Population and sample characteristics, United Kingdom.

Census Sample Sex and Age Male 18-34 years 0.14 0.13 Male 35-55 years 0.17 0.24 Male 56+ years 0.17 0.12 Female 18-34 years 0.14 0.20 Female 35-55 years 0.18 0.20 Female 56+ years 0.19 0.11 Geography North East 0.04 0.05 North West 0.11 0.12 Yorkshire and the Humber 0.08 0.08 East Midlands 0.07 0.08 West Midlands 0.09 0.09 East 0.09 0.10 London 0.13 0.10 South East 0.14 0.14 South West 0.08 0.08 Wales 0.05 0.05 Scotland 0.08 0.10 Northern Ireland 0.03 0.02 Education ISCED Lv0-4 0.61 0.52 ISCED Lv5-8 0.39 0.48 Vote choice Conservative 0.44 (0.37) 0.35 Labour 0.32 (0.27) 0.32 Liberal Democrats 0.12 (0.10) 0.09 SNP 0.04(0.03) 0.03 Other 0.09 (0.08) 0.05 Did not vote NA (0.16) 0.16

34 Table SM9: Population and sample characteristics, USA

Census Sample Sex and Age Male 18-34 years 0.15 0.15 Male 35-55 years 0.17 0.24 Male 56+ years 0.16 0.10 Female 18-34 years 0.15 0.18 Female 35-55 years 0.17 0.22 Female 56+ years 0.19 0.11 Geography Northeast 0.17 0.20 Midwest 0.21 0.23 West 0.24 0.21 South 0.38 0.36 Education ISCED Lv0-4 0.42 0.28 ISCED Lv5-8 0.58 0.72 Vote choice Republican 0.46 (0.34) 0.32 Democrats 0.48 (0.35) 0.33 Other 0.06 (0.04) 0.08 Did not vote NA (0.27) 0.27

35 A.3 Descriptive statistics

Table SM10: Descriptive statistics

Mean SD Min Max N Avoidant behavior 0.92 0.13 0.00 1.00 26504 Preventive behavior 0.59 0.44 0.00 1.00 26508 Psychological correlates Worry 0.66 0.25 0.00 1.00 26506 Self-eﬃcacy 0.80 0.17 0.00 1.00 26298 Interpersonal trust 0.49 0.28 0.00 1.00 26507 Institutional trust 0.63 0.29 0.00 1.00 26507 Demographics Sex (female) 0.52 0.50 0.00 1.00 26508 Age 46.05 16.18 18.00 99.00 26283 ISCED Lv0-4 0.49 0.50 0.00 1.00 26508 ISCED Lv5-8 0.51 0.50 0.00 1.00 26508 Income 0.30 0.21 0.00 1.00 23936 Employed 0.56 0.50 0.00 1.00 26508 Under education 0.08 0.27 0.00 1.00 26508 Outside the labor market 0.15 0.36 0.00 1.00 26508 Retired 0.20 0.40 0.00 1.00 26508 Right 0.37 0.48 0.00 1.00 26508 Left 0.37 0.48 0.00 1.00 26508 Did not vote 0.26 0.44 0.00 1.00 26508 Single 0.34 0.47 0.00 1.00 26508 In relationship 0.23 0.42 0.00 1.00 26508 Married 0.43 0.50 0.00 1.00 26508 No children 0.51 0.50 0.00 1.00 26508 Children 0.49 0.50 0.00 1.00 26508

36 Table SM11: Descriptive statistics, Denmark

Mean SD Min Max N Avoidant behavior 0.95 0.10 0.17 1.00 7390 Preventive behavior 0.71 0.40 0.00 1.00 7391 Psychological correlates Worry 0.62 0.25 0.00 1.00 7391 Self-eﬃcacy 0.88 0.13 0.05 1.00 7333 Interpersonal trust 0.61 0.26 0.00 1.00 7391 Institutional trust 0.79 0.21 0.00 1.00 7391 Demographics Sex (female) 0.54 0.50 0.00 1.00 7391 Age 48.71 18.34 18.00 92.00 7320 Income 0.35 0.24 0.00 1.00 6189 ISCED Lv0-4 0.34 0.47 0.00 1.00 7391 ISCED Lv5-8 0.66 0.47 0.00 1.00 7391 Employed 0.50 0.50 0.00 1.00 7391 Under education 0.13 0.33 0.00 1.00 7391 Outside the labor market 0.09 0.29 0.00 1.00 7391 Retired 0.29 0.45 0.00 1.00 7391 Right 0.35 0.48 0.00 1.00 7391 Left 0.50 0.50 0.00 1.00 7391 Did not vote 0.15 0.36 0.00 1.00 7391 Single 0.32 0.47 0.00 1.00 7391 In relationship 0.24 0.43 0.00 1.00 7391 Married 0.44 0.50 0.00 1.00 7391 No children 0.51 0.50 0.00 1.00 7391 Children 0.49 0.50 0.00 1.00 7391

37 Table SM12: Descriptive statistics, Sweden

Mean SD Min Max N Avoidant behavior 0.85 0.18 0.00 1.00 3023 Preventive behavior 0.54 0.44 0.00 1.00 3025 Psychological correlates Worry 0.61 0.25 0.00 1.00 3025 Self-eﬃcacy 0.81 0.16 0.00 1.00 2991 Interpersonal trust 0.50 0.27 0.00 1.00 3025 Institutional trust 0.56 0.30 0.00 1.00 3025 Demographics Sex (female) 0.51 0.50 0.00 1.00 3025 Age 47.12 16.97 18.00 86.00 3007 Income 0.33 0.21 0.00 0.64 2666 ISCED Lv0-4 0.62 0.49 0.00 1.00 3025 ISCED Lv5-8 0.38 0.49 0.00 1.00 3025 Employed 0.54 0.50 0.00 1.00 3025 Under education 0.09 0.28 0.00 1.00 3025 Outside the labor market 0.12 0.32 0.00 1.00 3025 Retired 0.25 0.44 0.00 1.00 3025 Right 0.43 0.50 0.00 1.00 3025 Left 0.36 0.48 0.00 1.00 3025 Did not vote 0.20 0.40 0.00 1.00 3025 Single 0.36 0.48 0.00 1.00 3025 In relationship 0.28 0.45 0.00 1.00 3025 Married 0.36 0.48 0.00 1.00 3025 No children 0.52 0.50 0.00 1.00 3025 Children 0.48 0.50 0.00 1.00 3025

38 Table SM13: Descriptive statistics, Germany

Mean SD Min Max N Avoidant behavior 0.94 0.12 0.17 1.00 2236 Preventive behavior 0.51 0.45 0.00 1.00 2236 Psychological correlates Worry 0.63 0.28 0.00 1.00 2235 Self-eﬃcacy 0.81 0.17 0.07 1.00 2222 Interpersonal trust 0.49 0.28 0.00 1.00 2236 Institutional trust 0.63 0.27 0.00 1.00 2236 Demographics Sex (female) 0.50 0.50 0.00 1.00 2236 Age 48.54 14.76 18.00 81.00 2231 Income 0.28 0.20 0.00 0.64 2096 ISCED Lv0-4 0.59 0.49 0.00 1.00 2236 ISCED Lv5-8 0.41 0.49 0.00 1.00 2236 Employed 0.58 0.49 0.00 1.00 2236 Under education 0.05 0.21 0.00 1.00 2236 Outside the labor market 0.13 0.33 0.00 1.00 2236 Retired 0.24 0.43 0.00 1.00 2236 Right 0.34 0.47 0.00 1.00 2236 Left 0.37 0.48 0.00 1.00 2236 Did not vote 0.29 0.45 0.00 1.00 2236 Single 0.35 0.48 0.00 1.00 2236 In relationship 0.18 0.38 0.00 1.00 2236 Married 0.47 0.50 0.00 1.00 2236 No children 0.63 0.48 0.00 1.00 2236 Children 0.37 0.48 0.00 1.00 2236

39 Table SM14: Descriptive statistics, France

Mean SD Min Max N Avoidant behavior 0.87 0.11 0.17 1.00 2319 Preventive behavior 0.51 0.46 0.00 1.00 2319 Psychological correlates Worry 0.68 0.26 0.00 1.00 2318 Self-eﬃcacy 0.65 0.21 0.00 1.00 2298 Interpersonal trust 0.39 0.28 0.00 1.00 2319 Institutional trust 0.51 0.30 0.00 1.00 2319 Demographics Sex (female) 0.52 0.50 0.00 1.00 2319 Age 46.22 15.32 18.00 99.00 2310 Income 0.29 0.21 0.00 0.64 2212 ISCED Lv0-4 0.59 0.49 0.00 1.00 2319 ISCED Lv5-8 0.41 0.49 0.00 1.00 2319 Employed 0.59 0.49 0.00 1.00 2319 Under education 0.05 0.21 0.00 1.00 2319 Outside the labor market 0.16 0.36 0.00 1.00 2319 Retired 0.20 0.40 0.00 1.00 2319 Right 0.28 0.45 0.00 1.00 2319 Left 0.36 0.48 0.00 1.00 2319 Did not vote 0.36 0.48 0.00 1.00 2319 Single 0.33 0.47 0.00 1.00 2319 In relationship 0.21 0.40 0.00 1.00 2319 Married 0.46 0.50 0.00 1.00 2319 No children 0.42 0.49 0.00 1.00 2319 Children 0.58 0.49 0.00 1.00 2319

40 Table SM15: Descriptive statistics, Italy

Mean SD Min Max N Avoidant behavior 0.96 0.10 0.11 1.00 3155 Preventive behavior 0.51 0.46 0.00 1.00 3156 Psychological correlates Worry 0.73 0.22 0.00 1.00 3156 Self-eﬃcacy 0.76 0.15 0.00 1.00 3140 Interpersonal trust 0.40 0.27 0.00 1.00 3155 Institutional trust 0.60 0.28 0.00 1.00 3156 Demographics Sex (female) 0.50 0.50 0.00 1.00 3156 Age 43.56 13.31 18.00 79.00 3139 Income 0.28 0.18 0.00 0.64 2855 ISCED Lv0-4 0.66 0.48 0.00 1.00 3156 ISCED Lv5-8 0.34 0.48 0.00 1.00 3156 Employed 0.60 0.49 0.00 1.00 3156 Under education 0.08 0.27 0.00 1.00 3156 Outside the labor market 0.24 0.43 0.00 1.00 3156 Retired 0.08 0.27 0.00 1.00 3156 Right 0.53 0.50 0.00 1.00 3156 Left 0.17 0.37 0.00 1.00 3156 Did not vote 0.30 0.46 0.00 1.00 3156 Single 0.28 0.45 0.00 1.00 3156 In relationship 0.22 0.42 0.00 1.00 3156 Married 0.50 0.50 0.00 1.00 3156 No children 0.46 0.50 0.00 1.00 3156 Children 0.54 0.50 0.00 1.00 3156

41 Table SM16: Descriptive statistics, Hungary

Mean SD Min Max N Avoidant behavior 0.93 0.12 0.22 1.00 2288 Preventive behavior 0.61 0.44 0.00 1.00 2288 Psychological correlates Worry 0.67 0.26 0.00 1.00 2288 Self-eﬃcacy 0.80 0.17 0.00 1.00 2280 Interpersonal trust 0.41 0.26 0.00 1.00 2288 Institutional trust 0.51 0.34 0.00 1.00 2288 Demographics Sex (female) 0.53 0.50 0.00 1.00 2288 Age 46.04 15.42 18.00 80.00 2249 Income 0.33 0.23 0.00 0.64 2110 ISCED Lv0-4 0.61 0.49 0.00 1.00 2288 ISCED Lv5-8 0.39 0.49 0.00 1.00 2288 Employed 0.54 0.50 0.00 1.00 2288 Under education 0.07 0.25 0.00 1.00 2288 Outside the labor market 0.17 0.37 0.00 1.00 2288 Retired 0.22 0.42 0.00 1.00 2288 Right 0.34 0.48 0.00 1.00 2288 Left 0.20 0.40 0.00 1.00 2288 Did not vote 0.46 0.50 0.00 1.00 2288 Single 0.31 0.46 0.00 1.00 2288 In relationship 0.29 0.45 0.00 1.00 2288 Married 0.41 0.49 0.00 1.00 2288 No children 0.46 0.50 0.00 1.00 2288 Children 0.54 0.50 0.00 1.00 2288

42 Table SM17: Descriptive statistics, United Kingdom

Mean SD Min Max N Avoidant behavior 0.94 0.13 0.00 1.00 3039 Preventive behavior 0.52 0.45 0.00 1.00 3039 Psychological correlates Worry 0.72 0.23 0.00 1.00 3039 Self-eﬃcacy 0.80 0.15 0.00 1.00 3023 Interpersonal trust 0.50 0.27 0.00 1.00 3039 Institutional trust 0.64 0.26 0.00 1.00 3038 Demographics Sex (female) 0.51 0.50 0.00 1.00 3039 Age 43.03 14.72 18.00 80.00 3025 Income 0.29 0.20 0.00 0.64 2872 ISCED Lv0-4 0.52 0.50 0.00 1.00 3039 ISCED Lv5-8 0.48 0.50 0.00 1.00 3039 Employed 0.67 0.47 0.00 1.00 3039 Under education 0.05 0.21 0.00 1.00 3039 Outside the labor market 0.17 0.37 0.00 1.00 3039 Retired 0.12 0.32 0.00 1.00 3039 Right 0.35 0.48 0.00 1.00 3039 Left 0.44 0.50 0.00 1.00 3039 Did not vote 0.21 0.41 0.00 1.00 3039 Single 0.33 0.47 0.00 1.00 3039 In relationship 0.23 0.42 0.00 1.00 3039 Married 0.43 0.50 0.00 1.00 3039 No children 0.49 0.50 0.00 1.00 3039 Children 0.51 0.50 0.00 1.00 3039

43 Table SM18: Descriptive statistics, USA

Mean SD Min Max N Avoidant behavior 0.91 0.16 0.00 1.00 3054 Preventive behavior 0.58 0.44 0.00 1.00 3054 Psychological correlates Worry 0.72 0.25 0.00 1.00 3054 Self-eﬃcacy 0.77 0.17 0.00 1.00 3011 Interpersonal trust 0.45 0.30 0.00 1.00 3054 Institutional trust 0.53 0.29 0.00 1.00 3054 Demographics Sex (female) 0.52 0.50 0.00 1.00 3054 Age 42.14 14.33 18.00 98.00 3002 Income 0.24 0.19 0.00 0.64 2936 ISCED Lv0-4 0.28 0.45 0.00 1.00 3054 ISCED Lv5-8 0.72 0.45 0.00 1.00 3054 Employed 0.58 0.49 0.00 1.00 3054 Under education 0.06 0.23 0.00 1.00 3054 Outside the labor market 0.25 0.43 0.00 1.00 3054 Retired 0.12 0.32 0.00 1.00 3054 Right 0.32 0.47 0.00 1.00 3054 Left 0.33 0.47 0.00 1.00 3054 Did not vote 0.35 0.48 0.00 1.00 3054 Single 0.44 0.50 0.00 1.00 3054 In relationship 0.16 0.37 0.00 1.00 3054 Married 0.40 0.49 0.00 1.00 3054 No children 0.55 0.50 0.00 1.00 3054 Children 0.45 0.50 0.00 1.00 3054

44 B Supporting Results

B.1 Correlations between protective behavior and demographics

Figure SM1: Correlations between protective behavior and demographics

Note: Filled blue circles (red triangles) show the estimated association between avoidant (preventive) behavior and our battery of demographics. Lines are the associated 95% conﬁdence intervals. Symbols without conﬁdence intervals are reference categories.

45 B.2 Full model

In this section, we replicate the main analyses, while including all psychological variables in the models at once (this implies that France drops out of these analyses as we do not observe our eﬃcacy measure in France). As can be seen below, our estimated associations are essentially similar to those of the main text. All French observations are left out because we do not observe knowledge eﬃcacy in France.

Figure SM2: Psychological correlates. Full model

Note: Correlations from full model that includes all psychological predictors at once. Filled blue circles (red triangles) show the estimated overall correlations between the each psychological predictor and avoidant behavior (preventive behavior). Error bars are 95% conﬁdence intervals.

46 Figure SM3: Country-speciﬁc deviations from the overall correlations between protective behavior and psychological correlate. Full model

Note: Correlations from full model that includes all psychological predictors at once. Filled blue circles (red triangles) show the estimated country-specific correlations between each psychological correlate and avoidant behavior (preventive behavior). Blue (red) dashed lines refer to the estimated overall associations in Figure SM2. Error bars are 95 % confidence intervals that show whether the country-specific correlations are statistically significantly different from overall associations.

47 Figure SM4: Moderations. Full model

Note: Solid black lines show predicted values at high levels of each moderator, black dashed lines show predicted values at medium levels, and solid grey lines show predicted values at low levels.

48 Figure SM5: Country-speciﬁc deviations from the overall estimated interactions. Full model

Note: Filled blue circles (red triangles) show the estimated country-specific moderations. Error bars are 95 % confidence intervals that show whether the country-specific moderations are statistically significantly different from overall moderations (see Figure SM4).

49 B.3 Self-eﬃcacy: Capability dimension

In this section, we rerun all the main analyses, while focusing only capability-dimension of the self-efficacy scale. That is, the questions that asks: To what extent do you agree or disagree with the following statement: I’m certain I can follow official advice to ”distance myself” from others if I want to. All results, replicate the self-efficacy results from the manuscript.

Figure SM6: Overall correlations. Capability-dimension of self-eﬃcacy

Note: Filled blue circles (red triangles) show the estimated overall correlations between the capability question andavoidant behavior (preventive behavior). Error bars are 95% conﬁdence intervals.

50 Figure SM7: Country-speciﬁc correlations. Capability-dimension of self-eﬃcacy

Note: Filled blue circles (red triangles) show the estimated country-specific correlations between the capability dimension and avoidant behavior (preventive behavior). Blue (red) dashed lines refer to the estimated overall associations in Figure SM6. Error bars are 95 % confidence intervals that show whether the country-specific correlations are statistically significantly different from overall associations.

51 Figure SM8: Moderations. Capability-dimension of self-eﬃcacy

Note: Solid black lines show predicted values at high levels of capability, black dashed lines show predicted values at medium levels, and solid grey lines show predicted values at low levels.

52 Figure SM9: Country-speciﬁc moderations. Capability-dimension of self-eﬃcacy

53 B.4 Behavior change as alternative outcome

In this section, we rerun all the main analyses, while shifting the outcome from our protective behavior index to the the variable ’behavior change’. Behavior change is measured by the question that ask: ”to what degree do you feel that the current situation with the Corona virus has made you change your behaviour to avoid spreading infection? Respondent answered on a

4-point scale from ’to a high degree’ to ’not at all’. We rescale this alternative outcome from

0-1. As can be seen below, our estimated associations are essentially similar to those of the main text. Italian observations are left out because we do not observe the alternative outcome in Italy.

Figure SM10: Correlations between behavior change and psychological correlates. Alternative outcome

Note: Correlations from our benchmark model. Filled black circles show the estimated association between the behavior change outcome and each of our psychological variables. Lines are the associated 95% conﬁdence intervals. Results are essentially similar to those of Figure 2.

54 Figure SM11: Country-speciﬁc deviations from the overall correlations between protective behavior and psychological correlate. Alternative outcome

Note: Filled black circles show the estimated association between the behavior change outcome and each of our psychological variables. Lines are the associated 95 % conﬁdence intervals. The overall pattern of the results is essentially similar to those of Figure 3.

55 Figure SM12: Moderations. Alternative outcome

Note: Solid black lines show predicted values at high levels of each moderator, black dashed lines show predicted values at medium levels, and solid grey lines show predicted values at low levels.

56 Figure SM13: Country-speciﬁc deviations from the overall estimated interactions. Alternative outcome

Note: Filled black circles show the estimated by country moderations. Lines are the associated 95% conﬁdence intervals. Dashed lines are the estimated overall correlations. Results are relatively similar to those of Figure 5.

57 B.5 Robustness of moderations

Figure SM14: Are the moderations linear? Binned estimator

Note: Solid lines display the linear interactions. Filled red circles the binned estimators from using interﬂex package (Hainmueller et al. 2019). The ﬁgure shows that the results of (Figure 4) are robust.

58 Figure SM15: Are the moderations linear? Kernel estimator

Note: Solid lines display the marginal effects estimated by the kernel smoother from using interflex package (Hainmueller et al. 2019). Grey filled area is the 95 % confidence intervals. The results of the figure are consistent with those of (Figure 4).

59 B.6 Individual fixed effects linear regression analyses: Effects of efficacy

Our panel sample consists of 10,569 individuals who were observed more than once in the period between March 13 and May 16. In total, the sample consists of 24,720 observations.

Unfortunately, we do not observe all measures from the cross-sectional sample in this period.

However, we do observe key measures that allow us to get causal leverage on the estimate of eﬃcacy on protective behavior by utilizing this temporal component. In particular, we have one repeated measure of eﬃcacy available: “To what degree do you feel that you know enough about what you as a citizen should do in relation to the Corona virus?”. Respondents answered this question on a 4-point scale from ”not at all” to ”to a high degree”.

On protective behavior, the following questions are available for the entire period: The question about avoidance of crowds (”Were you in a room with more than 10 people yesterday?”), the question about hand hygiene (”How many times do you estimate that you washed your hands or used hand sanitiser yesterday?”), and the alternative outcome that directly assess whether respondents feel they have changed behavior in order to avoid spreading the infection

(”To what degree do you feel that the current situation with the Corona virus has made you change your behaviour to avoid spreading infection?”). Crowd avoidance is coded 1 if respondents indicated that they were in a room with 10 or more people yesterday, and 0 if they were not. The hand hygiene question is coded 1 if respondents indicated that they washed their hand 10 times or more, and 0 if they washed their hands less. The behavior change questions was coded 1 if respondents answered ”to a high degree” and 0 otherwise. We summarize these three measures into a modiﬁed index of protective behavior scaled to range from 0-1, where high values indicate behavior that is compliant with protective advice.

To assess the causal eﬀect of self-eﬃcacy on protective behavior, we utilize the panel sample.

The first column of Table SM19 shows the estimated effect of efficacy on protective behavior using the two-way fixed effects estimator. As the column shows, there is a substantial effect of about 5.5 percentage points (p < 0.0001). Note that this estimate reflects moving the full range on efficacy. If we instead focus on the within-individual standard deviation on efficacy, it yields an effect size of about a ½ percentage points increase in compliance with public COVID-19 health advice as knowledge efficacy increases by one standard deviation (which corresponds to an 11 percentage points increase in efficacy).

60 Table SM19: Individual fixed effects linear regression analyses: Effects of efficacy on modified protective behavior index

Two-way FE model Lead model Individual-specific time trend 0.0551*** 0.0915*** 0.0487 Efficacy (0.0116) (0.0189) (0.0295) -0.0056 Efficacy (lead) (0.0243) Individual fixed effects Yes Yes Yes Data round fixed effects Yes Yes Yes Individual linear trends No No Yes Observations 24,720 6,430 24,720 Individuals 10,569 2,848 10,569 Note: Unstandardized regression coefficients from two-way fixed effects analyses. Standard errors are two-way clustered by individual and time (in parentheses). *** p < 0.001.

The two-way fixed effects estimator gives an unbiased estimate of the causal impact of self- efficacy on protective behavior on the assumption that the protective behavior of individuals had followed parallel trends in the absence of changes in efficacy (Angrist and Pischke 2008).

In other words: Absent a change in self-efficacy, all individuals would have experienced similar developments in protective behavior. The primary way in which this assumption can be violated is reverse causality (or simultaneity), causing bias in the estimated double differences between those who experience large changes in self-efficacy and those who experience little or no change.

In this application, the concern is that individuals who changed their behavior as a consequence of this change come to feel a sudden larger degree of self-efficacy. To test the robustness of the parallel trends assumption, we run models with a lead on the effect of self-efficacy (see the second column of Table SM19). Crucially for the plausibility of the parallel trends assumption, the estimated coefficient on the lead is very close to 0 and far from conventional levels of statistical significance.

Another common way to gauge the plausibility of the parallel trends assumption is to include unit-speciﬁc time trends (i.e., interactions with individuals and time). Including such time trends is in part problematic because some of the information used in the estimation of the unit-speciﬁc trends is post-treatment and thus potentially lead to post-treatment bias in the estimate (see, e.g., Wolfers (2006) for a detailed discussion of this issue). However, the results are informative because they can indicate whether the parallel trends assumption is plausible

61 or not. If the estimated eﬀect remains similar after inclusion of the individual-speciﬁc trends, it corroborates the plausibility of the parallel trends assumption. The third column of Table

SM19 shows an effect estimate of about 5 percentage when including the individual-specific time trends. Although this the effect estimate is statistically indistinguishable from 0, it remains substantively similar to the effect estimate in column 1.

Taken together, these findings provide evidence that individuals did not begin increasing their level of compliance in advance of a change in self-efficacy and thus corroborate the parallel trends assumption that underpins the causal interpretation of the estimated effect. Furthermore,

Figure SM16 shows the estimated effects of efficacy on protective behavior, both when pooling all country samples and for each country, respectively. Overall, Figure SM16 suggests that self- efficacy causally influences protective behavior in relatively similar ways across the countries.

Denmark provides a lower bound on the eﬀect (βDenmark = 0.04, p = 0.065), while the US is an upper bound (βUSA = 0.12, p = 0.025).

62 Figure SM16: Individual fixed effects linear regression analyses: Effects of efficacy on protective behavior index

Note: Filled black circles show the within estimates. Thin lines are the associated 95% conﬁdence intervals. Thick lines are the associated 90% conﬁdence intervals. Standard errors are two-way clustered on individuals and time.

63 B.7 Societal threat

In this section, we show the robustness of our ﬁndings from Figure 4, when changing the focus from feelings of personal threat to feelings of societal threat. In particular, we exploit that our survey holds the question ”To what degree are you concerned about the consequences of the Corona virus . . . for your country”. Respondents answered this question on a scale ranging from 1 ”not at all” to 4 ”To a high degree”. We use this questions to tap into respondents’ degree of fear that comes from societal concerns.

Figure SM17: Moderations

Note: Solid black lines show predicted values at high levels of each moderator, black dashed lines show predicted values at medium levels, and solid grey lines show predicted values at low levels.

64 B.8 Overall correlations, by agreeablenss

In this section we replicate our results while splitting the sample based on into participants above and beyond the median of the agreeableness scale. The logic of these analyses is to distinguish between people who are comfortable and uncomfortable admitting that they do not comply. Empirically, we thereby show that conclusion are fundamentally similar between those who are most and least likely to disclose their actual behavior.

Figure SM18: Overall correlations, by agreeableness

Note: Filled blue circles (red triangles) show the estimated overall correlations between each of the psychological variables and each of the protective behavior index among participants above (below) the median on the agreeableness scale. Error bars are 95% conﬁdence intervals.