Stress and worry in the 2020 coronavirus pandemic: relationships to trust and compliance with preventive measures across 48 countries in the COVIDiSTRESS global survey

1. Background and research questions

This registered report (RR) presents the COVIDiSTRESS global survey, which collects early data on human responses to the 2020 COVID-19 pandemic (caused by the coronavirus SARS-CoV-2) in 48 countries. This open science study was co-designed by researchers from numerous universities across the world (for a full list of universities, see §2; for a full list of consortium co-authors, see electronic supplementary material, co-authorship statement and appendix) to investigate how psychological and behavioural responses to the COVID-19 pandemic differed across countries and cultures, and how this impacted social behaviour, coping and trust in government efforts to slow the spread of the coronavirus.

Global data gathering efforts like the COVIDiSTRESS project are of pressing importance during the COVID-19 crisis, in order to understand how different government measures succeed or fail in achieving objectives to change public behaviour. By generating data rapidly, and disseminating them continuously as open data on the Open Science Framework (OSF), the international research network behind the COVIDiSTRESS project aims to yield rapid insights for the use of practitioners such as health communicators, government officials and policy makers at all levels, as well as members of the academic community. Currently, there are many psychologists and social scientists making a great effort to better understand the psychological effects and implications of the COVID-19 pandemic (e.g. [1–3]) and provide suggestions for practitioners (e.g. [4–6]). To our knowledge, there is currently no other project comparing human responses to COVID-19 across such a large number of countries. Further, the COVIDiSTRESS project specifically attempted to leverage the convenience of unpaid online populations to quickly acquire a large number of responses from 48 countries by using a snowball sample via online and traditional media platforms, rather than pre-specified samples from paid subject pools.

• 1. What is the impact of psychological stress on trust in and compliance with government efforts to prevent COVID-19 compared across countries?

• 2. To what degree is psychological stress related to concerns over COVID-19 itself, and what proportion of the variance in psychological stress can be explained by other factors?

This registered report (RR) analyses data from a swiftly executed global survey study meant to allow both hypothesis testing and exploratory cross-country comparisons. As an open science project, the COVIDiSTRESS effort as a whole intends to supply data to the public, allowing researchers and other stakeholders to independently conduct their own analyses [7]. As such, the present RR investigates 10 central hypotheses, focusing on people's experiences and behaviours during the early COVID-19 crisis. The present RR does not include a number of other variables from the COVIDiSTRESS global survey.

In early 2020, the COVID-19 pandemic quickly became a matter of serious concern for public health throughout the global community. To prevent the spread of coronavirus, governments imposed a range of measures, which included quarantines, physical distancing and limits to civil liberties. Inevitably, these changes have generated a variety of psychological responses [8], which in turn shape the level of compliance with preventive efforts. Extant research on the factors that shape willingness to comply with public health efforts has highlighted the importance of psychological responses such as anxiety and risk aversion [9–11] as well as trust in state authorities [12–14]. A recent study on risk and compliance during the COVID-19 pandemic in Singapore found that high levels of trust in the government led to lower risk perception and non-compliance with preventive measures [11]. Along the same lines, a study on the implications of COVID-19 in New Zealand found an increase in levels of institutional trust in the early stages of the lockdown pointing to the rally-round-the-flag effect [2]. This varied body of findings clearly signifies the urgent need for further exploration of the psychological mechanisms at play that shape responses to preventive measures.

Not only do pandemics foster a fear of contagion, but also misinformation, uncertainty and the lack of clarity about how to react [15,16]. Concern over a disease outbreak has typically been found to be positively correlated with preventive behaviours or compliance with health guidelines [17–19]. However, this relationship is complex; worry, for instance, has been found to be related to trust in disease-related information and media [20–22], the perception of openness and reliability of governments and health organizations [23].

Concern over one's medical risk during a pandemic can also be a source of ongoing stress (e.g. in H1N1, [24]; in MERS, [22]). For instance, stress was widespread during the 2003 SARS epidemic [25,26]. This relationship is potentially exacerbated by the fact that even some of the most efficient methods of slowing the spread of a disease, such as self-isolation and quarantine, also take an immediate and potentially long-lasting psychological toll among affected populations [8,27,28]. Surveys of quarantined healthcare workers [29] and members of the general population [30] following the 2003 SARS outbreak found evidence of acute stress responses in cases with a longer duration of quarantine and greater perceived difficulty in compliance with recommendations. Similarly, studies of patient groups typically find a negative relationship between experienced stress and compliance with prescribed behaviours, often with direct medical consequences [31–33], and this correlation has been extended to compliance with measures intended to minimize the spread of virulent disease [34].

Both the medical situation and the psychological effects of isolation and confinement [35,36] thus need to be taken into consideration when prolonged periods of quarantine are implemented, in order to understand the acceptance of government measures and compliance with preventive measures. Compliance with medical guidelines has been shown to decrease not just as a result of heightened stress levels [37], but also of minor everyday stressors such as workplace conflict or household responsibilities [38]. Efforts such as closing down schools and workplaces and calls for people to self-isolate in their homes are likely to constitute a source of both existential and practical stress, unrelated to the fear of contracting the disease. For instance, some stress is known to result from an inability to participate in work or from interpersonal problems in the family [39,40]. Moreover, a pandemic may also exacerbate existing financial precarity, generating another source of stress for those experiencing financial hardship [10]. Prolonged states of emergency and the chronic psychological, social and economic stressors related to them [8,27] may decrease compliance with behavioural objectives during pandemics. Sources of stress and worry may thus have direct implications not simply for well-being, but also for the effectiveness of collective efforts and the degree of public compliance with self-isolation measures. For this reason, we tested the hypotheses that stress and concern over the coronavirus will be adversely correlated with compliance and preventive behavioural measures suggested in each country.

In sum, it seems that worry, trust and situational stress are associated with each other and jointly influence behaviour during pandemics. It is thus an interesting question whether the known negative relationship between stress and compliance apply to the population at large, especially since, for example, physical distancing measures and hand-washing are recommended to protect not only the individual in question, but also other people in their surroundings. The extent of compliance with preventive measures has repeatedly been found to be influenced by the perceived gravity of the pandemic, as well as trust in the preventive measures against it [18,24,41,42]. Thus, we also expected that trust in the preventive measures taken in each country will lead to better compliance and less worry over the COVID-19 medical situation.

In this context, it is also worth investigating what factors may alleviate perceived stress, and thus, positively impact not just well-being, but also the likelihood of acceptance and compliance with behavioural guidelines. Social and practical support from groups such as one's family, friends and colleagues can moderate the effect of concern for the disease or other sources of stress on one's psychological well-being [43,44]. Indeed, interpersonal communication has been shown to affect the relationship between risk perception and actual behaviour during health crises, where mass media messages may fall short [45]. Various theoretical dimensions of social support, including social integration and reassurance of worth [46] are captured by the Social Provisions Scale (SPS) and its shorter versions (SPS-10; [47]). According to the stress-buffering hypothesis [48], social support through, for example, interpersonal communication buffers people from other negative consequences of distressing events. For instance, those who were ‘isolated but not lonely' [3] might be less susceptible to negative effects of isolation or other stressors related to a pandemic by maintaining strong social contacts while physically distanced. Further, social support could promote less threatening interpretations of stressful situations [47,49], and thus, possibly more positive appraisals of institutional and governmental efforts to stop the spread of a pandemic. Empirical evidence shows that perceived social support has helped build up resilience against disaster or catastrophic events, and exerts its protective function on alleviating psychological distress [50]. Furthermore, social support has been found to correlate positively with compliance in patient populations [51].

It is thus reasonable to investigate if we can extend the previous findings related to the stress-buffering hypothesis to the context of acceptance and compliance with behavioural guidelines during the pandemic. Here, we explored the relationship between experienced stress and trust in the efforts and behavourial guidelines set up by governments to curb the spread of COVID-19, as well as in relation to social provisions. In particular, we hypothesized that social provisions would negatively moderate the negative effect of stress on compliance with behavioural objectives during the pandemic.

So far, little is known about the cross-national differences in reactions to preventive measures during the 2020 pandemic. However, cross-cultural differences have been observed in responding to stressors [52,53] and to various countermeasures against pandemic outbreaks [54]. Further, it has been suggested that cultural variables such as the degree of collectivism, or social ‘tightness', may lead to different attitudes regarding compliance with social norms [3], and one study observed that cultural differences in relational mobility, or the perceived easiness to engage with strangers and freely choose friends, was associated with the spread of COVID-19 [55]. Thus, optimal institutional responses to the COVID-19 crisis may require culturally and situationally appropriate public health interventions, both to alleviate human distress and to communicate the need for preventive measures in an effective manner. In the present study, we took a multilevel analytic approach with individual responses nested within each country in order to estimate the variability in various dependent variables (DVs) residing at both individual and country levels [56]. By taking into consideration non-independence among individual responses within a country, as well as the differences between countries, we can better assess the extent to which individual-level variables measured in the present study explain variance in various criterion variables at the individual level and country level, respectively.

2. Methods

A global survey study was conducted with the purpose of mapping the psychological and behavioural responses to the COVID-19 pandemic across countries, and how these have impacted social behaviour and trust in government efforts to slow the spread of the coronavirus.

The COVIDiSTRESS survey (COVIDiSTRESS global survey network, 2020—available at https://osf.io/z39us/) consists of two parts. The first part collected general demographic data including proximate effects of the COVID-19 pandemic (e.g. isolation status, first-hand experience, attenuated risk), personality traits and self-reported variables, such as loneliness and perceived stress [57], daily behaviours including compliance with general and social preventive measures. The second part contains sets of more specific items related to distress, coping and social provisions. Firstly, we inquired about people's experiences of distress and worry during the ongoing outbreak of coronavirus, e.g. access to amenities, loss of work, adapting work, education and social interactions on digital platforms, the social stresses of confinement with adults and children. We also asked questions about the factors people experience as soothing and positive to coping (e.g. social contact, staying informed, dedicating oneself to preparation, hobbies, religion) and also included the Social Provisions Scale (SPS; [47]). Validated short versions of established measures were used whenever possible and available in local languages. In order to protect participants' data and avoid sensitive information, participants were not asked about COVID-19 symptoms or other data with direct medical implications that would allow third-parties to identify them.

Translations were completed for 46 languages and locations (Afrikaans, Albanian, Arabic, Bangla, Indonesian, Bosnian, Bulgarian, Chinese [Simplified and Traditional], Croatian, Czech, Danish, Dutch [Belgium, Netherlands], English, Spanish [Argentina, Colombia, Cuba, Mexico, Spain], Filipino, Finnish, French, German, Greek [Greece and Cyprus], Hebrew, Hindi, Hungarian, isiXhosa, isiZulu, Italian, Japanese, Korean, Lithuanian, Nepali, Persian, Polish, Portuguese [Brazil, Portugal], Romanian, Russian, Slovakian, Serbian, Turkish, Urdu, Vietnamese) with the possibility for more translations in future waves. Translations were completed by a forward translator from the original English version and then validated through both panel and back-translation process by separate translators when possible.

Given the urgent call for COVID-19 research, the survey received a waiver to commence data collection from the IRB office at Aarhus University, Denmark, followed by formal approval in June 2020 (case number: 2019-616-000009). The survey and data extraction schedule was pre-registered at the Open Science Framework [58]. The full annotated dataset has also been published in Scientific Data [59].

3. Key independent and dependent variables

The COVIDiSTRESS global survey contains a number of variables (full survey at https://osf.io/z39us/). The Royal Society Open Science registration files including the in principle accepted stage one analysis protocol and electronic supplementary material are available at the Open Science Framework https://osf.io/vz73d/. The final version of the cleaned and preprocessed dataset is available at https://osf.io/cjxua/. This urgent call RR focuses on 10 hypotheses (see §§1 and 4) and comparisons between countries. Table 1 gives an overview of all variables relevant for our hypotheses.

4. Hypotheses

First, this RR was intended to draw a map of differences in the psychological impact of COVID-19, and the trust in government measures and compliance with preventive guidelines in the countries represented in the first data extraction on Monday, 6 April 2020 (see inclusion criteria in §6). We expected the COVIDiSTRESS global survey to show significant differences between countries.

—	H1a: Perceived stress will differ between countries.
—	H1b: Concern over the coronavirus will differ between countries.
—	H1c: Trust in government efforts to slow the spread of coronavirus will differ between countries.
—	H1d: Compliance with behavioural guidelines to slow the spread of coronavirus will differ between countries.

Exploratory cross-country comparisons and descriptive statistics are provided in the Results section for H1 and electronic supplementary material.

Second, the present analyses focused on how both stress and concerns about the COVID-19 pandemic (e.g. isolation, changes in public life and worry over the virus itself) are related to compliance with preventive guidelines and trust in government efforts intended to reduce the spread of coronavirus. Compliance with behavioural guidelines:

—	H2a: Across countries, perceived stress will be negatively correlated with compliance to behavioural guidelines to slow the spread of coronavirus.
—	H2b: Across countries, concern over the coronavirus will be positively correlated with compliance to behavioural guidelines to slow the spread of coronavirus.

Trust in government efforts:

—	H3a: Across countries, trust in government efforts to slow the spread of coronavirus will be negatively correlated with perceived stress.
—	H3b: Across countries, trust in government efforts to slow the spread of coronavirus will be negatively correlated with concern over the coronavirus.

Third, we analysed the sources of stress during the global COVID-19 crisis. We assumed that concern over the coronavirus, arising from the direct health threat of COVID-19 and uncontrollable nature of the pandemic, will be a direct stressor.

—	H4: Across countries, concern over the coronavirus will be a predictor of perceived stress.

Fourth, based on the stress-buffering hypothesis, we analysed whether social support can alleviate the negative effect of perceived stress on the extent to which people comply with preventive measures. While stressful situations may be unavoidable during a pandemic, the availability of support systems may promote less threatening interpretations of the situation, and thus, lead to a more positive appraisal of efforts to combat the coronavirus.

—	H5: Across countries, availability of social provisions will negatively moderate the effect of perceived stress over coronavirus on compliance with behavioural guidelines to slow the spread of coronavirus.

Lastly, individual-level control variables/covariates were included in our multilevel analyses that examine hypotheses 2–5. These variables/covariates are respondent gender, age and education level. According to Coffé & Bolzendahl [62], women were less interested and involved in politics than men. Because trust in government can be easily affected by salient political issues (such as government's efficacy to handle the pandemic), we expected that respondent gender may predict trust in government efforts. In addition, we expect that respondents of different ages may react differently in terms of their experienced stress and worry [63]. As for respondent education, Dalton [64] indicated that education level was related to a decrease in trust in government. Based on the above reasons, respondent age, gender and education level served as individual-level covariates/control variables when testing our hypotheses 2–5. In addition, we added several macro-level (country-level) control variables that may be associated with the outcomes. We were particularly interested in each country's socio-economic growth (e.g. GDP, Gini coefficient and educational attainment). Country-level socio-economic factors have been reported to be related to psychological well-being indicators (e.g. stress, negative affect), so they have a merit to be included in our models [65]. Moreover, country-level education attainment was also considered given the negative association between the overall educational level and poor psychological well-being at the country level [66].

5. Number of observations to be collected and rule of termination

Given that this large-scale COVIDiSTRESS survey was spread by numerous researchers all over the world, we had limited control over the total number of responses per country. Thus, we defined a set of stopping rules that are practically feasible.

Statistical power in multilevel design is rather complex and greatly depends on the nested structure of the data [67]. Therefore, a general rule of thumb (i.e. 30–30 rule; [68]) is unlikely to be applicable to our data structure. Thus, we sought to refer to a more specific guideline to plan sample sizes according to the estimates acquired from pre-existing databases.

Due to a lack of previous research that could be used for reasonable power analyses, we ran a planned multilevel analysis for H5 (specified in §9), as there was a dataset at hand, which we perceived as sufficiently suitable. We used the 2018 European Social Survey (ESS) database to estimate the effect sizes of fixed and random effects as well as intraclass coefficients (ICCs). The ESS database consists of 43 215 respondents from 23 European countries (on average 1878 respondents per country). Although the variables available in the ESS database are not identical to those in the COVIDiSTRESS survey, given their similarity in the underlying constructs, we used the health condition variable (a higher value indicating poorer health condition) and satisfaction about the government in the ESS data as proxies for stress, compliance and trust in government efforts for our original variables of interest, respectively. The ICC from the observed ESS data was 0.26, indicating that 26% of variability comes from between-country differences. Since random effects are often distorted while the variables are standardized, the random effects were estimated with grand-mean-centred continuous variables. The same multilevel analyses were run again with standardized variables only in order to obtain the standardized estimates of fixed effects. Standardized estimates of fixed effects of the individual-level predictor (individual's health condition: β = 0.09; social provisions: β = 0.21) and its country-level predictor (mean health condition in each country: βs = −0.35) were deemed to have small and medium effect sizes, respectively. Their cross-level interaction effects had only small effect sizes (βs = 0.03–0.10).

Subsequently, we referred to the guidelines from Arend & Schäfer [67], which provided a fast and frugal power estimation for each combination of effect size, the value of ICC, as well as the size of random slope variance. Based on Arend & Schäfer's Table 8 of power simulation results for required sample size and group size (in our case, the number of countries) to detect such effects with 80% statistical power (p. 17), we planned to recruit at least 30 participants per country so as to detect both the effects of individual- and country-level predictors. However, since our data collection process was still ongoing at the time of stage one submission, the potential number of group size and sample size were greater than we had initially reported in this section.

6. Inclusion criteria

All participating researchers received the same task of starting to distribute the survey in their language on 30 March, 2020 (exception: Denmark started on 26 March, 2020, see §7). All possible channels were allowed, e.g. social media, panels, e-mails to friends or organizations, use of media contacts such as the newspaper or TV, websites of organizations such as universities or NGOs involved in health communication. Moreover, we also asked all participants to help spread the survey after they finished their own responses.

Data was extracted each Monday and made available at the Open Science Framework. In order to be considered for the present analysis, however, a country needed at least 30 respondents at the time of acceptance of the RR stage 1 (i.e. the most recent data extracted at time of acceptance). In order to be considered as a valid participant for the present analyses, a respondent must have reported their country of residence and submitted valid responses to the variables treated in each analysis (see also §7).

7. Exclusion criteria

We defined exclusion criteria on the country and on the respondent level.

8. Quality checks

Regarding quality checks, we checked for floor/ceiling effects within the following scales/item batteries: Perceived Stress Scale (PSS-10), Corona concerns, Compliance, SPS-10. We evaluated floor/ceiling effects on the base of the percentage of the respondents with the minimum/maximum scores. Therefore, we provided the percentage (%) and n for respondents with the minimum/maximum scores. Floor/ceiling effects will be considered as present if minimum/maximum scores occur in 15% or more of the respondents, following the previous recommended threshold (see [73]).

Furthermore, we tested the measurement invariance of the included scales: the PSS-10 (see [57,74]), Social Provisions Scale (SPS, short form of PSP-10; [47]) and coronavirus concerns (self-reported measure, constructed for the purpose of this study). The measurement invariance test was required since we wanted to compare scores measured by the aforementioned scales across different countries. Further details regarding the measurement invariance test procedures are described in the next section¹. We implemented the examination of the equivalence of measures at the suggestion of a reviewer.

Given that the number of countries to be analysed directly influences the power of the planned multilevel modelling, we conducted the multilevel modelling analyses as follows:

• 1. If our planned analyses can be conducted without any issue (e.g. failed convergence) with the dataset that passed the measurement invariance test, then we would use composite scores and report the findings.

• 2. If 1 failed, then we would apply the alignment method and use aligned factor mean scores instead of composite scores for the planned analyses. If 1 failed and findings from 2 were reported, we would discuss limitations and caveats regarding the interpretation of the findings.

For quality check reasons we provide histograms of the main variables in the electronic supplementary material. Specifically, we depict histograms for perceived stress (electronic supplementary material, figure S1), concern over the coronavirus (electronic supplementary material, figure S2), trust in government efforts (electronic supplementary material, figure S3) and compliance with behavioural guidelines (electronic supplementary material, figure S4).

9. Analysis plan

10. Data processing

Data collection started on 30 March 2020 in all participating countries, with a pre-launch in Denmark on 26 March 2020 to test if the survey worked in practice and get a first impression on the response rate. Prior to submitting the RR stage 1, we had not accessed or analysed any data, except following response rates to give all COVIDiSTRESS participant researchers feedback on the recruiting progress.

This RR was submitted under the COVID-19 rapid response call for papers in Spring 2020. Upon acceptance of the stage 1 protocol, we analysed the newest weekly data extraction that fit minimum criteria (see above). The full dataset collected is available to any other researcher for further exploration or hypotheses testing via the Open Science Framework.

The version of the COVIDiSTRESS dataset analysed here was extracted on 30 May 2020. A total of 173 429 responses were recorded and cleaned using the R-code recommended by the COVIDiSTRESS team (available at https://osf.io/z39us/files/). The cleaned dataset consisted of 125 306 responses.

Following our pre-registered plan of analysis, we first checked whether the equivalence of invariance of our measures across countries could be reached. For invariance models to converge, we initially excluded 2538 participants (2.03%) (i) from countries with less than 100 participants and (ii) who did not indicate the country of residence. Next, 1223 participants (0.99%) were excluded who completed the survey in less than 10% of the Qualtrics estimated survey duration (i.e. less than 2 min and 12 s). When the actual duration was examined, the mean duration was 2118.38 s. The 1223 excluded participants completed the survey in the top 0.99% fastest speed. Further, we used Mahalanobis distance to detect multivariate outliers due to random or carelessly invalid responses [71,72]. A total of 5189 (4.27%) participants with a p < 0.001 in the chi-square test were excluded. After these exclusions, the dataset consisted of 116 356 participants (see electronic supplementary material, table S1).

In the second step, following our pre-registered quality check plan, we investigated floor and ceiling effects for our independent variables, testing how many participants' self-reported compliance with preventive guidelines as well as composite scores of perceived stress (PSS-10), available social provision (SPS) and concern over coronavirus reached the minimum or maximum value. If more than 15% of respondents reached either the maximum or minimum value for a measure, the ceiling/floor effect was considered present. For SPS, a ceiling effect was detected for two countries: Brazil (N = 81, 16.0%) and Panama (N = 78, 16.1%). For concern over coronavirus, ceiling effects were found for 11 countries: Mexico (N = 1505, 19.2%), United States (N = 322, 16.0%), Panama (N = 103, 16.1%), Philippines (N = 89, 17.9%), Brazil (N = 112, 18.3%), Colombia (N = 28, 17.2%), Malaysia (N = 137, 29.7%), South Africa (N = 24, 21.4%), Greece (N = 105, 18.8%), Pakistan (N = 40, 15.2%) and Portugal (N = 153, 17.0%). In the third step, we assessed measurement invariance. As detailed in the RR, we expected that the COVIDiSTRESS global survey would show significant differences between countries, and five hypotheses were proposed. In order to test our hypotheses, we needed to examine the extent to which participants from all countries completed the focal measures similarly. Using a classical approach, we reached partial invariance in the case of PSS-10 and SPS-10, but not for concern over coronavirus. Thus, we proceeded with the alignment method (see OSF for the results of measurement invariance tests and alignments; links: https://osf.io/qp96b/ and https://osf.io/hu683/). The three measures examined were the corona concern scale, the PSS and the SPS (table 1).

We followed four main analytical steps in Mplus for measurement testing. First, we conducted an exploratory factor analysis using maximum-likelihood estimation and geomin rotation for each scale to confirm their dimensionality. One-factor structures were confirmed for all three scales. Second, we examined the extent to which the one-factor model was well-fitting across the entire sample using a pooled dataset. This was achieved by adjusting variables so that cases from each country have equal weighing on the covariance matrix. Third, we conducted a multi-group confirmatory factor analysis testing a configural invariance model to examine the measurement invariance of the one-factor model across countries. Fourth, we tested measurement invariance more properly with the alignment optimization method (see detailed description below). Finally, we conducted Monte Carlo simulations to provide a validation check of the alignment results, indicating whether the rank order of latent factors across countries is trustworthy.

The latent factor scores for each scale obtained from the alignment analyses were combined to be used in the key hypotheses testing analyses. We examined data from 104 332 participants from 48 countries. Sample size ranged from 109 (New Zealand) to 20 165 (Finland), with a mean of 2173 and median of 921 participants per participating country. Based on the results reported in detail below, we deemed that measurement invariance of the three measures was achieved. In brief, the results provided strong evidence for the one-factor structure of all three measures, that this unifactorial factor structure was confirmed across all 48 countries, that participants from all countries responded to the items relatively similarly, and the Monte Carlo simulations confirmed the trustworthiness of the latent factor scores obtained.

The final dataset consisted of 116 356 participants from 48 countries with the mean age of 39.4 (s.d. = 14.1), 84 942 (73.00%) of participants identified as female, 29 869 (25.51%) as male, and 1283 (1.10%) provided other responses or declined to respond. For detailed demographic statistics, see electronic supplementary material, (electronic supplementary material, table S1).

For the multilevel models (H1a to H5), we dummy coded individual-level categorical variables (participant gender and education level) and centred the continuous variables (predictors and control variables) on grand means (see electronic supplementary material, table S1 to S4 with descriptive statistics). Because the ICCs showed that a considerable amount of variability of outcome variables was accounted for by between-country difference (see the Results section), country-mean variables for individual-level predictors were created by aggregating individual scores at the country level and centred on the grand means in order to separate variance of country part from the variance of the individual part (see [56]).

Country-level covariates (population size, GDP per capita, Gini coefficient and educational attainment) were acquired from the World Bank World Development Indicators database [92]. We used PPP-adjusted GDP per capita, which represents each country's level of economic activities controlling for purchasing power (i.e. the differences in price levels between countries; Eurostat). Gini coefficient measures the extent to which the distribution of income among individuals in a given country deviates from a perfectly equal distribution (ranging from 0 to 100—0 represents absolute equality, whereas 100 represents absolute inequality; [92]). Educational attainment referred to the percentage of 25 years old or above in the population who completed at least upper secondary education.

11. Results