During Pandemics, We Trust:COVID-19 Risk Exposure and Health System Trust


 We examine the extent to which exposure to higher relative COVID-19 mortality (RM), influences health system trust (HST), and whether changes in HST influence the perceived ease of compliance with pandemic restrictions during the COVID-19 pandemic. Drawing on evidence from two representative surveys covering all regions of 28 European countries before and after the first COVID-19 wave and using a difference in differences strategy together with Coarsened Exact Matching (CEM), we document that living in a region with higher RM during the first wave of the pandemic increased HS. However, the effect is driven by individuals over 45 years of age, and the opposite is true among younger cohorts. We find that a higher HST reduces the costs of complying with COVID-19 restrictions, but only so long as excess mortality does not exceed the average by more than 20%, at which point the ease of complying with COVID-19 restrictions significantly declines, offsetting the positive effect of trust in the healthcare system. Our interpretation of the estimates is that RM is interpreted as a risk signal among those over 45, and as a signal of health-care system failure among younger age individuals.



Introduction
Given that clinical processes and quality are complex and poorly understood by the public, users' trust in the health system can serve as a behavioural guide. People's beliefs about the efficacy and effectiveness of health care services can be critical in navigating a crowded healthcare system (Ramalingam et al., 2020) 1 . Nonetheless,under pandemic circumstances like COVID-19, cooperation with pandemic regulations depends heavily on people's goodwill. As a result, health system trust (HST) becomes a low-cost heuristic for users deciding whether or not to comply with  restrictions and treatment compliance (O'Malley et al. 2004;Ozawa and Sripad 2013;Kittelsen andKeating 2019, Rogers andPrentice-Dunn, 1997;Voeten et al., 2009;van der Weerd et al., 2011). . According to Hall et al. (2001), HST refers to a person's belief that healthcare institutions and professionals in general are concerned about their health.
However, it is founded on normative value judgements derived from knowledge of other people's experiences and information disseminated through the media, rather than solely on personal experience (Thiessen, 2009) In a pandemic, HST can influence the perceived cost of compliance with social distancing (Bargain and Aminjonov, 2020;Clark et al., 2020) 3 , as well as individual's 1 In extreme cases, excess reliance on trust can crowd out preventive healthcare behaviours Including skipping breast and cervical cancer screenings (Yang et al., 2011), reducing contact with doctors (Trachtenberg et al., 2005;LaVesit et al., 2009), or disregarding medical advice (Egede and Ellis, 2008). 2 According to Zengh et al. (2003) trust can be explained by four dimensions (both inter-personal and public): (i) fidelity, or upholding the patient's interests above all else, (ii) competence, or ability to produce the best possible outcomes, (iii) honesty, avoiding deliberate misrepresentation, and (iv) confidentiality, or the correct use of sensitive information. 3 Similarly, some studies establish an association between institutional trust and ease of compliance with recommendations in the context of the H1N1 epidemic in the UK (Rubin et al., 2009) and Italy (Prati et al., 2011), SARS in Hong Kong (Tang andWong, 2003) and Ebola in Liberia (Morse et al., 2016). likelihood of reporting a positive test, and more generally, adhering to self-isolation or quarantine requirements (Gilson, 2003, Department for International Development, 2020. These were historically unprecedented interventions limiting individuals' freedoms 4 . However, so far, we know little how does the severity of a pandemic influence HST. To date, it is unclear how individuals interpret changes in a country's relative COVID-19 mortality, whether it is a signal of higher risk calling for further confidence in the healthcare system experts given the complexity of the pandemic causes, or instead it is interpreted as a sign of failure of the health system regulations to stop the pandemic. This paper adds to the literature by shedding light on how HST changes with exposure to COVID-19 mortality, which was a piece of information heavily communicated. We disentangle whether they are interpreted as a proxy of further risk, or health system failure. Next, we examine whether HST impacts on individuals' perceived ease of compliance with lockdown restrictions 5 . We exploit evidence from two representative survey datasets from 28 European countries from before and after the first wave of the pandemic, as well as regional level NUTS-2 6 mortality data. We use a difference-in-difference strategy combined with Coarsened Exact Matching (CEM); a Relatedly, lack of trust in health institutions is associated with increased difficulties in dealing with bioterrorism threats (Meredith et al., 2009;McKee et al., 2009). 4 These include the effects on loneliness, unemployment, educational interruption, and interrupted healthcare, especially undeserved individuals. Indeed, some evidence suggests that whist early spring 2020 lockdown in Europe and the United States reduced mortality by 10.7%, later lockdowns did not (Herby et al, 2021). 5 We do not assume that COVID -19 exposure is a risk for everyone, but for the average individual. For instance, it might well be that younger people exposure might develop natural immunity and are thereby better able to protect the vulnerable people the interact with. However, this is not the case for most of the population.
The variable ease of compliance with lockdown is measured using the following question: "Thinking about the measures taken to fight against the Coronavirus outbreak, in particular the lockdown measures, would you say that it was an experience easy or difficult to cope with?: (1) very easy to cope with, and even an improvement to your daily life", (2) fairly easy to cope with, (3) both easy and difficult to cope with, (4) fairly difficult to cope with, (5) very difficult to cope with, and even endangering your mental and health". We define the variable "ease of compliance with lockdown restrictions" (COMPLY) inverting the Likert scale, so that (5) corresponds to "very easy to cope with" and (1) corresponds to "very difficult to cope with".
Explanatory variables. Based on the previous literature (Listhaug and Jackobsen 2017; Newton et al., 2017) we include controls for age, gender, nationality, marital status, occupation, age when finishing full-time education, household composition, difficulties in paying bills, level in society and Internet use. In addition, country-specific data includes size of municipality and region of residence. The lack of specific information of income and wealth is compensated using the difficulty to pay bills and the self-reported level in the society. Descriptive statistics are shown on Table A3. We are constrained by data availability. Eurobarometer datasets do not collect information on the full composition of the household, beyond dependents under 15, hence we can't identify the presence of older individuals in the household. We also do not have information on self-reported health status or whether they suffer from any chronic disease.
We draw on regional data on COVID-19 Excess mortality, measured as the excess mortality in 2013 and in 2020 with respect to the average of 2016-2019 and the J o u r n a l P r e -p r o o f average 14-day case rate of new COVID-19 cases per 100,000 inhabitants 13 . In the field of environmental pollution, a positive relationship has been documented between the risk perception of individuals exposed to pollution and local mortality records (Interdonato et al. 2014, Janmaimool and Watanabe 2014, Wachinger et al. 2013.
Although pollution affects people far more equally than COVID-19 and is more visible to individuals, the reporting on individual COVID-19 cases and deaths made it more visible and it was presented as if the risk could affect everyone. In this paper, we pose a similar hypothesis in probability of contagion and risk of death.
Relative mortality in 2020 ( 2020, ) is computed using average weekly registered deaths (all causes) between week 11 ( 11−2020 ) and week when respondent was interviewed ( 93.1 ) with respect to average weekly deaths between years 2016 and 2019 by NUTS-2, using the same dataset as before. The variable estimates the community deaths directly or indirectly attributed to COVID-19. (2) 13 These measures have been calculated with reference to the region of residence (NUTS-2) except for Cyprus, Estonia, Latvia, Lithuania, Luxembourg and Malta for which the country as a whole has been taken as a reference. In total, regional information is available for 197 NUTS-2 and 6 countries. 14 We cannot confirm that the information reported in the press coincides exactly with that appearing in these databases, but we have found that the countries used in this work meet the criteria of reliability and absence of manipulation examined in several studies (Sambridge and Jackson, 2020;Farhadi, 2021;Farhadi and Lahooti, 2021).
The average cases is defined as the average of 14-day case rate of newly reported COVID-19 cases per 100,000 population by week and territorial units (14 , ) between week 11 ( 11−2020 ) and the week when the respondent was interviewed ( 93.1 ). The sources consulted to compute the "Average Case Rate" by NUTS-2 are listed on Table B2.  Figure A2 in the appendix shows a map of European territorial units shaded in red according to RM in 2020 with respect to the 2016-2019 average (higher intensity indicates higher RM), which suggests an association between lower regional RM and higher trust in the healthcare system 15 . Similarly, Figure A3 displays the relationship between ease of compliance with lockdown restrictions and trust in the healthcare, suggesting an association between a region's healthcare system trust and ease of compliance with restrictions 16 . Finally, Figure A4 n the appendix maps the spatial distribution of the perceived ease of compliance with lockdown restrictions and the average number of COVID-19 cases per 100,000 inhabitants and displays that in regions with a higher incidence rate, there is greater dispersion in ease of compliance with restrictions 17 .

Exposure to COVID-19 and Healthcare System Trust
COVID-19 may have been a one-of-a-kind pandemic in terms of risk information exposure. Indeed, since the outbreak of the pandemic, the media has played a critical role in reporting on cases and deaths (Anwar et al., 2020;Tsao et al., 2021).
One way to capture exposure to the pandemic is by examining the effects of regional (excess) mortality in 2020 compared to the periods immediately before the pandemic (2016 to 2019). We hypothesise that individuals' trust in the healthcare system may be affected by relative mortality (RM).
To assess the impact of the pandemic on trust in the healthcare system, we propose a difference-in-difference-in-differences or triple (DiDiD) model, which compares trust in regions with excess mortality versus all other regions, and in 2013 versus 2020. A DiDiD model addressed the potential endogeneity form three types of unmeasured confounders: those that vary over time but affect people in a similar fashion 16 Regions showing the greatest ease of compliance with mobility restrictions are Danish (Sjaelland (3.76), Syddanmark (3.68), Nordjylland (3.65) and Hovedstaden (3.64)). Malta (3.61), Overijssel (3.53) and Zeeland (3.50) in the Netherlands also stand out. In these regions, confidence in the healthcare system is well above average (32% in the Danish regions, 29% in Malta, 25% in the Dutch regions). In contrast, the greatest difficulties are concentrated in Cantabria (Spain; 1.33) and several Italian regions (Marche, 1.71;Toscana,1.73;Liguria,1.87). In these regions, trust in the healthcare system is well below average (52% in Cantabria and 39% in the Italian regions). 17 The highest average number of confirmed cases per 100,000 inhabitants corresponds to several Spanish regions (Aragon,168.10;Madrid,132.87,La Rioja,117.60) and Smalland Med Arna (Sweden, 119.69). In these regions, the ease of compliance with the restrictions is above average, except in Madrid where it is 6% below average. In contrast, the lowest average infection rate is observed in Northern Ireland (2.17), Scoltland (2.18) and Pohjois-ta Ita-Suomi (Finland, 2.49). In these regions, the ease of compliance with restrictions is above average (13%, 11% and 28%, respectively).
(e.g., changes in the healthcare system between 2013 and 2020), those that vary across people but remain constant over time (e.g., fundamental differences among age cohorts), and those that vary over time but affect people differently (e.g.,, mortality) The DiDiD specification allows for differential trends across regions and by respondent's age . Following this assumption, we estimate the following DiDiD equation using ordinary least squares (OLS): where is the level of trust in healthcare system (according to the Likert scale, from (4) corresponding to "totally trust the health system" to (1)  regional and country fixed effects. They capture long-term NUTs-specific differences and common changes that occurred in all states in the same year (i.e., those linked to the economic cycle). Robust standard errors are obtained with clusters at regional level 18 .
Parallel trends assumption. An important limitation of a DiDiD analysis is that that the outcomes in the treatment and control groups would have followed parallel trends in the absence of the pandemic. For this purpose, we have relied on coarsened exact matching. Coarsened exact matching (CEM) is a matching strategy developed by Iacus et al. (2012), which reduces the impact of confounding on observational causal inference. The strategy consists of simultaneously matching using a set of possible confounders which are "coarsened", reducing the number of possible matching values for a given covariate with the aim of increasing the number of matches achieved 19 .
After applying the CEM method, a weighting variable is obtained to equalise the number of observations within the comparison groups, which takes values between 0 and 1. To check the balance of two comparison groups, the multivariate imbalance measure L1 is used, whose size depends on the dataset and the selected covariates, and which takes values between 0 (perfect overall balance) and 1 (maximum imbalance), i.e. a larger value represents a larger imbalance between two groups. When good matching occurs, a substantial reduction in L1 is obtained (Green et al., 2015) 20 . 18 In additional specifications, we also show that our results are robust to using the Donald and Lang (2007) method to calculate standard errors.
19 CEM works as follows. First, it makes a copy of the set of covariates chosen for matching. Second, the variables are broken down into different meaningful strata (i.e., into equal intervals of the same size or into intervals of different dimension from each other), through user choice automatically or through the CEM algorithm. Third, a unique stratum is created for each observation and each observation is placed in a stratum. The strata created are reassigned to the original data set, and any strata that does not contain at least one treated and one control unit is removed. Thus, the treatment effect is based on the matching provided by the algorithm, since the difference between treated and control units is obtained from the difference in the outcome variable between units belonging to the same strata. Finally, the higher the coarsening (higher number of strata), the lower the imbalance, as well as the lower the number of matches provided by the CEM. 20 See Table A2 for L1 statistics before and after CEM.
In our study, CEM has been used to make the two groups of respondents to the  (2015) also showed that optimal performance is warranted only when the vector of important confounders is relatively small (fewer than 10), which is fulfilled in our case. 22 As the weights are proportional to the residuals from a regression of treatment on country, region and year effects, we have checked that the residuals from a regression of the outcome variable on region and year fixed effects are linearly related to the residuals from a regression of treatment on region and year fixed effects and the slope of this linear relationship does not differ between the treatment group and the comparison group (results available upon request).
The DiDiD is an intention-to-treat analysis in which the coefficient 7 represents the effect of the pandemic on trust among older respondents in regions with higher RM.
To interpret the DiDiD effect as the causal effect of COVID-19, the incidence of the pandemic must be uncorrelated with other time-varying determinants of trust in healthcare in our sample. This assumption would be violated if the pandemic induced selection into our sample (for example, if the level of trust in the healthcare system of those who died between the two waves of the Eurobarometer were not randomly distributed, which would affect the sample of respondents in 2020).
To evaluate the plausibility of these concerns, we present the results from regressions that estimate the DiDiD model using observable respondent characteristics as dependent variables (and thus omitting the controls ). As we do not include individual-level controls in these regressions, we collapse the data to respondent's ageregion/year level. Results in Tables A4-A6 suggest that the pandemic is fundamentally uncorrelated with the explanatory variables Therefore, it seems unlikely that differential demographic trends drive the results shown in section 5.1.

Effect of trust in healthcare over ease of compliance with lockdown restrictions
where measures the ease of compliance with lockdown restriction of individual i living in region r of country c (using the Likert scale from (4) which corresponds "very easy to cope with" to (1) which denotes "very difficult to cope with").
, and are defined as in the previous model. As in the DiDiD model, RM is entered in the regression either as a binary variable or as a continuous variable.
denote the average of 14-day case rate of newly reported COVID-19 cases per 100,000 inhabitants for region r of country c (since the onset of the pandemic until the day of the interview).
Further, we examine the so-called "Cummings effect" to support the causal effect of trust in health authorities. This effect is named after Dominic Cummings, senior aide to the British Prime Minister, who was caught not complying with lockdown regulations, traveling with his wife (a COVID-19 suspect) and his son. Numerous scientists expressed their concern that such actions could undermine confidence in the health authorities 24 . Similar regulation breaches have been detected in Greece 25 , New Zealand 26 , Norway 27 , Spain 28 , which can undermine trust and individuals' behaviours, contributing to further outbreaks 29 , and relaxing their adherence to health recommendations, which may lead to further outbreaks (Wong and Jensen, 2020).
24 Fancourt et al. (2020) analysed 220,755 interviews conducted with 40,597 individuals between April 24 and June 11, 2020, in England, Scotland and Wales, and reported a reduction in confidence in government in England, starting on May 22 nd , although no comparable behaviour was found for confidence in the governments of the devolved nations. A knock-on effect of such actions was a decrease in public adherence to the guidelines of the health authorities (Marien and Hooghe, 2011). Fancourt et al. (2020) shows that before the Cummings breach became known (on May 22) there had been a relaxation in compliance, but the gap between England and Wales and Scotland widened in the weeks that followed. 25 Greek PM accused of breaking coronavirus lockdown rulesagain -POLITICO  (Lunn et al., 2020). However, this is an empirical question as a very strict lockdown may increase the likelihood of breaking the rules. Therefore, the potential endogeneity of the variables relative mortality ( ) and average case rate ( ) should be considered.
In equations (7) and (8), the vector refers to exogenous variables. In this paper, we use as instrumental variables the classification of the 28 countries into quartiles according to the INFORM Covid Risk Index 30 , which relies on three dimensions: "Hazard and Exposure", "Vulnerability" and "Lack of Coping Capacity" which focus on structural factors 31 . Using the value of the index, we classify the values into quartiles (very low risk, low risk, moderate risk and high risk) as reported on Table C1 (see Appendix C for more information of the items included in each dimension) 32 . Table C2 displays the average values of HST by RM and the number of confirmed cases in 2020. 30 The INFORM COVID-19 Risk Index is an adaptation of the Inform Epidemic Risk Index that tries to identify: "countries at risk from health and humanitarian impacts of COVID-19 that could overwhelm current national response capacity, and therefore lead to a need for additional international assistance" (Poljanšek, 2020). 31 Each of the 3 dimensions is measured on a scale between 0 and 10 in which a higher value indicates that the country faces more adverse conditions. The aggregation of the indicators has been performed following the INFORM model (De Groeve et al., 2014). 32 The use of the INFORM Covid-19 Risk Index might raise some doubts about its suitability, if one suspects that countries with higher values of this index, and therefore less preparedness to face a health emergency, would have opted to impose more restrictive mobility measures. However, this hypothesis does not seem at all plausible for three reasons. First, the INFORM Covid-19 Risk Index was published on April 20 th , 2020, e.g., when the first wave of the pandemic had already begun 32 . Second, Table C3 shows the chronology of mobility and containment restrictions approved in all the countries analysed, and all countries had enacted severe containment measures before the date of publication of this index. Third, Figure C1 shows the relationship between the INFORM Covid-19 Risk Index and the average Oxford Covid-19 Stringency Index during the first wave of the pandemic, showing that there is no positive relationship between the two variables. Using the RM as a continuous variable, Figure 1  [Insert Figure 1 about here] Table 2 reports the heterogeneous effects of several relevant covariates extending the specification M6 of Table 1. In 2020, a higher relative exposure to regional excess COVID-19 mortality gives rise to a sharp increase in HST among nationals (46-64 and 65+). Similarly, we find a compatible effect when we evaluate the effect among migrants, but the effect is significantly lower. Next, we examine the level of difficulty in making ends meet, and we document that, as expected, the negative effect is higher among the cohort aged 65+ years, which exceeds by more than 10 times that of those who have no difficulties at all. Thus, in 2020 and in the presence of excess mortality, lower income households have been more prone to reduce HST, and the gap J o u r n a l P r e -p r o o f between households' income increases with age 33 . Finally, we document significant heterogeneity by educational attainment. We observe a stronger decrease in HST for all age cohorts among those that left school before the age of 16.

Heterogeneity
[Insert Table 2 about here]

Ease of compliance with COVID-19 restrictions
Next, we examine how variations in HST affect people's ease of compliance with lockdown restrictions drawing on an instrumental variable strategy 34 .
[Insert Table 3 about here] Table 3 displlays the OLS and IV estimation results for the degree of ease in complying with lockdown constraints. IV estimation are performed by 3SLS, which uses GLS and provides more efficient estimates than a simple GLS (Greene, 2008). The upper part of the table shows the results for the variables considering that RM is a binary variable (1 if it is positive, 0 otherwise). The lower panel of the table shows the 33 These estimates are consistent with estimates of the self-reported social class: in the cohort aged 31-45 years, we document a different effect among those regarding themselves as "working class", an effect nearly 30 times higher than that of those who consider themselves to be "higher class". 34 First, we verify that the referred instruments satisfy two conditions: (1) relevance or being sufficiently correlated with the suspected endogenous variable, and (2) exogeneity or being distributed independently of the error process. The results presented in Table C4 strongly reject the null hypothesis of underidentification. To detect weak instruments, there are several informal procedures, such as the first-stage partial R 2 , which measures the contribution of the excluded instruments to explain variation in the endogenous variable, and the first-stage partial F-statistics on the excluded instruments. All the Fstatistics are above 10 and the partial R 2 suggesting that our instruments are relevant and strong. Since the Cragg-Donald-based test for weak instruments assumes homoscedastic errors, we also present the Kleibergen-Paap Wald rk F-statistic, which is valid in case of non-i.d. errors (Kleibergen and Paap, 2006). We find that the Cragg-Donald and Kleibergen-Paap Wald rk F statistics reject the weakness of the instruments. As the number of instruments is larger than the number of potential endogenous variables, we test for over-identification using the Hansen-Sargan (Hansen, 1982). The null hypothesis is that the instruments are valid (e.g., uncorrelated with the error term) and that the excluded instruments are correctly excluded from the estimated equation. The test statistics show that exogeneity is rejected at the 5% significance level. All three instrument options have been validated. results considering that RM is a continuous variable. For both, OLS and IV estimations, we proceed by a progressive incorporation of explanatory variables.
The OLS estimates underestimate the effect of HST but overestimate the effect of RM and the case rate. Estimates suggest that individuals are more likely to comply with lockdown restrictions if they live in high contagion or high mortality regions after the pandemic. However, IV estimates reveal that this is not the case. The real underlying motivation lies in HST. So, if the epidemiological situation leads health authorities to recommend a lockdown, and individuals understand that the underpinning reason is to protect their health by avoiding a health system collapse, they are more likely to cooperate in complying with mobility restrictions. One standard deviation increase in HST increases the probability in complying with the restrictions, and the effect ranges between 0.0046 and 0.0065 points (IV).
Similarly, when we focus on OLS estimates, we find that a one standard deviation increase in HST gives rise to an increase in ease of compliance with COVID-19 regulations of 0.0007 points, but such effect becomes negligible in the IV estimation.
Living in a region with excess mortality (binary variable) increases ease of compliance with mobility constraints by 5.22% compared to the mean value (that is, one percentage point smaller compared to M4 in the OLS estimation). Similarly, when we consider the continuous dimension of this variable, estimates suggest that a one standard deviation increase in RM increases the perceived ease of compliance with the pandemic regulations by 0.0006 points (compared to 0.0103 in the OLS estimation). Finally, its worth noting that we find a nonlinear effect in older cohorts (46-64, 65+ years), namely an initial decline (easier compliance with lockdown), but only up to a RM of 105. From this point on, the ease of compliance rises to a RM of 120. That is, when COVID-19 mortality exceeds 120, all cohorts, we find a decrease in the perceived ease (or increase in the difficulty) of compliance with restriction irrespectively of the age of the respondent. Although we hypothesised that the high mortality rate could be interpreted as an increased risk of contagion and, as a result, a greater preference to seek safety at home, our estimates suggest the opposite effect, probably indicating that higher level of relative COVID-19 are a signal of health system failure to control the pandemic. Table 4 shows the results of the IV estimates of the effect of HST on ease of compliance by age, nationality, age at leaving school and two measure of socioeconomic status, namely difficulties in making ends meet and self-reported social class.

Heterogeneity for ease of compliance perceptions
[Insert Table 4

about here]
We find that the effect of HST increases with restrictions, and 35 is 45% higher among older cohorts and. with years of education 36 . Indeed, the effect of HST is 105% higher among the highest educated (compared to the lowest educated) 37 . That said, more educated people may have higher expectations about the performance of political institutions (Cook and Gronke, 2005) and might be less tolerant with corruption (Hakhverdian and Mayne, 2012) 38 .
Next, we turn to examine the effects by nationality, and we find that nationals exhibit a 10% higher perceived ease of compliance than migrants, though HST reduces such ga 39 As expected, we find that the effect of trust on the ease of compliance is greater for households that do not face financial constraints 40 , and consistently, the average degree of ease of compliance with restrictions is almost 16% among those who consider themselves belonging to higher class compared to working class. This result is consistent with Newton et al. (2017)

Mechanisms
37 Most studies that have addressed the relationship between trust and education have focused on trust in political powers. Some studies (Hetherington, 1998;Anderson and Singer, 2008) document a positive relationship between trust and education. 38 Hence, education is a proxy variable for both cognitive skills and information processing ability and is found to reinforce the effect of trust in the healthcare system on ease of compliance to a greater extent than biological age. 39 The survey does not provide information on the health coverage of respondents, but it could be that unequal access to healthcare between nationals and immigrants is the cause of the effect among immigrants. 40 One standard deviation increase in trust in the healthcare system increases the ease of compliance with restrictions by 0.0107 points if there are no difficulties in making ends meet, compared to 0.058 points for households that always struggle to make ends meet (i.e., almost twice).
Finally, we propose two mechanisms to help explain our effect, namely the compulsory nature of the restrictions 41 , which might not be seen as justified, and the effect of the restrictions on the economy.
We rely on two questions from Eurobarometer 93.1. The first question refers to whether the restrictions impact on the country's economy 42 . We define three binary variables to represent the three possible responses: 41% of respondents thought there was a balance between health and economic protection, while 35% thought it was too focused on health at the expense of the economy (see Table A3). Table A7 displays the results of the OLS regressions for each of the three binary variables defined above. For each dependent variable, eight different specifications have been estimated, four using RM t as a binary variable and another four using RM t as a continuous variable, and in turn, in each of these four models the explanatory variables were introduced progressively. One standard deviation increase in HST decreases the probability of believing that measures are too much focused-on health by 0.0012 pp, or too much focused-on economy by 0.0021 pp. In contrast, one standard deviation increase in HST, increases the probability of believing that there is a good balance between health and economy by 0.0030 pp.
An increase in RM in 2020 or an increase in relative case rate is consistently associated with a decrease in the perception that measures are overly focused on health versus the economy. Living in a region with high mortality raises the perception that restrictions (are overly focused on the economy and lowers the perception that restrictions are overly focused on health. The second question asks the extent to which the respondent regards restrictions to be justified: "Thinking about the measures taken by the public authorities in your country to fight the Coronavirus and its effects, would you say that the limitations to public liberties were: (1) absolutely justified, (2), somewhat justified, (3) not very justified or (4) not at all justified?". 44% reveal that the measures were absolutely justified whilst 37% reveal they were quite justified (see Table A3). Table A8 shows the results of the OLS regressions for each of the three binary variables defined above. For each dependent variable, eight different models have been estimated, four using RM t as a binary variable and another four using RM t as a continuous variable, and in turn, in each of these four models the explanatory variables were introduced progressively. We find that one standard deviation increase in HST increases the probability of believing that lockdown measures are absolutely justified by 0.0026 pp. Importantly, living in a region with excess mortality increases (decreases) the beliefs that measures are somewhat justified (not very justified) by 49.8% (57.8%) 43 .

Conclusion
This paper has examined whether changes in relative COVID-19 mortality (RM) builds or weakens healthcare system trust (HST), and whether HST influences how costly it was for individuals to comply with COVID-19 regulations. We document three sets of findings. 43 In other words, one standard deviation increases in relative mortality in 2020 with respect to average 2016-2019 increases the probability that containment measures are absolutely justified by 0.002 pp. and decreases the probability of believing that measures are not at all justified by 0.004 pp.
First, we find that on average that RM increased health system trust (HST), and that HST reduces the costs to comply with COVID-19 restrictions. However, the effect is non-linear, as we show that 20% above average mortality reduces significantly, the propensity to comply with regulations, offsetting the positive effect of trust in the healthcare system. Second, HST increases with age and the effect of RM on HST during the pandemic was heterogeneous across individuals age groups. That is, it increased HST among people 45-64 and 65 and over as they were mostly affected by the pandemic, but it decreases it among younger cohorts.
Third, we find that a one standard deviation increase in HST leads to an increase in the perceived ease of compliance with COVID-19 restrictions which was heterogeneous across age groups and varied between 0.0086 points (18-30 years) and 0.107 points (over 65 years). That is, the effect of HST and perceived ease of compliance is 45% stronger for the older cohort.
There are several explanations for these results including higher economic difficulties among younger individuals, as proxied by an effect of individuals reporting "difficulty in making ends meet" and "self-reported social class". We document that the effect of HST on the ease of compliance is weaker among households that face financial constraints. The negative effect of RM on this groups can be explained as blaming the health system for the spread of the pandemic and the consequences it has had for their lives, jobs or businesses.
These results suggest that higher RM strengthens HST among individuals that are perceived to be more vulnerable. However, even such effect it only holds so long as it does not exceed 20% of the average RM. This evidence suggests that the pandemic was J o u r n a l P r e -p r o o f especially challenging among younger age groups, for whom RM is not necessarily entailed higher risk exposure, for whom higher RM is interpreted as a sign of failure that weaken their trust in the health system. Altogether, these estimates suggest that building HST is important and can make a difference to the perceived costs of compliance of the regulations necessary to fight future pandemics, and provides an explanation for the heterogeneous costs of compliance in regulation across age groups, which might suggest that in the event of the pandemic, younger age individuals out to be compensated, if HST is expected to remain strong among such an age group Schaeffer, K. (2020). A look at the Americans who believe there is some truth to the conspiracy theory that COVID-19 was planned. Pew Research Center July 24, 2020. Wachinger, G., Renn, O., Begg, C., Kuhlicke, C. (2013). The risk perception paradox-implications for governance and communication of natural hazards. Risk Analysis 33 (6) Table 1) Estimations have been performed using the final sample after CEM. Predicted trust in healthcare system after estimating a DiDiD model with interactions between age cohort, year 2020 and relative mortality with respect to average 2016-2019, and the following explanatory variables: sex, marital status, years of education, nationality, relation with economic activity, household size, number of household members (aged 15 and older, between 10 and 14 year, less 10 years), size of municipality of residence, difficulties for making ends meet, having internet at home, self-reported social class, territorial unit. Robust standard errors clustered at NUTS-2 level.  The table reports estimate of a canonical of a triple difference in differences specification based examining the effect of relative mortality and age cohorts on health system trust (HST). We report in bold the effect of one standard deviation increase over dependent variable for continuous regressors or percentage increase over average dependent variable for binary regressors. The estimations have been performed using the final sample after CEM.All regressions include as explanatory variables: sex, marital status, years of education, nationality, relation with economic activity, household size, number of household members (aged 15 and older, between 10 and 14 years, less 10 years), size of municipality of residence, difficulties for making ends meet, having internet at home, self-reported social class, territorial unit. Robust standard errors clustered at NUTS-2 level. Models M1, M2 and M3 include the continuous variable relative mortality in year t (t=2013, 2020) with respect to average 2016-2019. Models M4, M5 and M6 include the binary variable: 1 if t > 0 and 0 otherwise. Bold figures correspond to the effect of one standard deviation increase of the regressor over the dependent variable (for continuous variables) or the percentage variation with respect to the mean (for binary variables) Standard deviations in brackets. *** p<0.01, ** p<0.05, * p<0.1  to the effect of one standard deviation increase of the regressor over the dependent variable (for continuous variables) or the percentage variation with respect to the mean (for binary variables). The upper part of the table report OLS and IV regressions using "Relative Mortality in 2020 with respect to average 2016-2019" as a binary variable (1 if relative mortality is above zero and 0 otherwise). Lower part of the able report OLS and IV regressions using "Relative Mortality in 2020 with respect to average 2016-2019" as a continuous variable. Model M1 includes as explanatory variables: age cohort, sex, nationality and region of residence. Model M2 includes the same explanatory variables as M1 and additionally marital status and age when finishing education. Model M3 includes the same explanatory variables than M2 and also relation with economic activity. Model M4 includes the same explanatory variables than M3 and also household characteristics (size and number of people younger than 10, between 10 and 15, aged 15 and older), difficulties for making ends meet, having internet and self-reported social class. Robust standard errors clustered at NUTS-2 level. IV regressions use four instruments (high risk countries, moderate risk countries, low risk countries and very low risk countries according to the Inform COVID-19 Risk Index) to instrument the potential endogenous variables (trust in healthcare, relative mortality in 2020 and average case rate). Standard deviations in brackets . *** p<0.01, ** p<0.05, * p<0.1 Estimates refer to IV estimates for ease of compliance with lockdown restrictions using four instruments (high risk countries, moderate risk countries, low risk countries and very low risk countries according to the Inform COVID-19 Risk Index) to instrument the potential endogenous variables (trust in healthcare, relative mortality in 2020 and average case rate). In all regressions, RM 2020 is a continuous variable. Covariates include age cohort include sex, nationality, region of residence, marital status, age when finished education, relation with economic activity, household characteristics (size and number of people younger than 10, between 10 and 15, aged 15 and older), difficulties for making ends meet, having internet and self-reported social class. Robust standard errors clustered at NUTS-2 level. Standard deviations in brackets. *** p<0.01, ** p<0.05, * p<0. Source: own work using data from Table A1.

Relationship between HST (green circles) and relative mortality in 2020 with respect to average 2016-2019 (red areas) by NUTS-2.
Note: Green circles denote trust in healthcare, the higher the intensity of the colour, the higher the level of confidence. Red areas denote the relative mortality 2020, considering average weekly registered deaths (all causes) between week 11 ( 11−2020 ) and week when respondent was interviewed ( 93.1 ) with respect to average weekly deaths between years 2016 and 2019 by NUTS-2. The higher the colour intensity, the higher the relative mortality in 2020 with respect to average 2016-2019. Source: own work using data from Table A1.

Figure A3. Relationship between HST (green circles) and ease of compliance with lockdown restrictions (purple areas) by NUTS-2.
Note: Purple areas denote the degree of ease for complying with lockdown restrictions, the higher the colour intensity, the easier it is to comply with the lockdown restrictions. Green circles denote trust in healthcare, the higher the intensity of the colour, the higher the level of confidence. Source: own work using data from Table A1.

Figure A4. Relationship between average 14-day case rate of COVID-19 cases per 100,000 inhabitants (red bricks) and ease of compliance with lockdown restrictions (purple areas) by NUTS-2.
Note: Purple areas denote the degree of ease for complying with lockdown restrictions, the higher the colour intensity, the easier it is to comply with the lockdown restrictions. Red bricks denote the average of 14-day case rate of average of 14-day case rate of newly reported COVID-19 cases per 100 000 population by week and NUTS-2 between week 11 ( 11−2020 ) and week when respondent was interviewed ( 93.1 ). Longer bricks denote higher case rate.
= ∑ , 93.1 = 11−2020 93.1 − 11−2020 Source: own work using data from Table A1.          The effect of relative mortality on different preferences with regards to the balance between health and the economy during the pandemic. In bold: effect of one standard deviation increase over dependent variable for continuous regressors. The left panel of the table reports OLS regressions using "Relative Mortality in 2020 with respect to average 2016-2019" as a binary variable (1 if relative mortality is above zero and 0 otherwise); the right part of the able report OLS regressions using "Relative Mortality in 2020 with respect to average 2016-2019" as a continuous variable. Model M1 includes as explanatory variables: age cohort, sex, nationality, and region of residence. Model M2 includes the same explanatory variables as M1 and additionally marital status and age when finished education. Model M3 includes the same explanatory variables than M2 and relation with economic activity. Model M4 includes the same explanatory variables than M3 and household characteristics (size and number of people younger than 10, between 10 and 15, aged 15 and older), difficulties for making ends meet, having internet and self-reported social class. Robust standard errors clustered at NUTS-2 level. The effect of relative mortality on different preferences with regards to COVID-19 restrictions during the pandemic. In bold: effect of one standard deviation increase over dependent variable for continuous regressors. the left part of the table report OLS regressions using "Relative Mortality in 2020 with respect to average 2016-2019" as a binary variable (1 if relative mortality is above zero and 0 otherwise); the right part of the able report OLS regressions using "RelativeMortality in 2020 with respect to average 2016-2019" as a continuous variable.Model M1 includes as explanatory variables: age cohort, sex, nationality and region of residence. Model M2 includes the same explanatory variables as M1 and additionally marital status and age when finished education. Model M3 includes the same explanatory variables than M2 and also relation with economic activity. Model M4 includes the same explanatory variables than M3 and also household characteristics (size and number of people younger than 10, between 10 and 15, aged 15 and older), difficulties for making ends meet, having internet and self-reported social class. Robust standard errors clustered at NUTS-2 level.

Appendix B. Epidemiological variables 
Comply with lockdown: "Thinking about the measures taken to fight the Coronavirus outbreak, in particular the lockdown measures, would you say that it was an experience easy or difficult to cope with? (5) very easy to cope with, and even an improvement to your daily life, (4) fairly easy to cope with, (3) both easy and difficult to cope with, (2) fairly difficult to cope with, (1) very difficult to cope with, and even endangering your mental and health".  Trust in healthcare: "Please, tell me if you tend to trust or not to trust overall healthcare in your country: (4) completely trust, (3) somewhat trust, (2) somewhat mistrust and (1)   Source: Own work using Eurostat. Regional statistics by NUTS-2 Demographic statistics (Database -Eurostat (europa.eu)) for "Relative mortality in 2013" and "Relative mortality in 2020"; EB80.2 and EB 93.1 for "Trust in healthcare"; EB93.1 for "Comply with lockdown". The sources consulted to compute the "Average Case Rate" by NUTS-2 are listed on Table B2.  The Inform COVID-19 Risk Index is an adaptation of the Inform Epidemic Risk Index that tries to identify: "countries at risk from health and humanitarian impacts of COVID-19 that could overwhelm current national response capacity, and therefore lead to a need for additional international assistance" (Poljanšek, 2020 The Vulnerability ( ) dimension describes how severely exposed people can be affected (i.e., health vulnerability due to the social, economic, ecological, migratory behavioural and hazards characteristics of the country).  The Lack of Coping Capacity ( ) dimension measures the shortfalls in physical infrastructure, healthcare system capacity, institutional and management capacity. Each of the 3 dimensions is measured on a scale of 0 to 10 and a higher value represents worse conditions. The aggregation of the indicators has been performed following the INFORM model (De Groeve et al., 2014). The Inform COVID-19 Risk Index is obtained as: The risk group has been calculated taking into consideration the 28 countries analyzed. Taking as a reference the maximum value (4.48) and the minimum (1.49), quartiles have been obtained, so that countries in the first quartil exhibit very low risk and countries in the fourth quartil exhibit high risk.