key: cord-0767744-nr5b2dja
authors: Laliotis, Ioannis; Minos, Dimitrios
title: Religion, social interactions, and COVID-19 incidence in Western Germany
date: 2021-11-30
journal: Eur Econ Rev
DOI: 10.1016/j.euroecorev.2021.103992
sha: 2c6d831f9f1765b0633cd35e929a2e298d11bb0f
doc_id: 767744
cord_uid: nr5b2dja

This paper investigates how social interactions, as shaped by religious denomination, are related to COVID-19 incidence and associated mortality in Western Germany. We observe that the number of infections and deaths during the early pandemic phase were much higher in predominantly Catholic counties with arguably stronger family and social ties. The relationship was confirmed at the county level through numerous robustness checks, and after controlling for a series of characteristics and county fixed effects. At the individual level, we confirmed that Catholics, relative to non-Catholics, have tighter and more frequent interactions with their family and friends. Moreover, the intensity of social interaction was able to partially explain the relationship between COVID-19 incidence and share of Catholics at the county level. Our results highlight the number of dimensions that have to be taken into account when designing and implementing mitigation measures in the early stages of disease outbreaks.

Since the onset of the COVID-19 pandemic, millions of cases and related fatalities have been recorded across the globe. In Europe, Italy and Spain were the first countries to be hit the hardest in early 2020, with France following suit. Initial containment measures were relatively homogeneous across countries, e.g. school closures, local and nation-wide lockdowns, social distancing etc. However, the numbers of cases and related deaths varied substantially by country. The reasons are difficult to uncover. The timing of both the pandemic onset and the various policies to constrain it clearly played a role in March 2020; for example, in Sweden, Belgium or the UK. However, evidence suggests that the virus was present in Europe long before it was initially believed, e.g. as early as December 2019 in France (Deslandes et al., 2020) . Yet, in the first phase some countries were more severely affected compared to others, resulting in vastly overburdened health systems and excess mortality rates, before most governments were able to enact their mitigation policies.

Variation in social preferences could partly explain differences in population health outcomes between countries. Such preferences are crucial in determining the degree of social distancing and social interactions and they are largely culturally inherent (Remland et al., 1995; Sorokowska et al., 2017) . Especially when vulnerable groups such as the elderly are concerned, lower levels of social interactions have been linked to a lower incidence of the virus in some countries during the early pandemic phase (Bayer and Kuhn, 2020).

The early COVID-19 literature promptly highlighted that, for as long as a medical treatment was not widely available, social capital and human behaviour are important in keeping the number of cases low (Bartscher et al., 2020; Borgonovi and Andrieu, 2020; Ding et al., 2020; Durante et al., 2020; Van Bavel et al., 2020) .

For example, Bargain and Aminjonov (2020) argue that compliance to containment policies was higher in European regions where people trust their politicians more.

Recent empirical evidence demonstrated that variations in the number of COVID-19 cases and related fatalities are crucially driven by the type and frequency of interactions among people, especially in early pandemic stages (Alfaro et al., 2020; Liu et al., 2020; Platteau and Verardi, 2020) . Dowd et al. (2020) argued that in countries like Italy and Spain, co-residence and intergenerational social interaction can partially explain variation in COVID-19 incidence. However, substantial differences in socioeconomics, demographics, health systems, testing and healthcare capacity, social distancing policies and various types of policy response at different time periods, and event data collection processes, render cross-country comparisons problematic (Georganas et al., 2021) . 1 There is also evidence on how genetic differences affect the cross-country variation of the number of cases (Delanghe et al., 2020) . Moreover, such cross-country differences did not allow for a coordinated virus mitigation and lockdown exit strategy at the EU level in the first phase (Platteau and Verardi, 2020) . Therefore, empirical research shifted towards subnational analyses in order to identify how social structures, and differences in societal and individual behaviour and attitudes are linked to within-country variations in the number of COVID-19 infections and deaths. Bayer and Kuhn (2020) observed a country-level positive correlation between the number of cases and the share of people 30-49 years old living with their parents, arguing that social integration might be the factor behind the death toll in Italy. However, subsequent research by Belloc et al. (2020) using Italian regions as their units of observation revealed that the correlation has the opposite sign, suggesting policy recommendations based on cross-country results should be read with caution. In Europe, policy responses were rather similar across several countries at the onset of the pandemic, yet they resulted in varying degrees of efficiency. Even though European governments implemented nation-wide measures during the early phases of the pandemic, the number of infections and the associated death toll exhibited a great degree of variation at the local level within countries.

For these reasons, we focused on Western Germany where concerns around heterogeneity are mitigated (Spenkuch, 2017) . Our main purpose is to examine whether the strength of social and family ties as determined by religious denomination played a role. More specifically, we focused on a religious divide within Western German counties, i.e. Catholics and non-Catholics. This setting has been widely used to explain how behaviour is affected by cultural differences related to religion (Iannaccone, 1998; Ekelund et al., 2002; Arruñada, 2009; Spenkuch, 2017; Becker and Woessmann, 2018; Spenkuch and Tillmann, 2018;  Becker and Pascali, 2019). 2 Rather than discussing religion or Catholicism per se, we explore whether differences in specific behaviours within an otherwise very similar population could have impacted groups differently in terms of infections and related deaths. Our paper is the first one to rely on regional variation of religious group composition as a measure of cultural norms and preferences in this context. Hence, it contributes to the literature on how social preferences are linked to the virus incidence within a country.

Some earlier attempts examined the implications of local variation in mitigation measures compliance.

Specifically, two studies for Switzerland focused on the cultural and behavioural gradient behind the regional variation of COVID-19 cases. Proxying culture through language and using the twenty-six Swiss cantons as their units of analysis, Mazzonna (2020) argued that in the early pandemic phase, elderly people living in Latin speaking cantons were more severely affected by the virus spread, relative to those living in German-speaking cantons. This was attributed to more frequent social contacts at that time, although people from all ages in Latin, relative to German, cantons strictly complied after the lockdown. This was in line

with Brodeur et al. (2020) , Durante et al. (2020) , and Deopa and Fortunato (2020) who also reported greater mobility declines in areas with higher levels of civic capital and trust. 3 In this paper we use religious denomination to proxy culture. Guiso et al. (2006) defined culture as those customary values and beliefs being unchangeably transmitted from generation to generation within ethnic, religious, and social groups. The geographic dispersion of Catholics in West Germany has deep historic origins and has been covered extensively (Becker and Woessmann, 2009; Spenkuch and Tillmann, 2018) .

Cultural and behavioural dimensions shaped by religion at the local level are inherited from previous generations, they can be considered lifetime invariant, and they may have manifested themselves as a local way of life that is resilient to changes (Becker, 1996; Guiso et al., 2006) . Our paper is the first to contribute to this literature by shedding light on how cultural variations, that can be considered as exogenous and largely unclouded by unobserved heterogeneity, affect the within-country variation in the number of COVID-19 cases and associated deaths.

Using data from the initial stage of the pandemic, we observe that COVID-19 infection and related death rates were strikingly correlated with the local share of Catholics at the county level. The relationship remained strong after partialing out the effects of socioeconomic characteristics, county fixed effects, time trends, and after ruling out explanations related to geographical proximity to Northern Italy, differences in mobility patterns, and differences to pre-pandemic total mortality trends between groups using a differencein-differences approach. Our OLS, instrumental variables and difference-in-differences estimates confirmed that Catholic counties were more affected by the virus, in terms of cases and associated fatalities during the early stages of the pandemic. Our argument behind this finding is that Catholics, compared to non-Catholics exhibit stronger family and social ties, triggering a possible mechanism behind the county-level relationship that we document. 4 Previous research in economics has established that Catholics are more bound to close social circles and networks, i.e. family and friends, and they exhibit different patterns of social interactions compared to non-Catholics (Arruñada, 2009; Glaeser and Glendon, 1998; Ekelund et al., 2002; Satyanath et al., 2017) . To formally test our argument, we relied on individual-level information. First, we used data from the European Social Survey (ESS) and the European Values Survey (EVS) to confirm that, relative to non-Catholics, Catholics exhibit systematic differences in social preferences and behaviours that are crucial determinants for the number of COVID-19 cases. Next, we showed that the correlation between the share of Catholics and the COVID-19 incidence at the county level, can be partially explained by the intensity of social interactions, as the latter was proxied by cohabitation with parents, closeness to family and friends, frequent religious-service attendance etc.

Our results could support epidemiologists and public health policy makers to better understand how cultural factors can influence the initial spread of pathogens in a society and incorporate such differences as well when designing their optimal response policies (Platteau and Verardi, 2020). The remainder of the paper is organised as follows: Section 2 offers a background regarding the early pandemic phase in Western German counties. Section 3 describes the data sources used in the empirical analysis. Section 4 presents some descriptive evidence and outlines the adopted empirical strategy. Section 5 discusses the results, and Section 6 concludes.

According to data released from the Robert Koch Institute, the first reported COVID-19 infections in In terms of testing, widespread Polymerase Chain Reaction (PCR) testing was made available on March 25, 2020, and guidelines restricting testing only to severe cases were lifted, so that more than 2 million people received a test by late April. As of April 23, 2020, more than 148,000 confirmed cases and more than 5,000 related deaths were reported, based on data from the Robert Koch Institute. Figure A. 1 displays how the cumulative numbers of reported cases and deaths per 100,000 population were scattered across the country at the end of April.

The early pandemic phase, spanning from January 28, 2020, to May 01, 2020, is the period we will focus on, as the initial stages of the pandemic largely determine the incidence of the virus across the country (Zhao et al., 2020) . This is important, if we consider the relatively long incubation period, i.e. up to two weeks, and the large number of asymptomatic carriers, due to which cases may rise during the early stage before mitigating strategies become effective (Gudbjartsson et al, 2020; Lauer et al., 2020) .

Information on COVID-19 cases and related fatalities were drawn from the Robert Koch Institute (RKI) which is the federal government research institute responsible for disease control and prevention. in Panels A and B, respectively. In both datasets, Catholic and non-Catholic individuals seem quite similar in terms of standard demographics. With respect to the social interaction variables in the ESS data, Catholics seem more likely to meet other people more than several times per week relative to non-Catholics, and they are also more likely to meet other people more frequently compared to individuals of the same age. Based on the EVS data, Catholics consider family and friends to be more important in their lives, they attend religious services more frequently, they consider religion to be more important, and they are more likely to believe that children should provide care for their parents, relative to non-Catholics. Also, youngest individuals within Catholic households are older compared to those within non-Catholic households, on average.

Finally, to investigate for differences in mobility by different means of transportation, we used Apple mobility data. They are obtained through GPS tracking and they are available after January 13, 2020, and relative to that date, at the city level and for eighteen major cities in total. They provide information regarding requests for directions by transport type, i.e. walking, driving and transit and they are useful in comparing mobility patterns between cities.

Appendix Figure A .1 depicts our empirical observation. Western German counties with a higher share of Catholics in their total population appear more severely affected by the virus. There is a striking similarity between the prevalence of the disease in the early pandemic and the dispersion of Catholics across the country. On the sample of Western German counties, the correlation coefficient between the Catholic share over the total county population and the cumulative COVID-19 cases and deaths per 100,000 inhabitants was 0.42 and 0.32, respectively. This difference in COVID-19 incidence between Catholic counties and non-Catholic counties is also shown in Appendix Despite the observed similarity in terms of demographic and socioeconomic characteristics, we inspect for differences in overall monthly mortality rates before the pandemic (Appendix Figure A. 2). This is to ensure that the observed pandemic-related mortality patterns in Western German counties are not picking up preexisting mortality differences between Catholic and non-Catholic counties. The time series run from January 2011 to December 2017. They begin considerably after the last major swine flu outbreak in 2009, with no other major public health threats that could be affected from social contacts occurring since then. Both series are quite close and they move in parallel implying that before the COVID-19 pandemic nothing noticeable caused Catholics to pass at a higher rate relative to non-Catholics. Any systematic differences in genetic predisposition, pre-existing conditions, overall health status, socioeconomic characteristics, local infrastructure, health infrastructure or behaviours such as risky attitudes would be likely reflected in systematically different past mortality rates between groups. This would also capture most of the unobserved heterogeneity that could drive mortality rates. Therefore, any divergence in mortality patterns between counties observed during the first half 2020 should be due to the incidence of COVID-19 cases and related deaths. 10 The non-existence of pre-pandemic systematic mortality differences was also validated in a regression-based framework. More specifically, we regressed (logged) annual mortality rates from all causes (2011-2017) on a set of controls to formally test the parallel trends assumption at the county level. As seen in Appendix Table A .3, there were no systematic differences in overall monthly mortality between counties before the pandemic, conditional on county observables and fixed effects.

Ideally, the existence of such differences should be formally tested under a difference-in-differences (DiD) framework using daily mortality data at the county level. However, such daily information is only available at the state level. Given this limitation, we used data from January 01, 2020, onwards to test whether mortality rates were higher in predominantly Catholic states relative to non-Catholic ones after the pandemic onset. Predominantly Catholic states were defined as those where Catholics are the majority in their population. 11 Having shown that the pre-pandemic mortality trends in Catholic and non-Catholic counties were similar (Appendix Figure A. 2), we estimated the following DiD specification at the state level:

where is the logged mortality rate (total deaths per 100,000 population) in the -th state on day of 2020, is a binary indicator switched on for predominantly Catholic states, and is a dummy variable switched on after the first pandemic-related death in that state was reported. The DiD parameter of interest in Equation (1) is and indicates whether differences in mortality rates between Catholic and non-Catholic states have changed significantly after the onset of the pandemic. Based on our hypothesis about a higher COVID-19 incidence in Catholic counties and hence, states, we expect that . Table 1 displays the DiD parameters obtained after estimating Equation (1). Column 1 indicates an increased mortality rate in Catholic states during the exposure period. This is confirmed when the estimation sample is limited until April, 22, i.e. when some restrictions were lifted, covering roughly one month since the nation-wide lockdown. This would allow to capture fatalities resulting from the initial stage of the pandemic. 12 Column 3 adds a linear time trend running from the day when the first related death was reported to capture the already higher spread in some states compared to others. Finally, we add state fixed effects in column 4. In every case, the DiD parameter is positive and significant. Given that pre-pandemic mortality trends were similar across all states and counties, these results provide further evidence that mortality increased in predominantly Catholic states after the pandemic onset, following a higher number of coronavirus infections in those areas.

[ Table 1 here] Moreover, in Figure 1 , there is a clear discrepancy on how the number of COVID-19 cases per 100,000 population during the early pandemic phase has evolved over time in Catholic counties. At the onset of the pandemic, COVID-19 incidence is higher in Catholic counties, and this is the reason behind higher levels of related fatalities per 100,000 population in those counties as well. Moreover, apart from aggregating across Western German counties, Figure 1 displays the results for individual states with a sufficient number of counties. The similar pattern emerges across states with a sufficient number of counties, with the difference being more pronounced in some states, e.g. in Bayern, Rheinland-Pfalz, and Niedersachsen.

[ Figure 1 here]

Another issue could be that the higher number of cases could be due to the carnival festivities that took place in North Rhine-Westphalia in February (Pluemper and Neumayer, 2020). We used Apple mobility data to examine if this could induce problems of any sort. We classified the 18 Western German cities available in those data into Catholic and non-Catholic ones and we calculated the mean mobility indicators for each group, weighted by the size of the local population. Figure 2 displays the trends by transportation type before and after the lockdown (March 22, 2020). The use of public transport spikes about 30 days before the lockdown, which coincides with the culmination of carnival festivities in late February. However, this spike appeared at the same period for both Catholic and for non-Catholic cities, and mobility patterns were almost indistinguishable from each other, implying that there were no noticeable mobility differences between Catholic and non-Catholic cities. If the number of COVID-19 cases during the early pandemic was due to carnival festivities in Catholic cities, we would observe substantially higher mobility in those cities compared to non-Catholic ones. However, despite carnival being a Catholic tradition, it has evolved into a nationwide celebration for the youth, so that the spike in mobility occurred throughout Western Germany, and thus transmission of an infectious disease is just as likely anywhere in the country.

[ Figure 2 here]

Having established that observable characteristics, mobility patterns and pre-pandemic mortality rates were similar in Catholic and non-Catholic counties and states, the main goal of our analysis is to demonstrate that the reported COVID-19 incidence varies with religion at the county level. This is challenging because demographic information at the county level does not arrive as frequently as data on infectious diseases do.

Hence, we start our empirical application by estimating variants of the following specification:

where is the logged count of COVID-19 cases (or deaths) per 100,000 local population in county {1,…,N} in day of the early pandemic phase, is the logged share of Catholics in county (based on the most recent data available before the virus outbreak), is a linear time trend starting from the day when the first COVID-19 case was reported in each county, is a set of state fixed effects so they are not perfectly collinear with demographic and economic predictors varying at the county level, and is the disturbance term. For robustness, Equation (2) is also estimated with Catholic being expressed as a binary variable that is switched on for counties where the share of Catholics over the total population is higher than their respective state average.

The variable S represents the proportion of susceptible individuals in each county, where their stock has been approximated as the number of people after removing those reported deceased from the virus in each day (Adda, 2016) . 13 The t parameter represents the incubation period and has been set equal to 14 days. 14 We do not consider any spatial variation in the model so the incidence rate in each county is solely determined by its own past realisations. Therefore, parameter α could be interpreted as an estimate of the within-county spread. All models control for the size of the local population and for a series for economic and demographic controls. 15 Models also include state-specific time trends in order to account for confounders or policies that may vary at the state level, given the country's federal system. Estimating parameter β through these empirical models will provide a first indication about how the count of COVID-19 cases and related fatalities per 100,000 population varied with the local share of Catholics in each county during the early pandemic, conditional on past incidence, observable characteristics, state fixed effects and time trends. We test the robustness of baseline estimates by considering factors that could arguably affect the incidence of COVID-19, i.e. including more control variables, controlling for distance from places that suffered a lot during the early pandemic, excluding extreme counties and counties close to any border, and controlling for weather conditions at the county level.

The problem with Equation (2) is that county-specific time-invariant unobserved heterogeneity is not partialled out, because including a county fixed-effect would not allow β to be estimated. Hence, the results based on Equation (2) are conditional only to state-level fixed effects. 16 To address this concern within a context where N is large and T is small and fixed, we applied the 2-stage method of Pesaran and Zhou (2018). In the first stage, the predicted residual cases (or deaths) per population from a fixed-effects estimation were averaged for each county over the entire period under consideration. In the second stage, they were used as the dependent variable in regressions over the cross-sectional sample of Western German counties, where models controlled for the local share of Catholics and other county-level characteristics.

More specifically, in the first stage we run the following daily panel data model using a fixed effects estimator: 13 We deviate from Adda (2016) by focusing on parameter β that measures level differences in the incidence of COVID-19 between Catholic and non-Catholic, rather than focusing on differences in transmission rates between counties through a Standard Inflammatory Response (SIR) model. However, we are also able to adapt their methodology as an additional robustness check in Appendix section A.1. 14 We have also considered alternative values for the incubation period, i.e. from 3 to 20 days (Lauer et al., 2020) . 15 In robustness checks, models control for other variables. For example, driving distance from Milan which was the first major pandemic centre in continental Europe, as well as factors that have been argued to affect the number of COVID-19 incidence, e.g. temperature, precipitation etc. 16 In fact, in the COVID-19 literature, most studies focusing on one country report estimates that either condition on some higherlevel fixed effects or they are unconditional to fixed spatial differences, e.g. Sa (2020), and Pluemper and Neumayer (2020).

In the second stage, the obtained residuals, ̂ , were averaged over time and they were regressed on timeinvariant county characteristics, using the sample of Western German counties: 17 The 14-day lagged COVID-19 incidence was multiplied by a factor representing the fraction of susceptible individuals in the county population. This factor is an approximation, and the county population is calculated as the population minus the cumulative daily number of COVID-19-related deaths. Therefore, it does not consider population changes due to county-specific fertility and mortality from other causes, and does not address any endogeneity concerns. The results are robust when alternative lag structures are considered, i.e. ranging from 3 to 20 days.

[ To address concerns regarding proximity to the Northern Italian border, an area that was hit the hardest at the pandemic outbreak in Europe, we calculated the fastest highway driving distance (in kilometres)

between the major city of each Western German county and Milan, Italy, using Google Maps. 19 Although this does not entirely rule out any geographic heterogeneity, it mitigates any concerns related to how infections were scattered across continental Europe. As seen in Table A .4, including the logged fastest highway driving distance, in kilometres, from Milan, did not affect the baseline coefficients. Driving distance from Milan could not explain any of the variation in the data either, as the associated point estimate was quite noisy and insignificant. In fact, closeness to borders in general, did not seem to offer a convincing explanation for the observed link. Excluding all counties right on any border area left the results unaffected.

We also tested the robustness of the baseline evidence by including weather-related variables, such as temperature and precipitation for 136 available counties taken from the Climate Data Centre (Deutscher Wetterdienst), that have been argued to be related to the number of cases, however, the same conclusions were reached. 20

Another concern could be that there might be some other groups, mostly concentrated in urban areas, such as Muslims, that were reportedly more exposed to the virus. Our test for this was to exclude cities with more than 200,000 population, in which such groups are more likely to be concentrated, but again the positive relationship was confirmed. This last test also helps in excluding the large number of non-religious individuals concentrated in urban areas, reducing potential biases from non-random migration and rendering the remaining counties more comparable. Finally, we performed a placebo test by replacing the local share of Catholics in the county with the share of Protestants; as well as the dummy indicating a Catholic county with similarly constructed one indicating a Protestant county. As seen in Table A .4, the effect of Protestants was not significant and negative, providing some further reassurance about the direction and magnitude of our empirical observation. Moreover, controlling for both Catholic and Protestant regressors, confirms the positive relationship between the Catholic variable and the COVID-19 incidence at the county level. In models controlling for both religious groups the coefficients of the Protestant variables are close to zero and not significant. In this last specification the reference category is the share of other/non-religious. The results imply that given the share of other/non-religious, a higher share of Catholics is positively associated with COVID-19 incidence, whereas the share of Protestants is not.

A limitation of the results discussed so far is that they were conditional only to state level fixed effects, otherwise the coefficient of interest could not be estimated. However, it needs to be ruled out that the local share of Catholics is not correlated with unobserved factors that influence the COVID-19 incidence at the county level. This is a common problem in the COVID-19 literature because local economic and demographic covariates do not vary at the daily level while the virus incidence does. In fact, they are time invariant in this context.

To address this issue, we use the fixed-effects filtered (FEF) estimator suggested by Pesaran and Zhou (2018) . This allows to uncover the effect of a time-invariant covariate when using large N, small T data. In the first stage, a FE estimator is applied to Equation (3) and the residual COVID-19 incidence is averaged over the period for each county. In the second stage, Equation (4), the mean residual is regressed on the local Catholics share and other characteristics using the cross-sectional sample of Western German counties.

The results from the second stage are reported in Table 3 . In Panel A, the outcome is the mean residual obtained from applying a FE estimator to the total period in the first stage. The share of Catholics is positively and significantly affecting both the residual number of infections and deaths per 100,000 population, and this holds after controlling for demographic and economic characteristics. This is confirmed by both OLS and 2SLS estimates, where the local share of Catholics in each county was instrumented by the (logged) distance between each county's capital and the town of Wittenberg as suggested by Becker and Woessmann (2009) . In the case of the 2SLS estimates, the first-stage result is particularly strong, suggesting that the current religious composition is strongly related to the historical one, an observation similar to Spenkuch (2017). In Panel B, the outcome is the mean residual when the estimation sample in the first stage is restricted to the period before the lockdown, in March 22, 2020. The effect of the local Catholics share in the county remains positive and significant on both cases and deaths after controlling for the usual set of county characteristics. The same conclusions are reached when the first stage estimation is conditioned on the period during the lockdown. In the second stage, the link between COVID-19 incidence and Catholics share is positive and significant according to both OLS and 2SLS estimates. Finally, in Panels D and E, we use cases and related deaths for those below and above 60 years old, respectively, in the first stage. The RKI data report the age for each COVID-19 case or fatality, so we were able to test for age-related nonlinearities. The results in the second stage confirm a positive relationship for both age groups and, as expected, the estimated coefficients of the local Catholic share are higher for those above 60 years old, especially when considering the number of deaths per 100,000 population. This is worth mentioning, because even though non-Catholic counties have a slightly higher share of elderly, the number of infections and related deaths among the more vulnerable increases with the local share of Catholics.

[ Table 3 here]

The county-level analysis so far, established a positive link between COVID-19 incidence and the prevalence of Catholics in Western Germany. Behind this observed relationship, differences in individual behaviours and preferences could play a role. More specifically, stronger family ties and tighter social networks, could provide a possible channel that partially explains the aggregate results. can affect behaviour in ways other than social contacts. Using this information, we constructed indicators on whether a respondent believe it is justifiable to avoid paying taxes, accept bribe, avoid paying fare in public transport, and whether they have confidence in their government.

Then we used those behavioural variables as outcomes in regressions where the variable of interest was a binary variable equal to one if the respondent was a Catholic and zero otherwise, alongside a series of demographic and other characteristics (i.e. age, gender, subjective health status, employment status, household size, household income, age of youngest household member, city size) and state fixed effects. Table 4 reports the results. In Panel A, we consider social interaction outcomes from the EVS data. The results are supportive of a positive relationship between belonging to the Catholic denomination and frequency of social interactions. When the outcome is the frequency of meeting other people, the coefficient of the Catholic dummy is barely not significant at the 10 percent level (t-statistic = 1.67), but the other two are, indicating that Catholics tend to have more frequent social interactions and exhibit stronger ties relative to non-Catholics. In Panel B, we turn to EVS data in order to see whether there is any systematic differentiation with respect to outcomes related to social and family ties. In line with Arruñada (2009), the results clearly indicate that Catholics, relative to non-Catholics, regard their family and friends as being very important, they trust their family members more, and they are more likely to reside in the same household with their parents or parents-in-law. This provides further support that the higher incidence of COVID-19

infections and related fatalities in predominantly Catholic counties, could be partially attributed to the existence of stronger social and family ties among Catholics. For robustness, we proceed with excluding all other religious groups from our specifications and only focussing on Catholics, Protestants and nonreligious. 22 These three groups account for roughly 92 percent of all EVS respondents in Western Germany.

The results are presented in Panel C of Table 4 and our findings remain largely unchanged.

To further strengthen our proposed channel, we employed another set of individual-level regressions using individual EVS data. The idea was to examine whether the observed differences in the COVID-19 incidence in Catholic and non-Catholic counties in the early pandemic are due to differences in behaviours other than through social contacts. Specifically, we test whether Catholics have a higher propensity to justify cheating behaviour in avoiding taxes, accepting bribes and avoiding fares in public transport. Further, we examine for differences in their confidence towards the government; even though recent research suggests that in most European countries confidence in governments and their implemented measures have increased during the early stages of the pandemic and that individuals tend to prioritise health (Bol et al., 2020; Hargreaves Heap et al., 2020) . This test should be indicative of whether the higher number of cases and fatalities in predominantly Catholic counties was due to an overall riskier behaviour or disobedience towards the authorities and the mitigation measures they introduced. This would likely have a differential impact on the number of COVID-19 cases even before the mitigation measures become effective. Panel D in Table 4 reports the results.

[ Table 4 here]

Broadly speaking, the data reveal that adherence to rules is very high in Western Germany. Moreover, Table   4 does not suggest a higher propensity for Catholics to disobey rules, relative to non-Catholics. If anything, Catholics are less likely to tolerate even the milder offense (avoiding fare) and exhibit marginally higher confidence in the government. Therefore, the higher count of COVID-19 cases during the early pandemic phase seems unlikely to have resulted from differences in risky behaviours and civil disobedience between 22 Having already shown, in Appendix Table A .4, that infections do not vary with the local share of Protestants, Appendix Table  A .5 further narrows down the comparison between Catholics and Protestants alone using individual-level data. Both the obtained OLS and unbiased parameter estimates suggest that Catholics seem to have more social interactions as compared to Catholics alone, e.g. they meet other people more than several times per week and they show greater trust towards their family members. In particular the unbiased coefficients point towards this direction in four out of six different outcomes we consider.

Catholics and non-Catholics. Based on the evidence from Panels A and B, stronger social and family networks seem more relevant.

An issue here is that unobserved heterogeneity may bias the results. This is often the case with crosssectional analyses of survey data. To partially correct for this, we follow the methodology proposed by Oster (2019). The idea is to simultaneously observe how the coefficient of interest co-moves with the when including the full set of controls. This allows to approximate the selection on unobservables relative to the selection on observables and obtain an unbiased coefficient of the Catholic indicator. We read the latter as a "conservative" estimate obtained under the standard assumptions that (a) unobservables matter as much as observables, and (b) that the maximum variance that can be explained is 30 percent higher than the one Table 4 displays the obtained unbiased parameters for the Catholic indicator.

For all the outcome variables considered in Panels A, B, C and D of Table 4 , the unbiased parameter estimate of the Catholic dummy suggests that if omitted variables were as important as the included observables, the bias would be against the estimated parameters in column 1. However, the estimated coefficients in column 1, Panel A, remain rather stable implying that Catholics are different in terms of social interactions relative to non-Catholics. The same applies to the outcomes considered in Panel B.

Catholics appear to have stronger family and social ties relative to everybody else. Among others, Catholics are more likely to live in the same household with their parents or parents-in-law. This provides some reassurance that the uncovered differences in behaviour and preferences between Catholics and non-Catholics are less likely to be driven by unobserved heterogeneity. Regarding the outcomes considered in Panel C, it seems that accounting for unobservables moves the unbiased estimates in the first three columns closer to zero. This further suggests that being a Catholic does not imply any differences in justifying cheating behaviour and complying with rules. When considering trust in the government as outcome, the coefficient remains remarkably stable indicating that Catholics exhibit indeed more confidence in the government relative to non-Catholics. Therefore, any observed differences in the initial number of cases and fatalities could be due to stronger social and family ties, rather due to riskier behaviour and non-compliance.

Within the context of this paper, a more direct test about how the more frequent social interactions of Catholics, relative to non-Catholics, can partially account for the higher number of COVID-19 cases in counties where the local share of Catholics is higher, would require controlling for variation in such behaviour at the county level, i.e. when estimating Equation (4). Originally this was not possible as both the ESS and the EVS datasets provide only state-level identifiers. After a licence agreement, county level identifiers for the 2018 EVS data were provided to us, allowing the aggregation of social interaction measures at the county level. However, surveyed individuals were sampled in a subset of 118 Western German counties. Therefore, our hypothesis was directly tested at the county level but under this restriction. To directly test whether increased social interactions among Catholics can explain (some of) the relationship between COVID-19 incidence and the local share of Catholics at the county level, we controlled for a social interaction index, in column 2. This index was constructed through a principal component analysis ( In column 2 of Table 5 , we estimated Equation (4) using OLS and controlling additionally for the predicted social interaction index (the first principal component). This reduces the magnitude of the coefficient associated with the local Catholic share at the county level, implying that the local share of Catholics is strongly correlated with our measure of social interactions. Even if other factors associated with the historic geographic dispersion may play a role, we find this strong association with social and family ties. Moreover, the relationship between the index and the number of COVID-19 cases per 100,000 county population is positive and significant at the 5 percent. This is also the case in column 4 reporting the 2SLS estimates of the local Catholic share. Compared to column 3, the 2SLS estimate in column 4 is reduced by nearly 28 percent when models control for the predicted social interaction. Instead of the constructed index, columns 5-7 control for the averages of the variables that were used to construct it instead. In all cases, the estimated coefficient of the local Catholic share at the county level is reduced, and the social interaction variables have the expected sign (column 7 controls for all variables used in the principal component construction).

Therefore, although based on a limited number of available observations, this direct test provides some support in our argument about more frequent and closer social interactions, cohabitation with parents etc., observed among Catholics, playing a role in explaining the positive relationship we observed at the county level.

[ Table 5 here]

Moreover, this relationship does not seem to be due to other behavioural differences between Catholics and non-Catholics. Looking again at the Apple mobility data in Figure 2 , there is a significant drop in mobility after the lockdown, however, there is not any notable differentiation on the basis of religion. A concern with those data could be that they refer to searched for directions to specific destinations via certain transportation means. This cannot be formally tested; however, it is plausible to assume that individuals would not necessarily look up the addresses of their family members and close friends before visiting them prior to the lockdown. Moreover, the Federal Statistical Office provided commuter mobility data at the county level gathered by Teralytics AG, a private service provider, for 30 million mobile phones in Western

Germany. The data refer to individuals crossing county borders during peak commuting hours between Monday and Thursday each week. The data are provided as monthly averages between January 2020 and May 2020, compared to the respective month of the previous year. We used those estimates as independent variables in regressions separately for each month as well as for the total period. As commuting could be an important factor in spreading the virus, this would indicate any related differences between Catholics and non-Catholics. However, the estimated coefficient for the local Catholics share is not significant, in Appendix Table A .6. Hence, there is no differential change in commuter mobility compared to the previous year depending on the local share of Catholics in the county. Therefore, the higher COVID-19 incidence in Catholic counties is more likely to be due to stronger family ties, cohabitation with parents, and closer social circles, rather than higher mobility of any sort before or after the lockdown.

The context within which social preferences are shaped is a crucial determinant of societal and individual behaviour. Such differences are difficult to be observed between otherwise comparable individuals and groups of people. It is often the case that they are unobserved and it is burdensome to isolate their potential impact on individual or group-level outcomes and relationships (Alesina and Giuliano, 2015) . Nevertheless, as they can lead to differential outcomes it is of high importance to take them under consideration. Such heterogeneities need to be taken into account not only in epidemiological modelling and parameterisation, but also to design optimal policy responses. Moreover, they do not only exist between, but also within countries and often go beyond standard demographic and socioeconomic characteristics, e.g. age, income or health status. This is particularly important when facing crises such as the COVID-19 pandemic, which has already claimed many lives both directly and indirectly (Vandoros, 2020) . Being able to explain diverging trajectories of countries in the early stages could provide invaluable insight to policymakers to design mitigation measures and policies around pandemics (Platteau and Verardi, 2020) . Previous attempts to highlight the importance of culture and how it can determine social interactions are marred by unobserved heterogeneity and it is difficult to disentangle the number of factors that need to be taken into account. In this paper we used data from Western Germany in early 2020 to highlight the relevance of cultural differences in shaping social interactions on the incidence of the disease and the associated mortality in the initial phase of the COVID-19 pandemic. The starting point of our motivation was the observation that the incidence of COVID-19 infections and associated fatalities during the early pandemic phase was strikingly similar to how Catholics are scattered throughout the Western German counties. After validating that

Catholic and non-Catholic counties are quite similar in terms of county-level characteristics, mobility patterns, and pre-pandemic mortality rates, we confirmed our empirical observation in a regression-based framework that accounted for differences in county-level covariates, county-level fixed effects and time trends. Our results survived a wide series of robustness checks and they were also confirmed using the predicted Catholic incidence under an instrumental variables framework. COVID-19 cases per 100,000 population were higher in predominantly Catholic counties in Western Germany. It followed that the associated mortality, as well as overall mortality, was higher in those counties.

Our main argument here is that the higher number of COVID-19 cases and related deaths in predominantly Catholic counties can be attributed to the fact that Catholics seem to exhibit a higher degree of social interactions. The historical geographic dispersion of Catholics across Western Germany closely resembles the contemporary one. As a result, some shared values and preferences may have firmly established themselves at the local level. To test its relevance, we first worked at the individual level, showing that

Catholics are systematically different from non-Catholics in terms of social interaction frequency, religious services attendance, cohabitation with parents and parents-in-law and a series of other indicators capturing social and family ties. Next, we used the individual-level data to construct a social interaction index, that was able to explain part of the observed positive relationship at the county level.

Our results are important and indicate that, especially in societies with stronger social and family ties, virus outbreaks should be managed carefully, promptly and in a targeted manner by the authorities in order to avoid rapid spread and, consequently, higher death tolls. It is this insight that will enable policymakers to better respond to public health crises with the potential to upend society, the economy and the political landscape. Notes: OLS estimates. Treated (Catholic) states are those where Catholics are the majority, i.e. Bayern, Baden-Württemberg, Saarland, Rheinland-Pfalz, and North Rhine-Westphalia. Control (non-Catholic) states are Schleswig-Holstein, Hamburg, Bremen, Niedersachsen, and Hessen. Standard errors in parentheses are clustered at the state level. Logged all-cause mortality rate since January 01, 2020, is the outcome variable. Days after first COVID-19-related death are state-specific. Regressions are weighted by the size of the state population. Clustered standard errors at the state level in parentheses. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. .050** (.022) .012** (.005) .012** (.005) .092** (.037) .019** (.008) Within county (α) .364*** (.013) .367*** (.014) .053*** (.004) .055*** (.004) .356*** (.031) .058*** (. Notes: OLS estimates. Outcomes are expressed as the logged count of cases (or deaths) per 100,000 local population. In Panel A, the local (logged) Catholic share is calculated relative to the county's total population. In Panel B, Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. Regressions are weighted by the size of the county population. Standard errors in parentheses are clustered at the county level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (2018) two-stage approach. In the first stage, the logged count of COVID-19 cases (or deaths) per 100,000 local population is regressed on a full set of county fixed effects, lagged number of cases and time trends. In the second stage, the dependent variable is the mean residual obtained from the fixed effects estimation in the first stage. The (logged) local Catholic share is calculated relative to the county's total population. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, population density, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. Regressions are weighted by the size of the county population. In columns [3] and [6], the local Catholic share is instrumented with the logged distance between the county's capital and Wittenberg, Germany. Robust standard errors in parentheses. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (2019). Standard errors in parentheses are clustered at the state level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (2018) two-stage approach. In the first stage, the logged number of COVID-19 cases per 100,000 local population is regressed on a full set of county fixed effects, lagged number of cases and time trends. In the second stage, the dependent variable is the mean residual obtained from the fixed effects estimation in the first stage. The local (logged) Catholic share is calculated relative to the county's total population. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, population density, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. County social interaction controls include the shares of those considering family as very important in their lives, those considering friends as very important in their lives, those living together with their parents and/or their parents-in-law, those who completely trust their family, those who completely trust people they personally know, those who live in households with maximum 2 members, those who regularly attend religious services, those who consider religion as very important in their lives, and those who believe that children should provide care for their parents. The predicted social interaction index was constructed from a principal component analysis using the social interaction controls. Regressions are weighted by the size of the county population. In columns [3] to [7], the local Catholic share is instrumented with the (logged) distance between the county's capital and Wittenberg. Estimation sample is restricted to Western German counties in which respondents were surveyed for the 2018 EVS wave. Robust standard errors in parentheses. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (March 22, 2020) . Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (2018); European Values Survey (2018). Notes: Catholics and non-Catholics are distinguished on the basis of self-reported religious affiliation in the ESS and EVS. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. Notes: OLS estimates. The local Catholic share is calculated relative to the county's total population. Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, population density, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. All variables are in logs unless mentioned otherwise. Standard errors in parentheses are clustered at the county level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. Outcomes are expressed as the logged count of cases (or deaths) per 100,000 local population. In columns 1-2 the local (logged) Catholic share is calculated relative to the county's total population. In columns 3-4 Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. We exclude counties that border on other countries and in a second specification all those that are within 20 km from a border. Temperature is measured in Celsius and Precipitation in centimetres and refer to the monthly average recorded in a weather station in a particular county. Specification also include state fixed effects and state-specific time trends. Regressions are weighted by the size of the county population. Standard errors in parentheses are clustered at the county level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. (2019). Standard errors in parentheses are clustered at the state level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively. Notes: OLS estimates. The local Catholic share is calculated relative to the county's total population. Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, population density, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita, and the number of vehicles per capita. All variables are in logs unless mentioned otherwise. Standard errors in parentheses are clustered at the county level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively.

Standard epidemiological modelling relies on variants of the Standard Inflammatory Response (SIR) model.

The idea behind SIR models is to predict the course of a pandemic over time based on the fractions of susceptible (S), infected (I), and recovered (R) individuals within a population, with . An important feature of such models is that they can be modified to estimate not only the within-county transmission rate, i.e. parameter α in Equation (2), but also differences in transmission rates based on some county-level observables. The seminal work of Adda (2016) popularised this sort of model in the economics literature by examining how policies, such as school closures, and economic interactions affect virus transmission. For example, if policies to mitigate the pandemic impact were enacted in some counties but not in others, their efficiency could be assessed by testing for differences in virus transmission rates between counties. This idea can be expanded to observe the differential impact of any kind of "treatment". Therefore, to make this adaptation in our setting, Equation (2) was modified to include an interaction between the term capturing the within-county spread of the virus and the Catholic county binary indicator. Although our paper focuses on differences in the observed incidence between Catholic and non-Catholic counties, documenting a difference in the spread of the virus using a modification of a SIR model adds credence to our results. A positive and statistical significant estimate of the parameter associated with this interaction term would indicate that the transmission rate was significantly higher in Catholic counties during the early pandemic.

The results are in Table A .6. Our outcome variable is the logged count of COVID-19 cases per 100,000 population, and the models are estimated using OLS. In columns 1-4, specifications include combinations of county controls, state fixed effects and state-specific time trends. As in Table 2 , past infections are significant drivers of current within-county COVID-19 incidence. Moreover, and in line with our results about Catholics exhibiting more frequent social interactions, the transmission rate is higher in predominantly Catholic counties, as indicated by the estimated coefficient of the interaction term. Notes: OLS estimates. Outcomes are expressed as the logged count of cases per 100,000 local population. Catholic (non-Catholic) counties are defined as those where the local share of Catholics is higher (lower) than their respective state average. County controls include the share of people over than 65 years old, the share of foreign-born, GDP per capita, the share of those completed secondary education, the average number of nights spent by tourists, the number of hospital beds per capita. Regressions are weighted by the size of the county population. Standard errors in parentheses are clustered at the county level. Asterisks ***, ** and * denote statistical significance at the 1%, 5% and 10% level, respectively.

Economic activity and the spread of viral diseases: Evidence from high frequency data

Source: Robert Koch Institute (RKI); Federal Statistical Office of Germany. Notes: Data on COVID-19 cases and deaths are as of April 20, 2020. Data on the local share of Catholics (at the county level) relative to the total county population refer to 2018. The correlation coefficient between the share of Catholics at the county level and the number of cumulative cases and cumulative deaths per 100,000 local population is .416 and .315, respectively.