Introduction

While researchers acknowledge the important adverse effects of violence on economies, it should be recognized that most causes of violence are still not well understood from a scholarly perspective (Heinemann and Verner 2006; Murdoch and Sandler 2004; World Bank 2003). This work aims to study the interconnection and mutual influence of a triad, namely, religion, violence, and sport. Despite many claims regarding the supposed decline of religion across the (Western) world in contemporary times, mainly put forward by proponents of secularization theory, and the expectation of a modernity completely devoid of any form of organized religion, religion continues to play a pivotal role in many people’s lives (Casanova 2006; Stark 1999). Recent estimates suggest that the vast majority (about 84%) of the world’s population sees itself as belonging to a particular religion (Pew Research Center 2017).

Violence is regarded in different ways across different cultures. The boundaries of what is considered acceptable and non-violent may shift greatly in different times or in different cultures. There are thus many reasons to believe that what is considered violent may also change with regard to religious faith, given that it is part of culture, and possibly one of great importance.

Sport is often defined with metaphors referring to war (Segrave 2000); some authors have defined sport as a “war without weapons” (Mangan 2003). It has also been suggested that in societies that are at peace, sport plays the national identity-building role that was traditionally played by war (Smith Porter 2004). Furthermore, many of the most passionate sports fans see strong analogies between sport and war (End et al. 2003). One strand of the literature (Blanchard, 2013; Mazurkiewicz 2011) suggests that sports are rituals that can be considered as the sublimation of ancient rites and actions that involved violence, such as sacrifice. All this seems to imply that the study of the relationship between violence and religion, even in the framework of sport, has greater scope and larger consequences than what is related strictly to sports events. Finally, it is interesting to note that it has been suggested that sport is a form of religion itself (Hoffman 1985); this makes it very interesting to study the interaction between religious faiths and this para-religion with regard to violence.

All the above indicates the value and interest of pursuing the main objective of this exploratory study, which is to shed some light on the relationship between national religion and violence, using sport as framework. Our research question can be stated as: “Does a given national religion increase, decrease or have no effect on violent behaviour?” We will investigate this research question with respect to Christianity, Islam, Judaism, and Buddhism, as well as to some of the major confessions within the first religion.

Although the empirical focus of this work is based on an analysis of sports performances, we believe we contribute to a wider literature, specifically the strand concerning the impact of religion on human behavior (see Guiso et al. 2003; McCleary and Barro 2006). Since Max Weber’s The Protestant Ethic and the Spirit of Capitalism, this literature has grown substantially in the social sciences. Over time, it has expanded to include more general references to culture, rather than those that refer only to religion. As regards the social sciences, this literature seems to have grown in sociology (Banfield 1958) more than it has in economics (see Guiso et al. 2006, for an example); this is probably due to the fact that it is harder in this setting to employ typical quantitative economic tools of analysis. Indeed, scholars have found it difficult to disentangle the effects of culture empirically from other possible determinants (such as legal institutions or poverty) that drive violence. Following the literature on the subject (Miguel et al. 2008), we believe that soccer, one of the most popular sports in the world (Boeri and Severgnini 2012), which fills stadia with hundreds of thousands of spectators every week, and which was watched by four billion people on television (Caruso et al. 2017), or five, as projected by Gianni Infantino for the Qatar 2022 World Cup, offers an interesting quasi-experimental setting (Altman et al. 2021) in which to investigate this. Indeed, apart from being of interest to sports science, since it may help to clarify the outcome and determinants of top-tier sports events, this study also contributes to the economic literature studying the relationship between religious affiliation and violence (Canetti et al. 2010; Murphy 2011).

It has been argued that social scientists, in building their models, are concerned mainly with three aspects, namely, incentives, institutions, and outcomes. At the same time, economists mostly agree that behavioral factors and personal idiosyncrasies may play a role in and impact the decision process. One of the factors most likely to influence this context is religion, the variable of interest for the present work. Even if its role has been analyzed in numerous studies (e.g., Guiso et al. 2003; McCleary and Barro 2006), its impact on human behavior remains unclear, given the ambivalent nature of human beings. As Altman (2012: p. 51) points out, “people aren’t calculating machines. Instead, they’re decision makers driven by both passion and reason.” Adopting this logic, a well-known topic derives from the link between religiosity and the probability of salvation, that is to say the influence of religious participation and beliefs on the chance of being saved in the afterlife (Azzi and Ehrenberg 1975). Nevertheless, a less often observed but very interesting phenomenon, as stated by McCleary and Barro (2006: p. 51), is “[…] the powerful force from afterlife beliefs [that] can promote anti-social actions, such as violence – the so-called ‘dark side of religion’.”

Our study aims to assess the impact that religious identity has on behavioral interaction between individuals, with specific reference to aggression and conflict. More relevant to our purposes is the specific context we set our research in, from which it is possible to derive more general and meaningful insights.

Among the prior studies that have had the greatest influence on our work, those that have adopted a framework most similar to ours are Miguel et al. (2008), Caruso et al. (2015), Caruso et al. (2017), and Alfano and Capasso (2020). Miguel and his co-authors perform an analysis on data from soccer club violence, proxied by penalty cards, and years of war in the home country. They found a strong relationship between a history of civil conflict in the home country and a propensity to violence on the field. The two studies of Caruso et al. 2015, 2017) investigate the effects of national identity on violence and on the determinants of aggressiveness. The second of these studies includes a short analysis regarding religious distinction, but this is operationalized as difference in religiosity, and not as difference in religious denominations. Both studies use data gathered from soccer matches. More specifically, Caruso and his co-authors build a dataset from the official matches of FIFA national teams, which is aimed at studying the country of origin’s effects on violence in one case, and the determinant of aggressiveness in the other. Alfano and Capasso (2020) perform, from a theoretical perspective, a similar analysis, applying a similar logic, but with a different aim: its purpose is to extend the results of an experiment run in a specific (sports) framework (in their case the table-soccer performances of top athletes by nationality) to a more general level (namely, predicting the performances of workers that come from countries with a different set of rules in an international environment). As should be clear, the first of these studies is of particular interest to our research; indeed, mutatis mutandis, the study of the effect that belonging to a given nationality has on violent behavior when playing soccer may be considered similar to studying the effect of belonging to a given religion. Please note that, to the best of our knowledge, there is currently no specific focus in comparable studies on the effect of religion (other than, as mentioned above, the very generic operationalization of religious distinction in Caruso et al. 2017), while adherence to a given confession and the nation’s degree of religiosity may be implemented in such an analysis with very interesting results.

As in the four papers just cited, in this work too we gather data from soccer to perform our analysis. There are many reasons for choosing soccer as the framework in which to apply our empirics. As highlighted, soccer is the most popular sport in the world, and its World Cup, held every 4 years, offers a great example of top-tier athletes of the discipline coming from all over the world to challenge each other.Footnote 1 This is seemingly an ideal framework from which to derive information from people coming from various backgrounds, including different religions. Furthermore, gathering data from national teams helps us to obtain a set of players among whom it is (more) acceptable to assume a certain degree of influence by the national religion, especially in terms of attitude and behavior.Footnote 2 This assumption would of course be much more difficult (not to say impossible) were it to be made in a soccer club context, where players from very different national backgrounds play together. Because of this, and the peculiar relationship that exists between soccer and violence, we consider soccer to be the most appropriate context in which to investigate our research questions. Indeed, it is a sport that allows contact between players, unlike what basketball rules prescribe, for instance; at the same time, it does not tolerate acts that would be considered as physical violence in everyday life, as in the cases of American football or rugby. In the rest of this work, by studying the impact that national religion exercises on a top-tier event of this sport, the FIFA World Cup (which is arguably soccer’s most important tournament, at least for national representatives), we will deduce more general relationships about the relationship between religion and violence.

More precisely, in line with the argument of recent literature that the sports context is valid as a quasi-experimental setting (Caruso et al. 2015, 2017; Alfano and Capasso 2020; Altman et al. 2021), in this work, we will shed some light on the impact of the national religion in a number of given countries (i.e., those participating in the final phase of all the soccerFootnote 3 World Cups held in history) on the violence experienced in the matches that their national teams played. While exploiting sports data in this setting may at first glance appear unusual, it is worth noting that there is a consolidated (and, more importantly, steadily growing) literature that uses this kind of quantitative method, especially in economics (Caruso et al. 2015 and 2017; Alfano and Capasso 2020, Altman et al. 2021).

But as well as accepting this growing academic trend in quantitative applied social sciences, several reasons lead us to believe that sports data are reliable, and that it is a good idea to use them to proxy various phenomena. It could be argued that sports players behave more alike than how a typical agent would behave in a real-life extreme situation, and thus, this type of data possibly gives researchers more realistic and accurate feedback than the data gathered from answers to a survey, for instance, done online or over the phone, where it is hard to imagine that the interviewee puts great effort into providing the best possible answers. Sports data also seem to be closer to a real-life response than the observations made by researchers about role-play games carried out in controlled settings in universities. All these situations (from which, it is important to note, data are regularly gathered and routinely used in research, and are far more common and popular in the literature than their sports counterparts) possibly provide data that are less close to reality than those obtained from behavior observable in a sports setting, especially a highly competitive one.

To shed some light on the (possible) impact of national religion on violent behavior, we take a quasi-experimental approach (Altman et al. 2021) that relies on religious data from Religious Characteristics of States Dataset Project: Demographics v. 2.0 (Brown and James 2019) and soccer data from the FIFA World Cup. More precisely, we analyze the last 14 tournaments of the most important soccer competition, from 1970 to 2022 (i.e., those for which data on foul cards including yellows are available). It includes a total of 83 teams from all the continents (a summary of the teams participating in each tournament is included in Appendix 1, while a summary of the representation of each continent in each tournament is included in Appendix 2).

The study of the possible impact of the national religion of the team playing, or the main confession of each faith investigated, on violence, may shed light on two different areas, namely:

  1. i)

    The general existence of a relationship between violence and national religion, measured in a quasi-experimental setting (such as sport);

  2. ii)

    The determinants of a soccer or—more generally—a sports match, which may be affected by the national religion of the players, as well as by other factors.

While it may be tempting to assume that the religious make-up of each team is broadly similar to the religious make-up of its nation, the absence of micro data does not allow us to make this assumption credibly. While for some countries with low levels of religious fractionalization, such as Italy, this may be acceptable, it becomes easily affected by biases for single players from a different religious background. Nonetheless, there are reasons to believe that the national religion does have an impact on this dynamic, regardless of the personal beliefs of the players. For instance, the teams could be responding to their fans, and hence playing according to expected norms, regardless of the specific religious leanings of the team and its individual members.

Following this introduction, the rest of the work is organized as follows. The next section defines the main concept used in this research, namely, violence. The third section deals with the data used and the methodology adopted in the empirical estimation of the relationship, describing the sources and the operationalization of the different variables. The fourth section presents the results of this empirical estimation. Finally, the last section concludes and indicates possible future lines of research.

On the Operationalization of Violence

To perform our analysis, it is vital to have a clear operationalization of what violence is, and in this regard, the sports setting seems highly suitable. Indeed, in sport, there are clear rules that state what is legitimate and what is not, and referees who oversee those rules (Goff and Tollison 1990). This has not always been the case: soccer’s various rules and procedures have developed over a long time through erratic and complex processes. Nowadays, fouls and misconduct are addressed in Law 12 of the so-called Laws of the Game that regulate soccer. These are acts committed by players deemed by the referee to be unfair and which may subsequently be penalized. In Fig. 1, an Euler diagram shows the relationship between fouls and misconduct in association soccer, with a few examples. Finally, the colored card system, which was introduced in soccer in 1966 to indicate the seriousness of the punishment committed, was used for the first time in the 1970 FIFA World Cup. This is also the first year that FIFA collected statistics about yellow cards, which were not available before. Thus, to avoid biases, our analysis will start in this year.

Fig. 1
figure 1

Source: https://commons.wikimedia.org/wiki/File:Foulsandmisconduct.svg

An Euler diagram showing the relationship between fouls and misconduct in association soccer, with examples.

Full size image

In soccer, a yellow card is a caution, issued to a player who has committed misconduct or some other serious offense. A total of six offenses can lead to a yellow card, namely, unsporting behavior, dissent by action or word, continuous infringement of laws, delaying play restarting, not standing the required distance away from a spot-kicker, and entering or re-entering the field. While not all of these refer to physical violence, all of them refer to some sort of violence, at least verbal, or psychological (indeed, the aim of some fouls is surely to pressure the adversary psychologically). Examples of this are insulting the opposing side’s supporters (an example of unsporting behavior), insulting or mocking the referee (dissent by action or word), continuous penalizable behavior (continuous infringement of laws), and more generally not respecting the opponent (delaying play restarting, not standing the required distance away, and entering or re-entering the field). Therefore, we believe it is appropriate to consider also a yellow card linked to violent behavior.

On the other hand, a red card in soccer means that the player has committed a serious foul or has committed serious misconduct in other ways. In this case, the player has to leave the field of play immediately. There are five principal reasons for a red card to be given, namely, serious foul play, spitting and abusive language, violence, deliberately fouling, or receiving two yellow cards. Once again, all of these reasons signal the use of violence by the player, possibly of a greater magnitude than what earns a yellow card.

Methodology and Data

It should be clear at this point that answering our research question via a quantitative analysis is a daunting task. Indeed, the scarcity of empirical analysis examining institutional controls on conflict (Caruso et al. 2017: p. 512) makes it very difficult to use these lenses to examine the topic under investigation here.

Nevertheless, it is also important to highlight that our understanding of the role of social and cultural norms in driving violence is particularly limited, and that this is due to inherent empirical challenges. As suggested by Miguel et al. (2008), controlling for non-cultural factors in a regression framework is very complicated, since both formal legal and informal cultural restraints on violent conduct reinforce each other. This means that it is plausible to assume that these factors are correlated across societies (Waldmann 2007). Laws and income levels (two variables that are very likely to have an impact on violence) differ across societies and are very likely to interact with culture in complicated ways (Becker 1968).

One of the possible strategies to reduce the extension of this problem is to adopt a quasi-experimental setting. As explained, there are many reasons to believe that sports events are very close to an ideal experimental setting, which is arguably the best source of data for quantitative analysis (Altman et al. 2021). Indeed, it has been argued that sports competitions produce cleanly generated data, and many potential sources of spurious relationships are controlled by rules and referees (Goff and Tollison 1990). At the same time, professional sport is typically a winner-takes-all competition, in which careers, pride, honor (and a lot of money) are at stake; this creates a huge incentive for each agent (i.e., the different players) to play at their very best.

In more formal terms, we empirically test the impact of religion on violence by estimating the following equation:

$${Vio}_{imt}=\alpha +{\beta }_{1}{X}_{it}+{\beta }_{1}{Z}_{t}+{\beta }_{3}{W}_{im}+\varepsilon$$
(1)

where our dependent variable Vio is a proxy of the violence in each soccer match m of team i in year t; X is a matrix composed of the variables that vary by each team i and year t (these are HSoc, DHost, DWin, Match, and Rel; more details on these variables further below); Z is a matrix composed of a set of dichotomous dummy variables that control for fixed effects of each year t and D32 (more details on these variables below); W is a matrix composed of the variables that vary by each team i and match m (these are AdvVio, DLos, DHR, DSFin, and DFin); more details on these variables below). Finally, as usual, \(\varepsilon\) is the error term.

Hence, assuming that the equation controls for all the relevant determinants of violence, the impact that national religion has on it is expressed by the relative coefficient, which expresses the magnitude of the impact of Rel on Vio.

To build our dataset to estimate Eq. (1) empirically, we gathered data from three different sources: the World Cup Archive, a website for data regarding teams’ behavior during World Cup matches (http://www.rsssf.com/tablesw/wcf-full-intro.html), integrated with the FIFA website for the 2018 and 2022 tournaments; the Religious Characteristics of States Dataset Project: Demographics v. 2.0, for data about the religions followed in various countries in various years; and finally Kaggle, for data on the historical tradition of soccer in various countries.

Data

Vio is operationalized in formal terms for each team i in each match m as:

$${Vio}_{im}={YCards}_{im}+{YYCards}_{im}+{RCards}_{im}$$
(2)

where YCards is the sum of yellow cards each player belonging to team i in match m received from the referee; similarly, YYCards is the sum of second yellow cards (i.e., yellow cards awarded to players who had already been punished with a yellow card in the same match) obtained by all the players playing for team i in match m; and, once again, RCards is the sum of red cards received by players of team i in match m.

Vio is thus a measure of how violent team i has been in match m. Nonetheless, violence may be operationalized in different ways: Appendix 3 deals with this issue, showing how different operationalizations of violence lead to equivalent results.

Is this a proper operationalization of the violence that occurs in a match? As previously highlighted, in soccer, red and yellow cards represent the referee’s recognition of violent behavior, whether physical or verbal, among the players. Unfortunately, we are not aware of any available statistics on the reasons why cards were issued; these would be useful in refining the data. Nevertheless, it seems quite acceptable that, at the very least, the total number of cards (or their alternative elaboration proposed in the appendix) is highly correlated with a team’s violent behavior. If this condition holds, please note that our results hold as well, since what we require to perform our analysis is a high correlation, in order to give the variable a valid operationalization. Vio thus seems to be a good proxy of the average amount of violence from team i in match m.

Moving on with the description of the variables, HSoc is operationalized as the number of years that have passed since the first game played by the national soccer team of a given country,Footnote 4 using data gathered from Kaggle. The underlying idea here is that the length of a tradition of playing soccer is a good proxy for the popularity of the sport in a given country (Alfano and Capasso 2020), and that this is to be preferred to FIFA Rankings, due to the bias of the latter (Macmillan and Smith 2007). This means that HSoc is a measure of the soccer tradition in country i, and thus arguably of the ability of the players from this country, and the sophistication of its strategies and techniques. For these reasons, we expect HSoc to have a negative coefficient, since more experienced and skillful teams should be less violent while playing, whereas teams with younger traditions may rely on more violent aggressiveness to try to win the match.

The next variable included in the equation is DHost. This is a dichotomous dummy variable which assumes the value 1 if in year t the country i is hosting the World Cup, and 0 otherwise. In this case too, the original information is obtained from the FIFA website and transformed into a dummy variable. The rationale behind the choice of including this variable in the equation is that playing at home, in front of many supporters, has been proven to create a social pressure that influences refereeing decisions (Pettersson-Lidbom and Priks 2010), and thus, it may have an impact on the degree of violence measured on the field. For this reason, it seems important to control for this variable, since it may capture an important variability. We expect the variable to have a negative coefficient.

Similarly to the previous variable, DWin is a dichotomous dummy variable that assumes the value 1 if in year t the country i wins the World Cup, and 0 otherwise. Once again, data are obtained from the FIFA website and are operationalized by the authors into a dichotomous variable. The team i that wins the World Cup is arguably the best in year t; for this reason, it seems important to include a dummy that captures this information and allows us to control whether this is a determinant of violence. We expect this variable to have a negative coefficient, since the best team should probably rely less on violence in its matches.

D32 is also a variable built from FIFA data and is equal to 1 if the given tournament includes 32 teams (i.e., all FIFA World Cups after 1998). It seems important to control for this information given that the number of teams playing and the number of matches played may both influence the dynamic of violence.

Data for the construction of the Match variables are, once again, derived from the FIFA website. These variables are operationalized as a matrix composed of four dichotomous dummy variables, namely, MATCHES PLAYED=4, MATCHES PLAYED=5, MATCHES PLAYED=6, and MATCHES PLAYED=7. This matrix discriminates through four dichotomous dummy variables for the five possible modalities that the variable MATCHES PLAYED may assume, namely, 3, 4, 5, 6, or 7. The modality equivalent to three matches played (i.e., the value that Matches assumes for all the teams that are eliminated after the first phase) is omitted, to avoid having perfect collinearity and thus falling into the so-called dummy trap issue. This set of variables should help us to discriminate between teams with very different levels of ability. Other than the soccer tradition, which we have already taken into account with the HSoc variable, it is possible that in a given year t, a specific team has one or more very strong players that may put the team in a better league than usual. On the other hand, it is also possible that in a specific tournament of the World Cup, a team that is usually very strong and comes from a country with an old soccer tradition arrives with a team that performs below its usual level due to injuries of key players or other contingencies. Furthermore, in each year, the path of each team playing the FIFA World Cup may also depend on the structure of the play-offs: the relative strength of the teams facing each other in the single elimination round is a (relative) matter of chance, and may play a role in the consequent values of Vio, explaining a part of its variance. In all these cases, the violence observed may vary. Thus, this objective measure of the path of each team i in a given tournament of the World Cup t gives us a precise measure of the relative strength of the teams, regardless of their history. Please note that as the modality referring to three matches is omitted, all the coefficients are expressed as relative distances from this modality.

The team that is losing a match may be more violent as it tries to find a way back into a match, or simply out of frustration. Unfortunately, we cannot control for the score at the moment of the card, due to lack of data, but we added DLos to the analysis, a dichotomous variable that assumes the value of 1 for the team that lost the match, to control for this dynamic.

Matches between historical rivals may also be more violent; for this reason, we included DHR, a dichotomous variable equal to 1 if the two teams have a rivalry.Footnote 5 We also expect this variable to be positively related to violence, due to acrimony.

Given that the further one advances in the competition, the greater the consequences of receiving a card become, both for the team and for the player personally, who may lose the opportunity of playing in a career-building match, we also included two dichotomous variables, DSFin and DFin, which assume the value of 1 if the match is a World Cup semi-final or final (either for first or for third place).

As the old saying goes, it takes two to tango. For this reason, we consider it appropriate also to control for the violence from the other team and measure this in AdvVio (which in the regressions is proxied, to avoid biases, in the same way that Vio is computed). We expect this to be positively related to our dependent variable. This is a potential source of endogeneity, since Vio may affect AdvVio. Nonetheless, a robustness check performed removing this variable from the estimation showed that this did not affect the results; hence, we are confident that this potential bias did not influence our findings.

Finally, Rel is the operationalization of the national religion of the team. As already discussed, we chose to apply our empirics to the FIFA World Cup as this is the most important competition for soccer, which is a very suitable game in which to investigate this relationship, and because the basic statistical units of observation are the different teams, which are national representatives. Unlike soccer clubs, national representatives are constituted of players that share the same nationality, and who thus generally come from the same country (with some minor exceptions, such as naturalized players, who we assume to be irrelevant to our analysis given the lack of precise data), or are otherwise subjected to the same national pressure, which is based on shared norms and culture, which are in turn influenced by religion.

It is tempting to assume that all the players of the national team share the same religion. Unfortunately, we do not have access to data that is sufficiently granular to support such an assumption. One may however assume that, while it would be unreasonable to state that the members of the national soccer team constitute a representative sample of the country, there are no specific reasons to believe that they differ from the general population with regard to their religious beliefs. The ability to play soccer is distributed as a random normal variable, or at least it should be, according to the Central Limit theorem (La Place 1814; Gnedenko and Kolmogorov 1954; Le Cam 1986), which is the sum of many different random normal variables.Footnote 6 Stochastically speaking, this should mean that if a random observation is extracted from this set (i.e., doing what the national team coach does when she or he selects a player to join the national representative), he should have the same characteristics as the rest of the population from which the unit is extracted (i.e., the whole country) for variables that are orthogonal (i.e., completely unrelated) to the ability to play soccer. Outliers should be irrelevant asymptotically, and their bias should be captured by the error term.

Nonetheless, in this study, we do not aim to measure the impact of personal religion on violence, but on the contrary the impact of the national religion. Some of the countries included in the analysis have very low religious fractionalization, and our assumption could hold in a place like Italy, for instance, which is mostly Catholic. Other countries are of course much more diverse, and in these, the assumption starts to become unrealistic. And yet there are reasons to believe that the national religion has an impact on this dynamic, regardless of the players’ individual religious affiliations. Players respond to their fans and hence are incentivized to play according to what their public expects, irrespective of the specific religious leanings of the team and its members.

Thus, if we measure the impact of the share of believers in a given religion in a country on violence, the relative coefficient should measure the impact of each of these religions on the dependent variable Vio. In other words, the relative coefficient answers the question: does a greater (lower) share of believers in religion X increase (decrease) violence?

This, essentially, is our research question. Please note that it is not possible in the empirical estimation to include all the shares of believers at once. This is for two reasons, principally: first, it is unrealistic to assume that the shares of believers in the four main religions in each country are variables without any inter-correlations: of course, the higher the share of believers in a monotheistic religion, the lower the share of believers in another will be. Second, including all four variables may easily lead to multicollinearity issues, since the combination of believers in the four main religions represents a share of the total population that is very close to the total in several countries. Accordingly, we have considered it more appropriate to include only one religion at a time in the estimates, operationalized in Rel in each specification of the model. In other words, while running the regression to estimate the coefficients empirically, rather than including several Rel for the different shares of religion in one regression, we re-define Rel as the share of each different religion in any single specification of the model, to estimate the relative impact of any given national religion on violence.

To operationalize these concepts into measurable variables, we gathered data from the Religious Characteristics of States Dataset Project: Demographics v. 2.0 (RCS-Dem 2.0), in its COUNTRIES ONLY specification. The RCS-Dem 2.0 dataset reports different estimates of religious demographics, country by country. According to Brown and James (2019), they created RCS-Dem 2.0 to fulfil the need for a dataset on the religious dimensions of countries of the world, with the state-year as the unit of observation. It covers 220 independent states, for every year from 2015 back to 1900. In this respect, it is important to highlight that the most recent data available (i.e., data from 2015) are in our dataset imputed to the World Cup of 2018 and 2022, given the lack of more up-to-date data. RCS-Dem 2.0 estimates populations and percentages of adherents of a hundred religious denominations, including second-level subdivisions within Christianity and Islam, along with several complex categories (such as Western or Eastern Christianity).

Considering that, other than what already discussed that as explained does not affect the results, we do not expect any other source of endogeneity, since the violence expressed on the field should not be influenced by the religion in any other way than the mechanism we have explained, our regression, performed with robust standard errors to correct for the potential heteroskedasticity, should not be affected by this potential source of bias. In the absence of endogeneity, the best linear unbiased estimator of the coefficients is given by the Ordinary Least Squares (OLS) estimator, provided it exists (Neyman 1934). At the same time, it may also be argued that our operationalization of violence, as a count variable (sums of cards), is not distributed as a normal variable, and a Poisson distribution would be more approximate (Kingman, 1992). We accordingly chose to estimate our model through a maximum-likelihood model for count data (Cameron and Trivedi 2013). This model allows us to analyze a count variable more effectively, i.e., a variable that measures an individual piece of count data.

While our data are gathered from FIFA World Cup tournaments held in different years, and we control for the tournaments with the matrix Y, our analysis is cross-sectional, since a time dynamic seems unlikely.

Slight differences can be seen in the number of observations in the different regressions. This is due to the fact that in some countries, Rel is equal to 0 for some religions; in other words, there are no people that belong to a given religion. Considering that we only included one religion at a time in each regression, as discussed above, to avoid multicollinearity issues in our analysis, it seems more appropriate to have the total number of observations vary slightly (considering that, as explained, this happens when the probability of having a member of this religion in a team is equal to 0), rather than proceeding with the listwise deletion of teams, which nevertheless may be useful and help to explain the violence of other religions. Table 1 reports descriptive statistics about the variables included in the analysis.

Table 1 Summary statistics
Full size table

Results

Estimates of Eq. (1) are reported in Table 2, concerning the four main religious groups present in our data, which are also four of the six most common religions in the world, namely, Christianity, Islam, Judaism, and Buddhism. Unfortunately, in our regression, we had to restrict the analysis to these four religions, since in the countries with teams that have participated in a FIFA World Cup other religious denominations do not have a representation that is substantial enough to provide a statistical power sufficient to justify its inclusion in these estimates. Furthermore, by including religions that are followed by only a tiny share of believers in our sample population, we may easily misinterpret the result.

Table 2 OLS–All denominations–Vio
Full size table

As can be seen in Table 2, in all four of the specifications, the variable HSoc is negative and statistically very significant, as we expected. This suggests that a longer soccer tradition in a country leads to its teams being relatively less violent in the World Cup.

The variable DHost is not statistically significant, while that for DWin is. This suggests that hosting the FIFA World Cup has no statistically significant effect on violence, while being the team that wins it has a negative impact on violence (a finding in line with our hypotheses). The finding related to DHost, meanwhile, may be due to the smaller influence of social pressure (identified by Pettersson-Lidbom and Priks 2010 as the mechanism through which home teams are less likely to receive cards) on the best possible referees, i.e., those presiding over the FIFA World Cup.

On the other hand, the D32 variable has a positive and statistically significant coefficient, suggesting that in the tournaments with more teams, there has been more violence. This may be due to the greater number of matches played by the teams, which means there are simply more occasions to receive cards. On the other hand, since the tournaments with 32 teams are the most recent (since 1998), it may also signal that in more recent years, there has been an increase in violent behavior. It is finally important to highlight that, under the assumption that the qualifiers select the best teams in the world, the higher the number of teams allowed into the final phase, the lower the average skill level will be. Hence, the sign of this coefficient may suggest that violence increases when teams with a lower skill level are added to the World Cup (possibly as a result of desperate measures taken to try to compete with technically superior teams).

As discussed above, while it is important to include the Matches matrix to control for this further characteristic of the teams, it is hard to derive theoretical insights from the empirical estimation of these variables, which are nevertheless statistically non-significant.

DLos is positive and statistically very significant, suggesting that being the team that loses has an effect on violence, and so is DHR, the dummy variable discriminating for matches between teams that have a rivalry, as we expected.

Contrary to what we expected, playing in the semi-final or the final, as captured by DSFin and DFin, does not have any statistically significant impact on violence.

Our analysis also suggests that it does indeed take two to tango, since AdvVio is positive and statistically very significant. The more violent one team is, the more violent its adversary, suggesting that there is a reciprocal relationship, as expected.

Finally, we come to the variable of greatest interest for the present study, which is the operationalization of Rel. Its four different specifications assume the share of Christians (2a.1), Muslims (2a.2), Jews (2a.3), and Buddhists (2a.4). We may observe that there is a statistically significant impact in two of the four specifications (namely, in 2a.1 and in 2a.4). Indeed, our analysis shows that the more Christians there are in a country (belonging to any possible confession in this operationalization), the more violent its soccer team is likely to be (as presented in 2a.1). While the magnitude of the coefficient may not seem especially big (0.240), it is statistically significant at 10%, which is interesting considering both the number of observations (1523) and the large number of controls included. It is also worth noting that Vio is operationalized for the number of cards (and thus, indirectly, the number of fouls) in a single game; this means that this statistically significant positive coefficient plays a role in the long run. Furthermore, as already explained, our main aim is not to derive insights into the dynamics of soccer, which are only a side result of this research, but instead to use soccer as a quasi-experimental setting from which to derive more general insights about the relationship between religion and violence. We may thus state that an initial finding suggests that the more Christians there are in the nation represented by a team, the more violence there will be in a game played by that team, and from this, it can be deduced that belonging to this religion (rather than the others) has an impact on violent behavior.

When operationalizing Rel as the share of Buddhists within a country (2a.4), we found the opposite result. Apparently, the share of Buddhists in the nation a team represents has a statistically significant negative impact on violence. This suggests that the more Buddhists a country has, the less violent its national team is likely to be. This coefficient has a higher statistical significance than the previous case (5%), and therefore, similar reasoning may be adopted. Once again, if we want to generalize the result, we may state that this finding suggests that the influence of Buddhism as a national religion implies being less violent (compared to other religions included in the analysis).

On the other hand, when Rel is operationalized as the shares of Muslims and Jews (in specifications 2a.2 and 2a.3, respectively), there is no statistically significant impact of the variable Rel on violence Vio, although both the coefficients are negative. This suggests that a greater share of people belonging to either Islam or Judaism does not have any statistically significant impact on the level of violence in games played by their national teams.

We decided to explore these initial results further by investigating whether there was any difference in impact on violence from the various confessions belonging to Christianity, the religion that reported a statistically significant impact on violence, and which has significant shares of confessions among the teams analyzed. These are presented in Table 3.

Table 3 OLS–Christian denominations–Vio
Full size table

In order to operationalize Rel with regard to the different Christians confessions, we rely on the RCS-Dem 2.0 distinction. It considers the Western Christian category as including Latin-Rite Catholics, Extended Protestants, and most Liminal Christian denominations. These three branches have several characteristics in common; for instance, they possess a common scriptural root (in the sense that they all consider the New Testament to be canonical). Furthermore, they dominate and coexist in largely the same geographic regions, and possibly more importantly, they share a common heritage (for instance, Protestantism and Liminal Christianity split off from their parent denominations about 500 and 200 years ago, respectively). Although there is no doubt that these confessions differ in certain beliefs, structures, and religious institutions (for example, it is clear that Catholicism is organized in a considerably more hierarchical way than Protestantism), they also have similar ethical structures, including political ethics. This is a difference that has been taken especially into account in the literature on security studies, which considers the war ethics of these confessions as being similar enough to treat all of them as a single group (Brown 2008; Cahill 1994; Johnson 1975, 1987). We thus consider it to be a matter of interest to test the impact on violence of belonging to this confession.

On the other hand, the various denominations of Eastern Christianity share fewer characteristics than the Western denominations. Nevertheless, it is included as a group, since it seems important to have a group to counterpose to the Western one, and thus, this may be read as its complement. To further investigate Western Christians, two important confessions are extrapolated from this agglomeration because of their relative importance, namely, Catholics and the sum of Protestants and Anglicans. Once again, while it is possible in theoretical terms to distinguish between several different branches in each of these two confessions, statistically speaking, it is very unlikely that we would obtain a meaningful result from a tiny share of believers. For these reasons, it seems appropriate to consider Protestants and Anglicans together as a unique group.

As can be seen in Table 3, when operationalizing Rel as the different shares of population belonging to the different Christian confessions, Eastern Christians (3a.2) have a positive coefficient that is statistically significant at 5%, while those of the other three confessions are not statistically significant. All the other independent variables keep the statistical significance and the signs that have already been commented on, suggesting some robustness in the model and its results. This suggests that Eastern Christianity has the most violent impact on religion among Christians.

Results of the regression estimated with Poisson, presented in Tables 4 and 5, show equivalent findings, and so does an estimation of Eq. (1) with ANOVA or Fractional Probit estimators.Footnote 7

Table 4 Poisson–All denominations–Vio
Full size table
Table 5 Poisson–Christian denominations–Vio
Full size table

Conclusions

In the present study, we have offered insights into the relationship between religion and violence by means of a quantitative analysis that employs quasi-experimental data from soccer. By studying the violence recorded in FIFA World Cup tournaments, we found a statistically significant correlation between violence and the shares of people belonging to a given religion in the country the team is from.

More precisely, we found that the more Christians there are in a country, the more violent that country’s national team will be; on the other hand, the more Buddhists there are, the less violence is observed when the national representative plays. There are no statistically significant results for the other two religions studied, Judaism and Islam. We also explored this relationship with regard to some of the different confessions within Christianity. In this case, Eastern Christians were the only group that appeared to be statistically significantly more violent.

It is important to highlight that while this study contributes to the literature using sport data to answer questions of wider relevance, some shortcomings may jeopardize the quality of our results, and we urge the reader to approach them with caution. Most importantly, our data are observational, and not the result of a sampling operation. For this reason, some biases may affect the results, which may be driven by some hidden dynamic that consistently keeps some teams in or out of the FIFA World Cup. Moreover, our estimation could be affected by endogeneity, even though a robustness check suggests that this source of bias should not influence our findings.

Nonetheless, accepting the fact that our soccer data are quasi-experimental and that the relevance of this analysis may not go beyond the sport in question (though even if that is the case the results may be of some interest and help in determining the results of a world-class soccer event), our results, accepting our assumptions and with all the caution that is due in such a study, show that on average being from a nation where Christianity is more prevalent is positively related to being more violent (compared to all the other religions included in the analysis), while on the contrary Buddhism is associated with being less violent.

This finding may be explained with a certain group of values that influence the way people look at life. Religion may be considered as an important part of culture, and without any doubt plays a role in how believers look at life and thus how they behave.

Future studies may aim to expand these findings and make our results even more robust by extending the analysis to include other sports or expanding our analysis by studying a larger number of tournaments in the FIFA World Cup. Also, analysis based on the interaction of the religion of the two teams, or of that of the referee, are interesting additions that could be worthy of study in the future.