Skip to content
Publicly Available Published by De Gruyter April 29, 2022

Thinning out spectators: Did football matches contribute to the second COVID-19 wave in Germany?

  • Kai Fischer ORCID logo EMAIL logo
From the journal German Economic Review

Abstract

The COVID-19 pandemic has decelerated substantial parts of economic and human interaction. This paper estimates football matches’ contribution to the spread of COVID-19 during Germany’s second infection wave in summer and autumn 2020. Exploiting the exogenous fixture schedules of matches across German counties in an event study design, we estimate that one additional match in a county on average raises daily cases by between 0.34 to 0.71 cases per 100,000 inhabitants after three weeks. Hence, this implies an increase of the seven-day incidence per 100,000 inhabitants by around three to seven percent. We do not find qualitatively different results for a subsample of German top league matches with the strictest hygiene regulations or matches with higher occupancy levels. Notably, the found effect is mediated by the incidence level at the day of the match with very few infections for matches at a seven-day incidence below 25. Using mobile phone data, we identify strong increases in the local mobility as an underlying mechanism. We finally show that the ban of away fans successfully limited the spread of COVID-19 beyond county borders. Our results alert that even outdoor mass gatherings can remarkably cause infections.

JEL Classification: I18; H12; Z20; Z21

1 Introduction

Since early 2020, countries worldwide have been fighting COVID-19 and its impact on their economy and public health. Even with vaccines, the pandemic’s end remains uncertain with regard to persistent risks such as new virus mutants. Similarly, deriving policy implications for potential future pandemics seems to be more important than ever before.

A persistent concern of businesses and policymakers throughout the crisis has been to find ways to adapt to the new circumstances and to reorganize public life. One of which is to allow events and activities in public but under additional, pre-emptive measures (McCloskey et al. 2020; Memish et al. 2020; Nunan and Brassey 2020; Parnell et al. 2020). Though indispensable for ensuring a safe increase of human interactions, up to now, there is only little evidence on the functioning and the limitations of such tools in most economic sectors. In particular, there is a lack of analyses on the role of the prevailing infection level or the number of participants. One example of such events are football matches. Their impact on COVID-19 is of specific interest for policy makers – given that they take place on a regular basis and all over the country. Also, they are known for enthusiastic supporters, packed stadiums and loud fan chants – factors that potentially boost the spread of COVID-19. Opposingly, matches are open-air and are mostly conducted with a thinned-out audience and with additional hygiene measures (e. g., distant seating, no alcohol, partially mandatory masks) which might mitigate transmissions. Hence, it is unclear what the overall effect of football matches on COVID-19 infections is.

This paper contributes to closing this gap for mass gatherings by studying German football matches in summer and autumn 2020.[1] After Germany shut down all mass events in response to the first wave in Spring 2020, counties, starting in early August 2020, individually and independently reallowed spectators to attend matches. Later, on September 15, 2020, the national government passed a federal directive for sports events to allow an occupancy of up to 20 % or a minimum of 1,000 spectators. These reopenings were bound to harsher restrictions in the presence of local case numbers of more than 35 cases per seven days per 100,000 inhabitants in a county (Hartmann 2020). Some federal states with very low case numbers even relaxed these rules. For example, Schleswig-Holstein allowed 25 % occupancy.[2] Over this time period, infections increased from around five (August 01, 2020) to more than 100 cases per 100,000 inhabitants over a seven-day time span (October 31, 2020). In response to these rapidly rising case rates from mid-October onwards, stricter rules were introduced on the national level (‘lockdown light’) from November 2, 2020 onwards. Also, stadiums were closed again confirming the uncertainty whether matches might drive infections.[3]

In this paper, we exploit the fixed match schedule across teams from different counties for identification. To account for the timing of matches, their spatial distribution, and delayed repercussions, we analyze the German matches in an event study setup as already applied in the context of COVID-19 by for instance Dave et al. (2021), Dave et al. (2020b), DeFilippis et al. (2020), Gupta et al. (2021), Isphording et al. (2021), Lange and Monscheuer (2021), and Mangeum and Niekamp (2022). Our baseline estimates indicate that one additional match in a county on average raises daily cases by between 0.34 to 0.71 cases per 100,000 inhabitants after three weeks. This corresponds to an increase of cases by 3–6.5 % after three weeks. Moreover, we find matches to be especially related to many cases if the local case rate was high on the matchday. The risk of infections for incidence levels below 25 cases per 100,000 inhabitants over seven days at the matchday is observed to be humble. We could not carve out an effect of short-term transmission to neighbouring counties or away team counties. We also show that most cases are mainly related to the age groups of 15–34 and 35–59 years with limited spillovers to the vulnerable, elderly population.

As a main mechanism for additional infections, we determine that mobility increases in counties where football matches take place. From these results, it seems to be likely that the applied policy, to reduce spectators in stadiums when a county’s seven-day incidence exceeds 35, was at least necessary or even insufficient.

Our results prove that outdoor mass gatherings with a thinned-out audience drive infections – even under social distancing measures and partially mandatory mask-wearing in the stadium. Policy makers should especially consider the increasing risk in the presence of higher local case rates and the transmission risk across counties if travel to events is unrestricted.

The remainder of this paper is structured as follows: In Section 2, we summarize previous research on the role of mass gatherings for the spread of COVID-19 before explaining our empirical setup in Section 3. We present our results in Section 4 and discuss them in Section 5. Subsequently, we conclude our findings in Section 6.

2 Literature

Due to the high number of contacts, research has identified mass gatherings – though indoor more than outdoor (Weed and Foad 2020) – and highly occupied event locations as one driver of the COVID-19 crisis. While Dave et al. (2020a) do not find regional inclines in COVID-19 cases to be related to a single rally in the US election campaign, Bernheim et al. (2020) provide evidence for eighteen rallies to have raised local case numbers. Similarly, Cotti et al. (2021) find long queues on election day in Wisconsin’s primary election to foster the positive test rate in counties with higher attendance in-person at polling locations.[4]

For Germany, Felbermayr et al. (2021) show that tourism to and from crowded Apres-Ski parties in Ischgl, Austria, has been a main driver of COVID-19. Fetzer (2022) hints at the relevance of restaurants for the spread of COVID-19 in the UK. By using spatial variation in the exposure to a governmental program subsidizing restaurant meals during certain times of the week, he finds this policy to have caused eight to seventeen percent of all cases in the respective time period. This achieves support in Chang et al. (2021). They argue that especially restaurants and shopping are risky locations or activities by using US cell phone mobility data.

Regarding sports events, research especially focussed on the contribution of events to the first wave in early 2020 (Ahammer et al. 2020; Olczak et al. 2020; Wing et al. 2021). These studies mainly investigate cross-sectional, time-invariant differences on the county level, only weakly control for the exact timing of matches, and importantly cannot refer to more recently introduced hygiene and occupancy restrictions. However, they contribute very important benchmark findings: Ahammer et al. (2020) analyze indoor events and find an additional NBA (professional basketball) or NHL (professional ice hockey) match to raise COVID-19 cases in the respective US county and its neighbouring counties by nine percent. They also estimate the effect’s development over time but do not account for the specific timing of matches. Their findings are in line with Wing et al. (2021), who consider an NHL/NBA match to cause almost 800 additional cases. They argue that college matches less strongly drive infections. Their results emphasize the relevance of match and league differences. Lastly, Olczak et al. (2020) analyse the impact of amateur and professional football matches, especially in March 2020 in the UK. They estimate six additional cases per 100,000 inhabitants to be related to every single football match. As they mainly find an increasing number of cases as a consequence of sports events, the hope that less occupied stadiums and adapted hygiene regulations can guide a way back for supporters has to be tested carefully.

In contrast, to our knowledge, there is almost no research on sports events’ influence on COVID-19 in Europe’s later waves and hence in the presence of social distancing measures yet.[5]

Note that there also is research on the spread of influenza at sports events by Cardazzi et al. (2020) and Stoecker et al. (2016). The former find that absolute influenza mortality increases in US cities where new top league sports teams are introduced. The authors conclude that this makes sports events a hotspot for the transmission of viruses. Similar results are presented in Stoecker et al. (2016), who present inclining influenza transmission in the presence of the Super Bowl. Finally, also note that it is likely that people’s behavior has not only changed outside of the stadium, see, for example, Mendolia et al. (2021) for voluntary changes in mobility, but that sports spectators also adapt due to the higher exposure to risk related to the pandemic. In fact, Gitter (2017) for the H1N1 virus in Mexico and Reade and Singleton (2020) for COVID-19 in Europe show that less people voluntarily attend matches to higher case rates. Reade et al. (2021) specifically analyse stadium demand in Belarus where football matches have been continued without any regulatory changes in response to COVID-19. They find that attendance dropped in early 2020 due to high uncertainty but recovered throughout the season.

3 Data and empirical strategy

To quantify potential effects of football matches on COVID-19 case rates in Germany, we construct a dataset mainly consisting of three components.

First of all, we collect data on daily COVID-19 cases on the German “Kreis” (county) level, which is provided by Germany’s main health monitoring institution, the Robert-Koch-Institut (RKI).[6] This data also includes information on the age group of an infected person and whether he or she recovered, died or still is infectious. In our analysis, we primarily focus on data ranging from August 10, 2020 to November 08, 2020. This time frame covers all professional football matches in the season 2020/2021, which were played in front of fans as they were banned from stadiums again from November 02, 2020, onwards due to rising COVID-19 cases. Hence, our sample across 401 counties over 91 days forms a balanced panel of 36,491 observations.

Secondly, we collect all (≈ 1,200) professional football matches which took place during this time in Germany. We include matches from the top four German men’s divisions (‘Bundesliga’, ‘2. Bundesliga’, ‘3. Liga’, ‘Regionalligen’[7]), the women’s top league (‘Bundesliga’), cup matches and friendly or test matches. In addition, there are a few matches from international competitions (‘Champions League’, ‘Europa League’) and the national teams, too. To obtain precise information on the crowd’s size relative to the stadium capacity – which is likely to be a relevant measure for the spread of the virus – we collect information on attendance and capacity, as well.[8] Descriptive statistics on the matches by leagues can be found in Table A1 in the appendix. We exploit substantial variation in the timing of the matches throughout summer and autumn 2020, as well as heterogeneity in the hygiene and spectator restrictions between counties and over time. Figures 1 and 2 present the number of matches played at a specific day and document the variation in observed attendance and occupancy across matches as well as the share of matches played without fans.

Overall, we observe matches in 127 of 401 German counties. Information on the spatial distribution of matches and the respective local case rates are plotted in Figure 3. As evident, matches took place in almost every region of the country. Infection numbers were lowest in the Northern and Eastern parts of Germany. This translated into less need for stricter NPIs and hence a higher share of matches being played in front of spectators in these regions. Overall, in about two thirds of all treated counties, matches were played with and without visitors. Out of the remaining counties, three counties only experienced ghost games and the rest only matches with visitors.

Figure 1 
Played Matches With Visitors Over Time.
Figure 1

Played Matches With Visitors Over Time.

Figure 2 
Visitor and Occupancy Distribution for Matches with Visitors Allowed.
Figure 2

Visitor and Occupancy Distribution for Matches with Visitors Allowed.

Figure 3 
Distribution of Matches, Visitors, Matches with Visitors and COVID-19 Cases Across Counties.
Figure 3

Distribution of Matches, Visitors, Matches with Visitors and COVID-19 Cases Across Counties.

Finally, we complement the dataset with Google Mobility Data[9] to track behavioral changes in the population on the state and day level which could have affected case rates.

Match occurrence highly varies across counties and over time. Also, several counties experience a repeated treatment (e. g., several home matches of one team). Therefore, it is important to assess the effect of a match independently and to disentangle the forces of individual matches. Moreover, the delayed effect of matches – as cases related to a match will be registered several days after the match date – demands for a fine-grained method to evaluate the role for the spread of COVID-19. To tackle these issues, we apply a dynamic difference-in-differences model with staggered treatment – an event study – as it especially considers the timing of matches and their effects’ development over time.

This setup will identify a causal effect of football matches on infection numbers if treated and untreated counties follow the same trend in the absence of a treatment (‘parallel trend assumption’). Flat pre-trends will indicate whether this assumption holds in this setting. Furthermore, flat pre-trends will support for the absence of a reverse causality bias. This is because a potential effect from case rates on the happening of football matches would then result in different trends of treated and untreated counties before the treatment.

We use the following regression equation as baseline:

(1) C O V I D i t = d = 15 , d 0 22 β d M a t c h i t d + X s t ζ + θ i + λ t + η s w + ϵ i t

where C O V I D i t is a measure of COVID-19 transmission in county i at time t. We mainly refer to the daily number of newly registered cases per 100,000 inhabitants here as there has been little variation in the number of deaths in Germany throughout the summer months until late October (s. Figure A4).[10] We do not use the reproduction number, which has been used in evaluation of non-pharmaceutical interventions (Haug et al. 2020), as this measure is past-dependent and very volatile on the county level. M a t c h i t is an indicator for county i’s exposure to football matches – which can be the number of matches per 100,000 inhabitants or the number of visitors/spectators per 100,000 inhabitants – on date t. Hence, we normalize the treatment with respect to the population size of counties. Therefore, we do not exploit a dichotomous event dummy here.[11] d gives the number of lags and leads of the treatment in days in the effect window. Throughout the paper, we mostly will bin leads and lags to groups of three days to eliminate noise from the regression. Note that both treatment indicators named contain different factors. While the match variable does not account for the exact number of fans in the stadium, it also considers effects outside of the stadium, such as private gatherings or infections on the way to the match (e. g., public transportation). On the other hand, the number of visitors, gives a better impression of the relevance of the actual number of people in the stadiums.

We observe the two weeks before and the three weeks after a county is treated as effect window. We will adopt a different effect window and different binning later on, too. Our results are robust to these modifications.

Matches typically are exogenous treatments as the weekend on which they take place are set in advance before the start of the season (s., e. g., Lichter et al. (2017)). Beyond this, we control for changes in the people’s mobility which we capture by X s t which includes time-variant Google mobility data.[12] Further, we consider county and date fixed effects θ i and λ t , which control for factors implying incidence heterogeneity in the cross-section or over time. This could, for example, include differences in the population density or age distribution of counties and trends in case numbers across entire Germany. Finally, as we observe treatments on the county level, we can include fixed effects for state s and week w interactions to control for underlying regional trends in our data – similar to Gupta et al. (2021), who include state fixed effects on the daily level. These fixed effects should especially consider the rapid development in COVID-19 cases in October, which started in federal states such as Bavaria or Berlin and took place with delay in some, mostly Eastern and Northern federal states. This would neither be suitably absorbed by county or date fixed effects. All regressions are estimated with standard errors clustered on the county level.

Note that professional football teams mostly come from urban regions. As this could have resulted in significant disparities between treated and untreated counties, we provide probit regression estimates on the socio-demographic and economic characteristics of the counties in Table A2.[13] Indeed, we find that a county’s number of inhabitants is a main determinant of hosting a football match from our sample which fits the hypothesis of matches taking place in especially large cities. We consider that an additional match or spectator will have a smaller impact on the local case rates per 100,000 inhabitants if a county’s population is high. Hence, we do not use the absolute number of matches or visitors as treatment variable but use the relative (normalized) impact of matches or visitors per 100,000 inhabitants. Also, we identify treated counties to have a lower share of inhabitants above 65 years which, if at all, could bias the number of registered cases and deaths downwards (Felbermayr et al. 2021). This makes it more difficult to find clear effects and hence can be interpreted as an additional hurdle that has to be overcome to show effects. Moreover, county fixed effects should account for substantial parts of this difference. Finally, note that the available income seems to be lower in treated counties. But this factor has not been found to be robustly associated with heterogeneity in German case numbers (Felbermayr et al. 2021; Krenz and Strulik 2021). Population density does not differ between both groups of states, which e. g. correlates with accessibility via road which has been a relevant driver of COVID-19 cases in Germany (Krenz and Strulik 2021). Other crucial measures for the fight against COVID-19, such as the hospital (bed) density, do not differ between treated and untreated counties.

Lastly, note that it is crucial to determine which other factors may have driven case rates. Besides controlling for regional trends, we generally analyse a time period in which measures against COVID-19 have not been changed much. Figure A3 presents the development of NPI coverage in Germany over time.[14] Most measures have persistently been in place or were not enforced before the end of our sample period. Later on, we also control for NPIs on the county-level as additional covariates for a robustness check.

4 Results

In the following, we present general findings for the effect of football matches on the transmission of COVID-19 in Germany across counties and over time. To better understand the forces which drive potential effects, we also conduct narrower analyses on specific age groups, spatial transmission, or the role of the prevailing incidence of infections for instance.

General findings

First of all, we analyse the effect of all football matches in the sample, which took place in front of spectators. For that, we normalize the number of matches per county with the inverse population size.[15] We then regress the number of daily cases per 100,000 inhabitants on the number of matches per 100,000 inhabitants and its leads and lags. Plotting the coefficients of the event study framework as explained in Section 3 gives us Figure 4.[16]

Figure 4 
Effect of Matches per 100,000 Inhabitants.
Figure 4

Effect of Matches per 100,000 Inhabitants.

Figure 5 
Effect of Visitors per 100,000 Inhabitants.
Figure 5

Effect of Visitors per 100,000 Inhabitants.

The same regression with the number of visitors per 100,000 inhabitants as treatment indicator is shown in Figure 5. Unsurprisingly, we do not find an effect in the first week after a match in both plots. This is because that potential infections will be symptomatic with delay and hence be registered several days after the match.[17] This is robust for both treatment indicators used. We find an increasing trend in the number of cases in both plots over the three weeks post-treatment with slightly significant coefficients from the match and visitor treatment indicator.[18] If there was one more match per 100,000 inhabitants, this would result in an increase of 1.59 daily cases per 100,000 inhabitants – or 0.18 standard deviations respectively – after three weeks. Similarly, an increase in the number of visitors per 100,000 inhabitants by one raises the number of daily cases per 100,000 inhabitants by 0.0008 after three weeks. Considering a county of about 100,000 inhabitants, this implies that 1000 additional spectators at a match increase the daily number of cases after three weeks by 0.8 or 0.09 standard deviations respectively.

Note that our approach, to bin daily effects, disables us to identify the immediate beginning of the rise in cases. Hence, we also provide both event studies without binning. These are available in Figure 6. The delayed rise of infections after about 1.5 weeks indicates that transmissions may be passed on after the matches.

Figure 6 
Effect of Football Matches and Visitors per 100,000 Inhabitants – No Bins.
Figure 6

Effect of Football Matches and Visitors per 100,000 Inhabitants – No Bins.

Age distribution

A main objective during the pandemic has been the protection of vulnerable groups such as elderly people – especially before vaccinations. As shown in Figure A5, up to 40 % of all COVID-19 cases in spring 2020 came from the age group of 60+, which largely contributed to the number of COVID-19 deaths throughout the first wave. Starting from August onwards, the share of COVID-19 cases among this age group increased again. If and to what extent football matches could have contributed to the growth in cases among different ages is shown in Figure A6. We intuitively find most cases to belong to the age groups of 15–34 and 35–59 year-olds. Children and elderly people seem to be less affected by football matches in the short-run. Similar patterns are found for the alternative treatment indicator visitors per 100,000 inhabitants of the respective age group (s. Figure A7). When considering that the age group 35–59 years has the highest absolute share of the German population (approx. 35 %), we find that most football-related infections are associated with this age group.

Incidence level

Especially relevant for policy makers is the question whether stricter regulations for spectators also work out in the presence of rising infections or whether their functioning is not secured beyond a certain threshold. Also, recent research has shown that contact tracing – which becomes more challenging with increasing infection numbers – is crucial (Fetzer and Graeber 2021). To understand the (in)sensitivity of such measures in football stadiums, we estimate the effect of matches at different incidence levels at the matchday. In particular, we calculate the 7-day case incidence per 100,000 inhabitants on the county level at each matchday and divide matches up into groups of incidences below 15, 15–25, 25–35, and above 35. Results of the regressions referring to this dispersion are provided in Figures 7 for visitors. Results for the match treatment indicator are not reported here but in the appendix (s. Figure A8). Results are qualitatively identical. For matches at an incidence level above 25 or even 35, we find strong effects of up to 0.01 daily cases per 100,000 inhabitants for an increase in visitors per 100,000 inhabitants by one. This impression is similar for matches per 100,000 inhabitants as treatment indicator. Here, we see strong increases of up to ten additional cases per additional match per 100,000 inhabitants for incidence levels above 35.

While it is not surprising that higher incidence levels provoke more infections, it is striking that we partly observe substantial increases of infection levels already above 25. Hence, these findings are to some extent in line with the administrative decision to limit higher occupancy levels to at least an incidence level like 35.[19]

Figure 7 
Effect of Visitors per 100,000 Inhabitants Under Different Incidence Levels.
Figure 7

Effect of Visitors per 100,000 Inhabitants Under Different Incidence Levels.

Away teams

An important question with regard to the spread of COVID-19 is the spatial transmission across county borders via for example travel (Adda 2016; Chinazzi et al. 2020; Coven et al. 2020). In early stages of the pandemic, this issue was addressed by closing borders or restricting holiday travel within Germany. With regards to football, away fans were excluded from German football matches in professional leagues.[20] Therefore, we will analyze the effect of matches on the number of newly registered cases in the county where the away team comes from. This, on the one hand, contributes to the discussion on the spread of COVID-19 throughout the country in relation to a match – even though away fans are not allowed in professionals’ stadiums and hence should avoid travelling. More importantly, it also provides insights to infections outside the stadiums – assuming that away fans cannot attend the matches in the stadiums, and hence could watch those (possibly some in groups) at home or in pubs which were open at this time. In Figure A10, we present the effect of an additional match per 100,000 inhabitants of the away county on the spread of COVID-19 there. We neither identify a significant effect of football matches on daily case numbers three weeks after the match nor a systematic pattern.[21] Hence, we suggest that this implies infections outside of the stadiums to be limited and below the numbers observed in home counties. To ensure that the smaller to non-existing effect in away counties is not driven by an on average too low interest for the match in away counties or for example no possibility to follow matches in the away county as no TV stream is available, we also run the regressions for only matches with a subsample of Bundesliga, 2. Bundesliga and 3. Liga (top league) matches which are all streamed. Here, we do not identify an effect either, which supports the overall findings.

Note that we can also interpret these results as a (quasi-)placebo test for the general effect from above. Admittedly, there may still be places where infections take place related to football matches (e. g., bars), but it serves as a simple confirmation of null effects outside the match counties.

Finally, we also provide additional evidence on the non-existence of COVID-19 transmission across county borders as an extension in the appendix. We discuss potential spreading to neighbouring counties and throughout commuting networks there (s. Section A.1 in the appendix).

Occupancy

To further investigate the relevance of spectator density on the number of cases, we also estimate the effect of different occupancy levels. Olczak et al. (2020) show that already low levels of stadium occupancy (e. g., below 20 %) raise COVID-19 deaths. Still, this finding covers matches where people were free to densely gather together. Hence, new distancing measures might be effective here. Also, note that, for example, Chang et al. (2021) suggest a non-linear relation between occupancy and infections as already a small reduction in case numbers can have a high impact on the reduction of cases in several public locations such as restaurants. In our analysis – to avoid an endogeneity problem in our regressions due to simultaneity between the daily case numbers and the occupancy – we only include matches in the regression sample which were played at an incidence level of up to 35. Above this mark, increasing cases cause a reduction in allowed visitors and hence occupancy. We present the results on different occupancy levels (0–5 %, 5–10 %, 10–15 %, 15 %+) in Figures A11 and A12. We find that there is no clear relation between a higher occupancy level and caused infections. It seems that the allowed occupancy levels are below a critical threshold so that social distancing measures remain similarly (in)effective – given the incidence level cap. At all occupancy levels, we see slight increases in cases to the end of the effect window. Moreover, by the fact that occupancy on this level does not drastically drive infections, we support suggestions by Olczak et al. (2020) that transmission typically does not take place at the seat but probably on crowded ways or queues in the stadium.

Ghost matches

Facing an increase in case numbers over time in the sample, almost one half of the matches in the data were played behind closed doors. We make use of these matches to cross-validate our results on infections inside and outside the stadiums. Figure A13 presents the event study coefficients of ghost matches. We do not find an increasing pattern in a county’s COVID-19 cases in response to a ghost match. This impression results in two conclusions: Firstly, we find evidence for ghost matches to fullfil their purpose to not additionally drive infections. Secondly, ghost matches would only allow for infections outside of the stadiums. As all fans are banned, private gatherings could even increase for the subsample of ghost games. Still, we cannot find any evidence for such infections. This underlines our claim from above that detected infections mainly originate from spectators on their trip to the arena or the ways in the stadium.

Top leagues

In addition, we shed light on the question of whether top league matches better mitigated potential effects on case numbers. Especially after politics announced the exclusion of football fans starting from November 2, 2020, many clubs complained that there is no evidence that top league football matches accelerated COVID-19 transmission.[22] To test this hypothesis, we run regressions of equation (1) for a subsample of top league matches – including Bundesliga, 2. Bundesliga, and 3. Liga matches. We present these results in Figure A14. Three weeks after a match, we identify an increase in the case numbers with a rising number of visitors or matches relative to the county’s population. Hence, we argue that even top league matches with the potentially best hygiene regulations cannot avoid slight increases in the case numbers. After three weeks, the found effect of around two additional daily cases per 100,000 inhabitants for an additional match per 100,000 inhabitants is similar to the estimates in the overall sample in Figure 4.

Deaths

Besides registered cases, the number of related deaths is an important indicator to determine the downsides of reopening football stadiums. As we saw that the age group 60+ did not experience a clear increase in cases, the question arises whether there is a fatality effect at all. We present event study findings on the effect of all matches and visitors in our sample on the respective counties’ number of deaths in Figure A15. Note as the RKI does not document the date of death but the registration date for a deceased case, we would expect to already see effects as soon as for cases. Interestingly, and in line with the age distribution of football-related cases, we cannot find a significant increase for both treatment indicators – matches and visitors. This could also hint at a self-selection effect that especially vulnerable fans do not attend matches anymore. Still, this explanation does not hold for infections after the match (for example a person who got infected in the stadium and infects another one a few days later).

Lockdown effects

For our analyses above, we relied on data until November 08, 2020, which is just until one week after the beginning of the German ‘lockdown light’, to ensure that our results are not biased by potentially less transmission opportunities. That the lockdown led to a reduction of the growth in COVID-19 infections can be easily seen in Figure A4. To assess whether the lockdown also reduced the infections originating from the latest football matches, we extend our observation period of COVID-19 cases to December 06, 2020 and reestimate our main regressions. As can be seen in Figure A16, the overall effect of an additional match per 100,000 inhabitants is corrected downwards and now turns out to be insignificant while still showing a positive but less steep upward trend. This indicates that the transmission from more recent matches could have been decelerated partly through the additional measures – even in the presence of higher infection levels. Note that the effect from an additional visitor per 100,000 inhabitants is still significant after three weeks. Nevertheless, the insignificance of the match effect after including lockdown COVID-19 cases indicates that the effect’s robustness may be weaker for effects of mass gatherings during the first COVID-19 wave with less hygiene regulations (Ahammer et al. 2020; Olczak et al. 2020; Wing et al. 2021).

5 Mechanism

While we cannot clearly determine where – i. e., in/out of the stadium – infections related to football matches happen, there should be a way to identify additional, football-induced mobility patterns which result in the observed increase in cases. Hence, we subsequently provide evidence for increased mobility in counties and on days which were exposed to football matches. For that, we make use of mobile phone mobility data provided by the RKI (for a detailed discussion of the data, s. Schlosser et al. (2020)[23]) which gives daily information on people’s mobility in a specific county.[24] In particular, the data represents the relative change in mobility in a county on a 2020 day compared to the average of all days in the same month of the preceding year which have been on the same weekday. To be precise, i. e., the mobility on Saturday, August 01, 2020, is, for example, compared to the average mobility on all Saturdays in August 2019.

If football matches caused an influx in the mobility, we should therefore find an increase in the mobility – or mobility change compared to 2019 – on a certain date in a county where a football match took place. We investigate this by estimating a simple differences-in-differences model following the equation:

(2) M o b i l i t y i t = γ 0 + γ 1 M a t c h i t + X i t + ζ + ϵ i t

where M o b i l i t y i t gives county i’s mobility change on date t in comparison to the previous year in percentage points. γ 1 is the coefficient of interest as it gives us the marginal effect of the match or visitor indicator on the mobility in county i on date t. X i t encompasses county-level covariates on the daily level such as the prevailing incidence level. Finally, ζ is a set of county and date fixed effects which are intended to capture heterogeneity in, for example, mobility shocks due to COVID-19 policies or weather fluctuations. Further, we again add interactions of the state and week to control for regional, temporary patterns and also introduce county linear time trends.

Note that as the outcome variable compares mobility during the pandemic to pre-pandemic mobility, the variation which we use for identification is the within-county variation in the mobility change during the pandemic. The county fixed effects take account of the general county-specific difference in mobility between before and during the pandemic. As the regression identifies mobility changes during the pandemic relative to before the pandemic, we are able to show mobility changes due to football matches. Mobility changes relative to pre-pandemic mobility directly indicate a change in absolute mobility, so that our estimates are meaningful for understanding how many more contacts people had on a certain day in a certain county. Figure 8 further shows that the mobility level in our sample period (up to early November 2020) was comparable to the preceding year. Only after the introduction of new NPIs in early November mobility dropped again.

All regressions are presented in Table 1. We use data for all counties between the dates on which the first and last match with visitors took place – August 01, 2020 and November 01, 2020.

Figure 8 
Mobility Change over the Sample Period.
Figure 8

Mobility Change over the Sample Period.

Table 1

Effects of Football Matches on Mobility Changes.

M o b i l i t y i t

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Matches/100,000 Inh. 1.142** 1.207** 1.221** 1.161** 1.062**
(0.568) (0.486) (0.502) (0.485) (0.458)
Matches/100,000 Inh. × 1[Small Crowd] 0.856*
(0.486)
Matches/100,000 Inh. × 1[Large Crowd] 2.912***
(0.890)
Matches/100,000 Inh. × 1[Top Leagues] 2.778**
(1.211)
Matches/100,000 Inh. × 1[Not Top League] 0.981*
(0.511)
Visitors/100 Inh. 1.101*** 1.178*** 1.172*** 1.102***
(0.374) (0.256) (0.279) (0.249)
Matches/100,000 Inh. (Away County) 0.383 0.700* 0.591 0.759* 0.763* 0.751* 0.353 0.378 0.693* 0.584 0.752*
(0.477) (0.421) (0.409) (0.413) (0.414) (0.414) (0.439) (0.473) (0.419) (0.407) (0.412)
7-Day Incidence (1-Day-Lag) −0.051*** −0.051*** −0.051*** −0.059*** −0.051***
(0.014) (0.014) (0.014) (0.013) (0.014)
7-Day Incidence ∈ [35, 50) −0.011 −0.012 −0.012 0.002 −0.012
(0.328) (0.328) (0.328) (0.476) (0.328)
7-Day Incidence ∈ [50, ) −0.233 −0.230 −0.231 −1.301* −0.228
(0.656) (0.656) (0.656) (0.781) (0.656)
County FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Date FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
State × Week FE No Yes No Yes Yes Yes Yes No Yes No Yes
Lin. County Trend No No Yes No No No No No No Yes No
Sample Full Full Full Full Full Full Weekend Full Full Full Full
Observations 37,694 37,694 37,694 37,694 37,694 37,694 11,228 37,694 37,694 37,694 37,694
Adjusted R2 0.821 0.868 0.864 0.870 0.870 0.870 0.853 0.821 0.868 0.864 0.870
  1. Note: *p < 0.1; **p < 0.05; ***p < 0.01. Heteroskedasticity-robust standard errors clustered on the county level.

In columns (1) to (4) of Table 1, we find that an additional match per 100,000 inhabitants increases mobility by about 1.2 percentage points which is equivalent to an incline of 0.065 standard deviations of the mobility change. This demonstrates a potential mechanism of infection effects from football matches. Increased mobility could cause additional COVID-19 cases. Similarly, more visitors of a football match raise mobility. In particular one visitor per 100 inhabitants increases the mobility in 2020 in relation to 2019 by about 1.1 percentage point or 0.06 standard deviations (s. columns (8)–(11)). The similar size of the effects fits the fact that an average match has about 1,000 spectators.

Heterogeneity analyses in columns (5) and (6) further prove that mobility increases more strongly for large crowds (above the mean of spectators for non-ghost matches). While mobility increases by almost three percentage points for such matches relative to days without matches, the effect is less than one percentage point for smaller crowds. Furthermore, mobility is higher for top league matches because average attendance and overall attention is higher. In column (7), we examine mobility changes in a subsample of weekend days (only Saturdays and Sundays) and find a very similar effect size. In doing so, we underline that our results are not driven by the fact that around 70 % of all matches took place on Saturdays or Sundays.

We also checked whether we find significant effects on the away county. Results show that there is an unrobust and smaller effect of football matches in the away counties. Only when inserting state-week interactions, we barely observe a significant effect. We interpret these results as support for our findings from above that there are none to much smaller effects of football matches in away counties which could be caused by the channel of a smaller reaction in the mobility.

Figure 9 
Mobility Change on Match Days.
Figure 9

Mobility Change on Match Days.

Note that these findings are robust to a variation in the used fixed effects and the inclusion of variables indicating the current incidence level. The latter should account for behavioral changes in response to the dynamic of the virus and potential changes in the measures against the virus.[25]

To provide some additional evidence for the causality behind our estimates, we also provide a binned event study in Figure 9. There we see a clear peak at the day of the match and find support for the estimated effect of approximately one percentage point more mobility related to a match per 100,000 inhabitants (or 0.001 p. p. to one additional visitor per 100,000 inhabitants). Daily event studies without bins are available in Figure A17 but naturally are noisier.

6 Discussion

Effect size

Our results show that there indeed is a relation between mass gatherings and COVID-19 transmission. This also holds in the presence of additional hygiene measures such as a limited number of fans allowed with social distancing measures at the seat. Our general findings show an increase of about 1.59 daily cases per 100,000 inhabitants for an additional match per 100,000 inhabitants and 0.0008 daily cases per 100,000 inhabitants for an additional visitor per 100,000 inhabitants. When performing a simple back-of-the-envelope calculation, we end with the following estimated effects and magnitudes: As the average match with audience in our sample implied 0.444 matches per 100,000 inhabitants and welcomed 1,045 spectators which translates to 407 spectators per 100,000 inhabitants, this equals an average increase of about 0.71 daily cases per 100,000 inhabitants for the match treatment indicator and 0.34 daily cases per 100,000 inhabitants for the visitor treatment indicator after three weeks.[26] In general, we interpret the finding from the match treatment indicator as an upper boundary as this treatment is more likely to pick up effects from out-of-stadium events. These could be even unrelated to professional football (e. g., amateur football matches or other sports events at the same days). The visitor treatment indicator should reveal a smaller upward bias. This variable mostly just considers variation in the attendance in the stadium and not activity outside of the stadium. Hence, we argue that on average a match seems to increase the seven-day incidence per 100,000 inhabitants by between 2.4 and 5.0 – keeping in mind that these numbers strongly vary with outer circumstances.

Further note that our findings suggest continuing keeping away fans out of the stadiums as this contributes to the spatial transmission of COVID-19. Moreover, our results suggest increasing sensitivity to the inclusion of visitors at rising incidence levels. Policy makers should consider the non-linear relation between incidence levels and cases.

Relation to first wave findings

Also recapitulate that the effects found here are hardly comparable to findings from the earlier phase of the pandemic as in Ahammer et al. (2020) or Olczak et al. (2020). Besides the general rules in the stadium, people’s behavior has changed endogenously. Our results indicate that a match with limited attendance does not have such devastating consequences in the short-run – especially in the presence of low incidence levels – as in Ahammer et al. (2020) where one sports event is related to up to 30 registered infections and 1.5 deaths per 100,000 inhabitants or in Wing et al. (2021), who find an NHL/NBA match to cause 783 COVID-19 infections. Interestingly, our short-run effects on registered infections do not seem to be much lower than the results for football matches in Olczak et al. (2020), as we find a match per 100,000 inhabitants to increase cases by about 0.34 to 0.71 per 100,000 inhabitants per day after three weeks, where the respective study only finds a match to increase cases by six per 100,000 inhabitants overall. This can be caused by the generally lower number of registered cases in their sample (for example, due to less testing at the beginning of the pandemic) or a more granular analysis of matches in lower leagues. Opposingly, we do not find strong effects on deaths while they identify two additional deaths per 100,000 inhabitants per match. When comparing our estimates to recent results by for example Fetzer (2022), who finds an UK policy subsidizing restaurants to account for eight to seventeen percent of all infections, or Lange and Monscheuer (2021), who find anti-COVID demonstrations to increase infection rates by more than 35 percent in treated counties, the effect of football matches seems to be rather small.

Limitations

Nevertheless, while being able to account for time-variant and regional trends, the event study approach suffers minor issues which should be kept in mind: First, event studies account for matches as additive treatments whose effects are insensitive of each other and are cumulatively summed up to get the overall effect. As a previous match which increased the infection level could act as a multiplier for the effect of a subsequent match, this could slightly underestimate the effect of earlier matches while overestimating the effect of later matches. Secondly, the model assumes a uniform trend in the infections after a match across treated counties. Nevertheless, this could be violated by the heterogeneous registration velocity of new cases as test results take longer in rural areas. Finally, we cannot completely explain the micro-mechanism behind infections. Whereas top league matches could cause infections outside of the stadium as well (e. g., people following the match together on TV), this is unlikely for lower league matches such as in the Regionalliga. On the other hand, less strict hygiene rules could cause proportionally more cases in the stadium.[27] Moreover, note that our results are based on a limited sample size of 1203 matches of which 660 took place with spectators. Especially the relatively small sample size for top league matches, due to the early decision to exclude spectators again, should be considered when generalizing this paper’s results.

Robustness checks

We performed several robustness checks to test our results’ sensity to configurations of the empirical strategy. For example, we also changed the regional time fixed effects from the weekly to daily level on the state level and also implemented time trends/fixed effects on the more granular NUTS2 (38 administrative regions) level (s. Figure A18).[28] We also tested for different suitable clusters of standard errors (s. Figure A19). We also modified the post-treatment effect to exemplarily four weeks (s. Figure A20). These adaptions did not change our results qualitatively – neither in the overall sample nor in the reduced top league sample.

In another robustness check, we made use of detailed data on the NPIs, which were implemented during this time period, which is provided for each county and date by the German Ministry for Economic Affairs and Energy. Controlling for the effects of such measures could have explanatory power beyond the regional time fixed effects as indicator for trends. In detail they capture 23 different measures of NPIs and track whether these measures have been in place or not. We match the one-week lag of this data to our sample and rerun our main event study regressions. Our results do not change (s. Figure A21).

We also checked whether additionally controlling for fluctuation in the population’s risk perception over time changes our findings. For that, we include state-level data on four measures of risk perception from the (mostly) bi-weekly data of the COSMO study (Betsch et al. 2020). In fact, we included variables which capture the perceived risk of infection, the perceived severity of an infection, the perceived ease of avoiding an infection, and the perceived affective risk perception across time and counties. Again, this does not change our results.

Moreover, we test the robustness of our results with regard to counties’ heterogeneous exposure to international holiday travellers.[29] Especially during the summer school holidays in July and August a substantial share of the national COVID-19 cases were related to exposition to the virus outside of Germany. Throughout our sample period, the RKI (2020b) documents that approximately 9 % of all cases were likely related to exposition outside of the country. This share was even higher during the summer travel season – though overall case rates were still low at this time. To account for counties’ travel exposure, we performed multiple robustness checks: Firstly, we dropped all counties with an international passenger airport. Secondly, we made use of monthly passenger numbers for each of these airports provided by the ‘Flughafenverband’ ADV (2020) which we implemented as additional control interacted with week fixed effects. Thirdly, we calculated the distance of each county’s geographical centroid to the geocoded location of the nearest international passenger airport and interacted the variable with week fixed effects to capture the remoteness of counties as covariate. Fourthly, we made use of the staggered rollout of school holidays across federal states to capture the exposure to travellers over time on the state-level by constructing a dummy if there have been school holidays in the county’s federal state seven days ago.[30] We again use interactions with week fixed effects to account for a heterogeneous relation over time. Fifthly, we use station data of all train stations in Germany by the Deutsche Bahn AG[31] and create a dummy turning one if a county has a train station connected to the intercity-express network. This allows for longer travels into holidays. We interact this dummy with week fixed effects again. Sixthly, we also control for the proximity of counties to national borders by adding the logged distance to the nearest border interacted with week fixed effects to the baseline regression. Seventhly, we make use of weekly number of cases which were likely due to exposition outside of Germany on the national level.[32] We interact these numbers with county fixed effect to proxy the regional exposure to cases from outside of Germany. Lastly, we add county-level trends to account for the spread of infections beyond state trends which might capture county-specific exposure to travellers as well. All of these robustness checks to travel exposure did not change our results qualitatively as can be seen in Figure A22.

Lastly, we tested the robustness of our results in a simplified difference-in-differences setting. We construct an absorbing dummy which turns one whenever a county experiences the first football match with visitors. This setting allows an easier interpretation of the results but does not account for the heterogeneous treatment across counties (for example different number of matches). We present the results in Table A3 and Figure A23. We again see an increase in cases in the treated counties of around 0.645 daily cases per 100,000 inhabitants. Again, the effect evolves over time (increases in later periods) and does not affect death rates.

7 Conclusion

Sports is at the heart of most societies. In the presence of a pandemic, a trade-off between this welfare contribution and potential impacts on society’s health might arise. To align both of these interests, policy makers introduced stricter hygiene regulations to reallow fans back into the stadiums. At the example of German professional football, we show that these hygiene concepts cannot fully avoid infections related to matches. The number of infections caused is mediated by the actual infection level at the day of the match in the respective county. In particular, only few infections emerge at a seven-day incidence below 25. As an underlying mechanism, we identify football matches to increase mobility in the respective counties. Our results cannot account for the exact infection dynamics (e. g., insufficient distance between fans, public transport to match) and are based on quite roughly defined treatment indicators. Still, we are confident to contribute relevant findings to optimizing living with the virus. Policy makers should take note of these results in embellishing hygiene regulations more efficiently, as we highlight the limitation of reopenings in the presence of too high infection levels. Nevertheless, as highlighted in a recent note by Singleton et al. (2021), “[p]olicy makers need more evidence on if, when and how it is safe to open sports stadiums as Covid-19 rages”. We are looking forward to more contributions on this issue.

Award Identifier / Grant number: 235577387/GRK1974

Funding statement: This project is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 235577387/GRK1974.

Acknowledgment

I am thankful for very helpful comments by Andreas Lichter, Benedikt Schmal and Simon Schulten as well as for insightful discussions at the Reading Online Sports Economics Seminar 2021 and the Public Health Conference at the International Centre for Economic Analysis (ICEA).

Appendix A

A.1 Extension: Transmission across county borders

We subsequently especially discuss the spatial transmission of potential infections from football matches across county borders. For that, we firstly examine neighbouring counties of treated regions (comparable to Ahammer et al. (2020)). Moreover, we then make use of commuter data provided by the German Employment Agency (Agentur für Arbeit).[33] The data includes information on the interrelation of counties with respect to commuting for work. We use this as proxy of general commuting behavior between counties and network interconnections. The advantage of this data is that the general reduction in mobility during the pandemic cannot totally be reduced in the world of work. Thus, this data should make up a relevant part of the remaining mobility and can also account for spreading beyond close regional borders to for example far distant federal states. Mense and Michelsen (2020) use the same data to investigate interregional interdependences as an infection mitigator in the first wave 2020. They find commuter networks to significantly explain changes in infection rates. While the effect decreased during the first lockdown, it nevertheless remained robustly existent.

To make use of the data, we follow a similar weighting strategy as Mense and Michelsen (2020) and create a treatment variable in the form of

T r a n s m i s s i o n i t = j = 1 , j i 401 [ M a t c h j t × I n c o m m u t e r i j + O u t c o m m u t e r i j P o p u l a t i o n i ]

which equals the sum of the match or visitor indicator (as used above) weighted with the exposure to commuting over all counties. In detail, we calculate the share of people who travel between county i and any other county j of the 401 counties relative to county i’s population which is I n c o m m u t e r i j + O u t c o m m u t e r i j P o p u l a t i o n i . This commuter density is the weight for the exposure to the treatment in county j. Summing this new treatment indicators up over all states gives the respective transmission indicator. As we include the exposure to the treatment in the treated counties j and the commuting behavior, we ensure that higher exposure to risk gives a higher value for T r a n s m i s s i o n i t . The variable can be interpreted as a likelihood or probability of infection spillover in case that football matches contribute to COVID-19 spreading. In comparison to the approach to use neighbouring counties, this procedure will account for example for the bias that the counties’ population is not suitably given by the number of border counties or geographical distance only.

To measure the potential effect of matches on counties which directly border treated counties, potential spillovers will be analyzed here. We plot the estimated coefficients for the effect of matches in a neighbouring county in Figure A1. We do not find any significant effects of a match in the neighbouring county. This, on the one hand, underlines the reliability of the detected patterns – as we hence do not only cover fractions of the development of cases over time – but it also indicates the limited spatial spreading due to less people in the stadiums.[34]

Figure A1 
Effect of Matches and Visitors per 100,000 Inhabitants on Neighbouring Counties.
Figure A1

Effect of Matches and Visitors per 100,000 Inhabitants on Neighbouring Counties.

Figure A2 
Effect of Matches and Visitors per 100,000 Inhabitants on Commuter-Exposed Counties.
Figure A2

Effect of Matches and Visitors per 100,000 Inhabitants on Commuter-Exposed Counties.

Here we now also provide results on our findings with respect to commuting. Applying the created transmission indicator as a treatment indicator in our event study setup should hint at the role of football matches for spatial transmission and can be seen as a robustness check to our findings on the limited transmission to neighbouring counties which are most likely to be locations to which commuting takes place. As can be seen in Figure A2, we cannot find a systematic effect of the transmission indicator on the infection numbers in states with a higher exposure to travel from and to counties with matches with visitors.

A.2 Figures and tables

Table A1

Descriptive Statistics on Included (Semi-)Professional Football Matches.

Competition # Matches ...of which no Ghost Games Mean # Visitors Mean Occupancy
Bundesliga 108 33 4809.3 11.97 %
2. Bundesliga 108 41 2515.3 10.41 %
3. Liga 149 44 2777.4 13.47 %
Regionalliga West 194 100 437.8 5.62 %
Regionalliga Suedwest 137 109 481.4 5.55 %
Regionalliga Nordost 121 116 751.4 9.73 %
Regionalliga Nord 98 71 431.9 10.23 %
Regionalliga Bayern 27 21 283.0 7.31 %
DFB-Pokal 46 21 1640.1 8.90 %
Others 110 52 695.0 6.31 %
Women’s Bundesliga 65 28 377.6 5.50 %
Women’s DFB-Pokal 40 24 189.6 6.87 %
1203 660 1045.1 8.26 %
  1. Note: “Others” include matches from the following competitions: Champions League, Europa League, State Cup finals (“Landespokal”), national team matches and friendlies. “Average # Visitors” and “Average Occupancy” give the respective values for the subsample of matches which were played in front of spectators. Match dates range from 2020-08-01 and 2020-12-23.

Table A2

Socio-Demographic and Economic Differences Between Treated and Untreated Counties.

N = 401 counties

127 treated, 274 untreated
(1 = Treated, 0 = Untreated)

(1) (2)
ln(Inhabitants) 0.322*** 0.356***
(0.052) (0.068)
Share Age ≥ 65 −4.096** −4.873**
(1.667) (2.100)
Population Density 0.015 0.016
(0.010) (0.010)
Share Foreigners −1.454* −0.760
(0.865) (1.232)
ln(Available Income) −0.594** −0.634*
(0.296) (0.373)
Share Protestants −0.013 0.341
(0.216) (0.572)
Share Catholics 0.087 0.406
(0.195) (0.508)
Share Households with Children −0.017 −0.022
(0.023) (0.026)
Average Household Size 0.071 0.173
(0.709) (0.779)
Hospital Density −1.256 −2.098
(2.507) (2.681)
Hospital Bed Density 0.019 0.020
(0.012) (0.013)
State FE No Yes
Observations 401 401
McFadden (Pseudo-)R2 0.294 0.315
  1. Note: The table gives marginal effects at the variables’ means of socio-demographic and economic differences between the treated and untreated counties estimated in probit regressions with heteroskedasticity-robust standard errors. Treated counties are all counties in which at least one match took place which is included in the sample. There are only three out of the 127 counties which did not have a match where visitors were allowed to attend but only ghost games.

Table A3

Absorbing Dummy Difference-in-Differences.

(Daily Cases per 100,000 Inh.)it (Daily Deaths per 100,000 Inh.)it


(1) (2) (3) (4)
1[Post First Match with Spectators] 0.645* 0.035 −0.003 0.002
(0.358) (0.231) (0.008) (0.005)
1[Post First Match with Spectators] × 1[≥ 10 Days Post First Match] 0.843*** −0.004
(0.315) (0.008)
County FE Yes Yes Yes Yes
Date FE Yes Yes Yes Yes
State × Week FE Yes Yes Yes Yes
Observations 36,491 36,491 36,491 36,491
Adjusted R2 0.658 0.658 0.135 0.135
  1. Note: *p < 0.1; **p < 0.05; ***p < 0.01. Heteroskedasticity-robust standard errors clustered on the county level.

Figure A3 
Development of NPI Density Across Counties.
Figure A3

Development of NPI Density Across Counties.

Figure A4 
Registered COVID-19 Cases and Deaths as of Registration Day in Germany.
Figure A4

Registered COVID-19 Cases and Deaths as of Registration Day in Germany.

Figure A5 
Distribution of Cases Across Age Groups and Over Time.
Figure A5

Distribution of Cases Across Age Groups and Over Time.

Figure A6 
Effect of Football Matches per 100,000 Inhabitants of Respective Age Groups.
Figure A6

Effect of Football Matches per 100,000 Inhabitants of Respective Age Groups.

Figure A7 
Effect of Visitors per 100,000 Inhabitants of Respective Age Groups.
Figure A7

Effect of Visitors per 100,000 Inhabitants of Respective Age Groups.

Figure A8 
Effect of Matches per 100,000 Inhabitants Under Different Incidence Levels.
Figure A8

Effect of Matches per 100,000 Inhabitants Under Different Incidence Levels.

Figure A9 
Attendance Depending on Incidence Level on Match Day.
Figure A9

Attendance Depending on Incidence Level on Match Day.

Figure A10 
Effect of Matches per 100,000 Inhabitants on Away Team Counties.
Figure A10

Effect of Matches per 100,000 Inhabitants on Away Team Counties.

Figure A11 
Effect of Matches per 100,000 Inhabitants (Occupancy Level).
Figure A11

Effect of Matches per 100,000 Inhabitants (Occupancy Level).

Figure A12 
Effect of Visitors per 100,000 Inhabitants (Occupancy Level).
Figure A12

Effect of Visitors per 100,000 Inhabitants (Occupancy Level).

Figure A13 
Effect of Ghost Matches per 100,000 Inhabitants.
Figure A13

Effect of Ghost Matches per 100,000 Inhabitants.

Figure A14 
Effect of Matches and Visitors per 100,000 Inhabitants (Top Leagues).
Figure A14

Effect of Matches and Visitors per 100,000 Inhabitants (Top Leagues).

Figure A15 
Effect of Matches and Visitors per 100,000 Inhabitants on Deaths.
Figure A15

Effect of Matches and Visitors per 100,000 Inhabitants on Deaths.

Figure A16 
Effect of Matches and Visitors per 100,000 Inhabitants on COVID-19 Cases (Extended Observation Period).
Figure A16

Effect of Matches and Visitors per 100,000 Inhabitants on COVID-19 Cases (Extended Observation Period).

Figure A17 
Mobility Change on Match Days (Daily Event Study).
Figure A17

Mobility Change on Match Days (Daily Event Study).

Figure A18 
Effect of Matches per 100,000 Inhabitants (Different Regional Inference Approach I).
Figure A18

Effect of Matches per 100,000 Inhabitants (Different Regional Inference Approach I).

Figure A19 
Effect of Matches per 100,000 Inhabitants (Different Clusters for Standard Errors).
Figure A19

Effect of Matches per 100,000 Inhabitants (Different Clusters for Standard Errors).

Figure A20 
Effect of Matches per 100,000 Inhabitants (4 Week Post-Treatment Effect Window).
Figure A20

Effect of Matches per 100,000 Inhabitants (4 Week Post-Treatment Effect Window).

Figure A21 
Effect of Matches and Visitors per 100,000 Inhabitants – Controlling for NPIs.
Figure A21

Effect of Matches and Visitors per 100,000 Inhabitants – Controlling for NPIs.

Figure A22 
Effect of Matches per 100,000 Inhabitants (Travel Robustness Checks).
Figure A22

Effect of Matches per 100,000 Inhabitants (Travel Robustness Checks).

Figure A23 
Effect of Matches per 100,000 Inhabitants (Absorbing Dummy Approach).
Figure A23

Effect of Matches per 100,000 Inhabitants (Absorbing Dummy Approach).

References

Adda, J. 2016. “Economic Activity and the Spread of Viral Diseases: Evidence From High Frequency Data.” Quarterly Journal of Economics 131(2): 891–941. 10.1093/qje/qjw005Search in Google Scholar

Flughafenverband ADV. 2020. Aktuelle Verkehrszahlen. https://www.adv.aero/aktuelle-verkehrszahlen/. Accessed: 2022-02-08. Search in Google Scholar

Ahammer, A., M. Halla, and M. Lackner. 2020. “Mass Gatherings Contributed to Early COVID-19 Spread: Evidence from US Sports.” COVID Economics 30: 44–62. Search in Google Scholar

Bernheim, B. D., N. Buchmann, Z. Freitas-Groff, and S. Otero. 2020. “The Effects of Large Group Meetings on the Spread of COVID-19: The Case of Trump Rallies.” Working Paper. 10.2139/ssrn.3722299Search in Google Scholar

Betsch, C., L. H. Wieler, and K. Habersaat. 2020. “Monitoring Behavioural Insights Related to COVID-19.” Lancet 395(10232): 1255–1256. 10.1016/S0140-6736(20)30729-7Search in Google Scholar

Cardazzi, A., B. Humphreys, J. E. Ruseski, B. P. Soebbing, and N. Watanabe. 2020. “Professional Sporting Events Increase Seasonal Influenza Mortality in US Cities.” West Virgina University Economics Faculty Working Paper Series 6-2020. 10.2139/ssrn.3628649Search in Google Scholar

Chang, S., E. Pierson, P. W. Koh, J. Geradin, B. Redbird, D. Grusky, and J. Leskovec. 2021. “Mobility Network Models of COVID-19 Explain Inequalities and Inform Reopening.” Nature 589: 82–87. 10.1038/s41586-020-2923-3Search in Google Scholar

Chinazzi, M., J. T. Davis, M. Ajelli, C. Gioannini, M. Litvinova, S. Merler, A. P. y. Pionotti, K. Mu, L. Rossi, K. Sun, C. Viboud, X. Xiong, H. Yu, M. E. Halloran, I. M. Longini, Jr., and A. Vespignani. 2020. “The Effect of Travel Restrictions on the Spread of the 2019 Novel Coronavirus (COVID-19) Outbreak.” Science 368: 395–400. 10.1126/science.aba9757Search in Google Scholar

Cotti, C. D., B. Engelhardt, J. Foster, E. T. Nesson, and P. S. Niekamp. 2021. “The Relationship Between In-Person Voting and COVID-19: Evidence from the Wisconsin Primary.” Contemporary Economic Policy 39(4): 760–777. 10.3386/w27187Search in Google Scholar

Coven, J., A. Gupta, and I. Yao. 2020. “Urban Flight Seeded the COVID-19 Pandemic Across the United States.” COVID Economics 54(4): 121–157. 10.2139/ssrn.3711737Search in Google Scholar

Dave, D., D. McNichols, and J. J. Sabia. 2021. “The Contagion Externality of a Superspreading Event: The Sturgis Motorcycle Rally and COVID-19.” Southern Economic Journal 87(3): 769–807. 10.3386/w27813Search in Google Scholar

Dave, D. M., A. I. Friedson, K. Matsuzawa, D. McNichols, C. Redpath, and J. J. Sabia. 2020a. “Risk Aversion, Offsetting Community Effects, and COVID-19: Evidence from an Indoor Political Rally.” NBER Working Paper 27522. 10.3386/w27522Search in Google Scholar

Dave, D. M., A. I. Friedson, K. Matsuzawa, J. J. Sabia, and S. Safford. 2020b. “Black Lives Matter Protests, Social Distancing, and COVID-19.” NBER Working Paper 27408. 10.2139/ssrn.3631599Search in Google Scholar

DeFilippis, E., S. M. Impink, M. Singell, J. T. Plozer, and R. Sadun. 2020. “Collaborating During Coronavirus: The Impact of COVID-19 on the Nature of Work.” NBER Working Paper 27612. 10.3386/w27612Search in Google Scholar

EPFL. 2018. European Leagues Fan Attendance Report. Accessed 08-25-2020. Search in Google Scholar

Felbermayr, G., J. Hinz, and S. Chowdhry. 2021. “Apres-ski: The Spread of Coronavirus from Ischgl Through Germany.” German Economic Review 22(4): 415–446. 10.1515/ger-2020-0063Search in Google Scholar

Fetzer, T. 2022. “Subsidizing the Spread of COVID-19: Evidence from the UK’s Eat-Out To-Help-Out Scheme.” Economic Journal 132(643): 1200–1217. https://doi.org/10.1093/ej/ueab074. 10.1093/ej/ueab074Search in Google Scholar

Fetzer, T., and T. Graeber. 2021. “Measuring the Scientific Effectiveness of Contact Tracing: Evidence from a Natural Experiment.” Proceedings of the National Academy of Sciences 118(33): e2100814118. 10.1073/pnas.2100814118Search in Google Scholar

Fuest, C., A. Peichl, and S. Siegloch. 2018. “Do Higher Corporate Taxes Reduce Wages? Micro Evidence from Germany.” American Economic Review 108(2): 393–418. 10.1257/aer.20130570Search in Google Scholar

Gitter, S. R. 2017. “The H1N1 Virus and Mexican Baseball Attendance.” Athens Journal of Sports 4(4): 263–275. 10.30958/ajspo.4.4.2Search in Google Scholar

Gupta, S., T. D. Nguyen, F. Lozano Rojas, S. Raman, B. Lee, A. Bento, K. I. Simon, and C. Wing. 2021. “Tracking Public and Private Responses to the COVID-19 Epidemic: Evidence from State and Local Government Actions.” American Journal of Health Economics 7(4): 361–404. 10.1086/716197Search in Google Scholar

Hartmann, U. 2020. “Das soll jetzt erst einmal eine Art Experiment werden.” Sueddeutsche Zeitung. Search in Google Scholar

Haug, N., L. Geyrhofer, A. Londei, E. Dervic, A. Desvars-Larrive, V. Loreto, B. Pinior, S. Thurner, and P. Klimek. 2020. “Ranking the Effectiveness of Worldwide COVID-19 Government Interventions.” Nature Human Behaviour 4(12): 1303–1312. 10.1038/s41562-020-01009-0Search in Google Scholar

Hess, P. 2020. “Eintracht-Vorstand Hellmann spricht von “Symbolpolitik”.” Frankfurter Allgemeine Zeitung. Search in Google Scholar

Isphording, I. E., M. Lipfert, and N. Pestel. 2021. “Does Re-Opening Schools Contribute to the Spread of Sars-Cov-2? Evidence from Staggered Summer Breaks in Germany.” Journal of Public Economics 198, 104426. 10.1016/j.jpubeco.2021.104426Search in Google Scholar

Krenz, A., and H. Strulik. 2021. “The Benefits of Remoteness – Digital Mobility Data, Regional Road Infrastructure, and COVID-19 Infections.” German Economic Review 22(3): 257–287. 10.1515/ger-2020-0068Search in Google Scholar

Lange, M., and O. Monscheuer. 2021. Spreading the Disease: Protest in Times of Pandemics. Working Paper. 10.2139/ssrn.3787921Search in Google Scholar

Lichter, A., N. Pestel, and E. Sommer. 2017. “Productivity Effects of Air Pollution: Evidence From Professional Soccer.” Labour Economics 48:54–66. 10.1016/j.labeco.2017.06.002Search in Google Scholar

Mangeum, D., and P. Niekamp. 2022. “JUE Insight: College Student Travel Contributed to Local COVID-19 Spread.” Journal of Urban Economics 127: 103311. https://doi.org/10.1016/j.jue.2020.103311. 10.1016/j.jue.2020.103311Search in Google Scholar

McCloskey, B., A. Zumla, G. Ippolito, L. Blumberg, P. Arbon, A. Cicero, T. Endericks, P. L. Lim, and M. Borodina. 2020. “Mass Gathering Events and Reducing Further Global Spread of COVID-19: A Political and Public Health Dilemma.” Lancet 395(10230): 1096–1099. 10.1016/S0140-6736(20)30681-4Search in Google Scholar

Memish, Z. A., R. Steffen, P. White, O. Dar, E. I. Azhar, A. Sharma, and A. Zumla. 2020. “Mass Gatherings Medicine: Public Health Issues Arising From Mass Gathering Religious and Sporting Events.” Lancet 393(10185): 2073–2084. 10.1016/S0140-6736(19)30501-XSearch in Google Scholar

Mendolia, S., O. Stavrunova, and O. Yerokhin. 2021. “Determinants of the Community Mobility During the COVID-19 Epidemic: The Role of Government Regulations and Information.” Journal of Economic Behavior and Organization 184:199–231. 10.1016/j.jebo.2021.01.023Search in Google Scholar

Mense, A., and C. Michelsen. 2020. “Räumliche Ausbreitung von COVID-19 durch interregionale Verflechtungen.” Wirtschaftsdienst 100(6): 416–421. 10.1007/s10273-020-2674-7Search in Google Scholar

Nunan, D., and J. Brassey. 2020. What is the Evidence for Mass Gatherings During Global Pandemics? A Rapid Summary of Best-Available Evidence. Oxford COVID-19 Evidence Service. Search in Google Scholar

Olczak, M., J. J. Reade, and M. Yeo. 2020. “Mass Outdoor Events and the Spread of an Airborne Virus: English Football and COVID-19.” COVID Economics 47:162–183. 10.2139/ssrn.3682781Search in Google Scholar

Parnell, D., A. Bond, P. Widdop, and D. Cockayne. 2021. “Football Worlds: Business and Networks during COVID-19.” Soccer & Society 221-2: 19–26. https://doi.org/10.1080/14660970.2020.1782719. 10.4324/9781003318811-5Search in Google Scholar

Parshakov, P. 2021. The Spread of COVID-19 and Attending Football Matches: Lesson from Belarus. Working Paper. 10.2139/ssrn.3764404Search in Google Scholar

Reade, J. J., D. Schreyer, and C. Singleton. 2021. “Stadium Attendance Demand During the COVID-19 Crisis: Early Empirical Evidence From Belarus.” Applied Economics Letters 28(18): 1542–1547. 10.1080/13504851.2020.1830933Search in Google Scholar

Reade, J. J., and C. Singleton. 2020. “Demand for Public Events in the COVID-19 Pandemic: A Case Study of European Football.” European Sport Management Quarterly 21(3): 391–405. 10.1080/16184742.2020.1841261Search in Google Scholar

RKI. 2020a. “Epidemiologischer Steckbrief zu SARS-CoV-2 und COVID-19.” https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_Coronavirus/Steckbrief.html. Accessed: 2020-12-14. Search in Google Scholar

RKI. 2020b. Täglicher Lagebericht des RKI zur Coronavirus-Krankheit-2019 (COVID-19). https://www.rki.de/DE/Content/InfAZ/N/Neuartiges_Coronavirus/Situationsberichte/Nov_2020/2020-11-03-de.pdf?__blob=publicationFile. Accessed: 2022-02-09. Search in Google Scholar

Schlosser, F., B. F. Maier, D. Hinrichs, A. Zachariae, and D. Brockmann. 2020. “COVID-19 Lockdown Induces Disease-Mitigating Structural Changes in Mobility Networks.” Proceedings of the National Academy of Sciences 117(52): 32883–32890. 10.1073/pnas.2012326117Search in Google Scholar

Singleton, C., A. Bryson, P. Dolton, J. J. Reade, and D. Schreyer. 2021. “Economics Lessons from Sports During the COVID-19 Pandemic.” In Research Handbook on Sport and COVID-19, edited by P. M. Pedersen. Edward Elgar Publishing. 10.4337/9781802207576.00008Search in Google Scholar

Stoecker, C., N. J. Sanders, and A. Barreca. 2016. “Success Is Something to Sneeze At: Influenza Mortality in Cities that Participate in the Super Bowl.” American Journal of Health Economics 2(1). 10.1162/AJHE_a_00036Search in Google Scholar

von Bismarck-Osten, C., K. Borusyak, and U. Schoenberg. 2022. “The Role of Schools in Transmission of the SARS-CoV-2 Virus: Quasi-Experimental Evidence from Germany.” Economic Policy. 10.1093/epolic/eiac001Search in Google Scholar

Weed, M., and A. Foad. 2020. “Rapid Scoping Review of Evidence of Outdoor Transmission of COVID-19.” Working Paper 30: 44–62. 10.1101/2020.09.04.20188417Search in Google Scholar

Wing, C., D. H. Simon, and P. Carlin. 2021. “Effects of Large Gatherings on the COVID-19 Epidemic: Evidence From Professional and College Sports.” Economics and Human Biology 43: 101033. 10.2139/ssrn.3657625Search in Google Scholar

Published Online: 2022-04-29
Published in Print: 2022-12-31

© 2022 Walter de Gruyter GmbH, Berlin/Boston

Downloaded on 28.3.2024 from https://www.degruyter.com/document/doi/10.1515/ger-2021-0060/html
Scroll to top button