The fundamental properties of the spread of SARS-CoV-2 have been known since early on. However, it remains difficult to this day to exactly quantify the impact of small or short-term changes in social behavior. The diversity of social behavior poses extreme challenges due to the variety of conditions and temporal changes of social interactions. In our study we make use of the UEFA European Soccer Championship 2020 (EURO2020, which took place in June and July 2021) as a laboratory for assigning COVID-19 cases to short-term behavioural changes. We find that the impact of behavioral changes due to such an event depends strongly on the epidemiological conditions at the start of the event. This quantification is only possible due to the availability of highly resolved data which underlines the importance of excellent data availability in future outbreaks and pandemics as well.
Previous studies have generally found inconclusive results or no impact of the EURO2020 championship on COVID-19 spread since they mostly focussed on local infections in the stadium, which are negligible on a national scale. Studies focussing on overall case numbers have often been inconclusive because it is difficult to attribute an observed longer-term change in the infection dynamics to the EURO2020. Instead of such approaches, we make use of established social habits — the gender imbalance amongst soccer fans, causing an expected asymmetry in COVID-19 infections between genders directly after matches — and the known time structure of matches. where we assume the strongest effect in the countries of the playing parties, and not in the country where the match happens. This requires modeling the weekday-dependent reporting delays in detail, in order to assign short-term surges in COVID-19 cases to match dates. We apply a Bayesian analysis, modeling the gender asymmetry, the infection and detection delay structure, the weekday-dependent reporting delays, and the time-variable underlying gender-symmetric infection dynamics, and including all secondary cases.
Using the time structure and gender asymmetry
It has been noticed already during the course of the EURO2020 in June '21 that male infection rates surged in Scotland and England.
The data is obvious (see Fig. 1) – there is a spike both in the total number of infections (a) and in the male to female gender ratio (b) of infections during and shortly after the time frame of the EURO2020. This observation in itself, combined with the "common knowledge" that on average there are more avid soccer fans among men than women, could be enough to conclude that indeed and undeniably mass events like the EURO2020 have a positive effect on transmissions. However, science does not stop here since many questions arise from that. How can this observation be generalized over all or at least many participating countries? Are the infections linked to visiting match venues, or more likely due to local and private social gatherings connected top watching the games? How can the social "common knowledge" of a more avid interest of men in soccer be incorporated in epidemiological modeling? Is there additional power in the time structure of the data, which can be used to link infections to individual matches (see Fig. 1 (c))? Already at first glance, the Scottish data suggests this link. However the weekday-dependent reporting delays (differing from country to country) certainly further complicate giving a certain answer. Finally, how many cases can be positively attributed to the EURO2020, and maybe most importantly: Is there a connection between the underlying pandemic situation in a country and the number of cases (or fraction thereof relative to all cases) which can be assigned to the EURO2020?
The key functionality of our Bayesian model is visualized in Fig.2, where it is applied to England. The underlying gender-symmetric pandemic course is modeled for each country separately using a change-point based SEIR-model. This allows describing the variation of the rate of infection (represented by the R-value R_base, see Fig. 2 (d)) from causes unrelated to the EURO2020. Furthermore, a gender asymmetric noise term (Fig. 2 (f)) allows to describe long-term differences between the infection rates of the genders, also from causes unrelated to the EURO2020. In order to extract the underlying pandemic dynamics from the data, its time structure needs to be addressed separately for each country. This is achieved using a delay between infection and detection, and its average width (Fig. 2 (g) and (h)), and a weekday dependent fraction of cases which are unreported on the day of discovery (Fig. 2 (k) to (q)). With these prerequisites in place, the model can be used to infer a change in the R-value on each day the team of a given country plays a match. Note that such change need not only be an increase, but could also be a decrease as, in principle, a match could have led to fewer out-of-household interactions, thus potentially reducing the infection rate. In Fig. 2 (e), the inferred changes ΔR_football are shown, where the triangles denote the matches played by England, and the spikes and its uncertainty band show the change in the R-value attributed to the social contacts in England when the English team played – irrespective of where in any European country the match was played.
The Bayesian model allows the parametrization of assumptions including their uncertainty. One example is the asymmetry between male and female participation in fan activity, which varies significantly between countries. Also, rather few social science studies seem to address this question quantitatively, which is in contrast to the ubiquitous assertion that men are much more interested in soccer than women. We assign a rather wide uncertainty to the prior assumption that women comprise about ⅓ of the fans (and further vary this assumption to a wide band around ½ in systematic studies of the robustness of our parametrization). In a country with a strong effect size like England, the data is strong enough to constrain both the gender asymmetry (See Fig. 2 (i)) and the delay structure ((g) and (h)) between infection and discovery. Thus, this study can also be regarded as a measure of female participation in fan activity in England, weighted by the relative transmission risk of a given activity, of which the result is 32% (95% Credible Interval: [28%, 38%]).
Additionally, the spikes in transmission on single days offer the interesting opportunity to cross-check the measurement of the average delay between infection and reporting. In England, a delay of around 4.6 (95% CI: [4.4, 4.9]) days is observed. It is not trivial that this agrees with the results from tracing studies – it would have been well possible that traceable cases impose a bias on the delay, compared to cases where tracing was not even attempted. Also, the strictness of measures might have an effect on the frequency of rapid testing and on the discipline of getting oneself tested in case of symptoms, thus inducing a time-dependence in the delay. Also, the transfer of tracing study results from one country to another one is not trivial, due to different testing and reporting regimes. This by-product of our analysis highlights an interesting aspect of the modeling and measurement of physical processes from which we as authors are coming: Physical systems and experiments often make it possible to validate the assessment of uncertainties, or even quantitatively measure systematic uncertainties, by comparing independent approaches to measuring the same physical quantity. Over-constraining our understanding of epidemiological processes and, thus, gaining the ability to quantitatively cross-validate many of our modeling assumptions is an important aspect of present activities.
The average time difference between matches of a given national team of 4 days is interesting also with regard to the transmission properties of SARS-CoV-2: After 4 days, many infected are already infectious, but not yet symptomatic. Also, the average delay between infection and detection is on average longer than the distance between matches. Accidentally, and unfortunately, the typical rhythm of matches is probably adjusted close to an optimum for COVID-19 spread.
The assignment of cases to the EURO2020 in our model is generally rather robust. Changing the prior assumptions on the delay structure, the gender imbalance or the time structure of the fitted transients of the underlying infection dynamics do not significantly alter the results. Counterfactually moving all match dates by an offset relative to the pandemic data yields first a corresponding counterfactual change in the delay parametrisation, and then, for counterfactual time shifts of 14 days or more, a breakdown of the sensitivity of the model. This shows that the model is not picking up underlying spurious variations of the gender asymmetry. Also, adding a local effect in the country where the match was played on the day of the match does not alter the results, and yields no significant effects for the countries of the match locations. This is in agreement with previous assessments of the EURO2020 pandemic effect, which focussed on local effects in or around the match venues and generally found no significant result.
The results are significant and variable between countries
The results vary significantly between 12 European participating countries, as shown in Fig.3. Overall, around 840000 (95% CI: [0.39M, 1.26M]) primary and secondary cases are assigned to the EURO2020 by the model. In addition, most countries under study show an increase in the underlying pandemic modeled by R_base during the course of the EURO2020. This hints towards an underestimating tendency of the model: Gender symmetric infections with nonspecific reporting time structure are assigned to the underlying pandemic developments and not to the fan activity related to the EURO2020, no matter whether they really come from fan activity or not. A possibility for the cause of such a gender-symmetric and smooth acceleration of the pandemic spread is an additional increase of the frequency of private gatherings, largely of couples, also to watch matches of other teams than the ones of their own nationality. Due to the random occurrence of such gatherings on any day during the championship – and not only on the days where the own national team plays, which induces a strong spike in fan activity – the time structure is smooth and the model is conservatively compelled to assign these cases to the underlying pandemic, either through R_base in case of gender-symmetric effects, or through R_noise in case of a slight gender asymmetry.
At first glance, the different results in different countries might seem inconclusive. However, the observation of these differences is the basis for studying the dependence of the EURO2020 related infections on the observed infection dynamics in each country. After all, it is an important goal to allow social interaction without negative health impact. Thus, understanding why in some countries the EURO2020 seems to have had a lesser effect than in others is an important step.
Predicting the impact and determining long-term mitigation strategies for large short-term events
When comparing the relative total pandemic effect of the EURO2020 to individual observables of the contact behavior in each country, we generally find no or little significant correlation. However, it can already be observed in Fig. 1 (a), that the largest part of the EURO2020 related infections are secondary infections (yellow). Compared to primary infections (red), there is generally a ratio of 4:1 to 5:1 between subsequent and primary infections in the countries under study. This means that the total impact must be compared to a combination of different effects, and not just a single observable: The number of matches played, the fraction of infected people at fan activities, and the potential for the gender-symmetric spread of subsequent infection chains as given by the underlying infection dynamics. Combining these three aspects with each other (using the R-value before the start of the EURO2020 as a proxy for the underlying dynamics during the championship) yields a strong correlation with the total impact (see Fig. 4). It is important to note that this is unlikely to be an accidental spurious correlation caused by trying out an arbitrary number of auxiliary observables and reporting the result with the highest correlation: The functional form used for calculating the "Potential for pandemic spread" in Fig. 4 is what an SEIR-model (with constant pandemic parameters) yields a priori for the expected impact of injections of additional infections.
This observation is a further confirmation for important long-standing policy recommendations: A low infection activity and low R-value ensure that the society can enjoy more freedom without negative health effects in a pandemic. Minimally invasive measures (e.g. very good and fast contact tracing forwards and backwards, masks in highly frequented enclosed areas, frequent (self) testing, etc) and non-invasive measures (e.g. sufficient ventilation and air cleaning in all highly frequented public spaces), which reduce the underlying pandemic spread, are thus a direct enabler of socially attractive behavior which might come with increased infection risks, such as coming together and shouting at a TV screen and rooting for your team in a crowded pub.
The importance of good data and being prepared for the next pandemic
One last question might remain: Why have only 12 countries and not all participants of the EURO2020 been studied? And why is there the result for the EURO2020 now, and no mention of the recent world cup in Qatar in December 2022? The answer is simple: data with a daily time resolution (necessary for the assignment of cases to matches, which follow at a succession of less than 7 days, on average) and resolved in gender was only available for 12 of the participating countries of the EURO2020. It would be much better if such data was available for all countries, and if the parameterization of the delay structure was not necessarily different for all countries, i.e. if a common data management allowed to study the data with much higher precision: many comparisons between countries are muddled because in each country not only the underlying infection dynamics, but also properties of the data collection and publishing have to be modeled separately.
But what about the World Cup in 2022? A brief look into the data sources show that this would not be possible. Gender resolved data is published in fewer and fewer countries, and many countries such as England moved to reporting weekly averages. This decline in precision and resolution might seem not very relevant at first glance; after all, efficient vaccines and high immunization have allowed those countries to return to almost pre-pandemic conditions. However, the conclusion cannot be that pandemics are over. To the contrary, the next pandemic will come for sure. Being prepared means that the detection and detailed understanding of the next pandemic through detailed modeling should be even faster than for SARS-CoV-2, and that instruments should be available which allow policymakers to instantly react to the modeling predictions for pandemic threats. A decline in data quality does not bode well for being prepared.