Abstract
The 2022 FIFA World Cup was the first major multi-continental sporting Mass Gathering Event (MGE) of the post COVID-19 era to allow foreign spectators. Such large-scale MGEs can potentially lead to outbreaks of infectious disease and contribute to the global dissemination of such pathogens. Here we adapt previous work and create a generalisable model framework for assessing the use of disease control strategies at such events, in terms of reducing infections and hospitalisations. This framework utilises a combination of meta-populations based on clusters of people and their vaccination status, Ordinary Differential Equation integration between fixed time events, and Latin Hypercube sampling. We use the FIFA 2022 World Cup as a case study for this framework. Pre-travel screenings of visitors were found to have little effect in reducing COVID-19 infections and hospitalisations. With pre-match screenings of spectators and match staff being more effective. Rapid Antigen (RA) screenings 0.5 days before match day outperformed RT-PCR screenings 1.5 days before match day. A combination of pre-travel RT-PCR and pre-match RA testing proved to be the most successful screening-based regime. However, a policy of ensuring that all visitors had a COVID-19 vaccination (second or booster dose) within a few months before departure proved to be much more efficacious. The State of Qatar abandoned all COVID-19 related travel testing and vaccination requirements over the period of the World Cup. Our work suggests that the State of Qatar may have been correct in abandoning the pre-travel testing of visitors. However, there was a spike in COVID-19 cases and hospitalisations within Qatar over the World Cup. The research outlined here suggests a policy requiring visitors to have had a recent COVID-19 vaccination may have prevented the increase in COVID-19 cases and hospitalisations during the world cup.
Author summary Mass Gathering Events (MGEs) can potentially lead to outbreaks of infectious disease and facilitate the dissemination of such pathogens. We have adapted previous work to create a framework for simulating disease transmission and mitigation at such MGEs. We use the 2022 FIFA World Cup as a test case for this framework. A policy of Pre-travel screenings of visitors was found to have little effect in reducing COVID-19 cases and hospitalisations. Pre-match screenings of spectators and match staff was found to be more effective. The most effective policy was to ensure that all visitors had a COVID-19 vaccination (second or booster dose) within a few months before departure. Qatar abandoned all COVID-19 related travel testing and vaccination requirements over the period of the World Cup. Our work suggests that the State of Qatar may have been correct in abandoning the pre-travel testing of visitors. However, there was a spike in COVID-19 cases and hospitalisations within Qatar over the World Cup. Given our findings, we suggest a policy requiring visitors to have had a recent COVID-19 vaccination may have prevented the increase in COVID-19 cases and hospitalisations during the world cup.
1 Introduction
The continuing COVID-19 pandemic, caused by an emerging coronavirus (1), has been affecting more than 200 countries since early 2020, profoundly overwhelming healthcare infrastructure worldwide (2; 3). Given the initial lack of availability of effective drugs and vaccines, in order to control and contain the pandemic, governments and authorities have implemented a package of public health interventions. Such interventions have collectively become known as NPIs (Non-Pharmaceutical Interventions) (4). In some countries, NPIs have included the ban of inter-household mingling and/or outdoor activities, particularly Mass Gathering Events (MGEs). As such, there is a large body of work pointing to the ban of MGE as an effective NPI (4). Furthermore, there are several examples of outbreaks of communicable diseases at MGEs occurring in the pre-COVID era, many of which contributed to the global dissemination of the pathogens responsible (5; 6; 7; 8; 9; 10; 11). The WHO defines MGEs as highly visible events attended by tens of thousands of people, such as pilgrimages and sporting events, and coordinates with member states on matters of pathogen control at such gatherings (12).
The resulting ban of MGEs in the wake of COVID-19 has affected the sporting world. Athletes have had to cope with unprecedented disruption, characterized by the loss of regular routine (e.g. training and matches), and the postponement or even cancellation of major national and international sporting events (Tokyo 2020 Olympic and Paralympic Games). During the pandemic, sports organizations have collaborated closely with the WHO and national public health bodies, assessing and implementing COVID-19-related risk reduction interventions to facilitate a safe step-wise return of sporting events (13). Generally, these measures have focused on three areas. First, lowering the risk from the actual sport itself: focusing on activities that can maintain physical distancing; holding matches outdoors; re-consider whether to allow contact sports (13). Second, lowering risks inherent to the size of an event, considering both participants and spectators. Third, reducing risks inherent to the geographic localisation of the event by considering the local epidemiological conditions such as COVID-19 community transmission and prevalence. There are many COVID-19 mitigation protocols that can be utilised in planning for MGEs, either sporting events or others. However, the effectiveness and performance of these protocols in controlling and reducing the risk of COVID-19 transmission and hospitalisations is not clearly established.
Here we build on previous work (14) to create a generalised framework simulating disease transmission specifically adapted for use in planning pathogen control at MGEs. The 2022 FIFA World Cup hosted in Qatar was the first multi-continental sporting MGE of the post COVID-19 era to allow foreign spectators (15; 16; 17). Denhing et al. (2023) and Subedi et al. (2022) (18; 19) highlighted the potential for disease spread at the world cup. Therefore, we chose to use the 2022 FIFA World Cup as a test case of the framework we developed, assessing various strategies to mitigate COVID-19 spread through match attendee testing and visitor vaccination requirements.
2 Methods
2.1 A Generalised Framework for Simulating Disease Transmission at Mass Gatherings
In order to model the spread of COVID-19 at MGEs we have built upon our previous work (14) and created a generalised deterministic model framework (see Fig 1, Eqs 1 and Tables 1 to 3). The general framework is that of a metapopulation stratified by clusters and vaccination groups, designated by subscripts i and v, respectively. Cluster composition is customisable to a range of MGEs. Specifically for the model outlined in this manuscript there are three main sets of clusters, the hosts, the visitor fans of team A and the visitor fans of team B. The effects of vaccination are controlled through parameters designated with a subscript v, effecting classes denoted within that vaccine group (see Vaccination Groups). ODE integration of this model framework is achieved through Scipy’s ODEint function (20).
A: all but the states with a * notation move between the vaccination groups depicted B at rates νv=unvaccinated, νv=effective or νv=waned.
2.1.1 Disease Stages
Within each population cluster and vaccination group susceptible individuals, Siv, can be infected through the force of infection λiv (see Fig 1A, Eqs 1 and Force of Infection). Infection leads to the early latent stages of infection Eiv, where infection is not detectable through RT-PCR or Rapid Antigen (RA) tests. From here infected individuals progress (ϵ1) to one of two later latent phases GIiv or GAiv, where infection is detectable through RT-PCR tests but not RA tests. Here an individuals infection pathway diverges either down a path leading to eventual symptoms at a proportion ps or asymptomatic infection at a proportion 1 − ps (with classes denoted with subscripts I and A, respectively) (see Fig 1A).
Infections become both transmissible, and detectable through RA tests, at rate ϵ2, moving to the incubating phases PIiv and PAiv (24; 40; 41). From this stage on the asymptomatic track, PAiv, people progress at rate ϵ3 to stages MAiv, then at rate γ1 to FAiv, finally recovering at rate γ2 to Riv. If on the symptomatic track PIiv people progress to the first stages of symptoms at rate ϵ3. Here there is a risk of people progressing down the hospitalisation pathway, moving to stage MHiv, at probability ph|s,v (see Fig 1A). Eventually individuals in MHiv are hospitalised at rate ϵH moving to compartment FHiv. It is assumed that those hospitalised do not contribute to the force of infection (see Force of Infection). Recovery from hospitalisation, FHiv, occurs at rate γH and leads to stage Riv. If a person does not move to the hospitalised pathway, 1 − ph|s,v, they remain on the symptomatic pathway develop symptoms and progress to stage MIiv. From here people progress to the final stage of infection FIiv at rate γ1 and then to recovered class, Riv, at rate γ2.
2.1.2 Vaccination Groups
All individuals start in the Unvaccinated group (indexed as 1 in Figure Fig 1B). After completing a primary series of vaccination people move to the Effective vaccination group, νv=unvaccinated (indexed as 2 in Figure Fig 1B). Several months after primary series of vaccination immunity wanes (26; 42) moving people from vaccine group Effective to Waned, νv=effective. The waned vaccination group is indexed as group 3. Note subscript v indexes the vaccination group not the number of doses of a vaccine. Individuals in the Waned vaccination group can receive a booster dose, at rate νv=waned, moving them back to the Effective vaccination group. Again after several months in the effectively vaccinated group immunity wanes, at rate νv=effective, moving people to the Waned vaccination group. In other words, after a primary series of vaccination people loop from the Effective vaccination group to the Waned vaccination group through the waning of immunity, νv=effective, and back again with booster doses, νv=waned. In concordance with many national public health agencies’ advice (43; 44) only non-symptomatic people (i.e. all classes but MIiv, MHiv, FIiv, and FHiv) can be vaccinated at rates νv=unvaccinated or νv=waned. The effectiveness of vaccination plays out in the different vaccine groups, through modification of force of infection (λiv) and hospitalisation (ph|s,v), see Eqs 2, Tables 2 and 3
2.1.3 Clusters
Clusters come under two main categories, visitor clusters and host clusters (see Table 4). In order to simulate COVID-19 screening, each of these main clusters have associated clusters for “RA Positive”, “Waiting for Positive RTPCR” and “RTPCR Positive”. Tests are simulated using the event queue system (see Event Queue). In the case of RA test events, a proportion of a clusters population (τRA) from states PIiv, MIiv, FIiv,MHiv, PAiv, MAiv and FAiv are moved to the associated RA Positive cluster (see Fig 1A). This detected proportion being based on the RA tests sensitivity (40). Those in the RA Positive cluster are isolating and thereby contribute less to transmission (see Table 4 and Force of Infection). RT-PCR tests are capable of detecting the presence of COVID-19 earlier in an infection (40), meaning that the proportion of a clusters population (τRT −P CR) is also drawn from states GIiv and GAiv (see Fig 1A). However, RT-PCR tests have a much longer turnaround time (40), typically a day or two (37; 38; 39). Therefore, the detected proportion from RT-PCR tests (τRT −P CR) will populate a “Waiting for Positive RT-PCR” cluster. All the classes in the “Waiting for Positive RT-PCR” cluster transition to the associated “RTPCR Positive” cluster at rate ωRT −P CR (RTPCR turnaround time). As with the RA Positive cluster, the RT-PCR Positive cluster is isolating and thereby contributes less to transmission (see Table 4 and Force of Infection).
2.1.4 Event Queue
In order to simulate changes in parameter values (such as increasing transmission) and the transfer of population between compartments (e.g. moving to isolation) an event queue system has been employed. This runs a model between events, then changes a parameter value, adds or deducts from compartments in a compartment model depending on the event. The code for this has been made freely available (see S1 Methods). A note of caution with making comparisons between scenarios with events at different times. If no event occurs at a time point in one scenario but there is an event in the other at that time, a null (do nothing) event must be inserted at that time point for simulations made without the event at that time. This is critical to ensure comparable accuracy of the integration for simulations of distinct scenarios.
2.1.5 Force of Infection
Force of infection is calculated for each cluster summing up the contribution from all clusters (including itself) (j) and their vaccination groups (v) (see Eq 3). As already mentioned states that do not display symptoms have their transmission modified by θ. Isolation is achieved in “RA Positive” and “RTPCR Positive” clusters by their κj = κ, for other clusters κj = 1.
The transmission term βij refers to transmission to cluster i from cluster j. For the majority of simulation time this is set at a baseline (βij = β). However, this can be changed using the event queue system to have βij = bβ for a period of time, b being a strengthening or weakening of transmission over that time period. Ni∗ represents the population in which the interaction between a susceptible individual of cluster i (Siv) and an infectious individual of cluster j (PIjv, PAjv, MAjv, FAjv, MIjv, FIjv, or MHjv) takes place. Similarly to the transmission term (βij), Ni∗ is typically set at the baseline value of the entire population being modelled (N). However, this can be changed using the event queue system allowing for transmission to be modelled through interactions taking place within certain sub-populations. Note the summation term
means to sum through all the infectious stages of all the vaccination groups of cluster j, in this case vaccine groups 1: Unvaccinated, 2: Effective and 3: Waned. Recall from §Vaccination Groups that the subscript v indexes the vaccination group not the number of doses of a vaccine.
2.2 Simulating FIFA 2022 World Cup Matches, as a Case Study
For a test case scenario of the generalised framework above (see A Generalised Framework for Simulating Disease Transmission at Mass Gatherings), we chose to model possible matches from the FIFA 2022 World Cup (not involving the Qatari team). Each match is seen as a 7 day MGE (see Uncertainty and Sensitivity Analyses).
2.2.1 Simulation of a FIFA 2022 World Cup Match
For each match there were five main clusters, one for hosts in general, one for host spectators, one for host staff and two clusters of visitor fans, one for each team, (see Tables 4 and 5). The eight stadiums hosting matches have estimated capacities ranging from 40,000 to 80,000 (45). We assume therefore that the population attending simulated fixtures ranges from 4,000 to 80,000 (NA). A proportion of tickets go to the host spectator cluster (0 <= ηspectators <= 0.5), meaning that the two visitor clusters made up the rest of the attendees, NA, split evenly. The host staff cluster population, NS, ranged from 4,000 to 20,000. The host general population cluster equaled the population of Qatar, 2,930,524 (46), minus the host spectator (NA ∗ ηs) and staff cluster (NS) populations.
No vaccination occurred during simulations
(νv=unvaccinated = νv=waned = νv=effective = 0), as the simulations occur over a short time frame. The host unvaccinated population was set at Qatar’s population minus the number of people fully vaccinated in Qatar as of 15/11/2022 (47). The hosts effectively vaccinated population was set as the total boosters given as of 15/11/2022 (47). The hosts waned vaccination group was populated with people fully vaccinated minus total boosters given.
Team A and B fans were assumed to have at least completed a primary series of vaccination. Prior to the world cup Qatar had travel restrictions requiring a primary series of vaccination to access public facilities (17; 49; 50). The proportion of effectively vaccinated in these two clusters therefore ranged between simulations, 0 <= vA <= 1 for Team A and 0 <= vB <= 1 for Team B fans. The remaining population of these two clusters was placed in the waned vaccination group.
The Host, Host Spectator, Host Staff, Team A and Team B Clusters were seeded with COVID-19 infections. For the three host clusters the starting prevalence, σH, was sampled from a range (see Table 5). The 7-day smoothed new cases per person for Qatar on 18/11/2022 (46; 47), multiplied by the lower and upper estimate of reported to actual infections for Qatar (48) to give this range. Similarly, the starting prevalence for Team A and B fans (σA and σB) was also sampled from a range based on the smoothed new cases per person 18/11/2022 (46; 47). The new smoothed cases per person for each nation was multiplied by the respective lower and upper estimate of reported to actual infections (48). The minimum and maximum from this set of values then informed the range for starting prevalences for Team A and Team B fans (see Table 5).
We chose a probabilistic approach to pick the seed infection stages. First we made a random draw to select which infection pathway (branch) a host is on, using the probability of being on an infection pathway. Then each infection stage of a pathway was assigned a weight. The weight was calculated as the inverse of the outflow rate from that compartment. We normalised the weights for each infectious stage by dividing by the sum of all weights, prior to using the result to draw the selection of infection stage. All draws were made using numpy’s multinomial function (51) see S1 Methods.
The baseline transmission term β was derived for a given value of R0, assuming no vaccination and a single cluster population. R0 was derived using Next Generation Matrix Methods (52) and sympy (53) (see S2 Methods for details). Simulation then proceeded as outlined in Table 6. We initiated the simulation 2 days prior to the actual MGE so as to capture pre-travel COVID-19 screenings (54). The simulation extended from 7 to 100 days post MGE without transmission, so as to capture the number of hospitalisations resulting from transmission during the MGE.
2.2.2. Uncertainty and Sensitivity Analyses
Parameters and starting variable values were either held fixed or sampled using Latin Hypercube Sampling (LHS) (55), using scipy’s LatinHypercube function (20) (see Tables 2, 3 and 5). LH sampling was done using uniform distributions and a sample size of 10,000. Partial Rank Correlation Coefficients (PRCCs) were then used to asses the effect of each sampled parameter on total hospitalised, peak hospitalised, total infected and peak infected. PRCCs were calculated using pingouin’s partial corr function (56).
2.2.3. Analyses of Testing Strategies
In order to asses the effect of different test strategies the same LH sample was run with each of the testing regimes described in Table 7. The effectiveness of the test strategies was measured through two sets of comparisons using the outputs total infections, peak infections, total hospitalisation and peak hospitalisation. The first set of comparisons were PRCC based. Each set of simulations made under a testing strategy was paired with the set of simulations made with no testing regime in place, as a control. For simulations under the test strategy a dummy parameter was given a value of 1. Simulations made without a testing regime in place were given a value of 0 for this dummy parameter. Thus, creating a parameter to base PRCC comparisons on. The second set of comparisons measured a testing regime’s percentage relative differences in outputs, using Eq 4, compared to the “No Testing” regime as a control. Regarding Eq 4, Rl is the percentage relative difference in an output O seen between a simulation with a treatment T and a control simulation C, where l is the LH sample used in the two simulations being compared.
2.2.4 Analyses of Travel Vaccination Restrictions
The proportions of visitor A and B effectively vaccinated (vA and vB) where both found to have significantly negative PRCCs with infections and hospitalisation (see Effects of Parameters and Starting Conditions Relating to COVID-19 control measures). This suggests that a policy restricting entry to those effectively vaccinated but no COVID-19 screening being enforced, was worth evaluating. Henceforth we will refer to such a policy as “effective visitor vaccination”. Therefore, a further LHS of size 10,000 was drawn, this time without vA and vB from Tables 2, 3 and 5 being sampled. The LHS parameter sets were then used to simulate a policy of “effective visitor vaccination” (vA = vB = 1 and no testing being in place). Calculations of percentage relative differences in total infections and hospitalisations between this policy, as a control, against simulations made under a different combination of testing regime and visitor effective vaccination (vA = vB) with the same LH sample set as treatments were made (see Eq 4). These combinations comprised of vA = vB = 0, vA = vB = 0.25, vA = vB = 0.5 or vA = vB = 0.75 with “No Testing”, “Pre-Travel RT-PCR”, “Pre-Match RA” or “RT-PCR then RA” testing regimes. Thereby, capturing a testing regime being in place with different background levels of visitor effective vaccination.
3 Results
Here we focus on an analysis of testing regimes, along with the parameters and starting conditions relating to COVID-19 control measures. The supplementary materials contain further analyses of the effects of other parameters and starting conditions that we varied through LHS. We note that for nearly all PRCCs of starting conditions, parameters and testing regimes against peak infections and hospitalisation followed the same trends as total infections and hospitalisation. The exceptions being rate of hospitalisation and recovery from hospitalisation. Similarly % relative differences caused by testing regimes follow the same trend when comparing the peak and total number of infections or hospitalisations (see S1 Results).
3.1 Effects of Testing Regimes
PRCCs from single testing regime showed much higher performance for pre-match testing over pre-travel testing in reducing infections and hospitalisations (see Fig 2). Furthermore, pre-travel screenings provide less reductions in infections and hospitalisation compared to pre-match screenings (see Fig 3). Single RT-PCR tests had a greater benefit in pre-travel testing, but less beneficial in pre-match testing compared with single RA testing. Therefore, it is unsurprising that the best overall testing regime for mitigating infections and hospitalisation was a pre-travel RT-PCR test and then a pre-match RA test, “RT-PCR then RA” (see Figs 2 and 3). However, “RT-PCR then RA” testing regime (pre-travel RA and pre-match RA) only narrowly outperformed the “Double RA” testing regime and was not much of an improvement over a single “pre-match RA” test (see Figs 2 and 3).
In calculating PRCCs Latin Hypercube (LH) sampling draws on the parameter space outlined in Tables 2, 3 and 5, using uniform distributions. Simulations are made with the resulting LH sample with each of the testing regimes outlined in Table 7. Every set of simulation made under a testing regime is given a dummy parameter value of 1, except ”No Testing” which is given a value of 0. Each testing regime’s effect on an output (Total Infections or Hospitalisation) is measured through calculating PRCCs using the dummy parameter comparing the 1 for the particular testing regime and 0 for its absence.
A: Boxplots Total Infections and Hospitalisation in simulations made with no testing regime. B: Boxplots of a Testing Regimes % Relative Differences in Total infections and Hospitalisation. For every parameter set produced under LHS the % relative difference in outputs simulated under a testing regime, Fig B, was calculated against the corresponding output from the “No Testing” regime simulations, depicted in Fig A, as a control (see Eq 4). The white dots are the means. The array of samples used in simulation was generated from Latin Hypercube sampling drawing upon the distributions outlined in Tables 2, 3 and 5. Details of testing regimes can be found in Table 7.
3.2 Effects of Parameters and Starting Conditions Relating to COVID-19 control measures
In terms of active control measures decreasing the transmission from isolating clusters would only be effective in testing regimes that included pre-match testing Fig 4. Note, pre-travel tests remove positive visitors from the model. However, greater reductions in infections and hospitalisations are seen through reduced transmission from pre-symptomatic and asymptomatic people. This could be achieved through many NPIs, such as encouraging or enforcing mask wearing, promoting hand sanitation and, when possible, social distancing.
Where, κ is the isolation transmission modifier (0-1), θ is the asymptomatic transmission modifier (0.342-1), and vA and vB are the proportion recently vaccinated visitors in group clusters A and B, respectively, (0-1). The array of samples used in simulation was generated from Latin Hypercube sampling drawing upon the distributions outlined above and in Tables 2, 3 and 5, using uniform distributions. Details of testing regimes can be found in Table 7.
The proportions of visitor clusters A and B effectively vaccinated (vA and vB) have a negative correlation with both infections and hospitalisation (see Fig 4). Differences in PRCCs can be transformed to z-scores, as outlined in (55). These methods were used to determine if the effects of Testing Regimes and vA and vB under the ’No Testing’ regime are significantly different (compare Fig 2 and 4). PRCCs of vA and vB compared to single pre-travel screening testing regimes demonstrate a significantly greater effect in reducing hospitalisations and infections (see one tailed p-values in S1 Table). vA and vB have a significantly greater effect in reducing hospitalisations than the “Pre-match RT-PCR” testing regime, but no significant differences are seen when it comes to infections. With all other testing regimes PRCCs for vA and vB demonstrate a significantly greater effect in lowering hospitalisations. However, in terms of infections negative associations with vA and vB are significantly less pronounced compared to the other testing regimes (see one tailed p-values in S1 Table).
3.3 Effects of Proportion of Recently Vaccinated as a COVID-19 Control Measure
It can be seen from Fig 5, that “effective visitor vaccination” (vA = vB = 1 and no testing being in place) outperforms the “Pre-Travel RTPCR” testing regime, reducing both hospitalisations and infections. When it comes to the “Pre-match RA” and “RT-PCR then RA” testing regimes, “effective visitor vaccination” outperforms for reductions in hospitalisations, but not infections. In terms of reducing infections on average (both mean and median), “Pre-match RA” and “RT-PCR then RA” testing regimes outperform “effective visitor vaccination”. At vA = vB = 0.5 the “Pre-match RA” and “RT-PCR then RA” testing regimes result in a similar number of infections as seen in “effective visitor vaccination”. As vA and vB reduce below 0.5 the “Pre-match RA” and “RT-PCR then RA” testing regimes are less effective in terms of controlling infections, compared to “effective visitor vaccination”.
A: Boxplots of Total Infections and Hospitalisation under ”effective visitor vaccination” (vA = vB = 1). B and C: Boxplots of % relative differences in Total Infections and Hospitalisation seen under various testing regimes at differing levels of effective vaccination for visitors compared to ”effective visitor vaccination” as a control. In B and C % relative differences are calculated between simulations made with the same Latin Hypercbe (LH) sample, see Eq 4. Testing regimes used in comparisons are “No Testing”, “Pre-Travel RT-PCR”, “Pre-Match RA” or “RT-PCR then RA” testing regimes (see Table 7). Levels of effective vaccination for visitors in the comparisons are vA = vB = 0, vA = vB = 0.25, vA = vB = 0.5 and vA = vB = 0.75. The white dots on the boxplots represent mean values. All parameters other than those relating to effective vaccination for visitors (vA and vB) are drawn using LH sampling from distributions outlined in Tables 2, 3 and 5.
4 Discussion
Major MGEs such as religious pilgrimages, festivals or sport competitions can generate a variety of health risks. In the context of an ongoing infectious disease pandemic, in addition to risks at the host site, risks of amplification or the dissemination of the pathogen to regions from which it was originally absent or close to it. MGEs have the potential to enable or favor the evolution and spread of novel variants of SARS-CoV-2 and other analogous pathogens (18).
To curtail these risks, host sites have at their disposal an arsenal of public health measures that they can used independently or concurrently. Such measures can act at three different stages: at entry, on-site and at exit. Exit controls are an efficacious way to disrupt the global spread of infectious pathogens (57). However, they are rarely explicitly used because the onus is then on the exit-screening country to treat the detected case. Regarding the world cup, visitors were returning to a large number of home locations, making the assessment of exit controls difficult. For these reasons exit controls were not included in our evaluation.
Instead, we focused our study on the role of the most common stages of control: entry and on-site controls. We used the example of the recent FIFA World Cup in Qatar to investigate the effect of different types of interventions, namely, vaccination, antigen and RT-PCR testing, with the testing taking place at different stages in a participant’s travel to or sojourn in the host location. We made the realistic assumption that travellers are vaccinated prior to their arrival in the host location. We focused the implementation of interventions on spectators and those staffing the MGE rather than on the whole population of the host location.
We found that pre-travel testing in the FIFA world cup has little effect on disease burden, potentially due to the pre-existence of community transmission and leakage of COVID-19 false negative visitors from abroad. Indeed, when community transmission is already taking place, the contribution of introduced cases is minimal (58). It can be inferred that in cases were disease is completely absent from the mass gathering site, pre-travel testing would prove beneficial, as has been observed in location which implemented a COVID-zero policy (59), but this was not evaluated here. We found that pre-match testing was more effective, particularly with RA testing a half day prior to the match. We found only marginal improvements in COVID-19 control if visitors had undergone a pre-travel RT-PCR test and match attendees had undergone a pre-match RA test.
We also investigated the relative roles of pre-match and pre-travel testing in comparison to requirements for visitors to be effectively vaccinated. We found that such a vaccination based policy generally outperformed testing regimes in controlling infections. When it came to reducing hospitalisation such a policy more consistently outperformed testing regimes and often to a much greater extent. As the background levels of effective vaccination amongst visitors decreased, the reduction in hospitalisation under various testing regimes paled in comparison to reductions under a requirement that all visitors be effectively vaccinated.
The state of Qatar decided to remove COVID-19 pre-travel testing and vaccination related travel restrictions for the period of the World Cup. Instead merely suggesting that all visitors in this period be fully vaccinated and up to date on their booster doses (17; 49; 50). Although, it should be mentioned that in order to access Qatari healthcare facilities visitors had to register their health status on the Ehetraz app (54). Fig 6 demonstrates that the number of COVID-19 cases and hospitalisations had been on a downward trajectory before the World Cup. An increase in the number of COVID-19 cases and hospitalisations starts towards the end of the group stages, peaking at the beginning of the quarter final stage of the tournament. Such an increase may support (19), who found that there was little effect on COVID-19 transmission associated with a nation hosting a UEFA 2020 match, but speculated that hosting an entire tournament such as FIFA 2022 could increase COVID-19 transmission. The increase in cases and hospitalisations is then followed by a decline, most probably reflecting less interest from certain fan-bases as their national side drops out of the tournament. Our work here would suggest that the State of Qatar’s removal of pre-travel testing may have been reasonable. However, the resulting spike in COVID-19 cases and hospitalisations may have been avoided with the enforcement of a policy requiring visitors to have had a second dose or a booster COVID-19 vaccination within a reasonable time-frame (e.g., 6 months to 14 days) prior to entry. Thus, ensuring COVID-19 vaccination among visitors was actually effective (26; 60; 61; 62).
The area between the yellow dotted lines is the time between the first world cup match and the final match. The area between the red dotted lines is the time between the last group stage match and the beginning of the quarter finals.
Daily Qatari data on the number of COVID-19 detections, hospitalisations and vaccinations differentiating between second and third (booster dose) required to fit our model is limited. The data-set from the (63) is missing data between 27-10-2021 and 29-6-2022, the data is patchy after 29-6-2022 and no record was made to indicate if a vaccine dose was a second or third booster. When it comes to Qatar ’Our World in Data’ (47) only lists COVID-19 case detections, missing the data on hospitalisations and vaccinations that were also required to fit our model. Furthermore, we did not have access to estimates of staffing (stadium or policing) and numbers of spectators for matches (including their composition by nationality) from the State of Qatar. Therefore, we chose to use a scenario analyses based on LHS. If the required data was available, our scenario analyses could have been based on a two stage approach. The first stage would have been to fit a single host cluster version of the model to Qatari COVID-19 detections, hospitalisations and vaccinations. The second stage would have been to use the parameters and variable estimates from the fitting in stochastic processes, such as τ -leap methods (64), to simulate the scenarios. The large number of LHS samples used for our scenario analyses rendered the use of such stochastic simulations impractical, at least with numbers of stochastic simulations large enough to control for the resulting aleatoric uncertainty. However, a fitting stage would have decreased the epistemic uncertainty, making such large numbers of stochastic simulations feasible.
A further limitation of our work is that we assumed isolation is as effective at reducing transmission from visitors as residents. Adding differential effects of isolation on transmission between hosts and visitors would have added more complexity to our model and increased the parameter space being sampled. A scenario with fewer clusters of people would present a more simplified and therefore ideal setting for assessing the effect differential group isolation on transmission, using our generalised model framework.
5 Conclusion
Our study demonstrates the feasibility of using modelling to assess disease control strategies at large MGEs, such as the FIFA World Cup 2022, in a time of COVID-19 and other pandemics. We find that requiring visitors to be effectively vaccinated is more effective than visitor pre-travel COVID-19 testing, and typically outperforms pre-event COVID-19 testing of attendees. Differing conclusions may be drawn if COVID-19 transmission was absent from the host nation (59). Therefore, the State of Qatar’s abandonment of pre-travel COVID-19 testing may have been reasonable. However, in light of the COVID-19 cases and hospitalisations seen over the world cup we conclude that pre-travel COVID-19 testing should have been replaced with required effective vaccination pre-entry. Put another way, all visitors should have completed a primary series of vaccination close to the time of entry or should have had a booster dose timed so as to ensure the fullest possible immunity.
Data Availability
Code used in simulations can be found at https://github.com/LIAM-COVID-19-Forecasting/Modelling-Disease-Mitigation-at-Mass-Gatherings-A-Case-Study-of-COVID-19-at-the-2022-FIFA-World-Cup.git
Supporting information
S1 Methods. Can be found at https://github.com/LIAM-COVID-19-Forecasting/Modelling-Disease-Mitigation-at-Mass-Gatherings-A-Case-Study-of-COVID-19-at-the-2022-FIFA-World-Cup.git
S2 Methods. Deviation of the Basic Reproductive Number, R0, and its relationship to the transmission term, β.
S1 Results. In depth examinations of % relative differences and PRCCs.
S1 Table. Differences in PRCCs between proportions of visitor clusters A and B effectively vaccinated (vA and vB) and different testing regimes. Differences are transformed to z-scores using methods outlined in (55).
Declaration of Potential Conflicts of Interest
Laurent Coudeville, Edward W. Thommes and Amine Amiche are Sanofi employees and may hold stock options within Sanofi.
Acknowledgments
There are multiple funders of this research to which we are grateful, as without their support this research would not have been possible. Martin Grunnill’s position was funded through the Fields Institute’s Mathematics for Public Health Next Generation program, grant number 72062654. Julien Arino is funded through the Discovery Grant program from the Natural Science and Engineering Research Council of Canada (NSERC), grant number RGPIN-2017-05466. L. Bourouiba’s work is supported, in part, by the US National Science Foundation (NSF). Ali Asgary is funded through the Advanced Disaster, Emergency and Rapid Response Simulation Initiative (ADERSIM), Ontario Research Fund 33270. Jianhong Wu’s work is also supported by the Canada Research Chairs program (230720), the Discovery Grant program from NSERC (105588), and the ADERSIM (Ontario Research Fund 33270). This work was supported by the NSERC-Sanofi Industrial Research Chair program in Vaccine Mathematics, Modelling, and Manufacturing (517504). These funding bodes had no role in the design of this study, its execution, the interpretation of the data or the decision to submit results.