key: cord-0891278-2k88ztfc
authors: Toumi, A.; Zhao, H.; Chhatwal, J.; Linas, B. P.; Ayer, T.
title: The Effect of NFL and NCAA Football Games on the Spread of COVID-19 in the United States: An Empirical Analysis
date: 2021-02-19
journal: nan
DOI: 10.1101/2021.02.15.21251745
sha: b3efb7b0e598378d6ad4c98a858244048d739bf7
doc_id: 891278
cord_uid: 2k88ztfc

Importance: In 2020 and early 2021, the National Football League (NFL) and National Collegiate Athletic Association (NCAA) had opted to host games in stadiums across the country. The in-person attendance of games has varied with time and from county to county. There is currently no evidence on whether limited in-person attendance of games has caused a substantial increase in coronavirus disease 2019 (COVID-19) cases. Objective: To assess whether NFL and NCAA football games with limited in-person attendance have contributed to a substantial increase in COVID-19 cases in the counties they were held. Design: In this time-series cross-sectional study, we matched every county hosting game(s) with in-person attendance (treated) in 2020 and 2021 with a county that has an identical game history for up to 14 days (control). We employed a standard matching method to further refine this matched set so that the treated and matched control counties have similar population size, non-pharmaceutical intervention(s) in place, and COVID-19 trends. We assessed the effect of hosting games with in-person attendance using a difference-in-difference estimator. Setting: U.S. counties. Participants: U.S. counties hosting NFL and/or NCAA games. Exposure: Hosting NFL and/or NCAA games. Main outcomes and measures: Estimating the impact of NFL and NCAA games with in-person attendance on new, reported COVID-19 cases per 100,000 residents at the county-level up to 14 days post-game. Results: The matching algorithm returned 361 matching sets of counties. The effect of in-person attendance at NFL and NCAA games on community COVID-19 spread is not significant as it did not surpass 5 new daily cases of COVID-19 per 100,000 residents on average. Conclusions and relevance: This time-series, cross-sectional matching study with a difference-in-differences design did not find an increase in COVID-19 cases per 100,000 residents in the counties where NFL and NCAA games were held with in-person attendance. Our study suggests that NFL and NCAA football games hosted with limited in-person attendance do not cause a significant increase in local COVID-19 cases.

The Centers for Disease Control and Prevention (CDC) has advised that mass in-person events have the potential for significant spread of coronavirus disease 2019 (COVID-19) [1] . To curb the spread of COVID-19, states have implemented non-pharmaceutical interventions (NPI) with varying intensity, including closure of work places, limit on indoor and outdoor gatherings, and travel restrictions [2] . Of note, sporting events have been banned or cancelled due to the potential of mass gatherings to become "super spreader" events [3] .

In early 2020, the National Basketball Association (NBA) and National Hockey League (NHL) temporarily suspended their 2019-20 seasons in an effort to limit the spread of COVID-19. A few months later, both leagues committed to resuming games in a "bubble" format where games were held in select sites with no fan attendance. The NBA suspended their season a second time due to a strike, while the NHL resumed their season with no interruption. In the late summer, the National Football League (NFL) and National Collegiate Athletic Association (NCAA) made the decision to play games in their respective stadiums, with many hosting in-person fans in a limited capacity. Despite various restrictions in placeon sporting and non-sporting events, -COVID-19 cases continued to increase nationally from 4.62 million in August 2020 to 13.84 million in December 2020; the increase in cases varied across counties and states [4] .

However, the effect of limited in-person attendance in football games on the spread of COVID-19 in the hosting counties is not well-understood. Despite the recent availability of COVID-19 vaccines, strict restrictions on holding large sporting events may remain in place until a herd immunity is achieved.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) Hence, quantification of the effect of in-person football games attendance on the spread of COVID-19 could inform appropriate management decisions in future.

The purpose of this study was to assess whether the historical NFL and NCAA football games with limited in-person attendance have contributed to a significant increase in COVID-19 cases in the counties they were held. We used a matching method and a difference-in-differences (DID) estimator suited for time-series, cross-sectional data to estimate the effect of in-person attendance on the spread of new, reported COVID-19 cases per 100,000 residents.

We extracted NFL and NCAA game information, COVID-19 case counts, and COVID-19 policies and interventions from public sources. We gathered information on NFL and NCAA games played in the 2020-2021 regular season, spanning 08/29/2020 to 12/28/2020 from Pro Football Reference [5] . Extracted data included game dates, in-person attendance (yes/no), and the stadiums where the games were held.

We obtained stadiums' longitude and latitude from Wikipedia [6] . Home stadiums' longitude and latitude were mapped to the corresponding U.S. county. We collected COVID-19 statewide non-pharmacological interventions and policies from the COVID-19 U.S. State Policy Database (CUSP) with respect to physical distance closures (closing non-essential businesses and closing restaurants except take-out), stayat-home orders (stay-at-home/shelter-in-place and end/relax stay-at-home/shelter-in-place), and second closures and reopening (closing bars, reopening bars, reopening restaurants, and reopening non-essential businesses) [7] . We obtained county-level populations from the 2019 U.S. Census Bureau Gazetteer Files [8] . We computed county-level new cases of COVID-19 per 100,000 residents using data from the COVID-19 Data Repository by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University [9] .

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 

We quantified the effect of interest by comparing daily changes in COVID-19 cases per 100,000 residents in counties that have held NFL/NCAA games with limited in-person attendance with those that did not hold NFL/NCAA games or have no attendance. This time series cross-sectional study used matching and a DID design to estimate the Average Treatment effect on the Treated (ATT) [10] .

The overall study design consisted of three major parts: (1) constructing initial matched set for each treated county based on history of games, (2) refinement of the matched sets based on additional control variables, and (3) ATT estimation using a DID estimator.

The treatment X !" is defined as whether county has had NFL and/or NCAA game(s) with in-person attendance on date . If there is a game with in-person attendance in county at date , X !" is set to be 1. If not, X !" is set to be 0. We define post-treatment period as the time period (days) following treatment, which we set to 14 days due to the incubation period of coronaviruses [11] . Similarly, we define pretreatment period to be 14 days. The outcome of interest was the change in new, reported daily COVID-19 cases per 100,000 residents from time to time + , that is from day 0, # , to day 14, $% .

For each treated county X !" , a set of matched counties were determined based on pre-treatment and posttreatment game history such that matched counties should have identical game history from time − to − 1. Note that is fixed for both treated and matching counties, so that with the same time trend, we can adjust for time-specific unobserved confounders later. Secondly, matched control units were excluded if they had a game with in-person attendance after time but before the outcome is measured at time + .

For each treated county !" with !" = 1 and !,"'$ = 0, the matched set of counties is defined as:

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.15.21251745 doi: medRxiv preprint ℳ !" = . Though county E also has the same history, it is not selected since it has treatment on 09/04. In the next matching step, county B is filtered out because it has a treatment within three days after 09/04. Hence, only county C is in the matched set of bolded blue . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.15.21251745 doi: medRxiv preprint treatment (with solid rectangles). In the next matching step, county B is filtered out because it has a treatment within three days after 09/04. Hence, only county C is in the matched set of bolded blue treatment (with solid rectangles). Similar matching procedures lead to no matched counties for bolded red treatment. In brief, counties are matched counties only if they first have dashed-dotted rectangles around past history and then have solid rectangles around future events.

In the initial matching step based on history of games played, other control variables that may play an important role in spread of the disease are ignored. However, in addition to the identical history of games, we also need to ensure that treatment and control groups are similar based on factors that may influence the disease epidemiology such as population size and presence of non-pharmaceutical interventions, as otherwise the required parallel trends assumption for DID would not hold. In this step, we refine the matching algorithm to account for such control variables and ensure the validity of parallel trends assumption.

Every treated county X !" has up to 10 control counties selected from the matched set ℳ !" with replacement. To this end, we use the average Mahalanobis measure to calculate the distance between treated county and each matched county over time adjusting for covariates. As specified in the data section above, covariates include COVID-19 statewide non-pharmacological interventions, policies, and county-level population. Formally, the average Mahalanobis distance over pre-treatment history in our study is defined as:

where ( ∈ ℳ !" is a matched county of X !" , COV !" is the time-varying covariates that we want to adjust for, and S .,/ is the covariance matrix of COV !" . In other words, given a control county in the matched set, . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.15.21251745 doi: medRxiv preprint we compute the standardized distance using the time-varying covariates, and average it over time. For each treated county, we choose control counties with top ten smallest Mahalanobis distance, if there are any, and assign equal weights to control counties in the refined matched set. Doing so implies that up to 10 most similar control counties under Mahalanobis distance are selected for the treated county:

($#) stands for the 10th smallest order statistic of !" ( ( ) among the original matched set

After the refined matched sets are obtained, we calculate the counterfactual outcome using the weighted average of control counties in ℳ !" * . For treated counties and observation days, the estimated average treatment effect on the treated (ATT) of the occurrence of a football game is defined as: Besides the time fixed effect, this model also accounts for county-level fixed effects as they are eliminated by the difference between < !,"65 − !,"'$ @ and < ! ! ,"65 − ! ! ,"'$ @. To compute the standard error of estimator I ( , ), we used a block-bootstrap procedure designed for matching with TSCS data by conditioning on the weight. Number of bootstrap iterations is set to 1000.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

We also performed sensitivity analyses by varying the pre-treatment period and post-treatment period. All preprocessing of data and statistical analyses were performed using R version 4.0.3 (R-Core Team, 2020).

This study was deemed exempt from review by the Institutional Review Board due to the use of publicly available data.

Out of 796 NFL and NCAA games played through 08/29/2020 to 12/28/2020, 528 games had in-person attendance. NCAA game attendance numbers are not available publicly. The median (IQR) attendance at NFL games through 08/29/2020 to 12/28/2020 was 11,133 (6,000 to 13,797) fans. Figure 2 shows the distribution of NFL attendance. Table 1 shows the estimated average treatment effect on the . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021.

and pre-treatment period are set to 14 days, along with standard error (SE) and 95% confidence interval. Table 1 and Figure 3 suggest that the effect of in-person attendance at NFL and NCAA games is not significant, as the average treatment effects are less than 5 daily new cases of COVID-19 per 100,000 residents. At 14 days after treatment, the average treatment effect estimate in daily new cases is 4.722 per 100,000 residents.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.15.21251745 doi: medRxiv preprint To evaluate the robustness of our results, we performed sensitivity analyses by varying the pre-treatment period and post-treatment period parameters ( , ) to (14, 7) , (14, 14) , (14, 7) and (21, 14) days. Across the different ( , ) parameter specifications, we found that the average treatment effects insignificant for the entire post-treatment period as they do not surpass 5 daily new cases of COVID-19 per 100,000 residents. Table 2 shows the estimated average treatment effect on the treated (ATT) with ( , ) set to (14, 21) , along with standard error (SE) and 95% confidence interval. 

In this time-series, cross-sectional study of U.S. counties with NFL and NCAA football games, we used matching and a difference-in-differences estimator to estimate the effect of games with in-person attendance on county-level COVID-19 spread compared to games held without any fans. The study considered the effect of both NFL and NCAA games since many counties had both an NFL or NCAA game in the pre-treatment or post-treatment periods, hence the effect of NFL games alone or NCAA games alone could not be studied separately. Games with in-person attendance were matched with a set of control counties that share the same game history from time − 14 to − 1 days. We selected 14 days based on numerous papers that found the incubation period for COVID-19 to be 2 to 14 days[11]. To control for post-treatment bias, we apply the idea of marginal structural models by excluding control counties that had a game with in-person attendance after time t but before the 14-day post-treatment period is complete. We found that the average treatment effects on the treated of in-person attendance at NFL and NCAA games on new, daily reported COVID-19 cases (per 100,000 residents) was not significant over a 14-day period as they did not surpass 5 new cases per 100,000 residents. We surmise that the NFL and NCAA policies with regards to limited in-person attendance, mask use and social plus physical distancing measures in stadiums does not lead to a significant increase in the community spread of COVID-19 [12, 13] . Additionally, an important number of NFL and NCAA football stadiums are . CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted February 19, 2021. ; https://doi.org/10.1101/2021.02.15.21251745 doi: medRxiv preprint outdoors or possess a retractable roof, which could've had an impact on mitigating spread [14] .

There are limitations to this study. The Stable Unit Treatment Value Assignment (SUTVA) assumption stipulates that each unit receives the same form or version of the treatment [15] . The treatment in this study, namely in-person attendance, was defined as a binary (yes or no fans) and therefore the results do not capture the possible impact of differing attendance numbers (Figure 2 ) on the outcomes. When the NCAA attendance numbers are made publicly available, additional sensitivity analyses can be performed to establish whether varying attendance numbers have an impact on COVID-19 spread. We also did not control for other large gathering events, such as political rallies, as some have been found to cause a local spread in COVID-19 cases [16] . We also did not account for the spillover effects to the counties adjacent to the ones hosting NFL/NCAA games. Finally, our analysis could not assess whether the Super Bowl LV game held on February 7 th , 2021 in Tampa, Florida is associated with an increase in the spread of COVID-19 cases because of post-game celebrations.

Our study provides new information on the effect of football games with in-person attendance on COVID-19 spread. This time-series, cross-sectional study with difference-in-differences design suggests that NFL and NCAA games held with limited in-person attendance do not cause an increase in COVID-19 cases in the counties they are held. Further research is needed to account for spillover effects to counties adjacent to the ones hosting games. While COVID-19 vaccination has started, restrictions will likely stay in place for several months and possibly until the start of the 2021-22 NFL and NCAA seasons. Our study provides evidence suggesting that in-person attendance of football games with social distancing and mask use could be resumed in the 2021-22 season.

. CC-BY-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted February 19, 2021. ;

Considerations for events and gatherings

The New York Times. See Coronavirus Restrictions and Mask Mandates for All 50

Mass gathering events and reducing further global spread of COVID-19: a political and public health dilemma

Coronavirus Pandemic (COVID-19)

Pro Football Reference. 2020 NFL Attendance Data

List of current National Football League stadiums

COVID-19 US State Policy Database

An interactive web-based dashboard to track COVID-19 in real time. The Lancet. Infectious diseases

2020. 11. Centers for Disease Control. Clinical Questions about COVID-19: Questions and Answers

How Many Fans Will Be At College Football Games This Year?

Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction

The Effects of Large Group Meetings on the Spread of COVID-19: The Case of Trump Rallies