key: cord-1030134-f1ihexts
authors: Baugh, Christine M.; Kroshus, Emily; Meehan, William P.; McGuire, Thomas G.; Hatfield, Laura A.
title: Accuracy of US College Football Players’ Estimates of Their Risk of Concussion or Injury
date: 2020-12-29
journal: JAMA Netw Open
DOI: 10.1001/jamanetworkopen.2020.31509
sha: 1e7684bf282ab2616711ea8a91139f904694bd2d
doc_id: 1030134
cord_uid: f1ihexts

IMPORTANCE: Despite increased concern about the health consequences of contact sports, little is known about athletes’ understanding of their own risk of sports-related injury. OBJECTIVE: To assess whether college football players accurately estimate their risk of concussion and nonconcussion injury and to identify characteristics of athletes who misestimate their injury risk. DESIGN, SETTING, AND PARTICIPANTS: In this survey study, questionnaires were given to 296 current college football players on 4 teams from the 3 of the 5 most competitive conferences of the US National Collegiate Athletic Association. Surveys were conducted between February and May 2017. Data were analyzed from June 2017 through July 2020. MAIN OUTCOMES AND MEASURES: Multiple approaches were taken to compare athlete perceptions of their risks of concussion and nonconcussion injury with individual probabilities of these risks, which were modeled using logistic regression. RESULTS: Of 296 male college-aged athletes from 4 football teams who participated in the survey, 265 (89%) answered all questions relevant for this study. Participating teams were similar to nonparticipating teams across nearly all measured characteristics. One hundred athletes (34%) had sustained 1 or more concussions, and 197 (68% of the 289 who responded to the question) had sustained 1 or more injuries in the previous football season. Logistic regression models of single-season injury and concussion had reasonably good fit (area under the curve, 0.75 and 0.73, respectively). Of the 265 participants for whom all relevant data were available, 111 (42%) underestimated their risk of concussion (χ(2) = 98.6; P = .003). A similar proportion of athletes (113 [43%]) underestimated their risk of injury, although this was not statistically significant (χ(2) = 34.0; P = .09). An alternative analytic strategy suggested that 241 athletes (91%) underestimated their risk of injury (Wilcoxon statistic, 7865; P < .001) and 167 (63%) underestimated their risk of concussion (Wilcoxon statistic, 26 768; P < .001). CONCLUSIONS AND RELEVANCE: The findings of this survey study suggest that college football players may underestimate their risk of injury and concussion. The implications for informed participation in sport are unclear given that people generally underestimate health risks. It is necessary to consider whether athletes are sufficiently informed and how much risk is acceptable for an athlete to participate in a sport.

We used two different methods to transform between modelled probabilities (continuous values between 0 and 1) and athlete perceptions (measured on a 7-point Likert scale).

First, we used existing literature 1 on people's qualitative interpretations of probabilities to transform athlete perceptions into numerical probabilities: Definitely won't=0, Very Unlikely=0.05, Unlikely=0.20, Middle Category=0.50, Likely=0.70, Very Likely=0.90, Definitely Will=1. Going in the other direction, we transformed modeled probabilities into seven ordinal categories using the following cut points: (0, 0.025, 0.075, 0.35, 0.6, 0.8, 0.95, 1).

Second, we used a transformation that minimized the differences between the modeled values and athletes' perceptions. We searched over the space of cut points to find one that minimized the sum of the absolute distance between the categorized probabilities and athlete perceptions. For a set of cut points, each modeled probability can be transformed into one of 7 categories, represented as an integer: Definitely won't=1, Very Unlikely=2, Unlikely=3, Middle Category=4, Likely=5, Very Likely=6, Definitely Will=7. We computed the absolute distance between the integer representation of the athlete's actual Likert response and the integer category of the transformed modelled probability. Summing these absolute differences over athletes, we get a score for how closely the categorized probabilities match the athlete ratings. Minimizing this score produced data-driven cut-points for injury (0.09, 0.24, 0.81, 0.93, 0.98, 0.99) and concussion (0.06, 0.11, 0.49, 0.8, 0.87, 0.93). In addition, we transformed athlete perceptions into numerical values using the midpoints of each category. 

A psychometric study of adolescent risk perception