Injuries/Violence
Causal inference for repeated events, case-crossover or case-control: application to sport injuries George Stefan* George Stefan Stefan Stefan Stefan Stefan Stefan University of Toronto / The Hospital for Sick Children
Recurrent events such as injuries are commonly encountered in health research but often not properly evaluated. We were interested in evaluating the ability to correctly estimate the causal effect of bye (rest) weeks in professional football using two commonly used designs—case-crossover and case-control.
Using simulations, we generated data that mimic real data from the Canadian Football League. We simulated eight teams with approximately 45 players each, competing over 14 weeks across 10 seasons. Based on current theory, we set the intervention—whether the week preceding the game week was a bye week—as well as season, players’ age, team, game order, and opposing team to be associated with individual player injury risk. We accounted for within-team correlation of players and within-player temporal correlation by using a nested random effects structure.
In the case-crossover setting, we matched (1) each individual to themselves and (2) the injury game with the most recent previous injury-free game. In the matched case-control setting, we matched players with an injury to players without an injury playing in the same game on the opposing team. As a secondary analysis, every case was randomly assigned to controls from any games. We estimated the effect of bye week on the presence of injury from conditional logistic regression models.
We found that the case-crossover and simple case-control designs recovered the true effect with minimal bias and coverage close to the nominal level, while the matched case-control design was heavily biased. When consecutive injury games occurred, case-crossover exhibited less bias when only using the first injury game in the series as opposed to all injuries, whereas the opposite was observed for the simple case-control design.
