Bayesian Methods For Estimating Injury Rates In Sports Injury Epidemiology

PURPOSE: The injury rate is commonly used as a measure of occurrence in sports injury surveillance and involves expressing injury incidence using both population at risk and exposure-time. Traditional methods of calculating injury rates employ a Frequentist estimation approach. We conceptualize the injury rate as a data generation process and offer practical approaches for estimating this parameter under a Bayesian inferential framework. METHODS: Using injury surveillance data captured within the National Collegiate Athletic Association Injury Surveillance Program during 2014/15-2018/19, we assessed the suitability of various count data models for injury rate estimation under a Bayesian framework. Models considered included standard Poisson, Negative Binomial, and zero-inflated Negative Binomial formulations, and we constructed models to estimate injury rates (for 23 sports separately) as a function of event type, competition level, injury diagnosis, academic year, and reporting school. Relationships between the injury rate and each explanatory variable represented parameters of interest, and prior distributions were specified for each parameter. Herein, we walk through our workflow including how we assessed model fit, and review the advantages offered by this analytical approach. RESULTS: We found that the following posterior predictive checks were most useful for assessing posterior distributions from our fitted models: plots of observed and fitted injury rates, plots of % zero injury cases in observed and fitted data, and probability integral transforms. Through an iterative process, we found that a zero-inflated Negative Binomial formulation was most appropriate in modeling surveillance-based injury rates. We also found that school-level information was a key factor to consider in modeling injury rates. CONCLUSIONS: Our findings indicate that the Bayesian framework offers natural characterizations of the injury rate, which complement traditional perspectives on estimating this measure. Key benefits of the Bayesian approach in this context seem to be the ability to test model suitability in a variety of creative ways, as well as the ability to generate plausible estimates with sparse data.