Time-to-event analysis for sports injury research part 2: time-varying outcomes

Abstract

BackgroundTime-to-event modelling is underutilised in sports injury research. Still, sports injury researchers have been encouraged to consider time-to-event analyses as a powerful alternative to other statistical methods. Therefore, it is important to shed light on statistical approaches suitable for analysing training load related key-questions within the sports injury domain.ContentIn the present article, we illuminate: (i) the possibilities of including time-varying outcomes in time-to-event analyses, (ii) how to deal with a situation where different types of sports injuries are included in the analyses (ie, competing risks), and (iii) how to deal with the situation where multiple subsequent injuries occur in the same athlete.ConclusionTime-to-event analyses can handle time-varying outcomes, competing risk and multiple subsequent injuries. Although powerful, time-to-event has important requirements: researchers are encouraged to carefully consider prior to any data collection that five injuries per exposure state or transition is needed to avoid conducting statistical analyses on time-to-event data leading to biased results. This requirement becomes particularly difficult to accommodate when a stratified analysis is required as the number of variables increases exponentially for each additional strata included. In future sports injury research, we need stratified analyses if the target of our research is to respond to the question: ‘how much change in training load is too much before injury is sustained, among athletes with different characteristics?’ Responding to this question using multiple time-varying exposures (and outcomes) requires millions of injuries. This should not be a barrier for future research, but collaborations across borders to collecting the amount of data needed seems to be an important step forward.

Improved reporting of overuse injuries and health problems in sport: an update of the Oslo Sport Trauma Research Center questionnaires

“…In effect, therefore, the severity score represents an ordinal-scale variable with 25 possible outcomes, not 100. Recent publications highlight the analytical benefits of representing various ‘states’ of an athlete’s health on an ordinal scale 16. Unfortunately, for this approach to be feasible with small samples (as is normally the case in sports medicine research), the number of potential states needs to be far fewer than 25 to reduce the risk of sparse data bias 17.…”

“…Such analyses are of special interest for elite athletic populations that, in general, are smaller in number but tend to experience multiple events (injuries and/or illnesses). When considering longitudinal analytical methods, users should be aware of the inherent challenges of these methods, including (i) missing data, (ii) time-varying exposures, outcomes, confounders, effect-measure modifiers and mediators, (iii) recurrent/subsequent events and (iv) competing risks 16 18 19…”

Injuries and illnesses in Swedish Paralympic athletes—A 52‐week prospective study of incidence and risk factors

“…20 Survival analyses using the Kaplan-Meier method were conducted to estimate the cumulative survival probability (SP) and the primary endpoints: median time to the first injury and illness, respectively. 22 Log-rank tests assessing the hazard function were used to compare differences in survival times between the subgroups, and Cox proportional hazard regression with corresponding hazard ratios (HR) were performed to analyze the actual risk of sustaining a first SRIIPS. 18 Univariate models assessing risks associated with each variable were first tested.…”

“…The median number of reported illnesses per athlete was 2 (IQR 1-4, min-max 0-11) (Table 4). Vertebral column 40 (22) 27 (25) 5 (19) 6 (35) 5 4611 (20) 11 (18) 2 (33) 17 (22) 24 (…”

The Garmin-RUNSAFE Running Health Study on the aetiology of running-related injuries: rationale and design of an 18-month prospective cohort study including runners worldwide

“…This approach was introduced based on experiences using the Yamato et al 52 consensus definition and the Oslo Trauma Research Center questionnaire 53. Importantly, the injury outcome should be considered as a time-varying covariate as described in Nielsen et al 54…”