Table 1

Key questions and associated key points that are covered in the article

Question 1
How to deal with time-varying outcomes?
Key question 1: a researcher collected data on sports injury status (in statistical terms: states) over time in a group of athletes to investigate the aetiology of Achilles tendinopathy. In weekly self-assessments, the athletes classified their injury severity into no Achilles injury, moderate Achilles tendinopathy and severe Achilles tendinopathy. The next step for the researcher is to analyse the data. Here, the question remains: is time-to-event a suitable analytical approach to deal with a time-varying outcome?
Key point 1: Time-to-event models allow for the inclusion of time-varying outcomes using the concept of multistate transitions. To date, there is no universally accepted way to classify sports injury into different outcome states. Sports injury researchers have the opportunity to use certain injury definitions, and have a degree of flexibility to choose the cut-offs that separate each injury state.
Question 2
How to deal withsubsequent injuries?
Key question 2: imagine a researcher having collected data on subsequent injuries (eg, athletes that sustained Achilles tendinopathy three times during the follow-up). The next step for the researcher is to analyse such data. Are there certain analytical approaches needed to deal with this type of data?
Key point 2: in time-to-event modelling, the researcher can consider subsequent injuries using the concept of shared frailty. This allows for correction for selection of ‘less-injury-prone’ athletes over time.
Question 3
How to deal withcompeting risk?
Key question 3: in your dataset, there are data on many different injury types (eg, Achilles tendinopathy, patella-femoral pain, iliotibial band syndrome, patellar tendinopathy). However, you may only be interested in studying Achilles tendinopathy. Should you just omit all other injuries (patella-femoral pain, iliotibial band syndrome, patellar tendinopathy) when analysing the data?
Key point 3: researchers should ‘stick to this world’ by including many injury types into the analysis using a competing risk setup. Excluding injuries of less interest is strongly discouraged as it will generate misleading results because the injury risk is overestimated.
Question 4
How to deal withassumptions and requirements?
Key question 4: you may speculate: What are the downsides of time-to-event modelling?
Key point 4: sports injury researchers need to calculate the event/variable ratio to avoid biased results. In addition, sports injury researchers should ensure there are at least five injuries in each exposure state to be analysed. Dealing with (multiple) time-varying exposures requires a considerable number of injuries to avoid violating the requirements underpinning the time-to-event analysis. Analysing data without consideration to number of injuries in each exposure state will easily lead to sparse data bias.
Question 5
Are there any considerations when designing my study?
Key question 5: I want to design a new study looking into the association between changes in training load and sports injury. What should I consider when I am designing my data collection?
Key point 5: researcher must consider: am I able to get the number of injuries needed in order to analyse changes in training load as a time-varying exposure to sports injury? How many injuries are likely to occur in each exposure state (or transition)? How many cut-offs to separate the exposure groups are suitable?
Question 6
Are there any alternative methods?
Key question 6: it is difficult to collect the amount of data needed to avoid violating the assumptions and requirements needed to perform a robust time-to-event analyses on a change in training load-related question. Accordingly, are there any alternative methods that could be considered?
Key point 6: the use of computational modelling could be considered as a complementary and alternative approach to time-to-event modelling in future sports injury research applications because no consideration to number of injuries is needed. However, unlike traditional statistical modelling, the assumptions underpinning computational models are often based on subject matter knowledge and other various forms of empirical evidence. If these are wrong, the results from the analyses will be questionable.