• Epidemiology
• Sporting injuries
• Statistical review

Finch and Cook1 recently requested feedback on their proposed classification for reinjuries. We applaud them for identifying key challenging areas, and summarising standard statistical approaches to time-to-event data.2–5 Developed within a Bayesian framework, their approach can be applied equally well within a frequentist framework. However, we do not believe their claims that the proposed new classification solves the current analytical challenges. Although we cannot offer solutions at the present time, we do believe the previous classification systems6 ,7 mentioned by Finch and Cook should remain the focus of analytical strategies at this time.

Previous reinjury classifications focused on the previous injury's healing state. If not healed, subsequent injuries are exacerbations. If healed, subsequent injuries are classified according to anatomic location. The proposed classification simply splits each of the previous classification categories6 ,7 into two groups: ‘related to index injury’ and ‘not related to index injury’. The same relationships could be modelled using the previous classifications by simply adding a dichotomous variable (yes/no) called ‘Relation to Index Injury’ in the categorisation scheme. Is including such a variable beneficial?

The proposed approach will likely underestimate the total causal effect of the index injury, and can lead to lost opportunities for injury prevention strategies. Consider a study on drugs, with the outcome as adverse events. Finch and Cook appear to focus on adverse events believed to be related to exposure (index injury), versus all adverse events. Such focus means we lose the ability to detect adverse events (and their mechanisms) that were not previously known. Further, if such an approach is used, those making the decision must be blinded to exposure status in order to avoid the obvious potential bias. In the authors’ example, the index injury is a fully healed fractured leg that occurred due to a blow. The subsequent injury is a second fracture after the index injury is fully healed. According to the authors, ‘this clearly fits into Category 6’ (an injury to the same body site and nature, ie, not related to the previous injury); the authors make no suggestion to verify if their categorisation of the injury is correct. What if proprioception unknowingly remained impaired after the fracture healed, and the blow in the subsequent injury occurred because the athlete lost his/her balance (ie, the index injury caused the subsequent injury)? Previous classification systems would explore statistical associations between: (1) the index injury and (2) all subsequent injuries, and reveal that these two injuries were related. In the proposed new classification, the true causal relationships would never be identified because a priori, there was no relationship.

Although not recommended by Cook and Finch, investigators could conduct analyses to verify the accuracy of category 6 injuries. A relation between the index injury and category 6 injuries leaves only two options. First, investigators could conclude that category 6 injuries (subsequent injuries unrelated to the index injury) are related to index injuries, which is internally inconsistent because by definition, the two are unrelated. Second, investigators could review and change the classifications of each injury so that some category 6 injuries would now be considered related to previous injury (categories 2–4), and their categories changed. These steps and the analysis would be repeated until category 6 was unrelated to previous injury. Such a process leads to a predefined conclusion, which is generally not a helpful process.

In addition to ‘related to injury’ status, Finch and Cook suggest that previous classification systems were ambiguous with respect to their categories 3 and 4 injuries. First, category 3 injuries clearly affect non-healed tissue, and represent exacerbations in the previous classifications cited.6 ,7 Second, a footnote says category 4 injuries refer to ‘overuse injuries with no acute onset of symptoms’. Previous classifications treat overuse and acute injuries similarly—they represent exacerbations for non-healed index injuries, and recurrent injuries for healed index injuries. In essence, category 4 incorporates a separate variable (mechanism of injury) into the classification, similar to the incorporation of ‘Relation to Injury’. We agree that analysing mechanism of injury may be beneficial for some research questions, and again, this is best incorporated into the categorisation scheme as a separate variable. Such an approach would have clearly and appropriately applied the mechanism of injury variable to all categories of subsequent injuries (mechanism of injury is not included for Finch's categories 7–10). In summary, their proposed classification combines type of reinjuries (four levels according to Hamilton classification),6 relation to index injury (two levels: related or unrelated) and mechanism of injury (two levels: acute or overuse), resulting in 17 categories (4×2×2 plus ‘no injury’). Separating mechanism of injury into more levels (eg, acute contact, acute non-contact and overuse) increases the number further.

In summary, statistical analyses help determine associations between variables. For example, in survival analysis, the base variables are time at first exposure (index injury), time to follow-up, event of interest (yes/no), competing event (yes/no) and censoring event (yes/no). In previous classification systems,6 ,7 these variables are kept separate, and then used for the analysis. In the proposed classification system, the variables for competing event and censoring event are combined, and then have to be reclassified later. It is not clear how this provides additional advantages. In addition, incorporating an a priori assessment of whether an event is related to the index injury is likely to lead to biased estimation of total causal effects, and requires assessors making subjective assessments be blinded to the research question.