• epidemiology
• recurrent
• injury
• statistics

In any research study, variables may be associated due to either ‘cause and effect’ or alternative reasons that are not causal. While all causal relationships are associational, not all associational relationships are causal, that is, correlation does not equal causation. Sports medicine clinicians are generally interested in causal relationships because they want to know whether an intervention will prevent, treat or rehabilitate injury. However, many sport medicine studies do not distinguish between causal and non-causal associations, although they may present their findings in a manner that implies causation. The purpose of this editorial is to help clinicians distinguish causal and non-causal associations to avoid faulty conclusions and misguided clinical decisions.

Generally, in a well-conducted randomised trial with a sufficient sample size, high adherence and minimal dropouts, one can assume that the change in the outcome was caused by the treatment. However, in observational studies, there are often additional hidden assumptions to be acknowledged before concluding the relationship is causal; for example, the absence of confounding or collider stratification bias. Causal diagrams such as those in figure 1 make many of these assumptions explicit.1 In addition, when random study results are absurd, such as an association between a sports team logo and concussion risk,2 a causal diagram forces researchers to either explain the association or to label it as a chance finding. Shrier and Steele provide an introductory overview of causal diagrams with reference to sport medicine.1

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

The four general ways that variables A and B can be associated. Association simply means that knowing the value of one variable provides information on the other variable. Diagrams I and II are causal relationships. Diagrams III and IV are non-causal relationships.

Figure 1 outlines the four general ways that two variables (A and B) can be associated. When changing the value of A increases the value of B, then A is a cause of B (figure 1-I). And, when changing the value of B increases A, then B is a cause of A (figure 1-II). The problem is that variables A and B can also be associated due to non-causal reasons.

Consider figure 1-III and a sport medicine study of the relationship between injuries  (variable A) and penalties (variable B). Variable C signifies ‘aggressive play,’ and this could be a confounding variable as it is a cause of both injuries and referees calling more penalties. Thus, injuries and penalties could be positively associated without a causal relationship. If so, then instituting an injury prevention programme that decreases injuries would be unlikely to result in fewer penalties, and instituting a programme where referees would call fewer penalties would not result in fewer injuries. However, an intervention that decreases aggressive play would decrease both injuries and penalties.

There is a second type of non-causal association shown in figure 1-IV, which often goes unrecognised. If variables A and B both cause something (C), they remain unassociated if the value of C is unknown. However, if the value of C is controlled, for example, through adjustment in a multiple regression analysis or through restricting the study to only those with C, then A and B become associated via a non-causal relationship. This phenomenon is known as ‘collider stratification bias’.3 This type of bias is seen in the sports medicine literature with reports of the so-called ‘obesity paradox’, which occurs when one restricts studies to only those with an obesity-related comorbidity such as heart failure. In these studies, obese subjects have better outcomes than non-obese subjects.4 Interpreting these results to be causal (ie, that obesity is the cause of better outcomes) is incorrect as the collider stratification bias creates an artificial and non-causal association. Unfortunately, some are now questioning the recommendation for weight loss in obese patients with obesity-related heart failure,5 even though the better outcomes in the obese patients are consistent with a non-causal association.

The subject of causal versus non-causal associations is especially complex when discussing issues such as recurrent injury, as in the studies summarised in the systematic review of observational studies by Toohey et al in this edition of BJSM.6 The studies reviewed by Toohey et al generally compared individuals with and without previous injury and concluded that previous injury increased the risk for subsequent injury. However, the analyses in the studies have two major assumptions that are unlikely to be true. First, with recurrent injuries, one expects different individuals to have different inherent injury risks that may cause both the first and the subsequent injury (eg, player frailty, different positions or styles of play).7 This type of non-causal association is outlined in figure 1-III. When Hamilton et al 7 used a more appropriate analysis to estimate causal effects in rehabilitated circus artists, previous injury did not causally increase the risk of subsequent injury. Second, there is an assumption that injury risk does not change over time.8 This assumption appeared reasonable in the population of circus artists studied by Hamilton et al 7 but is unlikely to be true over the course of a sports season where training intensity, fitness and fatigue may vary.

Conflating causal and non-causal associations is a problem in many studies and we use the review by Toohey et al merely for pedagogical reasons. First, the stated purpose and conclusion of the article was to guide injury prevention.6 This immediately leads the reader to believe the review is trying to evaluate causal relationships even though this is not possible from the literature reviewed. Second, there is ambiguous phrasing. For example, in the abstract and text, the authors say, ‘Previous history of an ACL injury was associated with an increased risk of subsequent hamstring injury…’. The word, ‘associated’ is appropriate because it includes both causal and non-causal relationships. However, ‘increased risk’ is likely to be interpreted as a ‘cause’ because if A increases the risk of B, the implication is that A causes B. We suggest wording such as ‘previous history of an ACL injury was positively associated with subsequent hamstring injury’ as it is accurate and less likely to be interpreted as a causal relationship.

In conclusion, distinguishing between causal and non-causal associations is essential for clinicians, since prevention and treatment programmes should be targeted at causal associations. When findings from original research can be explained by non-causal associations, authors should use language that is as unambiguous as possible rather than implying causation. Analysis of observational studies requires making assumptions about relationships between variables. We encourage authors to include a causal diagram in manuscripts evaluating sports injury aetiology, prevention or treatment to help the reader recognise and understand these underlying assumptions.