A study published in PLOS ONE and accepted in August has examined the relationship between training behaviour and running-related injury among performance-oriented parkrunners, using both conventional statistical methods and machine learning. Participants completed an online survey covering personal characteristics, training habits and any injuries sustained over the previous 12 months. The headline result runs against a good deal of received wisdom: participation in strength and conditioning was not associated with a lower likelihood of injury.
The associations that did emerge pointed the other way. Resistance training and stretching or yoga were both linked to a greater likelihood of reporting a running-related injury. The machine learning classifiers, working across the same data, produced a different emphasis again, favouring a high frequency of interval sessions, the exclusion of stretching and yoga, and having training prescribed by a qualified coach.
It is worth being precise about what this design can and cannot show. The study is cross-sectional and retrospective, and injury status is self-reported over a year in arrears. That opens a wide door to reverse causation: runners who pick up an injury are the runners most likely to be sent to a physiotherapist and to come back doing resistance work and mobility drills. A survey taken afterwards records the strength training and the injury together, and cannot tell which came first.
That caveat matters because the trial evidence points in the opposite direction. Randomised and prospective work on structured strength training in runners has generally found a protective effect, and the practical guidance that has followed from it — two sessions a week, progressively loaded — is not overturned by an association drawn from a questionnaire. What the new study does usefully is show how easily observational data on injury can produce a counter-intuitive result, and how differently a machine learning model and a regression can read the same responses.
The more durable contribution is methodological. parkrun offers a sampling frame of a size and consistency that running research rarely gets near, with hundreds of thousands of participants a week across more than 2,000 events, and the authors are explicit that injury is a threat to continued participation rather than an abstract clinical outcome. The obvious next step is a prospective cohort drawn from the same population, tracking training and injury forward in time. Until that exists, the sensible reading is that this study raises a question about how strength work is prescribed and sequenced, not an argument for abandoning it.