In the following video I am reviewing very interesting article by Windt, Gabbett et al.: “Training load–injury paradox: is greater preseason participation associated with lower in-season injury risk in elite rugby league players?”. Tim Gabbett is leading researcher in the field of training load and injuries, and this paper represents continuation of his work on this topic. All papers (co)authored by Tim are highly recommended reading for coaches in general and sports scientists in particular.
The purpose of this video is to provide some of critiques and recommendations for future authors of similar studies, reviewer and journal editors. Please bear in mind that I am not an expert in this field nor expert in statistics/machine learning, so take my opinions with grain of salt.
Some of my main critiques are the following:
- Predictive vs. retrodictive models (no hold-out data sets, no cross-validation)
- Lack of any reporting of the overall model fit (deviance, AIC, BIC, WAIC, ROC)
- Lack of tuning of the model parameters: ACT, CTL and injury lag, as well as no description/rationale of using weekly binning vs. rolling windows
- Pooled injuries and not differentiation between types and locations – for example building multiple models for estimating likelihood of hamstring strain, groin strain and so forth
- No reporting of model comparison (univariate vs. multivariate). No comparison/use of model models such as random forest, LASSO, ElasticNet, SVM and so forth
- No discussion regarding practical significance of predictors association (statistical significance is not practical significance). Plus, hard to judge coefficients from logistic regression intuitively and without intercept
- More counterfactual graphs needed
- No discussion of random effects
- No code or data available – the need for reproducible research
- And much more