PURPOSE: To evaluate common modelling strategies in training load and injury risk research when modelling continuous variables and interpreting continuous risk estimates; and present improved modelling strategies.
METHOD: Workload data were pooled from Australian football (n=2,550) and soccer (n=23,742) populations to create a representative sample of acute:chronic workload ratio observations for team sports. Injuries were simulated in the data using three pre-defined risk profiles (U-shaped, flat and S-shaped). One-hundred datasets were simulated with sample sizes of 1000 and 5000 observations. Discrete modelling methods were compared to continuous methods (spline regression and fractional polynomials) for their ability to fit the defined risk profiles. Models were evaluated using measures of discrimination (area under ROC curve) and calibration (Brier score, logarithmic scoring).
RESULTS: Discrete models were inferior to continuous methods for fitting the true injury risk profiles in the data. Discrete methods had higher false discovery rates (16-21%) than continuous methods (3-7%). Evaluating models using the area under the receiver operator characteristic (ROC) curve incorrectly identified discrete models as superior in over 30% of simulations. Brier and logarithmic scoring was more suited to assessing model performance with less than 6% discrete model selection rate.
CONCLUSIONS: Many studies on the relationship between training loads and injury that have used regression modelling have significant limitations due to improper discretization of continuous variables and risk estimates. Continuous methods are more suited to modelling the relationship between training load and injury. Comparing injury risk models using ROC curves can lead to inferior model selection. Measures of calibration are more informative when judging the utility of injury risk models.