The applied impact of 'naïve' statistical modelling of clustered observations of motion data in injury biomechanics research.

OBJECTIVES: Appropriate statistical analysis of clustered data necessitates accounting for within-participant effects to ensure results are repeatable and translatable to real-world applications. This study aimed to compare statistical output and injury risk interpretation differences from two statistical regression models built from a clinical movement sidestepping database. A "naïve" regression model, which does not account for within-participant effects, was compared with an appropriately applied mixed effects model.

DESIGN: Comparative study.

METHODS: Three-dimensional unplanned sidestepping joint angle data (trunk, hip, and knee) from 35 males (112 observations) were used to model peak knee valgus moments and anterior cruciate ligament injury risk during the impact phase of stance. Both statistical models were cross-validated using a k-fold analysis.

RESULTS: The naïve regression returned inflated goodness of fit statistics (R2 =0.50), which was evident following cross-validation (predicted R2 =0.43). Following cross-validation, the mixed effects model (predicted R2 =0.40) explained a similar amount of variance, despite containing three less predictors. The naïve model produced inaccurate parameter estimates, overestimating the effects of certain kinematic parameters by as much as 79 %.

CONCLUSIONS: A regression model naïvely applied to clustered observations of sidestepping data resulted in erroneous parameter estimates and goodness of fit statistics which have the potential to mislead future research and real-world applications. It is important for sport and clinical scientists to use statistically appropriate mixed effects models when modelling clustered motion capture data for injury biomechanics research to protect the translatability of the findings.