"Outlier detection: Underused in sport science despite outliers' impact on inference and prediction".

Abstract

Outlier detection is critical in statistical analysis and predictive modelling, but it is often overlooked in research, leading to potentially inaccurate conclusions. This study aimed to (1) assess the prevalence of outlier detection strategies in sport science publications, (2) examine the effect of outliers on statistical inference using general linear mixed-effects models with longitudinal data, and (3) evaluate the impact of outliers on regression predictive models using the same dataset. A systematic literature search of 4,622 articles published in 2023 in Q1 journals in "sport science" category found that only 9.05% (99% CI: 4.87%-14.99%) reported employing outlier detection practices. A comprehensive outlier detection framework was designed to enhance data quality before formal statistical analysis using general(ised) linear mixed-effects models. The framework was applied to publicly available longitudinal data from rugby union small-sided games. Inclusion of outliers resulted in inaccurate maximum likelihood estimates, inflated standard errors, and biased conclusions regarding model parameters. Removing error outliers significantly improved the accuracy of predictive models, as evidenced by reduced root mean square errors. These findings underscore the importance of outlier detection in sport science research and demonstrate that appropriate handling of outliers enhances the validity of statistical inferences and predictions.