INTRODUCTION: Training load terminology can be confusing. Parameters like the "normalized distance" are commonly used to describe the load profile in sports games. The interpretation of such parameters is not always straightforward and depends on the question asked and the data processing. Causal effect estimates are often the target, but this is even harder for observational studies (e.g., analysis of tournament data) compared to experiments. In handball, research interest was placed on positional differences and decreasing intensity over time. Our aim is to highlight methodological pitfalls when interpreting observational data from intermittent sports games and show techniques to overcome t METHODS: The data in this example were obtained from the European Handball Championship 2020 and recorded with the Kinexon Local Positioning System (LPS). The raw data were processed to obtain summary statistics for each player in each game, which were used for statistical modelling. A three-step procedure is used to interpret the causal effects of playing position, intensity and time played: First, I discuss a simple definition of volume, time and intensity. Second, I am going to identify the underlying causal effects using a graphical causal model. Third, I discuss and compare different models to estimate causal effects for positional differences and decreasing intensity over time. RESULTS: The terminology of "time-normalized distance" conceals the correct interpretation: It`s the average velocity and, thus, a measure of intensity by definition. A mediation model guides the regression analysis: To identify the direct effects of intensity change over time and positional differences in intensity, we must include position and time in the same regression model. Thus, the direct effect for time yields an intensity decrease of 0.21 W/kg [CI95% [0.17; 0.25] per 10 minutes played. In the same model, positional differences were observed between wings, mid backs, left and right backs and pivots. The direct effects differed substantially from the total effect, e.g. mid backs had lower intensity estimates as direct effect (8.17 CI95% [7.93; 8.41] vs. 7.85 CI95% [7.68; 8.03]). Importantly, modelling individual observations as random effect yields larger confidence intervals compared to a simple factorial (ANOVA) design, e.g. the mean standard error was 39% higher. The random slope/intercept model outperformed other models in fit parameters (information criteria, p-values). CONCLUSION: Causal identification strategies should generally be used when interpreting observational data from sports games. Current practice is to present simple means. This can be misleading and marginalized effects should be presented according to the assumed causal model structure. Violation of the independence assumption may lead to narrower standard errors and might thus increase the risk of false-positive results. Also, visualizing raw data and model predictions should be a standard rather than an exception.
© Copyright 2022 27th Annual Congress of the European College of Sport Science (ECSS), Sevilla, 30. Aug - 2. Sep 2022. Julkaistu Tekijä Faculty of Sport Science - Universidad Pablo de Olavide. Kaikki oikeudet pidätetään. / Kääntänyt https://www.DeepL.com/Translator