Transfering Targeted Maximum Likelihood Estimation for Causal Inference into Sports Science

Area of Science:

• Sports Science
• Causal Inference
• Statistical Modeling

Background:

• Causal inference is valuable in various fields but underutilized in sports science.
• Targeted Maximum Likelihood Estimation (TMLE) is an advanced causal inference method known for its robustness to model misspecification and integration with machine learning.
• Generalized Linear Models (GLM) are commonly used but can be sensitive to model assumptions.

Purpose of the Study:

• To introduce TMLE and a roadmap for its application in sports science.
• To compare the performance of TMLE against GLM for estimating effect sizes in sports science.
• To investigate the influence of player substitutions on the total physical performance of soccer teams using causal inference.

Main Methods:

• Development of a causal model and a misspecified causal model.
• Utilizing simulation datasets and an observed tracking dataset from 302 elite soccer matches.
• Comparison of effect size estimation between TMLE and GLM.

Main Results:

• TMLE demonstrated superior performance over GLM in estimating the effect size of substitutions on total physical performance in simulation data.
• TMLE exhibited greater robustness against model misspecification in both simulation and real-world tracking datasets.
• Analysis of the tracking dataset indicated that substitutions positively impact the overall physical performance of soccer teams, irrespective of the statistical method used.

Conclusions:

• TMLE is a valuable and robust tool for causal inference in sports science, particularly for estimating the impact of interventions like substitutions.
• The study provides a practical framework for applying TMLE in sports science research.
• Substitutions are confirmed to enhance the collective physical performance of soccer teams.