Research Papers

Ben Jenkins

This paper presents an interference-aware causal inference framework for evaluating defensive performance in American football, utilizing player tracking data. We model all 22 players jointly with a multi-agent Transformer to estimate counterfactual Expected Points Saved (EPS) for individual defenders while explicitly accounting for coordination, substitution, and other violations of independence. Using doubly robust estimation, adversarial representation balancing, and Bayesian uncertainty quantification, we recover causal defensive value that traditional tackle-based and predictive models systematically miss. Extensive synthetic validation, sensitivity analyses, and real-game examples demonstrate that explicitly modeling defensive interdependence substantially reduces bias and reveals hidden contributions across positions.