Beyond Match Results: Bayesian Mediation of Ability, Performance, and Winning in Professional Tennis

Abstract

Match results measure competitive ability, but they do not explain how ability is expressed during play. Match statistics offer a natural mediator, however performance is usually measured over the same match as the outcome: a full-match statistic uses events that help determine the winner, while a first-set statistic is earlier but rests on fewer service points and still contributes to the result. Comparing the two therefore mixes timing and precision. We develop a Bayesian mediation framework for this problem. Dynamic abilities on hard, clay, and grass courts are estimated from previous seasons only and are linked to both players' service performance, and a Bradley--Terry model relates ability and performance to the match result. A control mediator samples the first set's number of service points from the full match, separating the measurement window from the point count. In 3,820 men's matches, the hard-court mediated share rises from 0.52 with a first-set mediator to 0.95 with a full-match mediator; the point-count contrast accounts for 0.30 of this gap and the window contrast for 0.13. The pattern replicates on the women's tour. Simulations and sensitivity analyzes examine the decomposition and assumptions needed for a causal interpretation.