CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Bayesian modelling of athletic performance Authors:  Jim Griffin - University College London (United Kingdom)
Maria Zafeiria Spyropoulou - University of Kent (United Kingdom) [presenting]
James Hopker - University of Kent (United Kingdom)
Abstract: Publicly available databases of sporting competition results have made it possible to model athletic performance across a wide range of sports accurately. A Bayesian hierarchical model was developed to analyse athletic sporting performance in track and field athletics and weightlifting, accounting for confounding factors such as age, month and year effects, as well as environmental conditions in athletic events. Bayesian variable selection was used to fit a spline model separately to the performance results of each athlete within a performance database. The specific focus of this project was on the development of an approximate algorithm which addresses the issue of computational intensity and lack of processing speed associated with MCMC. This is a two-stage algorithm where the first step is to estimate parameters shared by all athletes, while the second step estimates individual athlete-specific parameters. For the first step, we implemented an EM algorithm in combination with variational Bayes. Then, for the second step, instead of using an MCMC algorithm for all parameters, we used an adaptively scaled individual (ASI) version of the MCMC algorithm to fit the athlete-specific spline model. This approach allows us to utilise parallel computing alongside the ASI algorithm to accelerate the processing speed further. We will illustrate the performance of the algorithm on several databases of sporting performance.