Measurement of sporting abilities using item response theory: an introduction and example using passing ability in soccer

Measurement of performance in sport has historically followed operationist tendencies and applications of Classical Test Theory (i.e., true score theory). However, while specific performance is often the focus, it is likely the underlying latent ability that gives rise to performance which is of greater importance particularly for those interested in things such as athlete selection, long term development, and intervention effectiveness. Item Response Theory (IRT; i.e., latent trait theory) can allow for the modelling of individuals latent abilities and their observed performance on a particular test but, despite early suggestions (Safrit, Cohen, and Costa, 1989, Research Quarterly for Exercise and Sport, 60(4), 325-335), has rarely been applied in sporting contexts. Thus, our aim is to re-introduce and explain this approach for measurement of sport abilities. We use the example of a specific sporting ability, passing in soccer, and apply IRT based models to a test of this ability. Existing data (data collection is ongoing) was used from 1,189 tests employing a 360 degrees random light target (24 targets) based passing task with 305 players representing an assumed heterogenous range of ability levels (i.e., ranging from professional to recreational soccer players). We sought to model the binary response data (i.e., whether they accurately passed to the target or not) as a function of both the underlying latent player ability, and characteristics of the target in terms of its absolute distance in degrees from the previous target (i.e., how far from the previous pass the current one was). Data were analysed using a Bayesian hierarchical regression framework (Bürkner, 2019, arχiv, DOI: 1905.090501) and a model selection process was conducted by fitting and comparing a range of models varying in complexity from one-parameter logistic (i.e., Rasch model) to four-parameter logistic models, both with and without person (age) and target (right or left turn, turn angle, target type) covariates, and also with response and response time jointly modelled. We use this example to highlight and explain the application of IRT based models to sporting data as a powerful and flexible approach to understanding both player abilities and test characteristics in addition to the effect of both person and item covariates.