Modelling football match scoring outcomes using multilevel models

Abstract

Multilevel modelling technique recognizes the existence of hierarchal structures in the data by allowing for random effects at each level in the hierarchy, thus assessing the variation in the dependent variable at several hierarchical levels simultaneously. Multilevel modelling is becoming an increasingly popular technique for analysing nested data with such popularity accredited to the computational advances in the last two decades. In many sports, including football, the game fixtures are nested within seasons, which in turn are nested within country leagues invoking a multilevel structure in the data. Many gaming companies engage in sport data analysis in a bid to understand the dynamics and patterns of the game. This will assist the gaming company in developing fantasy sport games that will enhance gamer engagement and augment revenue to the company. This paper presents a comprehensive description of two and three level models, which are applied to a real football data set accessed from an online free football betting portal. The aim is to examine the relationship between the number of goals scored during a football match and several game-related predictors. These multilevel models, which assume a Poisson distribution and a logarithmic function, are implemented using the facilities of GLLAMM (Generalized Linear Latent and Mixed Models), which is a subroutine of STATA.