A Machine-Learning Approach to Understanding Performance of Canadian Nonprofit Sport Organizations

Previous approaches to model the performance of nonprofit organizations and their determinants largely rely on linearity and monotonicity assumptions. This research note makes a methodological contribution by jointly using descriptive and predictive models, particularly advanced machine-learning algorithms that allow for the consideration of non-linear or non-monotonic relationships, to understand the relevance of factors associated with the performance of nonprofit sport clubs as well as the nature of relationships. Data were collected via an online survey with 126 representatives of Canadian sport clubs, in which four performance domains were considered: Member relationship, service quality, financial stability, and sporting success. Explanatory linear regressions and four machine-learning models (i.e., ridge regression, bagged regression, random forest, and gradient boosting machine) are used. The results reveal that machine-learning models increase the explanatory power compared to linear models. The random forest outperforms the other models in terms of root mean squared error and, partly, mean absolute error, and R square (even though absolute levels of R square are low at times, particularly for financial stability and sporting success, where the presence or absence of a high-volume donor or high-performance sports mission might help or hinder performance). Non-linear relationships are found for several predictors across the four dimensions that were considered, such as the use of outside knowledge, trust, coopetition, age, and tenure of the club representative. We showcase the use of joint computational techniques in nonprofit research to serve two relevant goals: enhance the explanatory power and maintain the interpretability of predictive models.