ABSTRACT
With the aim to explore the potential of machine learning for nonprofit research, this article contrasts traditional linear regression with four contemporary supervised machine learning approaches. Concretely, we predict (1) reputation ratings and (2) the total number of volunteers for 4021 non-profit organizations in the U.S. Combining two distinct sources of data, 56 predictors spanning financial characteristics, governance practices, fields of activity, and emotional framing categories are used for the predictive tasks. Results indicate a notable predictive superiority of machine learning models, particularly for Random Forests and (Extreme) Gradient Boosting. Ranking predictors according to their relative predictive importance, we find that financial indicators and governance practices are most decisive for the prediction of organizations’ reputation rating. No clear-cut conclusions can be drawn from the predictor ranking for the prediction of the number of volunteer, which highlights the inherent complexity of the process of volunteer attraction. We discuss these findings as well as the learning points of our supervised learning approach to evaluate potential avenues for leveraging machine learning in future nonprofit research.