Using machine learning to model interaction effects in education: A graphical approach

Abstract

Educational systems can be characterized by a complex structure: students, classes and teachers, schools and principals, and providers of education. The added value of schools is likely influenced by all these levels and, especially, by interactions between them. We illustrate the ability of Machine Learning (ML) methods (Regression Trees, Random Forests and Boosting) to model this complex 'education production function' using Hungarian data. We find that, in contrast to ML methods, classical regression approaches fail to identify relevant nonlinear interactions such as the role of school principals to accommodate district size policies. We visualize nonlinear interaction effects in a way that can be easily interpreted.