Università degli studi di Modena e Reggio Emilia

Abstract
Student learning is a complex process whereby inputs are converted into outputs. Heterogeneity in student performance arises from various sources, including the students themselves, the class, the school and the geographical area. We propose tree-based mixed-effects models to analyze educational data with the goal of untangling the effects stemming from different levels of grouping and to flexibly and efficiently model the educational production function. In particular, we develop and describe innovative Classification And Regression Trees and Random Forest for hierarchical data and generalized response variables. We then leverage the potential of these methods in two primary contexts: analyzing 15-year-old students across countries using the OECD-PISA dataset and predicting student dropout at the university level using the Politecnico di Milano dataset. Results demonstrate that tree-based mixed-effects methods enable the modeling of non-linearities and interactions among covariates, capturing diverse forms of the educational production function. Moreover, these methods produce interpretable and transparent evidence while still accounting for the nested structure of the data.