Heterogeneous Returns to Higher Education: An Estimation Based on Generalized Random Forests

Abstract

This study investigates the heterogeneous causal effects of higher education on earnings by employing generalized random forests (GRF), a nonparametric machine learning method designed for estimating treatment effect heterogeneity. While existing literature predominantly focuses on average returns to education, substantial variation in individual benefits remains underexplored due to methodological limitations in capturing complex interaction effects. Using data from the National Longitudinal Survey of Youth (NLSY), we address potential selection bias through a causal forest framework that incorporates a rich set of socioeconomic, demographic, and cognitive characteristics. Our results reveal significant heterogeneity in returns, with estimated annual income increases ranging from 5% to 28% across individuals. Key moderators include family socioeconomic status, pre-college academic ability, and local labor market conditions. These findings underscore the limitations of average treatment effects and highlight the importance of personalized educational policy. The application of GRF demonstrates its utility in uncovering nuanced patterns of treatment heterogeneity where conventional parametric methods fall short.