Advancing textbook evaluation with debiased machine learning: a theoretical and empirical approach

Introduction Textbooks can substantially influence student achievement, but common evaluation approaches (e.g., linear regression) often depend on strong functional-form assumptions that may misstate causal effects. This study presents Double/Debiased Machine Learning (DML) as a more flexible framework for estimating the causal impact of textbooks on learning outcomes. Methods We use DML to estimate textbook effects while allowing high-dimensional, non-parametric modeling of outcome and treatment assignment processes. We (1) derive the theoretical advantages of DML-particularly its robustness to model misspecification and its use of orthogonalized estimating equations-and (2) apply the approach to an existing large-scale elementary school mathematics curriculum dataset. We compare DML estimates to those produced by Ordinary Least Squares (OLS) regression and Kernel matching, focusing on precision and efficiency of causal effect estimation. Results Across the empirical application, DML yields more precise and efficient estimates of textbook effects than OLS and Kernel matching. The approach reduces reliance on restrictive linearity assumptions and improves the stability of estimated causal impacts in settings where relationships between covariates, curriculum assignment, and outcomes are complex. Discussion These findings indicate that DML is a robust alternative for evaluating educational materials, offering clearer evidence to inform curriculum selection and adoption decisions. More broadly, the study contributes methodologically to learning and intelligence research by strengthening the tools used to measure educational interventions' effects on achievement.