Comparison of Statistical and Machine Learning Approaches for Predicting Mathematical Literacy: Evidence from PISA 2022 Türkiye

This study compares statistical and machine learning methods for predicting mathematical literacy among students in Türkiye who participated in PISA 2022. Using data on 6,427 students and 13 standardized predictors capturing cognitive, affective, and contextual dimensions, we evaluated multiple linear regression, least absolute shrinkage and selection operator, random forests, extreme gradient boosting, artificial neural networks, and a stacking ensemble within a 10-fold cross-validation design. Ensemble approaches outperformed linear methods: the stacking model achieved the highest accuracy (out-of-fold R² = .319; RMSE = .777), followed closely by extreme gradient boosting (R² = .313) and random forests (R² = .304). Linear models yielded weaker results (multiple linear regression R² ≈ .270; least absolute shrinkage and selection operator R² = .273). Mean absolute error values were nearly identical across models (≈ .633–.658), with minimal between-fold variation due to rounding at three decimals. Residual analyses indicated that ensemble models produced more stable error structures, whereas linear methods showed stronger heteroskedasticity. Across all approaches, socioeconomic status consistently emerged as the strongest predictor, followed by mathematics self-efficacy and disciplinary climate, underscoring the dual roles of student beliefs and classroom environment. These findings highlight the advantages of ensemble methods for predictive performance and variable-importance estimation, emphasizing the ongoing impact of socioeconomic inequalities on educational outcomes.