Stacking: An ensemble learning approach to predict student performance in PISA 2022

Abstract

Predicting student performance in international large-scale assessments (ILSAs) is crucial for understanding educational outcomes on a global scale. ILSAs, such as the Program for International Student Assessment and the Trends in International Mathematics and Science Study, serve as vital tools for policymakers, educators, and researchers to examine the effectiveness of educational systems worldwide. By accurately predicting student performance, policymakers and educators can identify trends, disparities, and areas for improvement in educational practices and policies. Researchers can utilize traditional statistical methods or machine learning algorithms to predict student achievement, enabling proactive interventions to support struggling learners and enhance overall educational quality. In this study, we employed stacking, an ensemble machine-learning algorithm, to predict student performances in large-scale assessments based on a wide range of predictors. Without filling in missing data or categorizing the outcome variable, we predicted student performances using the stacking method and then compared the results with those generated by three boosting algorithms and blending. Our findings revealed that stacking outperformed the boosting and blending methods, yielding more stable and accurate predictions. Our analysis encompassed the 80 countries that participated in the administration of PISA 2022. Compared to the three boosting algorithms and blending, we found that stacking demonstrated superior performance with the lowest error metrics for most countries. Robust linear mixed-effects models also indicated that stacking produced significantly lower MAPE, MAE, and MSE values than boosting and blending. Overall, our findings emphasize that stacking is one of the most accurate methods to predict student performance in large-scale assessments.