Recurrent and system-specific model-attribution patterns in PISA 2022 science performance across nine education systems.

Abstract

Using science data from the 2022 Programme for International Student Assessment (PISA), we examined survey-weighted descriptive association and model-attribution patterns across a fixed, non-probability cohort of nine education systems (78,685 students). Because PISA science plausible values are drawn from a latent population distribution generated with a conditioning model that uses background information, the analysis describes survey-weighted descriptive associations rather than independent individual predictions. Seven machine-learning pipelines were evaluated descriptively across 20 repeated school-grouped held-out partitions in each system. The genetic algorithm-optimized backpropagation neural network (GA-BPNN) was retained as the focal architecture for Shapley additive explanations (SHAP), without claiming systematic or universal superiority. Alternative training-only median/mode imputation produced system- and metric-specific performance changes. The four sensitivity-supported variables-ESCS, HOMEPOS, MATHEFF, and ST255Q01JA-formed a conservative subset of the eight variables that recurred under the primary top-35 screening procedure. Feature-screening stability was assessed under the high-missingness quartile stress test, whereas held-out performance and SHAP magnitude/rank variation were assessed only under the full alternative median/mode imputation workflow; the sample-restriction analysis was not used to assess changes in SHAP ranks. A cross-architecture check showed substantial but incomplete agreement between GA-BPNN SHAP and XGBoost TreeSHAP. These results describe recurrent and system-specific attribution patterns across the nine education systems. They do not support causal inference, individual diagnosis, or generalization to all PISA systems; missing-data analyses did not resolve missing-not-at-random (MNAR) mechanisms or residual selection bias.