Comparing the Predictive Importance of Mathematics Self-Efficacy, Socioeconomic Status and ICT Access: A SHAP Analysis of PISA 2022 Across 19 Education Systems

Digital access occupies a prominent place in educational policy debates, yet its predictive contribution to mathematics achievement, relative to psychological and socioeconomic factors, remains insufficiently quantified. Existing evidence often relies on linear models, examines predictor blocks separately, or considers technological resources without placing them alongside psychological constructs within a common framework. This study compares the predictive importance of psychological, socioeconomic, demographic and ICT-access indicators for mathematics achievement in PISA 2022. The analysis used data from 141,563 students across 19 education systems and included seven predictors: mathematics self-efficacy, mathematics anxiety, sense of school belonging, economic, social and cultural status (ESCS), gender, ICT resources at home and ICT resources at school. Weighted gradient boosting models were fitted separately to each of the ten plausible mathematics values and interpreted using TreeSHAP; a weighted random forest with permutation importance was used as a robustness check. The full model explained 39.2% of the weighted test-set variance in the plausible-value outcomes (R2 = 0.3922, SEtotal=0.0066, and RMSE = 77.95 score points). Mathematics self-efficacy ranked first under both criteria (42.3% of SHAP importance; 57.8% of permutation importance), ahead of socioeconomic status (30.9%; 32.0%), while the ICT-access block contributed 6.9% and 2.1%, respectively and added 0.0122 to test-set R2. The two importance rankings were identical (Spearman’s ρ = 1.000). The SHAP ranking was unchanged across all 800 plausible-value × replicate-weight runs, matched XGBoost permutation importance, and was reproduced under a school-grouped train–test split. The model also detected a non-monotonic association between school belonging and predicted achievement, together with a MATHEFF × ESCS interaction pattern, supported by a direct-outcome interaction model, in which the modelled association between self-efficacy and achievement was stronger at higher ESCS levels. Because PISA 2022 is cross-sectional and plausible values are designed for population-level inference, these findings should be interpreted as predictive and associational rather than causal. The results suggest that, in systems where ICT access is already widespread, reported access to technological resources contributes comparatively little to prediction once psychological and socioeconomic indicators are considered, although sensitivity analyses indicate that home ICT access is partly affected by missingness patterns.