Cross-economy stability of predictor rankings for adolescent school belonging and mathematics anxiety: an explainable machine-learning analysis of PISA 2022

Introduction School belonging and mathematics anxiety reflect distinct relational and motivational processes, but cross-economy comparisons can blur distinctions among psychological relationships, model-based importance, and questionnaire availability. Methods We analyzed Program for International Student Assessment (PISA) 2022 data from 613,744 students; outcome-valid samples comprised 561,339 students in 78 economies for school belonging and 475,792 students in 76 economies for mathematics anxiety. Per-economy LightGBM models and TreeSHAP rankings estimated relative predictive salience. Secondary post-hoc analyses assessed economy-level predictor availability, excluded unadministered modules, and fitted complementary association models. Official individualism scores covered 74 economies; outcome-eligible reference sets covered 73 for belonging and 70 for anxiety. Results Eight of 12 focal combinations retained stable relative predictive salience and formed two psychologically interpretable predictor patterns. Perceived school safety, exposure to bullying, and student-teacher relations formed a relational-institutional pattern for belonging, while mathematics self-efficacy, 21st-century-skills self-efficacy, achievement, growth mindset, and persistence formed an efficacy-achievement-engagement pattern for mathematics anxiety. In separate weighted economy fixed-effects samples of 241,786-476,803 students, safety (β = 0.323) and student-teacher relations (β = 0.282) were positively associated with belonging, exposure to bullying was negatively associated (β = 0.253), and the five focal mathematics predictors were negatively associated with anxiety (β = −0.080 to −0.340). Continuous individualism interactions were small (absolute β ≤ 0.047), inconsistent, and multilevel sensitivity models retained the direction of every main association. Four combinations involving stress resistance, assertiveness, and cooperation failed the availability criterion because these indices were much less available in low-individualism economies and unevenly distributed across IDV terciles. Discussion The evidence therefore supported stable relative predictive salience across economies more strongly than individualism-related differentiation. Together, the predictive rankings and association models provided complementary evidence for these two patterns but did not establish psychological mechanisms, causal effects, or transferable intervention benefits.