关键词: 青少年, 处境不利学生, 学业韧性, 社会情感能力韧性, 机器学习

Abstract: Children and adolescents who attain favorable developmental outcomes despite socioeconomic adversity are often characterized as resilient. While prior work has predominantly focused on academic resilience, less attention has been paid to social-emotional skills resilience, defined here as positive social-emotional adaptation among socioeconomically disadvantaged students. The present study aimed to identify the key influencing factors of both forms of resilience and to estimate their quasi-causal effects, while examining how these effects vary across developmental stages and gender.
Data were drawn from the OECD Survey on Social and Emotional Skills (SSES; N = 54, 729 students from nine countries, ages 10 and 15). Resilient students were operationalized within each country as those whose family socioeconomic status fell in the bottom quartile while their academic achievement (academic resilience) or social-emotional skills score (social-emotional skills resilience) fell in the top quartile. Missing data were addressed using multiple imputation by chained equations (five imputed datasets with ten iterations). Five machine learning algorithms, including Lasso regression, Random Forest, XGBoost, Classification and Regression Trees, and LightGBM, were used to screen candidate predictors across individual, family, and school domains. The top-ranked predictors were then entered into inverse probability of treatment weighting (IPTW) analyses to estimate average treatment effects (ATEs) under the assumptions of ignorability and overlap, with cluster-robust standard errors at the school level. Cross-cultural robustness was assessed via country fixed-effects models and random-effects meta-analytic pooling of country-specific ATEs. Subgroup analyses were conducted by age (10 vs. 15) and gender.
After covariate balancing, school belonging, interpersonal closeness with peers, and teacher-student relationships showed consistent positive quasi-causal effects on both forms of resilience, whereas parent-child conflict and academic anxiety showed consistent negative effects. For academic resilience, social-emotional skills exhibited the largest positive effect; for social-emotional skills resilience, extracurricular participation, self-academic expectations, friendship quality, and emotional growth mindset further showed domain-specific positive contributions. These patterns remained directionally stable across countries in fixed-effects and meta-analytic models, with effect magnitudes varying across contexts. Subgroup analyses suggested that 10-year-olds were more sensitive to school-context and emotional factors than 15-year-olds, and that relational support was relatively more salient for girls while academic motivation and structural variables played a relatively larger role among boys.
Together, these findings suggest that academic and social-emotional skills resilience are supported by partially overlapping but domain-specific protective configurations, in which relational resources function as shared foundations and domain-specific factors offer differentiated support. The integration of machine learning with IPTW provides a feasible analytic pathway combining variable selection with quasi-causal estimation. Findings should be interpreted as quasi-causal estimates conditional on observed covariates, and may offer directional implications for designing supportive school environments and developmentally tailored interventions for disadvantaged students.

Key words: adolescents, disadvantaged students, academic resilience, social-emotional skills resilience, machine learning