Identifying Family Predictors of Adolescents’ Future Educational Attainment Across Different Race/Ethnic Groups: A Machine Learning Study
Family experience during adolescence plays a critical long-term role in educational attainment (Sun et al., 2020). Ecological models (Bronfenbrenner & Morris, 2006) and family systems theory (Minuchin, 1985) emphasize the role of multiple, interacting family factors in adolescent development. However, most studies consider only a subset of family experience predictors and rely on traditional statistical methods—limiting the ability to capture how diverse family factors jointly shape outcomes, dismissing complex, non-linear data patterns. Further, the cultural ecological model (García Coll et al., 1996) posits that findings based on predominantly White samples may not generalize to other groups. The salience of different family factors may vary across racial/ethnic groups, along with persistent disparities in educational attainment. This study applies machine learning (ML) to identify key predictors separately among White, Black, and Latinx youth groups. Restricted-access data from the U.S. National Longitudinal Study of Adolescent to Adult Health was used to examine 119 potential adolescent family predictors at Wave I (grades 7 to 12) of educational attainment at Wave IV (14 years later). Several ML models were trained with 5-fold cross-validation, tuned via 100-iteration random search, and evaluated on a 20% test set using R² and mean squared error (MSE). Predictor importance was assessed using SHapley Additive exPlanations (SHAP) in the best-performing models. Optimal ML algorithms differed across groups, highlighting the benefit of stratified modeling. SHAP showed parental education, childhood exposure to violence, and family income were strong predictors across all groups. However, there were differences among the most salient predictors for different race/ethnic groups which emphasizes the need for subgroup-specific modeling. In conclusion, integrating ML with family theories can uncover shared drivers and unique culturally embedded pathways often overlooked by traditional methods. Future work will extend this analysis to a larger group of predictors, including individual and school domain predictors.