Nonlinear dynamics of attrition in adolescent panels: Identifying differential dropout patterns using random survival forests

Abstract

Panel attrition in adolescent research is not merely a methodological nuisance but often reflects substantive developmental discontinuities during critical transitions. Unlike adult populations, adolescents face unique attrition risks associated with institutional changes and shifting developmental trajectories that traditional linear models often fail to capture. This study investigated the predictors of panel attrition using data from the Korean Education Longitudinal Study 2005, which covers the extensive developmental period from early adolescence to early adulthood. Random survival forests (RSF) were used to capture the complex, nonlinear trajectories of adolescent behavior obscured by traditional linear models. We found that structural instability, specifically school transfer experience, was the strongest predictor of attrition. Crucially, RSF uncovered distinct nonlinear patterns: a threshold effect for academic achievement, whereby risk was concentrated among low achievers, whereas maternal education showed a U-shaped pattern with elevated risk at both extremes. These findings highlight that attrition is associated with divergent developmental mechanisms such as academic disengagement and strategic parental gatekeeping, emphasizing the need for developmentally sensitive methodologies in longitudinal research.