Latent Profiles and Regression-Based Analysis of Socio-Emotional Development: Evidence from Longitudinal Data and Interpretable Machine Learning

Abstract

Socio-emotional development in early childhood serves as critical foundations for later psychological well-being, academic achievement, and social adjustment. While previous studies have identified various factors that influence these developmental outcomes, there remains a need for data-driven approaches that simultaneously capture both categorical profiles and continuous variation in socio-emotional devel opment. Accordingly, the present study draws on longitudinal data from the Panel Study on Korean Children to identify key predictors of socio-emotional development in early childhood. Specifically, independent variables measured between ages two and five were used to predict socio-emotional outcomes at age six. The analysis focused on three domains of socio-emotional development: problem behavior, self-esteem, and social competence. Initially, latent profiles were derived through principal component analysis and K-means clustering, and classification methods were applied to identify predictors of profile membership. In parallel, Ridge regression models with Shapley Additive Explanations (SHAP) were employed to estimate the relative contribution of each predictor to individual outcomes. The classification model achieved an accuracy of 86.7% in predicting latent socio-emotional profiles, and Ridge regression models explained 25.1% to 36.2% of the variance in the three outcome variables. SHAP-based interpretation across both models revealed that peer interaction, maternal warmth, and early sociability functioned as protective factors, whereas disruptive behavior and parenting stress were associated with negative adjustment. These findings grounded in statistically robust model performance and interpretable machine learning analysis, highlight the practical utility of this approach for early screening and intervention.