Identifying prospective temperament predictors of callous-unemotional traits using machine learning

Children with callous-unemotional (CU) traits (i.e., low guilt, restricted empathy) are at high risk for disruptive behavior disorders (DBD) across development. The Sensitivity to Threat and Affiliative Reward (STAR) model posits that low fear and low affiliation (i.e., disrupted social bonding motivation) are temperament dimensions that increase risk for CU traits. However, prior tests of the STAR model are limited by the lack of prospective longitudinal studies and reliance on short-term, single-timepoint, single-measure assessments. We applied machine learning to repeated observational measures of temperament across infancy and preschool to test whether STAR model features (i.e., fear, affiliation), alongside other temperament constructs (e.g., frustration, activity, persistence), predicted CU traits at age 7. Data were from the Family Life Project (FLP), a birth cohort study (N = 1,292) that oversampled families with low household incomes. We used random forest models to predict CU traits and conduct disorder (CD) symptoms at age 7 using 39 features derived from observed temperament measures assessed at ages 6, 15, 24, 35, and 48 months. Models explained only 2% of the variance in CU traits at age 7, with behavioral observations of positive affect and persistence at 48 months among the strongest predictors. Although temperament measures relevant to affiliation were modestly predictive of CU traits, findings overall provide weak evidence for the STAR model.