Predicting adolescent conduct problems: A machine learning approach using early family and child predictors.

Early recognition of predictors of adolescents' conduct problems is crucial for timely intervention. The use of traditional regression models, however, is limited in the ability to identify non-linear relationships between predictor variables. This study employed multiple machine learning algorithms to predict adolescent conduct problems from earlier family and child functioning variables, comparing algorithm performance to identify which methods and early indicators most accurately forecast later problems. Data from 242 families were analysed using eight machine learning algorithms. The algorithms were used to predict conduct problems in adolescence, as assessed by the SDQ Conduct Problems scale, from 15 predictor variables assessed 1 year earlier, including but not limited to interparent conflict, emotional security, depression, anxiety and somatic concerns. Support vector regression emerged as the best-performing model (R2 = .343, RMSE = 1.381), outperforming baseline prediction. Feature importance analyses indicated that disengagement in response to interparental conflict was the strongest predictor of later conduct problems. Insecurity in the interparental relationship, depressive symptoms and aggressive behaviour also contributed substantially to prediction. This pattern suggests that children's dysregulated coping and perceived lack of safety within the family system are more proximal predictors of later conduct problems than exposure to conflict alone.