Development and validation of a predictive risk model for runaway among youth in child welfare

Existing predictive risk models in child welfare tend to focus on child protection and prevention of youth’s entry to the child welfare system. Less explored is the potential contribution of predictive risk models to improve the care and meet the placement needs of youth in the child welfare system. This study demonstrated how an empirical model predicted youth’s risk for running away from child welfare placement to inform preventive interventions for youth in care. We used administrative data on youth’s demographic, child welfare, and clinical characteristics to develop a Cox proportional-hazards model. The model predicted time-to-first runaway among 8,255 legal custody spells of 12- to 17-years-olds in the care of the public child welfare agency of one large Midwestern state over an eight-year period. We validated the model internally using 200 bootstrap resamples from the development sample, and externally using a more recent, but smaller sample of 1,836 legal custody spells for youth whose information was not used to develop the model. Internal validation indicated stable predictor estimates, significant predictors that corroborate with relevant empirical literature on runaway, and strong discriminative ability and prediction accuracy (bias-adjusted c-statistic = 0.78). Through external validation, we demonstrated one preventive application of the model by predicting the risk of running away within the first 90 days of a legal custody spell. The model maintained the expected relationships between model-based predicted risks of runaway and actual runaway, as evidenced by accuracy metrics of precision and recall. We conclude by describing how child welfare agencies can apply predictive risk models such as ours to identify youth’s placement needs and estimate system capacity to prevent runaway.