Exploring Machine Learning to Support Decision-Making for Placement Stabilization and Preservation in Child Welfare

The Family First Prevention Services Act requires youth’s placement in residential care to be clinically appropriate, time-limited, and only when youth’s needs cannot be met in family-like settings in foster care. State child welfare agencies can benefit from upstream, empirical decision support to preempt youth’s placement disruption, coordinate proactive placement stabilization services, prevent unnecessary step-up to residential care, and improve outcomes for the youth. This statewide case study explores the potential benefit to child welfare decision support for placement stabilization and diversion from residential care, by comparing predictive machine learning (ML) models with conventional regression models. We analyzed child welfare spells of 12,621 youth in one large Midwestern state between January 2017 and January 2020. Caseworkers could refer youth to a placement stabilization and preservation program. To predict youth’s monthly program need in the next 6 months, we developed and validated a wide grid of ML models—random forest, regularized logistic regression, decision tree, dummy classifier—and a conventional unregularized logistic regression model, using literature-informed predictors from child welfare administrative data. We retrained, retested, and compared all models over time using temporal hold-out sets. Based on anticipated program capacity, model evaluation focused on accuracy in identifying the 100 highest-need youth, fairness, and equity of resource allocation. Random forest models produced the best performance with a precision (positive predictive value) 10 times greater than baseline precision. Common important predictors across models included youth’s age, history of placement changes, and emotional/behavioral needs. We discuss potential applications of ML to support preventive child welfare decisions, adapt to policy changes, and allocate limited resources.