Predictors of Placement Instability in Foster Care: A Machine Learning Analysis of AFCARS Data

Placement instability, when a child in foster care experiences multiple placement changes, is associated with many things, including disrupted schooling, damaged relationships, and compounded trauma. This study makes use of restricted data from the federal Adoption and Foster Care Analysis and Reporting System (AFCARS) and works to identify factors associated with placement instability among children who exited foster care during federal fiscal year 2024, drawing on an analytic sample of N = 131,315 children with complete data on the selected features, out of 179,420 children who exited that year. Placement instability was defined as three or more placement episodes within a child's current removal episode, a threshold consistent with prior foster care research. Three models were compared in this project: logistic regression, XGBoost, and a class-weighted XGBoost model. The class-weighted model achieved the strongest recall on the unstable class (0.60). It correctly identified 60 percent of children who experienced instability, compared to 43 percent for the unweighted model. Using SHAP analysis, age at entry and most recent case plan goal emerged as the two strongest predictors. Older age at entry was associated with higher instability, and children with a case plan goal of long-term foster care (72.3 percent instability) or emancipation (70.6 percent) had substantially higher instability rates than those with a goal of reunification (30.2 percent). These findings are correlational; because case plan goal is recorded as of a child's most recent AFCARS record rather than fixed at intake, there is likely a bidirectional relationship between case plan goal and instability rather than a one-directional predictive one. This analysis is best understood as a retrospective, cross-sectional look at which case characteristics are associated with instability among children who have already exited care, not as a prospective screening tool. Therefore, it is intended to complement direct casework and professional judgment.