Association for Public Policy Analysis & Management

Child welfare agencies operate in an environment that increasingly requires using limited resources to meet nearly limitless demands. As a result, agencies are increasingly searching for new opportunities that most effectively leverage their limited resources to improve both child welfare policy and practice. Recent advances in data availability and computing technology have brought increased attention to the potential of using predictive analytics, and placement instability is a critical area where predictive analytics approaches can be used to improve child welfare policy and decision-making. Children that experience multiple placements during their time in the child welfare system are at-risk for impaired development and psychological well-being, greater uncertainty regarding their futures, and a greater likelihood of emancipating from the child welfare system. Earlier identification of children that are at-risk for high levels of placement instability can allow child welfare staff to more effectively target their resources to minimize the number of placements that a child is likely to experience while placing an increased focus on identifying stable placements that promote permanency. This paper utilizes a random sample of 15,000 children from the Adoption and Foster Care Analysis and Reporting System (AFCARS) dataset. A theoretically driven approach consisting of 21 variables along with a data-driven approach consisting of 60 variables are used to run a collection of predictive analytics models that determine whether children were likely to experience a high number of placement moves while in the care of a child welfare system. The sample was partitioned into training and test sets, with a test set of 11,250 children used to train linear and nonlinear classification models, and classification trees and rule-based models, to identify whether a child was likely to experience four or more placement moves while in care. After training the models on the training set, the models were run on a test set of 3,750 children to predict which children were likely to experience a high number of placements while in care. Results found that the prediction accuracy of the models ranged from 85% on the low end to greater than 90% for optimal models. The paper then identifies a common set of variables across all models that child welfare agencies can place greater focus on in their efforts to improve the experiences and outcomes for children at high-risk for placement instability. The paper concludes with a discussion of the strengths and limitations of the various models, the tradeoffs associated with model accuracy and interpretability, and a collection of best practices for government agencies and policymakers interested in implementing predictive analytics models.