Abstract
Child welfare agencies are increasingly required to leverage their limited resources to meet nearly limitless demands. As a result, agencies are searching for new opportunities to efficiently improve policy and practice, and advances in data availability and technology have brought increased attention to the utility of predictive modeling. While the literature has often highlighted the considerable potential of predictive models leveraging “big data”, discussions of the methodology and the associated best practices remain critically absent. To address this gap, this paper provides an illustrative case involving the development and testing of models used to predict the probability of whether U.S. foster children would achieve legal permanency. The models were trained and tested using a national administrative dataset of 233,633 foster care children that discharged from state child welfare systems in 2013. The optimal model, a boosted tree, predicted whether children would achieve permanency with 97.66% accuracy. The paper concludes with a discussion of best practices detailing how agencies can utilize predictive modeling to enhance policy and practice.