Predictive Algorithms in the Frontlines of the Administrative State: Empirical Evidence from Child Protection Screening

Abstract

Administrative agencies increasingly deploy predictive algorithms to support frontline decision-making, yet evidence on how such tools affect real-world practice remains limited. This study examines the introduction of a predictive risk model (PRM) used by Child Protection Services (CPS) in a large U.S. county to assist screening decisions for allegations of abuse or neglect. A technical constraint in the county’s case management system generated quasi-experimental variation in access to the PRM: the score could be displayed in morning screening sessions but not in afternoon sessions. Using approximately 130,000 client-referral observations from 2020–2023 and a difference-in-differences design, we estimate the intention-to-treat effect of PRM availability on screening and investigation-track assignments. We find modest but systematic impacts. Access to the PRM reduced screening-in rates for low-risk referrals and increased the likelihood that screened-in cases were assigned to the less intrusive assessment track. Effects did not extend to higher-risk referrals and attenuated over time, suggesting that organizational routines and worker discretion mediate algorithmic influence. The PRM’s presence improved alignment between predicted and assigned investigation tracks, though effects were small in magnitude. These findings highlight the importance of evaluating predictive tools within the conditions of field deployment rather than in isolation. In hybrid human-algorithm systems, the effects of predictive models depend not only on their statistical performance but also on the institutional, temporal, and behavioral contexts in which they are embedded.