Algorithms and Decision-making

Abstract

Around 40% of children experience maltreatment (Finkelhor et al. 2013), with harmful outcomes and high social costs. Child protection decisions are complex, and potentially biased and/or prone to errors due to underfunding and overload. Our randomized controlled trial showed algorithmic tools sped up decision-making but didn't significantly change child outcomes, perhaps because COVID disruptions limited outcome analysis. Results are suggestive that risk aversions played a role: high-risk cases flagged by the tool were more likely screened in, while low-risk cases weren’t more likely screened out. Time savings from the tool could enable caseworkers to spend more time directly with families.