Predictive Risk Scores in the Public Sector: Experimental Evidence from Child-Protection Investigations
NBER Working Papers · 2026 · E. Jason Baron、Arkadev Ghosh、Richard Lombardo
中文摘要
公共部门的许多决策都需要在不确定性下分配稀缺的关注资源。我们考察算法风险评估能否改善儿童保护决策——在此类决策中，主管人员需要判断哪些案件应受到更严格的审查。我们在北安普顿县对14个月内4,752起儿童转介案件进行了一项随机评估，主管人员在获得常规案件记录的同时，还获得了算法风险评分。获取该评分提高了预测风险最高儿童的寄养安置率与服务提供率，而对较低风险案件几乎没有影响，并且降低了后续的虐待转介。我们未发现该评分扩大了决策或结果中种族差异的证据，这表明算法可以在保留人工裁量权的同时改善资源投放的精准性。
Abstract
Many public-sector decisions require allocating scarce attention under uncertainty. We examine whether algorithmic risk assessments improve child-protection decisions, where supervisors decide which cases need closer scrutiny. In a randomized evaluation of 4,752 child referrals over 14 months in Northampton County, supervisors received an algorithmic risk score alongside standard case records. Access to the score increased foster-care placements and services for children at highest predicted risk, with little change for lower-risk cases, and it reduced subsequent maltreatment referrals. We find no evidence that the score widened racial disparities in decisions or outcomes, suggesting algorithms can improve targeting while preserving human discretion.