当预言机误导：风险评估工具中使用可观测结果而非潜在结果的后果建模
统计方法学
2021-04-06 v1
摘要
风险评估工具（RAIs）被广泛用于预测医疗和刑事司法等领域的不良结果。RAIs通常在观测数据上训练，并优化以预测可观测结果而非潜在结果，后者是指若不存在特定干预时将会发生的结果。相关的潜在结果示例包括：若不加治疗患者病情是否会恶化，或被告若在审前获释是否会再犯。我们说明了训练用于预测可观测结果的RAIs如何导致更差的决策，恰好造成其旨在预防的那类危害。即便预测变量是贝叶斯最优的且不存在未测量的混杂，这种情况也可能发生。
引用
@article{arxiv.2104.01921,
title = {When the Oracle Misleads: Modeling the Consequences of Using Observable Rather than Potential Outcomes in Risk Assessment Instruments},
author = {Alan Mishler and Niccolò Dalmasso},
journal= {arXiv preprint arXiv:2104.01921},
year = {2021}
}
备注
6 pages, 3 figures. Presented at the workshop "'Do the right thing': machine learning and causal inference for improved decision making," NeurIPS 2019