主题: Statistical vs. Operational Model Selection: The Value of Model Misspecification in Algorithmic Hiring

中文题目：统计与运营视角下的模型选择：模型错设在算法招聘中的价值

时间: 2025年12月5日 上午 9:00-10:00

地点: 管理科研楼一楼第一教室

主讲人: Dr. Peng Xuefeng, Postdoctoral Fellow, Business School, University of Hong Kong

Bio:

Xuefeng Peng has been a Postdoctoral Fellow at the University of Hong Kong Business School since 2024. He received his Ph.D. in Management Science and Engineering from the University of Science and Technology of China, with a half-year study & RA at the University of Toronto. He focuses the intersection of OM and Marketing, with a close eye on AI-driven decision making. He has papers in POM, DSS, JORS, and revisions at MS and POM.

Abstract:

Algorithmic screening plays an increasingly central role in modern recruitment, yet how firms should select model complexity to improve hiring outcomes remains unclear. This study reframes model selection as a strategic design choice shaped by two foundational trade-offs: the classical bias–variance trade-off from statistical learning and a benefit–penalty trade-off that links classification outcomes—true and false positives and negatives—to operational payoffs. We develop a theoretical framework in which firms estimate candidate expected quality using models that may be overspecified (including non-casual features) or underspecified (excluding causal ones) relative to the ground-truth model. We show that overspecification can improve outcomes in moderately selective environments by amplifying quality dispersion and increasing true positives, while underspecification can reduce false positives and false negatives in high-selectivity settings by acting as a conservative filter. We further identify conditions under which a simpler or more complex model outperforms the other regardless of the ground-truth specification. Specifically, a simpler model is preferred when misclassification costs are high; a more complex model is preferred when the benefit of accurate identification of qualified candidates is high; and the correctly specified model is optimal when both misclassification costs and the benefit of accurate identification of qualified candidates are high. These findings overturn the classical presumption that correct specification is always optimal and demonstrate that model complexity should be treated as a context-sensitive design choice—one that links statistical estimation to operational payoff.