A1788
Title: Asymptotics for treatment decisions with partial identification
Authors: Jose Luis Montiel Olea - Cornell University (United States)
Chen Qiu - Cornell University (United States) [presenting]
Joerg Stoye - Cornell University (United States)
Abstract: A common approach in the study of treatment choice problems with partial identification assumes a normal location shift model and seeks minimax regret optimal decisions by jointly accounting for point-identified and set-identified parameters. Within this framework, substantial progress has been made when the parameter space is convex and centrosymmetric. Yet, in practice, data may not be normal and the parameter space may not be convex or centrosymmetric. A new, practical and scalable asymptotic approach to treatment decisions with partial identification is proposed. This approach justifies the common approach as in fact providing local asymptotic optimality results when the parameter space is convex and centrosymmetric, and can accommodate a large class of models even beyond convexity or centrosymmetry. The approach is also computationally cheap, offering an effective way for practitioners to find large-sample optimal decisions in many models.
