Bayesian Rapid Optimal Adaptive Design (BROAD): Method and application distinguishing models of risky choice

Abstract

Economic surveys and experiments usually present fixed questions to respondents. Rapid computation now allows adaptively optimized questions, based on previous responses, to maximize expected information. We describe a novel method of this type introduced in computer science, and apply it experimentally to six theories of risky choice. The EC² method creates equivalence classes, each consisting of a true theory and its noisy-response perturbations, and chooses questions with the goal of distinguishing between equivalence classes by cutting edges connecting them. The edge-cutting information measure is adaptively submodular, which enables a provable performance bound and lazy evaluation which saves computation. The experimental data show that most subjects, making only 30 choices, can be reliably classified as choosing according to EV or two variants of prospect theory. We also consider whether subjects should and could manipulate by misreporting preferences, and find little evidence of manipulation.