Working Paper: NBER ID: w21387
Authors: Jonathan D. Ketcham; Nicolai V. Kuminoff; Christopher A. Powers
Abstract: Neoclassical and psychological models of consumer behavior often make divergent predictions for the welfare effects of paternalistic policies, leaving wide scope for researchers’ choice of a model to influence their policy conclusions. We develop a framework to reduce this model uncertainty and apply it to administrative data on consumer decision making in Medicare Part D. Consumers’ choices for prescription drug insurance plans can be explained by Abaluck and Gruber’s (AER 2011) model of utility maximization with psychological biases or by a neoclassical version of their model that precludes such biases. We evaluate these competing hypotheses using nonparametric tests of utility maximization and a trio of model validation tests. We find that 79% of enrollment decisions in Medicare Part D from 2006-2010 satisfied basic axioms of consumer preference theory under the assumptions of full information, zero transaction cost, and no measurement error. The validation tests provide evidence against widespread psychological biases. In particular, we find that precluding psychological biases improves the structural model’s out-of-sample predictions for consumer behavior.
Keywords: Consumer Decision Making; Medicare Part D; Utility Maximization; Paternalistic Policies
JEL Codes: D12; I11; I38
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Consumer choices can be evaluated against a utility function (D11) | Consumer decision-making quality in Medicare Part D (I11) |
| Deviations from expected utility maximization (D81) | Genuine consumer mistakes or model misspecifications (C50) |
| Precluding psychological biases (D91) | Enhances predictive power of structural models (C51) |
| Observed deviations from expected utility maximization (D81) | Not predominantly due to consumer mistakes but rather model specifications (L68) |
| Evidence of choice inconsistencies (D91) | Not robust when alternative specifications are considered (C59) |
| Structural model's out-of-sample predictions (C52) | Improved when psychological biases are excluded (D91) |