Scholarlink

<p>The theory of rational inattention posits that individuals maximize expected utility net of information costs, which are often modeled using mutual information. A key prediction of this framework is that posterior beliefs are invariant to prior probabilities. We test this prediction using a novel auditory-information experiment. Our results reveal a clear violation of this prediction: under low-stakes conditions, we observe a simple behavioral asymmetry that is inconsistent with invariance. We then introduce a reduced-form hybrid specification, defined as a convex combination of Bayesian updating–based mutual-information costs and hypothesis testing–based log-likelihood-ratio costs, to quantify the relative empirical contributions of the two components. Within this hybrid specification, the estimated marginal cost of Bayesian learning is positively related to the marginal cost of Type I decision errors, and the estimated weight on the mutual-information component is approximately 0.60 for the modal group of subjects.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Rational inattention: The interplay of stakes and prior beliefs in a laboratory study

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Zexuan Wang

摘要

The theory of rational inattention posits that individuals maximize expected utility net of information costs, which are often modeled using mutual information. A key prediction of this framework is that posterior beliefs are invariant to prior probabilities. We test this prediction using a novel auditory-information experiment. Our results reveal a clear violation of this prediction: under low-stakes conditions, we observe a simple behavioral asymmetry that is inconsistent with invariance. We then introduce a reduced-form hybrid specification, defined as a convex combination of Bayesian updating–based mutual-information costs and hypothesis testing–based log-likelihood-ratio costs, to quantify the relative empirical contributions of the two components. Within this hybrid specification, the estimated marginal cost of Bayesian learning is positively related to the marginal cost of Type I decision errors, and the estimated weight on the mutual-information component is approximately 0.60 for the modal group of subjects.