Imprecision Attenuates Updating

Martin Vaeth · 2025

Read Imprecision Attenuates Updating by Martin Vaeth (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-02238.txt.

Introduction

Martin Vaeth shows how fuzziness in information pulls beliefs toward what was already thought. His paper appeared in 2025. When people update on a noisy signal, the posterior mean lands between the prior belief and the signal. Vaeth asks when one noise source counts as more precise than another. He defines a precision order through exactly this pull: one noise wins if its posterior always lands closer to the signal, across all signal values and a broad family of priors.

The characterization is clean and complete for symmetric location experiments with well-shaped noise. He then applies the order to three staples of information economics: how prior precision interacts with learning, the value of information, and voting. The paper gives those comparative statics a single rigorous footing.

Catalog record

CREATORMartin Vaeth
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-02238

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Imprecision Attenuates Updating Martin Vaeth Abstract This paper studies how imprecision in noisy signals attenuates Bayesian posterior means toward the prior mean, a property underlying many comparative-statics results in information economics. We introduce the precision order, characterized by the attenuation effect across symmetric location experiments: noise density g ~ \tilde{g} is more precise than g g if and only if the posterior mean under g ~ \tilde{g} is closer to the signal than under g g for all signal realizations and all symmetric, log-concave priors. We apply the precision order to derive comparative statics for prior precision, the value of information, and voting. Keywords: signal extraction, underreaction, information orders, comparative statics, cognitive imprecision JEL Classifications: C11, D83, D84 I am grateful to Xiaosheng Mu for his advice on this project. For helpful discussions I thank Roland Bénabou, Christopher Chambers, Martin Cripps, Maxime Cugnon de Sévricourt, Loren Fryxell, Benjamin Enke, Faruk Gul, Navin Kartik, Alessandro Lizzeri, Pietro Ortoleva, Wolfgang Pesendorfer, and Fedor Sandomirskiy. I gratefully acknowledge financial support from the William S. Dietrich II Economic Theory Center. All errors are my own. 1 Introduction In many contexts, agents’ responses to fundamentals are attenuated: consumers underreact to non-salient taxes ( Chetty, Looney, and Kroft, 2009 ), numerical estimates are biased toward reference values ( Tversky and Kahneman, 1974 ), and, beyond economics, perceptual judgments regress toward the mean of the stimulus distribution ( Hollingworth, 1910 ; Stevens and Greenbaum, 1966 ). Such attenuated behavior is often explained through imprecise information. In the standard account, agents base their choices on noisy signals of the fundamentals, where the noise reflects imperfect observation of the state. A more recent literature on cognitive imprecision locates the noise in cognition instead: even when agents observe the fundamentals, the cognitive process mapping them to actions is noisy ( Woodford, 2020 ; Enke and Graeber, 2023 ; Ilut and Valchev, 2023 ). Both versions of the explanation share a core mechanism: noise attenuates Bayesian updating, compressing behavior toward a default action. The imprecise-information explanation is usually formalized with the normal-normal model. The state x is drawn from a normal prior with mean \mu , and the agent observes a signal s=x+\varepsilon with independent normal noise \varepsilon . The posterior mean lies between the signal and the prior mean, and is farther from the signal for less precise signals (larger noise variance). 1 1 1 If x\sim\mathcal{N}(\mu,\sigma_{x}^{2}) and \varepsilon\sim\mathcal{N}(0,\sigma_{\varepsilon}^{2}) independent of x , then \mathbb{E}[x\,|\,s]=\lambda\,s+\left(1-\lambda\right)\mu with \lambda=\sigma_{x}^{2}/(\sigma_{x}^{2}+\sigma_{\varepsilon}^{2}) . For any signal realization s , larger \sigma_{\varepsilon}^{2} moves the posterior mean farther from the signal. We call the comparative static in signal precision imprecision attenuates updating , or attenuation effect for short. The formalization is usually completed by assuming actions track the posterior mean, under which the attenuation effect carries over to behavior. A key piece of evidence for cognitive imprecision rests on the attenuation effect: subjects reporting higher cognitive uncertainty exhibit more attenuated behavioral responses ( Enke and Graeber, 2023 ; Enke, Graeber, Oprea, and Yang, 2024 ). Despite the theoretical and empirical relevance of the attenuation effect, it was unknown whether it extends beyond the normal distribution. Normality can be justified in particular cases, through the central limit theorem or in rational inattention with a normal prior and quadratic loss, but these cases are limited. Whether the effect is special to the normal distribution or holds more broadly therefore bears on how far the empirical evidence can be interpreted as a consequence of imprecise information. We characterize exactly when the attenuation effect holds and which assumptions drive it. We prove the attenuation effect holds in all additive-noise models (location experiments) with symmetric noise under a symmetric, log-concave prior: for any signal realization, the posterior mean moves toward the signal realization as the signal becomes more precise (Theorem 1 ). The conditions are nonparametric, replacing the normal functional form with two shape restrictions, symmetry and log-concavity. To state the result we introduce the precision order . Noise density \tilde{g} is more precise than noise density g if the likelihood ratio \tilde{g}(x)/g(x) decreases as x moves away from 0, placing the absolute noise term of the more precise signal closer to 0 in the sense of likelihood-ratio domination (see Figure 1 ). Scaling up a log-concave noise term produces a less precise signal under this definition. The precision order is not only sufficient for the attenuation effect, but also necessary, so the effect characterizes the order. If noise density \tilde{g} is not more precise than noise density g , then for some symmetric, log-concave prior f and some signal realization, \tilde{g} induces more attenuation than g . Moreover, the attenuation effect does not hold in general if we drop symmetry or log-concavity of the prior. Our main result extends the attenuation effect to a nonparametric class of distributions and pins down which distributional assumptions drive the effect. An essential assumption is symmetry, which is already needed for the more basic property that the posterior mean lies between the prior mean and the signal. Chambers and Healy (2012) show that without symmetric prior and noise densities, the posterior mean can easily fall outside this interval, so agents overreact or update in the opposite direction. More generally, the normal-normal model is pervasive in economics and is often used for comparative statics in signal precision. We show how our main result generates such comparative statics without parametric assumptions through several applications. We establish comparative statics for the median posterior mean conditional on the state, with an application to voting. 2 2 2 The Online Appendix includes comparative statics of the average posterior mean conditional on the state. We also study how prior precision, rather than signal precision, shapes updating. A more precise prior increases attenuation and lowers the value of information, the opposite of a more precise signal. Related Literature This paper introduces a new information order and provides comparative statics for signal structures with additive independent noise, called location experiments. Prior work has provided orderings on location experiments related to the value of information. Boll (1955) shows one location experiment Blackwell-dominates another if and only if the latter can be derived by adding independent noise to the former. This condition is very restrictive, making location experiments comparable only in rare cases. In response, Lehmann (1988) introduced the less restrictive Lehmann order to rank location experiments by their value in all monotone decision problems. We show the precision order strengthens the Lehmann order and neither implies nor is implied by the Blackwell order. Other comparative statics results have been established for location experiments with log-concave noise distributions, which satisfy the monotone likelihood-ratio property. Milgrom (1981) shows the strict monotone likelihood-ratio property is equivalent to higher signals being “good news,” in the sense of yielding first-order stochastically dominant posteriors, for any prior. Conversely, Dawid (1973) examines violations of the “good news”-property under heavy-tailed noise distributions: if the noise is sufficiently heavy-tailed relative to the prior, extreme observations are “rejected,” yielding posteriors that revert to the prior. Together with our results, this highlights log-concave location-scale experiments are a well-behaved class of signal structures that retains key properties of the normal-normal model. Other work studies properties of the distribution of posterior means induced by signal structures, with

Read the complete text