📈 economics

Imprecision Attenuates Updating

This paper introduces a "precision order" for noisy signals, defined by the attenuation of Bayesian posterior means toward the prior, to establish comparative statics regarding prior precision, the value of information, and voting outcomes.

Original authors: Martin Vaeth

Published 2026-07-28

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Martin Vaeth

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess the temperature outside, but you can't see the thermometer. You have a "gut feeling" based on the season (your prior belief), and then you get a single, fuzzy reading from a broken thermometer (a noisy signal). In the world of economics and psychology, this is a classic puzzle: how do we combine what we already think we know with new, imperfect information? Usually, we assume our brains act like perfect statisticians, blending the old guess and the new clue to find a middle ground. But real life is messy. Sometimes we ignore the new clue too much, and sometimes we overreact. Scientists call this "updating," and they've long wondered why our brains often "underreact," pulling our final guess too close to our original hunch instead of trusting the new data enough.

This paper dives into that exact mystery, but it swaps the usual math-heavy assumptions for a more flexible, rule-based approach. Instead of assuming our brains or the world always follow a perfect "bell curve" (the normal distribution), the author asks: what are the absolute minimum rules needed for this "underreaction" to happen? The paper introduces a new way to rank how "precise" a piece of information is, not just by how small the error is, but by how the errors are shaped. It turns out that for our brains to consistently underreact to noisy signals, two things must be true: the noise must be symmetrical (errors are just as likely to be too high as too low), and the shape of our beliefs must be "log-concave" (a fancy way of saying our confidence peaks in the middle and tapers off smoothly, like a hill, rather than having weird bumps or flat plateaus).

The Core Discovery: The "Precision Order"

The main finding of this paper is a mathematical proof that confirms a specific intuition: imprecise information always pulls our updated beliefs closer to our original guess. The author calls this the "attenuation effect." Think of it like a rubber band. Your original belief is one anchor, and the new signal is the other. If the signal is fuzzy (imprecise), the rubber band is loose, and your final answer stays close to your original anchor. If the signal is sharp (precise), the rubber band is tight, and your answer snaps closer to the new signal.

The paper proves that this rubber-band behavior holds true for any situation where the noise is symmetrical and the beliefs are shaped like a smooth hill (log-concave). It doesn't matter if the noise looks like a bell curve, a triangle, or a box; as long as it's symmetrical and the beliefs are smooth, a "more precise" signal will always move your final guess closer to the truth than a "less precise" one. The author defines a new ranking system called the precision order to measure this. If one noise pattern is "more precise" than another, it means its errors are clustered tighter around zero in a very specific mathematical way. The paper shows that this ranking is the exact condition needed for the attenuation effect to work. Crucially, the paper proves that symmetry is not just helpful, but strictly necessary: if the noise isn't symmetrical, the reliable "underreaction" rule breaks down completely across all possible signals, and the brain might overreact or update in the wrong direction entirely. Similarly, if your beliefs aren't smooth (log-concave), the math doesn't hold up. The paper also clarifies that this new "precision order" is different from older, stricter ways of ranking information (like Blackwell domination). You can have a signal that is "better" in the old sense but doesn't trigger this specific underreaction, and vice versa.

What the Paper Rules Out

It is crucial to note what this paper says doesn't work. The author explicitly rules out the idea that this effect is a special trick that only happens with the famous "bell curve" (normal distribution). While the bell curve is a common tool in economics, this paper proves the effect is much more robust and applies to a wide variety of shapes. However, the paper also draws a hard line: symmetry is non-negotiable. If the noise is lopsided (asymmetrical), the neat "underreaction" rule vanishes. You can't just assume the brain will underreact if the errors are biased one way. The paper also clarifies that this new "precision order" is different from older, stricter ways of ranking information (like Blackwell domination). You can have a signal that is "better" in the old sense but doesn't trigger this specific underreaction, and vice versa.

Real-World Implications: From Voting to Voting

The paper doesn't just stay in the realm of abstract math; it applies these findings to real-world scenarios like voting and economic forecasting.

Voting: Imagine a group of people voting on a policy. If the voters are unsure and their information is fuzzy, they tend to stick closer to their "default" political leaning rather than shifting toward the true state of the world. The paper shows that if the voters perceive their information as more precise (even if it's just a feeling of confidence), they will make fewer mistakes. They will shift their votes closer to the actual truth, reducing the chance that the group picks the wrong candidate just because everyone was too cautious.
Value of Information: The paper also flips the script on how we value information. Usually, we think having a better prior (a smarter starting guess) is always good. But here, the author finds that if you start with a very precise prior, you actually become less responsive to new signals. Your "rubber band" is already so tight to your original guess that new information barely moves you. This means that in some cases, having a super-confident starting point can actually lower the value of new information because you stop listening to it.

How Sure Are We?

The author is extremely confident in these results because they are mathematically proven, not just guessed or simulated. The paper provides rigorous theorems that show the "attenuation effect" is a necessary and sufficient condition for the precision order under the stated assumptions. This means the result isn't a lucky guess; it's a fundamental property of how these specific types of information systems work. The paper doesn't just suggest this might happen; it proves that if you have symmetrical, log-concave noise, the underreaction must happen, and if you don't have those properties (especially symmetry), the underreaction cannot be guaranteed.

In short, this paper gives us a new, more flexible rulebook for understanding why we sometimes ignore new evidence. It tells us that as long as our errors are balanced and our beliefs are smooth, the more precise the news, the more we'll listen to it. But if the news is lopsided or our beliefs are weirdly shaped, that reliable pattern breaks, and our brains might do something much stranger.

Technical Summary: Imprecision Attenuates Updating

Problem and Motivation
In information economics and cognitive science, agents often exhibit "attenuated" responses to fundamentals: their actions or estimates regress toward a default value (such as a prior mean or reference point) rather than fully tracking new signals. While the standard normal-normal model formalizes this via the "attenuation effect"—where a less precise signal (higher noise variance) pulls the posterior mean closer to the prior mean—this result relies heavily on parametric assumptions. The paper addresses a critical gap: it is unknown whether this attenuation effect holds beyond the normal distribution. If the effect is specific to normality, empirical evidence of underreaction (e.g., in cognitive uncertainty studies) cannot be robustly interpreted as a consequence of imprecise information. The paper seeks to characterize exactly when the attenuation effect holds and identify the minimal distributional assumptions required.

Methodology and Framework
The paper studies location experiments, a class of Blackwell experiments where the signal $s$ equals the state $x$ plus independent noise $\varepsilon$ ( $s = x + \varepsilon$ ). The agent holds a prior density $f$ and observes a signal to form a posterior mean $m(s, f, g)$ , where $g$ is the noise density.

To generalize the attenuation effect, the author introduces the Precision Order:

Definition: A symmetric noise density $\tilde{g}$ is "more precise" than $g$ if the likelihood ratio $\tilde{g}(x)/g(x)$ is non-increasing for $x > 0$ (assuming symmetry around 0). This implies that the absolute noise term of the more precise signal is closer to 0 in the sense of likelihood-ratio domination.
Assumptions: The analysis relies on two shape restrictions rather than parametric forms:
1. Symmetry: Both the prior $f$ and noise $g$ must be symmetric.
2. Log-concavity: The prior $f$ must be log-concave (which implies quasi-concavity).

The core technical challenge addressed is that a more precise noise density has two opposing effects on the posterior mean: it upweights values closer to the signal on one side of the signal but downweights them on the other. The proof overcomes this by transforming the posterior mean integral using the symmetry of the noise, recasting the problem into an integration of a monotone function against the absolute noise term, which the precision order ranks via likelihood-ratio domination.

Key Results
The paper's central contribution is Theorem 1, which establishes an equivalence between the precision order and the attenuation effect:

Sufficiency: If $\tilde{g}$ is more precise than $g$ (under the precision order) and the prior $f$ is symmetric and log-concave, then for every signal realization $s$ , the posterior mean under $\tilde{g}$ lies strictly closer to $s$ than the posterior mean under $g$ . Specifically, if $s \geq \mu_f$ , then $s \geq m(s, f, \tilde{g}) \geq m(s, f, g) \geq \mu_f$ .
Necessity: The precision order is necessary. If $\tilde{g}$ is not more precise than $g$ , there exists a symmetric, log-concave prior and a signal realization where the attenuation ordering fails (i.e., the more "precise" noise induces more attenuation).
Role of Assumptions:
- Symmetry: The paper proves that symmetry is necessary for the attenuation effect to hold robustly across all signals. Without symmetry, the posterior mean can fall outside the interval between the signal and the prior mean, leading to overreaction or updating in the wrong direction.
- Log-concavity: The prior must be log-concave; quasi-concavity alone is insufficient to guarantee the result.
- Scaling: The paper shows that scaling up a log-concave noise term (increasing variance) always results in a less precise signal under this order, covering distributions like Normal, Logistic, Laplace, and Uniform.

Applications and Comparative Statics
The paper applies the main theorem to derive non-parametric comparative statics in three areas:

Median Posterior Mean and Voting: Under symmetric, strictly log-concave noise, the median posterior mean equals the posterior mean evaluated at the true state. The paper applies this to common-interest voting models. It demonstrates that a more precise noise density shrinks the set of states where the electorate elects the alternative closer to the prior mean rather than the true state, thereby improving information aggregation.
Prior Precision: Using a duality property ( $m(s, f, g) = s - m(s, g, f)$ ), the paper shows that making the prior more precise has the opposite effect of making the signal more precise: a more precise prior increases attenuation, pulling the posterior mean closer to the prior mean.
Value of Information: In mean-measurable decision problems, a more precise prior lowers the value of information for any location experiment with symmetric, log-concave noise. This contrasts with standard results where signal precision increases value; here, a more informative prior reduces the marginal benefit of acquiring new signals.

Significance and Claims
The paper claims to provide a non-parametric foundation for the "imprecise information" explanation of attenuated behavior. By showing that the attenuation effect holds for a broad class of distributions defined only by symmetry and log-concavity, the paper argues that empirical observations of underreaction can be robustly interpreted as a consequence of imprecise information, without requiring the restrictive assumption of normality.

Furthermore, the paper introduces the Precision Order as a new information ordering that strengthens the Lehmann order (for log-concave experiments) but is distinct from the Blackwell order. It clarifies that while Blackwell domination is too restrictive for location experiments (requiring noise to be a sum of independent variables), the Precision Order captures the intuitive notion of "less noise" relevant for Bayesian updating in a wider range of economic contexts. The results pin down symmetry and log-concavity as the specific drivers of the attenuation effect, offering a precise characterization of when and why Bayesian updating compresses behavior toward a default.

More like this