📈 economics

Network Beliefs and Behavior with Peer Effects

This paper introduces a unified framework called "Iterative Belief Centrality" to demonstrate how individuals' heterogeneous and correlated beliefs about incomplete network structures systematically shape peer-driven behavior, amplifying action differences between more and less connected agents compared to models assuming complete information or homogeneous beliefs.

Original authors: Promit K. Chaudhuri, Matthew O. Jackson, Sudipta Sarangi, Hector Tzavellas

Published 2026-06-08

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Promit K. Chaudhuri, Matthew O. Jackson, Sudipta Sarangi, Hector Tzavellas

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 at a massive, chaotic party. You want to know how loud to shout to get your point across. Do you shout softly, or do you scream?

In the real world, you don't know the whole layout of the party. You only see the people standing right next to you. You have to guess what the rest of the room looks like based on your tiny slice of it.

This paper, "Network Beliefs and Behavior with Peer Effects," by Chaudhuri, Jackson, Sarangi, and Tzavellas, is a mathematical story about exactly this situation. It explains how our guesses about who is connected to whom shape our actions, often more than the actual connections do.

Here is the breakdown of their findings using simple analogies.

1. The "Echo Chamber" of Guesses

Most old theories assumed one of two things:

The Omniscient Guest: Everyone knows the exact map of the whole party (Complete Information).
The Clueless Guest: Everyone is equally confused and guesses the same thing about the rest of the party (Homogeneous Beliefs).

The authors say: "Neither of these is true." In reality, you are a Partial Guest. You see your immediate neighbors, and you use that to guess the rest. But here's the twist: You also guess what they are guessing.

The Analogy: Imagine you see three people talking loudly next to you. You think, "Wow, this party is rowdy!" So, you shout louder.
The Twist: You also realize that those three people are looking at their neighbors. If their neighbors are also loud, they will think the party is even rowdier than you do.
The Result: You aren't just reacting to the people next to you; you are reacting to your guess of what they are guessing. This creates a "belief echo" that bounces back and forth.

2. The "Iterative Belief Centrality"

The authors give a fancy name to this process: Iterative Belief Centrality.

Think of it like a game of "Telephone," but instead of distorting a message, the message is about how influential you are.

In a perfect world, your influence is measured by how many friends you have and how connected they are (this is called Katz-Bonacich centrality).
In this paper's world, your influence is measured by how many friends you think you have, and how connected you think they are.

Because everyone is making their own unique guesses based on their own tiny slice of the party, everyone ends up with a different "Centrality Score," even if they are standing in the same spot.

3. The "Rich Get Richer" (and the Poor Get Poorer)

The paper finds a surprising pattern: Your local view amplifies your behavior.

The High-Degree Agent (The Popular Person): Imagine you have 10 friends at the party. You look around and think, "Wow, I have so many friends! The whole party must be super connected." You assume everyone else is also very popular. So, you act more aggressively (shout louder, invest more, participate more).
The Low-Degree Agent (The Wallflower): Imagine you only have 1 friend. You look around and think, "This is a quiet, sparse party. Nobody is really connected." You assume everyone else is also isolated. So, you act less aggressively.

The Big Surprise: The gap between the "Popular" person and the "Wallflower" becomes wider than it would be if everyone knew the true map of the party.

If everyone knew the truth, the Popular person would just be slightly more active.
But because they believe the whole world is connected, they go overboard.
Because the Wallflower believes the world is disconnected, they underperform.

The authors call this amplification. Your local view doesn't just inform you; it distorts your reality in a way that makes the rich (in connections) act even richer, and the poor act even poorer.

4. The "Mixing Rate" (How Fast Do We Forget?)

The paper introduces a concept called the Mixing Rate of the belief system.

Fast Mixing: Imagine the party is a giant, open room where everyone is mixed up. If you look at your neighbors, you quickly realize, "Oh, they are just like the average person." Your specific local view gets "washed out" quickly. Everyone ends up acting similarly.
Slow Mixing: Imagine the party is divided into tight-knit cliques. If you are in a clique of loud people, you stay in that bubble. You keep thinking, "Everyone here is loud!" and you keep shouting. Your local view sticks with you for a long time.

The authors show that Slow Mixing leads to huge differences in behavior. If people stay in their "belief bubbles," the differences between the loud and the quiet become extreme.

5. Why This Matters (According to the Paper)

The paper explains several real-world weirdnesses without needing to say people are "irrational":

The Virality Paradox: Why do some posts go viral while others don't? It's not just the content; it's that people with many connections think the network is dense and viral, so they share more. People with few connections think it's sparse, so they don't.
Political Participation: If you think your friends are all active in a movement, you think the whole country is active, so you join in. If you think your friends are quiet, you think the whole country is quiet, so you stay home.

Summary

The paper argues that we don't act based on the network we are in; we act based on the network we imagine we are in.

Because we all have different views of the world, we all imagine different networks. These imagined networks create a feedback loop where our guesses about others' guesses make us act more extremely than we would if we just knew the facts. The more "sticky" our local views are (slow mixing), the more extreme our behavior becomes.

It's not that we are crazy; it's that we are rational people trying to guess the shape of a giant puzzle while only holding a single piece. And sometimes, that single piece makes us see a dragon where there is only a bird.

Technical Summary: Network Beliefs and Behavior with Peer Effects

Problem Statement
Individuals often act within social networks without possessing full knowledge of the network's global structure. While existing literature on network games typically assumes either complete information (where agents know the entire adjacency matrix) or homogeneous incomplete information (where agents know their own degree but share identical beliefs about the rest of the network), these frameworks fail to capture the systematic heterogeneity in beliefs that arises from local observations. Empirical evidence suggests that individuals hold subjective, often divergent, representations of network connectivity (e.g., Krackhardt, 1987; Kilduff et al., 2008), which significantly influence behaviors such as leadership, political participation, and online virality. The paper addresses the gap in formalizing how these heterogeneous, position-dependent beliefs about network structure shape equilibrium behavior in peer-interaction games.

Methodology
The authors develop a game-theoretic framework based on the Harsanyi (1967) approach to incomplete information, applied to a linear-quadratic network game (Ballester et al., 2006).

Information Structure: A network $g$ is drawn from a commonly known prior distribution. Agents observe partial information about the network (e.g., their own direct links, or potentially neighbors' links) and update their beliefs using Bayes' rule. The model allows for arbitrary priors over networks, not just degree distributions.
Belief Hierarchy: Agents form beliefs not only about the network structure but also about what other agents know and believe. This creates a hierarchy of higher-order beliefs (beliefs about beliefs).
Equilibrium Concept: The authors analyze the Bayesian Nash Equilibrium (BNE) in pure strategies. They establish existence and uniqueness under standard conditions on the strategic complementarity/substitutability parameter $\lambda$ relative to the maximum degree.
Key Analytical Tool: The equilibrium is characterized via an infinite expansion over expected walks in the network. This leads to the definition of Iterative Belief Centrality, a measure of perceived influence that aggregates expected strategic spillovers across all distances, weighted by the probability of encountering specific types of agents and their subsequent beliefs.

Key Contributions and Results

Iterative Belief Centrality: The paper characterizes equilibrium actions as a function of Iterative Belief Centrality. Unlike the Katz–Bonacich centrality in complete-information settings (which depends on realized walks), this metric depends on expected walks derived from the agent's local observations and their inferences about others' inferences.
- Result: Under complete information, belief iteration collapses, and Iterative Belief Centrality coincides with Katz–Bonacich centrality. Under symmetric incomplete information, it collapses to a degree-based statistic. In intermediate cases, it generates heterogeneous behaviors driven by belief hierarchies.
Amplification of Behavioral Differences: The authors demonstrate that rational Bayesian updating from local observations can lead to systematic and amplified behavioral differences.
- Mechanism: Under "Row-Based Dominance" (a condition where higher observed connectivity shifts beliefs toward higher connectivity throughout the belief hierarchy), agents with higher realized degrees expect stronger spillovers not just directly, but at every distance in the belief chain.
- Result: In games with strategic complements, high-degree agents exert higher actions and low-degree agents exert lower actions relative to homogeneous-belief benchmarks. The gap between high and low-degree behavior is larger than what is predicted by models assuming homogeneous beliefs or complete information.
Microfoundation for Degree-Based Models: The paper provides conditions under which the complex Iterative Belief Centrality collapses to a simple degree-based representation.
- Condition: If agents observe only their direct neighbors and the prior over networks is "isomorphic" (invariant to node relabeling), then agents with the same degree possess isomorphic types.
- Result: In this specific environment, degree becomes a sufficient statistic for behavior. This offers a microfoundation for existing degree-based models, showing they emerge as a special case of a more general information structure where belief hierarchies collapse due to symmetry.
Role of Belief Correlation and Mixing: The paper introduces a belief transition matrix ( $P$ ) that describes how agents of a given degree believe they are connected to others of various degrees.
- Assortativity: The authors define path-based assortativity measures to quantify how degree information persists along belief paths.
- Mixing Rate: The spectral properties of the belief transition matrix (specifically the second-largest eigenvalue $\rho^*$ ) govern the persistence of belief heterogeneity.
- Result: Slow mixing (high $\rho^*$ ) allows belief differences to echo and amplify, increasing both the mean and variance of equilibrium actions. Fast mixing causes beliefs to converge rapidly to the stationary distribution, yielding outcomes closer to the homogeneous benchmark. The paper shows that heterogeneous beliefs can lead to non-monotonic actions in degree and increased dispersion in aggregate outcomes, even when the underlying degree distribution is fixed.

Significance and Claims
The paper claims to provide a tractable theory of peer-driven network behavior under incomplete information that unifies complete-information and degree-based models as special cases. Its primary significance lies in demonstrating that:

Rationality does not imply homogeneity: Even fully rational agents with identical priors can form heterogeneous beliefs and exhibit divergent behaviors solely due to differences in their local network positions.
Beliefs amplify structural effects: The interaction between local observations and higher-order beliefs can amplify behavioral differences beyond what is explained by network position alone.
Microfoundations for existing models: It explains why degree-based models work (symmetry collapsing belief hierarchies) while also identifying the conditions under which they fail (heterogeneous priors or asymmetric information).
Policy and Empirical Implications: The framework offers a theoretical mechanism for empirical phenomena such as the "friendship paradox," the "virality paradox," and variations in civic participation, suggesting that perceived influence (shaped by belief hierarchies) is a key determinant of behavior, distinct from actual structural influence.

The authors maintain a modest scope, focusing on the linear-quadratic formulation to ensure tractability, but note that the logic of Iterative Belief Centrality extends to broader network games with strategic spillovers. They do not claim to solve all network uncertainty problems but rather to establish a general framework for how local information propagates into global behavioral patterns.