Group Wisdom at a Price: Jury Theorems with Costly Information

Matteo Michelini, Adrian Haret, Davide GrossiIJCAI 2022 · DOI

TL;DR

This paper investigates the Condorcet Jury Theorem (CJT) in a more realistic setting where voters incur a cost (effort) to acquire the information needed to vote competently. They find that this cost can lead to equilibria where the "wisdom of crowds" effect breaks down, meaning larger groups don't necessarily lead to better decisions. However, they also show that social network structures can sometimes mitigate this loss of wisdom.

What Problem Does This Paper Address?

The Condorcet Jury Theorem (CJT) is a foundational result in social choice theory, suggesting that a group of independent voters, each slightly better than random at identifying the truth, can achieve near-perfect collective decision-making through majority voting, especially as the group size grows. This theorem has been used to justify democratic processes and collective decision-making. However, the CJT relies on strong assumptions, including that voters' competence is readily available without cost.

This paper challenges this assumption by introducing the concept of "costly information." In their model, voters must expend effort to become competent enough to vote. This effort is a disincentive, and voters must balance the desire for a correct group decision against the personal cost of acquiring the necessary information. The core problem is to understand how this cost of information affects the "wisdom of crowds" phenomenon and whether the optimistic conclusions of the CJT still hold when voting is not frictionless.

Methods

The authors model the scenario as a game where agents choose between voting (incurring effort and gaining competence) and abstaining (no effort, no competence). They analyze this game by looking for Nash Equilibria (NE), where no single agent can improve their outcome by unilaterally changing their strategy. They also consider Strong Nash Equilibria (SNE), where no coalition of agents can benefit from deviating.

They explore four variations of the basic model:

  1. Uniform Stakes and Competence: All agents have the same probability of being correct and value the outcome equally.
  2. Heterogeneous Stakes: Agents value the correct outcome differently.
  3. Heterogeneous Competences: Agents have different probabilities of being correct.
  4. Social Networks: Agents' awareness of others' actions is limited to their neighbors in a graph, rather than the entire group.

The core of their analysis involves calculating the probability of a correct group decision (group accuracy) given a voting profile and then determining the conditions under which agents are incentivized to vote or abstain, considering the effort cost.

Datasets

Not discussed in the paper. The paper is a theoretical model and does not use empirical datasets for training or evaluation.

Key Insights and Ideas

  • Effort is a Double-Edged Sword: While effort is necessary for competence, it also acts as a disincentive to vote. This creates a tension where agents might abstain even if their vote could contribute to a better group outcome.
  • Wisdom of Crowds Can Be Lost: The introduction of effort can lead to Nash equilibria where the number of voters is limited, and increasing the group size does not necessarily improve the collective decision-making accuracy. This directly undermines the CJT's guarantee of near-infallibility for large groups.
  • Pivotal Voters are Key: Agents are primarily incentivized to vote if they believe their vote can be "pivotal" – meaning it can break a tie and directly influence the outcome. The cost of effort can make this incentive insufficient.
  • Social Networks Can Save Wisdom: In the social network setting, limiting agents' awareness to their neighbors can, surprisingly, restore some of the wisdom of crowds. Agents might see themselves as pivotal within their local network, leading to more participation than they would if they were aware of the entire group's potential actions.
  • Topology Matters: The structure of the social network (its topology) significantly influences whether the wisdom of crowds can be recovered. Certain graph structures are more conducive to maintaining collective intelligence.

Key Results

  • In the basic model with uniform competence and stakes, if the effort cost e is too high (specifically, e > p - 1/2, where p is the individual competence), no one votes, and the group accuracy is essentially random (1/2). If the effort is lower, equilibria exist where the number of voters is capped by a threshold ne, determined by the effort and competence.
  • With heterogeneous stakes or competences, similar patterns emerge: high effort leads to abstention, and even with lower effort, the number of voters and thus the group accuracy can be capped, preventing the CJT's asymptotic guarantees.
  • For specific graph structures like paths, cycles, and trees with even branching factors, the authors show that the minimal size of Nash equilibria grows with the group size. This implies that, in these network structures, the wisdom of crowds effect is preserved, even in the worst-case scenario (minimal equilibrium size).

Limitations and Future Work

  • Simplified Effort Model: The paper assumes a fixed effort cost for voting. Future work could explore models where agents can choose the amount of effort to exert, leading to varying levels of competence.
  • Specific Graph Structures: While the paper provides strong results for paths, cycles, and certain trees, future work could investigate wisdom of crowds on more general or random graph structures. The authors conjecture that on random graphs, the average number of voters at equilibrium is around n/2.
  • Incentive Design: The paper highlights the problem of disincentivized participation due to costly information. Future research could focus on designing mechanisms or incentive structures to encourage sufficient voter participation and thus ensure wise crowds, even with deterring factors.
  • Stronger Equilibria: While SNEs are considered, further exploration of other equilibrium refinements could be beneficial.

Summary updated

BibTeX
@inproceedings{michelini2022group,
  author = {Matteo Michelini and Adrian Haret and Davide Grossi},
  title = {{Group Wisdom at a Price: Jury Theorems with Costly Information}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/60},
}
Figure