Can Children and Adults Balance Majority Size with Information Quality in Learning from Preferences?

Rebekah Gelpí, Andrew Whalen, Thomas L. Griffiths, Xu Fei, Daphna BuchsbaumPsyArXiv 2023 · DOI

TL;DR

While both children and adults prefer information from people with first-hand experience, children are uniquely swayed by the "wisdom of the crowd"—even when that crowd is just repeating hearsay. This study reveals that while 3- to 5-year-olds can evaluate the quality of a source, they possess a significant conformity bias that leads them to overweight the size of a majority, whereas adults more rationally balance group size against the actual number of independent observations.

What Problem Does This Paper Address?

In social learning, we often rely on two main cues: consensus (how many people agree) and source quality (how they know what they know). Usually, a majority is a reliable signal because more people typically mean more independent pieces of evidence. However, this fails when a majority is formed through "hearsay"—a game of telephone where many people simply repeat the opinion of a single, perhaps unreliable, individual. This is known as the "illusion of consensus."

The researchers sought to determine when humans develop the ability to discount redundant information. Specifically, they investigated whether children and adults can navigate the tension between a large group with low-quality information (hearsay) and a small group with high-quality information (direct observation) within the domain of personal preferences.

Methods

The researchers conducted six experiments involving 189 children (ages 3–5) and 241 adults. The core task involved choosing between two boxes containing hidden snacks or toys based on the "preferences" of groups of informants (dolls for children, cartoon characters for adults).

  • Experimental Design: Informants were categorized as having Direct Knowledge (they looked inside the boxes) or Indirect Knowledge (they heard which was better via "whispers" or hearsay).
  • Computational Modeling: The authors developed three Bayesian models to predict participant behavior:
  • Source Knowledge Model (Rational): A normative model that treats hearsay as redundant, essentially only "counting" the primary source.
  • Conformity-Biased Model: A model that treats every endorsement as independent, regardless of its source.
  • Mixed Model: A 50/50 hybrid of the two.
  • Statistical Analysis: They used generalized linear mixed models (GLMM) and compared the log-likelihood of their computational models against actual behavioral data.

Datasets

The "datasets" in this paper consist of original behavioral data collected by the authors:

  • Children: $N = 189$ (3- to 5-year-olds from the US and Canada).
  • Adults: $N = 241$ (US residents recruited via Amazon Mechanical Turk).
  • Open Science: The authors provided a link to the anonymized data and analysis code on the Open Science Framework (OSF).

Key Insights and Ideas

  • Children are not "blind" conformists: In Experiments 1–3, when group sizes were equal (e.g., 4 vs. 4), children consistently chose the box endorsed by the group with direct visual access. They clearly understand that seeing is better than hearing.
  • The "Majority" Trumps Quality for Kids: The "aha moment" occurs in Experiments 4–5. When a larger group of "hearsay" informants (e.g., 7 people) was pitted against a single "direct" informant, children switched their preference to the majority. Even though both groups relied on exactly one primary observation, the mere presence of more people biased the children's judgment.
  • Adults are (mostly) Normative: Unlike children, adults in Experiment 6 behaved almost exactly like the "Source Knowledge Model." They recognized that seven people repeating one person's opinion is no better than one person speaking for themselves.
  • The Mixed Model Fit: Children’s behavior was best explained by the Mixed Model. This suggests they are performing a "tug-of-war" between their ability to evaluate source quality and an innate drive to follow the crowd.

Key Results

  • Direct vs. Hearsay (Equal Size): Children chose the direct group significantly more than chance ($OR = 2.92, p = .019$).
  • The 1 vs. 7 Conflict: In the most extreme conflict (1 direct observer vs. 7 hearsay informants), children chose the box endorsed by the indirect majority significantly more often ($p = .034$).
  • Adult Performance: Adults chose the direct group in the 4 vs. 6 and 3 vs. 5 conditions with high significance ($p < .001$), showing no majority bias when quality was lower.
  • Model Comparison: For children, the Mixed Model provided a significantly better fit (log-likelihood -250.91) than the purely rational Source Knowledge Model (log-likelihood -279.04).

Limitations and Future Work

  • Theory of Mind: The authors suggest children’s conformity might stem from an emerging "Theory of Mind"—they may not yet fully grasp that beliefs are generated from specific evidence and that hearsay doesn't add "new" evidence.
  • Social Motivation: Children might be motivated by a desire to affiliate with the larger group rather than just seeking the "best" snack.
  • Domain Specificity: The study focused on preferences. The authors note that children might weigh consensus differently for "facts" (where there is a ground truth) versus "preferences" (which are subjective and often shared).
  • Prestige Bias: It remains unclear if children perceive the "source" of a hearsay chain as more "prestigious" because others are listening to them, which could be an alternative explanation for the bias.

Summary updated

Abstract

We investigate how 3- to 5-year-old US and Canadian children (N = 189) and US adults (N = 241) balance the number of endorsements for a given option with the quality of the informants’ source of information when deciding which of two boxes contains the better option. When choosing between two different boxes endorsed by groups of equal sizes, both children (Experiments 1–3) and adults (Experiment 6) tend to choose boxes endorsed by informants with visual access to the boxes over informants with hearsay. However, children’s choices were biased towards the larger group when the size of the group conflicted with the quality of the source of the groups’ information (Experiments 4–5), while adults more often chose the option endorsed by the group with the higher quality information (Experiment 6). Children were more likely to conform to a majority opinion when compared to both adults and to a normative computational model that endorses a group proportional to the number of independent, direct observations made by that group’s informants. These findings suggest that, while adults balance the size of a majority with the quality of the informants’ information source, preschoolers can evaluate when groups differ in the source of their information, but may assume that the presence of a majority endorsing an option is inherently informative over and above the information source group members’ testimony relied on.
BibTeX
@misc{gelpi2023can,
  author = {Rebekah Gelpí and Andrew Whalen and Thomas L. Griffiths and Xu Fei and Daphna Buchsbaum},
  title = {{Can Children and Adults Balance Majority Size with Information Quality in Learning from Preferences?}},
  year = {2023},
  doi = {10.31234/osf.io/pa3g4},
  howpublished = {PsyArXiv},
}
Figure