pith. sign in

arxiv: 2605.17662 · v2 · pith:YQYDIXUOnew · submitted 2026-05-17 · 💰 econ.TH

Learning Through Imitation: An Experiment

Marina Agranov , Gabriel Lopez-Moctezuma , Philipp Strack , Omer Tamuz This is my paper

Pith reviewed 2026-05-20 12:31 UTC · model grok-4.3

classification 💰 econ.TH
keywords social learningimitationinformation aggregationeconomic experimentsherdingoptimal actionsgroup size
0
0 comments X

The pith

Observing others' actions leads to more optimal choices even without new information.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper compares information aggregation in two repeated social learning environments. Agents who see a public dataset plus the past actions of others select the optimal action more often than those who see only the public dataset. This occurs even though the observed actions add no payoff-relevant information and despite risks such as herding, free riding, or information overload. The study further varies group size and tests a version where agents also receive private data.

Core claim

Despite actions containing no additional payoff-relevant information, agents take the optimal action more often when they can observe and imitate the past actions of others.

What carries the argument

Experimental comparison between a public-data-only condition and a public-data-plus-observed-actions condition.

If this is right

  • Larger groups increase the frequency of optimal choices when actions are observable.
  • Imitation still raises optimal play rates when agents also receive private signals.
  • Action observation mitigates free-riding and overload problems in repeated social learning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Platforms that display others' choices may improve aggregate decision quality beyond raw data access alone.
  • Models of rational herding could be extended to predict net gains from imitation in finite groups.
  • Future work could test whether noisy or delayed action observations preserve the reported benefit.

Load-bearing premise

The experimental protocol isolates the effect of observing actions without confounding changes in incentives, presentation, or subject selection.

What would settle it

A replication experiment in which agents who observe others' actions do not select the optimal action at a higher rate than those who see only the public dataset would falsify the main claim.

Figures

Figures reproduced from arXiv: 2605.17662 by Gabriel Lopez-Moctezuma, Marina Agranov, Omer Tamuz, Philipp Strack.

Figure 1
Figure 1. Figure 1: Aggregate statistics in the all and signals treatments Notes: Panel (a) presents the average frequency of correct actions in each round, averaged across games. Panel (b) depicts the evolution of consensus in each round, i.e., the relative size of the majority, averaged across games. For panel (b) we exclude cases with an equal number of green and red signals. Shaded regions represent 95% confidence interva… view at source ↗
Figure 2
Figure 2. Figure 2: Learning from signals in the signals and all treatments Notes: Panel (a) depicts the fraction of red actions as a function of the difference between the number of red and green signals. The size of the dot corresponds to the number of observations in each bin. Panel (b) shows the estimated probability of an optimal action (i.e., reporting the color of the majority of signals) as a function of the treatment… view at source ↗
Figure 3
Figure 3. Figure 3: Learning from others’ actions in the all treatment Notes: This figure depicts the probability of choosing red as a function of the share of red actions of other group members. The estimates are obtained from a Bayesian logistic regression of subjects’ actions on the share of others’ actions in the previous round conditional on signal strength and session random effects. Shaded regions represent 95% confide… view at source ↗
Figure 4
Figure 4. Figure 4: Responsiveness to signals and actions, individual level data [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Frequency of correct actions, by information structure and group size [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulation results: Evolution of the correct actions (frequency) [PITH_FULL_IMAGE:figures/full_fig_p025_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Simulation results: Learning from others’ actions in the [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Estimation results: Correct actions and learning from others [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Simulation results: Learning from others’ actions in the [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Behavioral Model Parameters (all treatment) as a function of IQ Parameter Effect on Probability β (Weight on S) γ (Weight on A) 0.5 1.0 1.5 2.0 0.1 0.2 0.3 0.4 0.5 High IQ Low IQ High IQ Low IQ Posterior Value Notes: The upper panels of this figure present posterior estimates of β and of the effect of Sit on choice probability by participants’ IQ level. The lower panels present the posterior distribution … view at source ↗
read the original abstract

We compare how well agents aggregate information in two repeated social learning environments. In the first setting agents have access to a public data set. In the second they have access to the same data, and also to the past actions of others. Despite the fact that actions contain no additional payoff-relevant information, and despite potential herd behavior, free riding and information overload issues, observing and imitating the actions of others leads agents to take the optimal action more often in the second setting. We also investigate the effect of group size, as well as a setting in which agents observe private data and others' actions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. This paper experimentally compares information aggregation in two repeated social learning environments. In the first, agents access only a public dataset; in the second, they access the same dataset plus others' past actions. The central claim is that observing and imitating actions increases the frequency of optimal choices despite actions containing no additional payoff-relevant information and despite risks of herding, free-riding, and overload. The study also varies group size and examines a private-data-plus-actions treatment.

Significance. If the design isolates the imitation channel, the result would be significant for the social learning literature in economics. It provides controlled evidence that imitation can improve aggregation even when theory highlights potential downsides, and the group-size and private-signal arms help map boundary conditions. The work supplies falsifiable empirical patterns that can discipline models of observational learning.

major comments (2)
  1. [Experimental Design / Abstract] The abstract and experimental-setup description do not specify whether the two environments are administered within-subjects or between-subjects, nor whether presentation order is counterbalanced, whether washout periods are used, or whether interface familiarity is controlled. Because the environments are repeated, any within-subjects exposure without these safeguards leaves open the possibility that higher optimal play in the action-observation condition reflects reduced cognitive load or cumulative interface learning rather than imitation per se. This directly undermines the causal attribution required by the central claim.
  2. [Results] The abstract states that agents take the optimal action more often in the second setting but supplies no information on sample sizes, session structure, statistical tests comparing frequencies across conditions, or adjustments for multiple comparisons. Without these details it is impossible to evaluate whether the reported difference is statistically reliable or driven by the imitation treatment.
minor comments (1)
  1. [Abstract] The abstract could usefully preview the exact number of subjects or sessions to convey scale to readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important issues of clarity and causal identification that we will address in revision. Below we respond point by point to the major comments.

read point-by-point responses

  1. Referee: [Experimental Design / Abstract] The abstract and experimental-setup description do not specify whether the two environments are administered within-subjects or between-subjects, nor whether presentation order is counterbalanced, whether washout periods are used, or whether interface familiarity is controlled. Because the environments are repeated, any within-subjects exposure without these safeguards leaves open the possibility that higher optimal play in the action-observation condition reflects reduced cognitive load or cumulative interface learning rather than imitation per se. This directly undermines the causal attribution required by the central claim.

    Authors: We agree that explicit description of the design is necessary to support causal claims. The experiment was run between-subjects: separate sessions were conducted for the public-data-only condition and the public-data-plus-actions condition, with no subject participating in both. Within each session, subjects completed a fixed number of rounds with the same information environment; practice rounds and a standardized interface were used to equalize familiarity. Because the design is between-subjects, order and washout issues do not arise. We will revise both the abstract and the experimental-design section to state these features clearly and to describe the session structure. revision: yes

  2. Referee: [Results] The abstract states that agents take the optimal action more often in the second setting but supplies no information on sample sizes, session structure, statistical tests comparing frequencies across conditions, or adjustments for multiple comparisons. Without these details it is impossible to evaluate whether the reported difference is statistically reliable or driven by the imitation treatment.

    Authors: The abstract is kept concise per journal norms, but we accept that key statistical information should be visible there. The full manuscript reports the number of subjects and sessions per treatment, uses session-level clustering, and presents both raw frequencies and regression results with appropriate tests. We will expand the abstract to include sample sizes and to note that the main difference is statistically significant at conventional levels (with details and any multiple-comparison adjustments provided in the results section). revision: partial

Circularity Check

0 steps flagged

No circularity: empirical experiment with direct behavioral comparison

full rationale

This is a laboratory experiment comparing subject behavior across two repeated social learning conditions (public data only vs. public data plus observed actions). The central claim rests on measured differences in the frequency of optimal actions, obtained from direct observation of choices rather than any derivation, fitted parameter, or equation that reduces to its own inputs. No self-definitional steps, no predictions that are statistically forced by construction, and no load-bearing self-citations appear in the reported protocol or results. The design is self-contained against external benchmarks because the outcome is falsifiable via subject data collected under the stated conditions, with no mathematical chain that collapses back to the inputs by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim rests on standard experimental economics assumptions about subject incentives and behavior plus the premise that the two environments differ only in the availability of action observations.

axioms (1)
  • domain assumption Subjects respond to monetary incentives in a manner consistent with the experimental design.
    The paper relies on this to interpret choices as information aggregation rather than other motives.

pith-pipeline@v0.9.0 · 5623 in / 1067 out tokens · 39652 ms · 2026-05-20T12:31:45.647736+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages

  1. [1]

    Agranov, B

    M. Agranov, B. Gillen, and D. Persitz. Behavioral and structural barriers to information aggregation in networks. working paper, 2025

    work page 2025

  2. [2]

    L. R. Anderson and C. A. Holt. Information cascades in the laboratory. The American economic review, pages 847--862, 1997

    work page 1997

  3. [3]

    Angrisani, A

    M. Angrisani, A. Guarino, P. Jehiel, and T. Kitagawa. Herd behavior in a laboratory financial market. American Economic Review, 95: 0 1427--1443, 2005

    work page 2005

  4. [4]

    Angrisani, A

    M. Angrisani, A. Guarino, P. Jehiel, and T. Kitagawa. Information redundancy neglect versus overconfidence: a social learning experiment. American Economic Journal: Microeconomics, 13 0 (3): 0 163--97, 2021

    work page 2021

  5. [5]

    Azrieli, C

    Y. Azrieli, C. Chambers, and P. Healy. Incentives in experiments: A theoretical analysis. Journal of Political Economy, 126: 0 1472--1503, 2018

    work page 2018

  6. [6]

    A. V. Banerjee. A simple model of herd behavior. The Quarterly Journal of Economics, pages 797--817, 1992

    work page 1992

  7. [7]

    D. J. Benjamin. Errors in probabilistic reasoning and judgment biases. Handbook of Behavioral Economics, edited by Doug Bernheim, Stefano DellaVigna, and David Laibson, 2019

    work page 2019

  8. [8]

    Bikhchandani, D

    S. Bikhchandani, D. Hirshleifer, and I. Welch. A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy, pages 992--1026, 1992

    work page 1992

  9. [9]

    Bikhchandani, D

    S. Bikhchandani, D. Hirshleifer, O. Tamuz, and I. Welch. Information cascades and social learning. Technical report, National Bureau of Economic Research, 2021

    work page 2021

  10. [10]

    Caplin, M

    A. Caplin, M. Dean, and D. Martin. Search and satisficing. American Economic Review, pages 2899--2922, 2011

    work page 2011

  11. [11]

    C elen and S

    B. C elen and S. Kariv. Distinguishing informational cascades from herd behavior in the laboratory. American Economic Review, 94 0 (3): 0 484--498, 2004

    work page 2004

  12. [12]

    Chandrasekhar, H

    A. Chandrasekhar, H. Larreguy, and J. Xandri. Testing models of social learning on networks: Evidence from two experiments. Econometrica, pages 1--32, 2020

    work page 2020

  13. [13]

    Chapman, M

    J. Chapman, M. Dean, P. Ortoleva, E. Snowberg, and C. Camerer. Econographics. NBER working paper \#24931, 2019

    work page 2019

  14. [14]

    D. L. Chen, M. Schonger, and C. Wickens. otree—an open-source platform for laboratory, online, and field experiments. Journal of Behavioral and Experimental Finance, 9: 0 88--97, 2016

    work page 2016

  15. [15]

    S. Choi, D. Gale, and S. Kariv. Social learning in networks: a quantal response equilibrium analysis of experimental data. Review of Economic Design, pages 135--157, 2012

    work page 2012

  16. [16]

    D. M. Condon and W. Revelle. The international cognitive ability resource: Development and initial validation of a public-domain measure. Intelligence, 43: 0 52--64, 2014

    work page 2014

  17. [17]

    Dasaratha and K

    K. Dasaratha and K. He. An experiment on network density and sequential learning. Games and Economic Behavior, 128: 0 182--192, 2021

    work page 2021

  18. [18]

    De Filippis, A

    R. De Filippis, A. Guarino, P. Jehiel, and T. Kitagawa. Non-bayesian updating in a social learning experiment. Journal of Economic Theory, 199: 0 105188, 2022

    work page 2022

  19. [19]

    Duffy, E

    J. Duffy, E. Hopkins, T. Kornienko, and M. Ma. Information choice in a social learning experiment. Games and Economic Behavior, 118: 0 295--315, 2019

    work page 2019

  20. [20]

    Duffy, E

    J. Duffy, E. Hopkins, and T. Kornienko. Lone wolf or herd animal? information choice and learning from others. working paper, 2020

    work page 2020

  21. [21]

    Enke and F

    B. Enke and F. Zimmermann. Correlation neglect in belief formation. Review of Economic Studies, pages 313--332, 2019

    work page 2019

  22. [22]

    Evdokimov and U

    P. Evdokimov and U. Garfagnini. Individual vs. social learning: An experiment. working paper, 2020

    work page 2020

  23. [23]

    Eyster and M

    E. Eyster and M. Rabin. Extensive imitation is irrational and harmful. The Quarterly Journal of Economics, 129 0 (4): 0 1861--1898, 2014

    work page 2014

  24. [24]

    Eyster, M

    E. Eyster, M. Rabin, and G. Weizsacker. An experiment on social mislearning. Available at SSRN 2704746, 2018

    work page 2018

  25. [25]

    Gale and S

    D. Gale and S. Kariv. Bayesian learning in social networks. Games and Economic Behavior, 45 0 (2): 0 329--346, 2003

    work page 2003

  26. [26]

    Gneezy and J

    U. Gneezy and J. Potters. An experiment on risk taking and evaluation periods. Quarterly Journal of Economics, pages 631--645, 1997

    work page 1997

  27. [27]

    J. K. Goeree, T. Palfrey, B. Rogers, and R. McKelvey. Self-correcting information cascades. Review of Economic Studies, 74 0 (3): 0 733--762, 2007

    work page 2007

  28. [28]

    Griffin and A

    D. Griffin and A. Tversky. The weighing of evidence and the determinants of confidence. Cognitive psychology, 24 0 (3): 0 411--435, 1992

    work page 1992

  29. [29]

    Grimm and F

    V. Grimm and F. Mengel. Experiments on belief formation in networks. Journal of the European Economic Association, 18 0 (1): 0 49--82, 2020

    work page 2020

  30. [30]

    Harel, E

    M. Harel, E. Mossel, P. Strack, and O. Tamuz. Rational groupthink. The Quarterly Journal of Economics, 136 0 (1): 0 621--668, 2021

    work page 2021

  31. [31]

    Huang, P

    W. Huang, P. Strack, and O. Tamuz. Learning in repeated interactions on networks. arXiv preprint arXiv:2112.14265, 2021

    work page arXiv 2021

  32. [32]

    Hung and C

    A. Hung and C. Plott. Information cascades: Replication and an extension to majority rule and conformity-rewarding institutions. American Economic Review, 91 0 (5): 0 1508--1520, 2001

    work page 2001

  33. [33]

    Kubler and G

    D. Kubler and G. Weiszacker. Limited depth of reasoning and failure of cascade formation in the laboratory. Review of Economic Studies, 71 0 (2): 0 425--441, 2004

    work page 2004

  34. [34]

    R. D. Luce. A probabilistic theory of utility. Econometrica: Journal of the Econometric Society, pages 193--224, 1958

    work page 1958

  35. [35]

    Mossel, A

    E. Mossel, A. Sly, and O. Tamuz. Asymptotic learning on bayesian social networks. Probability Theory and Related Fields, 158 0 (1-2): 0 127--157, 2014

    work page 2014

  36. [36]

    Mueller-Frank and C

    M. Mueller-Frank and C. Neri. A general model of boundedly rational observational learning: Theory and evidence. working paper, 2015

    work page 2015

  37. [37]

    Parikh and P

    R. Parikh and P. Krasucki. Communication, consensus, and knowledge. Journal of Economic Theory, 52 0 (1): 0 178--189, 1990

    work page 1990

  38. [38]

    M. E. Roberts, B. M. Stewart, D. Tingley, and E. M. Airoldi. The structural topic model and applied social science. In Advances in neural information processing systems workshop on topic models: computation, application, and evaluation, volume 4, pages 1--20. Harrahs and Harveys, Lake Tahoe, 2013

    work page 2013

  39. [39]

    Scheibehenne and R

    B. Scheibehenne and R. Greifeneder. Can there ever be too many options? a meta-analytic review of choice overload. Journal of Consumer Research, pages 409--425, 2010

    work page 2010

  40. [40]

    Smith and P

    L. Smith and P. S rensen. Pathological outcomes of observational learning . Econometrica, 68 0 (2): 0 371--398, 2000

    work page 2000

  41. [41]

    X. Vives. How fast do rational agents learn? The Review of Economic Studies, 60 0 (2): 0 329--347, 1993

    work page 1993

  42. [42]

    Weizsacker

    G. Weizsacker. Do we follow others when we should? a simple test of rational expectations. American Economic Review, pages 2340--236, 2010

    work page 2010

  43. [43]

    Ziegelmeyer, F

    A. Ziegelmeyer, F. Koessler, B. J., and E. Winter. Fragility of information cascades: An experimental study using elicited beliefs. Experimental Economics, 13 0 (2): 0 121--145, 2010

    work page 2010

  44. [44]

    Ziegelmeyer, C

    A. Ziegelmeyer, C. March, and S. Krugel. Do we follow others when we should? a simple test of rational expectations: Comment. Experimental Economics, 103 0 (6): 0 2633--42, 2013

    work page 2013