Open data · CC BY 4.0 what each one claims, and what replicated no ads · no tracking
Bias Atlascognitive biases

Information cascade

also known as — Informational cascade · Rational herding · Observational learning
When people choose in sequence and can see each other's choices but not each other's evidence, it becomes rational to ignore private evidence, so a crowd converges on very little of it.
Does it replicate? Mixed Sequential urn experiments keep producing cascades, and the one direct replication of Anderson and Holt reports cascade behaviour in 77.5 percent of rounds. What fails is the model's account of why: subjects lean far too hard on their own signal, and in long sequences cascades break and re-form instead of lasting.
VerdictMixed
First published1992
FieldSocial and self
Sources11
Last checked2026-09-21
What it claims

What does Information cascade mean?

This is a formal model rather than a described bias, and the people in it are all rational. Each person sees a private signal about which of two states holds, observes what everyone before them did, and cannot observe what anyone before them knew. Once the public record leans far enough, the next person's own signal cannot overturn it, so the payoff-maximising move is to copy.

From that point onward every choice is uninformative, learning stops, and the whole crowd rests on the handful of signals seen before the copying began. Nobody in the model wants to conform, nobody gains from conformity, and nobody is confused. The consequences follow from the information structure alone. Two things travel under adjacent names and are not this.

The bandwagon effect is a claim that popularity itself raises demand, and it is entry 74 here. The availability cascade adds reputational silence and press coverage to the informational mechanism, and it is entry 150.

Verdict

Repeats disagree, or the effect holds in some conditions and not others. The other 248.

The field
94 of the 170 entries in social and self share this verdict.
Does it replicate?

Has Information cascade been retested?

Mixed Sequential urn experiments keep producing cascades, and the one direct replication of Anderson and Holt reports cascade behaviour in 77.5 percent of rounds. What fails is the model's account of why: subjects lean far too hard on their own signal, and in long sequences cascades break and re-form instead of lasting.

Hung and Plott ran four experiments at Caltech, each with ten undergraduates and 28 to 55 periods, using computers and different payoffs. Under the individualistic institution, the one that copies Anderson and Holt, they observed cascade behaviour in 77.5 percent of rounds and state that the earlier results replicate. Lisa Anderson also repeated the design herself, varying the payoff scale, in Economic Inquiry for October 2001.

That article is closed at the publisher and no repository holds a copy, so no figure from it is quoted here. Goeree, Palfrey, Rogers and McKelvey ran thirteen sessions at Caltech and UCLA with sequences of 20 or 40 deciders rather than six, and signals correct with probability 5/9 or 6/9. Cascades formed in every sequence in all four treatments, which is what the theory predicts.

They then collapsed. Pooling the two 40-person treatments, a cascade arose and persisted in only 8 of 116 sequences, under 7 percent, and the paper describes cascades as ephemeral, cycling through formation, collapse and reformation. Breaks were informative rather than random: averaged over the four treatments, 84 percent of the breaks came from someone whose private signal contradicted the cascade.

Across the four treatments the chance that a reversed cascade was a self-corrected one, switching from the wrong state to the right one, was 63 percent. A correct cascade that broke reversed to an incorrect one in less than 6 percent of cases, while an incorrect one that broke led to a self-corrected cascade in more than 21 percent.

Those figures are read from Caltech Social Science Working Paper 1197 of April 2004, not from the 2007 journal version, which could not be obtained. Whether the copying is Bayesian is a separate question, and the answer is mostly no. Huck and Oechssler set three cascade problems inside a written microeconomics final exam, splitting 63 students into three equal groups that each faced one problem.

Students were told their hypothetical predecessors had all decided rationally, and a second exam question asked them to explain their answers. Only 53 percent of decisions matched Bayes' rule, which is not distinguishable from coin flipping, and the heuristic 'follow your own signal' fitted 71.4 percent. Cascades occurred in 31.0 percent of the cases where they should have.

Not one student set out the correct Bayesian reasoning for the two tasks that required joining a cascade. Re-examining Anderson and Holt's asymmetric sessions, they isolated 79 harder decisions where Bayes' rule and counting disagree, and found 39 in line with Bayes and 52 in line with following one's own signal. Kübler and Weizsäcker ran 66 Boston-area students through 4,140 decisions at Harvard Business School, charging for signals in three of the four treatments.

Equilibrium there requires only the first player to buy a signal and everyone after to follow blindly. Too many signals were bought, and all six players behaved as predicted in 12, 3 and 4 percent of games in the three costly treatments. In their free-signal control, which reproduces Anderson and Holt's setting, subjects disregarded a contradicting signal in 78 percent of cases, 63 of 81, and all six played the equilibrium in 41 percent of 90 rounds against 60 percent in the original.

Weizsäcker later pooled 13 of these experiments, 29,923 decisions made by 2,813 participants, to ask how much the reluctance costs. Where the data show it was better to contradict one's own signal, players took that action in fewer than half the cases, earning 53 percent of the high prize against the 64 percent they could have had. Those two figures are read from the July 2008 IZA discussion paper, because the published article of December 2010 could not be obtained.

The published abstract carries neither number. It does carry the other headline result: the average player contradicts her own signal only once the empirical odds of that signal being wrong pass 2:1, where rational expectations would put the threshold at 1:1.

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

Examples

What are some examples of Information cascade?

The two restaurants on one streetillustration

Both are empty. The first diner has a faint recommendation for the left one and takes it. The next two each have a hint favouring the right, see someone seated on the left, and follow. From then on no private information enters the record.

The paper everyone citesillustration

An early reader judges a finding solid and cites it. Later readers see the citations and not what any reader thought of the data, and cite it too. The record grows without one further reading of the evidence behind it.

The tool a department standardises onillustration

Each team can see which tool the teams before it picked and not why. Once two have chosen, a third with a mild preference for something else does better by copying, and the standard rests on two people's reasons.

Cases marked as illustrations describe nobody in particular. They are there to make the pattern recognisable, not to report that a particular thing happened.

The experiments

What experiments is Information cascade based on?

The design

The 1992 paper is mathematics. Its binary example has each person receive a high or low signal that is correct with probability p, and the probability that no cascade has begun falls exponentially in the number of people who have already chosen. With a signal barely better than a coin, the paper reports that this probability after only ten individuals is below 0.1 percent.

Cascades therefore form almost at once and, absent fresh public information, never end. They are also frequently wrong. At p of 0.75, the model gives 0.81 for the chance of settling into the correct cascade, and the paper's figure 1 is captioned to say that the wrong cascade stays considerable even for large p. Adding public releases repairs this: at the same p, one release per thousand individuals lifts 0.81 to 0.86, and ten releases lift it to 0.98.

The laboratory test came five years later. Anderson and Holt recruited 72 undergraduates from economics courses at the University of Virginia, paid $5 for turning up and $2 for each correct prediction, and ran twelve sessions of fifteen periods. Six subjects predicted in random order which of two urns a private draw had come from, each hearing the predictions made before theirs, with a seventh subject serving as monitor.

Over all twelve sessions cascades formed in 87 of the 122 periods in which they were possible. Six sessions used symmetric urns and gave 41 of 56; six used an asymmetric design, in which counting the signals and applying Bayes' rule give different answers, and gave 46 of 66. Reverse cascades, where early unrepresentative draws start a run of wrong predictions, made up 13 of the 41 and 18 of the 46.

In the symmetric sessions subjects followed their own draw against the Bayesian inference in 26 percent of the cases where the two conflicted, and 4 percent of decisions matched neither. Efficiency, scored so that optimal play is 100 and random play is 0, averaged 91.4 percent under symmetric urns and 67.6 percent under asymmetric ones.

3 primary research8 replication and reanalysis

Origin

Who first described Information cascade, and when?

Year1992
WhoSushil Bikhchandani, David Hirshleifer and Ivo Welch
WhereJournal of Political Economy

The paper ran in volume 100, issue 5, pages 992 to 1026, and coins the term. It defines an informational cascade as occurring "when it is optimal for an individual, having observed the actions of those ahead of him, to follow the behavior of the preceding individual without regard to his own information". Abhijit Banerjee published a model of the same structure in the Quarterly Journal of Economics for August 1992, volume 107, issue 3, pages 797 to 817, under the name herd behaviour.

Banerjee's abstract states that optimising individuals will end up "doing what others are doing rather than using their information", and that the resulting equilibrium is inefficient. Neither paper reports data.

The claim, and the replication this entry reads its figures from.

Where it runs out

When does Information cascade not apply?

Every experiment above is a two-state, two-action urn game with a public sequence and paid accuracy. Real crowds rarely see a clean ordering of who chose when, and almost never know that everyone faces the same question. Alevy, Haigh and List took the design to traders on the floor of the Chicago Board of Trade, and ran the student sessions at the University of Maryland.

Across 20 sessions with 55 traders and 54 students they collected 1,647 decisions, of which 78 percent matched the perfect Bayesian prediction. Cascade opportunities arose in 441 of those decisions, and cascades were realised in 245 of them, 56 percent. Traders were not less prone to cascade overall, but under asymmetric urns only 8 of their 66 cascades were reverse cascades, against 25 of 99 for students, a difference the authors report as significant at p below 0.05 on a Mann-Whitney test.

Those figures are read from NBER working paper 12767 of December 2006, the journal version not being obtainable. Field evidence is thinner and cannot isolate the mechanism. Cai, Chen and Fang randomised what a restaurant's menu displayed and report demand for the five most popular dishes rising 13 to 20 percent when their ranking was shown, with no significant saliency effect, which shows people using others' choices as information but not that anybody stopped using their own.

The randomised popularity experiments most often cited in this connection, by Salganik and colleagues and by Muchnik, Aral and Taylor, are treated in entry 74. They manipulate a displayed count rather than a sequence of observable individual decisions, so they do not test the model's inference step. No study was located that identifies an information cascade in an economically important episode by measuring the private signals people held.

Searching on 21 September 2026 covered Crossref by bibliographic query, OpenAlex records for the 1992 papers and for the experiments cited here, and web search pairing the term with field, identification, private signals and empirical.

Often confused with

Which biases is Information cascade confused with?

Each of these is named in this entry's own prose as something it gets taken for. Where the two came out differently, that is worth knowing before you quote either.

Commonly misread as

How is Information cascade commonly misread?

The model is not about a taste for conformity, and it does not need one. Everyone in it maximises their own payoff and would break the pattern for a strong enough signal. That is what separates it from the bandwagon effect, where wanting what others want is the claim itself. Anderson and Holt's 87 of 122 is quoted as showing the theory works, and it shows the behaviour appears.

Their own asymmetric sessions show a third of subjects deciding by counting signals rather than weighting them, and in the cases where counting and Bayes' rule disagree the correct decision came only half the time, 41 of 82. Huck and Oechssler's reading of that same data is the sharper one: the theory looks strong largely where its prescription happens to coincide with following one's own signal.

Rates of cascade formation and rates of Bayesian play are different numbers, and they are often swapped. Anderson and Holt's 87 of 122 counts cascades. The 78 percent in Alevy and colleagues counts Bayesian decisions, and their 245 of 441 counts cascades again. A second misreading runs the other way. Cascades in the model are permanent unless new public information arrives, and readers take the laboratory's ephemeral cascades as a refutation.

Goeree and colleagues instead show the breaks are systematic and mostly beneficial, since a reversal corrects the state 63 percent of the time. The model also does not predict that crowds are usually wrong. At a signal accuracy of 0.75 its own figure for ending in the correct cascade is 0.81. What it predicts is that the crowd is confident on thin evidence and can be flipped by a little more of it.

Sources

Sources for Information cascade

  1. 1Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy, 100(5), 992-1026. Full text read for the definition, the exponential result, the 0.81 figure, the figure 1 caption and the public-information numbers.primary
  2. 2Banerjee, A. V. (1992). A simple model of herd behavior. The Quarterly Journal of Economics, 107(3), 797-817. Read for the abstract and opening; source of the independent model published the same year.primary
  3. 3Anderson, L. R., & Holt, C. A. (1997). Information cascades in the laboratory. American Economic Review, 87(5), 847-862. Full text of the published article read; source of the 72 subjects, the twelve sessions, the 87 of 122 periods, the 41 of 56 and 46 of 66 splits, the reverse-cascade counts, the 26 and 4 percent figures, the efficiency figures and the 41 of 82. The copy linked here is the one read, because the econweb.ucsd.edu copy cited previously serves an incomplete certificate chain and cannot be fetched.primary
  4. 4Hung, A. A., & Plott, C. R. (2001). Information cascades: Replication and an extension to majority rule and conformity-rewarding institutions. American Economic Review, 91(5), 1508-1520. Published text read from the CaltechAUTHORS copy linked here; source of the four experiments, the ten Caltech undergraduates in each, the 28 to 55 periods and the 77.5 percent.replication
  5. 5Anderson, L. R. (2001). Payoff effects in information cascade experiments. Economic Inquiry, 39(4), 609-615. A repeat of the design by one of its own authors. Citation confirmed against Crossref on 21 September 2026; the full text is closed at the publisher and OpenAlex records no repository copy, so nothing from it is quoted on this page.replication
  6. 6Goeree, J. K., Palfrey, T. R., Rogers, B. W., & McKelvey, R. D. (2007). Self-correcting information cascades. Review of Economic Studies, 74(3), 733-762. All figures quoted here are read from the authors' Caltech Social Science Working Paper 1197 of April 2004, linked below. The journal version could not be obtained on 21 September 2026: OpenAlex records it as closed with no open copy anywhere, and academic.oup.com answers with a Cloudflare bot challenge rather than an article page.replication
  7. 7Huck, S., & Oechssler, J. (2000). Informational cascades in the laboratory: Do they occur for the right reasons? Journal of Economic Psychology, 21(6), 661-671. Figures read from the working paper of 23 January 1999 linked here. The published version could not be obtained on 21 September 2026: OpenAlex records it as closed with no repository copy, and Crossref confirms the journal, volume, issue and pages given.replication
  8. 8Kubler, D., & Weizsacker, G. (2004). Limited depth of reasoning and failure of cascade formation in the laboratory. Review of Economic Studies, 71(2), 425-441. Full text of the published article read for the 66 subjects, the 4,140 decisions, the equilibrium-play percentages, the 63 of 81 herding rate and the 60 percent the paper reports for Anderson and Holt.replication
  9. 9Weizsacker, G. (2010). Do we follow others when we should? A simple test of rational expectations. American Economic Review, 100(5), 2340-2360. The published abstract was read at the publisher and carries the 2:1 threshold but neither percentage. The 53 and 64 percent figures, the 29,923 decisions and the 2,813 participants are read from IZA Discussion Paper 3616 of July 2008, linked here, whose abstract differs from the published one. The published full text could not be obtained on 21 September 2026: pubs.aeaweb.org answers with a Cloudflare bot challenge, papers.ssrn.com returns a Content Blocked page, and the LSE Research Online deposit is the April 2008 working paper rather than the article.replication
  10. 10Alevy, J. E., Haigh, M. S., & List, J. A. (2007). Information cascades: Evidence from a field experiment with financial market professionals. The Journal of Finance, 62(1), 151-180. Figures read from NBER working paper 12767 of December 2006, linked here. The journal version could not be obtained on 21 September 2026: onlinelibrary.wiley.com answers with a Cloudflare bot challenge rather than an article page.replication
  11. 11Cai, H., Chen, Y., & Fang, H. (2009). Observational learning: Evidence from a randomized natural field experiment. American Economic Review, 99(3), 864-882. Published abstract read at the publisher; source of the 13 to 20 percent and of the absent saliency effect. The full text sits behind the same bot challenge as the other AEA articles here.replication

Every claim on this page was held against these sources on 2026-09-21. Nothing here is written from memory. Written and checked by Krishna Chagti; how entries are written and corrected. Found a mistake? Report it, with the sentence and a source.

About Information cascade

Is Information cascade real?

Sequential urn experiments keep producing cascades, and the one direct replication of Anderson and Holt reports cascade behaviour in 77.5 percent of rounds. What fails is the model's account of why: subjects lean far too hard on their own signal, and in long sequences cascades break and re-form instead of lasting. The effect sizes from the original study and from the replication are printed side by side above, so the comparison is visible rather than asserted.