Skip to content
mentua.org

What Is an Information Cascade?

Line drawing of balls on a staircase going down, the first two held together in a filled band, the later ones dotted, with a dotted arrow following the stepsLine drawing of balls on a staircase going down, the first two held together in a filled band, the later ones dotted, with a dotted arrow following the steps

5 min read≈ 1,190 words3 questions · up to 3 points

What this is about

Why would a long line of sensible people make the same choice when most of them privately had reason to doubt it? You learn why, after only two choices, everyone else may stop using their clue, and how one piece of news can turn the whole row around. You also learn how to tell a crowd that knows something from a crowd that only follows.

In a nutshell· skip ahead

An information cascade happens when people decide one after another, see what earlier people chose, and find it sensible to copy them instead of using their own information. From then on their choices reveal nothing new, so the group stops learning. A few early choices can then decide for everyone, and the result may be wrong.

Two restaurants stand side by side in a street you do not know. One is empty and the other has a few guests, so you choose the busy one, and the next passer-by sees you inside and makes the same choice. Each step looks sensible, but the busy restaurant may only have been lucky with its first two guests. How can a line of reasonable people end up following a choice that almost nobody had good reason to make?

What is an information cascade?

An information cascade happens when people decide one after another, can see what the earlier people chose, and find it sensible to copy them instead of using their own information. From that point on, their choices no longer depend on what they know, so the choices tell later people nothing new. In 1997 the economists Lisa Anderson and Charles Holt tested the idea in the laboratory with a simple game, which shows the mechanism step by step.

The game uses two urns. Picture urn A with two red balls and one blue ball, and urn B with two blue balls and one red. One urn is chosen in secret. Each player in turn draws one ball in private, puts it back and says out loud which urn they think it is, and a correct guess earns money. Each private draw is a signal, a clue that points in the right direction two times out of three.

  1. Ann draws red. Red is more likely to come from urn A, so she says “A”. Everyone can now work out that she drew red.
  2. Bob draws red too and says “A”. Had he drawn blue, his ball and Ann’s would have cancelled out, and Anderson and Holt assumed that a player who is undecided follows their own ball, so his guess also reveals what he drew.
  3. Carol draws blue. She knows of two red balls and one blue, and that is still more likely to come from urn A, so she says “A” against her own ball.
  4. Everyone after Carol is in the same position, whatever they draw, so they all say “A”. Their balls stay private, and the group’s knowledge stops growing after the second player.

After two red draws, the chance that the urn is A is 80 percent, which is enough to outweigh any single draw. But how often does this go wrong? If the urn is really B, there is a one-in-nine chance that the first two players both draw red, and then everyone after them is wrong. Taking all the ways the game can start, a cascade settles on the wrong urn about one time in five.

If a hundred players could instead show each other their balls, a majority vote would pick the right urn more than 99.9 percent of the time. Copying throws away almost all of the group’s information.

The same arithmetic explains why cascades are fragile. A cascade is built on only a little more information than one person’s clue, so a single new fact that everyone hears can make people use their own information again and turn the whole line around. The founders of the theory used this to explain fads, which spread quickly and stop just as quickly.

What an information cascade teaches you

When you see many people make the same choice, ask how many of them used their own information and how many only followed the people before them. A hundred followers can carry less evidence than three people who judged alone.

  • Ratings and charts. In 2006 the researchers Matthew Salganik, Peter Dodds and Duncan Watts built an online music market in which 14,341 participants could download 48 unknown songs. Some saw how often others had downloaded each song, in eight separate “worlds”, and others did not. Seeing the counts made success more unequal and harder to predict: the best songs rarely failed and the worst rarely won, but the songs in between could end up almost anywhere.
  • Meetings and juries. If people speak in turn, the first two opinions can decide the meeting. Asking everyone to write down a view before the discussion keeps each person’s information in the group. This is the advice of “The Wisdom of Crowds” by James Surowiecki: Summary and Review, whose chapter on independence explains how a crowd turns into a herd.
  • Health. In the United States, a donated kidney is offered to waiting patients in order. In 2010 the marketing researcher Juanjuan Zhang found that patients read earlier refusals as a sign of poor quality and refused in turn, so kidneys were used badly although there are far too few of them. Her simulations suggest that kidneys would be used better if the reasons for earlier refusals were shared.

Where does the theory not fit real people?

A common misunderstanding is that a cascade means people are foolish or easily led. In the model, everyone reasons carefully, and the cascade comes from the order in which decisions are made and from what can be seen, not from weakness. It is also different from social pressure, where people agree in order to fit in; in a cascade they copy because the earlier choices look like better evidence.

Real people do not follow the model exactly. In Anderson and Holt’s experiment, cascades formed in about 70 percent of the rounds in which the theory predicted one. In a similar experiment published in 2000, the economists Steffen Huck and Jörg Oechssler found that the simple rule “follow your own signal” explained their players’ choices better than careful probability reasoning did. People can also err in the other direction: when they treat the choices of earlier people as separate pieces of evidence, although those people were copying each other, they copy even more readily than the model predicts.

The strongest objection is that cascades are hard to prove outside the laboratory. When many people make the same choice, they may be copying, but they may also share the same news, fear standing out, or gain from choosing the same thing as others. A cascade also depends on people seeing only choices, not reasons. When people can explain why they chose, or say how sure they are, later people can often learn much more.

Learn more

The idea has no single originator. In 1992 the economists Sushil Bikhchandani, David Hirshleifer and Ivo Welch named “informational cascades” in the Journal of Political Economy, and in the same year Abhijit Banerjee, later a Nobel laureate in economics, described the same pattern as “herd behavior”. Welch also showed in 1992 that when shares in a new company are sold to investors one after another, the sale tends to succeed or fail quickly.

To see why a crowd of people who judge alone is often right, read What Is the Wisdom of the Crowd?, which is the opposite case of this one. The probability reasoning that Carol uses at step three is explained in Bayes’ Theorem, Explained.

Quiz

3 questions · up to 3 points

Open a question and pick one answer. Only your first answer counts: +1 if it is right, −1 if it is wrong.

  1. Why does a group stop learning once an information cascade has started?
    Show answer

    Later people choose the same thing whatever their own information says, so their choices tell others nothing new. A choice is only informative if it depends on what the person knows. In a cascade it no longer does, so the clues of all later people stay private.

  2. A new book has three five-star ratings on the first day, and soon hundreds of readers buy it and rate it highly. What should you conclude, if the idea of a cascade applies?
    Show answer

    The ratings may carry much less information than their number suggests, because many later readers may have followed the first ones. Once people follow earlier choices, extra choices add little new information. Hundreds of copied ratings may contain no more information than the opinions of a few early readers.

  3. Which statement about information cascades matches the article?
    Show answer

    Everyone in a cascade can be acting sensibly, and the group can still end up with the wrong answer. Each person makes the best bet given what they can see. The problem is that the later people's private information never becomes visible to the group.

The takeaway

An information cascade happens when people who decide in a row copy the visible choices of earlier people instead of using their own information. In the urn game, two equal guesses are enough to make everyone after them ignore their own draw, so the group learns almost nothing from later players and settles on the wrong urn about one time in five. Cascades break easily, and real people copy less exactly than the theory says. When many people make the same choice, ask how many of them used their own judgment.

Key quotes

“An agent is said to be in an information cascade if it is optimal for her to choose an action which is independent of her private signal.”
Sushil Bikhchandani, David Hirshleifer, Omer Tamuz and Ivo Welch, "Information Cascades and Social Learning", NBER Working Paper 28887 (2021), Definition 1

References

  1. A Theory of Fads, Fashion, Custom, and Cultural Change as Informational Cascades. Sushil Bikhchandani, David Hirshleifer and Ivo Welch, Journal of Political Economy 100 (5), 992–1026, 1992

    The paper that named informational cascades and argued that they explain local conformity and why mass behavior can change suddenly.

  2. A Simple Model of Herd Behavior. Abhijit V. Banerjee, The Quarterly Journal of Economics 107 (3), 797–817, 1992

    The second founding paper of the same year, which calls the pattern herd behavior and shows that the result is inefficient.

  3. Information Cascades and Social Learning. Sushil Bikhchandani, David Hirshleifer, Omer Tamuz and Ivo Welch, NBER Working Paper 28887; published in Journal of Economic Literature 62 (3), 1040–1093 (2024), 2021

    A free review by the founders. Section 2 walks through the simple two-choice model step by step, and later sections cover fragility, kidneys and courts.

  4. Information Cascades in the Laboratory. Lisa R. Anderson and Charles A. Holt, American Economic Review 87 (5), 847–862, 1997

    The first laboratory test with two urns and private draws, the game used as the worked example here.

  5. Experimental Study of Inequality and Unpredictability in an Artificial Cultural Market. Matthew J. Salganik, Peter Sheridan Dodds and Duncan J. Watts, Science 311 (5762), 854–856, 2006

    The online music market with 14,341 participants and 48 unknown songs, in which seeing others' downloads made success more unequal and less predictable.

information-cascadeinformational-cascadeinformation-cascadesherd-behaviorherdingsocial-learningconformityfadsbikhchandani-hirshleifer-welchabhijit-banerjeeanderson-and-holturn-experimentmusiclabbayesian-reasoningbehavioral-economicswisdom-of-the-crowddecision-making