Conjunction fallacy

The stargazing claims adjuster

Jordan spent college in the astronomy club, builds his own telescopes and runs monthly stargazing nights. Which is more likely? (a) Jordan is an insurance claims adjuster. (b) Jordan is an insurance claims adjuster who volunteers at the local planetarium.

Many people pick (b), because it’s the answer that sounds like Jordan. But every claims adjuster who volunteers at a planetarium is a claims adjuster, so (a) covers everyone (b) covers and more. This is the structure of Tversky and Kahneman’s “Linda” problem (see Evidence).

The favorite’s comeback

At a club chess final, the top-rated player faces a newcomer in a best-of-five match. A spectator says the likeliest story is “she’ll drop the first game, then win the match”, and rates that as more likely than “she’ll drop the first game”.

Dropping the first game and then winning is one of the ways of dropping the first game. The comeback story is appealing because it matches what a strong player does, but it can’t be more probable than the plain event it contains. Tversky and Kahneman found the same pattern in predictions about a champion tennis player.

Two estimates that never meet

In a planning meeting, a team estimates a 30% chance that their app launch slips past March. A week later, in a risk review, a different group from the same team estimates a 45% chance that the launch slips past March because the app store review takes longer than expected.

Nobody compared the two statements side by side, so nobody chose “A and B” over “A”. But the second estimate describes one specific way the first could happen, and it received the higher number. This is the less visible form: a detailed scenario rated as more likely than the bare outcome it’s part of, with the inconsistency spread across separate judgments.

• Direct: both statements appear together and the combination is picked or ranked above its part. The first example above.
• Indirect: the combination and its part are judged separately, often by different people, and the combination gets the higher estimate. Tversky and Kahneman called violations found this way conjunction errors, reserving conjunction fallacy for the direct comparison.
• Scenario detail: a plausible cause is added to an outcome (“the launch slips because...”). The cause makes the outcome easier to imagine and rates higher, although the outcome alone must be at least as likely. Tversky and Kahneman also noted that overestimating chains of events that must all go right feeds unwarranted optimism about whether a plan will succeed, one link to the Planning fallacy.

• When the question is really conditional. “If Jordan is a claims adjuster, is he likely to volunteer somewhere science-related?” is a sensible question, and the answer can be yes. The fallacy is treating how well a detail fits as the probability of the whole combination.
• When the plain option is reasonably read as excluding the detail. If “Jordan is a claims adjuster” is understood as “a claims adjuster who does not volunteer at a planetarium”, then ranking the other option higher breaks no rule. Whether people actually read it that way is a question the research has tested (see Evidence).
• When the more detailed claim is preferred for usefulness, not probability. A specific forecast can be more worth making than a vague one, even though it’s less likely to be right.

The test: is every case where both happen also a case where the first one happens? If so, the combination can’t be the likelier of the two.

“If the front arrives”

A forecaster says there’s a 30% chance of rain tomorrow. Then she adds: “If the cold front arrives by morning, the chance of rain is 70%.”

It sounds like a detailed scenario (“front and rain”) getting a higher number than the plain event (“rain”). But 70% is the chance of rain given that the front arrives, not the chance that the front arrives and it rains. A conditional probability can be higher than the unconditional one; the probability of both events together can’t. This is the conditional question condition.

The narrower forecast

A contractor tells a client the kitchen renovation will take five to six weeks. Asked why he didn’t say “three to ten weeks”, which is more likely to be right, he says the wider range wouldn’t help anyone plan.

The specific forecast is contained in the wider one, so it’s less probable, and the contractor knows it. He isn’t claiming it’s more likely; he’s trading some probability for a forecast the client can use. Tversky and Kahneman made this point themselves: a narrow estimate can be more valuable than a wide one. That’s preferring detail for usefulness, stated openly.

Tversky and Kahneman’s explanation is the representativeness heuristic: judging how probable something is by how well it matches a description or a stereotype. “Claims adjuster who volunteers at a planetarium” resembles Jordan more than “claims adjuster” does, so it feels more likely. In their studies, people’s rankings by probability and by resemblance were almost identical.

A detailed story also hangs together better, and coherence is easy to mistake for likelihood. That is the habit described by “what you see is all there is”: confidence tracks how well the story fits, not how many ways it could fail.

Ralph Hertwig and Gerd Gigerenzer proposed a different account: in everyday speech “probable” can mean plausible or credible, and a participant trying to make sense of the question may reasonably take it that way. On that reading, many answers counted as errors aren’t errors of reasoning at all. Later studies designed to remove those ambiguities still found the effect (see Evidence), but the account helps explain why its size varies so much with wording.

• Count cases instead of rating a single story. Picture 100 people who fit the description, and ask how many are claims adjusters, then how many of those also volunteer. Frequency wording has been tested repeatedly and usually reduces the error substantially, though it doesn’t reliably eliminate it:
  • In Tversky and Kahneman’s health-survey problem, 65% of respondents violated the rule when asked for percentages; 25% did when asked “how many of the 100 participants”.
  • Asking for the frequency of the single category first cut violations from 65% to 31%, and combined with the frequency wording to 11%.
  • Hertwig and Gigerenzer found 13% violations for a frequency version of the Linda problem, against 88% for the probability version, in small groups.
  • In a planned adversarial collaboration, frequency formats alone did not remove the effect. A 2021 preregistered replication still found it in frequency estimates for the Linda profile.
• Ask what the extra detail adds. If it makes the story more convincing but must also be true for the story to hold, it lowers the probability.
• Stakes help a little. A 2023 meta-analysis of about 3,300 participants found a small benefit of incentives (d = 0.19). In Tversky and Kahneman’s betting version of Linda, 56% still bet on the combination.

Status: replicates robustly. Violations of the conjunction rule have been reproduced many times, including by the effect’s leading critics and in a preregistered replication. Its size depends heavily on format, and its interpretation was disputed for years, but most of the tests built to settle that dispute found a real error rather than a misreading.

• Tversky and Kahneman (1983) described Linda, a 31-year-old philosophy graduate concerned with discrimination and social justice. Asked which was more probable, “Linda is a bank teller” or “Linda is a bank teller and is active in the feminist movement”, 85% of 142 undergraduates chose the combination. To check whether people were reading “bank teller” as “bank teller and not a feminist”, a new group rated both statements on a 9-point scale, where there’s no reason for that reading; 82% of 119 still rated the combination higher. In ranking tasks, statistically naive students and doctoral students in decision science violated the rule at similar rates (89% and 85% for Linda). At a 1982 forecasting conference, professional forecasters who rated a diplomatic breakdown preceded by a plausible cause gave it a higher probability than another group gave the breakdown alone. The authors cautioned that their problems were built to elicit errors and don’t estimate how common the errors are.
• Hertwig and Gigerenzer (1999) argued that “probability” has non-mathematical meanings and showed that people infer them in the Linda problem. In their studies, frequency wording cut violations sharply, which they attributed to the word “frequency” pinning down a mathematical reading.
• Mellers, Hertwig and Kahneman (2001) ran an adversarial collaboration on two of Hertwig’s claims: that frequency formats eliminate the effect, and that people read “and” as “or”. Frequency formats alone did not eliminate it. When filler items were removed, the effect disappeared with Hertwig’s rephrased conjunctions (“and are”, “who are”). The two sides interpreted the results differently.
• Sides, Osherson, Bonini and Viale (2002) used bets and unambiguous conjunctions to reduce both possible misreadings. Conjunction fallacies were as frequent under betting instructions as under probability instructions.
• Chandrashekar and colleagues (2021) ran a preregistered close replication of the 2001 study with 1,032 online participants. For the Linda profile, the combination received higher frequency estimates than the single category (d of about 0.5), and “and” versus “and are” made no difference. For a second profile, “James”, there was no conjunction effect, which the authors relate to evidence that the effect is sensitive to context.

The interpretation dispute doesn’t make the effect “contested” under this site’s criteria, because it was tested directly. Betting instructions, clarified conjunctions and frequency formats each left the error in place in at least some well-controlled studies, including a preregistered one; the effect disappeared only under particular combinations of wording and format. That makes wording a strong moderator, not evidence that the effect is a misreading.

What remains uncertain is its size and reach. No multi-lab replication of the classic probability version turned up in reviewing this entry; the one preregistered replication cited here used a frequency format and found the effect for only one of its two profiles; and rates vary widely with wording. Researchers also still debate which account (representativeness or something else) best explains it.