Base Rate Fallacy¶

The systematic human tendency to underweight or ignore the prior probability of a hypothesis when vivid, specific evidence is available, so a posterior estimate collapses toward the likelihood instead of tracking Bayes' rule.

Core Idea¶

The base rate fallacy is the systematic tendency for human reasoners to underweight or ignore prior probability — the base rate — when a vivid, specific piece of evidence is available, producing posterior probability estimates that track the specific evidence far more than Bayes' rule warrants. The canonical form: a test for a disease that affects 1 in 1000 people has a 99% sensitivity and a 5% false-positive rate; a positive result still yields a true positive probability below 2%, but most people, including trained clinicians, report intuitions near 95%. The specific evidence (this test is highly accurate; the result was positive) crowds out the prior (the disease is rare) almost entirely.

The proximate mechanism is representativeness (Kahneman and Tversky): reasoners judge the probability of a hypothesis by how well the specific evidence resembles or fits the hypothesis category, treating match as a proxy for probability. Because similarity to the hypothesis does not encode the hypothesis's prior probability, the base rate drops out. This is compounded by the relative vividness of the two information types: a positive test result or a detailed case description is concrete and imageable; a population prevalence statistic is abstract and easily set aside. Gigerenzer demonstrated that presenting the same information as natural frequencies — "10 out of 1000 people have the disease; of those 10, 9 test positive; of the 990 without it, 49 test positive" — dramatically reduces the fallacy, because the frequency format makes the prior population structure explicit and visible in a way that abstract percentages do not.

Structural Signature¶

Sig role-phrases:

the reasoner forming a posterior — a person estimating how probable a hypothesis is given evidence, the substrate of the fallacy
the vivid specific evidence — a concrete, imageable signal (a positive test result, a detailed case) with a high likelihood under the hypothesis
the dull rare prior — the base rate, an abstract population-prevalence statistic easily set aside, often low
the representativeness substitution — judging probability by how well the evidence resembles the hypothesis category, using match as a proxy for probability
the dropped base rate — the prior silently omitted, because resemblance does not encode prior probability
the likelihood-tracking posterior — the resulting estimate collapsing toward the likelihood (e.g. ~95% from a 99%-accurate test) rather than the true Bayesian value (<2%)
the danger-zone trigger — the failure acute precisely when a rare base rate meets a vivid high-likelihood signal; common priors or dull evidence recede the danger
the natural-frequency repair — recasting the figures as frequencies that expose the prior population structure, the representational (not exhortatory) fix that the fallacy resists and that exhortation alone cannot achieve

What It Is Not¶

Not the claim that a high-accuracy test warrants a high posterior. A 99%-accurate positive test for a 1-in-1000 disease yields a true-positive probability below 2%, because the rare prior dominates. Confusing the likelihood (how characteristic the evidence is if the hypothesis holds) with the posterior (how probable the hypothesis is given the evidence and its prior) is the core move the fallacy names.
Not evidence that the test is inaccurate. The low posterior is not a flaw in the instrument; a perfectly good 99%-sensitivity test still produces mostly false positives when the condition is rare, because the prior is what pins the answer low. Blaming the test misreads a correct Bayesian consequence as a measurement failure.
Not an error in Bayes' rule. The mathematics was always right and fully general; the fallacy is the human failure to apply it — representativeness substituting resemblance-to-the-hypothesis for probability, so the prior silently drops out. In non-human systems that integrate the prior by construction there is no fallacy at all, only the rule correctly used.
Not mere innumeracy or "ignoring statistics." The failure is specific: underweighting the prior while integrating it with a vivid likelihood, not a general inability with numbers. People who can compute fluently still commit it when the evidence is concrete and the base rate abstract, because the mechanism is representativeness, not arithmetic incompetence.
Not curable by warning people to be careful. Because the error is structural rather than a lapse of effort, telling reasoners the answer is counterintuitive leaves their estimate near the likelihood; only a representational fix — recasting the figures as natural frequencies that expose the prior population structure — moves it toward the true posterior. Exhortation does not restore the dropped prior.
Not the same as its sibling representativeness errors. Underweighting the prior when integrating it with a likelihood is distinct from confusing the likelihood with the posterior outright (the inverse-probability fallacy) and from judging a conjunction more probable than its conjunct (the conjunction fallacy). Same heuristic family, different specific errors — a repair tuned to one is not automatically licensed for the others.

Scope of Application¶

The base rate fallacy lives in one domain — human judgment and decision-making — and the contexts below are application settings of that single substrate (a human reasoner forming a posterior), not structurally distinct systems. The Bayesian rule it violates is fully substrate-general and travels everywhere as the parent; but where the prior is integrated by construction there is no fallacy, so non-agent uses are analogy and stay out of this map.

Medical screening — the canonical mammography / HIV-test problem: a positive result on a high-sensitivity test for a rare condition is over-read by patients and clinicians alike.
Criminal forensics — DNA and other random-match probabilities presented without population context produce wrongful confidence in a courtroom.
Hiring and admissions — a vivid interview signal is weighted without anchoring on the base rate of comparable candidates, driving over-confident predictions.
Security and counter-terrorism screening — profiling and anti-terror systems where rare-event base rates dominate the math and false positives swamp true ones.
Intelligence analysis — vivid specific intelligence overrides cool prior probabilities, a recurring structural pattern behind analytic failures.
Personal risk self-assessment — "I know someone who had X happen" overrides population statistics in everyday probability judgments.

Clarity¶

Naming the fallacy separates two quantities that intuition routinely collapses into one: the likelihood — how characteristic the evidence would be if the hypothesis were true (a positive result from a 99%-accurate test) — and the posterior — how probable the hypothesis is given that evidence together with its prior. The felt answer to "the test is positive, how worried should I be?" reports the likelihood while seeming to report the posterior, and without the label the gap between them is invisible; the rare-disease case simply feels like a near-certainty. With it, a reasoner can recognize that a high likelihood says nothing about how common the hypothesis was to begin with, and that omitting the prior is precisely the move the fallacy names.

That distinction converts a fuzzy worry into a structured triple — prior probability, likelihood ratio, posterior — whose arithmetic forces the usually-missing term into view. The sharp question it licenses is diagnostic: am I integrating the base rate, or is the vividness of this specific result overriding it; what would my estimate be if I were made to write the prior down first? And it points at the corresponding repair, which is representational rather than exhortatory: recasting the same numbers as natural frequencies makes the prior population structure explicit and visible, so the base rate can no longer quietly drop out. The contribution is not new mathematics — Bayes' rule was always there — but a name for the specific place human inference fails to apply it, and a check that localizes the failure to the neglected prior.

Manages Complexity¶

Over-confident inferences recur across medical screening, DNA forensics, hiring, security profiling, and intelligence analysis, and each arena seems to need its own account of why the experts there get the numbers wrong. The fallacy compresses all of them to one diagnosis with a fixed shape: whenever a posterior is being formed, vivid specific evidence crowds out the prior because representativeness substitutes resemblance for probability, and the base rate silently drops out. An analyst then needs to track only the same three quantities everywhere — prior probability, likelihood, posterior — to predict in advance where intuition will fail (rare base rate plus a vivid, high-likelihood signal is the danger zone) and by roughly how much, rather than re-litigating each domain's confidence error from scratch. The compression extends to the repair: the failure localizes to a single neglected term, and one representational move — recasting the figures as natural frequencies that expose the prior population structure — addresses it identically across screening, courtroom, and forensic settings. A wide catalogue of "everyone here is over-confident" puzzles thus reduces to one missing-prior structure with a known trigger condition and a uniform fix.

Abstract Reasoning¶

The fallacy licenses a tight set of inferences about when a confident posterior should be distrusted, what to change to repair it, and which problems are dangerous in advance.

Diagnostic — from an over-confident estimate back to the dropped prior. The governing move runs from a surface signature (a reasoner reports a posterior near the likelihood — "the test is 99% accurate and positive, so I'm about 95% likely to have it") to a hidden error (the prior was never integrated; representativeness substituted resemblance-to-the-hypothesis for probability, so the base rate silently dropped out), against the naive reading that a high-accuracy test warrants a high posterior. The tell is a posterior that tracks the specific evidence while the prior is rare and abstract: when the felt answer to "how worried should I be?" equals the test's accuracy, the prior has been omitted by construction. The discriminating self-probe is built in — would my estimate change if I were made to write the base rate down first? — and a posterior that does not move when the prior is forced into view is the fingerprint of the fallacy rather than of a correctly computed near-certainty.

Interventionist — what to change, and the predicted effect on the estimate. Because the failure localizes to a single neglected term, the corrective lever is representational, not exhortatory: recast the same figures as natural frequencies that expose the prior population structure ("10 of 1000 have it; of those 10, 9 test positive; of the 990 without it, 49 test positive"), and predict that the estimate drops sharply toward the true posterior because the prior can no longer quietly disappear. The model is specific about direction and about what will not work — telling reasoners the answer is counterintuitive, or urging them to be careful, leaves the estimate near the likelihood, because the error is structural rather than a lapse of effort; only making the prior visible moves it. It also predicts the magnitude of the correction scales with how rare the base rate is: the rarer the prior, the larger the gap the reframing closes.

Boundary-drawing — the danger zone, and where the inference does not extend. The fallacy marks its own trigger condition, turning "people get probabilities wrong" into a regime prediction: the failure is acute precisely when a rare base rate meets a vivid, high-likelihood signal (a positive test for a 1-in-1000 disease, a DNA match presented without population context, a detailed case that fits a category). Where the prior is common, or the evidence dull and abstract, intuition tracks Bayes more closely and the danger recedes. The effect also fixes an internal boundary against its neighbors that licenses or blocks inference: it names underweighting the prior when integrating it with a likelihood — distinct from confusing the likelihood with the posterior outright (the inverse-probability fallacy) and from judging a conjunction more probable than its conjunct (the conjunction fallacy) — same representativeness family, different specific errors, so a repair aimed at one is not automatically licensed for the others.

Predictive — anticipating the error before it is made. Holding the three quantities — prior, likelihood, posterior — the model predicts in advance who will be over-confident and roughly by how much, without studying each arena separately: clinicians reading rare-disease screens, jurors hearing random-match statistics, analysts weighing vivid intelligence against cool priors all land in the same danger zone, and all over-state the posterior by approximately the amount the omitted prior would have discounted it. The same structure predicts the form of the residual error after a partial fix — an estimate still biased toward the evidence-only value whenever the prior is present but underweighted rather than fully restored — and so tells a designer to build the prior into the representation at the point of judgment rather than trusting it to be recruited on demand.

Knowledge Transfer¶

Within judgment and decision-making the diagnosis transfers as mechanism, intact, with the standing caveat that its "domains" are application contexts of one substrate — a human reasoner forming a posterior — not structurally distinct systems. With that understood, the prior-likelihood-posterior triple, the danger-zone trigger (rare base rate plus vivid high-likelihood signal), and the natural-frequency repair carry without translation across medical screening (the mammography/HIV-test problem), criminal forensics (DNA random-match probabilities read without population context), hiring and admissions (interview signal weighted without candidate base rates), security and counter-terror screening (rare-event base rates dominating the math), and intelligence analysis (vivid specific intelligence overriding cool priors). The vocabulary (base rate, likelihood, posterior, representativeness), the diagnostic (a posterior that tracks the evidence while the prior is rare and abstract; the self-probe "would my estimate change if I wrote the prior down first?"), and the intervention (recast as natural frequencies that expose the prior population structure; predict the estimate drops sharply, and that exhortation alone will not move it) all move freely — because the same representativeness-substitutes-for-probability machinery is at work in every case.

Beyond the human reasoner the fallacy does not travel — and the reason is sharper here than for most cognitive entries. The underlying mathematics (Bayes' rule, the prior-likelihood-posterior integration) is fully substrate-general and genuinely recurs across biology, physics, statistics, and computation — but in those settings there is no fallacy, only the rule correctly applied; a Bayesian filter or a population-genetics calculation integrates the prior by construction. So what travels cross-domain is the parent — Bayesian updating, the normative structure the fallacy violates — not "the base rate fallacy," which names specifically the human deviation from that structure. Invoking "base rate neglect" for a system that fails to weight a prior is therefore an analogy unless that system is itself a reasoning agent prone to substituting resemblance for probability; absent a representativeness-style mechanism, the right description is just "it underweighted the prior," and the load-bearing content is Bayesian updating plus the general lesson a posterior requires integrating a prior with a likelihood, and absent that integration the posterior collapses to the likelihood. The fallacy's own cargo — the representativeness mechanism, the vividness asymmetry, the natural-frequency debiasing, the clinician-and-juror failure cases — stays bound to human cognition, where it sits as one named violation under bayesian_updating, representativeness, and the broader heuristics-and-biases family. The honest cross-domain account: as mechanism the fallacy stays inside human judgment; the Bayesian rule it violates travels everywhere as the parent; and any use of the fallacy outside a reasoning agent is borrowing the name. Carry Bayes, not the fallacy (see Structural Core vs. Domain Accent).

Examples¶

Canonical¶

The rare-disease screening problem is the defining instance. A disease affects 1 in 1,000 people; a test has 99% sensitivity and a 5% false-positive rate. Given a positive result, what is the probability the person is actually sick? Work it in natural frequencies over 1,000 people: 1 person has the disease and (at 99% sensitivity) tests positive — about 1 true positive. Of the 999 healthy people, 5% test positive — about 50 false positives. So roughly 51 people test positive and only 1 is sick: the true-positive probability is about 1/51 ≈ 1.9%, under 2%. In a much-cited 1978 study, Casscells, Schoenberger, and Graboys posed essentially this problem to staff and students at a Harvard teaching hospital; the most common answer was 95%, and only a small fraction gave the correct answer near 2%.

Mapped back: The 1-in-1,000 prevalence is the dull rare prior, easily set aside; the positive result from a 99%-accurate test is the vivid specific evidence. The modal 95% answer is the likelihood-tracking posterior, collapsing onto the test's accuracy while the base rate is dropped. Working the problem as "1 sick and positive versus 50 healthy and positive" is the natural-frequency repair that exposes the prior population structure.

Applied / In Practice¶

The wrongful conviction of Sally Clark in England shows the same error in a courtroom. Clark was convicted in 1999 of murdering her two infant sons after the pediatrician Roy Meadow testified that the chance of two cot deaths (SIDS) in one family was about 1 in 73 million — a figure he reached by squaring a single-death rate, itself compounding the statistical error. The vivid "1 in 73 million" number was heard as the probability of her innocence. But the relevant comparison was between two rare hypotheses — double SIDS versus double murder — and the base rate of double infant murder is also extremely low; ignoring it (the prosecutor's fallacy) made a rare-but-real natural explanation look impossible. The Royal Statistical Society formally objected, and the conviction was quashed in 2003.

Mapped back: The prior probability of the murder hypothesis, itself rare, is the dropped base rate; the "1 in 73 million" statistic is the vivid specific evidence. Treating that figure as the chance of innocence is the likelihood-tracking posterior, and the setup — a rare event paired with a striking, concrete number — is exactly the danger-zone trigger the fallacy predicts.

Structural Tensions¶

T1: Genuine error versus rational discounting (not every ignored prior is a fallacy). The fallacy names underweighting a base rate that should be integrated — but not every neglected prior is an error, and treating all prior-discounting as the fallacy over-applies it. A prior can be drawn from the wrong reference class, be non-stationary, or be causally irrelevant to the individuating case at hand, and where the specific evidence is genuinely more diagnostic, letting it dominate is correct, not fallacious. The literature shows people do use base rates that are causal or clearly relevant, and rightly discount incidental ones. The tension is that the concept must both flag the real failure (a rare, relevant prior crowded out by a vivid signal) and not indict the reasonable discounting of a prior that does not apply — and the line between them is a substantive judgment, not a formula. Diagnostic: Is the neglected base rate the correct, relevant prior for this hypothesis (neglect is the fallacy), or an incidental or ill-fitting reference class the individuating evidence rightly overrides?

T2: The base rate versus which base rate (the reference-class problem). The fallacy is stated as though the prior exists, waiting to be integrated — but a base rate is only defined relative to a reference class, and the "true posterior" shifts with the class chosen. Is the relevant prior the disease's prevalence in the whole population, in this patient's age-and-risk group, or in those presenting these symptoms? Each yields a different, defensibly "correct" answer. The concept's crispness — integrate the prior and the estimate drops sharply — quietly presupposes a settled reference class, yet selecting it is itself a judgment with no unique answer. The tension is that the fallacy indicts dropping the prior while offering no principle for which prior, so a reasoner can integrate a base rate faithfully and still be wrong if the reference class was mischosen. Diagnostic: Has the reference class defining this base rate been chosen to actually fit the case, or is a convenient population statistic being integrated as though it were the uniquely correct prior?

T3: Natural-frequency repair versus its fragility (the fix depends on a clean frame). Recasting the figures as natural frequencies reliably moves estimates toward the true posterior because it exposes the prior population structure the abstract percentages hid — the concept's signature intervention. But the fix is representational, not conceptual, and it is fragile: it works only when the problem can be cleanly cast as frequencies over a well-specified reference class, and a wrong reference class makes the frequency framing confidently misleading rather than merely omitted. People debiased by one frequency framing often fail to transfer the correction to a differently-framed problem, and many real judgments admit no tidy count. The tension is that the repair's power comes from a specific representational format whose availability and correctness are not guaranteed at the point of judgment. Diagnostic: Can this problem be cast as natural frequencies over the right reference class — or is the frequency framing unavailable, or built on a reference class that will mislead as confidently as the percentages did?

T4: Structural error versus the reach of the fix (expertise does not inoculate). Because the failure is structural — representativeness substituting resemblance for probability — it is exhortation-proof: warning reasoners that the answer is counterintuitive, or that they should be careful, leaves the estimate near the likelihood, and fluent, numerate, expert reasoners (the Harvard clinicians, the testifying pediatrician) commit it anyway. That is the concept's sharp claim, and its uncomfortable consequence: since neither training nor effort restores the dropped prior, the entire corrective burden falls on engineering the representation at the moment of judgment — which a designer cannot always control. The tension is that the same structural nature that makes the fallacy robust against caution also makes it robust against expertise, so the only reliable defense (building the prior into the representation) is precisely the one not available wherever the judgment context is uncontrolled. Diagnostic: Is there a way to build the prior into the representation at the point of decision — or is the setting one where, expertise and warnings being useless, the fallacy has no available structural fix?

T5: Autonomy versus reduction (a named cognitive fallacy or the human violation of Bayes' rule). The base rate fallacy is a genuine, canonically studied cognitive failure with its own cargo — the representativeness mechanism, the vividness asymmetry, the natural-frequency debiasing, the clinician-and-juror cases — and within human judgment it transfers as literal mechanism across screening, forensics, hiring, security, and intelligence, because all are one substrate (a reasoner forming a posterior). But its cross-domain reach is unusually clean to diagnose: what travels everywhere is the parent, Bayesian updating — the normative structure the fallacy violates — not the fallacy itself, because in non-human systems that integrate the prior by construction there is no fallacy, only the rule correctly applied. Calling a filter that underweights a prior "base rate neglect" is analogy unless the system is a reasoning agent prone to substituting resemblance for probability; absent that mechanism, the right description is just "it underweighted the prior." Diagnostic: Resolve toward the parent (Bayesian updating) whenever the lesson must travel beyond a human reasoner; toward the named fallacy only when representativeness and the vividness asymmetry are producing a human's over-confident posterior in situ.

Structural–Framed Character¶

The base rate fallacy sits at the framed-leaning position — the standard home for a named cognitive fallacy, patterning with the other judgment entries and sitting clearly framed of a substrate-general structure like isostasy, even though the structure it violates is unusually portable. The five criteria run mostly framed, with one structural pull, and the split is sharper here than for most cognition entries. On evaluative weight it reads firmly framed: "fallacy" is a verdict — the concept names a violation of Bayes' rule, an over-confident posterior that departs from the normatively correct value, so it flags an error against a precise standard rather than describing a value-neutral mechanism. On human-practice-bound it reads firmly framed: the fallacy exists only in a reasoning agent prone to substituting resemblance for probability, and the entry is emphatic that "in non-human systems that integrate the prior by construction there is no fallacy, only the rule correctly applied" — remove the human reasoner and nothing remains but Bayes, correctly used. On institutional origin it reads structural: this is a discovered empirical regularity of cognition (Kahneman and Tversky's representativeness; the Casscells 1978 demonstration), a fact about how minds drop priors, not an artifact any survey invented. On vocab-travels it reads framed for the named concept: the signature cargo — representativeness, the vividness asymmetry, natural-frequency debiasing, the clinician-and-juror failure cases — is pinned to human cognition, while the mathematics beneath it is universal. On import-vs-recognize the profile is bimodal and, the entry stresses, "sharper here than for most cognitive entries": within human judgment the same mechanism is recognised across screening, forensics, hiring, security, and intelligence, but beyond the reasoner only the parent rule travels, and calling a filter that underweights a prior "base rate neglect" is analogy unless the system is itself a resemblance-substituting agent.

The portable structural skeleton is Bayesian updating — the normative prior-times-likelihood-to-posterior integration, and the general lesson that a posterior requires combining a prior with a likelihood or it collapses to the likelihood — with representativeness as the proximate mechanism of the human deviation. That skeleton is fully substrate-spanning — indeed a pure mathematical structure that recurs across biology, physics, statistics, and computation — but it is exactly what the base rate fallacy violates and thereby instantiates from its umbrella primes bayesian_updating and representativeness, not what lets "the base rate fallacy" itself travel: the entry is explicit that "what travels cross-domain is the parent — Bayesian updating — not the base rate fallacy, which names specifically the human deviation." So the cross-domain reach belongs to bayesian_updating, while the domain-accented specifics — the resemblance-for-probability substitution, the vividness asymmetry, the frequency-format repair — stay home and reach non-agents only by borrowing the name. Its character: an error-defining cognitive fallacy, a discovered regularity of the human reasoner, whose substrate-spanning content is the pure Bayesian rule it violates and whose distinctive machinery is bound to representativeness-prone cognition — framed-leaning, structural only in the universal updating rule it fails to apply.

Structural Core vs. Domain Accent¶

This section decides why the base rate fallacy is a domain-specific abstraction and not a prime, and it carries the case for its domain-specificity — there is no separate section for that.

What is skeletal (could lift toward a cross-domain prime). Strip away the human reasoner and a thin relational structure survives: a posterior estimate requires integrating a prior with a likelihood, and where the prior is omitted the posterior collapses toward the likelihood. The portable pieces are abstract and, in fact, purely mathematical — a prior probability, a likelihood, their Bayesian combination into a posterior, and the failure signature of an estimate that tracks the likelihood alone. That skeleton is not merely substrate-portable but substrate-universal: Bayes' rule recurs across biology, physics, statistics, and computation, and any system that integrates a prior by construction realizes it. Which is exactly why it sits as the parent primes the entry instantiates — bayesian_updating (the normative prior-times-likelihood-to-posterior integration) with representativeness as the proximate mechanism of the deviation. But note the peculiar relation: the fallacy does not use that structure, it violates it. The portable core is the rule the entry fails to apply, not what makes the entry distinctive.

What is domain-bound. Everything that makes the concept the base rate fallacy in particular is judgment-and-decision-making furniture that does not survive extraction: the human reasoner forming a posterior; representativeness substituting resemblance-to-the-hypothesis for probability, so the prior silently drops; the vividness asymmetry (a concrete positive test or detailed case crowds out an abstract prevalence statistic); the danger-zone trigger (rare base rate meets vivid high-likelihood signal); and the natural-frequency repair — a representational fix that exhortation cannot substitute for. These are the worked vocabulary, the mechanism, and the empirical cases — the Casscells 1978 Harvard clinicians answering 95%, the Sally Clark 1-in-73-million miscarriage of justice — the field actually studies. The decisive test, sharper here than for most cognitive entries: remove the reasoning agent — hand the same numbers to a Bayesian filter or a population-genetics calculation that integrates the prior by construction — and there is no fallacy at all, only the rule correctly applied. The error is constituted by a mind prone to representativeness; without that mind, "underweighting a prior" is a plain description, not this named failure.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose cross-domain transfer is recognition of the same mechanism, not analogy. The base rate fallacy's transfer is bimodal, and its cross-domain split is unusually clean to diagnose. Within human judgment the fallacy travels as literal mechanism — the prior-likelihood-posterior triple, the danger-zone trigger, and the natural-frequency repair carry without translation across medical screening, criminal forensics, hiring, security screening, and intelligence analysis, because all are one substrate, a reasoner forming a posterior. Beyond the reasoner the fallacy does not travel: what recurs everywhere is the parent, bayesian_updating, the normative structure the fallacy violates — and in those systems there is no fallacy, only Bayes correctly used. Calling a filter that fails to weight a prior "base rate neglect" is analogy unless the system is itself a resemblance-substituting agent; absent that mechanism, the right description is simply "it underweighted the prior." So when the bare structural lesson is needed cross-domain — a posterior requires combining a prior with a likelihood — it is already carried, in fully general and indeed mathematical form, by bayesian_updating. The cross-domain reach belongs to that parent; the "base rate fallacy," as named, keeps its representativeness-and-vividness machinery bound to human cognition and reaches non-agents only by borrowing the name. Carry Bayes, not the fallacy.

Relationships to Other Abstractions¶

Current abstraction Base Rate Fallacy Domain-specific

Parents (3) — more general patterns this builds on

Base Rate Fallacy is a kind of Bias Prime

Base-rate fallacy is bias specialized to a posterior estimate that is systematically displaced toward vivid likelihood evidence because its prior is underweighted.
Base Rate Fallacy is part of Representativeness Heuristic Domain-specific

Base-rate fallacy contains representativeness because resemblance or vivid likelihood evidence supplies the proxy estimate that displaces the prior.
Base Rate Fallacy presupposes Bayesian Updating Prime

The base-rate fallacy presupposes Bayesian updating because it is defined by a posterior that fails to integrate the prior with the likelihood.

Hierarchy paths (9) — routes to 7 parentless roots

Base Rate Fallacy → Bias

Show alternative paths (8)

Not to Be Confused With¶

Prosecutor's fallacy / confusion of the inverse (inverse-probability fallacy). Directly transposing the conditional — reading the likelihood P(evidence | hypothesis false), such as a small random-match probability, as the posterior P(hypothesis false | evidence). This differs from the base rate fallacy's specific move of underweighting or omitting the prior while integrating it with a likelihood; the inverse fallacy skips the integration altogether by equating two different conditionals. They frequently co-occur (the Sally Clark case involves both), which is exactly why they are confused. Tell: is the error dropping the base rate while combining it with the evidence (base rate fallacy), or transposing the conditional — reading P(E | not-H) as P(not-H | E) — with no prior in view at all (prosecutor's/inverse fallacy)?
Conjunction fallacy. Judging a specific conjunction more probable than one of its constituents — the Linda-the-bank-teller error, where "feminist bank teller" is rated likelier than "bank teller." It shares the representativeness family with the base rate fallacy but is a different specific error: it concerns violating the conjunction rule of probability, not neglecting a prior when forming a posterior. Tell: is the mistake rating a conjunction more probable than its own conjunct (conjunction fallacy), or underweighting a prior while estimating a posterior (base rate fallacy)?
Representativeness heuristic. The proximate mechanism — judging probability by how well evidence resembles or fits a category, using match as a proxy for probability. The base rate fallacy is one consequence of representativeness (the prior drops out because resemblance does not encode it), not the heuristic itself; representativeness also produces the conjunction fallacy and other errors. The relation is mechanism-to-symptom. Tell: is the reference to the general judge-by-resemblance shortcut (representativeness), or specifically the dropped-prior posterior error it produces (base rate fallacy)?
Conservatism bias. The opposite updating failure: under-weighting new evidence and clinging too closely to the prior, so the posterior moves too little when data arrive (Edwards). Where the base rate fallacy collapses the estimate onto the likelihood and ignores the prior, conservatism anchors on the prior and under-uses the likelihood. A pure contrast case — same Bayesian machinery, opposite direction of error. Tell: is the estimate stuck too near the prior, barely moving on new evidence (conservatism), or collapsed onto the likelihood with the prior ignored (base rate fallacy)?
Availability heuristic. Estimating a probability or frequency by the ease with which instances come to mind, so vivid or recent cases inflate the estimate. It overlaps with the base rate fallacy's vividness asymmetry and can co-drive everyday misjudgments ("I know someone who had X happen"), but availability is about retrieval fluency supplying the estimate, whereas the base rate fallacy is about failing to integrate a known prior with specific evidence. Tell: is the judgment driven by how easily examples are recalled (availability), or by omitting a known base rate when weighing a specific signal (base rate fallacy)?
Bayesian updating (the parent rule it violates). The substrate-universal normative structure — combine a prior with a likelihood to obtain a posterior — that the base rate fallacy violates, and which the catalog carries as bayesian_updating. This is not a peer error but the correct rule the fallacy fails to apply; a Bayesian filter or population-genetics calculation that integrates the prior by construction realizes the rule with no fallacy at all. Calling such a system's under-weighting "base rate neglect" is analogy unless it is a resemblance-substituting agent. Tell: strip away the human reasoner and the representativeness mechanism and what remains is the correct prior-times-likelihood integration — at which point you are using bayesian_updating, and there is no fallacy to name. (Treated fully in Structural Core vs. Domain Accent and Knowledge Transfer.)

Neighborhood in Abstraction Space¶

Base Rate Fallacy sits in a moderately populated region (50^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Unclustered & Miscellaneous (309 abstractions)

Nearest neighbors

Conservatism Bias — 0.89
Conjunction Fallacy — 0.87
Cherry Picking — 0.85
Ad Nauseam (Argument from Repetition) — 0.83
Suppressed Evidence — 0.83

Computed from structural-signature embeddings · 2026-07-12