Insensitivity to Sample Size¶
Elicited judgments differentiate sample sizes too little relative to how a stated sampling model changes a sample statistic's probability or spread.
Core Idea¶
Insensitivity to sample size is a judgment error observed when probability or dispersion estimates elicited under different sample-size conditions distinguish those conditions less than a specified sampling model warrants. One task may ask the same person to compare sizes; another may assign different people to the size conditions. The inference must respect which comparison was made. When observations are comparable and sufficiently independent, a sample proportion or mean ordinarily varies less at larger \(N\). An extreme proportion that is fairly possible in a small sample can be rare in a much larger one. The error lies in the judgment's weak response to that difference, not in small samples having variable outcomes.[1][2]
Kahneman and Tversky observed the error in two materially different tasks: asking a person to choose which of two hospitals would have more days above a fixed birth-proportion threshold, and asking separate participant groups to assign probabilities across bins of adult male sample-mean height at \(N=10,100,1000\). A later direct replication again found little separation among elicited sampling distributions, while its results for the extreme-outcome choice family were more mixed. These controlled tasks warrant a bounded psychological entry. They do not establish that every person, task, or statistical inference ignores sample size, or that one proposed heuristic is the unique cause of each answer.[1][2]
Structural Signature¶
- Reference population or generator. A specified process and its rate or mean provide the parameter against which a sample statistic is compared. If the underlying populations change too, a difference in outcomes may come from that change rather than \(N\). A literal unit-selection frame is not part of this cognitive error.[1]
- Comparable unequal sample sizes. Two or more \(N\) conditions are compared under a sampling model that permits a meaningful size contrast. The same judge need not see all values when the study assigns different groups to different sizes. Dependence or altered selection can change the expected relation and must be assessed before assigning the bias label.[1]
- Size-sensitive statistic and event. A proportion, mean, or other defined statistic has a model-based dispersion or fixed-tail probability that changes with \(N\). The exact statistic and event matter: a tail above 60% boys is not the same question as the probability of one exact birth count.[1]
- Elicited human judgment. A person estimates a distribution or chooses which sampling outcome is more likely. An extreme value in a dataset, with no such judgment, is a statistical observation rather than an instance of this cognitive error.[1][2]
- Insufficient size adjustment. Elicited judgments across the \(N\) conditions differ less than the relevant sampling model requires, measured at the task design's actual unit of analysis. This comparison identifies the failure without assuming that resemblance to a prototype was experimentally isolated as its sole mental cause.[1][2]
The recognition test is therefore a comparison of two relations: what the model predicts as \(N\) changes and what a judge reports in a within-person task or what separately assigned groups report in a between-condition task. If the sampling assumptions do not support the modeled contrast, or the observed judgments adequately differentiate the size conditions, this particular error has not been shown.
What It Is Not¶
- Not the law of small numbers itself. Small denominators can generate broad empirical spreads. That fact does not say whether anyone misjudged those spreads.[3]
- Not unrepresentative sampling. Sampling (Representativeness) concerns the selection of units and coverage of a target population. A well-selected sample can still be misjudged by its size, and a poorly selected sample can be misdesigned without anyone making the defining judgment.
- Not every use of the Representativeness Heuristic. That broader entry addresses typicality substituted for probability in several forms, including base-rate and set-inclusion errors. The present identity is the narrower failure to calibrate a sample-statistic judgment to \(N\); the observed discrepancy need not prove typicality was the unique cause.
- Not a universal inability to use \(N\). The original authors noted that size invariance may not hold exactly when category counts vary or small samples permit enumeration; the later replication did not find an equal-likelihood modal answer in every extreme-outcome comparison.[1][2]
Scope of Application¶
The entry applies to human judgments about sampling distributions, extreme sample results, or reliability when the population, statistic, unequal sample sizes, and normative comparison are specified. Laboratory vignette and elicitation tasks provide its direct evidence here. The form of the statistic can vary from a binary proportion to a continuous sample mean; the elicited judgments must still underweight the model's \(N\) effect at the study's unit of comparison. Within practical data interpretation, the same diagnostic question can be asked, but a county map or school-score plot alone cannot establish that a particular observer committed the cognitive error.[1][2][3]
The scope has a model boundary. A comparison of raw counts with different denominators, correlated measurements, changing populations, or differently selected samples cannot be labeled size insensitivity merely because one \(N\) is smaller. First determine what the stated sampling process predicts for the chosen statistic and event. The controlled studies support task-bound response patterns, including counterexamples to an invariant all-people claim.[1][2]
Clarity¶
The concept separates three things often compressed into “small samples are misleading.” The statistical relation says how a specified statistic varies with \(N\). The observed judgment records what a person predicts, or what separate assigned groups predict, about that variation. The error exists when the two diverge in the size direction. A researcher may correctly warn that small counties have volatile rates without observing any psychological bias in a map reader. Conversely, a person can see the two hospital sizes and still judge their extreme-day frequencies about equal.[1][3]
It also makes the height example precise. The original article predicted that a particular height interval might be judged similarly across sample sizes, but the actual height task asked different groups for probabilities over seven bins of daily sample-mean height. Its plotted subjective height distributions changed little with \(N\); the paper did not plot an objective height-distribution curve alongside them. Repeating the seed's “above six feet” phrasing as an observed single-threshold result would misstate the experiment.[1]
Manages Complexity¶
Many sample-size questions can be organized with five variables: the reference population, \(N\), the statistic, the modeled distribution or fixed-tail event, and the elicited judgment. The hospital task uses a binary birth proportion and an over-60% tail; the height task uses a continuous sample mean and a binned distribution. Different surface stories become comparable only after those variables are filled. This compact inventory also exposes what is missing in an application case: county and school plots have denominators and observed extremes, but no measured judgment across controlled \(N\) conditions.[1][3]
Response format remains a real variable rather than noise to be suppressed. In the 2025 replication, Problem 6 subjective distributions remained close across sample sizes, but “about the same” was modal in only three of six Problems 7–9 extreme-outcome comparisons. The entry can represent both findings without claiming an invariant answer pattern.[2]
Abstract Reasoning¶
To diagnose a proposed case, specify the generating population and statistic, then calculate or otherwise establish how the chosen probability or spread should change across the given \(N\) values. For a fixed boy-birth rate near one half, the binomial chance that a day's proportion exceeds 60% is greater when about 15 births are observed than when about 45 are observed. Compare that direction with a participant's answer. “About the same” underweights the relevant size change in the original hospital task.[1]
The same reasoning changes what can be inferred from a striking small-area rate. Both very high and very low rates may cluster among small denominators even when the underlying rates do not differ so dramatically. That statistical check can prevent an overconfident policy story, but it is an application of sampling variability. To diagnose this named psychological error in a policy reader, one would still need evidence of the reader's judgment and an appropriate normative comparison.[3]
Knowledge Transfer¶
Within cognitive judgment research, the diagnostic relation transfers from a hospital's binary daily proportions to distributions of continuous sample means. The population, \(N\), statistic, modeled size effect, and reported judgment recur while the materials and response formats change. The replication tests those task families again and shows both persistence and limits in their response patterns.[1][2]
In epidemiology or education, the sample-size mathematics travels literally to suitable rate or score analyses, as the county and school illustrations show. A claim that people in those settings displayed the bias requires separate measured judgment evidence. Moving a statistical caution between fields does not automatically transfer the psychological finding or turn the named entry into a domain-free Prime.[3]
Examples¶
1. Hospital births, an extreme proportion. In Kahneman and Tversky's vignette, the small hospital had about 15 births a day, the large hospital about 45, and both drew from a population described as roughly half boys. Participants judged which hospital would record more days on which boys exceeded 60% of births. In the original more-extreme form, “about the same” was the modal answer; the smaller hospital was correct under the binomial model. Mapped back: the birth process is the reference generator, 15 and 45 are the unequal \(N\) values, the daily proportion above a fixed threshold is the statistic and tail event, the choice is the elicited human judgment, and its weak size differentiation is the discrepancy. The 2025 replication's broader choice-task results limit any claim that equal-likelihood responses always dominate.[1][2]
2. Daily mean adult male height, a distribution. Different participant groups in the original study estimated how often the average height of 10, 100, or 1,000 men would fall in each of seven height bins. The distributions elicited from those separate groups changed little across the \(N\) conditions; this does not show that any one participant compared all three sizes. The direct replication again detected no size-condition differences in mean probability estimates for its Problem 6 height condition; that failure to detect differences is not proof that all individual judgments were exactly equal. Mapped back: the stated male population supplies the reference mean, 10/100/1000 are the sizes, daily sample-mean height across bins is the statistic, the assigned bin probabilities are the elicited judgments, and the limited between-group contraction is the discrepancy from finite-variance sampling expectations. This was a binned-distribution task, not an observed single probability for a mean above six feet.[1][2]
Application boundary, not a third psychological positive. Wainer and Zwerling display county cancer-rate and school-score extremes among small denominators or enrollments. The plots show why size matters to interpretation, but they do not measure a person's probabilities for matched small and large samples. Their school result also differs by grade, so it cannot support a blanket claim that school size has no effect on achievement.[3]
Structural Tensions¶
The original and replication studies identify size-insensitive response patterns and limits that depend on task form.[1][2] They do not establish two opposed pressures with distinct costs that cannot be jointly addressed. “Intuitive typicality versus explicit calculation” names a proposed explanation and a correction, not a demonstrated two-sided trade-off. No structural tension is asserted on that evidence.
Structural–Framed Character¶
The entry is structural within cognitive judgment research because hospital proportions and adult-height means preserve the same model-versus-judgment relation despite different statistics and response formats. It is domain framed because its identity requires elicited human probability assessments, an appropriate sampling model, and a discrepancy between them at the design's unit of analysis. A data table cannot instantiate the error by itself.[1][2]
Evaluative weight: “error” evaluates a judgment against a specified model; it is not a moral judgment about the person. Human-practice dependence: eliciting and interpreting a probability estimate are human cognitive activities, though the binomial or mean-variance relation is mathematical. Institutional origin: experiments and statistical teaching have studied the pattern, but an institution does not create the response discrepancy. Vocabulary travel: sample size and dispersion travel through many sciences; the psychological claim travels only where a judge and comparable task are present. Import versus recognition: the hospital and height tasks are literal instances; calling an algorithm or a county map “insensitive” without an elicited human judgment imports an analogy or names a different calibration defect.
The portable skeleton is an agent underweighting a scale-relevant input when judging an outcome, but a live Prime must carry its own full signature and evidence before it can be a parent. Its character: a model-relative, task-sensitive cognitive error with a reusable within-domain diagnostic relation, whose human judgment and sampling commitments keep the named entry domain-specific.
Structural Core vs. Domain Accent¶
The structural core is a comparison between the size effect predicted for a specified sample statistic and the smaller effect expressed in elicited human judgments, whether compared within one person or between assigned groups. The core includes an observed answer; otherwise a small county's extreme rate would count as a bias even if no one interpreted it. It includes a valid normative size contrast; otherwise mere disagreement between two answers has no directional error to diagnose.[1][3]
The domain accent is constitutive: probability or dispersion judgments by human reasoners, finite-sample statistics, and conditions under which a sampling distribution changes with \(N\). The original papers propose representativeness as an explanation, but the narrow observed error does not require the unique mental proxy to have been identified. The school and county cases carry the size-dependent statistical relation while missing the psychological judgment role.[1][2][3]
The named entry does not clear the Prime bar. An agent in another substrate may ignore a scale parameter, but that abstract resemblance alone does not establish this specific human sampling-judgment effect. A future general underweighting-of-scale Prime would need independent cross-domain admission and source tests. Any portable reach belongs to that separately established identity, not automatically to this psychological name.
Instantiates / Related Primes¶
Insensitivity to sample size has no broader abstraction in the encyclopedia yet; no existing entry has been shown to contain the full identity of this judgment error. Bias is nearby but distinct: it requires a persistent process-level signed offset, while a task-level sample-size error can be recognized from a single within-person comparison without proving that persistence. The 1972 researchers proposed representativeness as an explanation, and Representativeness Heuristic lists sample-size errors among its manifestations, but the observed responses alone do not establish its necessary unregistered typicality substitution in every case. Sampling (Representativeness) instead concerns selection design and population coverage. Clustering Illusion concerns spurious sequence structure rather than the distribution comparison here. None of these covers the full identity across the qualifying tasks.
Neighborhood in Abstraction Space¶
Insensitivity to Sample Size sits in a sparse region of the domain-specific corpus (97th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Descriptive Statistics & Correlation Measures (40 abstractions)
Nearest neighbors
- Genetic Load — 0.77
- Null distribution — 0.76
- Sampling error — 0.76
- Standardized Rate — 0.76
- Student's t-Test — 0.76
Computed from structural-signature embeddings · 2026-10-17
Not to Be Confused With¶
- Representativeness Heuristic: a proposed broader typicality-based process that can yield several errors; the present entry diagnoses a particular model-versus-\(N\)-judgment discrepancy without proving that cause in each individual.[1]
- Sampling (Representativeness): a sample-selection and target-population relation; it can succeed or fail before anyone estimates an extreme outcome's probability.
- Clustering Illusion: perceiving random clusters as meaningful structure; a person may commit it without comparing sample-size-dependent statistic distributions.
- Small-area variation: high and low rates among small denominators are expected statistical possibilities, not an observed cognitive error until a person's size-insensitive inference is measured.[3]
References¶
[1] Daniel Kahneman and Amos Tversky (1972), “Subjective probability, A judgment of representativeness,” Cognitive Psychology 3: 430–454, doi:10.1016/0010-0285(72)90016-3. Complete original article consulted, especially printed pp. 437–442, Figure 1c, and the hospital problem on p. 443. The scan is hosted by Radboud University; the publisher identifies the work. The printed title uses a colon after “probability”; the comma transcription preserves the complete title for reference binding. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v
[2] Lewend Mayiwar, Kai Hin Wan, Erik Løhre, and Gilad Feldman (2025), “Revisiting representativeness heuristic classic paradigms, Replication and extensions of nine experiments in Kahneman and Tversky (1972),” Quarterly Journal of Experimental Psychology 78: 707–730, doi:10.1177/17470218241255916. Full author manuscript consulted; PDF viewer pp. 43–47 give the Problem 6 distributions and Problems 7–9 comparisons, and pp. 57–58 discuss their mixed implications. The printed title uses a colon after “paradigms”; the comma transcription preserves the complete title for reference binding. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n
[3] Howard Wainer and Harris L. Zwerling, “Evidence That Smaller Schools Do Not Improve Student Achievement,” Phi Delta Kappan 88(4) (2006): 300–303, doi:10.1177/003172170608800411. Complete publisher-archive article consulted; pp. 300–302 show small-denominator county and school extremes. County maps reprint Gelman and Nolan's material; this article does not measure a person's probability judgment, and its school findings differ by grade. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j