Skip to content

Anscombe's Quartet

Four hand-built bivariate datasets that share nearly identical low-order summary statistics yet have radically different scatter geometries, demonstrating that summaries are lossy projections and that one must visualize before granting any parametric summary authority over the data.

Core Idea

Anscombe's quartet (Francis Anscombe, 1973) is a set of four small bivariate datasets, hand-constructed so that they share nearly identical summary statistics — mean and variance of \(x\), mean and variance of \(y\), Pearson correlation, fitted linear regression slope and intercept, and residual sum of squares — while exhibiting radically different scatter geometries: a roughly linear cloud, a smoothly curved (quadratic) relationship with essentially no residual scatter, a linear cloud with a single outlier that pulls the fitted line away from the true linear trend, and a degenerate structure in which all \(x\) values are identical except one high-leverage point that alone determines the estimated slope. The datasets were constructed as a demonstration object: by exhibiting four datasets indistinguishable at the level of standard first- and second-order summaries but visually unmistakable as different, Anscombe made the claim that summary statistics are lossy projections of distributional shape vivid and unforgettable in one figure, establishing that any complete data-analysis workflow must include direct scatterplot inspection before treating parametric summaries as a faithful representation of the data's geometry.

The structural point the quartet demonstrates is that low-dimensional summaries — means, variances, covariances, and the statistics derived from them — define equivalence classes over distributions, and those classes can contain qualitatively distinct generators. Two distributions that are equivalent under a set of summary statistics are not equivalent in their higher-order structure (curvature, leverage, outlier pattern, conditional heteroscedasticity), and the higher-order structure is often exactly what determines whether a fitted model is well-specified, whether its assumptions hold, and whether its predictions are trustworthy. The quartet's pedagogical force is that this is not a theoretical observation but an empirical one: here are four distributions you cannot distinguish by looking at the numbers, and you can distinguish at a glance by looking at the plots. The practical prescription the quartet motivates — visualize before summarizing; check residual plots, not just fit statistics — has been a standard element of applied statistics instruction since Tukey's Exploratory Data Analysis (1977) and is reproduced in essentially every modern applied statistics text. The Datasaurus Dozen (Matejka and Fitzmaurice, 2017), which extends the quartet to thirteen datasets with identical mean, variance, and correlation including one shaped as a literal dinosaur, is its direct lineal descendant.

Structural Signature

Sig role-phrases:

  • the shared summary vector — the low-order statistics (x and y means and variances, Pearson correlation, fitted slope/intercept, residual sum of squares) all four datasets are built to match
  • the four constructed datasets — small bivariate point sets hand-built as the moment-equivalent pre-image of that summary vector
  • the divergent scatter geometries — the engineered reveal: a linear cloud, a smooth curve, an outlier-dragged line, and a single-leverage-point structure, visually unmistakable yet summary-identical
  • the equivalence-class fact — the structural point the exhibit burns in: low-dimensional summaries name an equivalence class of distributions, not one distribution
  • the discarded higher-order structure — curvature, leverage, outliers, heteroscedasticity that vary freely within the class and decide whether a fitted model is well-specified
  • the visualize-first prescription — the operational consequence: inspect the scatter and residuals before granting any parametric summary authority over the data
  • the no-better-summary boundary — its characteristic limit: because the moments are constructed to project the structure away, no refinement of the summaries can recover it — only looking can

What It Is Not

  • Not a found dataset or empirical discovery. The four datasets are hand-constructed to share their summaries while differing in shape; they are a designed demonstration object, not data collected from the world. The point is engineered, which is precisely why it is so clean — and why the Datasaurus extension can construct a dinosaur to the same specification.
  • Not "identical statistics" in every sense. Only the low-order summaries match — x and y means and variances, Pearson correlation, fitted slope and intercept, residual sum of squares. Higher-order structure (curvature, leverage, outlier pattern, heteroscedasticity) differs freely. The agreement is confined to the moments the construction targets, not to all statistics.
  • Not a lesson that more or better summaries fix it. Because the discarded structure is exactly what the moments are constructed to project away, no refinement of the summary statistics can recover it. "Compute better numbers" is ruled out; the only admissible remedy is to look at the scatter and residuals — visualization, not more computation.
  • Not the correlation-is-not-causation point. The quartet is about summary-loss of distributional shape within a single dataset, not about confusing association with causation. All four panels can carry the same correlation while a line is the right model for one and badly wrong for another — a claim about specification, not about causal direction.
  • Not Simpson's paradox. Simpson's paradox concerns the reversal of an association under stratification across subgroups; the quartet concerns shape-loss in the summaries of a single, unstratified dataset. Different pathology, different fix — the quartet has no subgroups whose pooling flips a sign.
  • Not Goodhart's law. Goodhart's is about a measure degrading once it becomes a target for optimization; the quartet is about a measure intrinsically discarding structural information regardless of any incentive. No one is gaming the statistics here — they are simply lossy by construction.

Scope of Application

Anscombe's quartet is a constructed demonstration object whose home is descriptive statistics, data visualization, and statistical instruction — its single role is to motivate visualization as a complement to summary statistics; its reach is bounded to that one substrate. The structural fact it exemplifies — low-dimensional summaries lose distributional structure — travels via compression / dimensionality_reduction (and model_misspecification/validity), not via the four hand-built datasets themselves, and stays out of this map.

  • Statistical pedagogy — the defining use: reproduced in essentially every applied-statistics text since Tukey's Exploratory Data Analysis (1977) to motivate visualizing before modeling.
  • Software defaults and visualization libraries — shipped as a built-in dataset (R's datasets::anscombe, seaborn's load_dataset('anscombe')) to demonstrate plotting and diagnostic functions.
  • Exploratory-data-analysis methodology — cited in EDA methodology to argue the visualize-before-model workflow, institutionalized further in diagnostic-plot culture (R's four-panel plot.lm() residuals).
  • Pedagogical descendants — the Datasaurus Dozen (Matejka & Fitzmaurice, 2017) extends the construction to thirteen datasets with identical mean, variance, and correlation, including a literal dinosaur scatter.

Clarity

The quartet's clarifying work is to convert an abstract caution — "summary statistics are lossy" — into a fact a practitioner cannot un-see, and in doing so it sharpens the distinction between a summary and the distribution that generated it. Stated as a proposition, the claim that means, variances, and correlations are lossy projections is easy to nod at and then ignore in practice; rendered as four datasets that agree on every standard first- and second-order statistic yet look nothing alike, it forces the analyst to treat a matching summary as defining an equivalence class of distributions rather than pinning down a single one. The sharp question it licenses is no longer "what are this data's summary statistics?" but "what else could share these exact summaries, and would I notice if it did?" — a question about the higher-order structure (curvature, leverage, outliers, heteroscedasticity) that the moments silently discard.

The quartet also makes legible why the discarded structure matters rather than being a mere curiosity: the very features that vary freely within a summary-equivalence class — a single high-leverage point, an unmodeled curve — are often exactly what decide whether a fitted regression is well-specified, whether its assumptions hold, and whether its predictions can be trusted. This localizes the methodological prescription with unusual precision. The fix is not "be more careful with the numbers" but "inspect the scatter and the residuals before granting any parametric summary the authority to stand in for the data," because no refinement of the summary statistics themselves can recover information they are constructed to project away. That a fit statistic can look pristine while the model is badly wrong is the confusion the quartet permanently dissolves.

Manages Complexity

The open methodological worry the quartet addresses — when can a parametric summary be trusted to stand in for the data that produced it? — is, stated generally, unbounded: a fitted line, a correlation, or a residual sum of squares could be masking any of an open-ended list of pathologies (an unmodeled curve, a single high-leverage point, an outlier dragging the slope, conditional heteroscedasticity), and an analyst trying to anticipate every way the numbers might mislead faces a sprawl with no natural stopping point. The quartet compresses that sprawl by supplying the organizing structural fact beneath it: low-dimensional summaries — means, variances, covariances, and everything derived from them — define equivalence classes over distributions, and the members of a class can be qualitatively distinct generators. Once the relationship between summaries and data is seen this way, the indefinitely many "ways a summary might mislead" collapse into one tractable object: the equivalence class of distributions consistent with the observed summaries, and the question of what varies freely within it.

What the analyst then has to track shrinks to a single dimension of vigilance and a fixed diagnostic rule. The dimension is the gap between a summary and its equivalence class; the rule is the operational consequence — inspect the scatter and the residuals before granting any parametric summary authority over the data — and it follows directly rather than needing to be re-justified for each new dataset or each new fit statistic. The branch structure is correspondingly simple: if direct inspection shows the scatter occupies the well-behaved region of the class (a genuine linear cloud), the summary may stand in for the data; if inspection reveals curvature, leverage, an outlier, or non-constant spread, the summary is known in advance to be an unreliable proxy and the model's specification, assumptions, and predictions are suspect. The quartet also fixes, once and for all, which intervention can help: because the discarded higher-order structure is precisely what the moments are constructed to project away, no refinement of the summary statistics themselves can recover it, so "compute better summaries" is ruled out and "look at the data" is the lone admissible move. A high-dimensional vigilance problem — guard against every possible way the numbers could lie — thereby reduces to one structural recognition (summaries are equivalence classes), one diagnostic act (visualize before trusting), and one branch read off that act (proxy trustworthy, or proxy suspect).

Abstract Reasoning

The structural fact the quartet burns in — that a vector of low-order summaries names an equivalence class of distributions rather than one distribution — licenses a characteristic anti-inference discipline: from "these data have slope 0.5, correlation 0.82, residual sum of squares \(S\)" the statistician refuses to conclude "the relationship is a clean line," and reasons instead FROM "I have only the summaries" TO "any member of the class consistent with these summaries — a curve, an outlier-driven fit, a single leverage point — remains live until I look." The move is to hold the generating geometry underdetermined by the numbers and to treat a matching summary as a question ("what else shares these, and would I notice?") rather than an answer. This is a predictive claim about one's own blindness: the analyst predicts that curvature, leverage, outliers, and heteroscedasticity can vary freely while every reported statistic stays fixed, and so predicts that a pristine-looking fit statistic is fully compatible with a badly misspecified model.

The diagnostic move is the scatter-and-residual read that resolves the underdetermination. Having granted that the summary fixes only the class, the statistician inspects the plot and classifies which member is in hand: a genuine linear cloud (the well-behaved region, where the summary may stand in for the data); a smooth curve (the linear summary is the wrong model and the slope is an artifact of forcing a line through a bend); a single far outlier dragging the fitted line off the bulk trend (the slope reflects the outlier, not the relationship); or a lone high-leverage point that alone determines the slope (the estimate rests on one observation). Each scatter geometry is read as a distinct diagnosis with distinct consequences for whether the model's assumptions hold and its predictions can be trusted — so the statistician reasons from the shape to the trustworthiness of the parametric summary, an inference the moments cannot deliver.

The boundary-drawing move fixes which remedy is even admissible. Because the discarded higher-order structure is exactly what the moments are constructed to project away, the statistician predicts in advance that no refinement of the summary statistics can recover it — "compute better summaries" is ruled out at a glance — leaving "look at the data, before granting the summary authority" as the lone available intervention, and fixing its order: visualization precedes, rather than supplements, the parametric summary. The same recognition draws a scope line around the lesson: it is a claim about shape-loss in a single dataset's summaries, so the statistician keeps it distinct from association-reversal under stratification (a different pathology with a different fix) and predicts that the construction generalizes — any agreed summary vector can be matched by visibly different scatters, which is precisely why the Datasaurus extension can hold mean, variance, and correlation fixed while drawing a dinosaur.

Knowledge Transfer

Anscombe's quartet is a constructed demonstration object rather than a causal mechanism, so within its home it transfers as an exhibit-plus-discipline rather than as machinery, and that transfer is genuine and wide inside statistics. The lesson — summaries name an equivalence class of distributions; visualize before granting a parametric summary authority — and the diagnostic discipline (scatter-and-residual read; the boundary ruling out "compute better summaries") propagate across statistical pedagogy (reproduced in essentially every applied-statistics text since Tukey's EDA), software defaults (shipped as datasets::anscombe and seaborn's load_dataset('anscombe'); the four-panel plot.lm() residual diagnostics and the broader diagnostic-plot culture institutionalize the prescription), and exploratory-data-analysis methodology, with the Datasaurus Dozen as its direct lineal descendant holding mean, variance, and correlation fixed while drawing a dinosaur. But this is all one substrate — descriptive statistics, data visualization, and statistical instruction — and the exhibit's single role is to motivate visualization as a complement to summary statistics; it is not machinery deployed on different substrates.

Beyond statistics the honest reading is case (B): what genuinely generalizes is not the quartet but the structural fact it exemplifies — low-dimensional summaries lose distributional structure, and any compressed representation can collapse qualitatively distinct generators onto the same summary — and that pattern is carried by existing primes, not by Anscombe's construction. The lossy-summary insight is compression; the projection-loses-structure insight is dimensionality_reduction; the consequence that a pristine fit statistic can coexist with a badly wrong model belongs to model_misspecification / validity in the research-reliability family. Those abstractions travel to any domain where a summary stands in for richer underlying variation; the quartet's specific cargo — the four hand-built bivariate datasets, the moment-equivalence construction, the visualize-the-scatter prescription — is a worked example bound to descriptive statistics. Treating the exhibit itself as portable would only multiply the catalog with named demonstrations (Datasaurus, McCabe's quartet) whose structural force is identical. So the honest cross-domain lesson should carry compression / dimensionality_reduction (and the misspecification/validity patterns), with the quartet retained as the concrete exhibit that argues for them, and kept distinct from look-alikes it is often confused with — Goodhart's law (a measure degrading once it becomes a target) and Simpson's paradox (association-reversal under stratification) are different pathologies with different fixes. See Structural Core vs. Domain Accent.

Examples

Canonical

Anscombe's original 1973 construction is the exhibit itself: four datasets of eleven (x, y) points each, built so that all four share, to two decimal places, the mean of x (9.0), the variance of x (11.0), the mean of y (about 7.50), the variance of y (about 4.12), the Pearson correlation (about 0.816), and the least-squares line y = 3.00 + 0.50x with the same residual sum of squares. Run the numbers on any of the four and they are interchangeable. Plot them and they are nothing alike: dataset I is a genuine noisy linear cloud; dataset II is a smooth concave curve for which a straight line is simply the wrong model; dataset III is a tight straight line marred by one outlier that tilts the fitted slope; dataset IV has every point at x = 8 except a single point at x = 19 that alone fixes the slope. The identical regression line means something completely different in each panel — trustworthy in one, an artifact in the other three.

Mapped back: The six agreed statistics are the shared summary vector; the four eleven-point sets are the four constructed datasets built as its moment-equivalent pre-image. The linear-cloud, curve, outlier, and single-leverage panels are the divergent scatter geometries, and their coexistence under one summary is the equivalence-class fact. Curvature, the tilting outlier, and the lone leverage point are the discarded higher-order structure the moments project away.

Applied / In Practice

The quartet's discipline is institutionalized in everyday statistical practice through residual-diagnostic tooling and its modern successors. Statistical software makes the "look before you trust the fit" prescription a default: fitting a linear model in R and calling plot() on it returns a standard four-panel set of residual diagnostics (residuals-versus-fitted, Q–Q, scale-location, residuals-versus-leverage) precisely so that curvature, heteroscedasticity, and high-leverage points cannot hide behind a clean R². Autodesk researchers Justin Matejka and George Fitzmaurice extended Anscombe's idea with the "Datasaurus Dozen" (2017), generating thirteen datasets — including one shaped like a dinosaur — that share the same mean, variance, and correlation to several decimals while looking wildly different, and used them to argue for animation and visualization in exploratory analysis. Both are the quartet's lesson turned into working practice.

Mapped back: The four-panel residual plots operationalize the visualize-first prescription and the no-better-summary boundary — no refinement of the fit statistics substitutes for looking. The Datasaurus reproduces the shared summary vector / divergent scatter geometries structure at higher resolution, confirming the equivalence-class fact: any agreed summary can be matched by radically different scatters, up to a dinosaur.

Structural Tensions

T1: Compression indispensable versus compression lossy (summaries you cannot trust yet cannot do without). The quartet's lesson is that low-order summaries are lossy projections that name an equivalence class rather than a distribution — a caution against trusting the numbers. But summaries are not optional: comparison, hypothesis testing, communication, and any analysis at scale require reducing data to a few statistics, and a raw scatter of a million overplotted points is as unreadable as a bare mean is uninformative. The quartet warns against the very compression that makes statistics possible, without offering a way to avoid it. The tension is that summarization is simultaneously the discipline's core tool and its core hazard, so "don't trust the summary" cannot mean "don't summarize" — it must mean something subtler about when the class is benign, which the exhibit dramatizes but does not resolve. Vigilance against lossiness cannot become refusal to compress. Diagnostic: Is the concern here that this summary is masking structure that matters, or a general distrust of compression that would leave no workable summary at all?

T2: Adversarial construction versus natural representativeness (a hand-built worst case). The quartet's unforgettable force comes from being hand-constructed as the moment-equivalent pre-image of a target summary — four datasets engineered to be maximally deceptive, culminating in a dinosaur drawn to spec. That very engineering is a double edge. Real data rarely conspires to match six statistics while concealing a smooth curve or a lone leverage point; the pathologies exist, but nature does not usually assemble them into the adversarial coincidences the quartet exhibits. So the demonstration proves that summaries can mislead with certainty while saying nothing about how often they do, and its constructed extremity can inflate the perceived frequency of the danger. The tension is that the exhibit's clarity and its representativeness pull apart: the more vividly it is built to deceive, the less it resembles the data an analyst typically meets, so the lesson's rhetorical power exceeds its base-rate warrant. Diagnostic: Is the risk in this dataset a plausible natural pathology, or is the quartet's hand-built extremity being read as evidence that ordinary summaries routinely hide such structure?

T3: Visualization as remedy versus visualization as another lossy projection (the fix that does not scale). The quartet rules out "compute better summaries" and installs "look at the scatter" as the lone admissible remedy, ordered before trusting any parametric summary. But a plot is not a neutral window onto the data: it is itself a projection whose axes, aspect ratio, binning, and point size can hide or manufacture structure, and it fails outright beyond the bivariate, small-n regime the quartet inhabits — overplotting swamps large n, and a scatter cannot show a high-dimensional relationship. So the prescribed cure replaces one fallible low-dimensional representation with another, not with ground truth, and does not generalize past the toy setting that makes it work. The tension is that "visualize before summarizing" presents looking as the escape from lossiness while looking is itself lossy and non-scalable, so the remedy inherits the very limitation it was meant to cure once the data leaves two dimensions. Diagnostic: Does visualization here actually reveal the discarded structure, or is the plot its own lossy projection — and does the data's dimensionality or size make looking infeasible anyway?

T4: Blanket vigilance versus usually-adequate summaries (a discipline that can be disproportionate). The equivalence-class fact motivates a standing rule: inspect scatter and residuals before granting any summary authority. This is cheap and prudent in the small-n bivariate case. But the quartet establishes only that the equivalence class is nonempty and diverse — that summaries can mislead — not that they usually do, and in fact one of its own four panels (the genuine linear cloud) is a case where the summary faithfully represents the data. Turned into universal practice, "always visualize before trusting" can impose vigilance out of proportion to the actual risk, treating every clean fit statistic as a probable deception. The tension is that the demonstration's certainty about possibility is easily over-read as a claim about prevalence, so the discipline it motivates ranges from essential (when structure lurks) to ritual overhead (when the summary was fine all along), with the exhibit itself not telling which case one is in. Diagnostic: Does the visualize-first discipline here catch a real masked pathology, or is it ritual vigilance against a class member that, for this dataset, does not differ from the well-behaved case?

T5: Autonomy versus reduction (Anscombe's quartet or the compression/dimensionality-reduction it exemplifies). Anscombe's quartet is a named demonstration object with concrete cargo — four hand-built bivariate datasets, the moment-equivalence construction, the visualize-the-scatter prescription. Within statistics it transfers as an exhibit-plus-discipline, reproduced in texts and shipped in software. But the structural fact it burns in — low-dimensional summaries lose distributional structure, and any compressed representation can collapse distinct generators — is carried by existing primes: the lossy-summary insight is compression, the projection-loses-structure insight is dimensionality_reduction, and the pristine-fit-yet-wrong-model consequence is model_misspecification/validity. Those travel to any domain where a summary stands in for richer variation; treating the exhibit itself as portable would only multiply named demonstrations (Datasaurus, McCabe's quartet) of identical force, and it must be kept distinct from Simpson's paradox and Goodhart's law. The tension is between a beloved concrete exhibit and the flatter compression/dimensionality-reduction facts that are what actually generalize. Diagnostic: Resolve toward compression/dimensionality_reduction (and misspecification/validity) when the lesson is about summaries losing structure in general; toward Anscombe's quartet when arguing visualize-before-modeling in descriptive statistics in situ.

Structural–Framed Character

Anscombe's quartet occupies an unusual place — mixed on the structural–framed spectrum — because the entry is a domain-bound pedagogical artifact that exemplifies a substrate-independent structural fact, and the two pull opposite ways. The criteria split accordingly. Evaluative_weight points structural: the quartet convicts nothing and prescribes only a methodological discipline ("visualize before granting a summary authority") — advice, not a verdict, carrying no normative charge about conduct or worth. Human_practice_bound is mixed: the four datasets, once specified, are just numbers whose statistics and scatter geometries hold observer-free, so the fact they embody is not practice-bound — yet the quartet's identity as a demonstration object is bound to the human practice of statistical instruction; it exists to be shown, and its whole point is a lesson for an analyst. Institutional_origin points framed for the named object: the quartet is hand-constructed — Anscombe engineered it in 1973 as the moment-equivalent pre-image of a target summary vector, and the Datasaurus extension builds a dinosaur to the same spec — so as an artifact it is a designed pedagogical construction, not a fact of nature; but it is a construction that exhibits a natural mathematical truth, so the institutional origin attaches to the exhibit, not to what it exhibits. Vocab_travels points framed for the named object: the four bivariate datasets, the moment-equivalence construction, and the visualize-the-scatter prescription are pinned to descriptive-statistics substrate, while the fact beneath them travels under other names. Import_vs_recognize is bimodal: within statistics the exhibit-plus-discipline transfers as recognition (pedagogy, software defaults, EDA methodology, the Datasaurus descendant), but beyond statistics only the structural fact travels — carrying the exhibit itself would merely multiply named demonstrations of identical force.

The portable structural skeleton is a single fact — a low-dimensional summary names an equivalence class of distributions, so a compressed representation can collapse qualitatively distinct generators onto the same summary — carried by compression and dimensionality_reduction, with the pristine-fit-yet-wrong-model consequence belonging to model_misspecification / validity. That skeleton is genuinely and deeply structural — a substrate-independent mathematical truth that travels to any domain where a summary stands in for richer variation — which is what gives the quartet its structural pull and keeps it from being a merely framed teaching gimmick. But it does not make the quartet itself a prime, because that fact is exactly what the quartet exemplifies and instantiates from compression/dimensionality_reduction, not what the four hand-built datasets uniquely carry: the cross-domain reach belongs to those parents, while the entry's distinctive cargo — the specific constructed datasets, the moment-equivalence engineering, the scatter-and-residual prescription, and the Datasaurus lineage — is the descriptive-statistics furniture that stays home. Its character: an evaluatively neutral, hand-constructed pedagogical exhibit of a genuinely substrate-independent compression fact — structural in the equivalence-class-of-a-lossy-summary skeleton it embodies and hands off to compression/dimensionality_reduction, but bound to descriptive statistics as the particular worked demonstration that argues for them.

Structural Core vs. Domain Accent

This section decides why Anscombe's quartet is a domain-specific abstraction and not a prime — an unusually clean case, because the entry is a concrete artifact whose whole reason for existing is to make one portable fact vivid, so the split between what lifts and what stays home is sharp.

What is skeletal (could lift toward a cross-domain prime). Strip the datasets away and a substrate-independent mathematical fact survives: a low-dimensional summary names an equivalence class of the objects it summarizes, so a compressed representation can collapse qualitatively distinct generators onto the same summary, and the discarded higher-order structure — not recoverable by refining the summary — is often exactly what decides whether a downstream model is well-specified. The abstract pieces travel: a many-to-one projection from rich objects to a short descriptor vector, an equivalence class of pre-images that vary freely in the projected-away dimensions, and the corollary that a pristine-looking summary is fully compatible with a badly wrong model of the underlying object. This is genuinely portable, which is exactly why it is already carried by compression (the summary is a lossy encoding), dimensionality_reduction (a projection discards structure), and model_misspecification / validity (a clean fit statistic coexisting with a wrong model). It is the core the quartet shares — the reason the exhibit is compelling at all — not what makes the quartet distinctive.

What is domain-bound. Everything that makes the entry Anscombe's quartet in particular is descriptive-statistics furniture, and none of it survives extraction. The four hand-built eleven-point bivariate datasets; the specific moment-equivalence construction (matching x/y means and variances, Pearson correlation, least-squares slope/intercept, and residual sum of squares to two decimals); the catalogue of pathologies keyed to regression diagnostics (the tilting outlier, the lone high-leverage point, the unmodeled quadratic, heteroscedasticity); the visualize-the-scatter-and-residuals prescription and its tooling (R's plot.lm() four-panel diagnostics, datasets::anscombe, seaborn's loader); and the Datasaurus lineage are all bound to the bivariate small-n regression substrate. The decisive test: remove that substrate and the quartet does not travel intact — it becomes a looser thing, the bare compression fact, which any number of other exhibits (McCabe's quartet, a dinosaur) instantiate equally well. Carrying "Anscombe's quartet" across domains would only multiply named demonstrations of identical structural force.

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 quartet's transfer is bimodal, and its bimodality is starker than most because it is an exhibit rather than a mechanism. Within statistics it transfers as recognition of the same discipline — reproduced in essentially every applied-statistics text, shipped as software defaults, institutionalized in EDA methodology and residual-diagnostic culture, and directly reproduced by the Datasaurus — because every use lives on the same descriptive-statistics substrate. Beyond statistics the quartet itself does not transfer at all; what generalizes is only the structural fact it exemplifies, and that fact is already carried, in more general form, by compression, dimensionality_reduction, and model_misspecification / validity. So the cross-domain reach belongs entirely to those parents; "Anscombe's quartet," as named, is the concrete worked demonstration bound to its home domain — invaluable as the exhibit that argues for the visualize-before-modeling discipline in statistics, but carrying descriptive-statistics baggage (the constructed datasets, the moment-equivalence engineering, the scatter-and-residual prescription) that should stay home. It clears the domain-specific bar decisively for statistics and data visualization, and its only substrate-spanning content is the compression fact its parents already own.

Relationships to Other Abstractions

Local relationship map for Anscombe's QuartetParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Anscombe's QuartetDOMAINPrime abstraction: Dimensionality Reduction — is a decomposition ofDimensionalityReductionPRIME

Current abstraction Anscombe's Quartet Domain-specific

Parents (1) — more general patterns this builds on

  • Anscombe's Quartet is a decomposition of Dimensionality Reduction Prime

    Anscombe's Quartet is a constructed demonstration of dimensionality reduction collapsing visibly different data geometries onto the same low-dimensional summary vector.

Hierarchy paths (4) — routes to 3 parentless roots

Not to Be Confused With

  • The Datasaurus Dozen. The quartet's direct lineal descendant — thirteen datasets sharing mean, variance, and correlation while drawing shapes up to a literal dinosaur. It is not a different concept but the same moment-equivalence construction at higher resolution and of identical structural force; multiplying such named exhibits is exactly what the entry warns adds nothing. Tell: both demonstrate the equivalence-class fact — if you are citing a specific constructed dataset-set to argue visualize-before-modeling, you are invoking the exhibit (quartet or Datasaurus), not a distinct structural claim.

  • Simpson's paradox. Concerns the reversal of an association under stratification across subgroups — a sign that flips when data are pooled or split. The quartet concerns shape-loss in the summaries of a single, unstratified dataset with no subgroups at all. Different pathology, different fix (disaggregate versus visualize). Tell: is there a grouping variable whose pooling flips the direction of an effect (Simpson's), or one flat dataset whose summary hides its geometry (the quartet)?

  • Ecological fallacy. Inferring individual-level relationships from group-level aggregates — a cross-level inference error, cousin to Simpson's paradox. The quartet makes no level-crossing claim; it is about summaries discarding shape within one level of data. Tell: does the mistake move from group statistics down to individuals (ecological fallacy), or from a summary to the distribution it compressed (the quartet)?

  • Goodhart's law. A measure degrades once it becomes a target for optimization — a failure driven by incentive and gaming. The quartet's summaries are lossy intrinsically, by construction, with no one optimizing against them. Tell: is the measure failing because someone is gaming it (Goodhart) or because it discards structural information regardless of any incentive (the quartet)?

  • Correlation is not causation. A caution about inferring causal direction from association. The quartet is silent on causation: all four panels can carry the same correlation while a line is the right model for one and badly misspecified for another — a claim about distributional shape and specification, not about cause. Tell: is the worry that an association might not be causal (correlation-causation), or that a summary statistic misrepresents the data's geometry (the quartet)?

  • The compression / dimensionality-reduction parent (with model_misspecification / validity). The substrate-independent fact the quartet exemplifies — a low-dimensional summary names an equivalence class of the objects it summarizes, so a compressed representation can collapse distinct generators, and a pristine fit statistic can coexist with a wrong model. That fact travels via compression, dimensionality_reduction, and model_misspecification / validity; the quartet is the concrete statistics exhibit that argues for them, treated more fully in Knowledge Transfer and Structural Core vs. Domain Accent. Tell: when the lesson is about summaries losing structure in general, carry the compression parent; "Anscombe's quartet" is the descriptive-statistics demonstration, not the portable fact itself.

Neighborhood in Abstraction Space

Anscombe's Quartet sits in a sparse region of the domain-specific corpus (71st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Paradoxes & Distributional Structure (11 abstractions)

Nearest neighbors

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