Probability Plot Correlation Coefficient Plot¶

A PPCC plot compares probability-plot correlations across a distribution family's shape values to locate plausible shapes and show how strongly the data favor them.

Version: v1 · 2026-10-04 · History
Domain-specific #: 13759
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Distribution Diagnostics → Experimental Design & Statistics
Aliases: Ppcc Plot

Core Idea¶

A probability plot correlation coefficient (PPCC) plot asks how straight a sample's probability plot would be under different values of one distributional shape parameter. For each candidate value, theoretical quantiles are paired with ordered observations and their correlation is computed. Graphing those correlations against the parameter exposes both a best-scoring shape and the width or flatness of the good-fit region. This is a comparison within a chosen family and plotting convention, not proof that the family generated the observations.^[1]^[2]

Structural Signature¶

Sig role-phrases: ordered sample; shape-indexed quantiles; correlation score; parameter scan; peak and profile.

Ordered observations: the empirical values supply one side of each probability plot.
Shape-indexed quantiles: a specified distribution family supplies the other side for each candidate parameter.
Linearity score: correlation compresses the quantile comparison at that parameter to a common scale.
Parameter scan: repeating the calculation creates a curve, rather than one goodness-of-fit number.
Peak and profile: the maximum proposes a shape; neighboring scores show whether that choice is sharply determined or nearly tied.

Location and scale shift or stretch the plotted quantile relation without changing Pearson correlation when the scale is positive. The shape search therefore focuses on variation beyond location and scale.^[1]^[2]

What It Is Not¶

Not a normal probability plot. That fixed-reference plot can test straightness, but the ordinary normal family has no free shape parameter to scan.^[1]
Not a universal distribution selector. A high correlation compares specified candidates; it does not exhaust all possible families or validate their causal interpretation.
Not a significance test or confidence interval. The plotted height is a diagnostic score, and its uncertainty needs a separate procedure.
Not automatically a unique estimate. Several adjacent shape values may have nearly equal scores.^[3]

Scope of Application¶

For reliability data, one can scan the shape of a Weibull distribution to see which Weibull probability plot is straightest. NIST also gives an explicitly hypothetical comparison of best Weibull and lognormal PPCC values to show how cross-family choice would work; it is not an observed reliability dataset. For a roughly symmetric sample, the Tukey-lambda family offers a shape continuum from long-tailed to short-tailed forms; NIST reports a PPCC plot for 100 generated normal values. These are unlike illustrative settings using the same repeated quantile-correlation operation, not two independently observed field studies.^[1]^[3]

Within a family, the scan helps choose a parameter for later modeling. Across several families, their best scores may be displayed together as an exploratory comparison. NIST explicitly cautions that close scores, model complexity, and scientific grounds can outweigh a marginal numerical lead.^[1]

Clarity¶

The plot separates two questions: “Does a probability plot for this shape look linear?” and “Which shape in this family looks most linear?” The first takes one quantile reference; the second repeats it and shows the answer as a function of shape. A peak near a familiar special case—for example, Tukey lambda near a normal-like value—suggests that case, but the profile and original probability plots remain relevant.^[1]

Manages Complexity¶

A collection of probability plots becomes one curve with a common ordinate. This makes a broad shape search inspectable while retaining information that a single winning parameter would discard: ties, flat regions, and poor scores throughout the range. The compression loses the detailed geometry of deviations, so a promising result should be followed by inspecting the underlying probability plots and other diagnostics.^[1]

Abstract Reasoning¶

Choose a defensible family and plotting-position rule; sort the data; compute theoretical quantiles at each shape; correlate each vector of quantiles with the ordered sample; inspect the resulting profile. If the maximum is flat, treat nearby shapes as unresolved. If the entire profile fits poorly, reconsider the family instead of tuning its shape harder. If several families fit similarly, use domain theory and parsimony as additional criteria.^[1]^[2]

Knowledge Transfer¶

The same plotted calculation moves from failure-time modeling to general exploratory statistics because the roles are sample order, shape-indexed quantiles, correlation, and a visual scan. Distribution meanings do not transfer: a Weibull failure-rate interpretation is not a Tukey-lambda tail interpretation. The live Visualization (graphics) parent captures the plot's visual encoding; a nearer probability-plot family remains a future catalog question.

Examples¶

Hypothetical Weibull-versus-lognormal reliability comparison¶

NIST asks readers to suppose a reliability dataset has best PPCC value 0.99 under a Weibull family and 0.94 under a lognormal family. For the Weibull curve, ordered failure times are paired with Weibull quantiles at a series of candidate shape values; the maximum within that curve supplies the 0.99 comparison value. The hypothetical lognormal family supplies a separate best-fit value of 0.94, so the Weibull has the stronger correlation criterion in NIST's illustration. NIST explicitly warns that a winning number alone does not settle scientific model choice. These numbers are a source-posed example, not measurements from an identified reliability dataset.^[1]

Mapped back: posited failure times = ordered observations → Weibull candidate quantiles = varying shape reference → per-shape correlation = vertical PPCC scores → 0.99 maximum = best within the hypothetical Weibull scan → 0.94 lognormal comparator = a different family, not a second point on the same curve.

Tukey-lambda tail exploration¶

NIST plots a Tukey-lambda PPCC curve for 100 generated normal random numbers and reports a maximum correlation of 0.997 at λ=0.099. Its reference table places the normal-like Tukey-lambda shape near λ=0.14; NIST accordingly describes the fitted symmetric distribution as nearly normal and not long-tailed, with a normal probability plot recommended as follow-up. The accessible handbook text does not say this particular curve has a broad near-best crest, so no such width is inferred from its reported maximum.^[1]

Mapped back: 100 generated values = ordered sample → λ changes Tukey-lambda reference quantiles → per-λ correlations form the plotted scores → 0.997 at 0.099 is the reported peak → near-normal tail interpretation is a bounded model suggestion, not proof of the generating law from PPCC alone.

Structural Tensions¶

Score versus model rationale. Maximizing PPCC supplies a reproducible linearity comparison, but an almost tied simpler or scientifically grounded model may be preferable. Choosing the numerical winner gains fidelity to this score while risking a model with poorer interpretability or weaker mechanism support; choosing the simpler or grounded model may sacrifice some fit by this score. Diagnostic: is the PPCC difference materially larger than the uncertainty and the value of the other model's justification?^[1]

Point estimate versus profile. Reporting only the maximizing shape is concise and useful for downstream fitting, but it hides whether adjacent shapes score almost as well. Inspecting and reporting the full profile preserves sensitivity information, at the cost of a less decisive summary and additional plot/model checks. Diagnostic: how fast does correlation fall away from the maximum, and would nearby choices change the intended analysis?^[1]

The score also compresses each probability plot's geometry. That is a method limit, not an additional intrinsic optimization tension: inspect the original probability plots before interpreting local deviations or declaring adequacy.^[2]

Structural–Framed Character¶

This is a mostly structural statistical method with a domain-bound interpretation. Its score and scan are explicit operations; evaluative weight enters in deciding whether the model is adequate. It depends on human modeling practice to select candidate families and plotting rules, not on an institution's decree. The vocabulary of quantiles and distributional shape does not travel literally to arbitrary problems; importing the method requires ordered observations and a meaningful parametric distribution, whereas recognizing a broad “compare scores across settings” resemblance is only analogy. Its character: structural inside distribution diagnostics, framed at the model-selection boundary.

Structural Core vs. Domain Accent¶

The skeletal relation is repeated scoring over a controlled parameter to expose an optimum and its robustness. Its visual encoding of score against parameter instantiates the live Visualization (graphics) genus. The domain-bound mechanism is theoretical-quantile linearity under a probability distribution family, calculated against ordered sample observations. Removing that statistical machinery leaves a generic parameter sweep, not a PPCC plot; the named plot therefore fails the cross-domain prime bar. A generalized score-scan skeleton is an explicit future-prime question, not an additional current parent edge established by resemblance.

This entry is a kind of Visualization (graphics).

The staged strict parent Visualization (graphics) supplies information-to-visual-mark mapping: a PPCC plot places correlation scores against scanned shape values so the peak and profile can be inspected. The live Normal Probability Plot is a related diagnostic, not a strict parent: it fixes the normal reference, whereas PPCC may scan Weibull or Tukey-lambda shape. A nearer Probability Plot genus could intervene later, but no such live entry was confirmed in this rematch.

Relationships to Other Abstractions¶

Current abstraction Probability Plot Correlation Coefficient Plot Domain-specific

Parents (1) — more general patterns this builds on

Probability Plot Correlation Coefficient Plot is a kind of Visualization (graphics) Domain-specific

A PPCC plot visually encodes correlation score against distribution-shape parameter for exploratory comparison.

Hierarchy path (1) — routes to 1 parentless root

Probability Plot Correlation Coefficient Plot → Visualization (graphics)

Neighborhood in Abstraction Space¶

Probability Plot Correlation Coefficient Plot sits in a sparse region of the domain-specific corpus (74^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Normal Probability Plot: a sample against fixed normal quantiles. Probability plot correlation coefficient: one scalar score for a specified plot. PPCC plot: a curve of those scores across shape values. Distribution fitting by maximum likelihood: a different objective and inferential procedure; its optimum need not match the straightest probability plot.^[1]^[2]

References¶

[1] NIST/SEMATECH, EDA Handbook §1.3.3.23, Probability Plot Correlation Coefficient Plot, method, examples, and model-selection caution. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m

[2] NIST/SEMATECH, EDA Handbook §1.3.3.22, Probability Plot, quantile construction and correlation interpretation. registry ↩a ↩b ↩c ↩d ↩e

[3] NIST, Dataplot graphics primer, Generate a PPCC Plot, Tukey-lambda and Weibull worked settings. registry ↩a ↩b