Quantile-Quantile Residual Check¶
Diagnostic plot — instantiates Residual-Driven Model Refinement
Plots ordered residuals against the quantiles of their assumed distribution, turning wrong tails and skew into a telltale bent line.
The Quantile-Quantile Residual Check interrogates the distribution the residuals are supposed to follow rather than their relationship to the predictors. Its defining move is a rank-for-rank comparison: it sorts the residuals and plots each against the value a perfectly well-behaved residual of the same rank would take under the model's assumed error law — usually the normal. If the assumption holds, the points fall on a straight line; every departure from that line names a specific distributional defect. A tail that bows away means the residuals are heavier- or lighter-tailed than assumed; both ends drifting the same way means skew; an S-shape means the wrong kurtosis. Where a residual-versus-fitted plot asks whether the mean function is right, this check ignores the predictors entirely and asks only whether the leftover has the shape the model claims for it.
Example¶
An insurer models the size of individual claims. The residuals have a mean near zero and the residual-versus-fitted plot looks like a tidy flat band, so the mean function seems sound. But a normal Q-Q plot of those residuals bends sharply upward at the right end: the largest observed claims are far larger than a normal error law would ever generate. The residual distribution is heavy-tailed, and the tail is exactly where the money is — a model that fits the ordinary middle will systematically under-reserve for the rare catastrophic claim.
Reading the bend as a distributional verdict, the team switches to a heavier-tailed error law (or models the log of claim size), then re-draws the Q-Q plot; the wayward tail points settle back onto the reference line. The check converted "the errors look centred" into the sharper and more consequential "the tail is wrong" — a defect no mean or variance summary would have surfaced.
How it works¶
- Rank and pair. Order the standardized residuals and pair each with the theoretical quantile of the assumed distribution at that rank, then plot the pairs.
- Read the departure by its shape. Both tails bowing outward means heavy tails; an S-curve means light tails; a consistent offset at both ends means skew; a discrete jump means a mixture or a cluster of outliers.
- Match the reference to the assumption. The plot only tests the distribution you compare against, so the reference law is fixed to what the model actually assumes — chosen before looking, never tuned afterward to make the points behave.
- Bracket the noise. Overlay pointwise simulation bands so that ordinary wobble in the extreme points is not mistaken for a real defect.
Tuning parameters¶
- Reference distribution — normal, Student-t, exponential, or the uniform of a probability-integral transform; it must match the model's assumed error law, or the plot tests the wrong thing.
- Residual standardization — raw versus standardized or studentized; unstandardized residuals from a heteroscedastic fit bend the line for a reason that has nothing to do with the distribution.
- Tail emphasis — plain quantiles versus a tail-stretching transform; how far to zoom the ends, where the consequential departures for risk and extremes actually live.
- Confidence band — whether pointwise or simultaneous envelopes are overlaid so that the always-jittery end points are not over-read.
When it helps, and when it misleads¶
Its strength is that it is the sharpest available look at tail behaviour and distributional assumptions — the very failures that govern risk, extremes, and the calibration of prediction intervals, and the very failures a residual-versus-fitted plot is blind to.[n1]
Its weaknesses are twofold. The extreme points always dance, because order statistics are noisiest in the tails, so the plot invites treating a couple of restless end points as a crisis. And it reports that the distribution is wrong without ever saying why — a heavy tail can mean a genuinely fat-tailed error process, an omitted predictor leaking into the residuals, or a handful of contaminating records. The classic misuse is swapping reference distributions until the points finally line up, thereby "passing" an assumption that was chosen after the fact. The discipline is to fix the reference to the model's real assumption before looking, judge departures against simulation bands, and hand the question of why to a root-cause review rather than reading a cause off the plot.
How it implements the components¶
residual_adequacy_criteria— it operationalizes "adequate" as distributional adequacy: whether the residuals follow the assumed law, judged by their straightness against the reference quantiles.residual_construction_rule— it depends on standardized or probability-transformed residuals so their ranks are comparable to the reference; that construction is half of what the method is.
It says nothing about structure versus the predictors — curvature and omitted terms belong to Residual-versus-Fitted Plot — nor about which subgroups carry the misfit, which is Subgroup Residual Heatmap; a formal head-to-head of two candidate fits is the Residual Comparison Test.
Related¶
- Instantiates: Residual-Driven Model Refinement — it supplies the distributional-adequacy verdict the loop needs before it trusts intervals or extremes.
- Sibling mechanisms: Residual-versus-Fitted Plot · Posterior-Predictive Residual Check · Autocorrelation and Whiteness Test · Heteroscedasticity and Scale Test · Residual Comparison Test · Subgroup Residual Heatmap · Residual Root-Cause Review
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Quantile Quantile Residual Check operates by pairs ranked residuals with theoretical quantiles and diagnoses tail weight, skew, mixtures, or outliers. That concrete deployed or enacted form is Analysis, Modeling & Optimization under the frozen taxonomy.
Nearest alternative: Assessment, Review & Assurance — Although Assessment, Review & Assurance can support this mechanism, the frozen evidence makes its operative form the act that pairs ranked residuals with theoretical quantiles and diagnoses tail weight, skew, mixtures, or outliers; the alternative is therefore secondary rather than defining.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Specialized
Rationale: Q-Q inspection of residuals against an assumed distribution is a canonical statistical model-diagnostic technique.
Related originating lineages:
- Data Science & Analytics — Production model validation materially extends its applied use.
Review resolution: Both blind reviewers agree on statistics_experimental_design as the primary origin. Explicit reconciliation resolves alternate_origin_disagreement. The merged alternate lineages retain only domains the reviewers identified as materially formative; domain_reach=specialized records later applicability separately from origin breadth.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Read this check after the residual-versus-fitted plot is clean, not before. A defect in the mean function leaks into the residual distribution and can masquerade as a distributional problem — a real curvature can bend the Q-Q line — so a non-normal Q-Q plot on top of an unresolved mean-structure defect points you at the wrong repair.
[n1] The Q-Q plot generalizes the normal probability plot: the points fall on a line only when the residuals actually follow the reference distribution being plotted against. That is why the reference must be pinned to the model's assumed error law rather than selected to make the picture pass. ↩