Does a coefficient sign change prove fraud?

No. Legitimate conditioning changes, confounding, and sampling variation can also alter signs.

Does the result justify using outcome-derived predictors?

No. Such construction can make apparent prediction circular.

Is this entry a full theorem proof?

No. It provides the core warning and an independently checked algebraic illustration.

AI Glossary / Statistics and probability / Advanced

Pyrrho's lemma

Definition

Pyrrho's lemma is a result about the vulnerability of regression conclusions to a deliberately selected additional regressor. With sufficiently unrestricted construction of that regressor, existing coefficients and other reported results can be manipulated extensively. Theo Dijkstra's account uses this result to motivate skepticism about conclusions produced through extensive specification searching without evidence of generalization. It is a warning about the freedom of model construction, not a claim that every added variable produces any desired result.

What it means

Regression coefficients describe relationships conditional on the variables included in a model. Adding another variable changes that conditioning and can alter existing coefficients. Pyrrho's lemma concerns the especially strong possibilities available when the added variable can be chosen from a sufficiently flexible set.

Its practical message is that impressive results on the data used to construct a model are not self-validating. The allowed construction and mathematical conditions matter. General slogans about obtaining any result should not be interpreted as removing ordinary least-squares constraints or as saying that a legitimate, predetermined explanatory variable is automatically suspicious.

How it works

An intuitive route begins with observed outcomes y and an existing design matrix X. Suppose an analyst wants a particular coefficient vector b. Define a new column z = y - Xb. Then y = Xb + z by construction. If the enlarged design has full column rank, an exact least-squares fit can reproduce the chosen coefficients alongside coefficient one on z.

This elementary construction reveals the danger: z contains the outcome that the model purports to explain or predict. It need not represent an independently meaningful quantity available before a future outcome occurs. The example is a special illustration, not a reproduction of every condition or extension of Dijkstra's theorem.

Real audits therefore examine feature provenance, the timing of measurement, the extent of specification search, and evaluation on appropriately separated data. A feature designed after seeing outcomes requires particular scrutiny. Testing must evaluate the whole selection process, rather than only the final fitted equation as though it had been fixed in advance.

Three practical examples

A regression review investigates why a coefficient changes sign after an unusual control variable is introduced.
An AI audit checks whether engineered features indirectly encode the target being predicted.
A research team documents specification searches instead of presenting the selected result as the only model considered.

Think of it this way

Imagine building a scoring rule and then adding a custom adjustment chosen after you already know everyone's final score. The adjusted rule can fit the existing scores beautifully, but that success may come from information built into the adjustment. The analogy captures why construction freedom matters. It does not imply that all adjustments are invalid or that a useful predictive model must explain outcomes perfectly.

Worked example

In an original construction, take x = (-2, -1, 0, 1, 2) and y = (-3, -4, 2, 0, 5). Both have mean zero. Regressing y on an intercept and x gives intercept zero and slope 2. The residuals are (1, -2, 2, -2, 1), with squared sum 14. The total squared deviation of y is 54, giving R-squared 1 - 14/54, approximately 0.741.

Now define z = y + 3x, obtaining (-9, -7, 2, 3, 11). Regress y on an intercept, x, and z. The exact identity y = -3x + z produces coefficients zero, -3, and 1. The enlarged design has full column rank, and its residual sum of squares is zero. The coefficient on x has changed from positive 2 to negative 3, while training R-squared becomes one.

This does not discover a negative causal effect. The new predictor was constructed using y. Its perfect fit is circular, and conventional nondegenerate t-statistic interpretations are not justified by this zero-residual illustration. The calculation demonstrates coefficient manipulation, not the theorem's complete claims about inferential quantities or a deployable forecasting method.

Illustrative case study

Fictional case study: Alder Metrics receives an impressive regression explaining support costs. A reviewer discovers that an engineered control combines the recorded cost outcome with an existing workload variable.

The team removes that circular feature from its prospective prediction pipeline and documents how the original specification was selected. It evaluates the revised process on later records whose outcomes were unavailable during development. This fictional case illustrates provenance checking; it does not establish that every sign change signals manipulation or that one validation exercise guarantees future success.

Common misconceptions and mistakes

The lemma does not mean any randomly chosen extra variable will force an arbitrary result. The ability to construct or select a suitable regressor is essential. Nor does it prove that regression is useless or that coefficient changes necessarily indicate misconduct.

Ordinary constraints still apply. For nested ordinary least-squares models on the same observations and outcome, adding regressors cannot increase the minimized residual sum of squares. With the same total sum of squares, training R-squared therefore cannot decrease. Adjusted R-squared and out-of-sample performance behave differently. An unrestricted claim that every statistic can independently take any numerical value would be misleading.

Business use

For business users, the lesson is to inspect how an explanation was obtained before trusting its apparent precision. A model selected after extensive experimentation may reflect flexible choices as well as stable relationships.

Maintain feature definitions, creation dates, search records, and evaluation boundaries. For prediction, verify that every feature is available at prediction time. For causal questions, use a defensible design and substantive assumptions; predictive validation alone does not establish causality. Transparency about selection makes a model easier to challenge, reproduce, and evaluate in the setting where it will be used.

Check your understanding

Can the constructed predictor z = y + 3x be used to predict an unknown y when only x is available?

Answer: Not as defined. Computing z requires the unknown outcome, so the exact fit does not supply a usable forecasting rule.

Related terms

Learn next

Study least-squares geometry and changes in conditional coefficients. Then examine post-selection inference, feature leakage, and evaluation designs that preserve independence from model development.

Least squares

Sources

Dijkstra's primary abstract, reproduced on his publication profile, supports the regression-manipulation warning: https://www.researchgate.net/profile/Theo-Dijkstra . Original paper: https://doi.org/10.1007/BF01894292 . Full-paper access remains unavailable. This article does not assert its complete theorem conditions or inferential extensions. The explicitly limited, full-rank algebraic construction is independently verified; the example and fictional case are original.

Previous: Pyramid vector quantization Next: Q-function