Variance-Inflation Review¶
A collinearity audit — instantiates Independent Generator Validation
Audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution untrustworthy.
Variance-Inflation Review is a retrospective audit of a model that has already been fit, aimed squarely at the downstream damage of redundancy. When candidates overlap, the model can still fit the data well — but it can no longer say which candidate did the work: the individual coefficients become unstable, wide, and prone to flipping sign. This review scores each candidate by how much the others inflate the variance of its estimated effect and flags the ones whose attribution can no longer be trusted. Its defining difference from the other diagnostics is that it reads redundancy through its consequence — whether the fitted model's per-candidate attribution is still identifiable — rather than through the geometry of the candidate set alone.
Example¶
A real-estate analyst fits a hedonic price model on square footage, number of rooms, lot size, bedrooms, and bathrooms. The model's overall fit is fine, yet the coefficient on number of rooms is large, negative, and swings wildly whenever a handful of sales are added or dropped. A variance-inflation review scores each predictor by regressing it on the others: rooms comes back with a high score (≈8, illustrative) because it is almost reconstructable from square footage plus bedrooms plus bathrooms. The verdict is not that the model predicts badly — it predicts fine — but that its rooms coefficient is untrustworthy for attribution; you cannot read it as "the effect of one more room." The recommended fix (merge the collinear size measures, or drop one) is handed off, not performed here.[n1]
How it works¶
What distinguishes it is a per-candidate score of coefficient-variance inflation, computed on a fitted model. For each candidate, regress it on the remaining candidates and take the R² of that auxiliary fit; convert it to a variance-inflation score, 1/(1−R²); compare against a chosen cutoff. High-scoring candidates are exactly those whose effects the model cannot separate from the rest. The review's product is a flagged list plus a trust verdict per coefficient — not a rebuilt model and not an orthogonalized residual.
Tuning parameters¶
- Flag cutoff — the score above which a candidate is called problematic; common rules of thumb sit near 5 or 10, but there is no universal right value. A lower cutoff flags more aggressively.
- Target-of-interest weighting — whether collinearity among control variables is tolerated while collinearity involving the variable of real interest is not. High inflation on a nuisance covariate can be harmless; on the key effect it is disqualifying.
- Scope — audit every candidate, or only the ones whose coefficients are actually being interpreted. Using the model to predict versus to attribute changes what must pass.
- Grouped variables — whether sets of dummy variables are scored jointly, via a generalized inflation measure, rather than individually — which otherwise mis-flags multi-category factors as collinear by construction.
When it helps, and when it misleads¶
Its strength is that it names the specific harm that matters when a model is used for explanation: not worse prediction, but untrustworthy attribution — and it points to the exact candidates responsible. It is the sibling that speaks the language of a fitted regression rather than raw linear algebra, which is why it lands with analysts who would never read a singular-value spectrum.
Its limit is that a high score is not automatically fatal: if you only care about prediction, or if the collinear variables are not the ones you interpret, inflated variance may be irrelevant. Conversely it detects only linear redundancy among the predictors and is blind to nonlinear or incidental dependence. The classic misuse is mechanically deleting the highest-scoring variable, which can discard a substantively important predictor and merely reshuffle the collinearity rather than resolve it. The discipline is to judge scores against the model's purpose — predict versus attribute — and to treat a flag as "attribution here is unreliable, investigate," not "delete."
How it implements the components¶
Variance-Inflation Review fills the components that concern the downstream cost of redundancy and the policy for flagging it:
downstream_uniqueness_constraint— it targets exactly this: whether the fitted model can still attribute a unique effect to each candidate, which collinearity destroys even when overall fit stays fine.tolerance_and_near_dependence_policy— the flag cutoff is a near-dependence policy stated in coefficient-variance terms rather than spectral ones.
It does not keep the residual signal itself (Residualization Contribution Test), count total dimension (Rank-Revealing Decomposition), or execute the fix it recommends (Independent-Axis Design Review and the Basis-Candidate Pruning Workflow).
Related¶
- Instantiates: Independent Generator Validation — it certifies whether a fitted model's per-candidate attribution can be trusted.
- Sibling mechanisms: Residualization Contribution Test · Feature Collinearity Heatmap · Singular-Value Threshold Scan · Condition-Number Dashboard · Rank-Revealing Decomposition · Basis-Candidate Pruning Workflow
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Variance-Inflation Review operates as a bounded evaluation of existing evidence or work that produces a finding or disposition because it audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution untrustworthy.
Independent corroboration: The frozen evidence defines Variance-Inflation Review as 'Audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution untrustworthy', so its operative form is Assessment, Review & Assurance.
Nearest alternative: Analysis, Modeling & Optimization — Variance-Inflation Review includes features of an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution, but its defining operation is a bounded evaluation of existing evidence or work that produces a finding or disposition.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Universal
Rationale: Both independent reviews identify statistics experimental design as the historical home of the operation—Audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution untrustworthy.. The retained alternates document formative adjacent traditions; the reach field, not the origin field, carries later applicability.
Related originating lineages:
- Data Science & Analytics — Data science's modeling, validation, and monitoring tradition contributes a separate formative lineage to the mechanism's variance inflation review logic.
- Mathematics — Mathematical modeling, proof, and abstract-structure practice supplies a parallel or contributing lineage for the mechanism's defining operation: audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution….
Review resolution: Both blind reviewers independently place the defining operation—Audits a fitted model for collinearity by scoring how much each candidate's redundancy inflates the variance of its estimated effect, flagging the ones that make attribution untrustworthy.—in statistics experimental design. Their queued differences are secondary: alternate_origin_disagreement, origin_mode_disagreement, domain_reach_disagreement, encyclopedia_synthesis_disagreement. Reviewer A uniquely contributes no additional alternate; reviewer B uniquely contributes ['mathematics']. I preserve the full evidence-supported union of 2 alternate domain(s), without a numeric cap. origin_mode=single_lineage reflects the more specific lineage judgment in reviewer B's evidence, while domain_reach=universal separately records present-day portability. The affirmative encyclopedia-synthesis finding is preserved, and confidence=high uses the more conservative reviewer level.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Two cautions worth carrying: the score captures redundancy among the predictors only and says nothing about a predictor's relationship to the outcome, so a high-score variable can still be essential and a low-score one useless. And because it summarizes the same auxiliary regression that Residualization Contribution Test runs, the two agree on where the redundancy lives and differ only in what they return — a trust verdict versus a usable residual. Reach for this review when the question is "can I trust this coefficient," and for residualization when it is "what unique signal can I keep."
[n1] The Variance Inflation Factor for candidate j is 1/(1−R²ⱼ), where R²ⱼ comes from regressing j on the other candidates; it multiplies the variance of j's estimated coefficient relative to the no-collinearity case. Cutoffs near 5 or 10 are common rules of thumb, not hard thresholds. ↩