Residualization Contribution Test¶
A per-candidate diagnostic — instantiates Independent Generator Validation
Regresses each candidate on all the others and keeps the residual, so what remains is exactly the part of that candidate the rest cannot reproduce.
Residualization Contribution Test examines one member at a time. It reconstructs candidate i as best it can from the others — fitting it on them — and then keeps the residual, the part of i the others could not explain. That residual is candidate i's unique contribution; if it shrinks to noise, i adds no independent direction. Its defining move is partialling out: independence is judged by what survives being explained away, not by a correlation coefficient or a determinant. And unlike the whole-set methods, it produces a per-candidate contribution signal you can keep and reuse, not merely a count or a flag.
Example¶
A marketing analytics team suspects their media-spend variables overlap. To test whether email spend is an independent lever, they regress the weekly email-spend series on TV, search, social, and promotions, and inspect the residual. Almost nothing is left: email had been scheduled to move in lockstep with the promotion calendar, so ≈90% of its week-to-week variation is reconstructable from the other channels (illustrative). Its unique contribution — the residual — is near noise, so email is not an independent lever in this plan and its separate coefficient cannot be trusted. Run the same test on search spend and a large residual survives: search carries variation nothing else reproduces, so it earns its place as a distinct axis. The test's verdict is per-candidate and comes with a usable by-product — the orthogonalized residual itself.[n1]
How it works¶
What distinguishes it is fit-then-subtract, applied member by member. For each candidate, regress it on the remaining set under the combination rule, compute the residual, and summarize the surviving magnitude — residual variance, or the fraction of the candidate's variation left unexplained. A residual near zero means the member is reconstructable, hence redundant; a large residual means it contributes a direction the rest cannot. The independence criterion here is fully operational: residual signal above a chosen floor counts as an independent contribution.
Tuning parameters¶
- Residual floor — how much surviving variation counts as a real contribution. A low floor keeps marginally-unique candidates; a high floor is aggressive about calling members redundant.
- Combination form — ordinary linear regression versus a richer reconstructor. A linear fit tests linear reproducibility only; a nonlinear reconstructor can expose a candidate that is redundant in a curved way a straight fit would miss.
- All-others versus subset — reconstruct each candidate from the entire remaining set, or from a chosen subset. Full-set is the strict test; a subset tests independence relative to one specific group of candidates.
- What to keep — the scalar summary (how unique the member is) versus the residual series itself, retained as an orthogonalized replacement to feed downstream.
When it helps, and when it misleads¶
Its strength is that it gives each candidate an individual, interpretable verdict and — uniquely in this set — an actual residual you can substitute in to break the collinearity, so it both localizes redundancy to the member and hands you the material to fix it.
Its limit is that it tests reproducibility only under the reconstructor you chose: a linear residualization blesses candidates that are nonlinearly redundant, and when many candidates are each near-collinear, the residuals themselves turn unstable. The classic misuse is orthogonalizing a variable and then reading its cleaned coefficient as its "true" effect, which quietly reassigns shared variance to whatever was held fixed rather than resolving the ambiguity. The discipline is to match the reconstructor to the real combination rule and to treat a small residual as evidence of redundancy to investigate, not an automatic deletion.
How it implements the components¶
Residualization Contribution Test fills the components that concern isolating and judging each member's unique contribution:
nonredundant_contribution_trace— the residual is this trace: the retained, quantified part of a candidate that the others cannot reproduce.independence_criterion— it operationalizes independence as "residual signal above a floor," applied member by member.
It does not count the set's total dimension (Rank-Revealing Decomposition) or return a whole-model collinearity score tied to coefficient variance (Variance-Inflation Review), and it produces no exact cancelling relation (Nullspace Dependency Certificate).
Related¶
- Instantiates: Independent Generator Validation — it supplies the per-candidate independence verdict, and an orthogonalized residual as a by-product.
- Sibling mechanisms: Variance-Inflation Review · Gram-Schmidt Orthogonalization Trace · Rank-Revealing Decomposition · Nullspace Dependency Certificate · Feature Collinearity Heatmap · Basis-Candidate Pruning Workflow
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Residualization Contribution Test operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it regresses each candidate on all the others and keeps the residual, so what remains is exactly the part of that candidate the rest cannot reproduce.
Independent corroboration: The frozen evidence defines Residualization Contribution Test as 'Regresses each candidate on all the others and keeps the residual, so what remains is exactly the part of that candidate the rest cannot reproduce', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Residualizing a candidate against other predictors is a canonical regression and partial-effect procedure.
Related originating lineages:
- Economics & Finance — Econometrics materially developed residualization for isolating unique explanatory contribution.
Review resolution: Both blind reviewers agree that statistics_experimental_design is the primary historical origin. Explicit reconciliation of alternate origin disagreement, domain reach disagreement adopts reviewer_a's evidence: Residualizing a candidate against other predictors is a canonical regression and partial-effect procedure. The selected record uses alternates=economics_finance, origin_mode=single_lineage, and domain_reach=multi_domain; the other review proposed alternates=data_science, mathematics, origin_mode=single_lineage, and domain_reach=specialized. The selected combination better preserves the mechanism-specific formative lineages and calibrated scope; broader present-day use is not treated as proof of additional historical origin.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Close kin, different keepsake: because it partials out, this test is one Gram-Schmidt orthogonalization step run per candidate, and it shares its exact auxiliary regression with Variance-Inflation Review, which summarizes the same fit as an inflation score. The three differ only in what each keeps — the residual signal, an orthogonal basis, or a variance-inflation number. Choosing among them is choosing what you need downstream, not running a different underlying test.
[n1] Partialling a variable out and working with its residual is the logic of the Frisch–Waugh–Lovell theorem: a coefficient in a multiple regression equals the coefficient obtained by regressing the residualized outcome on the residualized predictor. It is why a near-zero residual means a candidate carries no information the others do not already supply. ↩