Residual Correlation Diagnostic¶
Diagnostic test — instantiates Shared-Source Variance Isolation
Recomputes the correlation matrix after the shared source has been stripped out and keeps only the associations that survive the adjustment.
A Residual Correlation Diagnostic asks a single after-the-fact question: once the shared source has been stripped out — residualized, stratified, or partialled away — do the dimensions still correlate? It recomputes the correlation matrix on the residuals (or within source strata), keeps only the associations that survive, and draws the interpretive line between "true cross-dimensional signal" and "shared-source artifact." Its defining idea is survival under adjustment: it does not estimate the source (it consumes an adjustment) and it does not assume contamination levels (it uses the actual data) — it tests which correlations remain once the source is gone.
Example¶
A bank's risk team notices that four early-warning indicators for a loan book — utilization spikes, payment-timing drift, a sector-sentiment score, and a macro-stress index — all correlate tightly, seemingly confirming a broad deterioration. Three of the four are pulled from the same third-party data vendor with a shared refresh cadence. The analyst residualizes each indicator on vendor-and-refresh-date fixed effects, then recomputes correlations among the residuals. Two of the six pairwise correlations survive at meaningful strength; the rest collapse toward zero, revealing they were tracking the vendor's update rhythm, not the borrowers. The diagnostic's output is a pruned correlation matrix and a two-tier verdict: utilization-versus-payment-timing is real residual signal, while the sentiment-and-macro cluster was largely a shared feed. The risk memo is rewritten to lean only on the surviving pair.
How it works¶
Take the source-adjusted data — residuals from a fitted model, or within-stratum subsets — and recompute pairwise correlations. Compare each against its raw counterpart: associations that stay materially non-zero are retained as signal; those that vanish are flagged as source artifacts. The distinctive move is the before/after contrast on the correlation structure itself — it operates on relationships between dimensions, not on variance shares or hypothetical scenarios. It leans on partial-correlation logic[1] and hands its retained-versus-dropped verdict to whatever claim rested on the raw correlations.
Tuning parameters¶
- Adjustment source — residualize on a fitted model versus stratify and compare within levels. Model residuals use all the data; stratification is assumption-light but data-hungry.
- Survival threshold — how much residual correlation counts as "surviving." A high bar prunes aggressively (risking loss of weak-but-real links); a low bar keeps borderline artifacts.
- Multiplicity handling — how the many pairwise comparisons are corrected. Stricter control curbs false "survivors" but can bury genuine weak signal.
- Comparison lens — the absolute residual correlation versus the drop from raw to residual. The drop highlights how source-driven a pair was; the level highlights what remains.
When it helps, and when it misleads¶
Its strength is that it answers the archetype's central question — "would these still look connected without the shared source?" — on real data, and it exposes dashboard double-counting by showing which links simply evaporate. Its failure mode is unstable dimension identity: if the residualization also removes real shared substance, the surviving-or-vanishing verdict is itself an artifact of over-adjustment, and the residual dimension may no longer mean what its label says. The classic misuse is residualizing on a source so hard that every inconvenient correlation disappears, then declaring independence. The guarding discipline is to confirm the residual still carries the intended construct (an informal identity check), report raw and residual matrices side by side, and treat a vanished correlation as a prompt to investigate rather than automatic proof of leakage.
How it implements the components¶
independence_diagnostic_panel— it is the panel's core check made concrete: the residual (or source-stratified) correlation matrix that shows what remains after a source challenge.residual_claim_boundary— its retained-versus-dropped verdict draws the firewall between surviving cross-dimensional signal and shared-source artifact.
It consumes an adjustment rather than producing one, so it does not estimate or precision-weight the shared component (common_variance_adjustment_rule, precision_weight_update) — that machinery belongs to [Common Factor or Random-Effect Model]. Unlike its nearest twin [Variance Partitioning Report], it tests which pairwise correlations survive rather than attributing each dimension's variance into shares; and it plants no known-null tripwire (negative_control_dimension) — that's [Negative-Control Outcome Probe].
Related¶
- Instantiates: Shared-Source Variance Isolation — it delivers the post-adjustment survival test the interpretation turns on.
- Consumes: Common Factor or Random-Effect Model supplies the source adjustment whose residuals this diagnostic inspects.
- Sibling mechanisms: Source Variance Audit Matrix · Multitrait-Multimethod Matrix · Negative-Control Outcome Probe · Batch, Rater, or Instrument Counterbalancing Protocol · Variance Partitioning Report · Leakage Sensitivity Grid
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Residual Correlation Diagnostic operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it recomputes the correlation matrix after the shared source has been stripped out and keeps only the associations that survive the adjustment.
Independent corroboration: The frozen evidence defines Residual Correlation Diagnostic as 'Recomputes the correlation matrix after the shared source has been stripped out and keeps only the associations that survive the adjustment', 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: Partial correlation after removing shared predictors is a canonical statistical diagnostic.
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: Partial correlation after removing shared predictors is a canonical statistical diagnostic. The selected record uses alternates=none, 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.
References¶
[1] Baba, K., Shibata, R., & Sibuya, M. "Partial Correlation and Conditional Correlation as Measures of Conditional Independence". Australian & New Zealand Journal of Statistics 46(4), 657–664 (2004). Supports using partial-correlation logic to assess association after adjustment for other variables. registry ↩