Skip to content

Causal Feature Review Panel

Expert panel — instantiates Shortcut-Reliance Mitigation

Convenes domain experts to judge which of a model's influential features are causally or semantically meaningful and which are artifacts, proxies, or coincidences — and to name the intended structure it should be using instead.

Some shortcuts are invisible to any statistic because the cue is genuinely predictive in the data — it is simply the wrong reason. Causal Feature Review Panel brings substantive domain knowledge to bear on that question. A standing group of experts examines the model's most influential features and, for each, asks whether there is a mechanism by which it causes the target or whether it is merely a correlate, a proxy, a post-treatment artifact, or a coincidence of this dataset. Its defining move is that it judges what a feature should be worth, not what the model found it to be worth: it answers "ought this to matter?" where a behavioral test only answers "does the model use it?" The panel also does the two jobs only human expertise can: it states the intended structure the model is supposed to rely on, and it registers any feature that proxies a protected attribute.

Example

A lender's loan-default model is up for review, and its strongest features are applicant ZIP code, the device used to apply, and time of day. A panel of credit-risk analysts and fair-lending counsel works down the list. ZIP code predicts well — but there is no mechanism by which a neighborhood causes an individual to repay; it is a proxy for area and, through it, for protected characteristics, so it goes straight into the risk register with a redlining flag. Device type turns out to be an artifact of the acquisition channel, spuriously correlated with default through marketing, not causation. Debt-to-income ratio and prior payment history, by contrast, have a clear repayment mechanism and are affirmed. The panel's output is threefold: a written statement of the intended structure (repayment capacity and willingness), a per-feature causal verdict, and a proxy risk register naming ZIP code and its correlates for removal or monitoring. The model may be accurate; the panel's question is whether it is accurate for reasons the institution can stand behind.

How it works

  • Judge each influential feature against a mechanism, not a correlation. For every top feature the panel asks for an articulated causal or semantic story — how it would drive the target — and treats "it just predicts well" as a red flag, not an answer.
  • Name the intended structure first. The panel writes down what the model should be keying on, giving every feature a reference to be judged against rather than rationalized toward.
  • Flag protected-attribute proxies explicitly. Features correlated with a protected class, or with a known proxy for one, are logged in a register regardless of predictive value.
  • Require evidence proportional to stakes. A plausible narrative suffices for low-risk features; contested or high-stakes ones demand interventional or mechanistic support.

Tuning parameters

  • Panel composition — the mix of domain, ethical, and legal expertise. Adding ethics/legal members is essential where proxies bite, but widens scope and slows review.
  • Evidence bar — whether a plausible causal story is enough or interventional evidence is required. A higher bar resists rationalization but is slower and can stall on genuinely uncertain features.
  • Feature coverage — top-k by importance versus every feature. Deeper coverage catches low-salience proxies but costs expert time.
  • Proxy sensitivity — how tight a correlation with a protected attribute must be to trigger a register entry. Lower thresholds catch subtle proxies but flag more benign features.
  • Independence — whether the panel is separate from the builders, which trades context for freedom from motivated reasoning.

When it helps, and when it misleads

Its strength is that it imports knowledge no dataset contains: it is the only step that can declare a statistically valid feature nonetheless illegitimate — a proxy for a protected class, or a post-treatment variable that could never be known at decision time. Naming the intended structure also gives every downstream test a target to check against.

Its failure mode is that judgment can be rationalized: faced with any correlate, a clever expert can supply a plausible-sounding causal story after the fact, and blind spots in the panel's own expertise pass silently. The classic misuse is convening the panel to bless a model already built and slated to ship, so it manufactures post-hoc justifications for whatever features the model happened to use. The discipline that guards against this is to fix the intended structure and expected features before seeing the importances, to demand mechanism or intervention evidence rather than narrative for anything contested, and to keep the panel independent of the build team. The danger it addresses — a feature that predicts a protected class rather than the target — is the core of proxy discrimination.[1]

How it implements the components

Causal Feature Review Panel fills the human-judgment components — the ones that require knowledge outside the data:

  • intended_structure_specification — the panel writes down the causal drivers the model should rely on, the reference the rest of the appraisal judges against.
  • causal_or_semantic_feature_review — its core work: a per-feature verdict separating meaningful predictors from artifacts, proxies, and coincidences.
  • protected_proxy_risk_register — features that proxy a protected attribute are logged into a tracked register for removal or monitoring.

It does NOT generate the raw candidate list by adversarial hunting (Artifact Red-Team Review); it does NOT trace where a feature's data came from (Data Leakage Audit); and it does NOT behaviorally test reliance (Counter-Correlated Holdout Set, Feature Ablation or Occlusion Test).

  • Instantiates: Shortcut-Reliance Mitigation — the panel decides which of a model's reasons are legitimate and which must be removed or watched.
  • Consumes: Artifact Red-Team Review — the panel reviews the features and cues the red team flags.
  • Sibling mechanisms: Artifact Red-Team Review · Data Leakage Audit · Counter-Correlated Holdout Set · Challenge-Set Refresh Cycle · Deployment Canary and Drift Sentinel · Domain-Shift Stress Test · Feature Ablation or Occlusion Test · Group-Stratified Validation · Invariance Probe · Hard-Negative Data Augmentation · Shortcut-Risk Model Card Section

Notes

The panel's verdicts are judgments, not measurements: a feature it deems illegitimate may still be one the model barely uses, and a feature it affirms may still be exploited via an artifact it did not foresee. Pair its rulings with a behavioral reliance test before trusting either the clearance or the condemnation.

References

[1] Proxy discrimination — when a model relies on a feature that stands in for a protected attribute (e.g., ZIP code for race), reproducing a disallowed distinction without naming it. Identifying such proxies requires domain and legal knowledge that no purely statistical validity check supplies. ↩