Spurious Correlations
A spurious correlation is a statistical relationship that helps prediction in one dataset but does not reflect a dependable causal or policy-acceptable relationship. Models are excellent at finding these patterns because they optimize predictive performance, not meaning.
Spurious correlations are dangerous when they become invisible decision rules. A hiring model might learn that a resume format, school, commute distance, or employment gap predicts success because of past selection patterns rather than future ability.
Audit a Hiring Signal
Each of these features improved prediction accuracy in historical data. Run each one through four checks before trusting it in production.
stability checkfail
Predictive power drops sharply on resumes from a new template rollout.
ablation checkfail
Removing this feature barely changes accuracy — it wasn't carrying real signal.
Proxy checkfail
Template choice correlates with which school's career center advised the applicant.
Domain reviewfail
No recruiter would defend format as a skill indicator.
Transparency helps by showing which features matter, how stable they are across time and groups, what happens when they are removed, and whether domain experts can defend their use. A signal should not be trusted simply because it is predictive in historical data.
Use ablation tests, temporal validation, subgroup validation, counterfactual checks, and domain review to distinguish useful predictors from brittle or inappropriate proxies.
Predictive Does Not Mean Permissible
A feature can improve accuracy and still be unacceptable because it proxies a protected trait, reflects past discrimination, or cannot be acted on fairly.
What is the best next step when a highly predictive feature has no plausible relationship to the outcome?