Normalised Sensitivity Ratio Enables Post-Hoc Causal Feature Identification
July 29, 2026
The Normalised Sensitivity Ratio (NSR) provides a model-agnostic diagnostic to distinguish causal from spurious features in trained models under structured shifts. It achieves exact identification in linear structural causal models with at least three environments by measuring the coefficient of variation in per-environment sensitivity.
HOW THIS AFFECTS YOU
●
researcherYou can use this to audit pre-trained models for spurious correlations without retraining.
●
policyThis provides a mathematical framework for verifying model robustness across different deployment environments.