Causal Attribution Score Improves AI Explainability for Intervention Effects

Traditional predictive explanation methods for AI models attribute an output but often fail to explain the causal effect of an intervention on real-world outcomes. This research presents the Causal Attribution Score (CAS), a compact score architecture designed specifically for causal explanation. CAS operates by starting with an identified interventional coalition game, then allocates the joint intervention contrast using causal Shapley contributions, converting these effects into Local CAS, Signed Local CAS, and two complementary Global CAS summaries. The core innovation lies in its local-to-global causal reporting layer, which explicitly targets intervention effects rather than just predictions. Benchmarking against known truths demonstrated that coalition-aware CAS achieved a mean Local CAS MAE of 0.107, outperforming other normalization methods. Furthermore, on empirical datasets like 401(k) eligibility and Pennsylvania reemployment, predictive SHAP/TreeSHAP rankings of feature importance differed significantly from Feature-CAS rankings of treatment-effect modifiers. This highlights CAS's unique ability to isolate the added value of distinguishing between what predicts an outcome and what explains heterogeneity in an estimated causal effect.