CAS: A Causal Attribution Score for Local and Global Explainable Artificial Intelligence
06:00 · August 15, 2026 · arXiv cs.AI RSS
Predictive explanation methods attribute a model output; they do not, by themselves, attribute an intervention effect on the real-world outcome. We introduce the Causal Attribution Score (CAS), a compact score architecture for causal explanation. CAS starts from an identified interventional coalition game, allocates the joint intervention contrast with causal Shapley contributions, and converts those raw outcome-scale effects into Local CAS, Signed Local CAS, and two complementary Global CAS summaries. The innovation is not a new Shapley formula, but a local-to-global causal reporting layer with an explicit intervention target. In the known-truth benchmark, eight repeated primary-interaction simulations (n = 2,200 each, three actions) gave mean Local CAS MAE of 0.107 for coalition-aware CAS, compared with 0.173 for one-at-a-time normalisation and 0.213 for a global normalised absolute ATE vector. The paired advantage over one-at-a-time normalisation increased from -0.003 under additivity to 0.091 under strong interactions. On both empirical DoubleML datasets, 401(k) eligibility/net financial assets (n = 9,915) and Pennsylvania reemployment bonus/unemployment duration (n = 5,099), predictive SHAP/TreeSHAP rankings differed materially from Feature-CAS rankings of treatment-effect modifiers. In Pennsylvania, dep1 (exactly one dependent) moved from predictive global rank 13 to Feature-CAS rank 2 and was the leading local Feature-CAS modifier. These results isolate the added value of separating what predicts the outcome from what explains heterogeneity in an estimated causal effect.
Summary
Predictive explanation techniques such as SHAP and LIME attribute a model’s output to input features, yet they do not directly quantify how interventions on those features alter a real-world outcome. The Causal Attribution Score (CAS) addresses this gap by anchoring explanations to an explicit interventional target rather than to a fitted predictor. It begins with an identified coalition game whose value function records the expected outcome under any combination of target and baseline intervention levels, then allocates the joint contrast across actions with classical Shapley values expressed in outcome units.
These raw contributions are subsequently normalised into comparable scores. Local CAS expresses each action’s share of the absolute causal mass at a given covariate profile, while the signed variant preserves direction and reveals cancellation or reinforcement among effects. Two complementary global summaries follow: one averages the local scores across individuals, and the other weights actions by their total absolute contribution to the population-level effect. When interactions are absent the procedure reduces to normalised conditional average treatment effects; its added value appears precisely when the effect of one intervention depends on which other interventions are active.
Empirical checks confirm the distinction. In repeated simulations with known ground truth, coalition-aware CAS recovered heterogeneous effects with lower mean absolute error than one-at-a-time normalisation or a globally scaled average-treatment-effect vector, with the advantage growing under strong interactions. On two DoubleML data sets the ranking of treatment-effect modifiers produced by Feature-CAS diverged from the ranking produced by predictive SHAP or TreeSHAP; in the Pennsylvania re-employment experiment, for example, the indicator for exactly one dependent rose from predictive rank 13 to causal rank 2 and emerged as the leading local modifier. These results illustrate how CAS isolates variables that genuinely modify an estimated causal effect from those that merely predict the outcome.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners focusing on ethical, transparent, and trustworthy AI. As the EU AI Act demands greater model explainability, CAS offers a mathematically rigorous framework to understand the actual causal impact of interventions rather than mere predictive correlations.
