Practical do-Shapley Explanations for Estimand-Agnostic Causal Inference

Explaining model predictions through a causal lens is increasingly important for responsible AI. When we combine do-Shapley explanations with an estimand-agnostic stance, we get a practical framework that highlights how interventions on features causally drive outcomes—without being tethered to a single, rigid estimand. This article lays out the concepts, workflow, and practical considerations you can apply in real-world settings.

Foundations: do-Shapley and Estimand-Agnostic Causal Inference

Do-Shapley blends two ideas. First, Shapley values from cooperative game theory assign a fair credit to each feature for a prediction, by averaging its marginal contributions across all possible feature orderings. Second, the do-operator from causal inference formalizes interventional distributions by setting a subset of variables to fixed values and examining the downstream effect on the outcome. Put together, do-Shapley explains a model’s prediction in terms of the causal impact of intervening features, rather than merely relying on associations observed in data.

• Shapley values provide a fair, additive decomposition of a prediction into feature attributions, accounting for all possible interactions among features.
• do-calculus defines interventional distributions P(Y | do(X = x)) by severing the causal influence of parent variables on X.
• Estimand-agnostic emphasizes explaining the causal effect via interventions without committing to a single endpoint (like ATE or ITE) upfront; the explanations reflect outcomes under plausible interventions regardless of the chosen estimand.

Why this matters for practitioners

In practice, teams may pivot between estimands or confront limited counterfactual data. A do-Shapley approach centers explanations on controllable interventions, yielding insights that remain meaningful as the estimand landscape shifts. It also helps distinguish causal influence from spurious associations, making explanations more policy-relevant and robust across modeling choices.

“The goal is to illuminate a model’s reasoning in terms of actionable interventions, not just statistical associations.”

A practical workflow you can follow

1. Specify your causal assumptions. Draft a simple DAG or narrative about which features can be intervened upon and how they influence the outcome. Note potential confounders and instruments, and align with the data-generating process.
2. Choose the scope of features. For feasibility, select a focused set of features to explain. Consider interactions and whether some variables should be held fixed as baselines.
3. Define the intervention strategy. Decide on the values to intervene on (e.g., policy-relevant levels or representative percentiles). Clarify whether non-intervened features will be drawn from their observational or a DAG-consistent distribution.
4. Compute interventional outcomes. For each subset S of features, apply do(X_S = x_S) and sample the remaining features from a plausible distribution. Evaluate the model’s predicted outcome under each intervention scenario.
5. Estimate Shapley contributions. Use permutation-based sampling to approximate the Shapley values with respect to the do-interventions. This yields local (instance-level) contributions and, by aggregation, global insights.
6. Assess identifiability and robustness. Examine how sensitive the attributions are to your causal assumptions and to alternative priors for the non-intervened features. Consider running explanations under multiple plausible DAGs.

Interpreting the results

Look for features with consistently high do-Shapley values across sampled instances. Those are your main levers for policy or decision-making. Remember that interactions matter: a feature’s contribution may depend on the values of others, and the Shapley framework captures these dependencies through the ordering of interventions.

Tip: When the causal graph is uncertain, compute do-Shapley values under several plausible DAGs and compare the explanations. Consistent top contributors across graphs boost confidence in their causal role.

Limitations and practical tips

• Reliable explanations require credible causal graphs and identifiability assumptions. Misspecification can distort attributions.
• Computational cost grows with the number of features. Use a focused feature set or approximate with a sensible sampling quota.
• Global explanations are aggregates of local explanations; report both perspectives and explore heterogeneity across subpopulations.

Ultimately, practical do-Shapley explanations translate causal interventions into intuitive feature credits, helping teams connect model insights with actionable policies while staying adaptable to different estimands. This approach keeps explanations relevant, transparent, and aligned with real-world decision-making.