Do-Calculus

Establishes when an observed variable can be ignored for an interventional query.

Mechanism: If a variable W is d-separated from the outcome Y by a set Z in a graph where incoming edges to the intervened variable X are removed, then:

P(y | do(x), z, w) = P(y | do(x), z)

• Key Insight: This rule allows the removal of irrelevant covariates, simplifying the estimand.
• Graphical Criterion: Tested using d-separation in the mutilated graph G_overline{X}.
• Practical Use: Reduces data requirements by identifying non-confounding variables that can be safely excluded from adjustment sets.

Provides the conditions under which an intervention can be replaced by conditioning on the same variable.

Mechanism: If a set Z blocks all back-door paths from X to Y after removing the outgoing edges of X, then:

P(y | do(x), do(z), w) = P(y | do(x), z, w)

• Key Insight: This is the formal justification for the back-door criterion. It converts a hard intervention into passive observation.
• Graphical Criterion: Tested in the graph G_overline{X}underline{Z}.
• Practical Use: Enables the estimation of causal effects from observational data by identifying valid covariate adjustment sets.

Determines when an intervention on a variable can be entirely removed from the expression.

Mechanism: If Z(W) denotes the set of Z-nodes that are not ancestors of W in the graph G_overline{X}, then:

P(y | do(x), do(z), w) = P(y | do(x), w)

• Key Insight: An intervention on Z is irrelevant if it only affects variables that are not ancestors of the outcome Y in the relevant subgraph.
• Graphical Criterion: Tested by deleting non-ancestor interventions in G_overline{X,overline{Z(W)}}.
• Practical Use: Simplifies complex nested counterfactual expressions by stripping out redundant or causally irrelevant interventions.

A foundational theorem proving the three rules are sufficient for identifying any identifiable causal effect.

Result: Shpitser and Pearl (2006) proved that if a causal effect is identifiable from a given causal graph and observational data, repeated application of the three rules of do-calculus will eventually transform the interventional query into a do-free expression.

• Key Insight: The calculus is not just sound; it is complete. There is no identifiable query that requires a fourth rule.
• Algorithmic Form: The ID algorithm systematically applies these rules to derive estimands automatically.
• Practical Use: Guarantees that automated causal discovery systems can, in principle, solve any solvable identification problem.

The visual and mathematical substrate upon which the rules of do-calculus operate.

Definition: A mutilated graph is a modified version of the original causal DAG where all incoming edges to an intervened variable are severed.

• Notation: G_overline{X} denotes a graph with incoming edges to X removed. G_underline{Z} denotes a graph with outgoing edges from Z removed.
• Key Insight: The rules are syntactic operations on these modified graphs. Rule 1 uses G_overline{X}, Rule 2 uses G_overline{X}underline{Z}, and Rule 3 uses G_overline{X,overline{Z(W)}}.
• Practical Use: Provides a rigorous, visual test for the applicability of each rule, moving beyond algebraic manipulation to graphical criteria.

Do-calculus bridges the gap between population-level interventions and individual-level counterfactuals.

Connection: While do-calculus directly computes interventional distributions P(y | do(x)) at the population level, these quantities form the necessary components for computing counterfactual probabilities under a Structural Causal Model (SCM).

• Three-Step Process: Counterfactual inference requires abduction (updating noise priors), action (applying do-operator), and prediction. Do-calculus solves the action step.
• Key Insight: Without do-calculus to reduce interventional expressions, individual-level counterfactual queries like 'What would have happened if I had taken the treatment?' cannot be estimated from data.
• Practical Use: Essential for algorithmic recourse and fairness systems that must answer 'what if' questions for specific individuals.

The formal mathematical framework that makes do-calculus computable. An SCM defines a system using structural equations of the form (X_i = f_i(PA_i, U_i)), where (PA_i) represents direct causes and (U_i) captures exogenous noise. Unlike a Bayesian network, an SCM supports the do-operator by encoding the mechanism for each variable independently, enabling the mutilation procedure required for interventional queries.

A Directed Acyclic Graph (DAG) that encodes qualitative causal assumptions. Nodes represent variables; directed edges represent direct causal influence. The graph's topology determines which do-calculus rules apply:

• d-separation in the mutilated graph validates conditional independence
• Back-door paths identify confounding structures
• Collider bias warns where conditioning opens non-causal paths Without a causal graph, do-calculus has no structural input to manipulate.

The three-step computational process that do-calculus ultimately enables:

1. Abduction: Update the exogenous noise distribution (P(U)) using observed evidence
2. Action: Apply the do-operator to modify structural equations
3. Prediction: Compute the outcome in the modified model This answers queries like 'What would have happened had we intervened differently?' — the gold standard for individual-level explanation and fairness auditing.

A causal definition of individual fairness that directly applies do-calculus. A predictor (\hat{Y}) is counterfactually fair if: (P(\hat{Y}{A \leftarrow a}(U) = y | X = x, A = a) = P(\hat{Y}{A \leftarrow a'}(U) = y | X = x, A = a)) In plain terms: the prediction would be the same in the actual world and a counterfactual world where only the sensitive attribute (A) were changed. This requires a full SCM to compute the nested counterfactual.

A graphical test that determines when an interventional distribution (P(Y | do(X = x))) can be estimated from observational data without full do-calculus derivation. The criterion requires:

• No descendant of X in the conditioning set
• The set blocks all back-door paths from X to Y When satisfied, (P(Y | do(X = x)) = \sum_Z P(Y | X = x, Z = z)P(Z = z)). This is the most common practical shortcut derived from Rule 2 of do-calculus.

A graphical identification strategy used when unobserved confounders block the back-door path. It requires a mediator (M) that:

• Intercepts all causal influence from X to Y
• Has no unblocked back-door paths from X to M
• Has its back-door path from M to Y blocked by X The resulting estimator chains two do-calculus derivations: (P(Y | do(X)) = \sum_M P(M | do(X))P(Y | do(M))). A classic example is estimating smoking's effect on lung cancer using tar deposits as the mediator.