Do-Calculus

Judea Pearl's complete axiomatic system consists of three transformation rules that govern how the do-operator interacts with probabilistic relationships:

• Rule 1 (Insertion/Deletion of Observations): Allows adding or removing a variable from the conditioning set when the intervention and outcome are independent given that set
• Rule 2 (Action/Observation Exchange): Permits replacing an intervention with a conditioning observation when the causal effect is identifiable from the observational distribution
• Rule 3 (Insertion/Deletion of Actions): Enables removing or adding an intervention when it has no causal effect on the outcome variable

These rules are complete—any identifiable causal effect can be derived through their sequential application.

Do-calculus provides a sound and complete algorithmic procedure for determining whether a causal effect can be estimated from observational data given a causal graph:

• The backdoor criterion identifies covariate sets that block spurious paths between treatment and outcome
• The front-door criterion handles cases where confounders are unobserved by using an intermediate mediator
• The do-calculus derivation systematically applies the three rules to transform P(y|do(x)) into an expression containing only observed variables

If no sequence of rule applications yields an expression free of do-operators, the effect is non-identifiable from the available data.

The fundamental distinction that do-calculus formalizes is between seeing and doing:

• P(y|x): The conditional probability of observing outcome y given that variable X was observed to take value x—a passive observation
• P(y|do(x)): The probability of outcome y given that X was actively set to value x through external intervention, severing all incoming causal arrows to X

This distinction is critical in supply chains: observing that expedited shipping correlates with on-time delivery does not imply that forcing expedited shipping will cause on-time delivery if both are driven by a common cause like order priority.

When a causal effect is identifiable, do-calculus yields the truncated factorization or g-formula:

• The joint distribution under intervention do(X=x) is obtained by removing the factor P(X|parents(X)) from the factorization of the joint distribution
• The remaining conditional probabilities are multiplied and marginalized over all variables except the outcome
• This produces an expression computable entirely from observational data

For supply chain disruption analysis, this allows estimating the effect of forcing a supplier switch without actually executing it, using historical operational data and the causal graph structure.

Do-calculus possesses two critical theoretical properties that distinguish it from ad-hoc causal adjustment methods:

• Completeness: If a causal effect is identifiable from a given causal graph and available data, there exists a finite sequence of do-calculus rule applications that derives the identifying formula
• Decidability: The question of whether a causal effect is identifiable can be answered algorithmically in polynomial time for Markovian models (no unobserved confounders)

For models with latent variables, identifiability becomes more complex but remains decidable through extensions like the ID algorithm, which operationalizes do-calculus for practical computation.

In autonomous supply chain intelligence, do-calculus enables counterfactual disruption analysis without experimentation:

• Root cause identification: Given an observed delivery delay, do-calculus can isolate whether the cause was a specific supplier failure versus a systemic logistics bottleneck
• Intervention planning: Estimate the causal effect of adding buffer inventory at a specific node without physically implementing the change
• What-if simulation: Compute P(on_time_delivery | do(supplier_switch=yes)) using only historical observational data and the causal graph

This transforms supply chain control towers from correlational dashboards into causal decision engines capable of prescribing optimal interventions.