Do-Calculus

Insertion and deletion of observations when the intervention and outcome are conditionally independent given a set of covariates.

• Formal expression: P(y|do(x), z, w) = P(y|do(x), w) if Y ⊥⊥ Z | X, W in the graph where all arrows pointing to X are removed (G_X̅).
• Biomedical intuition: If a biomarker Z provides no additional information about the outcome Y once you already know the treatment X and covariates W, you can safely ignore Z.
• Example: In a drug efficacy study, if a patient's hair color is independent of their recovery given the drug and their age, hair color can be removed from the causal estimand without bias.

Transforming an intervention into a passive observation when the causal effect of the intervention is fully mediated through observed variables.

• Formal expression: P(y|do(x), do(z), w) = P(y|do(x), z, w) if Y ⊥⊥ Z | X, W in the graph where arrows pointing to X are removed and arrows pointing from Z are removed (G_X̅Z̲).
• Biomedical intuition: If the effect of a secondary intervention do(z) on the outcome Y is entirely captured by observing Z and conditioning on X and W, you can replace the experimental manipulation with observational data.
• Example: In a Mendelian randomization study, if a genetic variant's effect on disease is fully mediated through a measured protein level, the do-operator on the variant can be exchanged for conditioning on the protein.

Adding or removing an intervention entirely when the intervention has no causal pathway to the outcome.

• Formal expression: P(y|do(x), do(z), w) = P(y|do(x), w) if Y ⊥⊥ Z | X, W in the graph where arrows pointing to X are removed and arrows pointing to any Z nodes that are not ancestors of W are removed (G_X̅Z(W)̅).
• Biomedical intuition: If manipulating Z has no effect on Y once X and W are accounted for, the do(z) operation can be eliminated from the expression entirely.
• Example: In a multi-drug interaction study, if a co-administered drug Z has no causal path to the primary outcome Y after accounting for the main treatment X and baseline covariates W, the intervention on Z can be deleted from the causal query.

The three rules are provably complete — any causal effect that can be identified from observational data and a causal DAG can be derived by repeated application of these rules.

• Identifiability criterion: A causal effect P(y|do(x)) is identifiable if and only if there exists a finite sequence of do-calculus rule applications that eliminates all do-operators from the expression.
• Algorithmic implementation: The ID algorithm systematically applies these rules to determine identifiability and derive the estimand.
• Biomedical significance: In drug-target Mendelian randomization, do-calculus proves whether a causal effect of a protein target on disease can be estimated from GWAS summary statistics without conducting a randomized trial.
• Limitation: Do-calculus assumes the causal DAG is correctly specified; unmeasured confounding not represented in the graph cannot be overcome.

Special cases of do-calculus that provide intuitive graphical conditions for causal effect identification without full algebraic derivation.

• Back-door criterion: A set of variables Z satisfies the back-door criterion relative to (X, Y) if Z blocks all back-door paths from X to Y and no node in Z is a descendant of X. The causal effect is then P(y|do(x)) = Σ_z P(y|x, z)P(z).
• Front-door criterion: Applies when unobserved confounders block the back-door path but a measured mediator M fully mediates the effect. The causal effect is P(y|do(x)) = Σ_m P(m|x) Σ_x' P(y|x', m)P(x').
• Biomedical example: In biomarker discovery, if a genetic variant (X) affects disease (Y) only through a measured protein (M), the front-door adjustment recovers the causal effect even with unmeasured confounding between the variant and disease.

Practical applications of do-calculus for deriving causal estimands from real-world biomedical data.

• Mendelian randomization: Do-calculus formally justifies using genetic instruments as proxies for do(x) operations, proving when SNP-outcome associations identify causal effects.
• Transportability: Extended do-calculus determines when causal effects estimated in one population (e.g., a clinical trial) can be transported to a different target population (e.g., real-world clinical practice).
• Selection bias correction: Do-calculus can derive expressions that recover causal effects from data subject to selection bias, such as hospital-based cohort studies where patients are selected based on disease severity.
• Software implementation: Libraries such as DoWhy and causaleffect in R implement do-calculus for automated identifiability checking and estimand derivation.

The fundamental mathematical notation in do-calculus, written as do(X=x), representing a surgical intervention that forces a variable to take a specific value. Unlike conditioning on X (observing X=x), the do-operator removes all incoming edges to X in the causal graph, simulating a randomized experiment. This distinction between seeing and doing is the cornerstone of Pearl's causal hierarchy, moving from associational (Layer 1) to interventional (Layer 2) reasoning.

A complete set of inference rules for transforming interventional probabilities into observational ones:

• Rule 1 (Insertion/Deletion of Observations): Allows adding or removing a variable from the conditioning set when it is independent of the outcome given other variables and the intervention
• Rule 2 (Action/Observation Exchange): Permits replacing an intervention with observation when the variable and outcome share the same set of confounding parents
• Rule 3 (Insertion/Deletion of Actions): Enables removing an intervention entirely when the variable has no causal path to the outcome

These rules are proven complete—any identifiable causal effect can be derived using them.

A graphical condition derived from do-calculus for identifying causal effects by blocking spurious associations. A set of variables Z satisfies the back-door criterion relative to (X, Y) if:

• No node in Z is a descendant of X
• Z blocks every path between X and Y that contains an arrow pointing into X (back-door paths)

When satisfied, the causal effect is identifiable as: P(Y|do(X)) = Σ_z P(Y|X, Z=z)P(Z=z). This is the theoretical foundation for covariate adjustment in observational studies.

An alternative identification strategy when unobserved confounders block the back-door path. A set of variables M satisfies the front-door criterion if:

• M intercepts all directed paths from X to Y
• There is no unblocked back-door path from X to M
• All back-door paths from M to Y are blocked by X

This enables identification through a two-step adjustment: first estimating X's effect on the mediator M, then M's effect on Y. The classic example is the smoking-tar-cancer model where tar mediates the causal pathway.

A causal effect P(Y|do(X)) is identifiable if it can be uniquely computed from the observational distribution and causal graph structure alone. Do-calculus provides the algorithmic machinery to determine identifiability through symbolic derivation. Key results include:

• The do-calculus is complete: if an effect is identifiable, the three rules can prove it
• The ID algorithm automates identifiability checking for any given causal DAG
• Some effects remain non-identifiable even with full observational data, requiring additional assumptions or experimental data

A three-level framework organizing causal reasoning tasks by increasing difficulty:

• Layer 1 (Association): P(Y|X) — seeing and observing correlations. Requires only joint distributions
• Layer 2 (Intervention): P(Y|do(X)) — doing and predicting the effects of actions. Requires causal DAGs and do-calculus
• Layer 3 (Counterfactuals): P(Y_x|X=x', Y=y') — imagining what would have happened under different conditions. Requires functional causal models

Do-calculus operates at Layer 2, bridging observational data to interventional predictions without requiring full structural equations.