Backdoor Criterion

In a causal graph (a directed acyclic graph or DAG), a set of variables Z satisfies the backdoor criterion relative to an ordered pair of variables (X, Y) if:

• No node in Z is a descendant of X.
• Z blocks every path between X and Y that contains an arrow into X (a 'backdoor path').

Conditioning on a set Z that meets this criterion d-separates X and Y along all non-causal, confounding paths, isolating the direct causal effect.

A backdoor path is any non-causal path between treatment X and outcome Y that remains open if we do not condition on any variables. These paths create spurious associations through confounders.

• Example: X ← Z → Y is a classic backdoor path via confounder Z.
• Blocking: Conditioning on Z (e.g., stratifying analysis by Z's values) blocks this path, removing the confounding bias.
• The criterion systematically identifies all such paths that must be blocked.

A key rule is that a valid adjustment set Z must not contain descendants of the treatment X. Conditioning on a descendant of X can:

• Introduce bias by opening new non-causal paths (e.g., through colliders).
• Block part of the causal effect if the descendant is on the causal pathway from X to Y (a mediator).

Common Pitfall: Adjusting for a variable affected by the treatment (like a post-treatment measure) often violates this rule and leads to biased effect estimates.

The backdoor criterion provides a practical method for moving from observation to intervention. If a set Z satisfies the criterion, then the causal effect of X on Y, denoted P(Y | do(X)), is identifiable and can be computed from observational data using the adjustment formula:

P(Y | do(X=x)) = Σ_z P(Y | X=x, Z=z) P(Z=z)

This formula stratifies or weights by the values of Z, effectively simulating a randomized experiment where Z is held constant.

The frontdoor criterion is an alternative identification strategy used when no set Z meets the backdoor criterion due to unmeasured confounding.

• Backdoor: Adjusts for confounders (common causes). Requires measuring all relevant confounders.
• Frontdoor: Uses a mediator variable M that is fully intercepted by X and affects Y only through M. It does not require measuring the confounder.
• Use Case: Backdoor is the first and most intuitive check. Frontdoor is applied when a valid backdoor adjustment set is not available in the data.

Applying the backdoor criterion involves:

1. Drawing a Causal Graph: Specifying assumed relationships based on domain knowledge.
2. Listing All Paths: Identifying all paths between treatment X and outcome Y.
3. Finding an Adjustment Set: Selecting measured variables Z that block all backdoor paths without including descendants of X.
4. Performing Adjusted Analysis: Using regression, matching, or weighting based on Z.

This process formalizes the common advice to 'control for confounders' and provides a rigorous test for whether an analysis is likely to produce an unbiased causal estimate.

The frontdoor criterion is an alternative graphical identification strategy used when a backdoor path is blocked by unmeasured confounding. It requires a mediator variable (M) that:

• Is causally influenced by the treatment (X → M).
• Influences the outcome (M → Y).
• Is not influenced by the unmeasured confounders affecting X and Y.
• Has no direct path from X to Y that bypasses M.

If these conditions hold, the causal effect can be identified by combining the effect of X on M and M on Y, even in the presence of unmeasured confounders between X and Y.

Do-calculus is a complete set of three formal inference rules developed by Judea Pearl for transforming expressions containing the do-operator. It provides a systematic, algebraic method to determine if a causal effect is identifiable from observational data and a causal graph.

The rules allow one to:

• Add or remove observations from a probability expression.
• Interchange interventions and observations under specific conditions.
• Add or remove interventions from an expression.

If repeated application of these rules can eliminate the do( ) operator, the causal query can be answered using observational probabilities. The backdoor and frontdoor criteria are specific, commonly used instances derivable from do-calculus.

Causal identifiability is the property that a causal quantity (like the Average Treatment Effect) can be uniquely computed from the available data—typically observational—given a set of assumptions encoded in a causal model.

A query is identifiable if, in principle, with infinite data, we could calculate its true value. The backdoor criterion is a sufficient condition for identifiability. If no valid backdoor adjustment set exists, the effect may be non-identifiable due to issues like unmeasured confounding, requiring alternative strategies (e.g., instrumental variables, frontdoor adjustment) or stronger assumptions.

D-separation (directional separation) is the fundamental graphical criterion for determining conditional independence in a Directed Acyclic Graph (DAG). It is the mechanism by which the causal Markov condition implies probabilistic independencies.

A path between two variables is blocked by a set of conditioned variables Z if the path contains:

• A chain (A → C → B) or fork (A ← C → B) where the middle node C is in Z.
• A collider (A → C ← B) where the middle node C is not in Z, and no descendant of C is in Z.

If all paths between X and Y are blocked by Z, then X and Y are d-separated by Z, implying conditional independence. The backdoor criterion relies on d-separation to ensure all spurious paths are blocked.

Confounding is the central problem the backdoor criterion solves. A confounder is a variable that:

• Causally influences both the treatment (X) and the outcome (Y).
• Creates a non-causal association (a backdoor path) between X and Y.

Unmeasured confounding occurs when a confounder is not observed, violating the backdoor criterion's requirement that all confounders be measured and conditioned upon. This often makes causal effects non-identifiable from observational data alone. A valid backdoor adjustment set is a set of measured variables that, when conditioned on, blocks all backdoor paths, thereby adjusting for or controlling for confounding.

Once a valid backdoor adjustment set Z is identified, the causal effect is estimated using an adjustment formula. The most common is the backdoor adjustment formula:

P(Y | do(X=x)) = Σ_z P(Y | X=x, Z=z) P(Z=z)

This formula stratifies the population by values of Z, computes the association between X and Y within each stratum, and then averages over the distribution of Z. This mathematically removes the confounding influence of Z.

This formula underpins common estimation methods:

• Stratification/Matching on Z.
• Inverse Probability Weighting (IPW) using propensity scores.
• G-computation via outcome modeling.