Frontdoor Criterion

This assumption requires that there are no unobserved confounders affecting both the mediator (M) and the outcome (Y). Graphically, this means the path from the treatment (X) to the outcome (Y) via the mediator (M) is not corrupted by confounding. Any such confounding would introduce a backdoor path between M and Y, violating the criterion. This is distinct from the backdoor criterion, as it allows for unmeasured confounding between X and Y (the backdoor path), but not between M and Y (the frontdoor path).

The treatment variable (X) must have a direct causal effect on the mediator (M), and there must be no other direct or indirect paths from X to Y that bypass M. In a causal graph, this means all directed paths from X to Y must pass through M; M acts as a complete mediator. If X affects Y through any other channel (e.g., a direct arrow X→Y or via another unblocked path), the frontdoor adjustment formula will yield a biased estimate of the total effect.

There must be no unmeasured common causes of the treatment (X) and the mediator (M). This is often trivially satisfied in experimental settings where X is randomized, but in observational data, it requires that all variables influencing both X and M are measured and controlled for. If this assumption is violated, the estimated effect of X on M is biased, which propagates error through the entire frontdoor calculation.

When the three core assumptions hold, the causal effect of X on Y is identifiable and can be computed using the frontdoor adjustment formula:

P(Y | do(X)) = Σ_m P(M=m | X) * Σ_x' P(Y | X=x', M=m) * P(X=x')

This formula works by:

• First estimating the effect of X on M (P(M | X)).
• Then estimating the effect of M on Y, while adjusting for X to block any backdoor paths (Σ_x' P(Y | X=x', M=m) * P(X=x')).
• Finally, combining these estimates by summing over the mediator's values. It effectively uses M as a pseudo-instrument to isolate X's effect.

The assumptions can be verified using d-separation on the causal graph. For a set of variables (X, M, Y) to satisfy the frontdoor criterion:

1. All directed paths from X to Y must go through M.
2. There is no unblocked backdoor path from X to M.
3. There is no unblocked backdoor path from M to Y, after conditioning on X. If these d-separation conditions hold in the graph, and the causal faithfulness assumption is met, then the statistical assumptions of the frontdoor criterion are satisfied.

A classic illustrative example involves Smoking (X), Tar deposits in lungs (M), and Lung Cancer (Y). An unmeasured genetic factor (U) may confound X and Y (smokers might have a genetic predisposition to cancer). The frontdoor criterion can be applied if:

• Smoking causes Tar deposits (X→M).
• Tar deposits cause Cancer (M→Y).
• There is no unmeasured confounder between Tar and Cancer (e.g., no separate environmental factor causing both tar and cancer).
• The genetic factor (U) does not affect Tar accumulation. Even with the confounding from U, the effect of Smoking on Cancer can be estimated by measuring the effect of Smoking on Tar, and Tar on Cancer (while adjusting for Smoking).

The backdoor criterion is the primary graphical method for identifying a causal effect. It states that a set of variables Z satisfies the backdoor criterion relative to (X, Y) if:

• Z blocks every backdoor path between the treatment X and the outcome Y.
• No node in Z is a descendant of X. Conditioning on Z then allows estimation of the causal effect P(Y | do(X)) via adjustment: P(Y | do(X)) = Σ_z P(Y | X, Z=z) P(Z=z). It is the standard approach when you can measure all common causes (confounders).

Do-calculus is a complete set of three inference rules developed by Judea Pearl for manipulating expressions containing the do-operator. It provides the formal mathematical machinery to determine if and how a causal effect can be identified from observational data and a causal graph.

• Rule 1: Insertion/deletion of observations.
• Rule 2: Action/observation exchange.
• Rule 3: Insertion/deletion of actions. The Frontdoor Criterion is proven using these rules. They transform P(Y | do(X)) into an estimable expression involving only observed probabilities when a valid frontdoor variable exists.

An instrumental variable (IV) is an alternative identification strategy used when unmeasured confounding exists and the backdoor criterion fails. An IV, Z, must satisfy three core conditions:

1. Relevance: Z is correlated with the treatment X.
2. Exclusion Restriction: Z affects the outcome Y only through X (no direct path).
3. Exchangeability: Z is independent of unmeasured confounders of X and Y. While both IV and Frontdoor address unmeasured confounding, they leverage different graphical structures. IV uses an exogenous variable affecting the treatment, while Frontdoor uses a mediator fully intercepting the treatment's effect.

Causal mediation analysis quantifies the extent to which a treatment's effect operates through an intermediate variable (mediator). It decomposes the total effect into:

• Natural Direct Effect (NDE): The effect of X on Y not through the mediator M.
• Natural Indirect Effect (NIE): The effect of X on Y that operates via M. The Frontdoor Criterion is a special, strong case of mediation where the mediator M is assumed to be the only pathway from X to Y. This strong assumption (no direct effect) is what enables identification despite unmeasured confounding between X and Y.

A Structural Causal Model (SCM) is the foundational framework for formal causal reasoning. It consists of:

• A set of endogenous variables (observed).
• A set of exogenous variables (background noise).
• A set of structural equations defining each endogenous variable as a function of its direct causes and noise.
• A corresponding causal graph (DAG). The Frontdoor Criterion is applied within an SCM. The assumptions of no direct effect (X→Y) and no unblocked backdoor path (X←U→Y) are encoded in the SCM's structural equations and graph, providing the rigorous basis for the identification formula.

Causal identifiability is the core property that a causal query (e.g., P(Y | do(X)) ) can be uniquely computed from the available data (observational or experimental) under a given set of assumptions. It answers the question: Can we learn this causal effect from what we can observe? The Frontdoor Criterion provides a sufficient condition for identifiability when standard backdoor adjustment is impossible. It demonstrates that even with unmeasured confounding, a causal effect can be identified if we can find a variable satisfying the specific frontdoor conditions, turning an otherwise non-identifiable problem into a solvable one.