Front-door adjustment identifies a treatment effect despite unmeasured treatment-outcome confounding by using a fully mediating variable whose treatment and outcome relations have appropriate observed adjustment structure.
The front-door criterion
Definition. A mediator \(M\) satisfies the front-door criterion relative to \(A,Y\) if:
- \(M\) intercepts every directed path from \(A\) to \(Y\);
- there is no unblocked back-door path from \(A\) to \(M\);
- all back-door paths from \(M\) to \(Y\) are blocked by \(A\).
The first condition excludes a direct effect not mediated by \(M\). The second identifies how \(A\) changes \(M\). The third identifies how \(M\) changes \(Y\) after adjustment for \(A\).
Front-door formula
Theorem. Under the front-door criterion, \[ p(y\mid\operatorname{do}(a)) = \int p(m\mid a) \left\{ \int p(y\mid m,a')p(a')\,\mathrm{d}a' \right\}\mathrm{d}m. \]
Proof. Decompose over the mediator: \[ p(y\mid\operatorname{do}(a)) =\int p(y\mid\operatorname{do}(a),m) p(m\mid\operatorname{do}(a))\,\mathrm{d}m. \] Condition 2 and action-observation exchange give \(p(m\mid\operatorname{do}(a))=p(m\mid a)\). Because \(M\) intercepts every directed path, intervening on \(M\) renders the action on \(A\) irrelevant for \(Y\), so the first factor can be represented through \(p(y\mid\operatorname{do}(m))\). Condition 3 permits back-door adjustment for the \(M\to Y\) effect using \(A\): \[ p(y\mid\operatorname{do}(m)) =\int p(y\mid m,a')p(a')\,\mathrm{d}a'. \] Substitute both identities.
Linear structural calculation
Example. Consider \[ M=\alpha A+\varepsilon_M,\qquad Y=\beta M+U+\varepsilon_Y, \] where an unobserved \(U\) confounds \(A\) and \(Y\), but \(\varepsilon_M\) is independent of \(A,U,\varepsilon_Y\). There is no direct \(A\to Y\) effect. Regressing \(M\) on \(A\) identifies \(\alpha\). Regressing \(Y\) on \(M\) and \(A\) identifies \(\beta\), because conditioning on \(A\) blocks the back-door path \(M\leftarrow A\leftarrow U\to Y\). The total effect is \(\alpha\beta\), even though the coefficient from regressing \(Y\) directly on \(A\) is confounded.
Positivity requirements
Definition. Front-door identification requires support for: \[ p(m\mid a)\quad\text{over mediator values induced by each }a, \] and \[ p(y\mid m,a')\quad\text{for treatment values }a' \text{ across those mediator values}. \] These are stronger and less familiar than ordinary treatment overlap.
If treatment deterministically fixes the mediator, the outcome relation at a common mediator value across treatment levels may not be observed, and the front-door functional can require extrapolation.
Observed covariates in generalized front-door adjustment
With baseline covariates \(X\), a conditional front-door formula may take \[ p(y\mid\operatorname{do}(a)) =\int p(x)\int p(m\mid a,x) \left\{\int p(y\mid m,a',x)p(a'\mid x)\,\mathrm{d}a'\right\} \mathrm{d}m\,\mathrm{d}x, \] provided the corresponding conditional graphical criteria hold. Covariates can block confounding of \(A\to M\) and \(M\to Y\), but post-treatment covariates require a more careful longitudinal graph.
Why mediator-outcome confounding matters
Counterexample. If an unmeasured \(V\) causes both \(M\) and \(Y\) independently of \(A\), then adjusting only for \(A\) does not identify the \(M\to Y\) relation. The inner integral \(\int p(y\mid m,a')p(a')\,\mathrm{d}a'\) retains mediator-outcome confounding and the front-door formula fails.
Direct effects violate interception
Counterexample. If \(Y=\delta A+\beta M+U+\varepsilon_Y\) with \(\delta\ne0\), \(M\) no longer intercepts every directed path. The front-door formula identifies only the mediated component represented by its assumptions and misses \(\delta\). No observed statistical diagnostic can prove that the direct path is absent.
Relation to mediation analysis
Front-door identification uses a mediator to identify the total
Estimation of the front-door functional
A plug-in estimator fits \(p(m\mid a,x)\) and \(\mathbb E(Y\mid m,a,x)\), then numerically evaluates the nested integrals. Semiparametric one-step and targeted estimators add influence-function corrections and can permit flexible nuisance learning under rate conditions. Because the functional combines several mechanisms, diagnostics should separately assess treatment-mediator overlap, mediator-outcome overlap, and sensitivity to violations of each front-door condition.
A credible-use checklist
Substantive use of front-door adjustment should explain why treatment has no effect outside the mediator, why there is no unmeasured \(A\)-\(M\) confounding, why adjustment for \(A\) removes all \(M\)-\(Y\) confounding, and whether mediator values overlap across treatment levels. The method is powerful precisely because these assumptions are strong.