Longitudinal Front-Door Criterion

• Longitudinal Front-Door Criterion is a dynamic causal identification framework that extends Pearl’s front-door method to time-ordered settings with latent confounding and unknown lags.
• It utilizes summary causal graphs and repeated-measures designs to intercept causal paths even when standard back-door adjustments fail.
• The methodology enables nonparametric efficient estimation with multiply robust approaches, ensuring valid inference in complex longitudinal data structures.

The longitudinal front-door criterion denotes a family of extensions of Pearl’s front-door identification strategy to temporally ordered settings in which standard back-door adjustment fails because of unmeasured exposure–outcome confounding, yet observed mediator structure still permits identification. In the recent literature, the term covers at least two technically distinct constructions. One extends the front-door criterion from fully specified causal DAGs to a dynamic setting in which only a summary causal graph is available, while allowing latent confounding and even cycles at the macro level, with target micro total effect $P(Y_t=y_t \mid do(X_{t-\gamma}=x_{t-\gamma}))$ (Assaad, 2024). Another treats repeated exposure and mediator measurements over time, with observed data $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$ and target parameter $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$, and develops nonparametric efficient estimators of the resulting longitudinal front-door functional (Breum et al., 23 Sep 2025). Both formulations generalize the classical point-treatment front-door theorem, whose standard identification formula and do-calculus proof are given in (Javidian et al., 2018).

1. Classical front-door basis

Pearl’s original front-door theorem concerns a semi-Markovian causal Bayesian network with an unobserved confounder $U$ between $X$ and $Y$, and an observed mediator set $Z$ on the causal path from $X$ to $Y$. The front-door criterion requires three graphical conditions: $Z$ blocks all directed paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$0 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$1; there are no open back-door paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$2 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$3; and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$4 blocks all back-door paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$5 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$6. Under these conditions, and if $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$7, the causal effect is identifiable through

$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$8

The notation $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$9 denotes the intervention $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$0 (Javidian et al., 2018).

The classical proof proceeds by identifying $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$1, then $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$2, and finally $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$3 using do-calculus. Conceptually, the argument shows why front-door identification is possible even when $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$4 and $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$5 are confounded by an unobserved $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$6: the mediator breaks the problem into an $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$7 component that is observationally recoverable because $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$8 is not confounded with $\Psi^{\bar a_T}(P)=\mathbb{E}\{Y(\bar a_T)\}$9, and a $U$0 component that is observationally recoverable conditional on $U$1. The 2018 proof does not introduce a separate longitudinal theorem, but it provides the foundational case that later temporal and repeated-measures constructions generalize (Javidian et al., 2018).

2. Temporal extension via summary causal graphs

A prominent longitudinal extension is formulated for dynamic structural causal models with variables indexed by time. In this setting, the true causal structure lives at the micro level as a full-time acyclic directed mixed graph over variables such as $U$2, $U$3, and $U$4. The model allows temporal causation with unknown lag, latent confounding represented by bidirected edges, stationarity, bounded lag through a maximum lag $U$5, and an assumption preventing a latent process from directly causing itself at another time point (Assaad, 2024).

Because the true full-time acyclic directed mixed graph is unknown, the analysis is conducted on a summary causal graph $U$6. Each macro vertex represents an entire time series rather than a single time point. A directed edge $U$7 means that for some lag $U$8, there is an edge $U$9 in the underlying micro graph, and a bidirected edge $X$0 summarizes latent confounding at the summary level. A central difficulty is that summary causal graphs may contain cycles: once time ordering is collapsed, the macro graph can display $X$1 and $X$2 even though the micro-level graph is acyclic (Assaad, 2024).

The target of inference is a micro total effect,

$X$3

with $X$4 in the paper’s setup. This formulation is explicitly temporal: the intervention is applied to a time-indexed variable $X$5, and the response is evaluated at time $X$6. The extension is therefore not merely a rephrasing of the static theorem, but a response to partial temporal knowledge, latent confounding, and macro-level feedback patterns that arise when lag structure is unknown (Assaad, 2024).

3. The front-door criterion for summary causal graphs

The summary-graph extension defines a front-door criterion for summary causal graphs. A set of macro vertices $X$7 satisfies the criterion relative to $X$8 if four conditions hold: $X$9 intercepts all activated directed paths from $Y$0 to $Y$1 in the summary causal graph; there is no activated back-door path from $Y$2 to $Y$3; all back-door paths from $Y$4 to $Y$5 are $Y$6-blocked by $Y$7; and either $Y$8, or $Y$9 (Assaad, 2024).

Conditions 1–3 are the analogues of Pearl’s original front-door conditions, while the cycle/lag restriction is new. The paper emphasizes that cycles require extra care because a summary graph can obscure whether an apparent parent of $Z$0 is really a predecessor, a descendant in feedback, or both. If $Z$1 is in a directed cycle and $Z$2, the adjustment set required for the mediator process can include descendants of $Z$3, which breaks standard front-door reasoning. When $Z$4, a separate lemma still allows identification even if cycles exist by using a different adjustment set built from ancestors of $Z$5 (Assaad, 2024).

The key bridge from macro to micro level is that if $Z$6 intercepts all directed paths from $Z$7 to $Z$8 in the summary graph, then the corresponding micro set

$Z$9

intercepts all directed paths from $X$0 to $X$1 in any compatible full-time acyclic directed mixed graph. The identification theorem is then obtained through two sequential adjustment steps: first identifying $X$2 using a back-door adjustment set, and then identifying $X$3 similarly. The resulting estimand is a do-free front-door-type formula involving the time-indexed mediator copies $X$4, the lagged sets $X$5, $X$6, and $X$7, and sums over mediator and treatment histories. The theorem is explicitly stated as sound, not complete: if the criterion holds, identifiability is guaranteed, but failure of the criterion does not imply non-identifiability (Assaad, 2024).

4. Repeated-measures longitudinal front-door functional

A second major formulation treats the front-door problem in a longitudinal setting with repeated exposure and mediator measurements. The observed data are

$X$8

where $X$9 denotes baseline covariates, $Y$0 is treatment at time $Y$1, $Y$2 is the time-varying mediator or covariate at time $Y$3, and $Y$4 is the final outcome. The target parameter is the mean potential outcome under an exposure history $Y$5,

$Y$6

This generalizes the point-treatment front-door target $Y$7 to a dynamic exposure regime (Breum et al., 23 Sep 2025).

Identification relies on a specific unmeasured-confounding structure involving an unmeasured variable $Y$8. The assumptions are: consistency; for each $Y$9, no unmeasured confounding between treatment and mediator given past,

$Z$0

outcome independent of treatment history given $Z$1, mediators, and baseline,

$Z$2

and treatment exchangeability given $Z$3 and observed past,

$Z$4

The paper interprets these assumptions as saying that all causal effects of $Z$5 on $Z$6 are mediated through future mediators $Z$7, that $Z$8 may confound treatment and outcome, and that $Z$9 does not directly affect the mediators (Breum et al., 23 Sep 2025).

Under these assumptions and two positivity conditions, the paper identifies $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$00 by the longitudinal front-door functional, called the F-functional: $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$01 Here $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$02, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$03, and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$04 is the conditional density of $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$05. The paper also gives inverse-probability weighted and mediator-density-ratio representations of the same functional, including a Bayes-theorem rewriting that avoids direct estimation of potentially high-dimensional mediator densities (Breum et al., 23 Sep 2025).

5. Efficient estimation and large-sample theory

The repeated-measures formulation is accompanied by nonparametric efficient estimation theory. The efficient influence function for $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$06 is written as

$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$07

with an equivalent representation using $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$08 in place of $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$09. On this basis, the paper proposes a one-step estimator that plugs initial nuisance estimates into the efficient influence function and averages, and a targeted maximum likelihood estimator built by targeted fluctuation along least favorable submodels (Breum et al., 23 Sep 2025).

The one-step estimator requires nuisance estimators for $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$10, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$11, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$12, and the sequential regressions $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$13 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$14. The TMLE starts with initial estimates of $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$15, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$16, and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$17, then updates $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$18 via weighted logistic regression with clever covariate $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$19, recursively updates $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$20 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$21, reconstructs the nested regression objects $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$22, and finally estimates $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$23 by plugging the updated regression into the initial-layer formula. For the binary mediator special case, the supplement gives an alternative TMLE that directly updates the mediator density $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$24 (Breum et al., 23 Sep 2025).

A central property is multiply robust consistency. For the one-step estimator, consistency holds if any of three sets of nuisance limits are correct: $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$25 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$26 are correct for all $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$27; $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$28 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$29 are correct for all $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$30; or the sequential regression limits satisfy the relevant nested regression targets, together with $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$31 correct. Under an appropriate Donsker condition for the efficient influence function class, positivity and boundedness, and nuisance estimators converging faster than $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$32, the estimator is asymptotically linear: $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$33 The paper explicitly allows machine learning for nuisance estimation, including Super Learner, and notes that the Donsker requirement can be relaxed using sample splitting or cross-fitting (Breum et al., 23 Sep 2025).

The simulation study uses $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$34 with binary $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$35, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$36, $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$37, and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$38, nonlinear dependence, and unmeasured confounding between exposure and outcome through $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$39. Approximate target values are $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$40 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$41. When nuisance models are approximately correct, all estimators behave well; sequential regression estimators show noticeable bias under regression misspecification; the one-step estimator and TMLE remain robust in the multiply robust scenarios predicted by theory; and TMLE tends to have slightly smaller empirical standard deviation than the one-step estimator. Wald intervals based on the efficient influence function have good coverage when the required conditions are satisfied, but coverage degrades under substantial misspecification (Breum et al., 23 Sep 2025).

6. Terminological variants, scope, and limitations

The phrase “longitudinal front-door” is used in more than one sense in the current literature, and the formulations are not interchangeable.

Formulation	Core structure	Target
Summary causal graph extension	Dynamic structural causal model; summary causal graph; latent confounding; possible cycles	$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$42
Repeated-measures front-door functional	$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$43	$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$44
Cost-aware staged sampling	Two-stage sequential measurement design with $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$45	$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$46

The cost-aware design literature uses “longitudinal” in yet another sense. In the multivariate linear front-door SEM of "Cost-Aware Optimized Front-Door Experimental Design" (Mareis et al., 23 Mar 2026), the longitudinal aspect is a two-stage sequential measurement design rather than longitudinal treatment over time: stage 1 observes $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$47, stage 2 conditionally observes $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$48, and full measurement observes $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$49. The paper derives the full-data efficient influence function, characterizes the observed-data augmentation geometry, obtains a closed-form optimal sampling policy under a budget constraint, and reports efficiency gains of $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$50 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$51 over naive full-sampling strategies (Mareis et al., 23 Mar 2026). This is a front-door problem with sequential observation, but it is not a time-varying-treatment longitudinal criterion.

A separate line of work weakens Pearl’s original graph conditions while retaining the same static front-door functional. "Generalization of Pearl’s Front-Door Criterion" (Wu et al., 16 Apr 2026) replaces Pearl’s conditions “$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$52 blocks all directed paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$53 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$54” and “$O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$55 blocks all back-door paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$56 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$57” with the weaker requirement that there are no open, proper front-door paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$58 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$59 given $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$60, together with no open back-door paths from $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$61 to $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$62. The paper states that condition (ii) is necessary, condition (i) is not necessary, and the new criterion strictly generalizes Pearl’s original one. It does not, however, formulate a dedicated longitudinal or time-indexed front-door theorem (Wu et al., 16 Apr 2026).

The longitudinal extensions also have explicit limitations. In the summary-graph setting, the criterion is sufficient but not complete, assumes stationarity and bounded lag, is sensitive to cycles when $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$63 and $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$64, and depends on the hidden-variable assumption that prevents a latent process from directly causing itself at another time point (Assaad, 2024). In the repeated-measures setting, the criterion requires strong structural assumptions, including the absence of direct $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$65 confounding, and positivity can be restrictive. Sequential regression can be difficult to specify with standard parametric models, and efficient-influence-function-based inference requires nuisance estimators to converge faster than $O=(L_0,A_0,M_0,\ldots,A_T,M_T,Y)$66, although machine learning may help achieve this rate (Breum et al., 23 Sep 2025).

Taken together, these developments relocate the front-door idea from a single-treatment, single-mediator, single-outcome theorem to a broader class of temporal problems. In one branch, the emphasis is graphical identifiability under unknown lag, latent confounding, and macro-level cycles; in another, it is semiparametric efficiency, multiply robust estimation, and valid inference for repeated exposure and mediator processes. The shared principle is unchanged: when observed mediators absorb the relevant causal transmission while satisfying the required confounding restrictions, causal effects remain identifiable even when ordinary adjustment fails (Assaad, 2024, Breum et al., 23 Sep 2025).