Counterfactual Causal Mediation Framework

• Counterfactual causal mediation is a framework that decomposes total treatment effects into natural direct and indirect effects via nested potential outcomes.
• It relies on key assumptions like consistency, positivity, and no unmeasured confounding to enable the identification of mediation pathways.
• The framework extends to multiple mediators and path-specific, interventional, and latent-variable settings, offering robust tools for causal inference.

Counterfactual causal mediation is a causal-inference framework that decomposes a total effect into components operating through one or more mediators and components operating outside those mediators by defining effects with potential outcomes such as $Y(a)$, $M(a)$, $Y(a,m)$, and nested quantities such as $Y(a,M(a'))$. In this formulation, mediation is not a property of regression coefficients but of contrasts between hypothetical interventions, and the same formalism extends from single-mediator natural direct and indirect effects to path-specific effects, multiple mediators, longitudinal mediation, and hidden-variable settings (Shpitser, 2012, Shpitser et al., 2020, Robins et al., 2020).

1. Core counterfactual definitions

In the classical single-mediator setting with treatment $A$, mediator $M$, and outcome $Y$, the total causal effect is defined nonparametrically as

$E[Y(1)-Y(0)].$

The central mediation quantities are then expressed with nested counterfactuals. A standard formulation defines the natural direct effect as

$E[Y(1,M(0))]-E[Y(0)],$

and the natural indirect effect as

$E[Y(1)]-E[Y(1,M(0))].$

These satisfy the decomposition

$M(a)$0

The same framework also includes the controlled direct effect

$M(a)$1

and the pure or natural direct effect

$M(a)$2

These definitions place mediation squarely in the counterfactual or potential-outcomes tradition rather than in linear-regression heuristics (Shpitser, 2012, Robins et al., 2020).

Alternative notational conventions preserve the same decomposition logic. In one widely used formulation,

$M(a)$3

$M(a)$4

so that $M(a)$5. In natural-experiment formulations the same estimands are written as average direct and indirect effects,

$M(a)$6

The notation varies across literatures, but the target remains a decomposition of a treatment effect into mediated and non-mediated components (Hong et al., 2021, Hogan-Hennessy, 7 Aug 2025).

A central historical claim of the framework is that the familiar product and difference methods are special cases of this broader causal formulation. Under linear, additive, no-interaction models, the indirect effect reduces to the product of the exposure-to-mediator and mediator-to-outcome coefficients, and the difference method coincides with that product. Outside those restrictions, regression coefficients need not equal causal direct or indirect effects (Shpitser, 2012).

2. Identification, assumptions, and graphical semantics

Counterfactual mediation estimands are not identified from observed data without additional assumptions. The standard assumptions are consistency, positivity, and no unmeasured confounding or sequential ignorability. In one formulation these appear as

$M(a)$7

together with the cross-world condition

$M(a)$8

For total effects, the required ignorability is weaker; for natural direct and indirect effects, the framework generally requires assumptions that connect incompatible hypothetical worlds (Shpitser, 2012, Gao et al., 2020).

Under those assumptions, the nested counterfactual mean $M(a)$9 is identified by the mediation formula

$Y(a,m)$0

An equivalent expression for the pure direct effect in a simple DAG $Y(a,m)$1 is

$Y(a,m)$2

This is the standard nonparametric identification result for natural direct effects in the absence of unmeasured confounding (Shpitser, 2012, Robins et al., 2020).

Graphical formulations clarify which independences are being assumed. Single-World Intervention Graphs (SWIGs) encode counterfactual variables directly and allow ordinary d-separation on the SWIG to imply counterfactual conditional independence under the FFRCISTG model. In that language, the interventional joint law obeys the extended g-formula

$Y(a,m)$3

and Pearl’s do-calculus can be reformulated as a potential-outcome or po-calculus. This makes mediation assumptions more transparent by distinguishing single-world independences from stronger cross-world structure (Shpitser et al., 2020).

A major distinction in the literature is between FFRCISTG and NPSEM-IE. Under FFRCISTG, some interventionist quantities are identified from single-world counterfactual independences, but the mediation formula for the pure direct effect does not generally follow. Under NPSEM with independent errors, additional cross-world independences such as $Y(a,m)$4 arise, and the pure direct effect is point identified under the usual no-confounding assumptions. This difference is one reason the interpretation of natural effects remains closely tied to structural assumptions rather than to observed-data functionals alone (Robins et al., 2020).

3. Path-specific effects, multiple mediators, and interaction decompositions

The framework generalizes from a single mediator to path-specific effects, in which one asks for the effect of treatment on outcome along a selected set of causal paths $Y(a,m)$5. In the path-specific formulation, value $Y(a,m)$6 is sent along the active paths in $Y(a,m)$7 and value $Y(a,m)$8 along the complement, yielding counterfactuals $Y(a,m)$9. Under NPSEM-IE, edge-consistent path-specific counterfactuals are identified by a path-specific analogue of the mediation formula, and on the edge-expanded graph the problem reduces to standard identification of an intervention distribution by the g-formula or ID algorithm. The central obstruction is the recanting witness, and with hidden variables its generalization is the recanting district (Shpitser, 2012, Robins et al., 2020).

This graphical theory is especially important in longitudinal mediation. In acyclic directed mixed graphs with hidden confounding, path-specific effects are identifiable as functionals of interventional densities if and only if there is no recanting district. The recanting-district criterion is therefore the longitudinal and hidden-variable generalization of the earlier recanting-witness condition. It formalizes the idea that a node or district cannot be required to transmit both active and baseline treatment values along different paths simultaneously (Shpitser, 2012).

With multiple mediators, the framework branches into several regimes. For uncausally related mediators $Y(a,M(a'))$0, the joint indirect effect is

$Y(a,M(a'))$1

while the indirect effect through a single mediator $Y(a,M(a'))$2 is

$Y(a,M(a'))$3

Identification is based on the Sequential Ignorability for Multiple Mediators Assumption (SIMMA), and the joint mediator distribution must be modeled when latent common causes induce spurious correlation among mediators (Jerolon et al., 2018).

A separate line of work addresses causally ordered mediators and interaction terms. For two mediators, sequential or non-sequential, the total effect can be decomposed using natural counterfactual interaction effects ($Y(a,M(a'))$4). In the non-sequential case the paper derives a 10-component decomposition, while in the sequential case it derives a 9-component decomposition because some reference interaction terms are not separately identifiable when mediator-to-mediator edges induce problematic nested counterfactuals. The central idea is that interaction should be evaluated with mediators at their natural counterfactual values rather than at arbitrary fixed constants (Gao et al., 2020, Gao et al., 2020).

When mediator ordering is unknown or high-dimensional, the literature often moves from natural effects to interventional effects. One formulation defines outcomes under a joint draw from the counterfactual mediator distribution,

$Y(a,M(a'))$5

or under separate marginal draws for each mediator,

$Y(a,M(a'))$6

These estimands support direct effects, joint indirect effects, separate mediator-specific indirect effects, a mutual-dependence component, and a remainder component, while avoiding the need to specify mediator ordering or a full joint mediator model in high-dimensional problems (Loh et al., 2020).

4. Interventionist and interventional reformulations

A major reformulation of mediation replaces the idea of “holding the mediator fixed” by interventions on an expanded causal graph in which treatment is decomposed into separable components. In the nicotine example, treatment $Y(a,M(a'))$7 is decomposed into $Y(a,M(a'))$8 and $Y(a,M(a'))$9 with deterministic relations $A$0. Under the expanded graph, the PDE-like contrast becomes

$A$1

which is an ordinary intervention contrast rather than a nested mediator counterfactual. Under structural no-direct-effect conditions linking the expanded graph to the original graph, one obtains

$A$2

and hence the classical pure direct effect is recovered as a special case (Robins et al., 2020).

A related but distinct move is to define interventional direct and indirect effects through stochastic interventions on mediator distributions. Instead of setting the mediator to an individual’s own counterfactual value, the mediator vector is randomly drawn from its counterfactual distribution under treatment or control. This weakens the assumptions relative to path-specific natural effects and remains meaningful with many mediators, unknown mediator ordering, and possible hidden common causes among mediators. Estimation can then proceed by Monte Carlo integration using resampled mediator values and a fitted outcome model (Loh et al., 2020).

Another reformulation grounds path-specific effects in interventions on information transfer rather than node values. In the information-account approach, each structural mechanism is decomposed into information transfer along edges and information processing at nodes, and path intervention explicitly routes the information $A$3 through a selected path $A$4. The corresponding $A$5-formula identifies the joint counterfactual distribution under the path intervention and, unlike the Robins et al. approach discussed in that paper, does not require the non-existence of a recanting witness. Closely related non-agency interventions define path-specific decompositions by altering the information propagated through the edges of the graph and are proposed specifically for settings with treatment-induced mediator-outcome confounding (Gong et al., 2021, Díaz, 2022).

A more recent distributional alternative is Sequential Transport, which combines optimal transport with a mediator DAG. Instead of cross-world counterfactuals such as $A$6, it constructs unit-level mediator counterfactuals by sequentially transporting each mediator toward its distribution under an alternative treatment while preserving the conditional dependencies encoded by the DAG. For numerical mediators the transport map is

$A$7

and the induced direct and indirect effects are interpreted causally when treatment is randomized or ignorable, or distributively otherwise (Fernandes-Machado et al., 16 Mar 2026).

5. Extensions for difficult data-generating settings

A large modern literature extends the framework to settings in which standard sequential ignorability is implausible or incomplete. One prominent example is post-treatment confounding of the mediator-outcome relation. In that case, a treatment-induced covariate $A$8 affects both mediator and outcome, but the counterfactual $A$9 is missing among treated units and $M$0 is missing among controls. A sensitivity-analysis approach addresses this by treating the missing counterfactual post-treatment confounder as latent and indexing its dependence by

$M$1

with weighting-based identification of NIE and NDE implemented through either imputation or integration (Hong et al., 2021).

Another extension treats the mediator itself as hidden. Proximal hidden mediation introduces two proxies $M$2 and $M$3 for an unobserved mediator $M$4, together with bridge functions $M$5 and $M$6 that solve conditional moment equations. This allows identification of the cross-world mean

$M$7

as well as hidden front-door and population intervention indirect effect functionals, under proxy conditional independences and completeness conditions (Ghassami et al., 2021).

A different response to latent structure is to infer a surrogate confounder from multiple mediator pathways themselves. In the latent multiple-mediator framework, a generalized structural equation model introduces structured latent factors that deconfound treatment, mediators, and outcome simultaneously. The mediator-sharing confounder assumption and the identification theorem then justify conditioning on a learned surrogate $M$8 rather than on an unobserved $M$9, with direct, mediation, and total effects expressed by integrating the conditional models over $Y$0 and $Y$1 (Yuan et al., 2023).

Specialized outcome and mediator structures have also produced dedicated mediation formalisms. For zero-inflated mediators, the mediator is reinterpreted as two sequential mediators, the numeric value $Y$2 and its binary status $Y$3, leading to the decomposition

$Y$4

where $Y$5 is attributable to the mediator’s numerical scale and $Y$6 to the binary change from zero to non-zero. For recurrent event data, the target becomes

$Y$7

with natural direct and indirect effects defined on the recurrent-event mean scale and estimated by a semiparametric triply robust procedure (Jiang et al., 2023, Chen et al., 2023).

Natural experiments pose a further difficulty because treatment may be ignorable while the mediator is chosen by agents. A control-function approach retains the standard ADE and AIE estimands but replaces mediator ignorability with a selection model,

$Y$8

plus mediator monotonicity, selection on unobserved gains, and an instrument for mediator take-up costs. The second-stage outcome model then includes control functions $Y$9 and $E[Y(1)-Y(0)].$0, and the indirect effect acquires an explicit complier-adjustment term (Hogan-Hennessy, 7 Aug 2025).

6. Interpretation, testability, and the status of cross-world assumptions

The framework has always carried an interpretive tension between mathematically defined nested counterfactuals and the dictum “no causation without manipulation.” The interventionist expanded-graph approach addresses that tension directly by tying direct and indirect effects to conceivable interventions on treatment components rather than to direct manipulations of the mediator. In principle, a future randomized experiment on those components could test whether an off-diagonal intervention such as $E[Y(1)-Y(0)].$1 agrees with the formula inferred from ordinary two-arm data. This makes mediation empirically testable in a way the original nested counterfactual expression does not (Robins et al., 2020).

A related controversy concerns whether alternative mediated-effect definitions truly test mechanisms. The non-agency intervention literature argues that popular randomized interventions on the mediator can fail the sharp mediational null, because an indirect effect can be nonzero even when there is no true structural mechanism through the mediator. By contrast, information-transfer path-specific effects are proposed as valid statistical tests of mechanisms because they vanish whenever there is no causal influence along the relevant path (Díaz, 2022).

The counterfactual layer itself has also become an object of formal study. Canonical representations of Markovian structural causal models replace structural equations by one-step-ahead counterfactual process laws $E[Y(1)-Y(0)].$2, thereby characterizing the counterfactual equivalence class of structural causal models. In that representation, the observational and interventional constraints are fixed by the causal graphical model, while the cross-world coupling remains a choice of counterfactual conception rather than an estimand learned from data. A plausible implication is that many mediation disagreements are disagreements about cross-world coupling rather than about the observational or interventional layer (Lara, 22 Jul 2025).

Taken together, these developments define the contemporary scope of the counterfactual causal mediation framework. It remains the standard language for natural direct, natural indirect, and path-specific effects; it is closely integrated with DAGs, SWIGs, ADMGs, g-formulas, and identification algorithms; and it now includes interventionist, interventional, proximal, sensitivity-analysis, transport-based, and latent-variable variants. The common core is unchanged: mediation is a question about how treatment effects propagate through specified intermediate variables or paths, but the meaning, identifiability, and testability of that question depend on the counterfactual structure one is willing to assume (Shpitser et al., 2020, Robins et al., 2020).