Reweighted Mediation Formula Overview

• Reweighted Mediation Formula is an identification method that expresses counterfactual outcomes as weighted averages of observed mediator distributions across different treatment regimes.
• It leverages reweighting techniques to transport mediator distributions, enabling estimation of natural direct and indirect effects within causal mediation analysis.
• The approach extends to complex settings like multiple mediators, selection bias, and continuous treatments, underpinning efficient balancing estimators in causal inference.

Searching arXiv for recent and foundational papers on reweighted mediation formulas, weighting estimators, and path-specific effects. The reweighted mediation formula is a family of identification and estimation results in causal mediation analysis that rewrite counterfactual quantities such as $E\{Y(a,M(a'))\}$ as weighted functionals of observed data. In the simplest single-mediator setting, it is the mediation formula associated with natural direct and indirect effects; in later developments, the same reweighting logic is extended to nonparametric balancing estimators, general treatment regimes, multiple causally ordered mediators, path-specific effects, and settings with sample selection bias (Steen et al., 2018, Huang et al., 2022).

1. Counterfactual targets and the role of cross-world quantities

Mediation analysis seeks to infer how much of the effect of an exposure on an outcome can be attributed to specific pathways via intermediate variables or mediators. The canonical path-specific quantities in the single-mediator case are the natural indirect effect and natural direct effect: $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$ These are defined through nested counterfactuals such as $Y(a,M(a'))$, which the literature characterizes as “cross-world” quantities because, for $a \neq a'$, they combine interventions that are contradictory in any single world (Steen et al., 2018).

A closely related notation writes

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$

so that, for a general treatment variable, the average treatment effect decomposes as

$$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$

Under this parameterization, the first contrast is a direct effect and the second is an indirect effect (Huang et al., 2022).

The technical difficulty is that identification of total effects and identification of path-specific effects are not parallel problems. Adjustment for confounding is generally not sufficient for path-specific effects, because their magnitude is also determined by the extent to which individuals who experience large exposure effects on the mediator tend to experience relatively small or large mediator effects on the outcome. This dependence is encoded in the cross-world counterfactual structure of $Y(a,M(a'))$ and is not resolved by ordinary back-door adjustment alone (Steen et al., 2018).

2. Canonical mediation formula as a reweighting identity

In the standard single-mediator setting, when covariate adjustment suffices, the identifying functional is Pearl’s mediation formula: $$p(Y(a,M(a'))=y)=\sum_{\mathbf c,m} p(y\mid a,m,\mathbf c)\,p(m\mid a',\mathbf c)\,p(\mathbf c).$$ Its two components are $p(y\mid a,m,\mathbf c)$, the outcome law under treatment $a$ at mediator level $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$0, and $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$1, the mediator law under the alternative exposure level $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$2 (Steen et al., 2018).

The reweighting interpretation is direct. The observed outcome distribution under treatment $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$3 and mediator level $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$4 is reweighted by how the mediator would have been distributed had treatment been $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$5, within levels of pre-treatment covariates $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$6. In this sense, the mediation formula does not merely integrate over the mediator; it transports the mediator distribution from one treatment regime into another. That interpretation motivates later weighting estimators, which target the same counterfactual mean by constructing pseudo-samples whose mediator distributions mimic the counterfactual regime (Steen et al., 2018).

Equivalent weighted representations also appear in estimation theory. For the natural direct effect component

$$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$7

one nonparametric representation is

$$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$8

which makes the reweighting structure explicit: treated outcomes are reweighted by a treatment weight and by a mediator density ratio that shifts the treated mediator distribution toward the control mediator distribution (Chan et al., 2016).

3. Identification assumptions, graphical criteria, and failure modes

The mediation formula requires stronger assumptions than those needed for total-effect identification. In the formulation summarized in the graphical-model literature, the key conditions are conditional cross-world independence $$\text{NIE} = E\{Y(1, M(0)) - Y(0)\}, \qquad \text{NDE} = E\{Y(1, M(0)) - Y(0, M(0))\}.$$9, ignorability of treatment $Y(a,M(a'))$0, and the requirement that no element of $Y(a,M(a'))$1 is affected by $Y(a,M(a'))$2 (Steen et al., 2018).

Graphically, the covariate set $Y(a,M(a'))$3 must adjust for all confounding of the $Y(a,M(a'))$4 relation, and no element of $Y(a,M(a'))$5 may be treatment-induced. The paper states this through a d-separation condition in the DAG with arrows from $Y(a,M(a'))$6 and $Y(a,M(a'))$7 removed: $Y(a,M(a'))$8 must block all back-door paths between $Y(a,M(a'))$9 and $a \neq a'$0 not through $a \neq a'$1. Only pre-treatment covariates can be used for this purpose. If a variable $a \neq a'$2 is both affected by $a \neq a'$3 and a common cause of $a \neq a'$4 and $a \neq a'$5, then ordinary mediation-formula identification fails (Steen et al., 2018).

This failure is often described as a recanting witness or, in the more general formulation, a recanting district. For path-specific effects in complex graphs, the complete graphical criterion is that the effect is identifiable if and only if there is no recanting district. When the criterion holds, the identifying functional is an edge g-formula; when it fails, some kernels require integration over conflicting treatment values, so nonparametric identification is impossible (Steen et al., 2018).

A related formulation in the interventionist literature reaches the same obstruction from a different direction. In edge-expanded graphs or separable-treatment representations, identification of path-specific counterfactuals is again characterized by the absence of a recanting witness or district, and the resulting algorithms are described as sound and complete for determining identification (Robins et al., 2020). This suggests that the reweighted mediation formula is best understood as one special case of a broader identification theory for path-specific functionals.

4. Weighting estimators, balancing conditions, and efficiency theory

A major line of work replaces plug-in modeling of $a \neq a'$6, $a \neq a'$7, and $a \neq a'$8 with direct construction of balancing weights. In one nonparametric approach, the target parameter is estimated by choosing treated-group weights $a \neq a'$9 so that weighted empirical moments of $\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$0 match corresponding moments in a reweighted control group. The core balancing property is

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$1

for any integrable function $\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$2, with $\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$3. The empirical estimator is then obtained from a convex optimization problem over basis functions $\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$4, with a dual solution

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$5

leading to

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$6

This estimator is proved to be globally semiparametric efficient, and the paper also gives a variance estimator based on the empirical influence function (Chan et al., 2016).

Another strand develops a “menu” of weighting and model-based estimators for marginal natural effects and shows that cross-world weighting can be written in three equivalent ways: $\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$7

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$8

and the novel stabilized form

$\mu(t,t') = \mathbb{E}[Y(t,M(t'))],$9

These parameterizations motivate both model-based estimation and direct balancing approaches, and they make it possible to diagnose covariate balance and joint $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$0 balance in the pseudo cross-world sample (Nguyen et al., 2021).

A later balancing-weights formulation addresses two limitations of standard mediation IPW and EIF estimators: instability from inverse propensity scores and finite-sample imbalance. The proposed estimators are based on weights that directly penalize weight dispersion while enforcing approximate covariate and mediator balance, using a two-step minimal-dispersion construction. Under the stated sieve conditions, the resulting EIF-type and IPW-type estimators are asymptotically normal and achieve the semiparametric efficiency bound (Kawato, 10 Dec 2025). A plausible implication is that the reweighted mediation formula has evolved from an identification formula into a design principle for constructing finite-sample-stable pseudo-populations.

5. Extensions to general treatments, ordered mediators, and alternative weighting schemes

For treatments that are binary, multi-valued, continuous, or mixed, the reweighting logic persists but the weights must be generalized. One nonparametric framework identifies

$$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$1

where $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$2 for $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$3 and $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$4. The corresponding weights are estimated by solving moment equations

$$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$5

approximated with basis expansions and a max-entropy problem, with dual solution

$$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$6

For discrete treatments, these estimators attain the semiparametric efficiency bound; for continuous treatments, their convergence rates are slower than $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$7, but the paper states that they are still more efficient than estimators constructed from the true weighting function (Huang et al., 2022).

With multiple causally ordered mediators $$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$8, the identifying object becomes the generalized mediation functional

$$\mu(t,t) - \mu(t',t') = \big(\mu(t,t)-\mu(t',t)\big) + \big(\mu(t',t)-\mu(t',t')\big).$$9

which generalizes Pearl’s mediation formula to the ordered-mediator setting. Weighting estimators can be written either in terms of mediator-density ratios or, via Bayes’ rule, in terms of treatment-model odds ratios. The most prominent semiparametric estimators in this framework are described as $Y(a,M(a'))$0-robust and locally semiparametric efficient (Zhou, 2020).

An alternative nonparametric scheme is ratio of mediator probability weighting. Its core weight for the counterfactual mean $Y(a,M(a'))$1 is

$Y(a,M(a'))$2

The method is designed to relax the assumption of no treatment-mediator interaction while avoiding outcome-model specification, and the paper states that it is suitable for handling a large number of pretreatment covariates (Hong, 3 Jun 2025). This suggests a convergence between mediation weighting and the broader balancing-weight literature: both aim to recover counterfactual mediator distributions without committing to restrictive outcome regressions.

6. Selection bias, interventionist reformulations, and specialized domains

The reweighted mediation formula has also been extended to samples subject to selection. In a DAG augmented with a binary selection node $Y(a,M(a'))$3, one identifying formula for the natural indirect or direct effect is

$Y(a,M(a'))$4

Here $Y(a,M(a'))$5 denotes baseline covariates observed in the full target population, while the remaining terms are estimated in the selected sample. The practical reweighting device is inverse probability of selection weighting,

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

The paper formulates sufficient DAG-based conditions by extending the Generalized Adjustment Criterion to mediation and path-specific effects under selection bias (Khan et al., 2 Sep 2025).

In the liver-transplantation application associated with that framework, the natural indirect effect in risk ratio was $Y(a,M(a'))$7 after selection adjustment and the natural direct effect in risk ratio was $Y(a,M(a'))$8; these were nearly identical to the naive estimates, which the paper interprets as showing that formal correction justified the lack of substantial bias in that dataset (Khan et al., 2 Sep 2025). The substantive importance is methodological rather than numerical: the reweighted mediation formula can separate nonresponse or evaluation-completion mechanisms from the mediation question itself.

A different reformulation avoids natural-effect cross-world counterfactuals by expanding treatment into separable components in an enlarged graph. In that interventionist approach, direct and indirect effects are defined by ordinary interventions on treatment components rather than by interventions on the mediator, yet when separability holds the identifying formula coincides with Pearl’s mediation formula. The paper proves that the interventionist and nested-counterfactual approaches remain tightly coupled under a Non-Parametric Structural Equation Model except in the presence of a recanting witness (Robins et al., 2020). This is not merely a philosophical variation: it changes whether the mediation question is, in principle, empirically testable in a future randomized design.

Specialized domains have prompted further variants. In microbiome mediation, where mediators are high-dimensional, zero-inflated, and dependent, and where exposure-induced mediator-outcome confounding is present, one paper develops a novel nonparametric identification formula for the interventional indirect effect and an inverse probability weighting algorithm for estimation. The key point is that this formula does not require the cross-world independence assumption needed for natural effects, and the motivating AML application suggests that the effect of induction chemotherapy intensity on infection is mainly mediated by patients’ gut microbiome (Zhang et al., 2021).

Across these developments, a recurrent misconception is that mediation formulas are simply ordinary adjustment formulas with extra algebra. The literature instead treats them as identification results for path-specific counterfactuals whose validity depends on cross-world independence, separability, or alternative interventionist assumptions; on the absence of recanting districts; and, in weighted implementations, on positivity and adequate balance. The reweighted mediation formula is therefore best regarded as a technically specific bridge between graphical identification theory and weighted statistical estimation, not as a generic synonym for mediation analysis (Steen et al., 2018)