Organic Indirect Effect: Causal Mediation

• Organic indirect effect is a causal mediation parameter that attributes part of a treatment’s effect to mediator changes induced by an organic intervention.
• It avoids cross-world counterfactuals by reproducing the mediator’s conditional distribution, enabling practical estimation using methods like G-computation and IPW.
• Extensions address post-treatment confounding and measurement error, with applications in HIV cure research and drug development.

Searching arXiv for the primary literature on organic indirect/direct effects and related mediation formulations. arxiv_search(query="organic indirect effect mediation Lok Bosch", max_results=10) arxiv_search(query="(Lok et al., 2019) causal organic indirect and direct effects product method binary mediators", max_results=5) Searching for the cited mediation-analysis papers on arXiv. arXiv search query: "(Lok et al., 2019)" Organic indirect effect is a causal mediation estimand that attributes part of a treatment effect to treatment-induced changes in a mediator without requiring the mediator to be set to unit-specific counterfactual values. In the formulation introduced by Lok, the key object is an organic intervention on the mediator: an intervention that reproduces the mediator’s conditional distribution under the opposite treatment while leaving the outcome mechanism unchanged except through the realized mediator. This framework was developed as an alternative to natural indirect and direct effects, which rely on cross-world counterfactuals such as outcomes under treatment with the mediator set to its value under no treatment. It is intended for settings in which setting the mediator to specific values is not feasible and has since been extended to effects relative to no treatment, to product-method derivations for linear and binary mediators, to settings with post-treatment common causes of mediator and outcome, and to applications in HIV cure research (Lok, 2015, Lok et al., 2019).

1. Conceptual position within causal mediation analysis

Mediation analysis began with Baron and Kenny’s decomposition of a treatment effect into mediated and non-mediated components. Later work on natural indirect and direct effects supplied a formal causal interpretation, but did so through cross-world counterfactuals such as $Y^{(1,M^{(0)})}$, which compare outcomes under treatment with a mediator value drawn from the no-treatment world. Organic indirect and direct effects were introduced to avoid this cross-world construction by replacing “set the mediator to its untreated value” with an intervention that reproduces the distribution of the mediator under the other treatment, conditional on pre-treatment common causes (Lok et al., 2019).

This shift in formulation has two consequences. First, the estimand can be defined even when the mediator cannot be deterministically set for each unit. Second, the interpretation is intervention-based rather than value-setting: the estimand asks what would happen if one altered the mediator-generating process so that, conditional on covariates, it behaved as it would under the comparison treatment. In the original treatment-fixed formulation, the organic indirect effect is

$\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$

where $I$ is an intervention on the mediator under treatment that makes the mediator distribution match that under no treatment (Lok, 2015).

A later generalization parameterizes the effect relative to $a\in\{0,1\}$, allowing the intervention to be combined either with no treatment or with treatment. This is especially important for the “organic indirect effect relative to no treatment,”

$E[Y^{(0,I=1)}]-E[Y^{(0)}],$

which was argued to be particularly relevant for drug development because it can be estimated by combining outcome data without the new treatment and a hypothesized treatment effect on the mediator (Lok et al., 2019).

2. Formal definitions and organic interventions

In the standard notation, $A\in\{0,1\}$ is treatment, $C$ denotes all pre-treatment common causes of mediator and outcome, $M$ is the mediator, $Y$ is the outcome, and $I\in\{0,1\}$ indicates an intervention on the mediator. Let $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$0 and $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$1 be the mediator and outcome under treatment level $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$2 and intervention status $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$3, with $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$4 and $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$5 used when $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$6 (Lok et al., 2019).

An intervention $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$7 is organic relative to treatment level $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$8 and covariates $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$9 if, for all $I$0,

$I$1

and

$I$2

These conditions encode two distinct requirements. The first is a distributional shift: conditional on $I$3, the intervention must make the mediator look as it would under the opposite treatment. The second is a no direct effect of the intervention condition: conditional on the realized mediator and $I$4, the outcome distribution under the intervention must be the same as under the fixed treatment regime without the intervention (Lok et al., 2019).

With such an intervention in hand, the organic indirect and direct effects relative to $I$5 are defined as

$I$6

and

$I$7

For $I$8, the indirect effect measures the outcome change obtained by shifting the mediator distribution from its no-treatment law to its treatment law while leaving the treatment itself fixed at no treatment. For $I$9, the sign convention reverses relative to the original 2016 presentation, because the comparison is then between the treated world and a treated world in which the mediator is shifted back toward its no-treatment distribution (Lok et al., 2019, Lok, 2015).

3. Identification and the mediation formula

Under randomized treatment, identification in the basic setting requires consistency, no post-treatment confounders of the $a\in\{0,1\}$0 relation, and the organic-intervention conditions. In observational studies one additionally assumes no unmeasured confounding for $a\in\{0,1\}$1 given measured covariates, written in the 2019 formulation as conditioning on $a\in\{0,1\}$2 when needed (Lok et al., 2019).

Under these assumptions, the mediation formula identifies the mean outcome under the organic intervention: $a\in\{0,1\}$3 For the original treatment-fixed formulation,

$a\in\{0,1\}$4

which yields an observed-data expression for the organic indirect effect by subtraction from $a\in\{0,1\}$5 (Lok, 2015).

This formulation clarifies why organic effects avoid cross-world counterfactuals. The identifying conditions compare $a\in\{0,1\}$6 with $a\in\{0,1\}$7 under the same treatment level and compare mediator distributions under different treatment levels, but they do not invoke a quantity such as $a\in\{0,1\}$8. That avoidance does not eliminate all causal assumptions: the existence of a meaningful organic intervention and the absence of mediator–outcome confounding remain essential (Lok et al., 2019).

Several estimation strategies follow directly from the mediation formula. The 2016 exposition describes G-computation (parametric), Inverse-probability weighting (IPW), and Doubly-robust/AIPW estimators. In each case, inference can proceed via the nonparametric bootstrap or sandwich-variance formulas (Lok, 2015). A later HIV application emphasizes an additional robustness property for the organic indirect effect relative to no treatment: because it uses only $a\in\{0,1\}$9 together with the treatment-induced shift in the mediator distribution, treatment–mediator interaction in the outcome model does not enter the identification formula (Herath et al., 15 Jul 2025).

4. Product methods and connections to classical estimators

A central result is that the familiar product method survives in the organic framework and extends beyond the classical no-interaction case. Consider the linear structural models

$E[Y^{(0,I=1)}]-E[Y^{(0)}],$0

$E[Y^{(0,I=1)}]-E[Y^{(0)}],$1

If there is no treatment–mediator interaction, $E[Y^{(0,I=1)}]-E[Y^{(0)}],$2, then for the organic indirect effect relative to no treatment,

$E[Y^{(0,I=1)}]-E[Y^{(0)}],$3

and the direct effect reduces to $E[Y^{(0,I=1)}]-E[Y^{(0)}],$4. If there is treatment–mediator interaction, $E[Y^{(0,I=1)}]-E[Y^{(0)}],$5, the same product formula still holds for the indirect effect: $E[Y^{(0,I=1)}]-E[Y^{(0)}],$6 while the direct effect now involves both $E[Y^{(0,I=1)}]-E[Y^{(0)}],$7 and $E[Y^{(0,I=1)}]-E[Y^{(0)}],$8 (Lok et al., 2019).

This matters because it shows that, in all linear-model cases considered, the organic indirect effect relative to no treatment admits the familiar product $E[Y^{(0,I=1)}]-E[Y^{(0)}],$9. When the outcome model is linear and has no $A\in\{0,1\}$0–$A\in\{0,1\}$1 interaction, the resulting estimator coincides with the original Baron–Kenny product. The organic framework therefore preserves the classical estimator in the setting where it is already standard while extending the interpretation to explicitly causal, intervention-based estimands (Lok et al., 2019).

For binary mediators, the product method takes a different form. If $A\in\{0,1\}$2, then at $A\in\{0,1\}$3,

$A\in\{0,1\}$4

$A\in\{0,1\}$5

Writing

$A\in\{0,1\}$6

$A\in\{0,1\}$7

the effect becomes $A\in\{0,1\}$8. This is the product method for binary mediators on the difference-in-means scale, and on the risk-difference scale when $A\in\{0,1\}$9 is binary (Lok et al., 2019).

5. Extensions: post-treatment common causes, censoring, and measurement error

Most mediation analyses assume that there are no post-treatment common causes of the mediator and the outcome. Organic effects were extended to the case in which a post-treatment variable $C$0 is a common cause of $C$1 and $C$2, with the intervention on $C$3 taking place after $C$4 has been realized. In this setting, an intervention is organic after $C$5 if, for all $C$6,

$C$7

and

$C$8

Under randomization or no unmeasured confounding of $C$9, consistency, positivity, and these organic-intervention conditions, one obtains

$M$0

This identification result makes organic indirect effects available in settings where natural indirect effects fail because post-treatment confounders block cross-world identification (Lok, 2015).

A worked linear example in this post-treatment-confounding setting uses

$M$1

and yields

$M$2

which coincides with the familiar product term up to sign (Lok, 2015).

Recent work also extends estimation to mediators subject to an assay limit and to classical measurement error. In the HIV setting, the observed mediator $M$3 is left-censored below a lower assay limit, and estimation proceeds by maximizing an observed-data likelihood built from a mediator model and an outcome probit model under $M$4. For classical measurement error, with

$M$5

analytic correction formulas relate the coefficients of a working probit on $M$6 to the corresponding coefficients for the true mediator model. Under a hypothesized left-shift $M$7 in the mediator distribution, the organic indirect effect can then be estimated by plugging the corrected parameters into a closed-form version of the mediation formula and using the Delta method for standard errors (Herath et al., 15 Jul 2025).

6. Applications, interpretation, and common points of confusion

A prominent application domain is HIV cure development. The 2019/2021 work considers ART-interruption data without curative HIV treatments and estimates the organic indirect effect of hypothetical curative treatments mediated by two HIV-persistence measures: single-copy plasma HIV-RNA (SCA) and cell-associated HIV-RNA (CA). The outcome is viral suppression by week 4 or week 8 post-ART, and estimation combines a fitted outcome model under $M$8 with a hypothesized treatment effect on the mediator distribution. For a binary mediator, the procedure fits logistic regressions for $M$9 and $Y$0, imposes a hypothetical treatment effect such as multiplying the mediator odds by a factor $Y$1, and plugs the resulting probabilities into the binary-mediator product formula. For a continuous log$Y$2 mediator, the same idea is implemented through the general mediation formula with a distributional shift by $Y$3 (Lok et al., 2019).

The reported numerical findings illustrate how the estimand functions as a screening tool. At week 4 with $Y$4, the indirect effect was approximately $Y$5 with $Y$6 CI approximately $Y$7 for SCA-RNA and approximately $Y$8 with $Y$9 CI approximately $I\in\{0,1\}$0 for CA-RNA. Week 8 effects were larger. For continuous $I\in\{0,1\}$1 shifts of CA-RNA, the indirect effect was approximately $I\in\{0,1\}$2 with $I\in\{0,1\}$3 CI approximately $I\in\{0,1\}$4 at week 4 and approximately $I\in\{0,1\}$5 with $I\in\{0,1\}$6 CI approximately $I\in\{0,1\}$7 at week 8. These results suggest CA-RNA may be a more promising mediator for curative interventions (Lok et al., 2019).

A later analysis incorporating assay limits and measurement error reported larger adjusted effects. For a $I\in\{0,1\}$8 downward shift, the measurement-error-adjusted organic indirect effect for caRNA was $I\in\{0,1\}$9 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$00 at week 4 and $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$01 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$02 at week 8, whereas for SCA it was $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$03 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$04 at week 4 and $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$05 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$06 at week 8. A two-log drop in caRNA yielded approximately $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$07 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$08 at week 4, compared with approximately $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$09 $\text{OIE}=E(Y_1)-E(Y_{1,I=1}),$10 for SCA (Herath et al., 15 Jul 2025).

Several recurrent misunderstandings are addressed by the literature. One is that “organic” means “natural”; in fact the organic estimand was introduced precisely to avoid the cross-world formulation of natural indirect effects. Another is that avoiding cross-worlds makes the analysis assumption-free; it does not, because one still needs an organic intervention and the relevant no-confounding or randomization assumptions. A third is that the effect always requires treated outcomes; the “relative to no treatment” formulation shows that one can estimate only the pure indirect effect or organic indirect effect by combining a hypothesized treatment effect on the mediator with outcome data without treatment, which is why the estimand has been presented as particularly relevant for selecting prospective treatments for further evaluation in randomized clinical trials (Herath et al., 15 Jul 2025, Lok et al., 2019).