Relative Indirect Effects in Mediation Analysis

• Relative Indirect Effects (RelIE) quantify mediation by comparing the natural indirect effect with total or direct effects in causal models.
• They facilitate the interpretation of indirect pathways in fields such as epidemiology, ecology, and network science, enhancing comparisons of intervention strategies.
• Robust estimation methods like RMPW and front-door identification support reliable inference under assumptions including no unmeasured confounding and positivity.

Relative Indirect Effects (RelIE) refer to a class of quantitative measures expressing the magnitude of mediated (indirect) effects relative to total or direct effects in causal inference, mediation analysis, intervention studies, and network/ecosystem models. These metrics facilitate the interpretation of mediation's contribution—distinguishing the proportion or comparative size of a causal effect transmitted through an intermediate variable (“mediator”) or via indirect transmission chains. RelIE are foundational for quantifying causal mechanisms, comparing intervention strategies, and determining system-level importance of indirect pathways across diverse scientific domains such as biostatistics, epidemiology, ecology, and network science.

1. Definitions and Key Variants

Relative Indirect Effect metrics summarize the importance of mediation on a relative scale. Two primary variants are encountered:

• Proportion mediated (RelIE₁): The fraction of the total effect explained by the indirect (mediator) pathway,

$\text{RelIE}_1 = \frac{\text{NIE}}{\text{TE}}$

where NIE (Natural Indirect Effect) and TE (Total Effect) are defined via counterfactuals (Hong, 3 Jun 2025, Peña, 2023).

• Indirect-to-direct ratio (RelIE₂): The ratio of indirect to direct effect,

$\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$

where NDE is the Natural Direct Effect.

In special cases, such as survival analysis and infectious-disease modeling, analogous ratios compare relative survival or infection probabilities under intervention scenarios. Ecological applications often use the ratio of indirect to direct flow (I/D ratio), which are mathematically equivalent in interpretation to RelIE.

2. Formal Frameworks and Identification

Causal Mediation

Let $A$ be a binary treatment, $M$ a mediator, $Y$ an outcome, and $X$ pre-treatment covariates. Using counterfactual notation (Hong, 3 Jun 2025, Peña, 2023):

• $M(a)$: mediator under intervention $A=a$.
• $Y(a, m)$: outcome under $A=a$, $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$0.

Effects are defined as:

• Total Effect (TE): $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$1
• Natural Direct Effect (NDE): $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$2
• Natural Indirect Effect (NIE): $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$3

Relative Indirect Effect quantifies the contribution of mediation via:

• $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$4
• $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$5

Front-door identification offers an alternative when direct backdoor adjustment is invalid, allowing relative indirect effects to be computed under the weaker front-door conditions (Peña, 2023).

Network and Epidemiological Systems

In systems with interference or ecological networks:

• Ecology: The I/D ratio is defined as the system-level sum of indirect flows divided by direct flows, usually using realized (empirically observed) input/output boundary conditions (Borrett et al., 2011).
• Infectious Disease: RelIE compares the number of infections averted indirectly (through herd effects) to those averted directly among vaccinees, both on population and per-capita bases (Lin et al., 2024).

Survival Analysis

For time-to-event settings with time-dependent mediators:

• Relative indirect effect on survival: $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$6 with $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$7 the survival function under various interventions, and $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$8, $\text{RelIE}_2 = \frac{\text{NIE}}{\text{NDE}}$9 estimated via additive hazards and sequential linear models (Aalen et al., 2020).

3. Estimation Methodologies

Ratio of Mediator Probability Weighting (RMPW)

The RMPW estimator (Hong, 3 Jun 2025) provides a non-parametric, interaction-robust procedure for estimating counterfactual means needed for RelIE. Key steps include:

• Estimation of treatment propensity $A$0.
• Fitting mediator density models conditional on $A$1, via logistic or kernel methods.
• Propensity stratification: grouping units by estimated mediator probabilities for robust weight calculation.
• Weighted regression or pseudo-sample stacking to obtain point estimators for TE, NIE, NDE, and thus RelIE variants.

Estimation formulas:

Effect	Estimator
$A$2	$A$3
$A$4	$A$5
$A$6	$A$7
$A$8	$A$9
$M$0	$M$1

Variance estimation uses robust sandwich estimators or bootstrap.

Ecological, Interference, and Epidemic Metrics

• Ecology: Realized I/D is computed using observed flows and steady-state solutions to network equations, with rigorous equivalence of input- and output-oriented measures (Borrett et al., 2011).
• Interference Trials: RelIE defined as the prevalence of units affected by others' treatment, with conservative lower bounds estimated via inverse-propensity weighted test statistics and convex programming (Choi, 2023).
• Epidemics: RelIE at the population level is quantified by the indirect/ direct ratio based on SIR final-size equations; per-capita ratios compare the average benefit for unvaccinated to vaccinated individuals (Lin et al., 2024).

4. Interpretation and Domain-Specific Applications

RelIE provides interpretable metrics of mediation and systemic indirect influence:

• Causal Mediation: $M$2 as the proportion mediated (e.g., "what portion of the drug effect operates via the enzyme pathway") (Peña, 2023). $M$3 contextualizes the indirect path's size relative to the direct intervention.
• Epidemiology: When $M$4, the population-level RelIE for vaccination exceeds unity, indicating greater total indirect benefit among unvaccinated than direct protection (e.g., herd immunity effects) (Lin et al., 2024).
• Ecology: The realized I/D measure captures the system's indirect flow reliance, robust to orientation under steady state—a key ecological network diagnostic (Borrett et al., 2011).
• Survival Analysis: Relative indirect effects describe temporal patterns in mediated risk reduction, facilitating dynamic path analyses for time-dependent exposures and mediators (Aalen et al., 2020).
• Interference: RelIE as prevalence of spillover allows quantification without strong structural modeling assumptions, supporting robust inference in randomized trials with potential spillover or network effects (Choi, 2023).

5. Assumptions, Limitations, and Extensions

Identification of RelIE relies on variant-specific assumptions:

• No unmeasured confounding of treatment$M$5mediator and mediator$M$6outcome (e.g., sequential ignorability in mediation analysis (Hong, 3 Jun 2025)).
• Positivity: Sufficient overlap in propensity distributions.
• Consistency/SUTVA: Each unit's observed outcome matches its counterfactual under its actual treatment/mediator exposure.
• Ecological steady-state: For realized I/D invariance (Borrett et al., 2011).
• No structural modeling: In prevalence-based RelIE for spillovers, only randomization is required (Choi, 2023).

Front-door and interventional definitions permit partial relaxation of no-unmeasured-confounding when graphical criteria are satisfied (Peña, 2023, Aalen et al., 2020). In survival and epidemic models, analytic derivations facilitate parametric and semi-parametric estimation. For count data in networks or interference-prone experiments, RelIE is not fully identified but conservative bounds are attainable.

Bootstrapping and sandwich estimators address variance estimation. Practical recommendations include model flexibility for propensity estimation, diagnostic checking for positivity/overlap, and sensitivity analysis across weighting strategies (Hong, 3 Jun 2025).

6. Comparative Overview and Cross-Disciplinary Synthesis

RelIE serves as a unifying metric across causal inference, network science, and dynamical systems. While counterfactual mediation provides foundational definitions, the concept is generalized in ecology by the I/D ratio and in epidemiology via infection averted ratios. Inference under interference aligns RelIE with population prevalence, trading off identification for minimal assumption robustness. All forms partition effect components such that indirect pathways are normalized or contextualized relative to an appropriate system-wide or direct-response baseline.

Field	Relative Indirect Effect Definition	Application Context
Causal mediation	$M$7, $M$8	Psychosocial, biomedical, public health
Ecology (ENA)	Realized I/D: total indirect/total direct flow	Ecosystem material/energy eddies
Epidemiology	$M$9	Vaccine herd effects
Survival analysis	Ratio of counterfactual survival probabilities	Dynamic mediation, chronic risk
Interference	Prevalence of outcome shifts due to others’ treatment	Randomized field trials, networks

7. Software and Implementation Resources

Automated tools exist for computation and inference of RelIE:

• RMPW (Stata, R): Implements nonparametric, robust estimation for treatment–mediator–outcome frameworks including weighted regression and output of confidence intervals for RelIE, TE, NIE, NDE (Hong, 3 Jun 2025).
• Convex programming: For lower-bound estimation under interference (Choi, 2023).
• Custom statistical routines: For realized I/D computation in ecological networks (Borrett et al., 2011).

Robust mediation analysis leverages flexible models for propensities and mediators, with sensitivity checks for positivity and inferential stability via bootstrap. In epidemiological modeling, analytical and numerical solutions to dynamical system equations underpin RelIE estimation.

---

The study of relative indirect effects encompasses a spectrum of methodologies unified by their normalization of indirect effects against direct or total causal baselines, enabling rigorous, interpretable partitioning of causal mechanisms across diverse scientific disciplines.