MedC-I Protocol for Clinical Mediation Analysis

• MedC-I is a statistical methodology that decomposes total effects into clinically actionable controlled direct and indirect effects for multiple mediators.
• It utilizes the potential outcomes framework with techniques like the g-formula and inverse-probability weighting to enable identification from observational data.
• MedC-I guides clinical decision-making by ranking manipulable mediators based on empirical effects, optimizing targeted interventions in health studies.

The MedC-I (Mediation via Controlled Indirect Effect) protocol is a statistical methodology for mediation analysis in clinical and epidemiological research, particularly targeting systems with multiple, potentially causally dependent manipulable mediators. The protocol addresses limitations of classical mediation frameworks by allowing the decomposition of total effects into clinically actionable controlled direct and indirect effects for each specific mediator, while remaining identifiable under standard observational data assumptions and eschewing reliance on cross-world counterfactuals. MedC-I particularly suits scenarios where the goal is to identify the optimal target among several manipulable mediators for improving outcomes in clinical interventions (Sun et al., 2020).

1. Notation, Setup, and Counterfactual Framework

MedC-I operates in the potential outcomes framework. Let $A\in\{0,1\}$ denote a binary exposure (such as treatment), $Y$ the outcome of interest, $\mathbf{M} = (M_1, ..., M_K)$ a vector of $K$ binary, manipulable mediators, and $L$ a vector of pre-exposure covariates (confounders).

Potential outcomes are denoted as follows:

• $M_k(a)$: value of mediator $k$ if $A$ is set to $a$.
• $Y(a, m_1, ..., m_K)$: outcome if $A$ and all mediators are set to specified values.
• $Y_k(a, m)$: shorthand for the outcome if $A$ is set to $a$, $M_k$ set to $m$, and other mediators to their "natural" values under $A=a$.

These variables are observed in $N$ i.i.d. samples $\{A^{(i)}, L^{(i)}, M^{(i)}, Y^{(i)}\}_{i=1}^N$.

2. Causal Identification Assumptions

The protocol requires the following assumptions for identification of effects from observational data:

1. Consistency: Each unit’s observed variables equal their potential outcomes under the realized exposure and mediator values.
2. Positivity: For all $a, m, l$ in support, $P(A=a|L=l)>0$ and $P(M_k=m|A=a,L=l)>0$.
3. Ignorability: No unmeasured confounding for (a) $A \perp M_k(a)|L$ and (b) $(A, M_k) \perp Y_k(a, m)|L$.

These correspond to assumptions underlying standard DAG-based mediation analysis, avoiding the interventional and cross-world independence assumptions needed for "natural" indirect effects (Sun et al., 2020).

3. Formal Effect Definitions and Decomposition

MedC-I distinguishes three estimands:

• Total Effect (TE): $TE = E[Y(1)] - E[Y(0)]$
• Controlled Direct Effect (CDE) at mediator value $m$: $CDE(m) = E[Y(1, m)] - E[Y(0, m)]$
• Controlled Indirect Effect (CIE) for mediator $M_k$, fixing exposure at $a'$: $CIE_k(m, m') = E[Y(a', m)] - E[Y(a', m')]$

For binary mediators, $CIE_k(a) = E[Y_k(a, 1)] - E[Y_k(a, 0)]$. MedC-I introduces the specific summary $$sCIE_k = M_k(1) \cdot CIE_k(1) - M_k(0) \cdot CIE_k(0)$$, where $M_k(a) = E[M_k(a)]$.

Decomposition for $K$ possibly dependent mediators ($k=1,...,K$):

$TE = CDE_k(0) + sCIE_k,\qquad \text{(for any $k$)}$

and in aggregate,

$$TE = \frac{1}{K} \sum_{k=1}^{K} CDE_k(0) + \frac{1}{K} \sum_{k=1}^{K} sCIE_k.$$

This decomposition allows prioritization of individual mediators for intervention, even in high-dimensional or non-independent settings.

4. Identification, Estimation, and Algorithm

Under the stated assumptions, all target parameters admit identification by either the extended $g$-formula or inverse-probability weighting (IPTW):

$g$-formula:

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

$E[M_k(a)] = \int E[M_k|A=a, L=\ell]\,dP(\ell)$

$$E[Y_k(a, m)] = \int E[Y|A=a, M_k=m, L=\ell]\,dP(\ell).$$

Weighting:

Let $w_A^{(i)} = \frac{\mathbbm{1}(A^{(i)}=a)}{P(A^{(i)}=a|L^{(i)})}$ and $w_{AM}^{(i)} = \frac{\mathbbm{1}(A^{(i)}=a, M_k^{(i)}=m)}{P(A^{(i)}=a|L^{(i)})P(M_k^{(i)}=m|A^{(i)}, L^{(i)})}$.

Estimation Algorithm:

1. Fit propensity, mediator, and outcome models to estimate $P(A|L)$, $P(M_k|A,L)$, and $E[Y|A, M_k, L]$.
2. Calculate predicted mediator and outcome values under all relevant interventions for each subject.
3. Estimate $CDE_k(0), CIE_k(a), sCIE_k$ and $TE$ by averaging across the sample.
4. Use bootstrap or influence-function resampling for confidence intervals.

Doubly robust approaches are enabled by combining regression and weighting, and model selection may use penalized regressions or ensemble machine learning (Sun et al., 2020).

5. Clinical and Empirical Applications

MedC-I is structured for direct clinical interpretability. It prescribes:

• Selection of manipulable, binary mediators (e.g., treatable co-morbidities).
• Covariate collection to bolster ignorability and ensure positivity.
• Use of sCIE$_k$ to rank mediators by expected outcome improvement from "treating" $M_k$.

Empirical demonstrations from (Sun et al., 2020):

• In the "framing" effects dataset with two independent mediators, sCIE for emotional response exceeded that for harm perception, indicating targeted intervention design.
• In the HIV-Brain Age cohort (three dependent comorbidities), hyperlipidemia yielded the largest (≈86% of TE) controlled indirect effect, guiding prioritization of preventative measures.

6. Methodological Advantages and Considerations

MedC-I avoids key pitfalls of classical or "natural" mediation frameworks:

• No cross-world counterfactuals: Only requires counterfactuals in single "worlds," thus avoiding non-identifiable quantities.
• Dimensionality reduction: Only marginalizes or conditions on one mediator at a time—even in high-dimensional settings.
• Robust to arbitrary mediator dependencies: No sequential ignorability required for mediator chains or DAGs.

Practical estimation issues relate to small sample instability, unmeasured confounding, and measurement error. The method supports bootstrap inference and doubly robust estimation. Sample size, rich covariate coverage, and appropriate parametric or nonparametric modeling are essential.

7. Recommendations, Limitations, and Future Directions

Best practices include careful covariate selection, relevant positivity checks (possibly via trimming), and use of robust ML-based estimators.

Pitfalls encountered are unmeasured confounding (invalidating identifiability), small sample bias, and measurement error. The fundamental limitation is that only observed variables and ignorability-valid counterfactuals can be evaluated; violations of these induce bias.

Prospective extensions include:

• Adaptations for continuous or multi-level exposures/mediators.
• Generalization to path-specific controlled effects in longitudinal settings.
• Sensitivity analysis for unmeasured confounding.
• Experimental designs (e.g., parallel encouragement) to empirically validate controlled effects (Sun et al., 2020).

MedC-I provides a pathway for scientifically and clinically grounded mediation analysis—directing single-mediator interventions, avoiding intractable high-dimensional modeling, and remaining valid under standard identification criteria.