Identification, Estimation, and Inference for Sequential Causally Ordered Mediation Pathways

Technical Summary: Identification, Estimation, and Inference for Sequential Causally Ordered Mediation Pathways

Problem Statement
Mediation analysis is essential for understanding how exposures influence outcomes through intermediate pathways. While methods for single-mediator settings are well-established, rigorous tools for handling multiple, sequentially ordered mediators (SCOMs) remain underdeveloped. This gap is critical in longitudinal cohort studies (e.g., developmental origins of health and disease) where exposures operate through complex, time-dependent chains of mediators. Existing literature has largely focused on estimation in SCOM settings but lacks robust hypothesis testing frameworks. Standard tests (e.g., Sobel's test, MaxP) for mediation effects suffer from overly conservative Type I error control due to the composite nature of the null hypothesis (where the product of parameters is zero via different configurations: $\alpha=0, \eta\neq0$; $\alpha\neq0, \eta=0$; or $\alpha=0, \eta=0$). These tests fail to account for the non-unique limiting distributions of test statistics under the null, leading to a lack of power or invalid inference.

Methodology
The authors propose a comprehensive framework for Sequential Causally Ordered Mediators of order $K$ (SCOM(K)), encompassing identification, estimation, and inference for both continuous and categorical outcomes.

1. Identification Framework:

   • The authors define a Generalized Structural Equation Model (G-SEM) where mediators $M_1, \dots, M_K$ follow a strict causal ordering ($M_1 \to M_2 \to \dots \to M_K$).
   • Using the potential outcomes paradigm, they decompose the total effect into path-specific effects. Due to identifiability constraints (specifically the "cross-world" independence assumption required for certain individual paths), the authors focus on composite pathway effects.
   • For each mediator $M_k$, they define a composite effect $\tau^\dagger_k(S \to M_k \rightsquigarrow Y)$, which aggregates all causal routes from exposure $S$ to outcome $Y$ that pass through $M_k$ (regardless of subsequent mediators).
   • Under standard sequential ignorability and consistency assumptions, they derive identification formulas for these composite effects using nested integrals over conditional distributions.

2. Estimation:

   • The paper develops Maximum Likelihood Estimation (MLE) procedures for linear models (continuous outcomes), probit models (binary outcomes), and log-linear models (count outcomes).
   • For binary outcomes, closed-form expressions are derived for probit models; for logistic models, a rare outcome approximation is utilized to obtain closed forms.
   • The authors establish the joint asymptotic normality of the estimators for the path coefficients ($\alpha_k$) and the composite path effects ($\eta_{M_k \rightsquigarrow Y}$), proving that the estimators for these two components are asymptotically independent.

3. Inference (SOMET):

   • To address the composite null hypothesis $H_0: \alpha_k \eta_{M_k \rightsquigarrow Y} = 0$, the authors introduce the Sequentially Ordered Mediation Effect Test (SOMET).
   • Mechanism: SOMET utilizes a $Q$-fold data-splitting strategy. The data is partitioned into $Q$ disjoint subsets. A Sobel-type statistic is computed for each subset. These statistics are averaged and transformed into a Student's $t$-statistic.
   • Theoretical Advantage: By using data splitting, the method disconnects the reliance on non-unique asymptotic variances. Theorem 3 establishes that under the composite null, the test statistic converges to a Student's $t$-distribution with $Q-1$ degrees of freedom, regardless of the specific configuration of the null (i.e., whether $\alpha_k=0$, $\eta=0$, or both). This ensures valid Type I error control without prior knowledge of the underlying null configuration.
   • Power Enhancement: To mitigate power loss from small sample sizes in splits, the authors propose repeating the splitting procedure $B$ times and aggregating $p$-values using a Cauchy combination test, which remains valid under arbitrary dependence.

Key Results

• Simulation Studies: Extensive simulations across linear, logistic, and probit models demonstrate that SOMET maintains valid Type I error control across all null configurations ($H_{01}, H_{02}, H_{03}$), whereas existing benchmarks (Sobel, MaxP, and their bootstrap variants) are either overly conservative or fail to control error rates under specific null scenarios (particularly when both parameters are zero).
• Statistical Power: SOMET consistently exhibits higher statistical power compared to benchmark methods in detecting composite mediation effects, particularly in scenarios where traditional tests are conservative.
• Empirical Applications:
  • Diabetes Risk (All of Us Data): Analyzing the effect of family diabetes history on diabetes onset mediated by BMI and physical activity. SOMET identified significant mediation effects for both BMI and physical activity with smaller $p$-values than competing methods.
  • Adolescent Obesity (ELEMENT Data): Assessing the impact of prenatal lead exposure on adolescent BMI mediated by birth weight and lipid metabolites. SOMET was the only method to detect a statistically significant mediation pathway through a specific lipid metabolite (FA 18:0 OH) after multiple testing correction, whereas other methods failed to reject the null.

Significance and Claims
The paper claims to provide the first rigorous framework that simultaneously handles identification, estimation, and valid inference for sequentially ordered mediators in diverse outcome settings. The primary contribution is the SOMET procedure, which resolves the long-standing technical challenge of hypothesis testing under a composite null in mediation analysis. By utilizing data splitting to achieve a unique limiting distribution, the method avoids the conservativeness of existing tests without requiring tuning parameters or knowledge of the true null configuration. The authors assert that their methodology offers reliable estimation, valid inference, and improved power for discovering novel mediation pathways in complex longitudinal systems. They acknowledge that while SOMET increases computational cost, it provides a principled foundation for scalable mediation analysis in structured causal pathways.