Primary Submission Category: Mediation Analysis, Mechanisms
Validating Causal Mechanisms through Replicability: A Unified Bayesian Framework for Mediation Analysis
Authors: Ester Alongi, Gianmarco Altoè, Giovanni Parmigiani,
Presenting Author: Ester Alongi*
How can we ensure that an estimated mediation effect has a causal interpretation? Causal mediation analysis is a powerful tool for understanding the mechanisms transmitting an effect from an exposure to an outcome. However, estimating the natural indirect effect relies on the nonrefutable assumption of sequential ignorability, specifically the absence of unmeasured mediator-outcome confounding. While various sensitivity analyses have been proposed to assess the robustness of mediation estimates against hypothetical assumption violations, they do not directly address whether such confounding is present.
We propose a systematic replicability approach as an empirical tool to detect whether an estimated mediation effect reflects sample-specific unmeasured mediator-outcome confounding. We introduce a unified Bayesian hierarchical framework embedding causal mediation within a multifaceted replicability structure evaluating four dimensions: natural indirect effect consistency across independent studies; meta-analytic inference implied by a common generative model; consistency between studies and the shared meta-analytic structure; consistency between an existing body of evidence and a new study from the same generative model.
We apply this framework to a Mendelian randomization case study using GTEx data, evaluating whether a genetic variant’s effect on downstream metabolic genes in subcutaneous adipose tissue is causally mediated by the expression of the KLF14 transcription factor.