General approach of causal mediation analysis for survival outcome under sequential mediators

Abstract

Causal multi-mediation analysis is critical in understanding why an intervention works, especially in medical research. Deriving the path-specific effects (PSEs) of exposure on the outcome through the mediators can detail the causal mechanism of interest. However, the current multi-mediator approaches with survival outcomes perform PSE derivation in the setting of fewer mediators via partial decomposition. Here, we propose a novel multi-mediator model for survival analysis to obtain the fully decomposed formula of PSEs in the case with an arbitrary number of mediators. There are two significant contributions in this study. First, we define the multi-mediation formulas by partial decomposition and interventional approach to obtain the partial PSE and interventional PSE in a general case. By assuming normally distributed mediators, these formulas perform the g-formula while mediators are weighted by a normally distributed variable. Second, we extend partial PSE and interventional PSE into the context of the survival analysis. Based on Aalen additive hazard models and Cox proportional hazard models, it derives the analytic forms for both PSEs. We further establish the asymptotic property for the partial PSEs and interventional PSEs. The simulation is conducted to evaluate the performance of estimation in several scenarios. Our proposed approach is applied to a Taiwanese cohort study for investigating the mechanism of liver cancer risk. The model is implemented in the R package “iPSEsv” and the algorithm is available for download.