The lay abstract featured today (for Heterogeneous Mediation Analysis for Cox Proportional Hazards Model With Multiple Mediators by Rongqian Sun, Xinyuan Song) is from Statistics in Medicine with the full Open Access article now available to read here.
How to cite
Sun, R. and Song, X. (2024), Heterogeneous Mediation Analysis for Cox Proportional Hazards Model With Multiple Mediators. Statistics in Medicine. https://doi.org/10.1002/sim.10239
Lay Abstract
In epidemiological and sociopsychological studies, researchers often aim to evaluate the effect of certain treatment or intervention on patients’ survival. Mediation analysis is a common technique used to explore the underlying mechanisms by which the treatment affects the survival outcome. This is done by looking at potential intermediate variables that link the treatment to the survival outcome and explain parts of the overall effect.
Existing mediation methods have provided well-developed quantification of the direct (not through a mediator) and indirect effect (through a mediator) at population level. However, it is increasingly recognised that such effects can substantially differ across individuals based on their specific characteristics and backgrounds. Recent advances in causal survival analysis enable identification and estimation of these “heterogeneous” effects across subgroups. But a major limitation is that they often focus on the total effect and overlook potential differences in how the effects are transmitted through underlying causal pathways. Additionally, the growing accessibility of large-scale datasets from clinical trials and observational studies introduces a multitude of pre-treatment and post-treatment covariates. These covariates can act as potential confounders or mediators that may explain the sources of heterogeneity across multiple causal pathways from the treatment to the survival outcome.
To address these challenges, this paper proposes a joint modelling approach to flexibly capture heterogeneity in direct and indirect pathways. It uses advanced Bayesian tree ensembles to identify the most relevant mediators and estimate the corresponding mediation effect on individual level. This method employs shared tree topologies with sparsity-inducing priors to account for overlaps in real moderators and confounders across multiple pathways. No restricting assumptions are made on the regression surfaces, and relevant variables are selected in a data-driven manner without requiring manual specification. The empirical performance in effect estimation and variable selection are verified through numerical studies. The proposed method is also applied to the ACTG175 study to demonstrate its practical utility.
In plain terms, this paper provides a powerful way to gain deeper insights into how and why treatment effect can be transmitted differently through multiple pathways across individuals based on their unique profiles. The authors also provide R code of the proposed method to facilitate similar analyses in real-world applications across domains.
More Details