CMStatistics 2022: Start Registration
View Submission - CMStatistics
Title: Feature identification on high-dimensional mediators using causal mediation tree model Authors:  Yao Li - University of Tortonto (Canada) [presenting]
Abstract: High-dimensional mediation analysis plays an important role in recent biomedical research as a large number of mediators, such as microbiomes, could modulate the effect of exposure to the outcome of interest. Most of the current studies focus on modeling independent mediators. However, these methods do not consider the correlation between the mediators and their non-linear interactive effect. On the other hand, identifying the mediators with significant effects from the high-dimensional mediator space is challenging. We propose an innovative non-parametric approach to build a causal mediation tree to select important mediators and assess their nonlinear effects. The data are recursively partitioned into subpopulations constructed by the mediators with the largest mediation effect. We aim to incorporate this nonlinear relationship into the mediation framework and evaluate the total effect. Simulation studies were conducted to assess the performance of our algorithm under different scenarios of the interactive mediation effects. We applied the method to analyze vaginal microbiome data from the reproductive-age women study. We investigated the causal relationship between ethnic groups and the virginal PH levels mediated by the virginal microbiomes. We identified two important microbiome taxa with strong mediation effects and estimated the total effect of the mediation tree model.