Statistical methods for analysis of high-dimensional mediation pathways

Abstract This proposal harnesses statistical theory and applications underlying mechanistic models to study mediation pathways involving high-dimensional omics markers on the children growth and development. This proposal aims to advance novel methodology, algorithms, and software to improve the understanding of mechanistic effects of environmental perturbations and socioeconomic stressors on biological processes related to children’s health outcomes such as adolescent obesity, cognitive function, and sexual maturation. This project is the first to systematically study the foundation of an emerging best-subset statistical estimation and inference in high-dimensional structural equation models (SEMs), and the resulting analytic toolboxes allow practitioners to simultaneously cluster, estimate, and validate key mediation pathways of clinical importance. (i) We develop a new analytic paradigm that can jointly process a large number of mediators (e.g. metabolites or DNA methylation CpG sites) to unveil mechanistic mediation pathways with well-controlled false discovery rate. The methodology innovation lies in a simultaneous operation of high-dimensional pathway clustering, parameter estimation and inference in the high-dimensional SEMs with little estimation bias and no compromise on false discovery. (ii) We develop an adaptive hypothesis testing methodology in high-dimensional SEMs to perform statistical inference for mediation pathways with a proper type I error control. This new method is deemed for significant power improvement over existing methods. (iii) We investigate mediation effects of the maternal blood lipidome and DNA methylation markers for the relationship of gestational environmental and socioeconomic exposures on children’s health outcomes. Moreover, discovered mechanistic mediation pathways will help develop potential interventions for better children’s health. (iv) We develop, test, distribute, and support freely available implementations of the proposed methods in this proposal. The developed statistical toolboxes can facilitate the translational clinical studies.