High dimensional mediation analysis via neural networks
Conference
64th ISI World Statistics Congress
Format: IPS Abstract
Session: IPS 483 - Recent Statistical Developments on High-Dimensional Causal Inference
Monday 17 July 10 a.m. - noon (Canada/Eastern)
Abstract
Mediation analysis draws increasing attention in many research areas such as epidemiology, genomics, psychology and social sciences, and it gets difficult when the dimension of potential mediators is larger than the sample size. In addition, the effect of potential confounders can be complicated on both mediators and the outcome, which is merely considered as linear in existing literature. In this paper, we propose a novel quantile mediation deep neural network (QMDNN) estimation and inference procedure for evaluating the indirect and direct effect in mediation models, where linear high-dimensional mediators and complex non-linear confounders are incorporated in both the mediator model and the outcome model. By applying penalization to partially linear regression model, the proposed procedure performs consistently in selecting active mediators and has strong type I error control of hypothesis testing on both effects. Compared to existing methods, the superiority of the proposed approach is demonstrated in various Monte Carlo simulations and one real data application of selecting DNA methylation Loci mediating childhood trauma and cortisol stress reactivity.