Debiased machine learning for ultra-high dimensional mediation analysis.

Abstract

In ultra-high dimensional mediation analysis, confounding variables can influence both mediators and outcomes through complex functional forms. While machine learning (ML) approaches are effective at modeling such complex relationships, they can introduce bias when estimating mediation effects. In this article, we propose a debiased ML framework that mitigates this bias, enabling accurate identification of key mediators and precise estimation and inference of their respective contributions. We construct an orthogonalized score function and use cross-fitting to reduce bias introduced by ML. To tackle ultra-high dimensional potential mediators, we implement screening and regularization techniques for variable selection and effect estimation. For statistical inference of the mediators' contributions, we use an adjusted Sobel-type test. Simulation results demonstrate the superior performance of the proposed method in handling complex confounding. Applying this method to Alzheimer's Disease Neuroimaging Initiative data, we identify several cytosine-phosphate-guanine sites where DNA methylation mediates the effect of body mass index on Alzheimer's Disease. The R function DML_HDMA implementing the proposed methods is available online at https://github.com/Wei-Kecheng/DML_HDMA.