A1737
Title: MediEncoder: Nonlinear factor learning via autoencoders for high-dimensional causal mediation analysis
Authors: Shi Bo - Boston University (United States)
AmirEmad Ghassami - Boston University (United States)
Debarghya Mukherjee - Boston University (United States) [presenting]
Abstract: Causal mediation analysis aims to understand the mechanisms through which a treatment affects an outcome by decomposing the total causal effect into indirect and direct pathways. Modern applications increasingly involve ultra-high-dimensional covariates and mediators, where observed variables are often driven by a smaller number of latent biological processes. Existing approaches typically rely on assumptions such as sparsity or linear factor models for dimension reduction, which can be restrictive when the relationship between latent factors and observed variables is nonlinear. To address this limitation, MediEncoder is proposed as a representation learning framework for nonlinear ultra-high-dimensional mediation analysis. The method jointly learns latent representations of covariates and mediators using an encoder-decoder architecture that captures their structural dependence, enabling flexible nonlinear factor modeling while preserving causal pathways for estimating mediation effects. Based on the learned latent representation, an estimator of the direct and indirect causal effects is developed using an efficient influence function that is multiply robust and semiparametrically efficient, without strong structural assumptions. Numerical analyses show that MediEncoder improves accuracy compared with existing dimensionality reduction approaches. Finally, its utility is demonstrated on a high-dimensional biomedical dataset.
