Causal Mediation Analysis with High-Dimensional Surrogates: Identification, Estimation and Inference under Latent Factor Models
Abstract
Causal mediation analysis is an important tool for understanding how a treatment affects an outcome through intermediate variables referred to as mediators. However, in many applications, these mediators and pre-treatment covariates are not directly observed and are instead measured through noisy, high-dimensional surrogates. Mediation analysis when such latent variables are observed only through these surrogates remains less explored. In this paper, we propose Factor-Adjusted Regression for Mediation Analysis (FARMA), a factor-based framework for estimating direct and indirect effects when either one or both of the mediators and relevant pre-treatment covariates are latent. FARMA uses the surrogate measurements to recover information about the underlying latent variables. Under a latent factor structure, we show that the direct and indirect effects are approximately identified, with an approximation error that vanishes as the surrogate dimensions grow. We further show that our estimator is √n-consistent and asymptotically normal. We combine the theory with extensive simulation studies and apply it to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, where we investigate whether the effect of depressive symptoms on cognitive impairment is mediated through the latent process underlying high-dimensional DNA methylation measurements.
Related
- Global Parameter Identification and Uncertainty Quantification of reduced-order ADM1 using Maximum Likelihood Estimation and Bayesian Inference
- Artificial intelligence surrogates for treatment effect estimation with before-and-after data
- Efficient Deep Gaussian Process Surrogates via Mini-Batch Latent MAP Estimation
- Learning on-the-fly quasi-stationarity diagnostics from low-dimensional surrogates
- A Unified Three-Stage Weighting Framework for Causal Inference and Mediation Analysis under Case-Control Sampling
BibTeX
@misc{yadav2026causal,
title = {{Causal Mediation Analysis with High-Dimensional Surrogates: Identification, Estimation and Inference under Latent Factor Models}},
author = {Himani Yadav and Shi Bo and AmirEmad Ghassami and Debarghya Mukherjee},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.03167},
eprint = {2611.03167},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.03167},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Causal Mediation Analysis with High-Dimensional Surrogates: Identification, Estimation and Inference under Latent Factor Models
AU - Himani Yadav
AU - Shi Bo
AU - AmirEmad Ghassami
AU - Debarghya Mukherjee
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.03167
UR - https://arxiv.org/abs/2611.03167
AB - Causal mediation analysis is an important tool for understanding how a treatment affects an outcome through intermediate variables referred to as mediators. However, in many applications, these mediators and pre-treatment covariates are not directly observed and are instead measured through noisy, high-dimensional surrogates. Mediation analysis when such latent variables are observed only through these surrogates remains less explored. In this paper, we propose Factor-Adjusted Regression for Mediation Analysis (FARMA), a factor-based framework for estimating direct and indirect effects when either one or both of the mediators and relevant pre-treatment covariates are latent. FARMA uses the surrogate measurements to recover information about the underlying latent variables. Under a latent factor structure, we show that the direct and indirect effects are approximately identified, with an approximation error that vanishes as the surrogate dimensions grow. We further show that our estimator is √n-consistent and asymptotically normal. We combine the theory with extensive simulation studies and apply it to data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) study, where we investigate whether the effect of depressive symptoms on cognitive impairment is mediated through the latent process underlying high-dimensional DNA methylation measurements.
ER -