Inverse probability weighting-based mediation analysis for microbiome data

Mediation analysis is an important tool for elucidating causal mechanisms in biomedical research and has recently gained increasing attention in microbiome studies. Motivated by a microbiome study of acute myeloid leukemia (AML) patients, we investigate whether the effect of induction chemotherapy intensity on infection status is mediated through gut microbial taxa. This problem poses substantial methodological challenges due to the high dimensionality, zero inflation, and dependence of microbial mediators, as well as the presence of an exposure-induced mediator-outcome confounder, antibiotic use.

To address these challenges, we propose a novel nonparametric identification formula for the interventional indirect effect, a causal mediation estimand that remains identifiable in the presence of exposure-induced mediator-outcome confounding. Building on this identification result, we develop an inverse probability weighting estimator and construct standard normal bootstrap confidence intervals for inference. Simulation studies demonstrate that the proposed method achieves accurate effect estimation and maintains appropriate type I error rates and power in finite samples. Application to the AML microbiome study provides evidence of a statistically significant interventional indirect effect of induction chemotherapy intensity on infection risk through alterations in the gut microbiome.

The R code for the simulation studies is available at https://github.com/YX-ZHANG-1/CausalMediationMicrobiome and archived on Zenodo at https://doi.org/10.5281/zenodo.21960802.