Causal Inference Methods for Microbiome Data

This dissertation introduces two complementary methodological frameworks for advancing causal mediation analysis in microbiome research. Microbiome data are characterized by high dimensionality, sparsity, overdispersion, and strong interdependencies, making conventional mediation approaches ill-suited for reliable inference. Existing methods often fail to deliver valid interval estimates or to capture longitudinal dynamics, especially when indirect effects are modest or when correlated mediator pathways are present. Our proposed framework addresses these challenges, providing tools that are broadly applicable across genomics, epidemiology, and other fields that require multiple mediator causal analysis. The first framework develops a fiducial inference approach for baseline microbiome mediation. By integrating mixed-effects zero-inflated generalized linear models (MEZIGLM) with natural effects models, this method accounts for overdispersion, zero inflation, and subject-level heterogeneity while constructing generalized confidence intervals for natural direct and indirect effects. A second-order fiducial construction combines bootstrap-based covariance estimation with Monte Carlo sampling from asymptotic distributions, yielding robust interval estimates that outperform delta-method and bootstrap alternatives in finite samples. The second framework extends mediation analysis to longitudinal microbiome-outcome processes, introducing joint models that accommodate time-varying exposures, sequential mediators, and dynamic outcomes. Factor-analytic random effects reduce the burden of high-dimensional correlations, ensuring unbiased estimation of direct and indirect effects. Simulations demonstrate that the proposed approach achieves stable coverage, improved sensitivity, and scalability in longitudinal contexts. Applications to data from the Cups and Community Health (CaCHe) study in Siaya County, Kenya, highlight the real-world utility of these methods. At baseline, the fiducial framework identified taxa such as Lactobacillus crispatus, Atopobium vaginae, and Sneathia sanguinegens as key mediators linking sexual behavior to bacterial vaginosis (BV). Longitudinally, the joint mediation models captured hormonal and microbial pathways shaping community state type (CST) transitions, providing biologically interpretable insights into BV risk. In summary, this dissertation makes two important contributions: (i) fiducial inference for valid interval estimation in baseline mediation with microbiome data, and (ii) longitudinal joint mediation models for dynamic microbial processes. Together, these frameworks fill critical methodological gaps, enhancing the rigor and interpretability of causal mediation in complex, high-dimensional biomedical data.

Author dc:creator

• Debarghya Nandi (12997422)

• Biology
• Biostatistics

Statement dc:rights

• In Copyright
• Open Access after 2028-01-01

DOI dc:identifier

https://doi.org/10.25417/uic.31451794.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451794