MZINBVA

MZINBVA implements multilevel zero-inflated negative binomial variational approximation models to estimate parameters and test associations in hierarchical microbiome count data.

Key Features:

• Multilevel Modeling: Employs multilevel zero-inflated negative-binomial models that accommodate hierarchical structures with repeated observations across time points (level 1), nested within body sites (level 2), and nested within subjects (level 3).
• Zero-Inflation Handling: Incorporates zero-inflated components to model excess zeros commonly observed in microbiome count data.
• Overdispersion Management: Models overdispersed count data arising from biological variability and measurement error.
• Variational Approximation Methodology: Uses a variational approximation approach for maximum likelihood estimation and inference, leveraging optimization techniques rather than sampling methods to approximate the log-likelihood.
• Robust Covariance Estimation: Provides robust estimates of the covariance of parameter estimates to support reliable inference.
• Wald-Type Test Statistic: Constructs a Wald-type test statistic for association testing between microbial taxa and environmental factors or clinical outcomes.

Scientific Applications:

• Association analysis in microbiome surveys: Identifying associations between microbial taxa and environmental factors or clinical outcomes.
• Analysis of hierarchical study designs: Analyzing longitudinal and nested body-site/subject structures in microbiome datasets.

Methodology:

Applies variational approximation for maximum likelihood estimation and inference using optimization rather than sampling to approximate the log-likelihood; evaluated via extensive simulation studies and real-world application to the Human Microbiome Project (HMP) dataset.