A marginalized two-part Beta regression model for microbiome compositional data.

Area of Science:

• Microbiology
• Statistical Modeling
• Bioinformatics

Background:

• Microbiome studies aim to detect differential microbial abundance across conditions.
• Microbiome data are compositional, exhibiting high skewness, [0, 1) bounds, and many zeros.
• Conventional two-part models lack marginal interpretation of covariate effects.

Purpose of the Study:

• To propose a marginalized two-part Beta regression model for microbiome data.
• To capture zero-inflation and skewness while enabling marginal covariate effect interpretation.
• To improve statistical power and Type I error control in microbiome analysis.

Main Methods:

• Developed a marginalized two-part Beta regression model.
• Utilized simulation studies to evaluate model performance.
• Applied the model to a real metagenomic dataset of mouse skin microbiota.

Main Results:

• The proposed model effectively handles zero-inflation and skewness in microbiome data.
• Covariate effects on the marginal mean of microbial abundance can be examined.
• The likelihood ratio test under the marginalized model shows improved Type I error control without power loss.

Conclusions:

• The marginalized two-part Beta regression model offers a robust approach for microbiome compositional data analysis.
• This method enhances the interpretability of covariate effects on microbial abundance.
• The findings suggest improved statistical rigor for microbiome research.