Regression Models For Multivariate Count Data

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Multivariate count data are common in modern applications, including RNA-sequencing.
• The standard multinomial-logit model has restrictive mean-variance assumptions, leading to errors with real-world data like RNA-seq.
• Over-dispersion and complex correlations in count data necessitate more flexible regression models.

Purpose of the Study:

• To introduce and analyze generalized linear models for multivariate count data with various correlation structures.
• To address the lack of flexible models in current literature, particularly for non-natural exponential family distributions.
• To provide a unifying framework for estimation, testing, and variable selection for these advanced models.

Main Methods:

• Development of generalized linear models accommodating diverse correlation structures for multivariate counts.
• Application of a unifying framework for statistical inference, including estimation, hypothesis testing, and variable selection.
• Comparative analysis using both synthetic datasets and real-world RNA-sequencing data.

Main Results:

• Demonstrated limitations of the multinomial-logit model in handling over-dispersion and correlation in RNA-seq data.
• Proposed generalized linear models show improved performance and flexibility for multivariate count data.
• The unifying framework effectively handles estimation, testing, and variable selection for the studied models.

Conclusions:

• Generalized linear models offer a superior alternative to the multinomial-logit model for analyzing complex multivariate count data.
• The developed methods are crucial for accurate analysis of RNA-sequencing and similar biological count data.
• This work provides essential tools for researchers dealing with high-dimensional count-based biological data.