Flexible bivariate correlated count data regression

Area of Science:

• Statistical modeling
• Biostatistics
• Genomics

Background:

• Multivariate count data are prevalent across various scientific fields.
• These datasets often display intricate positive or negative dependencies between variables.
• Accurate modeling is crucial for understanding complex biological relationships.

Purpose of the Study:

• To propose and evaluate three novel Bayesian methodologies for modeling bivariate count data.
• To simultaneously account for covariate-dependent means and correlation structures.
• To compare the performance of these approaches using simulation and real-world genomic data.

Main Methods:

• A direct Bayesian approach using a bivariate negative binomial distribution.
• An indirect Bayesian approach employing a bivariate Poisson-gamma mixture model.
• A bivariate Gaussian copula model for capturing dependency structures.

Main Results:

• Simulation analyses indicated that the indirect and copula approaches offered superior model fitting.
• These methods also excelled in identifying covariate-dependent associations.
• The direct approach showed comparatively weaker performance.

Conclusions:

• The indirect and copula Bayesian models are recommended for analyzing bivariate count data with complex dependencies.
• These methods provide robust tools for exploring covariate effects in high-dimensional biological data.
• Application to RNA-sequencing data in cancer genomics highlights their practical utility.