Bayesian Variable Selection for Gaussian copula regression models | JoVE Visualize

Area of Science:

Statistics
Computational Statistics
Biostatistics

Background:

Regression models are essential for analyzing relationships between variables.
Handling multiple response variables of different types (e.g., continuous, discrete) presents significant statistical challenges.
Existing methods often struggle to simultaneously model complex dependencies and select relevant predictors.

Purpose of the Study:

To develop a novel Bayesian statistical method for selecting important predictors in regression models with multiple, diverse response types.
To address the challenge of multivariate dependencies among various response variable types.
To provide an efficient computational strategy for model estimation and predictor selection.

Main Methods:

A sparse Gaussian copula regression model is employed to capture multivariate dependencies between predictors and diverse response types.
Parameter expansion for data augmentation is utilized to construct a Markov chain Monte Carlo (MCMC) algorithm.
A fixed-dimensional proposal distribution within a Metropolis-Hastings step facilitates efficient exploration of the predictor model space.

Main Results:

The proposed Bayesian method effectively selects important predictors in regression models with mixed-type responses.
The MCMC algorithm, incorporating parameter expansion and a novel proposal, efficiently estimates model parameters and latent variables.
The method demonstrates robust performance on simulated data and is successfully applied to real-world datasets with complex response structures.

Conclusions:

The developed Bayesian sparse Gaussian copula regression model offers a powerful and flexible framework for high-dimensional variable selection with multiple, heterogeneous response variables.
The computational strategy ensures efficient model fitting and predictor identification, enhancing the applicability of complex statistical models.
This approach advances the analysis of complex datasets common in fields like biostatistics and econometrics.