Bayesian Inference for General Gaussian Graphical Models With Application to Multivariate Lattice Data

Area of Science:

• Statistics
• Computational Statistics
• Spatial Statistics

Background:

• Gaussian graphical models are crucial for understanding conditional independence relationships in multivariate data.
• Existing methods for inference and model determination can be computationally intensive, especially for complex structures.
• The G-Wishart prior offers a flexible framework for Bayesian analysis of precision matrices in Gaussian graphical models.

Purpose of the Study:

• To introduce efficient Markov chain Monte Carlo (MCMC) methods for inference and model determination in multivariate and matrix-variate Gaussian graphical models.
• To extend these sampling algorithms to a novel class of conditionally autoregressive models for sparse estimation in multivariate lattice data.
• To demonstrate the application of these methods in analyzing spatial data and real-world scenarios like disease surveillance.

Main Methods:

• Development of novel MCMC algorithms for Bayesian inference in Gaussian graphical models using the G-Wishart prior.
• Extension of sampling techniques to conditionally autoregressive models for sparse estimation in lattice-structured data.
• Application of the proposed methods to simulated data and a real-world cancer mortality surveillance dataset.

Main Results:

• Efficient MCMC methods were established for inference and model determination in multivariate and matrix-variate Gaussian graphical models.
• The framework accommodates both decomposable and non-decomposable graphs.
• The extended models demonstrated flexibility in estimating cross-outcome correlations and spatial structures, enabling adaptive smoothing.

Conclusions:

• The proposed MCMC methods provide efficient tools for analyzing complex Gaussian graphical models and related spatial data structures.
• The developed framework allows for flexible modeling of correlation structures, crucial for applications in fields like epidemiology and spatial statistics.
• The study offers practical computational solutions and demonstrates their utility through simulation and a real-world cancer surveillance example.