Approximate Predictive Densities and Their Applications in Generalized Linear Models.

Area of Science:

• Statistics
• Computational Statistics
• Bayesian Inference

Background:

• Exact calculation of model posterior probabilities is often analytically intractable.
• Predictive densities are crucial for Bayesian inference but difficult to compute.
• Existing methods like the Laplace approximation have limitations.

Purpose of the Study:

• To propose novel approximations for predictive densities.
• To contrast these new methods with the Laplace approximation.
• To demonstrate their utility in Bayesian generalized linear models (GLMs) and variable selection.

Main Methods:

• Development of new approximation techniques for predictive densities.
• Comparison of proposed methods with the Laplace approximation through theoretical analysis.
• Numerical studies to evaluate accuracy and efficiency across various hyperparameters.
• Application within GLMs with informative priors on regression coefficients.

Main Results:

• The proposed approximation methods are easy to implement and computationally efficient.
• They achieve high accuracy over a broad range of hyperparameters.
• The methods facilitate posterior computation in GLMs under informative priors.
• Demonstrated feasibility and usefulness in a real-world Bayesian variable selection scenario.

Conclusions:

• The novel approximations offer a practical and efficient alternative for computing predictive densities.
• These methods enhance the feasibility of Bayesian variable selection procedures, particularly in GLMs.
• The approach provides a valuable tool for researchers dealing with analytically intractable models.