Parameter-expanded data augmentation for analyzing multinomial probit models

Area of Science:

• Statistics
• Econometrics
• Computational Statistics

Background:

• Multinomial probit models are widely used for nominal categorical data analysis.
• Computational complexity and model identification challenges hinder their practical application, especially with maximum likelihood estimation and Markov chain Monte Carlo (MCMC) sampling.
• Existing methods often require restricted covariance matrices, complicating estimation and sampling.

Purpose of the Study:

• To address the computational and identification challenges in multinomial probit models.
• To develop novel parameter-expanded data augmentation methods for improved MCMC sampling.
• To enhance the convergence and mixing properties of MCMC algorithms for these models.

Main Methods:

• Constructing a non-identifiable multinomial probit model.
• Developing parameter-expanded data augmentation techniques.
• Utilizing a Gibbs sampler to sample an unrestricted covariance matrix, avoiding complex Metropolis-Hastings algorithms.

Main Results:

• The proposed methods significantly improve the convergence and mixing of MCMC components.
• The new approach circumvents the need to sample a restricted covariance matrix.
• Simulation studies and a consumer choice data application demonstrate the effectiveness of the proposed methods.

Conclusions:

• The developed methods offer a more computationally efficient and stable approach to analyzing nominal categorical data using multinomial probit models.
• These advancements facilitate broader application of multinomial probit models in statistical and econometric research.
• The improved MCMC sampling performance provides a practical advantage for complex data analysis.