Multinomial Models with Linear Inequality Constraints: Overview and Improvements of Computational Methods for...

Area of Science:

• Psychological modeling
• Statistical inference
• Computational statistics

Background:

• Psychological theories often translate into linear inequality constraints on multinomial distribution parameters.
• These constraints have two representations: solution sets of linear inequalities or convex hulls of extremal points.
• Bayesian analysis of such models requires specialized sampling techniques.

Purpose of the Study:

• To develop a general Gibbs sampler for Bayesian inference in multinomial models with linear inequality constraints.
• To present methods for estimating Bayes factors using the encompassing Bayes factor approach.
• To introduce the R package multinomineq for accessible and efficient implementation.

Main Methods:

• A general Gibbs sampler is described for drawing posterior samples from models with linear inequality constraints.
• Alternative sampling methods for Bayes factor estimation are summarized, utilizing the encompassing Bayes factor method.
• The R package multinomineq is developed to provide a user-friendly interface for these computational techniques.

Main Results:

• The developed Gibbs sampler enables Bayesian analysis for a wide range of psychological theories operationalized as constraints.
• The methods facilitate the estimation of Bayes factors for model comparison in these constrained models.
• The multinomineq package offers an efficient and accessible computational tool for researchers.

Conclusions:

• The proposed Gibbs sampler and associated methods provide a robust framework for Bayesian analysis of constrained multinomial models.
• The multinomineq package democratizes the application of these advanced statistical techniques in psychological research.
• This work bridges theoretical psychological constraints with practical computational tools for data analysis.