Simulating conditionally specified models

JOURNAL OF MULTIVARIATE ANALYSIS

卷:167

Article

Kuo, Kun-Lin¹  Wang, Yuchung J.² 

[1] Natl Univ Kaohsiung, Inst Stat, Kaohsiung, Taiwan

[2] Rutgers State Univ, Dept Math Sci, Camden, NJ 08102 USA

关键词: Dependence network;    Faster convergence;    Multiple imputation;    Non-full conditional specification;    Partially collapsed Gibbs sampler;    Valid scan order;   

DOI  :  10.1016/j.jmva.2018.04.012

来源: Elsevier

【 摘 要 】

Expert systems routinely use conditional reasoning. Conditionally specified statistical models offer several advantages over joint models; one is that Gibbs sampling can be used to generate realizations of the model. As a result, full conditional specification for multiple imputation is gaining popularity because it is flexible and computationally straightforward. However, it would be restrictive to require that every regression/classification must involve all of the variables. Feature selection often removes some variables from the set of predictors, thus making the regression local. A mixture of full and local conditionals is referred to as a partially collapsed Gibbs sampler, which often achieves faster convergence due to reduced conditioning. However, its implementation requires choosing a correct scan order. Using an invalid scan order will bring about an incorrect transition kernel, which leads to the wrong stationary distribution. We prove a necessary and sufficient condition for Gibbs sampling to correctly sample the joint distribution. We propose an algorithm that identifies all of the valid scan orders for a given conditional model. A forward search algorithm is discussed. Checking compatibility among conditionals of different localities is also discussed. (C) 2018 Elsevier Inc. All rights reserved.

【 授权许可】

Free