报告题目：Recent Developments of Conditionally Specified Distributions
报告人：Yuchung J. Wang
报告人简介：Prof. Yuchung J. Wang got his PhD degree in Rutgers University. He is a Professor in Statistics, Department of Mathematical Sciences, Rutgers University, Camden, New Jersey. His research fields include Categorical data analysis, multivariate dependence, conditionally specified distributions, social networks, statistical computation algorithms, machine learning, and quality engineering.
报告摘要：One modeling strategy that has received more and more attentions recently is the conditionally specified distribution (CSD), which uses a collection of conditional models/densities to identify a d-dimensional joint distribution. Bayesian network (BN) is a special kind of CSD because its flow of causality is forward always. We will discuss the general CSD, also known as dependence network (DN), in which the flows of reasoning form a feedback loop. DN uses pseudo-Gibbs sampler to approximate the joint distribution. When the conditionals are not derived from one joint distribution, the resulting conditional densities and models are said to be incompatible. The Gibbs sampler using incompatible conditionals is called a pseudo-Gibbs sampler. Kou and Wang (2019, ASIM) studied the stationary distributions of pseudo-Gibbs sampler, known as the pseudo-Gibbs distributions (PGD). Multiple imputation via chained equations (MICE) performs multiple imputations using PGD, because it is next to impossible to be certain that the regression models used for imputation are compatible. In the past ten years, I published several paper on CDS. I will concentrate on three recent researches: (a) the shared marginal distributions among d! PGD; (b) among d! scan sequences, which visiting schedules will lead to the correct stationary distribution (Kou and Wang, 2018, JMA); (c) efficient computations of exact PGD and verification of compatibility among the conditional models (Kou and Wang, 2019, submitted). In this talk, we show how our study of incompatible CSD has led to better understanding of the original Gibbs sampling using compatible conditionals.