报告题目：Bayesian Regularization Methods For (Unspecified) Loadings in Partially Confirmatory Factor Analysis

报告时间：2024年8月4日下午13：00

报告地点：南湖校区老图书馆四楼左侧研究生5-1学习室

主办单位：数学与统计学院

报告人：付志慧

报告人简介：付志慧，博士，闽南师范大学数学与统计学院硕士生导师，校龙江学者特聘教授，入选福建漳州市高层次 D类人才、沈阳市第一批高层次人才-拔尖人才，美国伊利诺伊大学香槟分校访问学者。主要研究方向为贝叶斯统计、教育统计与心理测量等。主持完成国家自然科学基金项目、国家社会科学基金项目、辽宁省自然科学基金和福建省自然科学基金项目、全国统计科学研究项目、辽宁省教育厅项目等课题；获第十一届全国统计科学研究优秀成果二等奖、辽宁省自然科学技术成果三等奖；科学出版社独立出版专著1部，在《British Journal of Mathematical and Statistical Psychology》、《Physica A: Statistical Mechanics and its Application》、《Multivariate Behavioral Research》等期刊上发表20余篇学术论文。兼全国工业统计学教学研究会青年统计学家协会常务理事、中国教育学会教育统计与测量分会理事、福建省统计学会常务理事等。

摘要：The application of regularization methods to factor analysis models is increasingly gaining popularity. This research proposes more valid Gibbs sampling algorithms based on the Partially Confirmatory Factor Analysis (PCFA) framework for unspecified loadings. Four Bayesian regularization methods were employed and compared, including Lasso, Spike-and-Slab prior (SSP), Horseshoe, and Horseshoe+. Simulation study results indicate that compared to the original Lasso in PCFA, the other three priors demonstrate better validity and robustness. In particular, SSP shows better parameter recovery performance even under extreme conditions. Moreover, all three priors showed better identification of factor correlation, Horseshoe+ and SSP showed better robustness at the exploratory and confirmatory step, respectively. Finally, results from real-data demonstrate that the Horseshoe provides the most parsimonious factor structure. The four different priors also offer developers more flexible options in empirical research.