A Bayesian Criterion for Re-randomization
Holder: Ke Deng(Department of Statistics and Data Science, Tsinghua University)
Time:2024-12-12 14:00-15:00
Location:Siyuan Lecture Hall,Zhi Hua Building-225
Abstract:
Imbalance of covariates between treatment groups is to be avoided even in randomized experiments, where this is achieved in expectation. Re-randomization based on Mahalanobis distance, referred to ReM, has been advocated as a method to achieve this. This basic method, as well as some extensions, allocates equal importance to each orthogonalized covariate. However, investigators often know a priori that some covariates are more important than the others for predicting the outcomes. Formulating such prior knowledge into a formal prior distribution and utilizing this to guide the design of re-randomization procedures can be used to establish a more efficient Bayesian framework for re-randomization. Theoretical analyses show that the re-randomization procedure induced by the Bayesian criterion, referred to as ReB, can enjoy attractive properties, and takes ReM, and many of its extensions, as special cases. A two-step procedure to implement ReB in practice is developed. The advantages of ReB are demonstrated via simulation and real data analysis.
About the Speaker:
邓柯是清华大学统计与数据科学系长聘副教授,主要从事贝叶斯统计方法和计算的研究,并致力于推动统计学与生物医学、人工智能、人文社科的交叉。2008年获得北京大学统计学博士学位,同年进入哈佛大学统计系从事研究工作,历任博士后、副研究员,2013年加入清华大学工作至今。2014年入选国家高层次人才计划青年项目,2016年获“科学中国人年度人物”荣誉称号,2018年受聘北京智源人工智能研究院担任数理基础方向的“智源研究员”,2024年当选“国际统计研究院”(International Statistical Institute,ISI) Elected Member。在统计学、数据科学知名期刊和会议发表论文五十余篇,主持多项科技部重点研发计划、国家自然科学基金、国家社科基金及北京市自然科学基金项目。在中文文本分析和数字人文方面的研究工作获得获“国际华人数学家大会”(ICCM)和“中国数字人文大会”最佳论文奖,在生物信息学方面的研究工作获得教育部“高校科学研究优秀成果奖”自然科学奖一等奖,在政务大数据分析方面的多项成果被政府采纳并应用。他历任国际计算统计学会亚太地区分会理事、中国现场统计研究会计算统计分会理事长、中国青年统计学家协会副会长、中国人工智能学会智慧医疗专业委员会副主任委员、国家抗肿瘤药物临床应用专家委员会委员,还担任国际统计学杂志 Statistica Sinica 副主编,以及《数字人文》、《应用概率统计》、《应用数学与力学》、《统计与精算》等期刊的编委。

Your participation is warmly welcomed!

欢迎扫码关注北大统计科学中心公众号,了解更多讲座信息!