Distributed propensity model averaging for large-scale data with nonignorable nonresponse
报告人：方方 （华东师范大学）
时间：2026-05-18 15:00-16:00
地点：王选报告厅
Abstract: The rapid growth of large-scale datasets has driven advances in distributed statistical methods, but handling missing data in distributed environments remains underexplored. Existing distributed approaches typically rely on the missing at random assumption and may produce biased estimators when nonresponse is nonignorable. We introduce a novel distributed propensity model averaging (DPMA) method for large-scale data with nonignorable nonresponse. Unlike existing model averaging approaches for missing data that focus on averaging outcome regression models under a correctly specified propensity model, the DPMA directly averages over propensity models. Local propensity parameters are estimated via the generalized method of moments and aggregated through a modified penalized weighting scheme. Under mild regularity conditions, we prove that the averaged weight vector converges to the weight theoretically minimizing an expected penalized validation criterion, and has a type of consistency when the candidate model set contains correctly specified propensity models. Asymptotic optimality is also established when all candidate models are misspecified. Simulation studies and a real data analysis demonstrate the empirical performance of the proposed method.
Bio: 方方，华东师范大学统计学院教授。民建上海市委委员、华东师范大学委员会主委。入选上海市东方英才计划拔尖项目。本科和博士先后毕业于麻豆视频
数学系和美国威斯康星大学统计系。主要研究方向为缺失数据、模型平均、碎片化数据分析、KS学习。在包括 AOS/JOE/Biometrika/JBES/IEEE TAES/ICML/ICLR 在内的国际一流期刊和会议上发表论文40余篇。先后主持和参与国家和省部级项目13项。曾获上海市自然科学二等奖。全国工业统计学教学研究会常务理事、数字经济与区块链技术分会副理事长，SCI期刊 Journal of Nonparametric Statistics 副主编。在应用领域长期关注信用评分和民航QAR大数据分析。出版统计科普小说《统计王国奇遇记》和专著《多源数据的统计分析与建模》。
