摘要：Model averaging is an important tool for treating uncertainty from model selection process and fusing information from different models, and has been widely used in various fields. However, the most existing model averaging criteria are proposed based on the methods of ordinary least squares or maximum likelihood, which possess high sensitivity to outliers or violation of certain model assumption. For the mean regression, no optimal robust methods are developed. To fill this gap, in our paper, we propose an outlier-robust model averaging approach by Mallows-type criterion. The idea is that we first construct a generalized M (GM) estimator for each candidate model, and then build robust weighting schemes by the asymptotic expansion of the final prediction error based on the GM-type loss function. So we can still achieve a trustworthy result even if the dataset is contaminated by outliers in response and/or covariates. Asymptotic properties of the proposed robust model averaging estimators are established under some regularity conditions. The consistency of our weight estimators tending to the theoretically optimal weight vectors is also derived. We prove that our model averaging estimator is robust in terms of having bounded influence function. Further, we define the empirical prediction influence function to evaluate the quantitative robustness of the model averaging estimator. A simulation study and a real data analysis are conducted to demonstrate the finite sample performance of our estimators and compare them with other commonly used model selection and averaging methods.

报告人简介：邹国华，首都师范大学特聘教授。博士毕业于中国科学院系统科学研究所，是国家杰出青年基金获得者、“新世纪百千万人才工程”国家级人选、中国科学院“百人计划”入选者、享受国务院政府特殊津贴，曾获中国科学院优秀研究生指导教师称号。

主要从事统计学的理论研究及其在经济金融、生物医学中的应用研究工作，在统计模型选择与平均、抽样调查的设计与分析、决策函数的优良性、疾病与基因的关联分析等方面的研究中取得了一系列重要成果，得到了国内外同行的好评与肯定，并被广泛引用。共出版教材2本，发表学术论文130余篇；主持和参加过近30项国家科学基金项目以及全国性的实际课题，提出的预测方法被实际部门所采用。

欢迎各位老师和同学参加！