复杂试验和大数据中因果推断的方法和理论

在大规模随机化试验和观察性研究中，研究者能够获得大量的高维的协变量信息，例如年龄、性别、基因信息、消费行为等，而且协变量的结构越来越复杂，以图片、文本、音频、视频等方式呈现。如何利用这些高维复杂的协变量信息探索因果关系成为大数据和人工智能领域研究的核心和热点问题。本项目主要研究复杂试验和大数据中因果推断的统计学方法和理论。具体而言，本项目基于潜在结果模型和随机化推断的框架，主要研究随机区组试验中的高维变量选择、稀疏建模、重新随机化和bootstrap等方法的理论性质；研究顺序随机化试验中的高维回归调整和重新随机化的方法和理论；研究裂区试验中平均因果效应的估计和推断方法；并研究大数据中异质性因果效应的统计推断方法及其在个性化医疗、精准营销等领域的应用。最后，应用这些方法和理论解决农业、医学、经济、社会科学等领域中的因果问题。

In large-scale randomized experiments and observational studies, researchers often observe a huge amount of high-dimensional covariates, such as age, gender, genetic, and consumer behaviors. Moreover, the structure of covariates becomes more and more complex, for example, image, text, audio, video and so on. How to use these covariates information to efficiently estimate and infer causal relationships is becoming an important and hot research problem in big data and Artificial Intelligence. This project mainly focuses on developing statistical theory and methods for drawing causal inference from complex experiments and big data. Specifically, under the potential outcomes causal model and randomization-based inference framework, this project investigates theoretical properties of high-dimensional variable selection, sparse modeling, re-randomization, and bootstrap in randomized block experiments; develops high-dimensional regression adjustment and re-randomization theory and methods in sequentially randomized experiments; studies statistical methods for estimating and inferring average treatment effect in split-plot experiments; and explores statistical and machine learning methods for estimating heterogeneous treatment effect from big data with applications in personalized medicine and precision marketing. Finally, this project will apply the developed theory and methods to solve causal problems in fields of agriculture, medical science, economics and political science.