Safe and Robust Causal Inference for High-Dimensional Complex Data

了解多元变量之间的因果效应是许多实证和实验类型研究的核心。在数据海量的时代，统计学家面临着从海量数据中得出因果推论的挑战。例如，数据分析可能因群体内潜在的混杂变量、异质性和时间依赖性的存在而变得复杂。这对现有的因果推理方法提出了巨大的计算和统计挑战。在现代数据集可用性的推动下，该研究项目旨在开发尖端的机器学习方法，以解决从海量数据中得出因果推论的理论、方法和计算挑战。拟议研究的发展不仅将推动因果推理理论的前沿，而且将使医学、流行病学、计算机科学和社会科学等各个领域的广泛研究人员受益。该项目还为研究生提供研究培训机会。 该项目的目标是开发一种新颖的高维因果推理框架，该框架相对于一类基本估计器具有影响函数双重鲁棒性和安全性。具体项目包括提出一种与现代机器学习技术相结合的协变量平衡方法，用于对高维数据进行有效的因果推理，针对海量异构数据的最优分布协变量平衡，以及针对纵向数据中的边际结构模型的顺序协变量平衡方法。贯穿全文的一个共同主题是协变量平衡方法的使用，该方法来自“倾向得分同义反复”，如果平衡了协变量，则估计的倾向得分是适当的。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力优点和更广泛的影响审查标准进行评估，被认为值得支持。

Understanding causal effects among multivariate variables is central to many empirical and experimental types of research. In an era where data is vast, statisticians are faced with the challenge of drawing causal inferences from massive data. For example, the analysis of data can be complicated by the presence of potentially confounding variables, heterogeneity, and temporal dependence within populations. This poses tremendous computational and statistical challenges to existing causal inference methods. Driven by the availability of modern datasets, the research project aims to develop cutting-edge machine-learning methods to address the theoretical, methodological, and computational challenges of drawing causal inferences from massive data. The development of the proposed research would not only push the frontier of causal inference theory but also benefit a broad range of researchers in a variety of areas, including medicine, epidemiology, computer science, and social science. This project also provides research training opportunities for graduate students. The goal of the project is to develop a novel high-dimensional causal inference framework that is influence-function doubly-robust and safe relative to a class of base estimators. Specific projects include proposing a covariate balancing methodology coupled with modern machine learning techniques for efficient causal inference with high-dimensional data, optimally distributed covariate balancing for massive heterogeneous data, and a sequential covariate balancing approach for marginal structural models in longitudinal data. A common theme throughout is the use of the covariate balancing methodology, which comes from the ``propensity score tautology,” the estimated propensity score is appropriate if it balances the covariates.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.