CAREER: Robust Causal And Statistical Inference In High Dimensional Structured Systems With Hidden Variables

理解因果关系在实证科学中是至关重要的，然而观察到的关联并不总是有明确的因果解释。例如，在医院接受抗生素治疗的病人也容易遭受机会性感染，尽管我们并不认为抗生素会导致感染。在这种情况下，可能的解释是，早期出现的传染性病原体导致了后一种感染，医生开了抗生素处方。从数据中找到有效因果关系的一个关键障碍是存在隐藏的但与因果相关的变量，如上述传染因子。该项目旨在开发新的方法，在具有隐藏变量的数据集中绘制有效的因果推理，同时避免现有方法的已知陷阱。除了在数据分析中的作用外，因果素养是公民和消费者做出知情选择的重要技能。研究者的目标是通过将因果方法纳入约翰霍普金斯大学现有的数据科学课程来提高这一技能，开发一门新课程，教授发现因果主张误解的方法，并开发一个教程，旨在弥合机器学习和统计学在讨论和研究因果推理方面的差距。有向无环图（dag）是对完全观察到的因果系统进行推理的一种优雅方法。本研究旨在为具有隐变量的因果系统提供一种新的形式主义，既保留了dag的优点，又消除了直接表示隐变量的缺点。这种形式通过混合图的规则模型捕获观察到的边际分布中的所有相等约束。该研究的成功将显著促进对隐含变量因果系统中所有主要任务的理解：识别、估计和计算效率的概率计算。作为方法学发展的试验台，研究者将使用马龙医疗保健工程中心和约翰霍普金斯外科部门合作获得的电子健康记录数据集。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Understanding causes and effects is crucial in empirical science, however observed associations do not always have clear causal explanations. For instance, hospital patients who were prescribed antibiotics also tend to suffer from opportunistic infections, though we don’t expect antibiotics to cause infections. The likely explanation in this case is that an early presence of an infectious agent caused the latter infection, and the doctor prescribing antibiotics. A key obstacle to finding valid cause-effect relationships from data is the presence of hidden but causally relevant variables, like the infectious agent above. This project aims to develop new methods for drawing valid causal inferences in datasets with hidden variables while avoiding known pitfalls of existing approaches. Aside from its role in data analysis, causal literacy is an important skill for making informed choices as citizens and consumers. The investigator aims to promote this skill by incorporating causal methods into existing data science courses at Johns Hopkins University, developing a new course that will teach methods for detecting misunderstandings of causal claims, and developing a tutorial aimed at bridging the gap between machine learning and statistics in discussing and working on causal inference.Directed acyclic graphs (DAGs) are an elegant method for reasoning about fully observed causal systems. The proposed research aims to provide a new formalism for causal systems with hidden variables that retains the advantages of DAGs, while dispensing with the disadvantages of representing hidden variables directly. This formalism captures all equality constraints in the observed marginal distribution via a regular model of a mixed graph. Success in the proposed research will significantly advance understanding of all major tasks in causal systems with hidden variables: identification, estimation, and computationally efficient probabilistic calculations. As a test bed for methodological developments, the investigator will use a dataset of electronic health records obtained in partnership with the Malone Center for Engineering in Healthcare and the Johns Hopkins Department of Surgery.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.