Leveraging Background Knowledge for Identification and Estimation of Causal Effects in the Presence of Latent Variables

该项目侧重于利用观测数据和专家知识进行因果推理。估计因果关系是许多科学努力的目标，通常，唯一可用的数据来自观察性研究。然而，仅根据观测数据估计因果关系并不总是可能的。利用专家知识和观测数据的混合方法可能有助于估计因果关系或缩小可能估计的范围。这种混合方法目前仅限于假设因果系统中的所有变量都被观察和测量。对于许多实际应用程序来说，这个假设通常过于严格。该项目将开发使用专家知识的方法，以帮助估计存在隐藏变量的观测数据的因果关系。该项目开发的方法将立即适用于广泛的科学学科，最著名的是流行病学、经济学、个性化医疗和算法公平性研究。本项目的研究基于概率图形模型，该模型可用于表示观察到的变量集上的条件独立关系。一般来说，仅根据观测数据，因果图模型只能识别到这种图的等价类。该项目的第一个目标是开发一套完整的规则，用于更新基于某些因果关系的专家知识的兼容图形模型集。该项目的第二个目标是根据更新的模型集制定因果关系识别和估计的图形标准。此外，研究者将在计算和统计效率方面研究更新后的等效类。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。

This project focuses on causal inference using observational data and expert knowledge. Estimating causal effects is a goal of many scientific endeavors, and often, the only available data are those from observational studies. However, estimating causal effects based on observational data alone is not always possible. A hybrid method that leverages both expert knowledge and observational data may help estimate a causal effect or narrow down a range of likely estimates. Such hybrid methods are currently limited to assuming that all variables in the causal system are observed and measured. This assumption is often too stringent for many real-world applications. This project will develop methods for using expert knowledge to help estimate causal effects from observational data in the presence of hidden variables. The methods developed in this project would be immediately applicable to an extensive range of scientific disciplines, most notably epidemiology, economics, personalized medicine, and the study of algorithmic fairness.The research in this project is grounded on probabilistic graphical models that can be used to represent conditional independence relationships on the observed set of variables. In general, based on observational data alone, the causal graphical model can be identified only up to an equivalence class of such graphs. The first goal of the project is to develop a complete set of rules for updating the set of compatible graphical models based on expert knowledge of certain causal relationships. The second goal of the project concerns the development of graphical criteria for causal effect identification and estimation based on the updated set of models. Furthermore, the investigator will study the updated equivalence class in terms of computational and statistical efficiencies.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.