CAREER: Approximate Causal Inference

因果关系是整个科学研究的核心。在没有因果信息的情况下，研究人员无法预测新干预措施的效果，无法估计追溯的反事实，或许更重要的是，无法构建对正在调查的现象的有意义的、深入的解释。随着史无前例的数据积累，找到有意义的解释的挑战可以归结为“数据融合”--即从在完全不同的、不可交换的条件下收集的实验研究和观察研究的组合得出因果解释(Bareinboim和PEAR，PROC)。娜塔莉。阿卡德。SCI。美国，2016)。尽管最近取得了所有进展，但在许多大规模环境中应用最先进的因果推理方法仍然不是一件微不足道的事情。特别是，科学家的现有知识并不总是符合理论的预期，理论也不接受作为输入(并产生作为输出)的更宽松的因果规范。鉴于理论的完备性，这些要求不能被严格放弃。然而，在现实中，一些研究人员即使在不满足所需条件的情况下也会继续提出他们的主张。在整个经验学科中，越来越多的人认识到，今天阐述的许多科学发现过于脆弱，无法抵抗更严格的审查，甚至无法重现。这个项目的目标是弥合理论所要求的条件(如果遵循，将产生强有力的、有科学依据的主张)和科学家手中可获得的知识之间的差距。具体地说，该项目寻求(1)描述现有数据和背景知识(科学理论)的组合与新假设的因果解释的强度之间的权衡，以及(2)构建允许粗略和不精确的输入，同时产生仍然具有因果意义的输出的近似方案。这项拟议的研究预计将为今天做出的大多数数据科学推断提供基础，这些推断将影响建立在因果关系基础上的几个数据密集型领域的实践，包括计量经济学、教育、生物信息学和医学。该项目还包含一个重要的教育部分。就像物理学和微积分在20世纪基础科学教育中的重要性一样，在一个数据丰富的现代社会，因果推理将成为本科生课程的重要组成部分。该项目开发了一个新的教育平台，专门向STEM学生传授因果推理的概念、原理和工具。这一新平台的主要目标是摆脱从普遍的基于回归的技术获得的非因果声明，以及模糊和不言而喻的声明，如“关联并不意味着因果关系”，并走向对支持因果声明所需条件的更基本的理解。这项建议的目标是为因果推理中的近似发展一个原则性的框架。有两个可能的近似维度，一个关于给定问题实例的输入，另一个关于给定问题实例的输出。首先，我们将开发充分和必要的识别条件，以接受未完全指定的模型(例如，因果DAG)作为输入，但只有对现象的较粗略描述可用。我们将进一步开发有效的程序，在给定关于潜在数据生成过程的结构性知识的情况下，确定是否可以从观察和实验数据集的组合中近似出因果量。该项目将进一步利用这两个结果来设计有效的学习算法，在宽松的假设下，输入只是部分指定的，输出可以是目标因果分布的近似。最后，我们将考虑当多个有偏见的数据集可用时，包括当受到选择偏差、混淆偏差和结构异质性的困扰时，学习因果解释的问题。其目标是开发一种近似因果推理的通用算法理论，能够产生更健壮、可重复和可概括的因果解释。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Causality is central to scientific inquiry across the sciences. Without causal information, researchers cannot predict the effects of new interventions, estimate retrospective counterfactuals, and, perhaps more importantly, construct meaningful, in-depth explanations of the phenomenon under investigation. With the unprecedented accumulation of data, the challenge of finding meaningful explanations can be summarized under the rubric of "data-fusion" -- namely, deriving a causal interpretation from a combination of experimental and observational studies collected under disparate, non-exchangeable conditions (Bareinboim and Pearl, Proc. Natl. Acad. Sci. U.S.A, 2016). Despite all the recent progress, it is still non-trivial to apply state-of-the-art causal inference methods in many large-scale settings. In particular, the scientist's available knowledge does not always match what the theory expects, and the theory does not accept as input (and generate as output) more relaxed causal specifications. Given the completeness of the theory, these requirements cannot be strictly waived. In reality, however, some researchers continue to make their claims even when the required conditions are not met. There is an increasing recognition throughout the empirical disciplines that many of the scientific findings articulated today are too fragile, incapable of resisting to a more rigorous scrutiny or even being reproduced. The goal of this project is to bridge the gap between the conditions entailed by the theory (which, if followed, would generate robust and scientifically-grounded claims) and the knowledge available at the hands of the scientist. Specifically, the project seeks (1) to characterize the trade-off between the combination of data and background knowledge (scientific theories) available versus the strength of newly hypothesized causal explanations, and (2) to construct approximation schemes allowing inputs that are coarse and imprecise, while generating outputs that are still causally meaningful. The proposed research is expected to offer foundational grounding for most of the data science inferences made today, which will impact the practice of several data-intensive fields that are built on cause-and-effect relationships, including econometrics, education, bioinformatics, and medicine. The project also contains a significant educational component. Similar to the importance of physics and calculus in basic science education in the 20th century, causal inference will be a vital component of the curriculum of undergraduate studies in a modern, data-rich society. The project develops a new educational platform tailored to teaching causal inference concepts, principles, and tools to STEM students. The primary goal of this new platform is to move from acausal claims obtained from pervasive regression-based techniques, as well as vague and self-evident statements such as "association does not imply causation", and go towards a more fundamental understanding of the conditions necessary to support causal statements. The goal of this proposal is to develop a principled framework for approximations in causal inference. There are two possible approximation dimensions, one regarding the input and the other the output of a given problem instance. First, we will develop sufficient and necessary identification conditions to accept as input a model that is not fully specified (e.g., a causal DAG), but only a coarser description of the phenomenon is available. We will further develop effective procedures for determining whether a causal quantity can be approximated from a combination of observational and experimental datasets, given structural knowledge about the underlying data-generating process. The project will further leverage both results to design efficient learning algorithms under the relaxed assumption that the input is just partially specified and the output can be an approximation of the target causal distribution. Finally, we will consider the problem of learning causal explanations when multiple biased datasets are available, including when plagued with selection bias, confounding bias, and structural heterogeneity. The goal is to develop a general algorithmic theory of approximate causal inference that is capable of producing more robust, reproducible, and generalizable causal explanations.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.