RI: Medium: Collaborative Research: Causal Inference: Identification, Learning, and Decision-Making

了解观察到的现象背后的因果机制是科学的主要目标之一。由于认识到统计关联本身不足以阐明这些机制，研究人员利用基于“因果推理”的技术来丰富传统的统计分析。然而，该领域最近的大多数进展都是在过于乐观的假设下进行的，而这些假设往往无法在实际的大规模情况下得到满足。本项目旨在发展一个健全和普遍的因果推理理论，以涵盖这些情况。目标是为智能系统的决策设计一个框架，包括(1)学习数据生成环境的因果表示（学习），(2)利用学习模型执行有效的推理（规划/推理），以及(3)基于(1)和(2)使用新的推断表示来决定下一步如何行动（决策）。这一新发现将有利于实证科学各个领域的研究人员，包括人工智能、机器学习、统计学、经济学以及健康和社会科学。预计该研究将从根本上改变数据科学在违反标准因果假设领域的实践（即缺失数据，选择偏差和混淆偏差）。有关决策的工作预计将为“自动化科学家”的设计铺平道路，即一个将观察和实验数据结合起来的程序，进行自己的实验，并决定行动和政策的最佳选择。该项目还通过(1)参与建立新的“数据科学”课程，其中因果推理起着核心作用，以及(2)为学生和公众开发新的教育材料，解释因果推理的实践（例如，书籍），帮助在整个科学中传播因果推理的原则。此外，该项目支持因果推理社区，通过促进一些教育倡议，如论坛、讲习班和创建新的激励措施，以开发教育材料（例如，“因果教育奖”）。宣称因果关系的存在（结构学习）、因果效应的大小（识别）和设计最佳干预（决策）是在数据驱动领域中发现的一些最重要的任务。本项目研究识别、学习和决策设置，其中(1)数据缺失不是随机的，(2)非参数估计不可行的，(3)聚合行为不能转化为个人层面决策的指导。具体而言，该项目考虑了测量系统扭曲（丢失数据）时的问题，这在统计文献中得到了大量关注，但在数据非随机丢失时，在因果推理的背景下基本上没有进行调查。该项目进一步旨在利用线性模型（非参数因果推理的最常见的第一近似）的特殊属性来阐明数据中的因果关系，并促进此类模型的敏感性分析。最后，该项目考虑了因果和反事实知识如何加速实验和支持原则决策的基本问题。目标是开发一个完整的算法理论，以确定何时可以从数据中学习到特定的因果效应，以及如何将学习到的因果知识（可能是通过实验）结合起来，以便可以在新的环境条件下摊销。

Understanding the causal mechanisms underlying an observed phenomenon is one of the primary goals of science. The realization that statistical associations in themselves are insufficient for elucidating those mechanisms has led researchers to enrich traditional statistical analysis with techniques based on "causal inference". Most of the recent advances in the field, however, operate under overly optimistic assumptions, which are often not met in practical, large-scale situations. This project seeks to develop a sound and general causal inference theory to cover those situations. The goal is to design a framework for decision-making of intelligent systems, including (1) learning a causal representation of the data-generating environment (learning), (2) performing efficient inference leveraging the learned model (planning/inference), and (3) using the new inferred representation, based on (1) and (2), to decide how to act next (decision-making). The new finding will benefit investigators in every area of the empirical sciences, including artificial intelligence, machine learning, statistics, economics, and the health and social sciences. The research is expected to fundamentally change the practice of data science in areas where the standard causal assumptions are violated (i.e., missing data, selection bias, and confounding bias). The work on decision-making is expected to pave the way toward the design of an "automated scientist", i.e., a program that combines both observational and experimental data, conducts its own experiments, and decides on the best choices of actions and policies. The project also helps to disseminate the principles of causal inference throughout the sciences by (1) engaging in the establishment of new "data science" curriculum where causal inference plays a central role, and (2) developing new educational materials for students and the general public explaining the practice of causal inference (e.g., book). Furthermore, the project supports the causal inference community by fostering a number of educational initiatives such as forums, workshops, and the creation of new incentives for the development of educational material (e.g., a "Causality Education Award").Making claims about the existence of causal connections (structural learning), the magnitude of causal effects (identification), and designing optimal interventions (decision-making) are some of the most important tasks found throughout data-driven fields. This project studies identification, learning, and decision-making settings where (1) data are missing not at random, (2) non-parametric estimation is not feasible, and (3) aggregated behavior does not translate into guidance for individual-level decision-making. Specifically, the project considers the problem when measurements are systematically distorted (missing data), which has received an enormous amount of attention in the statistical literature, but has not essentially been investigated in the context of causal inference when data are missing not at random. The project further aims to leverage the special properties of linear models, the most common first approximation to non-parametric causal inference, to elucidate causal relationships in data, and to facilitate sensitivity analysis in such models. Finally, the project considers the fundamental problem on how causal and counterfactual knowledge can speed-up experimentation and support principled decision-making. The goal is to develop a complete algorithmic theory to determine when a particular causal effect can be learned from data and how to incorporate causal knowledge learned (possibly by experimentation) so that it can be amortized over new environmental conditions.