CAREER: Robust Matching Algorithms for Causal Inference in Large Observational Studies

这项教师早期职业发展计划(Career)拨款将通过利用大数据的力量在大规模观察性研究中推断因果关系，从而促进国家健康、繁荣和经济福利。在许多情况下，特别是在公共卫生领域，设计对照研究来评估有效的公共政策可能是困难的或昂贵得令人望而却步。随着大规模数据收集的增加，通过将观察划分为适当的集合来推断治疗和结果之间的因果关系的方法的设计已成为一种有吸引力的替代方案。当前的因果推理方法面临着几个根本性的挑战，这些挑战可能导致次优的政策选择。该项目将开发易于处理的计算方法，以促进更好的政策决策。作为一个重要的用例，该项目将使用美国大规模医疗保健数据评估改善阿片类药物使用障碍(OUD)治疗质量的政策。综合教育和研究计划将吸引和参与不同的学生群体，从高中到研究生院，从事研究和实践。通过与社区学院和HBCU等合作组织的积极参与，该项目将为工程学中代表性不足的群体提供机会，以满足紧迫的社会需求。使用现代最优化观点，该项目将通过开发包括推理和匹配的理论和计算框架来改进现有的因果推理方法，以从观察性研究中识别因果关系。研究的目标是(1)建立一个稳健的因果推理框架，通过匹配方法减少不确定性，(2)确保高维空间中的协变量平衡，(3)开发最优协变量平衡技术，通过确保期望的分布特性来减少偏差和模型依赖，以及(4)基于该框架评估和推进美国医疗保健政策。为此，将采用严格的优化框架来明确考虑因果推理中的不确定性，在具有匹配要求的低维数据中维护高维数据的邻域结构，并确保观测数据的最佳分布特性。利用问题结构将开发出高效的精确解算法。可伸缩性将通过具有理想收敛特性的算法方案和基于数据结构的分解方法来解决。这些算法预计将对各种优化问题有用，如二次分配、凸-非线性可行性和二元可行性。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。

This Faculty Early Career Development Program (CAREER) grant will advance the national health, prosperity, and economic welfare by utilizing the power of big data to infer causality in large-scale observational studies. In many situations, particularly in the public health domain, it may be difficult or prohibitively expensive to design controlled studies to evaluate effective public policies. As large-scale data collection increases, the design of methods to infer causality between treatment and outcome by partitioning observations into appropriate sets has become an attractive alternative. Current methods underlying causal inference suffer from several fundamental challenges that may lead to sub-optimal policy selection. This project will develop tractable computational approaches to facilitate better policy decision making. As an important use case, the project will evaluate policies for improving treatment quality of Opioid Use Disorder (OUD) using large-scale U.S. healthcare data. The integrated education and research plan will attract and involve a diverse student body, from high school through graduate school, in research and practice. Through active engagement with partnering organizations, including community colleges and an HBCU, the project will provide opportunities for members of underrepresented groups in engineering to address pressing societal needs. Using a modern optimization perspective, this project will advance existing methods for causal inference by developing a theoretical and computational framework that encompasses both inference and matching to identify causality from an observational study. The research objectives are to (1) establish a robust causal inference framework with matching methods to reduce uncertainty, (2) ensure covariate balance in high dimensional space, (3) develop optimal covariate balance techniques to reduce bias and model dependency by ensuring desired distributional properties, and (4) evaluate and advance U.S. healthcare policies based on this framework. To this end, a rigorous optimization framework will be employed to explicitly account for uncertainties in causal inference, maintain neighborhood structures of high dimensional data in low dimensions with matching requirements, and ensure optimal distributional properties of observational data. Efficient exact solution algorithms will be developed exploiting problem structure. Scalability will be addressed through algorithmic schemes with desirable convergence properties and data structure-based decomposition methods. These algorithms are expected to be useful to a wide variety of optimization problems such as quadratic assignment, convex-nonlinear feasibility, and binary feasibility.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.