CAREER: CCF: CIF: Randomized Experimentation for Systems with Time-varying Dynamics and Network Interference

随机实验被广泛用于估计在自然科学和社会科学、工程、医疗保健和科技行业等一系列领域提出的治疗方法的因果效应。然而，许多现代系统涉及个人和时间之间的复杂依赖关系，称为干扰。这种干扰违反了对传统因果估计器的性能至关重要的假设。如果不对干扰因素进行调整，现有的方法可能会导致错误的结论，从而对政策决策产生不利的下游影响。该项目侧重于开发一种新的因果推理方法工具包，以解决网络相互作用和时间动态引起的干扰。该项目将开发一个因果推理软件包，作为一种工具，将研究整合到教育中，并向大学预科学生、本科生研究人员和行业合作者提供服务。该项目将考虑由邻居的治疗对个体结果的直接影响产生依赖性的设置，以及通过状态变量调节影响的设置，这样治疗效果可能会随着时间的推移在整个网络中传播。研究人员将探索不同的建模假设，以促进干扰复杂性的估计，同时旨在提供可推广到广泛类别的模型的鲁棒解决方案和分析。开发的技术将通过利用模型假设、随机实验设计和估计器的选择以及可实现的保证之间的相互作用，阐明已知结构和可用测量的丰富性对估计可行性的影响。该项目的教育部分还将包括构建一个软件包，使个人和研究人员能够模拟各种应用程序中的随机实验和观察研究，以基准新的解决方案概念。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。

Randomized experiments are widely used to estimate causal effects of proposed treatments in a range of domains, including the natural and social sciences, engineering, health care, and the tech industry. However, many modern systems involve complex dependencies across individuals and time, referred to as interference. The interference violates assumptions that are critical to the performance of conventional causal estimators. Without adjusting for interference, existing approaches can result in incorrect conclusions that have detrimental downstream impact on policy decisions. This project focuses on developing a new toolkit of methods for causal inference that address interference arising from network interactions and time dynamics. The project will develop a causal inference software package that will be used as a vehicle to integrate the research into education and outreach to pre-college students, undergraduate student researchers, and industry collaborators.The project will consider settings where the dependencies arise from direct impact of the treatments of neighbors on the outcome of an individual, as well as settings where the impacts are mediated through state variables such that the treatment effect may propagate across the network over time. The investigator will explore different modeling assumptions to facilitate estimation given the complexities of interference, while aiming for robust solutions and analyses that can be generalized to broad classes of models. The developed techniques will illuminate the implications of known structure and richness of available measurements on the feasibility of estimation by exploiting the interaction between model assumptions, the choice of the randomized experimental design and estimator, and achievable guarantees. The educational component of the project will also involve building a software package that enables individuals and researchers to simulate randomized experiments and observational studies across various applications to benchmark new solution concepts.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.