Models of Complex Experimentation: Attribute Discovery, Contextual Experimentation, and Experimentation on Causal Graphs

该奖项用于经济理论方面的研究。目标是分析复杂环境中的学习和战略实验，在这些环境中，信息很重要，但也不总是很容易被发现。实验是技术、公共政策、市场和科学创新和进步的关键。大数据的日益普及和实验方法的广泛使用意味着，企业和政府会进行试验或试验性工作，以快速评估从营销计划到政府推广工作的一切。然而，如此丰富和复杂的数据也为选择性和扭曲的学习留下了空间。这项研究将建立一个框架，在三个复杂的环境中研究这种扭曲：社会项目的多地点评估；联邦政府系统中跨州的小规模实验和政策传播；以及因果关系的动态学习。该研究试图为现有的最优属性学习、因果实验和具有相关结果的战略实验做出贡献。PI将介绍和分析在实验对象复杂的三个不同环境中的最优实验。第一个项目研究了通过选择性地发现共同感兴趣的多属性对象的属性来进行实验和说服的委托代理模型。在概念上，这项研究将最优属性抽样与对社会项目的小规模多地点评估进行了类比。在方法论上，它提供了一个基于高斯过程的新的理论框架。第二个项目将使用这个高斯框架来建立联邦系统中实验的动态模型。联邦联盟内各个州进行的试验是小规模基于背景的试验和大规模政策采纳之间相互作用的一个显著例子。该分析试图解释横向和纵向政策扩散的经验模式。第三个项目的动机是关于在实证工作中将机制实验作为传统政策实验的替代方案的辩论，该项目研究随机变量因果图上的战略实验。这项分析将寻求确定机制实验主导政策实验的形式条件，将因果图论的工具与经典的多臂强盗问题相结合。这一奖项反映了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 award funds research in economic theory. The goal to analyze learning and strategic experimentation in complex environments where information is important but is also not always easily discovered. Experimentation is key to innovation and progress in technology, public policy, markets, and science. The growing availability of big data and wide spread use of experimental methods mean that businesses and governments conduct trial or experimental efforts to quickly evaluate everything from marketing plans to government outreach efforts. Yet such data abundance and complexity also leaves room for selective and distorted learning. This research will develop a framework to study such distortions in three complex environments: multi-site evaluation of social programs; small-scale experimentation and policy diffusion in across states in a federal system of government; and dynamic learning of causal relationships.The research seeks to contribute to existing work in optimal attribute learning, causal experimentation, and strategic experimentation with correlated outcomes. The PI will introduce and analyze optimal experimentation in three distinct environments in which the object of experimentation is complex. The first project studies a principal-agent model of experimentation and persuasion through selective discovery of attributes of a multi-attribute object of common interest. Conceptually, the study draws an analogy between optimal attribute sampling and small-scale multi-site evaluation of social programs. Methodologically, it offers a novel theoretical framework based on Gaussian processes. The second project will employ this Gaussian framework to build a dynamic model of experimentation in federal systems. Experimentation performed by individual states within a federal union is a striking example of the interplay between small-scale context-based experimentation and large-scale policy adoption. The analysis seeks to explain empirical patterns of horizontal and vertical policy diffusion. Motivated by the debate on mechanism experiments as an alternative to traditional policy experiments in empirical work, the third project studies strategic experimentation on a causal graph of random variables. The analysis will seek to identify formal conditions under which mechanism experiments dominate policy experiments, by marrying tools of causal graph theory with classic multi-armed bandit problems.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.