New Statistical Methods for Randomized Experiments in Political Science and Public Policy

自20世纪20年代以来，治疗分配的随机化一直是现代科学发展最有力的工具之一，随机化实验在自然科学中被广泛用于检验假说。政治学和公共政策领域长期以来一直由观察性研究主导，但现在它们显示出越来越多地使用实验研究。这项拟议研究的目标是开发和评估随机试验的统计分析的新方法。在社会科学中，实验经常在实验室外进行，以增加结论的概括性。然而，这是以完全控制实验参与者对实验治疗的暴露为代价的。因此，为了确定有效的因果关系，通常必须进行统计调整。在拟议的研究中，将开发和评估新的统计方法，以解决随机实验中的丢失数据问题，当丢失数据机制依赖于变量的未观测值时，并在随机化单元为一群个体时实施有效的实验设计。这项提议的学术价值在于，提议的方法在政治学和公共政策内外具有广泛的适用性。PI正在进行的战略是在具体实验的背景下开发统计方法。使用这些振奋人心的例子使PI能够发现统计分析必须改进的领域，并需要开发新的方法。将分析五个实验：德国和日本的投票(调查)实验，墨西哥全民医疗保险计划的随机评估，非洲关于慎重决策的实地实验，以及关于种族启动对政治态度影响的调查实验。PI在因果推理的方法论研究方面进行了广泛的工作，特别强调在政治学中的应用，拟议的研究建立在早期工作的基础上。拟议的方法及其实质性应用作出了原创性的贡献。这个项目的更广泛的影响有几个。首先，上述随机试验的分析将取得实质性进展。将开发新的方法，以更好地了解政策信息和心理操纵对投票行为的影响；墨西哥全民医疗保险计划的健康和财务影响；领导人在协商民主中的作用；以及最近关于种族启动的学术辩论。拟议的方法还可以应用于政治学和公共政策之外，并应用于医学界在进行和分析群组随机试验方面的最佳做法建议。拟议的研究表明，有必要对这项研究的现行标准提出质疑。最终，改进的方法将提供给对统计理论和计算知识有限的应用研究人员。

Randomization of treatment assignment has been one of the most powerful tools for the development of modern science ever since the 1920s, and randomized experiments have been widely used to test hypotheses in the natural sciences. The fields of political science and public policy were long dominated by observational studies, but they now show growing use of experimental studies. The goal of the proposed research is to develop and evaluate new methods for the statistical analysis of randomized experiments. In the social sciences, experiments are often conducted outside of a laboratory to increase the generalizability of conclusions. However, this comes at the expense of having complete control over experimental participants' exposure to experimental treatments. Thus, statistical adjustments often must be made in order to ascertain valid causal effects. In the proposed research, new statistical methods will be developed and evaluated that address missing data problems in randomized experiments when the missing data mechanism depends on unobserved values of variables and implement efficient experimental designs when the unit of randomization is a cluster of individuals. The intellectual merit of the proposal lies in the wide applicability of the proposed methods within and beyond political science and public policy. The PI's ongoing strategy is to develop statistical methods in the context of specific experiments. Use of these motivating examples allows the PI to discover areas in which statistical analysis must be improved and new methods need to be developed. Five experiments will be analyzed; voting (survey) experiments in Germany and Japan, randomized evaluation of the Mexican universal health insurance program, a field experiment about deliberative decision-making in Africa, and a survey experiment about effects of racial priming on political attitudes. The PI has worked extensively on methodological research for causal inference with a particular emphasis on applications in political science, and the proposed research builds on the earlier work. The proposed methods as well as their substantive applications make original contributions. The broader impacts of this project are several. First, substantive progress will be made on the analyses of the aforementioned randomized experiments. New methods will be developed to better understand the effects of policy information and psychological manipulation on voting behavior; the health and financial effects of the Mexican universal health insurance program; the role of leaders in deliberative democracy; and the recent academic debate about racial priming. The proposed methods can also have applications beyond political science and public policy and into medical community recommendations for best practices in the conduct and analysis of cluster-randomized trials. The proposed research suggests the need to question the current standards for this research. Ultimately, improved methods will be made available to applied researchers with limited knowledge of statistical theory and computing.