CAREER: Robust Econometric Inference

该项目开发了新的计量经济学和统计工具，用于在现有技术不可靠的情况下对经济关系进行推断。在许多经济背景下，可用的数据包含的关于经济利益关系的有用信息相对较少，或者被现有的计量经济模型描述得很差。在这种情况下，现有的推理技术可能是不可靠的，使用这些技术的研究人员可能会在经济变量之间的关系以及拟议政策的可能影响方面得出错误的结论。该项目开发了在这种环境下进行可靠推理的新技术，使研究人员能够更准确地量化经济关系和政策建议的不确定性。为了增加这项研究的影响，调查人员组织、记录和公开发布了一系列讨论建议的方法的讲座，还发布了为可公开访问的软件存储库开发的方法的代码。此外，研究人员还参加了由美国经济协会为少数族裔背景的研究生组织的指导计划。该计划由四个部分组成。第一部分利用微分几何中的工具来研究具有异质处理效应的工具变量模型中的弱辨识-稳健推理，填补了现有文献中关于辨识-稳健推理的一个重要空白。第二个组件使用一种新的集成似然函数族来推导具有潜在辨识失败的矩相等模型的最优检验和估计器，所述集成似然函数族积分出关于功能干扰参数的先验。第三个部分表明，通过利用许多部分识别的矩不等式应用程序的共同结构，可以获得对感兴趣的低维参数的强大的、计算简单的推断。第四个也是最后一个部分考虑了局部错误指定的模型，在局部错误指定下的推理和弱识别下的推理之间建立了联系，并提出了允许错误指定的技术-对感兴趣的结构或因果参数进行稳健推理。

This project develops new econometric and statistical tools for drawing inferences about economic relationships in contexts where existing techniques are unreliable. In many economic contexts, the available data contain relatively little usable information about the relationships of economic interest, or are poorly described by available econometric models. In such cases, existing techniques for inference can be unreliable, and researchers using these techniques risk drawing spurious conclusions about relationships among economic variables and the likely effects of proposed policies. This project develops new techniques for reliable inference in such settings, allowing researchers to more accurately quantify uncertainty about economic relationships and policy proposals. To increase the impact of this research, the investigator organizes, records, and publicly posts a series of lectures discussing the proposed methods, and also posts code for the methods developed to publicly accessible software repositories. Further, the investigator participates in a mentoring program for graduate students from underrepresented minority backgrounds organized by the American Economic Association.The project consists of four components. The first component uses tools from differential geometry to study weak identification-robust inference in instrumental variables models with heterogeneous treatment effects, closing an important gap in the existing literature on identification-robust inference. The second component derives optimal tests and estimators for moment equality models with potential identification failure, using a novel family of integrated likelihoods which integrate out a prior on a functional nuisance parameter. The third component shows that by exploiting structure common to many partially-identified moment inequality applications one can obtain powerful, computationally simple inference for low-dimensional parameters of interest. The fourth and final component considers locally misspecified models, develops a connection between inference under local misspecification and inference under weak identification, and proposes techniques which allow misspecification-robust inference on structural or causal parameters of interest.