CAREER: The Design-Based Perspective of Causal Inference in Complex Experiments

点击翻译按钮获取中文摘要

Randomized experiments have been widely used in agriculture, industry, and clinical trials. R. A. Fisher formally discussed the value of randomization in experiments: it balances observed and unobserved covariates on average and serves as a basis for statistical inference. Classical results, however, are limited to simple experiments without rich covariates and complex time and hierarchical structures. Modern applications stemming from the social sciences and technology companies have richer covariates and more complex time and hierarchical structures. Motivated by these new applications, the PI will advance the theories and methodologies for the design and analysis of modern experiments for robust treatment effect estimation in various settings. Highlighting the role of the design of experiments, the PI will take a coherent design-based perspective of causal inference. In particular, the PI will propose various new experimental designs that can better balance covariates across experimental groups and develop statistical methods for these designs that are robust to model assumptions on the outcome generating processes. These theoretical results for experiments will also shed light on principled analyses of observational studies where controlled experiments are infeasible.  The training component includes graduate and undergraduate course work as well as the development of software through the help of both undergraduate and graduate students. This constitutes a strong plan to integrate research and education. The design-based perspective of causal inference does not assume any strong outcome modeling assumptions and focuses on the treatment assignment mechanism that can be determined by the experimenters. Under this perspective, the PI will improve existing experimental designs to have better covariate balance and evaluate many model-based procedures when the corresponding model assumptions can be violated. The PI will first propose and analyze rerandomization in blocking, sequential and factorial settings, focusing on repeated sampling properties of the treatment effect estimators and discussing the estimators with and without covariate adjustment. The PI will then propose and analyze linear and nonlinear covariate-adjusted estimators for treatment effects, including the cases with and without noncompliance. Moreover, the PI will calibrate randomization tests with targeted weak null hypotheses and propose randomization tests with robust and efficient covariate adjustment, based on detailed analyses of completely randomized experiments with covariates and finely stratified experiments. The PI will also establish randomization-based inferential frameworks and procedures for experiments with time and hierarchical structures. Finally, the PI will develop and disseminate open-source R software packages that implement the methodologies.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.