Collaborative Research: Theoretical and Methodological Frameworks for Causal Inference of Peer Effects

点击翻译按钮获取中文摘要

Understanding how friends or peers interact and affect each other is often of great interest in biomedical studies and the social sciences. However, it is not entirely clear how to quantify and develop inference for peer effects using a formal statistical framework. This project will focus on the development of a statistical causal inference framework to address these challenges, with the goal of developing both theoretical and methodological tools for a wide class of questions involving inference of peer effects. The methods will be applied to investigate peer effects among university students with different academic backgrounds. The research could provide important guidance for decision and policy makers.The classical potential outcomes framework for causal inference assumes no interference among experimental units. In some empirical studies, interference is a nuisance that complicates analysis and should be avoided by careful experimental design. In many applied fields, however, group or network structures exist and could cause interference among units. Interference is no longer a nuisance in these applications, because studying the pattern of causal effects with interference is the scientific question of interest with important implications for policy or decision making. The existing literature discusses external interventions on the units, where the networks, clusters or groups that induce interference are known a priori. The new framework allows for the development of inferential tools for causal inference with interference from the Fisherian, Neymanian, and Bayesian perspectives. Under the Fisherian view, randomization tests will be used to detect deviations from the sharp null hypothesis without imposing further structural assumptions. Under the Neymanian view, randomization-based point and interval estimators, which serve as the basis for finding optimal treatment assignments will be developed. Under the Bayesian view, hierarchical models will be developed to accommodate complex structures of real-life data and incorporate information from multiple groups with longitudinal outcomes. The project will also lead to the development of open-source R software. This project is supported by the Division of Mathematical Sciences and the Methodology, Measurement, and Statistics (MMS) Program in the Directorate for Social, Behavioral, and Economic Sciences.