Primary Submission Category: Synthetic Control Method
A new perspective on synthetic controls
Authors: Yujin Jeong, Dominik Rothenhäusler,
Presenting Author: Yujin Jeong*
In statistics and machine learning, we often want to quantify uncertainty and frame optimality with respect to sampling uncertainty. However, if we combine evidence from different data sets, sampling uncertainty might be lower order than the distribution shift between the data sets. This raises the question of how to optimally estimate in a data fusion setting. To address this issue, we model distributional shifts as a superposition of numerous random changes. We then develop tools for measuring the similarity between randomly perturbed distributions, estimating parameters of perturbations, and predicting outcomes for new distributions. Interestingly, these tools share a close connection to synthetic controls. Our framework provides a new language for distributional shifts and offers a fresh perspective on synthetic controls. Moreover, we evaluate its performance on real-world data sets and demonstrate that our new language and tools significantly improve estimation accuracy.