Primary Submission Category: Synthetic Control Method
History versus Unobservable Confounding in Causal Inference with Panel Data: A Design-based Perspective
Authors: Ye Wang, Yiqing Xu,
Presenting Author: Ye Wang*
Should researchers control for past values of variables or unobservable factors such as the fixed effects when attempting to establish causality in panel data, or is it possible to account for both? The paper investigates this question from a design-based perspective, which distinguishes identification assumptions on the treatment assignment process from structural restrictions for estimators to yield interpretable results. We propose a framework to unify two identification assumptions that are commonly invoked to justify the choice, sequential ignorability and strict exogeneity, and review available methods under each. We argue that fixed effects models are compatible with the presence of dynamic effects when treatment status does not reverse for any unit. In such scenarios, many existing methods automatically account for both types of confounders and generate estimates that agree with each other. Otherwise, structural restrictions on the persistence or heterogeneity of dynamic effects must be imposed for researchers to rely on fixed effects models to address the challenge of causal identification. We provide guidance for applied researchers to choose from potential options under various circumstances and propose a novel estimator when a long pre-treatment period exists. We substantiate these propositions through Monte Carlo simulations and a case study examining the effect of democracy on economic development.