Ye Wang presents History versus Unobservable Confounding in Panel Data Analysis: A Design-Based Perspective

Abstract: Should researchers control for the observable history of variables, such as lagged dependent variables, or unobservable confounders, such as fixed effects, when attempting to establish causality in panel data? In this paper, we review two commonly invoked assumptions to justify such choices: sequential ignorability and strict exogeneity, along with the methods available under each framework. We argue that when treatment status never reverses for any unit or when the outcome is unaffected by past treatment statuses, methods under both assumptions often produce comparable estimates, rendering the choice inconsequential. Otherwise, methods relying on sequential ignorability fail in the presence of unobservable confounders, whereas methods based on strict exogeneity, such as fixed-effects models, cannot identify meaningful causal quantities without additional restrictions on the data-generating process. We provide practical guidance for researchers navigating these scenarios and substantiate our claims through Monte Carlo simulations and two replication exercises. Our replication of Acemoglu et al. (2019) shows that applying fixed-effects models to the dataset with both treatment reversal and dynamic treatment effects leads to an overestimation of the impact of democracy on economic development.