Causal inference with time-series cross-sectional data : with applications to positive political economy

Abstract

Time-series cross-sectional (TSCS) data are widely used in today's social sciences. Researchers often rely on two-way fixed effect models to estimate causal quantities of interest with TSCS data. However, they face the challenge that such models are not applicable when the so called "parallel trends" assumption fails, that is, the average treated counterfactual and average control outcome do not follow parallel paths. The first chapter of this dissertation introduces the generalized synthetic control method that addresses this challenge. It imputes counterfactuals for each treated unit using control group information based on a linear interactive fixed effect model that incorporates unit-specific intercepts interacted with time-varying coefficients. It not only relaxes the often-violated "parallel trends" assumption, but also unifies the synthetic control method with linear fixed effect models under a simple framework. The second chapter examines the effect of Election Day Registration (EDR) laws on voter turnout in the United States. Conventional difference-in-differences approach suggests that EDR laws had almost no impact on voter turnout. Using the generalized synthetic control method, I show that EDR laws increased turnout in early adopting states but not in states that introduced them more recently. The third chapter investigates the role of informal institutions on the quality of governance in the context of rural China. Using TSCS analysis and a regression discontinuity design, I show that village leaders from large lineage groups are associated with considerably more local public investment. This association is stronger when the groups appeared to be more cohesive.