Essays in International Trade

Abstract

This dissertation studies economic geography and international trade. It discusses topics on productivity spillovers, agglomeration economies, transportation infrastructure, labor market dynamics, economic growth, comparative advantage and machine learning.

Chapter 1 measures productivity spillovers across cities by using the development of highspeed railways (HSR) in China as a natural experiment. HSR shortens inter-city passenger travel time, makes face-to-face communication easier and thus facilitates knowledge spillovers. I develop a dynamic spatial general equilibrium model that features intra- and international trade, frictional domestic migration and dynamics in labor markets. My structural estimation on the productivity spillover parameters show that production externalities are substantial but become negligible between cities that require more than 2 hours of travel time. I then calibrate the model to 2010 Chinese economy and characterize the out-of-steady-state dynamics of cities’ employment and income. Quantitative results indicate that the HSR network completed in mainland China before 2015 will affect the location choice of 1.33% of the total workforce in the long run. It benefits southern and southeastern regions where both cities and HSR routes are densely located substantially more than the northern or western regions in terms of labor inflow, regional productivities and real income.

Chapter 2 (joint work with Dominick Bartelme) uses machine learning techniques to examine whether comparative advantage (CA) structure predicts GDP growth. We first show that Hausmann et al. (2007)’s EXPY, an aggregate index widely used by policy makers as GDP growth predictor, fails to have predictive power out of sample. We then examine if the failure of EXPY was due to a loss of information during the aggregation process by directly investigating the linkage between export-revealed sector-level CA structure and GDP growth while controlling for foreign demand shocks. To handle the high dimensionality problem, we adopt machine learning techniques exemplified by Random Forest. We find the sector-level CA structure outperforms EXPY when predicting GDP growth; nevertheless, its predictive power becomes limited after controlling for a few additional standard macro variables.