Determinants of PTA design: Insights from machine learning

Abstract

Preferential trade agreements (PTAs) have emerged as the dominant form of international trade governance. Provisions included in PTAs are increasingly numerous, broad in their purview, deep in their scope, and varied between agreements. We study the economic, political, and geographic determinants of PTA design differences. For each of the hundreds of classified PTA provisions, we consider 287 country-pair characteristics as potential determinants, covering many individual mechanisms the literature has studied. We employ random forests, a supervised machine learning technique, to handle this high dimensionality and complexity. We use a robust variable importance measure to identify the most critical determinants of the inclusion of each PTA provision. Contagion due to competition for export markets, geographic proximity, and governance quality emerge as essential determinants of PTA design. These results motivate future exploration of individual mechanisms our exercise points to.