Thesis title: Think Outside the Black Box! Leveraging Machine Learning Predictions for Policy Insight
This thesis explores three distinct applications of ML to inform policy, each demonstrating how predictive algorithms can generate actionable insights while remaining cognizant of their limitations.
Chapter 1 employs an Elastic Net algorithm to construct "what-if" scenarios in the context of European green transition policies. The green transition represents a paradigmatic multifaceted process, requiring the simultaneous pursuit of multiple socioeconomic outcomes, many of which are collinear, across contexts marked by substantial local heterogeneity. Machine learning approaches are well-suited to this challenge: collinearity poses no obstacle to regularized regression methods, andthese algorithms explicitly exploit heterogeneity to improve predictive performance. We extend the standard Elastic Net framework to accommodate spatial effects, accounting for spillover and agglomeration dynamics that have long been documented in regional development literature. This approach enables us to anticipate potential winners and losers of the transition and to compare different policy interventions based on their predicted effects across diverse regional contexts.
Chapter 2 trains an ensemble of ML algorithms to predict regional unemployment rates across the EU. Through this exercise, we explore both the relevance
and functional form of relationships between diverse predictors and regional unemployment rates. The results provide data-driven insights into numerous questions raised by economic theory, from the role of inflation dynamics to the impact of environmental policies on labor market outcomes, paving the way for future impact assessments. Most significantly, we demonstrate that tree-based ML algorithms can predict unemployment with considerable precision, offering policymakers valuable early warning capabilities in detecting emerging regional vulnerabilities.
Chapter 3 explores the potential of ML in causal inference by delving into the Matrix Completion for Causal Panels (MCP) framework, a novel approach to
counterfactual imputation. This method leverages ML’s predictive capacity to impute counterfactuals for treated units, enabling a previously inaccessible level of disaggregation in treatment effect estimation. Crucially, MCP implicitly assumes interdependence among units, effectively overcoming the traditional Independent and identically distributed (i.i.d.) assumption. This aspect is embedded in the algorithm’s functioning, yet it represents a paramount shift compared to more traditional methods, entailing important economic implications. We apply MCP to assess the labor market impacts of minimum wage policies, demonstrating its value as an analytical tool while also highlighting important limitations. Thanks to the imputation of counterfactual quantities, we can test previously untestable hypotheses, including parallel trends, providing with additional powerful insights.
Together, these three studies underscore a crucial lesson. ML is a powerful tool, but its responsible application requires clearly delineating its appropriate scope, accompanying quantitative results with contextual understanding, and maintaining transparency about the trade-offs inherent in choosing prediction over explanation.