Think outside the black box! Leveraging machine learning predictions for policy insight

This paper predicts regional unemployment in the European Union by applying machine learning techniques to a dataset covering 198 NUTS-2 regions, 2000 to 2019. Tree-based models substantially outperform traditional regression approaches for this task, while accommodating reinforcement effects and spatial spillovers as determinants of regional labor market outcomes. Inflation—particularly energy-related—emerges as a critical predictor, highlighting vulnerabilities to energy shocks and green transition policies. Environmental policy stringency and eco-innovation capacity also prove significant. Our findings demonstrate the potential of machine learning to support proactive, place-sensitive interventions that aim to predict and mitigate the uneven social and economic impacts of structural change across regions. This study advances the understanding of a recent causal machine learning tool through its adaptation to two prominent studies. Specifically, the tool employed is the Matrix Completion method for Causal Panels (MCP), developed by Athey et al. (2021a), and the studies replicated are Card and Krueger (1994) and Callaway and Sant’Anna (2021). The contribution we provide is twofold, both economic and methodological. First, we seek to reconcile divergent evidence on the impacts of minimum wage on (low-skilled) employment and explore potential heterogeneous effects. While most of the literature relies on difference-in-differences (DiD) designs, these approaches rest on the common trends assumption, which is basically nontestable, involving unobservable quantities. Therefore, our second contribution is to explore whether a matrix completion approach relying on an entirely different set of assumptions corroborates previous findings. We find proof of a small, yet significant, negative and amplifying effect of a minimum wage introduction on the employment of targeted groups, in line with the prevailing literature. Second, we delve deep into MCP functioning, highlighting its potential and limitations. Last, but not least, we show that, in principle, MCP offers the possibility to test for untestable hypotheses.