Econometric vs. Causal Structure-Learning for Time-Series Policy Decisions: Evidence from the UK COVID-19 Policies

Researchers Bruno Petrungaro and Anthony Constantinou have published a significant paper comparing causal discovery methods for time-series data, using the complex real-world example of UK COVID-19 policy decisions. The study addresses a critical gap in causal machine learning, which has traditionally focused on cross-sectional data rather than time-ordered sequences. By evaluating 15 different methods—11 causal ML algorithms and 4 econometric techniques—the researchers provide a comprehensive analysis of how different approaches recover cause-and-effect relationships from temporal data, with implications for policy makers who need to understand the impact of interventions over time.

The key finding reveals a trade-off between methodological approaches: econometric methods offer clearer, rule-based temporal structures that are easier to interpret, while causal ML algorithms explore a wider space of possible graph structures, often producing denser networks that capture more identifiable causal relationships. The researchers have made their work practical by providing code to translate econometric results into bnlearn, the most widely used Bayesian Network library in R. This bridges the gap between traditional econometrics and modern causal ML, potentially improving how analysts model complex systems like pandemic responses, economic policies, or climate interventions where understanding temporal causality is essential.