Testing and interpreting the effectiveness of causal machine learning---an economic theory approach

Abstract

This paper demonstrates how causal machine learning (CML) can personalize treatment assignments (targeting) to improve intervention effectiveness. In a field experiment with nearly 500,000 participants at an online fashion retailer, we show that CML-based targeting transforms an otherwise ineffective loss-framing intervention into one that generates an 11\% revenue increase. Effective targeting is achieved using generic digital footprints without relying on context-specific historical data.

By combining data from the RCT with a behavioral measurement experiment, we find that CML-predicted individual treatment effects are strongly correlated with individual loss aversion, a core concept in behavioral economics. This alignment shows that CML implicitly captures established theoretical constructs, enhancing both the interpretability and transparency of its outputs. Furthermore, CML outperforms targeting policies based directly on measured loss aversion, demonstrating its ability to uncover heterogeneity beyond existing models.

Keywords: Behavioral Measurement, Loss Aversion, Causal Machiene Learning