Targeting Treatment Assignment by combining Behavioral Measurement and Machine Learning

Abstract

Randomized experiments are now standard in data-rich environments, yet average treatment effects often provide limited guidance when responses are heterogeneous and interventions can be deployed selectively. We study how behavioral economics and machine learning can be combined to construct effective and interpretable targeting rules. In a large field experiment at a major German online fashion retailer ( 500,000 visitors), a loss-framed discount message has essentially zero average effect on purchases and returns. We therefore implement a nested incentivized measurement study (N=651) to elicit individual loss aver- sion, a mechanism-consistent moderator for loss framing, and train a prediction model to impute loss-aversion categories from commonly observed digital footprints at scale. Targeting exposure to visitors with high predicted loss aversion yields statistically significant revenue gains, while indiscriminate treatment and predictive targeting based on revenue levels do not. Benchmarks using causal forests identify heterogeneity but deliver smaller, less robust improvements. The results illustrate how combining behavioral economics and machine learning can improve both the performance and interpretability of algorithmic treatment assignment.