Using Machine Learning to Target Treatment: The Case of Household Energy Use
We use causal forests to evaluate the heterogeneous treatment effects (TEs) of repeated behavioral nudges towards household energy conservation. The average response is a monthly electricity reduction of 9 kilowatt-hours (kWh), but the full distribution of responses ranges from -30 to +10 kWh. Selective targeting of treatment using the forest raises social net benefits by 12-120 percent, depending on the year and welfare function. Pre-treatment consumption and home value are the strongest predictors of treatment effect. We find suggestive evidence of a "boomerang effect": households with lower consumption than similar neighbors are the ones with positive TE estimates.
Leila Safavi and Paula Meloni provided outstanding research assistance. We thank Hunt Allcott and seminar participants at Carnegie Mellon, UC Berkeley, University of Connecticut, Yale, and MIT for valuable feedback. Alberto Abadie, Jonathan Davis, Peter Christensen, Stefan Wager, and Susan Athey gave valuable advice on implementing the causal forest algorithm. This research would not be possible without the work of Amy Findlay and colleagues at Eversource, who supplied the necessary data and background on the Home Energy Report Program. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Bureau of Economic Research.