Abstract:

“Using Machine Learning to Better Target Energy Conservation”
Christopher Knittel (Massachusetts Institute of Technology) and Samuel Stolper (University of Michigan)

We use machine learning to evaluate the heterogeneous treatment effects of Home Energy Reports, a widely-used behavioral nudge that relies on repeated social comparisons to encourage household energy conservation. The average treatment effect 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 based on the machine-learning results raises social net benefits by 12-120 percent, depending on the year and the welfare function used. Pre-treatment consumption and home value are the strongest predictors of treatment effect. We find suggestive evidence of a “boomerang effect”: the households that we predict to have raised their consumption in response to treatment are overwhelmingly likely to have used less energy than their (similar) neighbors to begin with.