Using Machine Learning to Target Treatment: The Case of Household Energy Use

We test the ability of causal forests to improve, through selective targeting, the effectiveness of a randomized program providing repeated behavioral nudges towards household energy conservation. The average treatment effect of the program is a monthly electricity reduction of 9 kilowatt-hours (kWh), but the full distribution of predicted reductions ranges from roughly 1 to 33 kWh. Pre-treatment electricity consumption and home value are the strongest predictors of differential treatment effects. In a pair of targeting exercises, use of the causal forest increases social net benefits of the nudge program by a factor of 3-5 relative to the status quo. Using earlier program waves to target in later ones, we estimate that the forest produces more benefits than five other alternative predictive models. Bootstrapping to generate confidence intervals, we find the forest’s advantage to be statistically significant relative to some but not all of these alternatives.