Code for: Machine Learning About Treatment Effect Heterogeneity: The Case of Household Energy Use

We estimate the heterogeneous impacts of Home Energy Reports, a widely used nudge towards household energy conservation, using causal forests. We estimate a range of treatment effects, from -40 kilowatt-hours per month to +10, with multiple statistical "modes" of response. The modes diverge over time, such that the households that reduced consumption in year one ramp up their reductions in year 2 and 3, while households with minimal responses in year 1 tend towards increased consumption in subsequent years. The most commonly-used household characteristics in our causal forests are pre-treatment electricity consumption and home value, which indicates that these variables in particular have significant predictive power.