Quarterly Journal of Economics, 2013, 128(4):1499-1545

Several results run counter to conventional wisdom: we nd no evidence of informational or behavioral failures thought to reduce program participation, there are large unobserved bene ts and costs that traditional evaluations miss, and realized energy savings are only 58 percent of predictions. In the context of the model, the two programs we study reduce social welfare by $0.17 per subsidy dollar spent, both because subsidies are not well-calibrated to currently-estimated externality damages and because of self-selection induced by subsidies that attract households whose par- ticipation generates low social value. However, the model predicts that perfectly-calibrated subsidies would increase welfare by $2.53 per subsidy dollar, revealing the potential of energy ef ciency programs.