Use of Synthetic Data in Dealing With Self-Selection

Because energy conservation programs are generally voluntary, participating households are different from nonparticipants in important, energy-related ways. This self-selection bias complicates efforts to estimate energy savings due to these programs. This article discusses several methods for dealing with self-selection. The choices include nonrandom sampling of program nonparticipants, binary choice models that explicitly treat house hold decisions to participate and to retrofit, or use of both methods. Because some of the methods discussed are new and have not yet been applied to analysis of energy conserva tion programs, we developed a "synthetic "data set. We conducted numerical experiments with this data to examine the performance of these different methods. These experiments show that the improved sample design and analytical techniques generally yield more accurate estimates of program energy savings. Our experience also suggests that a small, well-defined synthetic data set is helpful in developing, debugging, and evaluating soft ware associated with new analytical approaches.