On the “Where” of Social Experiments: Selecting More Representative Samples to Inform Policy

Most social experiments are conducted in samples of sites that are not formally representative of the population of policy interest. These studies may produce impact estimates that are unbiased for the sample but biased for the population from which the sample was selected. Recent research has estimated the bias associated with nonrandom inclusion. Although some research has focused on solutions to the problem at the design stage or the analysis stage, research on ways to address this problem is still sparse. This paper provides four recommendations to help researchers obtain more representative samples in impact studies. The fundamental challenge is that, in most impact studies, sites are not required to participate if selected. Therefore, obtaining a sample that adequately represents the population of policy interest can be difficult, and the resulting impact estimates may suffer from external validity bias. The recommendations in this chapter address this challenge to help researchers obtain more representative samples when obtaining a perfectly representative sample is not possible. The recommendations, which are based on standard survey sampling methods, demonstrate that researchers can take practical steps to obtain impact estimates that are more generalizable from the study sample to the broader population of policy interest—and therefore more relevant for informing policy decisions.