Designing Randomized Experiments to Predict Unit-Specific Treatment Effects

When evaluating a program or policy, a randomized experiment is typically designed to test a single confirmatory hypothesis about the average treatment effect, although subgroup and moderator effects may also be explored. The resulting average treatment effect estimate is then reported in research clearinghouses and used to inform policy and practice decisions for units not in the study. This use suggests that the purpose of these randomized trials is not only the testing of hypotheses, but rather the prediction of treatment effects for a broad set of units in a population. In this article, we consider the optimal design of a randomized experiment focused on the prediction of unit-specific effects. We consider how different sampling processes and models affect the mean squared error of these predictions. The results indicate, for example, that problems of generalizability—differences between study samples and target populations—can greatly increase prediction error. We also identify the conditions under which the best unit-specific treatment effect is the average treatment effect estimate. Throughout, we use simple regression models to connect the predictive and hypothesis testing literatures and to provide implications for the design of randomized experiments.