Association for Public Policy Analysis & Management

The U.S. has spent billions of dollars to evaluate the impacts of policies and interventions by deploying large social experiments on important questions in health, education, and social policy, among others. Much of this evidence is focused on assessing average causal effects of the treatment/intervention. To the extent that heterogeneous effects are examined, the approaches are often simple and ad hoc. Left unexplored is the possibility of complex effect heterogeneity. New methods that combine policy evaluation and machine learning algorithms have been developed that allow a principled, rigorous, and more complete assessment of possible heterogeneity. This paper uses these new methods on the large scale class size reduction program in Tennessee from the mid 1980s, Project STAR, and shows evidence of complex interactions in the effects of class size reduction based on teacher, student, and school characteristics.