Addressing validity and generalizability concerns in field experiments

Abstract

In this paper, we systematically analyze the empirical importance of standard conditions for the validity and generalizability of field experiments: the internal and external overlap and unconfoundedness conditions. We experimentally varied the degree of overlap in disjoint sub-samples from a recruitment experiment with more than 3,000 public schools, mimicking small scale field experiments. This was achieved by using different techniques for treatment assignment. We applied standard methods, such as pure randomization, and the novel minMSE treatment assignment method. This new technique should achieve improved overlap by balancing covariate dependencies and variances instead of focusing on individual mean values. We assess the relevance of the overlap condition by linking the estimation precision in the disjoint sub-samples to measures of overlap and balance in general. Unconfoundedness is addressed by using a rich set of administrative data on institution and municipality characteristics to study potential self-selection. We find no evidence for the violation of unconfoundedness and establish that improved overlap, and balancedness, as achieved by the minMSE method, reduce the bias of the treatment effect estimation by more than 35% compared to pure randomization, illustrating the importance of, and suggesting a solution to, addressing overlap also in (field) experiments.