Association for Public Policy Analysis & Management

External validity is a fundamental concern when we use social experiment in policy evaluation and is usually assessed by comparing the experiment to other data. This paper applies recently developed marginal treatment effect (MTE) method (Bjorklund and Moffitt, 1987; Heckman and Vytlacil 1999, 2005, 2007) to examine whether the treatment effect that the social experiment identifies will be externally valid in other contexts. In many experiments, individuals assigned to the control group can decide to receive treatment, and treatment group members can decide to drop out of the intervention. Following the insights of Kowalski (2016), we utilize the information from the endogenous post-randomization selection into treatment to examine treatment effect heterogeneity within the experiment, which informs the external validity of the experimental estimates. Building on the local average treatment effect (LATE) framework and the estimates of average treatment effects for compliers, always takers, and never takers, we examine the external validity of a large scale self-employment training demonstration, Project GATE, by analyzing the treatment effect heterogeneity within the experiment. This paper also bridges the literatures of endogenous subgroup analysis and external validity, utilizing the information on selection into the treatment post-randomization, which is within a social experiment, to inform if the treatment effect will likely to vary across different contexts.