Improving Generalizations From Experiments Using Propensity Score Subclassification: Assumptions, Properties, and Contexts

As a result of the use of random assignment to treatment, randomized experiments typically have high internal validity. However, units are very rarely randomly selected from a well-defined population of interest into an experiment; this results in low external validity. Under nonrandom sampling, this means that the estimate of the sample average treatment effect calculated in the experiment can be a biased estimate of the population average treatment effect. This article explores the use of the propensity score subclassification estimator as a means for improving generalizations from experiments. It first lays out the assumptions necessary for generalizations, then investigates the amount of bias reduction and average variance inflation that is likely when compared to a conventional estimator. It concludes with a discussion of issues that arise when the population of interest is not well represented by the experiment, and an example.