External validity is a fundamental challenge in treatment effect estimation. Even when researchers credibly identify average treatment effects - for example through randomized experiments - the results may not extrapolate to the population of interest for a given policy question. If the population and sample differ only in the distribution of observed variables this problem has a well-known solution: reweight the sample to match the population. In many cases, however, the population and sample differ along dimensions unobserved by the researcher. We provide a tractable framework for thinking about external validity in such cases. Our approach relies on the fact that when the sample is drawn from the same support as the population of interest there exist weights which, if known, would allow us to reweight the sample to match the population. These weights are larger in a stochastic sense when the sample is more selected, and their correlation with a given variable reflects the intensity of selection along this dimension. We suggest natural benchmarks for assessing external validity, discuss implementation, and apply our results to data from several recent experiments.
↧