Treatment for natural experiments: How to improve causal estimates using conceptual definitions and substantive interpretations

Abstract

The local average treatment effect (LATE) estimator has enabled a wide range of empirical studies using natural experiments to estimate causal effects from observational data. This empirical literature has overlooked a crucial assumption regarding the definition of the treatment—a part of the stable unit treatment value assumption called the consistency assumption in epidemiology. The consequence of ignoring this assumption has been that results have been misinterpreted and over-generalized. I illustrate these problems using examples from seminal studies of the natural experiments literature and present how to improve definitions in future studies. Correctly interpreted LATEs are much more specific than previously acknowledged but reclaim their internal validity. By providing clear and careful definitions of the treatment and interpretations of the results, future studies will increase their usefulness by presenting the causal effect that is actually estimated.