Doing More When You're Running LATE: Applying Marginal Treatment Effect Methods to Examine Treatment Effect Heterogeneity in Experiments
The local average treatment effect (LATE) estimated from an experiment is not globally externally valid if the treatment effect varies across individuals. 脗 The LATE gives the average treatment effect for compliers. 脗 I recover bounds on average treatment effects for always takers and never takers. These bounds can reject global external validity of the LATE. Under stronger assumptions required to identify a marginal treatment effect (MTE) with a discrete instrument, I develop weights to recover average treatment effects for always takers, never takers, and other discrete groups. Using the weighted MTE, I generalize the comparison of OLS to LATE when the treatment effect is heterogeneous. I relate treatment effect heterogeneity to observables, and I develop approaches to extrapolate treatment effects. 脗 I apply these methods to the Oregon Health Insurance Experiment. I find that the treatment effect of insurance on emergency room utilization decreases from always takers to compliers to never takers. Previous ER utilization explains a large share of the treatment effect heterogeneity across these groups. Extrapolation of these treatment effects shows that a different policy could increase or decrease ER utilization, depending on which individuals it induces to gain coverage.脗 脗