WorksEconomics, Econometrics and Finance
Informative Local Average Treatment Effects Using Propensity Scores
Abstract
In instrumental variable (IV) models, comparisons of conditional local average treatment effects across covariates can mix effect heterogeneity in observed covariates with changes in the complier population, that is, those whose treatment responds to the instrument. Compliers at different covariate values may differ in unobserved characteristics that may also drive effect heterogeneity. We propose informative conditional local average treatment effects (ICLATEs), which condition on a subset of covariates of interest and on the treatment propensity scores under both values of a binary instrument. Fixing these scores holds constant the range of unobservables that drive treatment selection and thus compliance, while the composition of the covariates not conditioned on may change. Under conditional IV assumptions and suitable support conditions, we establish nonparametric identification and characterize the relationship of the ICLATEs to marginal treatment effects. We distinguish effect homogeneity in the unobservables driving treatment selection from effect homogeneity across excluded covariates. We clarify their roles in extrapolating complier effects to population effects and show that, jointly, they yield a testable implication even with a single binary instrument. For estimation, we combine cross-fitted machine learning of the nuisance functions with pooled local linear regression of doubly robust scores. Under undersmoothing and regularity conditions, we establish pointwise asymptotic normality, which a simulation study supports. As an empirical illustration, we apply the method to the Job Corps program to assess how the earnings effects of training participation vary with age.