Discovering Heterogeneous Treatment Effects in Regression Discontinuity Designs
This work addresses the challenge of discovering treatment effect heterogeneity in causal inference for researchers and practitioners, representing an incremental advancement by adapting existing tree-based methods to regression discontinuity designs.
The paper tackles the problem of identifying heterogeneous treatment effects in regression discontinuity designs by proposing a causal supervised machine learning algorithm that builds an honest regression discontinuity tree to discover relevant pre-treatment covariates without invalidating inference. The method is evaluated through Monte Carlo simulations and applied to uncover heterogeneity in the impact of attending a better secondary school in Romania, though no concrete numerical results are provided in the abstract.
The paper proposes a causal supervised machine learning algorithm to uncover treatment effect heterogeneity in sharp and fuzzy regression discontinuity (RD) designs. We develop a criterion for building an honest ``regression discontinuity tree'', where each leaf contains the RD estimate of a treatment conditional on the values of some pre-treatment covariates. It is a priori unknown which covariates are relevant for capturing treatment effect heterogeneity, and it is the task of the algorithm to discover them, without invalidating inference, while employing a nonparametric estimator with expected MSE optimal bandwidth. We study the performance of the method through Monte Carlo simulations and apply it to uncover various sources of heterogeneity in the impact of attending a better secondary school in Romania.