Regression discontinuity designs (RDD) are a common quasi-experimental tool for causal inference, but estimating conditional average treatment effects (CATE) in this setting has been awkward. Earlier Bayesian methods struggle to accommodate CATE, while recent frequentist approaches assume a known basis expansion—a strong specification requirement. A new paper proposes a Bayesian additive regression tree (BART) variant that avoids that requirement.
The model uses linear leaf-level regressions on the running variable and a treatment dummy, including their interaction. The tree structure adaptively partitions covariate space into regions where the slope on the running variable differs appreciably, allowing the model to learn heterogeneous treatment effects near the cutoff. The authors say this yields interpretable Bayesian inference on CATE in the RDD framework.
The work is methodological and does not include empirical applications in the abstract, so its practical performance beyond the proposed framework is not detailed here. The authors are Rafael Alcantara, P. Richard Hahn, and Hedibert F. Lopes; the paper is available on arXiv as 2503.00326.