|
Abstract:
|
The regression discontinuity design (RDD) is a popular technique for estimating causal effects from observational data by exploiting changes in treatment assignment according to hard thresholds of a continuous variable. However, identifying candidate RDDs is a largely manual process requiring domain knowledge, so it would be useful to have data-driven methods for finding interpretable RDDs. Moreover, disparities could exist in how different sub-groups are assigned treatments – for instance, pain thresholds for treatment referral could vary, implicitly and often unfairly, by gender or ethnicity. The authors propose a method for an automated multi-dimensional RDD discovery through a tree-based partitioning approach, and validate its effectiveness through simulated and real claims data.
|
