Rank‐based estimation of propensity score weights via subclassification

Abstract Propensity score (PS) weighting estimators are widely used for causal effect estimation and enjoy desirable theoretical properties, such as consistency and potential efficiency under correct model specification. However, their performance can degrade in practice due to sensitivity to PS model misspecification. To mitigate this, we draw inspiration from subclassification methods, which are often more robust to such misspecification. We first reveal an intrinsic connection between the seemingly distinct weighting and subclassification approaches. Leveraging this connection, we construct robust PS weights via subclassification and propose full classification weights and associated estimators for causal effect estimation in observational studies. These estimators retain consistency while gaining robustness to model misspecification, thereby combining the strengths of both frameworks. Numerical studies demonstrate that our methods perform favourably against existing alternatives.