A hybrid generalized propensity score approach for observational studies

Abstract

The overlap assumption (OA) of the generalized propensity score (GPS) is often violated when the number of treatments increases (Lin et al., 2019), especially when the GPS is estimated using machine learning (ML) algorithms. Although the ML based GPS (ML-GPS) shows better performance than the multinomial logistic regression (MLR) based GPS (MLR-GPS), the ML-GPS frequently suffers from the OA violation, which causes difficulty in estimating average treatment effects (ATE) using weighted estimators. Thus, we propose a hybrid GPS that is easy to implement in practice to overcome the OA violation. The hybrid GPS simply combines the strengths of the ML-GPS and MLR-GPS. We conduct a Monte Carlo simulation to compare MLR-GPS with several hybrid GPS for estimating ATEs in terms of bias and mean squared error (MSE). Results show that, overall, the hybrid GPS performs better than the MLR-GPS.

Keywords:

• Causal inference
• Machine learning
• Observational studies
• Overlap