An overview of propensity score matching methods for clustered data

Area of Science:

• Epidemiology
• Biostatistics
• Observational Studies

Background:

• Propensity score matching (PSM) is vital for causal inference in observational studies.
• Clustered data present unique challenges for standard PSM methods.
• Existing methods struggle to balance measured and unmeasured confounders effectively in clustered settings.

Purpose of the Study:

• To provide an overview of PSM techniques for clustered data.
• To explore the utility of machine learning for estimating propensity scores in clustered data.
• To compare the performance of various PSM methods for clustered data.

Main Methods:

• Overview of PSM for clustered observational data.
• Application of generalized boosted models for propensity score estimation.
• Simulation studies comparing PSM methods under different clustering scenarios.
• Illustrative analysis using aspirin's effect on hearing deterioration.

Main Results:

• PSM can account for measured and unmeasured cluster-level confounders.
• Machine learning methods (e.g., GBM) for propensity scores can degrade with high clustering.
• Alternative strategies combining PSM with fixed-effects models can mitigate performance issues.
• Simulation results highlight the performance variations of different PSM approaches.

Conclusions:

• PSM is adaptable for clustered data, offering control over various confounders.
• Careful consideration of methods is needed when using machine learning with highly clustered data.
• Hybrid approaches may enhance causal effect estimation in complex observational studies.
• The study provides practical insights for analyzing clustered observational data.