Synthetic Negative Controls: Using Simulation to Screen Large-scale Propensity Score Analyses.

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Propensity scores are standard for controlling confounding in healthcare database studies.
• Comparing large-scale propensity score analyses with differing methods presents challenges.
• Current balance diagnostics lack guidance on variable selection and quantifying residual imbalance impact.

Purpose of the Study:

• To propose a framework supplementing balance diagnostics for comparing large-scale propensity score analyses.
• To screen analyses for bias signals caused by measured confounding.
• To enhance the reliability of healthcare database research.

Main Methods:

• Conducting and reporting results for multiple analytic choices.
• Utilizing balance diagnostics alongside synthetically generated control studies.
• Employing a treatment assignment model to create pseudo-treatment groups for synthetic data generation without outcome simulation.

Main Results:

• The proposed framework effectively screens analyses for potential bias due to measured confounding.
• Synthetic negative control studies approximate the study population's confounding structure under the null.
• The method is illustrated through simulations and an empirical example, demonstrating its utility.

Conclusions:

• The framework provides a robust method for comparing propensity score analyses and identifying potential bias.
• It addresses limitations of traditional balance diagnostics in large-scale studies.
• This approach enhances the validity and interpretability of findings from healthcare database research.