Bayesian Sensitivity Analysis for Causal Estimation With Time-Varying Unmeasured Confounding

Area of Science:

• Statistics
• Epidemiology
• Biostatistics

Background:

• Causal inference requires the untestable assumption of no unmeasured confounding.
• Sensitivity analysis is crucial for quantifying the impact of unmeasured confounding on causal estimates.
• Existing methods like latent confounder and sensitivity function approaches have limitations.

Purpose of the Study:

• To develop and extend Bayesian sensitivity analysis methods for time-varying treatment effects.
• To address time-varying unmeasured confounding in longitudinal observational data.
• To provide practical guidance on implementing these advanced sensitivity analysis techniques.

Main Methods:

• Developed Bayesian sensitivity analysis with latent confounding variables.
• Extended the Bayesian sensitivity function approach.
• Applied methods to longitudinal observational data with time-varying unmeasured confounding.
• Conducted simulation studies to evaluate performance.

Main Results:

• The developed Bayesian methods effectively estimate time-varying treatment effects under unmeasured confounding.
• Simulation studies demonstrated the robustness and performance of the proposed approaches.
• Application to a pediatric disease registry provided practical insights.

Conclusions:

• The extended Bayesian sensitivity analysis methods offer a robust framework for causal inference with unmeasured confounding.
• These methods are valuable for analyzing longitudinal observational data in complex health research.
• The study provides practical guidance for implementing sensitivity analysis in real-world settings.