A Bayesian Approach to Estimate Causal Average Treatment Effects Under Unmeasured Confounding.

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Causal Inference

Background:

• Unmeasured confounding is a significant bias source in clinical trials, impacting causal inference.
• Existing methods struggle to accurately estimate causal effects when confounders are not measured.

Purpose of the Study:

• To propose a practical Bayesian modeling approach to adjust for unmeasured confounding.
• To obtain precise causal average treatment effect estimates in two-arm randomized controlled clinical trials.

Main Methods:

• Developed an innovative Bayesian modeling approach incorporating unmeasured confounders.
• Utilized model reparameterization to address non-identifiability issues.
• Implemented an iterative algorithm for robust inference and prior sensitivity analysis.

Main Results:

• The proposed approach effectively adjusts for unmeasured confounding effects.
• Achieved robust average treatment effect estimates with correct statistical significance conclusions.
• Demonstrated efficacy using a real clinical data example, even without adjusting for measured confounders.

Conclusions:

• The Bayesian method offers a practical solution for unmeasured confounding in clinical trials.
• The approach is generalizable to various study designs and challenging data collection scenarios.
• Enhances the reliability of causal inference in the presence of unmeasured variables.