A flexible, interpretable framework for assessing sensitivity to unmeasured confounding.

Area of Science:

• Biostatistics
• Causal Inference
• Epidemiology

Background:

• Unmeasured confounding and model misspecification are significant challenges in estimating causal effects.
• Existing methods often struggle to address both issues simultaneously, leading to potential bias in results.

Purpose of the Study:

• To develop a unified semi-parametric sensitivity analysis framework to simultaneously address unmeasured confounding and model misspecification.
• To provide an interpretable method for assessing the impact of unmeasured confounders on causal effect estimates.

Main Methods:

• Incorporation of Bayesian Additive Regression Trees (BART) within a two-parameter sensitivity analysis framework.
• Assessment of the sensitivity of posterior distributions of treatment effects to varying sensitivity parameters.
• Development of open-source software (treatSens package in R) for practical implementation.

Main Results:

• The proposed method effectively assesses sensitivity to unmeasured confounding while limiting modeling assumptions.
• Evaluated through large-scale simulations and real-world application using high blood pressure data (NHANES III).
• Demonstrated an easily interpretable framework for bias assessment in causal inference.

Conclusions:

• The novel approach offers a robust solution for handling unmeasured confounding and model misspecification in causal effect estimation.
• The integrated open-source software facilitates broader application and adoption in statistical research.
• This method enhances the reliability of causal effect estimates in observational studies.