Hierarchical priors for bias parameters in Bayesian sensitivity analysis for unmeasured confounding

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Unmeasured confounding poses a significant challenge in observational studies, potentially biasing results.
• Existing Bayesian methods for adjusting unmeasured confounding often require complex prior elicitation for high-dimensional parameters, limiting their practical application.

Purpose of the Study:

• To propose a novel Bayesian methodology for adjusting unmeasured confounding in observational studies with binary covariates.
• To develop a method that derives default priors for bias parameters, simplifying the adjustment process for users.

Main Methods:

• The proposed methodology treats confounding effects of measured and unmeasured variables as exchangeable within a Bayesian framework.
• A log-linear model with pairwise interactions is used to model the joint distribution of covariates.
• Hierarchical priors are employed to constrain the magnitude and direction of bias parameters, with the unmeasured confounder's conditional distribution following a logistic model.

Main Results:

• The developed method offers a simplified approach to adjusting for unmeasured confounding by providing default priors.
• The conditional distribution of the unmeasured confounder demonstrates equivalence with previously proposed methods.
• Application in a pharmacoepidemiology data example illustrates the method's utility and the impact of prior choices.

Conclusions:

• The novel Bayesian methodology effectively addresses unmeasured confounding by simplifying prior specification.
• This approach enhances the practicality and accessibility of bias adjustment techniques in observational research.
• The method provides a valuable tool for researchers, particularly in fields like pharmacoepidemiology, for more robust causal inference.