unmconf

unmconf implements Bayesian multilevel regression models that account for one or two unmeasured confounders from the normal or binomial distribution, supporting normal, binary, Poisson, or gamma response types for probabilistic sensitivity analysis in observational studies.

Key Features:

• Bayesian unmeasured confounding model: Fits Bayesian multilevel regression models that incorporate unmeasured confounders into the analysis framework.
• Probabilistic sensitivity analysis: Performs probabilistic sensitivity analysis to investigate the impact of unmeasured confounding, in contrast to deterministic sensitivity analysis methods.
• Multiple response distributions: Supports normal, binary, Poisson, and gamma response types for flexible modeling across different data types.
• Multiple confounder distributions: Accounts for one or two unmeasured confounders drawn from normal or binomial distributions.
• Validation data integration: Accommodates varying levels of internal and external validation data for improved inference.

Scientific Applications:

• Observational study bias correction: Reduces bias in parameter estimates and improves uncertainty assessments when unmeasured confounders are present in observational studies.
• Causal inference sensitivity analysis: Investigates the robustness of causal conclusions to potential unmeasured confounding through probabilistic methods.
• Epidemiological and clinical research: Enables valid Bayesian modeling in studies where confounders may be unobserved or unrecorded.

Methodology:

unmconf uses Bayesian multilevel regression with Markov chain Monte Carlo computation performed on the back end. The package was validated through novel simulation studies demonstrating that credible intervals achieve near nominal coverage probability with smaller bias when modeling unmeasured confounders.