Bayesian Posterior Interval Calibration to Improve the Interpretability of Observational Studies

Area of Science:

• Health data science
• Biostatistics
• Causal inference

Background:

• Observational healthcare data can estimate medical product causal effects but suffer from systematic errors.
• Standard confidence intervals and p-values from observational studies do not account for systematic error.
• This leads to unreliable operating characteristics, hindering valid interpretation.

Purpose of the Study:

• To develop a Bayesian statistical procedure for posterior interval calibration.
• To address systematic error in observational studies using negative and positive controls.
• To restore nominal statistical characteristics, such as confidence interval coverage.

Main Methods:

• Proposed a Bayesian statistical procedure for posterior interval calibration.
• Utilized negative controls (falsification hypotheses) and positive controls.
• Adjusted confidence intervals and p-values based on detected bias from control analyses.

Main Results:

• The posterior interval calibration procedure successfully restored nominal operating characteristics.
• Demonstrated restoration of 95% coverage of the true effect size by the 95% posterior interval.
• Indicated improved reliability and interpretability of results from observational studies.

Conclusions:

• The proposed Bayesian calibration method effectively addresses systematic error in observational data.
• This approach enhances the validity of causal effect estimations for medical products.
• Restoring nominal characteristics ensures more trustworthy confidence intervals and p-values.