A Bayesian approach to estimate and validate the false negative fraction in a two-stage multiple screening test

Area of Science:

• Biostatistics
• Diagnostic Test Evaluation
• Medical Informatics

Background:

• Gold standard procedures are essential for diagnostic test accuracy but often infeasible for all subjects.
• Partial verification, where only a subset undergoes gold standard testing, presents a challenge in estimating error rates like the false negative fraction (FNF).
• Accurate estimation of FNF is vital for understanding diagnostic kit performance and clinical utility.

Purpose of the Study:

• To estimate the false negative fraction (FNF) in a two-stage screening test under partial verification using beta-binomial and Bayesian logistic models.
• To validate and compare the predictive performance of these two statistical models for diagnostic accuracy estimation.

Main Methods:

• Employed a Bayesian approach to estimate the FNF using both beta-binomial and Bayesian logistic models.
• Validated the models by assessing their out-of-sample predictive capabilities on independent data.

Main Results:

• The beta-binomial model yielded a median posterior estimate of FNF at 26.4% (95% CI: 0.123-0.650).
• The Bayesian logistic model provided a median posterior estimate of FNF at 23.3% (95% CI: 0.124-0.375).
• Model validation indicated that the beta-binomial model demonstrated slightly superior predictive accuracy compared to the Bayesian logistic model.

Conclusions:

• Bayesian statistical approaches are effective for estimating the false negative fraction (FNF) in diagnostic evaluations with partial verification.
• Comparing out-of-sample predictive performance is a robust method for validating and selecting the best-fitting statistical model.