Bayesian nonparametric inference for the overlap coefficient: With an application to disease diagnosis

Diagnostic accuracy is a central concern in medical testing. Traditional measures like sensitivity and specificity provide partial insights, but they do not capture the full overlap between diseased and nondiseased populations. The overlap coefficient offers a more comprehensive summary of diagnostic accuracy by quantifying the shared area between two distributions. Prior research has shown that this coefficient is useful for comparing diagnostic tests and evaluating their performance across different populations. However, estimating the overlap coefficient in a flexible and robust way remains a challenge. Existing methods often rely on parametric assumptions that may not hold in real-world data. That uncertainty drove the development of new nonparametric approaches. No prior work had resolved the issue of estimating the overlap coefficient while accounting for covariates in a Bayesian framework. This gap motivated the exploration of Dirichlet process mixture models to provide more accurate and flexible estimates.

This study aims to develop a new statistical framework for estimating the overlap coefficient using Bayesian nonparametric methods. The primary objective is to introduce two estimators based on Dirichlet process mixtures that do not require strong parametric assumptions. The researchers propose a general approach that can be applied to various diagnostic tests and populations. A secondary goal is to extend the overlap coefficient to account for covariates, allowing for more nuanced assessments of diagnostic accuracy. The authors seek to provide a flexible and robust alternative to traditional parametric methods. The study also aims to validate the proposed estimators through simulation and real-world applications. The ultimate purpose is to improve the reliability and interpretability of diagnostic accuracy measures in clinical practice. The researchers emphasize the need for methods that can adapt to the complexity of medical data.

The authors employ Bayesian nonparametric inference using Dirichlet process mixtures to estimate the overlap coefficient. This approach allows for flexible modeling of the underlying probability distributions without assuming a specific parametric form. The first estimator is based on a standard Dirichlet process mixture model. The second estimator incorporates additive normal models to improve performance in specific scenarios. The covariate-specific overlap coefficient is estimated using a Bayesian nonparametric approach that integrates additive normal components. A simulation study is conducted to evaluate the empirical performance of these estimators under various conditions. The simulations assess bias, variance, and coverage of the proposed methods. The researchers also apply their methods to two real-world datasets: one related to ovarian cancer biomarkers and another involving diabetes diagnosis. The study compares the proposed estimators with existing parametric and nonparametric approaches.

The simulation study reveals that the proposed Bayesian nonparametric estimators perform well in terms of accuracy and robustness. The first estimator shows lower bias and higher coverage compared to traditional parametric methods. The second estimator, which uses additive normal models, further improves performance in covariate-specific settings. The overlap coefficient estimates are consistent across different sample sizes and distributional assumptions. The covariate-specific overlap coefficient provides more accurate assessments of diagnostic accuracy when age or other factors are considered. In the ovarian cancer example, the proposed method identifies potential biomarkers with higher precision than conventional approaches. The diabetes application demonstrates improved age-specific accuracy of glucose as a diagnostic marker. The results suggest that the Bayesian nonparametric approach is a viable alternative to traditional methods in diagnostic accuracy research.

The authors conclude that the proposed Bayesian nonparametric estimators offer a flexible and reliable approach to estimating the overlap coefficient. The results from simulations and real-world applications support the effectiveness of these methods. The study highlights the advantages of using Dirichlet process mixtures for diagnostic accuracy assessment. The covariate-specific overlap coefficient allows for more detailed and accurate evaluations of diagnostic tests. The authors emphasize the importance of nonparametric methods in capturing the complexity of medical data. The proposed approach provides a robust alternative to traditional parametric techniques. The study demonstrates that the Bayesian framework can accommodate various diagnostic scenarios and populations. The authors suggest that these methods can be used to improve the interpretation and comparison of diagnostic tests in clinical practice.