Semiparametric Joint Inference for Sensitivity and Specificity at the Youden-Optimal Cut-Off

Siyan Liu, Qinglong Tian, Chunlin Wang, Pengfei LiarXiv 2026 · DOI · arXiv

This paper, recently published by researchers from the University of Waterloo and Xiamen University, introduces a more efficient way to measure the accuracy of medical diagnostic tests. By using a semiparametric approach, the authors provide a method to jointly estimate sensitivity and specificity that is more robust than traditional models and more precise than purely data-driven ones.

TL;DR

The authors propose a semiparametric framework using the Density Ratio Model (DRM) to jointly estimate sensitivity and specificity at the Youden-optimal cut-off. This method outperforms existing parametric and nonparametric techniques by providing tighter, more accurate confidence regions, especially in moderate sample sizes.

What Problem Does This Paper Address?

In medical diagnostics, biomarkers (like blood sugar or antibody levels) are often continuous values. To make a clinical decision, doctors must choose a "cut-off" point to classify a patient as "diseased" or "healthy." The Youden Index is a popular tool for finding this "optimal" threshold by maximizing the sum of sensitivity (true positive rate) and specificity (true negative rate).

However, there is a statistical catch: because the optimal cut-off is estimated from the same data used to calculate sensitivity and specificity, the two measures become statistically dependent. Most existing methods for calculating confidence regions for these measures are either:

  1. Parametric: They assume the data follows a specific distribution (like Normal). If this assumption is wrong, the results are biased.
  2. Nonparametric: they make no assumptions. While "safe," these methods often require very large samples to be precise and can be numerically unstable.

The authors aim to bridge this gap with a "best of both worlds" semiparametric approach.

Methods

The core of the paper is the Density Ratio Model (DRM). Instead of assuming a specific distribution for both the healthy ($f_0$) and diseased ($f_1$) populations, the DRM only models the ratio between the two densities: $$f_1(x) = \exp\{\alpha + \beta^\top \mathbf{q}(x)\} f_0(x)$$ Here, $f_0$ is left completely unspecified (the nonparametric part), while the relationship between the two groups is modeled via a parametric component.

Key Technical Contributions: * Maximum Empirical Likelihood (MEL): The authors use MEL to estimate the parameters. This allows them to leverage information from both the healthy and diseased samples simultaneously to gain efficiency. * Optimal Cut-off Estimation: They show that under the DRM, finding the optimal cut-off $c$ simplifies to solving the equation $\hat{\alpha} + \hat{\beta}^\top \mathbf{q}(x) = 0$. * Joint Asymptotic Normality: The paper mathematically proves that the estimators for sensitivity and specificity follow a joint normal distribution as the sample size grows. * Logit Transformation: To ensure that confidence regions don't suggest impossible values (like a sensitivity of 1.1), they apply a logit transformation to keep the results within the [0, 1] range.

Datasets

The researchers evaluated their method using:

  1. Synthetic Simulations: They tested the model against Lognormal, Gamma, and Beta distributions to see how it handled different shapes and skews.
  2. COVID-19 Antibody Data: A real-world dataset consisting of IgG antibody titers from 100 healthcare workers (42 infection-positive, 58 infection-negative) following vaccination.

Key Insights and Ideas

  • The "Aha" Moment: The authors realized that the Youden-optimal cut-off is exactly the point where the density ratio equals one. This makes the DRM a mathematically "natural" fit for this specific problem.
  • Efficiency through Structure: By assuming a structural relationship (the ratio) between the two groups, the model "borrows strength" from the entire pooled dataset. This leads to much smaller confidence regions than fully nonparametric methods, which treat the two groups as entirely unrelated.
  • Robustness to Skew: Unlike common parametric methods that rely on Box-Cox transformations, the DRM handled highly skewed Beta distributions without significant bias.

Key Results

  • Higher Precision: In simulations, the proposed method consistently achieved the smallest Mean Squared Error (MSE) compared to five other state-of-the-art methods.
  • Tighter Confidence Regions: In the COVID-19 data analysis, the proposed method's 95% confidence region was significantly smaller (Area: 3.889) than the fully nonparametric empirical likelihood method (Area: 40.605).
  • Better Coverage: The method maintained a 95% coverage probability (the "true" value fell inside the estimated region 95% of the time) even in scenarios where other methods failed as the Youden index increased.

Limitations and Future Work

  • Basis Function Selection: The user must still choose a "basis function" $\mathbf{q}(x)$ (e.g., $x$ or $\log(x)$). While the authors used AIC/BIC criteria to select this successfully, it adds a step of model selection.
  • Multi-class Extension: This paper focuses on the binary case (Healthy vs. Diseased). The authors note that extending this to multi-class problems (e.g., Healthy vs. Early Stage vs. Late Stage) is a promising direction for future research, as diagnostic accuracy then involves more complex ROC surfaces.

Summary updated

Abstract

Sensitivity and specificity evaluated at an optimal diagnostic cut-off are fundamental measures of classification accuracy when continuous biomarkers are used for disease diagnosis. Joint inference for these quantities is challenging because their estimators are evaluated at a common, data-driven threshold estimated from both diseased and healthy samples, inducing statistical dependence. Existing approaches are largely based on parametric assumptions or fully nonparametric procedures, which may be sensitive to model misspecification or lack efficiency in moderate samples. We propose a semiparametric framework for joint inference on sensitivity and specificity at the Youden-optimal cut-off under the density ratio model. Using maximum empirical likelihood, we derive estimators of the optimal threshold and the corresponding sensitivity and specificity, and establish their joint asymptotic normality. This leads to Wald-type and range-preserving logit-transformed confidence regions. Simulation studies show that the proposed method achieves accurate coverage with improved efficiency relative to existing parametric and nonparametric alternatives across a variety of distributional settings. An analysis of COVID-19 antibody data demonstrates the practical advantages of the proposed approach for diagnostic decision-making.
BibTeX
@misc{liu2026semiparametric,
  author = {Siyan Liu and Qinglong Tian and Chunlin Wang and Pengfei Li},
  title = {{Semiparametric Joint Inference for Sensitivity and Specificity at the Youden-Optimal Cut-Off}},
  year = {2026},
  eprint = {2602.23045},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.23045},
}
Figure