Bayesian Inference for Joint Tail Risk in Paired Biomarkers via Archimedean Copulas with Restricted Jeffreys Priors

We propose a Bayesian copula-based framework to quantify clinically interpretable joint tail risks from paired continuous biomarkers. After converting each biomarker margin to rank-based pseudo-observations, we model dependence using one-parameter Archimedean copulas and focus on three probability-scale summaries at tail level $α$: the lower-tail joint risk $R_L(θ)=C_θ(α,α)$, the upper-tail joint risk $R_U(θ)=2α-1+C_θ(1-α,1-α)$, and the conditional lower-tail risk $R_C(θ)=R_L(θ)/α$. Uncertainty is quantified via a restricted Jeffreys prior on the copula parameter and grid-based posterior approximation, which induces an exact posterior for each tail-risk functional. In simulations from Clayton and Gumbel copulas across multiple dependence strengths, posterior credible intervals achieve near-nominal coverage for $R_L$, $R_U$, and $R_C$. We then analyze NHANES 2017–2018 fasting glucose (GLU) and HbA1c (GHB) ($n=2887$) at $α=0.05$, obtaining tight posterior credible intervals for both the dependence parameter and induced tail risks. The results reveal markedly elevated extremal co-movement relative to independence; under the Gumbel model, the posterior mean joint upper-tail risk is $R_U(α)=0.0286$, approximately $11.46\times$ the independence benchmark $α^2=0.0025$. Overall, the proposed approach provides a principled, dependence-aware method for reporting joint and conditional extremal-risk summaries with Bayesian uncertainty quantification in biomedical applications.

💡 Research Summary

This paper introduces a Bayesian copula‑based framework for quantifying joint tail risk of two continuous biomarkers, a problem of growing importance in clinical and epidemiological research where simultaneous extreme values often signal heightened disease risk. The authors first remove marginal effects by converting each biomarker to rank‑based pseudo‑observations, thereby ensuring that inference focuses solely on dependence. Dependence is modeled with a one‑parameter Archimedean copula—specifically the Clayton copula for lower‑tail dependence and the Gumbel copula for upper‑tail dependence.

Three probability‑scale tail‑risk functionals are defined at a pre‑specified tail level α (e.g., 0.05):

• Lower‑tail joint risk (R_L(θ)=C_θ(α,α)), the probability that both biomarkers fall below their α‑quantiles.
• Upper‑tail joint risk (R_U(θ)=2α-1+C_θ(1-α,1-α)), the probability that both exceed their (1‑α)‑quantiles.
• Conditional lower‑tail risk (R_C(θ)=R_L(θ)/α), the probability that the second biomarker is also in the lower tail given that the first is.

These summaries are directly interpretable for clinicians: they describe the chance of simultaneous extreme readings and the conditional chance of one extreme given the other.

For Bayesian inference, the authors adopt a restricted Jeffreys prior on the copula parameter θ. The Jeffreys prior, derived from the square root of the Fisher information determinant, provides a non‑informative baseline, while the restriction (θ>0) respects the natural domain of Archimedean copulas. This prior yields a proper posterior that can be evaluated exactly on a finite grid of θ values. The posterior is approximated by evaluating the likelihood at each grid point, multiplying by the prior density, and normalizing across the grid. Because the risk functionals are monotone transformations of θ, the posterior distribution of each functional can be obtained analytically from the grid‑based posterior of θ, avoiding any additional Monte‑Carlo error.

Simulation studies were conducted for both Clayton and Gumbel copulas across three dependence strengths (weak, moderate, strong) and two tail levels (α=0.05, 0.10). For each scenario, 1,000 data sets of size n=500 were generated. The 95 % credible intervals for (R_L), (R_U), and (R_C) achieved empirical coverage rates of 94–96 %, confirming that the Bayesian procedure provides reliable uncertainty quantification even in the extreme tails where data are scarce.

The methodology was then applied to real data from the 2017‑2018 National Health and Nutrition Examination Survey (NHANES). The authors examined fasting glucose (GLU) and glycated hemoglobin (HbA1c, GHB) in 2,887 adults. Using α=0.05, they fitted both Clayton and Gumbel copulas. The Gumbel model, which captures upper‑tail dependence, yielded a posterior mean θ≈2.3 (95 % CI 1.8–2.9), indicating a pronounced tendency for high glucose and high HbA1c to co‑occur. Consequently, the posterior mean upper‑tail joint risk was (R_U=0.0286), roughly 11.5 times larger than the independence benchmark α²=0.0025. The lower‑tail joint risk was also elevated relative to independence but to a lesser extent, and the conditional lower‑tail risk (R_C≈0.57) suggests that when one biomarker is low, there is a 57 % chance the other is also low. These findings illustrate that ignoring dependence would dramatically underestimate the probability of simultaneous adverse biomarker profiles.

Key contributions of the paper include:

1. A clean separation of marginal and dependence modeling via rank‑based pseudo‑observations, enabling the use of any marginal distribution without affecting tail‑risk estimates.
2. The introduction of a restricted Jeffreys prior that yields a proper, analytically tractable posterior for the copula parameter.
3. A grid‑based posterior computation that provides exact posterior distributions for the tail‑risk functionals, eliminating the need for MCMC diagnostics.
4. Demonstration of accurate coverage and tight credible intervals in extensive simulations, and a compelling real‑world application that quantifies clinically relevant joint extremal risk.

Limitations are acknowledged. The framework is confined to single‑parameter Archimedean copulas, which may be insufficient for more complex multivariate dependence patterns or for data exhibiting asymmetric tail behavior not captured by Clayton or Gumbel. Moreover, the grid approach, while exact, can become computationally intensive for finer grids or higher‑dimensional extensions. Future work could explore vine copulas or other flexible multivariate constructions, and adopt variational Bayes or adaptive quadrature to retain computational efficiency while expanding model flexibility.

In summary, the authors deliver a principled, Bayesian, dependence‑aware method for reporting joint and conditional extremal‑risk summaries. By providing both point estimates and rigorous uncertainty quantification, the approach equips biomedical researchers and clinicians with a powerful tool for risk assessment when multiple biomarkers are jointly considered.