Joint Modeling of Multiple Longitudinal Biomarkers and Survival Outcomes via Threshold Regression: Variability as a Predictor

Mingyan Yu, Zhenke Wu, Michelle M. Hood, Carrie Karvonen‐Gutierrez, Sioḃán D. Harlow, Michael R. ElliottarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed a new Bayesian joint model that proves biomarker "noise"—the fluctuations often ignored by statisticians—is actually a powerful predictor of health outcomes. By combining multivariate longitudinal tracking with threshold regression, the model successfully predicts the age of natural menopause using hormone variability, outperforming traditional survival models that struggle with non-proportional hazards and censored data.

What Problem Does This Paper Address?

In clinical studies, we often track biomarkers (like hormone levels or blood pressure) over time to predict a major health event (like menopause or a heart attack). Traditionally, researchers focus on the mean trajectory—is the level going up or down? Any jitter or fluctuation around that line is usually dismissed as "nuisance" or measurement error.

However, in biological systems, increased variability often signals a loss of stability or "system dysregulation." This paper tackles three major hurdles in modeling these events:

  1. Variability as Information: Incorporating within-subject fluctuations as a predictor rather than ignoring them.
  2. The "Wall" Problem: Standard survival models (like Cox Proportional Hazards) assume the relative risk between groups stays constant over time. In reality, biological risks often shift as people age.
  3. Vanishing Data: Some biomarkers, like the hormone AMH, eventually drop below a "limit of detection" (LOD). Standard models often fail when data points simply disappear.

Methods

The authors propose a Bayesian Joint Model consisting of two interconnected components:

  • Component 1 (Multivariate Longitudinal): This tracks multiple biomarkers simultaneously using a mixed-effects model. Crucially, it uses a "Tobit-like" imputation method to handle biomarkers that fall below detection limits, treating them as censored values rather than missing data.
  • Component 2 (Threshold Regression): Instead of the standard Cox model, the authors use a First-Hitting Time (FHT) approach based on a Wiener process (a mathematical model of random motion). Imagine a subject’s health status as a particle moving toward a boundary; when the particle hits the boundary, the event (menopause) occurs. This allows for "non-proportional hazards," meaning the risk can change dynamically over time.

The entire system is implemented in Stan using Hamiltonian Monte Carlo, allowing for robust estimation of complex parameters and their uncertainties.

Datasets

The model was evaluated using:

  1. Synthetic Simulations: To prove that the joint model reduces bias and improves accuracy compared to "two-stage" models (where you model biomarkers first and then plug the results into a survival model).
  2. SWAN (Study of Women’s Health Across the Nation): A multi-site longitudinal cohort study. The authors analyzed data from 976 women (6,295 total visits), focusing on two primary hormones:
  3. FSH (Follicle-Stimulating Hormone): Typically increases during the menopausal transition.
  4. AMH (Anti-Müllerian Hormone): Typically declines and often becomes undetectable.

Key Insights and Ideas

  • Noise is a Signal: The most striking finding is that hormone variability is a significant predictor of the Final Menstrual Period (FMP). It’s not just about how high your FSH is, but how much it "bounces around."
  • Opposing Effects of Variability: Interestingly, variability in different hormones points in different directions. Higher variability in FSH is associated with an earlier menopause, while higher variability in AMH is associated with a later menopause.
  • The Superiority of Joint Modeling: The paper demonstrates that modeling the biomarkers and the survival event at the same time is far superior to doing them sequentially. Sequential (two-stage) models tended to produce biased estimates, particularly for the intercept terms.
  • Flexibility via Physics: By using a Wiener process (threshold regression), the model naturally accounts for the fact that some women may experience a "deterioration" or "improvement" in their latent health status before finally hitting the threshold.

Key Results

  • AMH Slope is King: The rate at which AMH declines was the strongest predictor of menopause age. A one-standard-deviation increase in the AMH slope (a slower decline) delayed the median age of menopause from 51.5 to 53.3 years.
  • Quantifiable Variability Impact:
    • Women with high FSH variability reached menopause roughly 0.5 years earlier than those with low variability.
  • Women with high AMH variability reached menopause roughly 0.4 years later than those with low variability.
  • Simulation Success: In tests, the proposed joint model achieved a 95% coverage rate for most parameters, whereas traditional two-stage methods often dropped to 0% coverage for certain coefficients, showing they were consistently and significantly wrong.

Limitations and Future Work

While the model is a significant step forward, the authors note a few areas for expansion:

  • Time-Varying Covariates: The current survival component uses "baseline" versions of the random effects. Future versions could allow the risk to change as the biomarkers evolve in real-time.
  • Non-Gaussian Data: The model assumes biomarker errors follow a normal distribution. In some clinical cases, data might be "heavier-tailed" or discrete, requiring different underlying distributions.
  • Computational Intensity: Bayesian joint models are computationally expensive. Scaling this to dozens of biomarkers or tens of thousands of patients would require further optimization of the sampling algorithms.

Summary updated

Abstract

Longitudinal biomarker data and health outcomes are routinely collected in many studies to assess how biomarker trajectories predict health outcomes. Existing methods primarily focus on mean biomarker profiles, treating variability as a nuisance. However, excess variability may indicate system dysregulations that may be associated with poor outcomes. In this paper, we address the long-standing problem of using variability information of multiple longitudinal biomarkers in time-to-event analyses by formulating and studying a Bayesian joint model. We first model multiple longitudinal biomarkers, some of which are subject to limit-of-detection censoring. We then model the survival times by incorporating random effects and variances from the longitudinal component as predictors through threshold regression that admits non-proportional hazards. We demonstrate the operating characteristics of the proposed joint model through simulations and apply it to data from the Study of Women's Health Across the Nation (SWAN) to investigate the impact of the mean and variability of follicle-stimulating hormone (FSH) and anti-Mullerian hormone (AMH) on age at the final menstrual period (FMP).
BibTeX
@misc{yu2025joint,
  author = {Mingyan Yu and Zhenke Wu and Michelle M. Hood and Carrie Karvonen‐Gutierrez and Sioḃán D. Harlow and Michael R. Elliott},
  title = {{Joint Modeling of Multiple Longitudinal Biomarkers and Survival Outcomes via Threshold Regression: Variability as a Predictor}},
  year = {2025},
  eprint = {2503.24146},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.24146},
  howpublished = {arXiv},
}
Figure