TL;DR
Researchers have developed a new Bayesian joint model that proves "noise" in biological data isn't just a nuisance—it’s a predictor. By accounting for the individual-level variability (variance) of sex hormones during menopause, rather than just their average levels, the model more accurately predicts changes in body composition while avoiding the statistical biases common in traditional two-stage methods.
What Problem Does This Paper Address?
In longitudinal health studies, researchers typically track biomarkers (like hormones) over time to predict a future health outcome. Most existing statistical tools focus on the mean trajectory—the average "path" a biomarker takes. However, the "wobble" or instability around that path (the subject-level variance) is often ignored or treated as measurement error.
In the context of menopause, hormones like estradiol (E2) and follicle-stimulating hormone (FSH) fluctuate wildly. Emerging evidence suggests this variability might be a marker of biological system instability, potentially predicting health risks like weight gain or bone loss. The challenge is that current methods to use variance as a predictor usually involve "two-stage" estimation: first you estimate the variance, then you plug that estimate into a regression. This approach ignores the statistical uncertainty of the first step, leading to biased results (usually underestimating the effect).
Methods
The authors propose a Joint Model with Individual Variances (JMIV). Unlike standard models, this approach estimates the mean trajectory and the residual variance-covariance matrix simultaneously within a single Bayesian framework.
Key technical highlights include:
- Hierarchical Priors: They use flexible hierarchical priors for the variance-covariance matrices, allowing these values to vary by subject.
- Correlation Modeling: For cases with multiple biomarkers (like E2 and FSH), the model accounts for the correlation between the "noise" of different markers.
- Hamiltonian Monte Carlo (HMC): Because the model lacks a closed-form solution, they implemented it using Stan, a probabilistic programming language that uses HMC to draw samples from the posterior distribution.
- Comparison: They compared JMIV against three common alternatives: Two-Stage Simple Linear Models (TSLM), Two-Stage Linear Mixed Models (TSLMM), and a Two-Stage Individual Variance (TSIV) model.
Datasets
- Simulation Data: The authors generated synthetic datasets with two and three biomarkers to test parameter recovery, bias, and credible interval coverage.
- SWAN (Study of Women’s Health Across the Nation): A real-world longitudinal cohort study. The authors analyzed data from 841 women with a total of 9,902 hormone measurements (E2 and FSH) taken over approximately 13 years, alongside body composition data (fat and lean mass) measured via DXA scans.
Key Insights and Ideas
- Variability is a Signal: The core "aha" moment is that subject-level variance is a meaningful biological predictor. In biological systems, increased variance often precedes a transition or indicates a loss of homeostatic control.
- The "Two-Stage" Trap: The paper demonstrates that two-stage models consistently suffer from "attenuation towards the null." Because they treat estimated variances as "known" facts rather than uncertain estimates, they fail to capture the true relationship between variability and health outcomes.
- Hormonal Coupling: The model revealed an interaction between E2 and FSH. If one hormone is highly variable while the other is stable, it predicts different health outcomes than if both are highly variable, suggesting that the balance of variability between coupled systems matters.
Key Results
- Simulation Performance: The JMIV model achieved over 90% coverage of true parameters in simulations, whereas two-stage models often had coverage as low as 0% to 15% for variance-related coefficients.
- Fat Mass Change: In the SWAN data, larger variability in either E2 or FSH was associated with higher levels of fat mass increase across the menopausal transition. Specifically, a one-unit increase in E2 variability was associated with a $7.03 \times 10^{-3}$ increase in fat mass proportion per year.
- Lean Mass Change: Similarly, higher variability in these hormones was associated with higher levels of lean mass change.
- Interaction Effects: Interestingly, when both E2 and FSH showed high variability, the rate of fat mass increase was actually moderated (lower), suggesting a complex interplay in how the body responds to hormonal instability.
Limitations and Future Work
- Time-Invariant Variance: The current model assumes a subject's variance is constant over the study period. However, hormone variability likely increases as a woman approaches the final menstrual period. Future work could model variance as a function of time ($S_{it}$).
- Missing Data: The study used a complete-case analysis for body composition. While missingness was low (<5%), future iterations could integrate models to handle intermittent missingness or dropout more robustly.
- Computational Complexity: As the number of longitudinal markers increases, the number of variance-covariance parameters grows quadratically, which may require new dimension-reduction techniques to remain computationally feasible.
Summary updated