TL;DR
This paper introduces a novel Bayesian joint modeling framework that simultaneously estimates individual-level hormone trajectories (mean and variability) and links them to body fat changes. This approach overcomes limitations of existing methods by properly accounting for uncertainty in estimated variability, revealing that higher variability in E2 and FSH hormones is associated with faster increases in fat and lean mass during the menopausal transition.
What Problem Does This Paper Address?
Epidemiological studies often collect longitudinal biomarker data (like hormone levels) and cross-sectional health outcomes (like body fat). The goal is to understand how these biomarkers predict health. Traditional methods primarily focus on the average trajectory of biomarkers. However, the variability around an individual's average trajectory might also hold crucial information about health risks. For instance, inconsistent hormone levels could signal underlying system instability. Existing statistical methods struggle to incorporate this subject-level variability as a predictor while accurately quantifying uncertainty. This gap makes it difficult to fully understand how individual differences in hormone fluctuations impact health outcomes, particularly during critical life stages like menopause.
Methods
The core contribution is a fully Bayesian joint model. This model has two main components:
- Longitudinal Marker Model: This part models the individual trajectories of biomarkers (E2 and FSH in this case). Crucially, it estimates not only the mean trajectory for each individual but also their individual-level variance-covariance matrix. This matrix captures how much an individual's hormone levels fluctuate and how these fluctuations are related over time.
- Outcome Regression Model: This component links the estimated individual-level mean and variance-covariance parameters from the longitudinal model to a cross-sectional health outcome (body fat changes).
What's novel here is the hierarchical prior specification for the individual-specific variance-covariance matrices. This allows for flexible modeling of between-subject differences in variability. By integrating these components within a single Bayesian framework, the model properly propagates the uncertainty from the longitudinal marker estimation to the outcome prediction, avoiding the biases seen in simpler two-stage approaches. The model is implemented using Hamiltonian Monte Carlo (HMC) via the Stan software.
Datasets
The primary dataset used for evaluation and application is the Study of Women's Health Across the Nation (SWAN). This is a longitudinal cohort study of women transitioning through menopause. The paper uses serum E2 and FSH measurements collected at multiple visits, along with body composition data (fat mass and lean mass) measured via DXA. The final analytical dataset included 841 individuals with 9,902 hormone measurements.
Key Insights and Ideas
- Variability Matters: The study strongly suggests that individual-level variability in hormone levels (E2 and FSH) is not just noise but a significant predictor of health outcomes, specifically body mass changes during menopause.
- Beyond the Average: Focusing solely on average hormone levels misses important information about individual health trajectories. Subject-specific fluctuations can reveal different aspects of physiological function.
- Uncertainty Propagation is Crucial: Traditional two-stage methods that treat estimated variances as known can lead to biased results. A joint model that accounts for the uncertainty in these estimates is essential for reliable inference.
- Menopausal Transition is Complex: The interplay between hormone variability and body composition changes during menopause is intricate, suggesting that disruptions in hormonal stability might have downstream effects on metabolic health.
Key Results
- Simulation Studies: The proposed joint model (JMIV) consistently outperformed alternative two-stage methods (TSLM, TSLMM, TSIV) in terms of bias, coverage, and interval length across various simulation scenarios. This validates the methodological advantage of the joint modeling approach.
- SWAN Data Analysis (Fat Mass):
- Higher variability in E2 was associated with a higher rate of fat mass increase per year (7.03 x 10⁻³).
- Higher variability in FSH was associated with a higher rate of fat mass increase per year (16.38 x 10⁻³).
- An interaction effect was observed: when both E2 and FSH variability were high, the rate of fat mass increase was moderated, suggesting a complex interplay.
- SWAN Data Analysis (Lean Mass):
- Higher variability in E2 was associated with a higher rate of lean mass increase per year (14.62 x 10⁻³).
- Higher variability in FSH was associated with a higher rate of lean mass increase per year (29.70 x 10⁻³).
- Similar to fat mass, a variance interaction term moderated the effects on lean mass change.
- Overall: Women with highly variable E2 and FSH tended to have slower fat mass and lean mass increases when both hormones were highly variable, compared to when only one was highly variable.
Limitations and Future Work
- Simplifying Assumptions: The model assumed that the longitudinal markers were independent given individual-specific parameters and that covariates (Wᵢ) were not included in the longitudinal marker model for simplicity in the application. The authors note that these assumptions could be relaxed.
- Time-Varying Variances: The current model assumes individual variances are constant over time. Future work could extend the model to allow individual variances to change over time (Sᵢₜ), which might better capture dynamic changes in hormone variability. This would likely require larger datasets.
- Missing Data: The analysis excluded individuals with missing hormone data or body mass outcomes. Future work could incorporate methods to handle missing data more comprehensively.
- Number of Trajectories: Extending the model to more than two longitudinal trajectories could lead to a complex outcome regression model due to the quadratic increase in variance-covariance parameters. Dimension reduction techniques might be necessary.
- Intermittent Missingness and Dropout: The study design and data collection can lead to intermittent missingness and dropout, which future work could address.
Summary updated