TL;DR
This paper introduces a novel statistical approach to model how individual variations in hormone levels (estradiol and FSH) and their variability over time predict bone mineral density (BMD) changes during the menopausal transition. By jointly modeling hormone trajectories and BMD, the study reveals that higher FSH variability, not just its average level, is linked to slower bone loss, offering a more nuanced understanding of bone health in midlife women.
What Problem Does This Paper Address?
Bone mineral density (BMD) naturally declines with age, with a significant acceleration during the menopausal transition. While the average changes in hormones like estradiol (E2) and follicle-stimulating hormone (FSH) are known to influence BMD, existing research has largely focused on population-level mean associations. This leaves a critical gap: how does the individual variability or fluctuation in these hormone levels, beyond their average trends, impact an individual's bone health trajectory? Understanding this subject-level variability is crucial for developing personalized strategies to maintain bone health and prevent osteoporosis, especially as women experience significant hormonal shifts during menopause.
Methods
The authors propose a joint modeling framework that simultaneously estimates individual-level trajectories for both hormone levels (the predictor) and BMD (the outcome). What makes this approach novel is its explicit modeling of individual residual variances for the hormone trajectories. This means the model doesn't just capture the average hormone level over time for each woman but also quantifies how much each woman's hormone levels fluctuate around her own average. These estimated individual variances are then directly incorporated as predictors in the BMD outcome model. This contrasts with traditional two-stage approaches, which often ignore or improperly handle the uncertainty associated with estimating individual-level variances, potentially leading to biased results. The statistical framework is Bayesian, implemented using MCMC sampling via Stan.
Datasets
The study utilizes data from the Study of Women's Health Across the Nation (SWAN), a multi-site longitudinal cohort study of women undergoing the menopausal transition. The analytic sample consists of 974 women with observed sex hormone and BMD data across multiple visits.
Key Insights and Ideas
- Variability Matters: The study highlights that individual variability in hormone levels, particularly FSH, is a significant predictor of bone health, not just the average hormone level. This shifts the focus from population averages to individual-specific patterns.
- FSH Variability as a Protective Factor: Counterintuitively, higher variability in FSH levels was associated with slower declines in BMD over time, suggesting a potential protective role that warrants further investigation.
- Joint Modeling Advantage: The proposed joint model effectively propagates uncertainty from the predictor (hormone) model to the outcome (BMD) model, leading to more accurate estimates, especially for variance-related parameters, compared to traditional two-stage methods.
- Menopausal Transition Nuances: The findings suggest that the impact of hormone variability on BMD can change over the course of the menopausal transition, with FSH variability showing a moderating effect over time.
Key Results
- FSH Variability and BMD: Higher FSH variability interacted with time was associated with slower declines in BMD across the menopausal transition. Specifically, higher FSH variance predicted an average 0.26% decline in BMD, moderated over time.
- E2 vs. FSH: The study suggests that FSH mean and variance might be stronger predictors of menopausal bone health than E2.
- Model Performance: The proposed joint model demonstrated superior performance in simulation studies, achieving over 90% interval coverage for parameters, outperforming naive two-stage alternatives that often failed to propagate uncertainty correctly.
- E2 Predictor Model: Predicted E2 residual at the visit before FMP was significantly associated with BMD at FMP. Higher E2 variability at FMP was negatively associated with BMD, though not statistically significant.
- FSH Predictor Model: Predicted FSH residual at the visit before FMP was associated with lower BMD at FMP. The FSH and time interaction was significant, indicating that the association between mean FSH and BMD becomes amplified over time. FSH variability at FMP was not significantly associated with BMD, but its interaction with time was significant, showing a moderating effect.
Limitations and Future Work
The authors acknowledge several limitations:
- Model Convergence: Attempts to fit a linear time-varying variance model for hormones on the SWAN dataset encountered severe convergence issues, suggesting the data might not have enough signal to support such complex individual variance specifications.
- Selection Bias: The reduction to an analytic sample of 972 women from the original 3,302 could introduce selection bias, particularly due to the exclusion of sites lacking DEXA technology or women with incomplete menstrual calendar data.
- Data Sparsity: The relatively small number of observations per individual and annual data collection might limit the ability to capture fine-grained time-varying variance trends.
Future work suggested includes:
- Modeling BMC and Bone Area: Extending the analysis to bone mineral content (BMC) and bone area trajectories, as BMD alone might not fully capture bone health.
- Complex Variance Modeling: Investigating more complex formulations of individual variances, such as decomposing them into short-term and long-term trends, which would require richer datasets.
- Multivariate Outcomes: Extending the joint model to simultaneously analyze multiple longitudinal markers (e.g., BMC and bone area) as multivariate outcomes.
Summary updated