Socioeconomic Status Fails to Predict Osteoporosis Once Sex Is Accounted For

A large longitudinal analysis of two national ageing cohorts has found that the apparent link between socioeconomic status and newly diagnosed osteoporosis largely evaporates once demographic factors, particularly sex, are taken into account. The study, published in Frontiers in Endocrinology, analysed data from the English Longitudinal Study of Ageing (ELSA) and the Health and Retirement Study (HRS) and concluded that a crude socioeconomic disparity in osteoporosis risk was visible only in the English cohort and only before adjustment.

The research team, led by investigators at the Affiliated Hospital of Southwest Medical University in Luzhou, China, drew on ELSA Waves 2 through 9, covering 2004 to 2018, and HRS Waves 11 through 15, covering 2012 to 2020. Participants aged 45 and older who already reported osteoporosis at baseline were excluded, along with those outside defined ranges for body mass index and blood pressure. After a complete-case filter on analysis covariates, the final samples comprised 5,041 ELSA participants with 423 incident osteoporosis events and 6,283 HRS participants with 734 events. Median follow-up was 12 years in ELSA, yielding 51,724 person-years, and 8 years in HRS, yielding 39,966 person-years. Cumulative incidence was 8.4 percent in ELSA and 11.7 percent in HRS.

Socioeconomic status was constructed as a composite score ranging from 0 to 5, combining education, scored 0 to 2, with within-cohort, within-wave quartiles of household economic resources, scored 0 to 3. Household economic resources summed total household income and total assets, including housing and financial assets. Baseline participants were then grouped into high, medium, and low SES categories. Osteoporosis was defined as self-reported physician diagnosis, ascertained at each survey wave, with participants free of the condition censored at their last confirmed osteoporosis-free wave. Because status was assessed only at discrete waves roughly two years apart, the exact onset date is interval-censored, a limitation the authors addressed in sensitivity analyses.

The headline result concerns how the estimated association changed with progressive covariate conditioning. In the crude ELSA model, participants with low baseline SES had a 54 percent higher hazard of incident osteoporosis than those with high SES, with a hazard ratio of 1.54 and a 95 percent confidence interval of 1.20 to 1.97. Yet after adjustment for age, sex, marital status, retirement, smoking, and alcohol consumption, the hazard ratio fell to 1.00, and in the fully adjusted model, which additionally included BMI, systolic blood pressure, and diastolic blood pressure, it was 1.14 with a confidence interval of 0.87 to 1.49, statistically non-significant. In HRS, no significant association appeared in any model; the fully adjusted hazard ratio for low versus high SES was 1.03.

A stepwise attenuation analysis in ELSA pinpointed where the crude association collapsed. Starting from an age-adjusted hazard ratio of 1.32, adding sex to the model dropped the estimate to 1.05, and further adjustment for marital status and retirement brought it to 1.00. Smoking, alcohol, BMI, and blood pressure produced only modest additional change. The authors interpret this pattern cautiously: because sex was specified a priori as a confounder rather than a mediator, the attenuation should not be read as evidence that sex lies on a causal pathway between SES and osteoporosis. Instead, they argue, the marked difference in sex composition across SES groups, with low-SES participants more likely to be female in both cohorts, plausibly confounded the crude association, given that women face both higher osteoporosis risk and lower lifetime earnings and wealth.

Importantly, the sex story does not extend to effect modification. Sex-stratified Cox models showed no statistically significant SES association among either men or women in either cohort, and formal SES-by-sex interaction tests were null, with a p-value for interaction of 1.000 in ELSA and 0.175 in HRS. In other words, the data provide no evidence that socioeconomic circumstances matter differently for osteoporosis in one sex, even though differences in sex composition across SES groups may explain the crude disparity.

The study also examined longitudinal SES trajectories rather than single-time-point measures. K-means clustering on standardised SES scores across all available waves identified two trajectory groups in each cohort, interpreted as persistently high and persistently low SES, with silhouette widths maximal at two clusters. In ELSA, the persistently high trajectory was associated with a 28 percent lower crude hazard of incident osteoporosis, hazard ratio 0.72, but this too attenuated to non-significance after adjustment, reaching 0.92 in the fully adjusted model. A strict pre-outcome landmark sensitivity analysis, reconstructing trajectories using only SES information collected before subsequent osteoporosis follow-up, confirmed the non-significant findings in both cohorts, guarding against future-information bias.

An exploratory mediation analysis tested whether baseline physical activity, quantified as MET-days per week from self-reported frequencies of light, moderate, and vigorous activity, mediated the SES-osteoporosis association. It did not. Average causal mediation effects were small and non-significant in both cohorts, and total effects were non-significant as well, so the proportion mediated was not interpreted. The authors note that a single baseline physical activity measure may not capture long-term patterns or the weight-bearing exercise types most relevant to bone health, and that the modest incidence rates limited statistical power for detecting small mediation effects. The analysis was explicitly framed as hypothesis-generating rather than causal.

Machine learning added a sobering note on prediction. Five algorithms were trained on baseline predictors, with SMOTE oversampling within cross-validation to address class imbalance. Logistic regression delivered the best balance of discrimination and case detection, achieving an AUC of 0.691 in ELSA with sensitivity of 0.754, and an AUC of 0.708 in HRS with sensitivity of 0.791. XGBoost posted a marginally higher HRS AUC of 0.711 but a sensitivity of only 0.032, meaning it missed nearly all cases. Several tree-based models had sensitivity below 0.10 and balanced accuracy near chance. The authors conclude these models are insufficient as standalone prediction tools, and that the findings do not support using SES alone for osteoporosis risk stratification or clinical decision-making.

The limitations deserve emphasis. Osteoporosis ascertainment relied on self-report rather than dual-energy X-ray absorptiometry, which may underestimate true occurrence and could introduce differential ascertainment if diagnostic access varies by SES. The authors speculate that cross-national differences, including the UK’s universal NHS coverage versus the more mixed US system, might partly explain why a crude association appeared in ELSA but not HRS, but they stress this remains a hypothesis, since insurance status, healthcare utilisation, DXA screening, and diagnostic access were not directly measured. Other limitations include interval censoring, substantial exclusions primarily for age and missing blood pressure data, the inability to perform a competing-risk analysis for death, residual confounding from unmeasured factors such as dietary calcium, vitamin D status, and medication use, and potential overadjustment because several covariates may lie downstream of SES. The fully adjusted model should therefore be read as a covariate-conditional association, not as evidence that SES has no overall relationship with osteoporosis. Future studies with standardised DXA-based ascertainment and direct measures of healthcare and diagnostic access are needed to separate socioeconomic differences in underlying disease risk from differences in who gets diagnosed.

Subject of Research: Longitudinal associations between socioeconomic status and incident osteoporosis risk in older adults in the ELSA and HRS cohorts

Article Title: Socioeconomic status and incident osteoporosis risk in older adults: a longitudinal analysis of the English longitudinal study of ageing and the health and retirement study

Article References: Huang, M., Wang, Q., Chen, Y., Lu, Y., Xing, F., & Wang, J. (2026). Socioeconomic status and incident osteoporosis risk in older adults: a longitudinal analysis of the English longitudinal study of ageing and the health and retirement study. Frontiers in Endocrinology, 17, Article 1831069. https://doi.org/10.3389/fendo.2026.1831069

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1831069

Keywords: osteoporosis, socioeconomic status, ELSA, HRS, longitudinal cohort study, sex differences, physical activity, machine learning prediction, ageing, epidemiology, bone health, health disparities