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

Background Socioeconomic status (SES) may influence osteoporosis (OP) risk, but evidence regarding incident OP remains limited. We examined associations of baseline SES and longitudinal SES trajectories with incident OP in two ageing cohorts and explored physical activity as a potential mediator. Methods We analysed data from the English Longitudinal Study of Ageing (ELSA, Waves 2–9, 2004–2018) and the Health and Retirement Study (HRS, Waves 11–15, 2012–2020). Participants with baseline OP were excluded. Cox proportional hazards models assessed associations between baseline SES and incident OP. SES was derived from education and household economic resources. SES trajectories were identified using K-means clustering. Sex-stratified and SES-by-sex interaction analyses were performed. Five machine learning algorithms predicted incident OP. Sensitivity analyses included discrete-time proportional hazards models and complete-case/multiple-imputation comparisons. Results The samples included 5,041 ELSA participants (423 OP events) and 6,283 HRS participants (734 events). In ELSA, low SES was associated with higher incident OP risk in crude analysis (M1 HR = 1.54, 95% CI 1.20–1.97, p<0.001), but the association attenuated after adjustment (M3 HR = 1.14, 95% CI 0.87–1.49, p=0.33), with the largest attenuation occurring after adjustment for sex. No significant association was observed in HRS (M3 HR = 1.03, 95% CI 0.84–1.28, p=0.75). Sex-stratified and interaction analyses provided no evidence that the SES–OP association differed by sex. Two SES trajectory groups were identified in both cohorts. The high-SES trajectory was associated with lower crude OP risk in ELSA (M1 HR = 0.72, 95% CI 0.59–0.87, p<0.001), but this association attenuated after adjustment. Physical activity showed no significant mediation. Logistic regression provided the best discrimination-sensitivity balance (AUC: ELSA 0.691; HRS 0.708), although XGBoost had a slightly higher HRS AUC (0.711) with low sensitivity (0.032). Sensitivity analyses supported the primary interpretation. Conclusions SES–OP disparities were observed in crude ELSA analyses but were substantially attenuated after covariate adjustment, particularly after adjustment for sex. Differences in sex composition across SES groups may contribute to the crude association, although there was no evidence of effect modification by sex. Cross-cohort differences may partly reflect healthcare and diagnostic contexts, but these explanations require further investigation using standardised OP ascertainment and direct measures of diagnostic access.