Predicted fat mass index and incident stroke in adults aged 65 years or older: heterogeneous associations across three national ageing cohorts

Background Predicted fat mass index (pBFMI) can be calculated when direct body-composition measurements are unavailable, but it is a deterministic transformation of body mass index (BMI), age, and sex. We examined whether its association with incident stroke was consistent across three national ageing cohorts. Methods We analysed stroke-free adults aged 65 years or older in CHARLS (2011–2020), ELSA (2004–2018), and HRS (2006–2020). Because stroke onset was known only between survey waves, the primary analysis used discrete-time complementary log-log models. Model 2 adjusted for age, sex, education, marital status, smoking, and an alcohol-use indicator; hypertension, diabetes, and heart disease were added in Model 3 as potential intermediates. Sex-specific standardization was recalculated within each cohort's age-65-or-older sample. We assessed mathematical coupling with BMI, compared apparent model fit and discrimination, and treated random-effects synthesis as descriptive. Results Model 2 included 3,241 CHARLS participants (317 strokes), 2,494 ELSA participants (205 strokes), and 3,614 HRS participants (413 strokes). Per sex-specific standard deviation higher pBFMI, hazard ratios were 1.247 (95% CI 1.165–1.334) in CHARLS, 0.924 (0.800–1.067) in ELSA, and 0.987 (0.892–1.091) in HRS. A one-stage discrete-time model showed cohort heterogeneity (joint interaction P < 0.001), which persisted when all Model 2 covariate effects were allowed to vary by cohort (P < 0.001). Hartung–Knapp descriptive synthesis gave an HR of 1.051 (0.709–1.560), I²=89.8%, and a 95% prediction interval of 0.491–2.250. pBFMI correlated strongly with BMI (r = 0.977–0.988) and did not materially improve apparent AIC or C-index. Conclusions Among adults aged 65 years or older, pBFMI was positively associated with incident stroke in CHARLS but not in ELSA or HRS. The findings do not support a consistent cross-population association and provide no evidence of materially improved apparent model fit or discrimination over BMI.