Joint functional-cognitive trajectories and hospitalization-associated within-person change in multimorbid older adults: a CHARLS and HRS analysis

In older adults with multimorbidity, functional disability and cognitive performance follow heterogeneous courses that baseline assessment alone misses. We characterised joint ADL-cognition trajectories, estimated the within-person association between wave-specific hospitalization reports and concurrent outcomes, and assessed how well baseline variables distinguished the fitted trajectory classes. We analysed adults aged at least 60 years who had multimorbidity, complete baseline ADL and cognition, and at least two observed waves, in CHARLS (2011–2020; n = 2,251) and in an independently analysed HRS sample (2012–2020; n = 8,139). Multivariate latent-class mixed models with class-specific quadratic trajectories were screened against prespecified criteria for convergence, cross-start reproducibility, classification quality, and class size. GEE-Mundlak models estimated within-person contrasts between waves with and without a hospitalization report. Baseline class distinction was evaluated with fixed five-fold out-of-fold multinomial predictions. CHARLS supported five classes of 288, 1,151, 222, 280, and 310 participants. HRS independently selected k = 5; eight of ten trajectory-direction comparisons matched after cohort-standardized Hungarian alignment, representing a descriptive, cohort-specific shape comparison. Model entropy was 0.780, class-specific APPA ranged from 0.764 to 0.892, and OCC from 7.87 to 53.06. The CHARLS k = 6 solution had a lower BIC but failed the prespecified stability criterion (minimum aligned kappa = 0.7845), and was therefore retained only as a shape-level sensitivity analysis. The within-person association between a hospitalization report and ADL was 0.214 points (95% CI 0.134 to 0.295) in CHARLS and 0.170 points (0.137 to 0.204) in HRS. Cognition estimates were negative but imprecise, at -0.150 (-0.333 to 0.032) and − 0.099 (-0.201 to 0.002), respectively. In CHARLS, P20 macro-AUC was 0.822 (conditional 95% interval 0.811 to 0.831), but recall for Class 1 was 0.017, indicating that aggregate baseline variables did not reliably distinguish Class 1, the minority class characterized by later functional worsening. Repeated observations resolved several descriptive functional-cognitive patterns. Hospitalization reports were associated with higher concurrent ADL limitation in both cohorts. Baseline variables provided aggregate class distinction but did not reliably identify one minority class, underscoring the value of repeated observations for characterizing longitudinal heterogeneity.