Longitudinal Trajectories and Multistate Transition Risks of Late-Life Depressive Symptoms Across ADL and Healthcare-Contact States

Source abstract

Depressive symptoms in later life develop alongside changes in physical function and contact with healthcare services, but average associations do not show how these features combine within individuals over time. We analyzed 6229 adults aged 60 years or older who participated in the 2015, 2018, and 2020 waves of the China Health and Retirement Longitudinal Study. Missing item data were handled with 20 multiply imputed datasets. Robust mean-shape latent growth models summarized population-average change in a six-component ADL score, a three-component healthcare-contact count, and CES-D-10 depressive-symptom scores. Functional status, healthcare contact, and depressive-symptom status were then combined into 18 joint states, with transition probabilities estimated separately for 2015–2018 and 2018–2020. Participant-level bootstrap resampling and Rubin pooling were used for interval-specific risks and prediction comparisons. ADL scores declined slightly, whereas healthcare-contact counts and CES-D-10 scores increased; the latter two changes were not constant across the two intervals. Next-wave depressive-symptom risk was driven mainly by prior depressive symptoms and was higher in functionally impaired states. The full transition distributions differed between intervals, although risk rankings among well-populated origin states were similar. Population-weighted analyses showed that common healthy-function pathways contributed more observed depressive-symptom cases than rarer severe-impairment pathways despite lower conditional risks. In temporal-transfer analysis, the full 18-state model improved AUC relative to a covariate-only model, but added little beyond prior depressive-symptom status. The multistate representation is therefore most useful for describing heterogeneous transition patterns and separating state-conditional risk from population pathway contribution, rather than as a stand-alone prediction algorithm.