Abstract / Summary
Abstract Background Latent multimorbidity phenotypes may be more clinically interpretable than disease counts, but phenotype studies are often vulnerable to unstable class solutions, classification error, and absence of validation. We derived physical multimorbidity phenotypes among adults with a history of stroke and tested their reproducibility and association with elevated depressive symptoms in non-overlapping participants. Methods The derivation sample comprised 486 participants reporting physician-diagnosed stroke in the 2011 baseline wave of the Harmonized China Health and Retirement Longitudinal Study (CHARLS). Bernoulli latent class models used 11 physical conditions. Stability was evaluated in 200 bootstrap samples, and outcome inference was repeated across 300 posterior pseudo-class draws. The 2011 class prevalences and conditional probabilities were then frozen and applied, without refitting, to 918 different participants whose earlier stroke reports were absent and whose first study report of stroke occurred in 2013, 2015, or 2018. Elevated depressive symptoms were defined as a 10-item Center for Epidemiologic Studies Depression Scale score ≥ 10. Adjusted logistic models included sociodemographic, behavioural, functional, and validation-wave covariates. Results The selected solution contained high-complexity, cardiometabolic, and low-comorbidity classes. Across bootstrap samples, median correlations with the original condition profiles were 0.966, 0.982, and 0.985, respectively. In the derivation complete-case model (n = 409), the high-complexity class was associated with elevated depressive symptoms versus the low-comorbidity class (adjusted odds ratio [aOR] 2.59, 95% CI 1.30–5.16); posterior pseudo-class inference gave aOR 2.35 (1.09–5.03). The association replicated in the temporal validation sample (n = 782; aOR 1.74, 1.17–2.60) and in a response-weighted analysis (1.93, 1.22–3.06). Cardiometabolic-class estimates were null. In a common complete-case subset, a simple disease-count model had marginally better fit and discrimination than the class model. Conclusions A high-complexity physical multimorbidity phenotype showed a reproducible association with elevated depressive symptoms in non-overlapping later-wave CHARLS participants. The classes add clinical interpretability but did not outperform disease count for statistical fit. Because validation remained within CHARLS and exposures and symptoms were measured concurrently within each wave, external prospective validation is still required.