Related Experiment Videos
Predicting transitions between functional and financial vulnerability states and death in older Chinese adults: a
Abstract:
Background Functional decline can increase care needs at the same time that out-of-pocket spending constrains continued treatment and rehabilitation. Existing studies usually predict these outcomes separately and therefore cannot identify older adults moving into joint functional and financial vulnerability. We developed a multistate model and tested whether its predictions remained informative in later CHARLS waves. Methods We used 2011-2018 Harmonized China Health and Retirement Longitudinal Study data. Adults aged at least 60 years contributed adjacent-wave observations. Four living states combined basic-activity difficulty and household catastrophic health expenditure, defined as out-of-pocket health spending at least 40% of non-food expenditure; death was the fifth state. Penalized multinomial regression and XGBoost were developed in 2011-2015 and evaluated without updating in 2015-2018. Penalized regression was primary because it matched XGBoost while remaining transparent. The locked model was applied to 2018 origins for exploratory 2020 validation of observable death and ADL components. Results The analysis comprised 12,866 transitions from 7,353 participants; 5,175 formed the 2015-2018 evaluation set, including 361 dual-vulnerability outcomes and 389 deaths. Penalized regression achieved a macro-AUROC of 0.722 (95% CI 0.712-0.734) and Brier score of 0.570 (0.557-0.583). AUROCs were 0.728 for dual vulnerability, 0.755 for death, and 0.629 for CHE alone. XGBoost did not improve macro-AUROC (0.720). Of 8,532 eligible 2018 origins, 8,004 had death or ADL status observed in 2020. Exploratory component AUROCs were 0.813 for death and 0.801 for ADL difficulty or death. The highest-risk 20% contained 53.5% of deaths and 46.1% of dual-vulnerability-or-death outcomes. Conclusions Joint functional-financial states were dynamic and prognostically distinct. A transparent penalized model provided moderate discrimination for dual vulnerability and death in the complete five-state evaluation and identified a manageable high-risk stratum for coordinated functional and financial review. The 2020 analysis supported only the observable clinical components. Isolated CHE remained difficult to forecast without detailed reimbursement, local-cost, and long-term-care information.