Predicting Later-Life Self-Rated Health from Childhood Living Conditions Using Machine Learning Approach

Background Previous studies have suggested that childhood living conditions contribute to health inequalities in later life. We employed 8 machine learning (ML) regression models to evaluate the importance and heterogenous associations of a wide range of childhood living conditions on self-rated health in later life. Methods We used data from the China Health and Retirement Longitudinal Study (CHARLS) on 59 childhood living conditions across 8 domains, along with self-rated health assessments from before (2018) and during (2020) the COVID-19 pandemic among older adults across 150 counties (N = 15,461; 53% women). Results CatBoost was the best-performing ML model. We mapped out the overall and domain-specific importance of the 59 childhood living conditions and identified the most critical conditions in the 8 domains such as food deficiency, self-rated health before age 15, exposure to civil war, and being bullied by neighborhood kids. For women, childhood family financial situation, male guardian upset and having a group of friends were more influential than among men. For rural residents, childhood family financial situation, relationship with the male guardian and male guardian upset played a more important role compared to urban residents. During the COVID-19 pandemic, childhood perceived neighborhood safety and neighborhood closeness played a more critical role comparing to before the pandemic. Conclusion Our findings suggest the potential for developing a quick screening tool for early intervention and targeted policy. Additionally, these findings may have implications for other aging countries in East Asia or other low- and middle-income contexts.