Tree-based analysis of longevity predictors and their ten-year changes: a 35-Year mortality follow-up

Abstract

Abstract Background Prior studies on longevity often examine predictors in isolation and rely solely on baseline information, limiting our understanding of the most important predictors and their dynamic nature. In this study, we used an innovative regression tree model to explore the common characteristics of those who lived longer than their age and sex peers in 35-years follow-up. We identified different pathways leading to a long life, and examined to how changes in characteristics over 10 years (from 1979 to 1989) affect the findings on longevity predictors. Methods Data was obtained from the “Tampere Longitudinal Study on Ageing” (TamELSA) in Finland. Survey data was collected in 1979 from 1056 participants aged 60–89 years (49.8% men). In 1989, a second survey was conducted among 432 survivors from the 1979 cohort (40.2% men). Dates of death were provided by the Finnish Population Register until 2015. We employed an individual measure of longevity known as the realized probability of dying (RPD), which was calculated based on each participant’s age and sex, utilizing population life tables. RPD is based on a comparison of the survival time of each individual of a specific age and sex with the survival time of his/her peers in the total population. A regression tree analysis was used to examine individual-based longevity with RPD as an outcome. Results This relative measure of longevity (RPD) provided a complex regression tree where the most important characteristics were self-rated health, years of education, history of smoking, and functional ability. We identified several pathways leading to a long life such as individuals with (1) good self-rated health (SRH), short smoking history, and higher education, (2) good SRH, short smoking history, lower education, and excellent mobility, and (3) poor SRH but able to perform less demanding functions, aged 75 or older, willing to do things, and sleeping difficulties. Changes in the characteristics over time did not change the main results. Conclusion The simultaneous examination of a broad range of potential predictors revealed that longevity can be achieved under very different conditions and is achieved by heterogeneous groups of people.

Keywords

• Mortality
• Relative measure of longevity
• Machine learning
• Regression tree
• Realized probability of dying