The Limits of Predicting Individual-Level Longevity: Insights From the U.S. Health and Retirement Study.

Statistical and machine learning models achieved comparable accuracy and relatively high discriminative performance but failed to account for most individual-level lifespan heterogeneity.

• 12 statistical and machine learning models were tested using more than 150 predictors from the U.S. Health and Retirement Study longitudinal data.
• Models showed 'relatively high discriminative performance' but could not explain most of the variation in individual lifespan.
• The finding held consistently across all model types, suggesting a ceiling on predictability rather than a limitation of any specific modeling approach.

Consistent inequalities in mortality predictability and risk discrimination were observed across sociodemographic groups.

• Lower prediction accuracy was found for men compared to other gender groups.
• Non-Hispanic Black individuals showed lower accuracy in mortality prediction.
• Low-educated individuals also demonstrated lower prediction accuracy.
• These inequalities were described as 'consistent' across models, suggesting a systematic rather than model-specific pattern.

Individuals in lower-predictability groups (men, non-Hispanic Blacks, low-educated) also showed lower accuracy in their own subjective predictions of lifespan.

• Subjective longevity predictions by study participants were compared to model-based predictions across sociodemographic groups.
• The same groups for whom statistical models performed worse also self-predicted their lifespan less accurately.
• This parallel between model-based and self-assessed predictability suggests underlying structural or informational disparities rather than purely statistical limitations.

Top predictive features were similar across sociodemographic groups, with variables related to habits, health history, and finances being the most relevant predictors.

• More than 150 predictors derived from the HRS were used in the analysis.
• Variables related to habits (e.g., smoking, physical activity), health history, and financial factors emerged as top features.
• The consistency of top features across groups suggests that differential predictability is not driven by different predictor relevance but likely by differential predictor-outcome relationships or unmeasured heterogeneity.

The study used the U.S. Health and Retirement Study, described as one of the richest longitudinal representative surveys in the United States, yet still encountered fundamental limits in individual-level mortality prediction.

• The HRS provided longitudinal data with more than 150 derived predictors across health, financial, behavioral, and demographic domains.
• Despite the richness of available data, models could not overcome fundamental limits in predicting individual-level mortality.
• The authors frame these limits as context-dependent and relevant for guiding future research and public policies.

The paper is positioned within a growing body of research on the predictability of life course events, drawing on that literature to contextualize mortality prediction challenges.

• Individual-level mortality prediction is described as 'a fundamental challenge with implications for life planning, health care, social policies, and public spending.'
• The study explicitly situates itself within 'the growing body of research on the predictability of life course events.'
• The authors provide 'baselines and guidance for future research and public policies' based on their findings.