Disparities in life expectancy across sociodemographic groups are large in the U.S. Still, using linked Census and mortality records for men born in 1910, Casey Breen and Nathan Seltzer find that sociodemographic characteristics explain less than 2% of variation in individual lifespan, highlighting the role of stochasticity in shaping individual-level mortality.
There are striking disparities in life expectancy across sociodemographic groups in the United States, shaped by structural forces such as racism, class inequality, and policy environments. To what extent do sociodemographic characteristics structure, or fail to structure, individual lifespans?
In a recent paper (Breen and Seltzer 2026), we used U.S. Census data linked to Social Security mortality records to assess how well early-adulthood social, economic, and demographic characteristics predict individual lifespan in a cohort of men born in 1910 and observed through their deaths between 1975 and 2005 (N = 121,000). These data come from the UC-Berkeley CenSoc project. Despite large group-level disparities, we found that sociodemographic characteristics measured in early adulthood explain less than 2% of the overall variation in individual lifespan. These findings reaffirm a central demographic regularity: variance in life expectancy between groups is small compared to variation in lifespan within groups.
Mortality demography, at its core, is concerned with population rates and group-level disparities. What does a predictive perspective add? First, the use of predictive performance as a diagnostic tool quantifies how much of the variation in lifespan within a cohort is captured by sociodemographic indicators of structural advantage and disadvantage. This approach generalizes traditional decompositions of lifespan variation (which in most empirical applications have examined one categorical covariate at a time) by translating multiple covariates into measures of explanatory power at the individual level. In short, prediction reframes classical decomposition as an inquiry into individual-level explanatory power, leveraging both categorical and continuous characteristics.
Second, limited predictability reveals how inequality and uncertainty coexist: stochasticity shapes the overall distribution of lifespans, while structural disparities shift its mean and skewness but remain limited in their ability to determine individual outcomes.
Finally, the low predictability of longevity based on sociodemographic characteristics is itself demographically meaningful. It provides empirical insight into a universal form of uncertainty, one that individuals consider when making life decisions (e.g., “Will I outlive my retirement savings?”). This perspective aligns with the emerging field of uncertainty demography (Trinitapoli 2023), which calls for making uncertainty more central to demographic inquiry and highlights the distinctive capacity of demographic approaches in illuminating the role of uncertainty in social life.
Why a cohort approach matters
Most studies of individual mortality prediction use a period design: researchers pool individuals of varying ages and predict if they will die in a fixed window, such as the next five years. In those models, age does most of the predictive work: older people die sooner. Even omitting age, covariates like self-rated health or homeownership status partly function as proxies for age. This makes period predictions useful for many practical applied forecasting purposes, but less well suited for understanding the extent to which observable sociodemographic characteristics explain variation in lifespan.
A cohort design solves this by holding age constant. We follow a single birth cohort: men born in 1910, observed in the 1940 Census at ages 29 or 30, with deaths captured in Social Security records between ages 65 and 95. This allows us to directly assess the age at which a given member of this cohort dies, given the observed sociodemographic characteristics in early adulthood. To our knowledge, this is the first U.S. cohort-based test of lifespan predictability.
Within-group vs. between-group variation
In our data, we observe a clear educational gradient. College-educated men in this cohort lived nearly three years longer at age 65 than men with only elementary schooling. We observe a similar gradient for wage and salary income. These between-group differences align with past findings and are consistent with well-documented patterns of mortality inequality.
However, our best predictive model, an ensemble model combining seven different machine learning algorithms, explained only 1.3% of the variation in age at death in our holdout set. Predictions clustered tightly between ages 76 and 82, indicating that the model could not reliably distinguish those who were dying earlier or later. Robustness checks confirm this is not an artifact of focusing on one male birth cohort: replicating the analysis for each birth cohort from 1900-1920 produced similar results, as did a separate sample including women.
Why are predictions so weak even in the presence of strong between-group disparities? Among men with college degrees, some died in their late 60s and others lived past 95. The same is true for every educational and income group. The roughly three-year between-group gap is substantial, but it is small relative to the 25+ year spread of ages at death within any group. This is a long-standing demographic regularity, which has been nicely documented using decomposition approaches considering one covariate at a time (Caswell 2023; van Raalte et al. 2012); our prediction exercise gives it a new empirical lens.
This highlights a core tension: there are substantial and important between-group disparities, but individual-level variation is great enough that we cannot use these sociodemographic characteristics to accurately predict lifespan for a given person. In other words, predictability is low not because between-group inequality is absent, but because within-group variation dominates overall variation in lifespan.
This echoes a classic result of Vaupel (1988): even in a hypothetical setting where all frailty (disciplinary jargon for mortality risk) is directly inherited from parents, frailty explains only 2–5 percent of lifespan variance. Lifespan is shaped, in large part, by stochasticity. More broadly, this finding also aligns with past research on the limited predictability of major life outcomes (Salganik et al. 2020).
Predictability also varies across subgroups. Our models explained a smaller share of lifespan variance among Black Americans, low-income men, and men with low education. The most disadvantaged face not only shorter average lifespans but also more unpredictable ones. This inequality matters because lifespan uncertainty plays into decisions about saving, retirement, family formation, and health investments.
Interpretation and takeaways
Low predictability should not be read as evidence that structural inequality does not matter. Between-group disparities remain substantial, policy-relevant, and central to mortality demography. Mortality is simply a highly stochastic process, not predetermined by individual-level sociodemographic characteristics.
These findings align with the emerging field of uncertainty demography (Trinitapoli 2023), which treats uncertainty as a constitutive feature of population processes rather than residual noise. In this sense, lifespan uncertainty is not merely noise to be explained away by better covariates and models, but a defining demographic feature that shapes how individuals and populations experience mortality.
References
Breen, C. F., and Seltzer, N. (2026). Structured Inequality, Uncertain Lifespans: Demographic Perspectives on Predicting Individual-Level Longevity. Population and Development Review. https://onlinelibrary.wiley.com/doi/10.1111/padr.70065
Caswell, H. (2023). The contributions of stochastic demography and social inequality to lifespan variability. Demographic Research 49: 309–354.
Salganik, M. J., et al. (2020). Measuring the predictability of life outcomes with a scientific mass collaboration. PNAS 117(15): 8398–8403.
Trinitapoli, J. (2023). An Epidemic of Uncertainty: Navigating HIV and Young Adulthood in Malawi. University of Chicago Press.
van Raalte, Alyson A., Anton E. Kunst, Olle Lundberg, Mall Leinsalu, Pekka Martikainen, Barbara Artnik, Patrick Deboosere, Irina Stirbu, Bogdan Wojtyniak, and Johan P. Mackenbach. 2012. “The Contribution of Educational Inequalities to Lifespan Variation.” Population Health Metrics 10(1):3. doi:10.1186/1478-7954-10-3.
Vaupel, J. W. (1988). Inherited frailty and longevity. Demography 25(2): 277–287.