Epigenetic clocks and mortality prediction: what cohort studies found
Prepared by: Research Desk
Prepared by: Research Desk. Sources are linked for verification.
Data point first: Researchers increasingly test epigenetic clocks against one hard outcome: survival. Several published human studies frame clock values as predictors of lifespan, survival or mortality risk. That framing is easy to misread. This explainer covers what recent cohort studies set out to measure, how to read the statistics they report, and why a population-level association is not the same thing as a forecast about one person.
If you are new to how clocks are built, start with our guide to how epigenetic clocks work. For the internal design of GrimAge specifically, see GrimAge explained. This page focuses on the outcome side: what happens when researchers line clock results up against survival data.
Why mortality is the test researchers care about
The 2013 Horvath estimator was built to estimate DNA methylation age across many human tissues and cell types [8]. Accuracy at estimating age, on its own, says little about whether a clock captures anything health-relevant. Mortality is a hard, unambiguous outcome, which is why many researchers use it when asking whether a clock tracks biological aging rather than calendar time.
Later clocks were published with lifespan and health outcomes explicitly in view:
PhenoAge was presented by its developers as an epigenetic biomarker of aging for lifespan and healthspan [4].
GrimAge was reported by its developers to strongly predict lifespan and healthspan [3]. A second version of the model was published in 2022 [9].
DunedinPACE was published as a DNA methylation biomarker of the pace of aging, so it expresses a rate rather than an age [5].
Because GrimAge was published as a predictor of lifespan and healthspan [3], the more informative questions are how well it holds up in new samples, how consistently it performs across populations, and how it relates to other aging biomarkers.
Beydoun et al. 2025: clocks and telomeres in one model
A 2025 study in Aging Cell examined telomere length, epigenetic age acceleration and mortality risk together in US adult populations, using an additive Bayesian network analysis [1].
Study type: observational analysis of human data from US adult populations [1].
About the method, in general terms: An additive Bayesian network is a type of statistical model that represents how several variables relate to one another at the same time. Models of this kind are generally used to explore whether one variable is linked to an outcome directly, or whether the link appears to run through other variables in the network. The specific variables, adjustments and structure used in this study are described in the paper itself.
Why the topic matters: Telomere length and epigenetic age acceleration are both proposed markers of biological aging. Placing them alongside mortality risk in one analysis, as the title describes [1], is a way of looking at these markers together rather than one at a time.
How to read it: This study is best read as evidence about how aging biomarkers relate to one another and to mortality at the population level [1]. Readers interested in exact effect sizes, cohort sizes and which clock measures were included should check the full paper rather than rely on secondary summaries, including this one.
Kuo et al. 2026: is change over time linked to survival?
Many clock-mortality analyses rely on a single blood draw per person. A 2026 Nature Aging study in the InCHIANTI cohort took a different angle, looking at change in clock values over time in relation to survival [2].
Study type: longitudinal observational human cohort study.
What it reported: as the title states, longitudinal changes in epigenetic clocks predicted survival in this cohort [2]. In other words, change over time was itself associated with survival. This is relevant to anyone who tests more than once, but it does not mean any individual's change between two consumer tests is meaningful. As a general caution, and not a finding of this study, repeat measurements carry some technical noise, a topic we cover in biological age test accuracy.
Other cohort evidence worth knowing
Specific populations. A 2021 study examined DNA methylation clocks side by side in Black South African men [6]. Work of this kind matters because most clocks were developed in other populations, and performance cannot be assumed to carry over.
Distinct biology. A 2026 study assessed the prognostic value of blood epigenetic aging biomarkers in people with well-controlled HIV-1 infection [10], an example of testing whether clock-outcome links hold in groups with distinct biology.
Beyond death. A 2025 Nature Communications analysis set 14 epigenetic clocks against 174 incident disease outcomes in a single unbiased framework [7], widening the benchmark from mortality alone to many specific diseases.
Related aging research. Separately, a 2025 Nature Aging study reported a Pace of Aging analysis of healthspan and lifespan in older adults in the US and UK [11].
How to read a hazard ratio
Many survival studies report results as hazard ratios (HRs) or log-hazard coefficients from survival models. A few general principles prevent most misreadings:
An HR describes relative risk at any given moment, averaged across a group. A hypothetical HR of 1.2 per standard deviation of clock acceleration would mean that, at any point during follow-up, people one standard deviation "older" had about a 20% higher instantaneous risk of death than otherwise similar people. It does not mean they lose 20% of their remaining years.
"Per unit" matters. Some papers report per year of age acceleration, others per standard deviation. The same clock can look stronger or weaker depending on the unit.
Adjustment matters. An HR adjusted for smoking, body size and education answers a narrower question than an unadjusted one. Because clocks differ in what they were built from, adjustment choices can change how one clock looks next to another.
Relative risk is not absolute risk. A higher relative hazard in a group with low baseline mortality can still mean a small absolute difference over a decade.
Population association is not a personal forecast
Even the strongest cohort result is a statement about averages across many people. Three gaps separate it from an individual reading:
Overlap. As a general statistical point, a clear group-level association can coexist with wide overlap between the clock values of people who died during follow-up and people who did not. A shift in average risk does not sort individuals into outcomes.
Measurement differences. As a general point, a published result is tied to the sample type, laboratory platform and clock version a study used. Consumer products may use different sample types, platforms or proprietary versions of a published clock, so a published HR may not transfer directly to a commercial report. Regulatory status adds another layer; see are biological age tests FDA approved?.
Population specificity. Clocks are being examined in distinct groups, such as Black South African men [6] and people living with well-controlled HIV-1 infection [10]. A result calibrated in one population is not guaranteed to hold in another.
The fair summary: GrimAge's developers reported that it strongly predicts lifespan and healthspan [3], a large 2025 analysis set 14 clocks against 174 incident disease outcomes [7], and InCHIANTI found that longitudinal change in clocks predicted survival [2]. None of this licenses reading a single number as a lifespan estimate.
Further reading
Readers who want depth beyond study abstracts often turn to books on epigenetics and the biology of aging. Useful criteria for choosing one: whether the author cites primary research, how recently it was published (clock research moves fast, and pre-2019 titles predate GrimAge), whether it separates observational association from causation, and whether it discloses commercial ties to any testing company.
Library catalogues and publisher sample chapters are a free way to check a book's sourcing and tone before purchasing.
Frequently asked questions
Does a high GrimAge result mean someone will die sooner?
No. GrimAge was published as a clock that predicts lifespan and healthspan [3], and in cohort research that kind of prediction is a statement about group averages. As a general point about statistics, a higher average risk in a group does not determine what happens to any one person.
How did the Beydoun study handle telomere length?
The study analysed telomere length, epigenetic age acceleration and mortality risk in US adult populations using an additive Bayesian network analysis [1]. Network models of this type are generally used to look at several variables together; the study's specific model and effect sizes should be checked in the full paper.
Do clocks predict mortality equally in every population?
Not necessarily. Researchers continue to examine clocks in specific groups, such as Black South African men [6] and people with well-controlled HIV-1 infection [10]. Performance can depend on the population and the measurement pipeline.
Is change between two results linked to survival?
The InCHIANTI analysis found that longitudinal changes in epigenetic clocks predicted survival in that cohort [2]. That is a research-cohort finding; a difference between two consumer reports may also reflect measurement noise rather than true change.
Related on this site
Biological age tests: how epigenetic clocks work and what they mean
GrimAge explained: how the mortality-trained epigenetic clock works
Biological age test accuracy: reliability, noise and repeat results
Are biological age tests FDA approved? How aging tests are regulated
Last updated: 20 October 2026.
Not medical advice. This page reports published research and regulatory records for information only.
Research-only notice: Content on this site is provided for research and educational purposes only and is not intended for human use. It is not medical advice.
Translation notice: Some content on this site may have been machine-translated using Google Translate and may contain errors.