A younger biological age: what the test can actually tell us

A repeat result

A healthy 50-year-old took a commercial DNA-methylation test a year ago and was told that her biological age was 52. She changed her diet, started a supplement and slept more. The repeat test says 49. Three years younger, the report notes, in twelve months.

The question is not whether the person feels better. It is what the number is, what would have to be true for a three-year change to be real, and what it would mean if it were. Those are questions about measurement, and they have precise answers.

What was measured

Epigenetic clocks read DNA methylation at a defined set of cytosine-guanine sites in blood and combine the values with weights fixed when the clock was trained. The output is a prediction, and the thing being predicted differs between clocks. First-generation clocks were trained to predict chronological age; their error against a calendar is the quantity they minimised. Second-generation clocks such as PhenoAge and GrimAge were trained against clinical biomarkers and mortality, so their output is a risk estimate re-expressed in years. A third kind, of which DunedinPACE is the best known, was trained on how fast a cohort's organ-system markers changed across two decades, and it reports a rate: years of physiological ageing per calendar year.

These are different targets. A 49 from one clock and a 49 from another are not interchangeable, and a change in one does not imply a change in the others. The figure below separates the three quantities a report may label as age.

years

Chronological age

Years since birth. Measured exactly; it is the reference the other two are trained against or compared with.

years, with error

Estimated biological age

A model's prediction of age or of mortality risk, re-expressed in years (for example PhenoAge, GrimAge). A single-sample snapshot.

rate, with error

Pace of ageing

A model trained on how fast many organ-system markers changed over time in a cohort (for example DunedinPACE). Reported as years of ageing per calendar year.

Clocks respond differently to the same intervention

Two randomised trials make this concrete, and both are post hoc analyses of stored blood rather than trials designed around the clocks. In CALERIE, 220 adults without obesity were randomised to a target of 25 percent caloric restriction or to their usual diet for two years (Waziry et al., Nature Aging 2023). Caloric restriction slowed the pace of ageing measured by DunedinPACE, by a few percent, and produced no significant change in the biological-age estimates from PhenoAge or GrimAge. The authors describe the effect sizes as small. One reading of this is that a two-year intervention changes the current rate of ageing before it changes the accumulated estimate, which is what one would expect of a speedometer and an odometer. Another is that different clocks simply capture different biology. Both readings are defensible, and neither supports treating a single clock as the measure of ageing.

In DO-HEALTH, a trial of vitamin D, omega-3 fatty acids and a home exercise programme in adults aged 70 and over, a post hoc analysis of 777 participants examined PhenoAge, GrimAge, GrimAge2 and DunedinPACE over three years (Bischoff-Ferrari et al., Nature Aging 2025). Omega-3 alone slowed PhenoAge, GrimAge2 and DunedinPACE, and the three interventions together had an additive effect on PhenoAge. The standardised effects ranged from 0.16 to 0.32 units, which the authors translate into 2.9 to 3.8 months over three years. Three features deserve emphasis. The effects were small. They were clock-specific, with the original GrimAge not responding. And the analysis was post hoc with several clocks and several treatment combinations, so some of the pattern will be chance, as the authors acknowledge.

Set against these trials, a three-year change in one person in one year is an outlier by an order of magnitude. That alone should prompt the next question.

How much of a change is noise

A measurement is only as informative as it is repeatable. When the same blood sample is run twice, standard epigenetic clocks can disagree with themselves by several years; one systematic study found deviations of up to nine years between technical replicates for six prominent clocks, and showed that recalculating the clocks on principal components of the methylation data brought most replicates within 1.5 years of each other (Higgins-Chen et al., Nature Aging 2022). Platform matters too: in 100 technical replicates compared across an array and a sequencing panel, principal-component clocks differed between replicates by around one year, while uncorrected versions of the same clocks differed by as much as 15 to 20 years (Shokhirev and Johnson, GeroScience 2025).

Technical repeatability is the easier half. A 2026 analysis of 18 clocks separated it from biological reliability, meaning agreement between samples taken from the same person hours or days apart, before and after meals, under acute stress and across environmental exposures (Sehgal et al., Aging Cell 2026). Most clocks were technically reproducible and only moderately stable biologically, stability fell further after adjusting for immune cell composition, and the technically robust clocks were not necessarily the biologically stable ones. The same study showed the practical consequence: less reliable clocks produced variable or misleading estimates of intervention response, while more reliable ones gave consistent results.

For an individual, the implication is direct. Unless the report states which clock was used, whether a reliability-improved version was applied, and what the test-retest error is in that laboratory, a change of three years cannot be distinguished from a change of zero. A single repeat measurement cannot establish a trend. Blood cell composition shifts with infection, sleep and time of day, and many clocks move with it.

Association is not surrogate validity

Second-generation clocks predict future disease and death in cohorts, and that is a genuine scientific achievement. It does not follow that a treatment which lowers a clock reading has lowered the person's risk. For a biomarker to stand in for a clinical outcome in that way, the treatment-induced change in the marker must be shown to track the treatment-induced change in the outcome, and that evidence can only come from trials in which both were measured. Blood pressure and LDL cholesterol earned that status over decades of outcome trials. No epigenetic clock has it. A reading that falls after an intervention may reflect a change in white cell proportions, a change in one methylation module among many, or genuine slowing of ageing processes, and the clock cannot say which.

For the same reason, a lower biological-age estimate should never be converted into years of life gained. The clock was trained to predict risk in a population; it was not trained to measure the effect of changing that risk in one person, and the arithmetic that turns three clock-years into three calendar years has no basis in any trial.

What a longevity claim would need to show

The due-diligence questions follow. Which clock, which version, and on which platform? What is the laboratory's documented test-retest error, and is the claimed change larger than it? Was the change measured in a randomised comparison, or in people who chose the product and were measured before and after? Were several clocks reported, or only the one that moved? Is the clock's response to this kind of intervention known from trials such as those above, and in the same direction? And, the question that decides everything else, has any trial shown that moving this clock with this intervention changed a clinical outcome?

Until the last question has an answer, a younger biological age is an interesting measurement with a wide error bar. It is a reason to look more closely, and not a result to act upon.

This note draws on the work of Professor Tomasz Guzik on vascular ageing; the evidence base is examined in the longevity programme. Teams assessing a product claim can request an independent evidence review.