Biological age tests have become the front door to longevity medicine: send a blood or saliva sample, receive a number, feel either vindicated or alarmed. The underlying science is real and the papers behind it are serious. The gap that matters is between what those papers established — that these measures predict mortality and disease across populations — and what the product implies, which is that the number is precise enough to track your own progress from one test to the next. On that second claim, the literature is unusually blunt.
What the clocks actually measure
DNA methylation is a chemical mark on the genome that changes with age in a patterned way. In 2013, Steve Horvath built a multi-tissue age predictor from 8,000 samples spanning 82 datasets and 51 healthy tissue and cell types, resting on 353 CpG sites. It read close to zero in embryonic and induced pluripotent stem cells, correlated with cell passage number, and showed significant age acceleration across all 20 cancer types examined — averaging 36 years.[1] That paper is the reason the field exists.
The first-generation clocks were trained to predict chronological age, which caps how useful they can be: a perfect predictor of the number on your birth certificate tells you nothing you did not already know. The second generation changed the training target. PhenoAge was built on composite clinical measures of phenotypic age, and outperformed earlier measures for all-cause mortality, cancers, healthspan, physical functioning and Alzheimer’s disease.[2] GrimAge went further still, assembling DNA methylation surrogates for seven plasma proteins plus a methylation-based estimate of smoking pack-years; it stands out among epigenetic clocks for predicting time-to-death, time-to-coronary-heart-disease and time-to-cancer.[3]
DunedinPACE is a different object again, and the distinction is worth holding onto. Rather than estimating how old you are, it estimates how fast you are aging — modeled from within-individual decline in 19 indicators of organ-system integrity tracked across four time points over two decades in the Dunedin 1972–73 birth cohort, then distilled into a single blood test. It was associated with morbidity, disability and mortality, with effect sizes similar to GrimAge.[4]
| Measure | Trained to predict | What it is good at |
|---|---|---|
| Horvath clock (2013) | Chronological age across tissues | Proving the concept; age acceleration in cancer tissue |
| PhenoAge (2018) | Composite clinical phenotypic age | All-cause mortality, cancers, healthspan, physical functioning |
| GrimAge (2019) | Plasma protein surrogates plus smoking pack-years | Time-to-death, time-to-coronary-heart-disease, time-to-cancer |
| DunedinPACE (2022) | Rate of decline across 19 organ-system indicators | Pace of aging; high test-retest reliability by design |
The number that should change how you read your result
In 2022 a team examined the technical reliability of the clocks themselves — not whether they predict mortality, but whether they give the same answer twice on the same sample. Technical noise alone produced deviations of up to 9 years between replicates across six prominent epigenetic clocks, which the authors described as limiting their utility.[5]
9 years
Maximum deviation between replicates of the same sample from technical noise alone
Higgins-Chen 2022, Nature Aging
1.5 years
Agreement between most replicates after the principal-component fix
Higgins-Chen 2022, Nature Aging
353
CpG sites in the original Horvath multi-tissue clock
Horvath 2013, Genome Biology
Sit with what that implies for a consumer product. If you test, spend six months on an intervention, and test again, a three-year improvement is entirely consistent with having changed nothing at all. The same paper offers the remedy — computing principal components from CpG-level data before predicting biological age brings most replicates within 1.5 years and improves detection of intervention effects — but a retail report rarely states which implementation produced your number.[5] DunedinPACE is a partial exception by construction, since it was trained on a methylation dataset restricted to exclude probes with low test-retest reliability, and reported high test-retest reliability as a result.[4]
How to read a test you have already taken
Three questions make a result interpretable. First, which clock produced it — a chronological-age clock and a mortality-trained clock answer different questions, and only the second is about health. Second, is it an age or a pace: DunedinPACE-style output is a rate, so a value near 1.0 means aging at roughly one year per year, and it is not comparable to a number of years. Third, was a reliability-corrected implementation used, because that determines whether a small change between two tests means anything.
And a fourth, practical point: nothing about a biological-age result tells you which intervention to take. The compounds most often sold alongside these tests have their own, much thinner evidence bases — see NAD⁺ precursors NR and NMN, rapamycin and the PEARL trial, metformin for longevity and taurine, or compare the whole field on our evidence matrix. The intervention with the least controversial evidence behind it remains the one that also treats age-related muscle loss.
For the trials that actually attempted to slow aging in humans and in animals rather than measure it, the reference points are the CALERIE caloric-restriction trial and the NIA Interventions Testing Program.
The honest bottom line
Epigenetic clocks are a genuine scientific achievement, and the second-generation ones predict mortality and disease well enough to be taken seriously as research tools. As a personal dashboard they are far weaker than they look: six prominent clocks disagreed with themselves by as much as nine years on replicate samples, which is larger than almost any effect an intervention could plausibly produce in a year. Test if you are curious. Do not let a change between two readings decide anything.