Biological Age Tests Disagree: What GrimAge and DunedinPACE Measure
Biological age tests can give the same person different answers. Here is what GrimAge, PhenoAge and DunedinPACE actually measure, and how to read a report without panic.
The report usually arrives as a PDF. You open it on your phone at the kitchen table on an ordinary Tuesday, and before reading a word your eye lands on one number in a large, friendly font: your biological age. A few years under your birthday, and you feel a small glow. A few years over, and the whole afternoon tilts.
I understand the pull. There is a young child in our house, and under almost every health article I read sits the same quiet question: how many good years do I get with this little person, and can I earn a few more? A single number that claims to sum up your whole body speaks straight to that question, which is exactly why it deserves a slower look.
This is not medical advice, and nothing here recommends buying or avoiding any test or company. If a result worries you, take it to a clinician along with the ordinary measurements they can act on.
What the big font leaves out is that there is no one biological age. There are several clocks, built to answer different questions, and they can disagree even on the same tube of blood.
What Each Clock Actually Measures
Every test in this family reads DNA methylation, small chemical tags on DNA that help decide which genes are active. Their pattern drifts with age consistently enough to model. A clock is an algorithm that turns readings from many of those sites into one output, and what it was trained to predict is what your number really means.
First generation: Horvath and Hannum
The Horvath and Hannum clocks both appeared in 2013, trained to predict chronological age from methylation. That sounds circular, since you already know how old you are, but the interesting part is the leftover. If the clock says 52 and your passport says 48, the gap is called age acceleration. The catch is that a model rewarded for matching the calendar learns to smooth over differences between same-age people, which are the differences you care about.
Second generation: PhenoAge and GrimAge
PhenoAge (Levine and colleagues, 2018) and GrimAge (Lu and colleagues, 2019) changed the target. They were trained against health phenotypes and mortality risk, so the question moved from how old you are to how old your risk looks.
GrimAge is the odder build. It combines methylation stand-ins for several blood proteins with a methylation estimate of smoking pack-years. GrimAge2 adds stand-ins for C-reactive protein (CRP), an inflammation marker, and HbA1c, a measure of average blood sugar. It behaves more like a composite risk score wearing a clock's clothes.
Third generation: DunedinPACE
DunedinPACE (Belsky and colleagues, 2022) was trained on how fast multiple organ systems declined over time in the same people in the Dunedin cohort. Its output is not an age. It is a pace, in biological years per calendar year.
Pace of Aging vs. Accumulated Damage: An Age Number vs. a Rate
The way I keep this straight is a car dashboard. First- and second-generation clocks work like an odometer: they give an age, and the gap from your calendar age reads like extra miles on the engine. DunedinPACE works like a speedometer. A reading of 1.0 means one biological year per calendar year. On the model's terms, 1.1 would be about 10 percent faster and 0.9 about 10 percent slower.
An odometer describes the whole trip so far, while a speedometer describes the stretch of road you are on now. A high-mileage car can be cruising gently today, so an older-looking age next to an ordinary pace is not necessarily a contradiction. The two numbers also live in different units, so resist the urge to convert one into the other.
In principle, a rate is the more natural thing to watch if you want to know whether something is changing. Whether any particular habit moves a particular clock is a separate question, and a single test cannot settle it.
Why Results Conflict, Even on the Same Sample
When two numbers disagree, the instinct is to ask which one is right. The better question is what each was built to measure, and how it was measured. Four things pull results apart.
Different training targets
A clock trained to guess calendar age and one trained to predict mortality risk are answering different questions. When they disagree about you, that is expected behavior, not a malfunction.
Sample type
Blood is made of immune cells. Saliva, according to Steve Horvath, who built the clock that carries his name and now works at Altos Labs, is roughly 65 percent immune cells and 35 percent cheek epithelial cells, and those cell types carry different methylation profiles. Apply a blood-trained clock to saliva without correction, he told The Scientist, and "the errors can be substantial."
Lab platform
Illumina methylation arrays and whole-genome bisulfite sequencing both read methylation, and their results are correlated but not identical. Horvath's larger point is sobering: it is not yet possible to send a sample to different validated labs and get the same readout. The same piece notes that most commercial companies don't disclose which clock they use.
Technical noise
This one surprised me most. In a 2022 Nature Aging study, Higgins-Chen and colleagues measured the same samples twice. Across six major clocks, technical noise alone produced deviations of 3 to 9 years between those replicates. Nothing about the person changed. Only the measurement did.
Software work left me with a reflex that fits here. When a test fails once, you don't rewrite the code. You rerun it, because flaky tests fail for reasons unrelated to your change. A single clock result is one run. The same team rebuilt the clocks from principal components, pooling signal across many sites instead of leaning on individually noisy ones, and those PC versions brought most replicates within 1.5 years. That is a real fix, and it shows how wide the band was before.
How to Read a Report Without Over-Reacting
Say the PDF is open and one number stings. Heartfulness practice has taught me something that helps here: let the first reaction rise and settle before deciding what it means. Then go slowly.
- Find out which clock made each number. An age and a pace are different quantities. If the report doesn't name the clock behind a result, give that result less weight.
- Check the sample type. Look for a statement that the clock was built or corrected for the tissue you sent.
- Treat one result as one run. A gap of a few years is smaller than the swings technical noise alone produced between replicates. It is a data point, not a verdict.
- Remember who the clocks were built for. Horvath describes them as tools designed for epidemiological studies, not for individuals. Population patterns are real, but your own path is not written in a single blood draw.
- Bring the basics to a clinician. Blood pressure, lipids, glucose and fitness are measurements a doctor can interpret and actually act on. If a clock finally gets you to book that appointment, it has done something useful.
A Consumer's Guide to the Main Clock Types
Most kits report several clocks, so treat this table as a decoder for the page you receive. The evidence column leans on the largest head-to-head comparison so far, published in Nature Communications on December 16, 2025, which tested 14 clocks against the 10-year onset of 174 diseases in 18,859 people from the Generation Scotland cohort.
| Clock | Trained to predict | What the number means | Evidence for health outcomes |
|---|---|---|---|
| Horvath (2013) | Calendar age | An age; the gap from your real age is age acceleration | Limited for disease; first-generation clocks showed little use in the 2025 comparison |
| Hannum (2013) | Calendar age | An age, read the same way | Limited for disease, for the same reason |
| PhenoAge (2018) | Health phenotypes and mortality risk | An age reflecting how your risk profile looks | Stronger; second-generation clocks significantly outperformed the first |
| GrimAge and GrimAge2 (2019 onward) | Mortality risk, via protein and smoking stand-ins (GrimAge2 adds CRP and HbA1c) | An age; a larger gap means a riskier-looking profile | Strongest single link: GrimAge2, hazard ratio 1.54 per standard deviation |
| DunedinPACE (2022) | Speed of decline across organ systems | A rate; 1.0 is one biological year per calendar year | Stronger; third-generation clocks also significantly outperformed the first |
That hazard ratio means people whose GrimAge2 acceleration sat one standard deviation higher developed the linked outcome at roughly a 54 percent higher rate. The study also found 27 diseases, including primary lung cancer and diabetes, where a clock's link to the disease was stronger than its link to death from any cause, and GrimAge's built-in smoking surrogate may help drive its ties to respiratory disease. All of this describes groups, not your personal odds.
A note on cost. The clocks were published in research journals. What you pay for is the kit, the lab work and the report, so price tells you about the product around a clock, not the strength of the science inside it. At the time of writing, 2026 review sites list TruDiagnostic's TruAge at about $499 one time, or about $249 on subscription, and Tally Health's TallyAge cheek-swab test at about $229. Prices change, and listing them is not a recommendation.
Frequently Asked Questions
Which biological age clock is the most accurate? Accurate for what? First-generation clocks were built to match calendar age. For links to future disease across large populations, second- and third-generation clocks did better in the 2025 comparison, with GrimAge2 strongest, but none was built to predict one person's future.
My report says I'm biologically older than my age. Should I worry? Not on the strength of one result. Technical noise alone moved standard clocks by 3 to 9 years when the same sample was measured twice. Use it as a nudge to check blood pressure, lipids, glucose and fitness with a clinician, not as a diagnosis.
Can I compare results from two different companies? Not reliably. Kits may differ in clock, sample type and lab platform, and even validated labs don't yet produce matching readouts. If you ever retest to track change, keeping the same test and sample type removes some variables, though not the noise.
Is a saliva or cheek-swab test worse than a blood test? Not automatically. Those cells carry different methylation patterns from blood, so the clock has to be built or corrected for them. The useful question is whether the clock was validated on the sample you are sending.
When I picture that PDF on the kitchen table, the number I would most want isn't on it. It is the count of ordinary evenings still ahead with a child who believes bedtime is open to negotiation. No clock reports that. It gets built slowly, out of the unglamorous numbers a doctor can already measure and help you move.