What a biological age test actually measures

TL;DR

• There is no single biological age - several different statistical models share the name, built from different inputs and trained against different targets.
• Generation matters - Horvath and Hannum were trained on chronological age; PhenoAge and GrimAge on mortality composites; DunedinPACE on rate of decline rather than level.
• Technical reliability was largely solved - principal-component clock versions get replicates agreeing within about a year, against earlier deviations approaching a decade.
• Biological reliability was not - a 2025 bioRxiv analysis finds clocks move between ordinary days with no intervention at all, which sets a floor on detectable change.
• Population association is not individual prediction - second-generation clocks track mortality across cohorts while the distributions overlap heavily person to person.
• No clock is a validated surrogate endpoint - VO2max, grip strength and gait speed predict mortality and disability as well, at a fraction of the cost.

---

The premise problem: "biological age" is not one thing

The phrase invites a specific mental model - that somewhere in the body there is a true age, distinct from the number of birthdays, and that a test reveals it. That model is wrong, and almost every downstream confusion follows from it.

What actually exists is a set of statistical models, each built by taking some biological measurement, fitting it against some target, and expressing the output in years because years are intuitive. The unit is a presentation choice, not a discovery. Two tests can both report "biological age" and be measuring different molecules, trained against different outcomes, with different error characteristics, and they will disagree - not because one is broken, but because they were never measuring the same construct. The anatomy of that disagreement - how much of the gap is model design, how much is the reference population, and how much is the assay moving on its own - is worked through in why two biological-age tests disagree.

So the first thing to establish about any result is what it is made of.

The families of test, and what each is actually measuring

The table repays a second look, because the pattern in the right-hand columns is the opposite of the pattern in marketing attention.

The generational structure of epigenetic clocks matters. First-generation clocks were trained to predict chronological age. They do it well - which is precisely why their deviation from chronological age is a difficult signal to interpret. If a model's job was to output your birthday and it output something else, that gap is partly meaningful biology and partly model error, and separating the two is not trivial. Second-generation clocks solved this by changing the target: instead of predicting age, they were trained against mortality and clinical-biomarker composites, and they predict health outcomes considerably better as a result. Third-generation pace-of-aging measures changed the question again, estimating rate rather than level - arguably the more useful quantity, since what most people want to know is which direction they are travelling and how fast.

Phenotypic clocks are underrated. They are composites of measurements already sitting in most people's medical records. They are cheap, they are interpretable - you can see which input moved and why - and every component is something a clinician already knows how to act on. They lack the molecular glamour of methylation, which is close to their only real disadvantage in practice.

Telomere length is the clearest case of reputation exceeding performance. It was foundational to the field and it remains biologically real. But measurement variability is high, and individual-level prediction is weaker than its cultural prominence suggests. A single telomere measurement carries far less information than the framing usually implies.

Functional measures are the quiet outlier. VO2max, grip strength and gait speed are not sold as biological age tests, and they are among the best-validated predictors of mortality and disability that exist. They are also directly modifiable by things a person can actually do, and the change is measurable in a way that does not depend on assay reliability. If you had to pick one number to track and one to change, the evidence points here rather than at a methylation array. Our epigenetic clocks piece goes deeper on the molecular side; this article is about what the whole category can and cannot support.

Reliability: the question vendors answer least often and the one that matters most

This is the section that should change how a reader reads any result, and it turns on a distinction that is almost never made in consumer-facing material.

Technical reliability

Technical reliability is reproducibility across replicate assays of the same sample. Split one tube of blood, run it twice, compare.

Most clocks do reasonably well here on standard arrays. But not uniformly: some show marked drops in reproducibility depending on slide position in the array and on DNA extraction protocol - that is, on laboratory handling rather than on the person. First- and second-generation clocks contain a substantial number of individually unreliable CpG sites, and deviations approaching a decade between technical replicates of the same sample have been reported.

The field's response was direct and largely successful. Principal-component (PC) versions of the major clocks were developed specifically to address this, aggregating across correlated CpGs rather than depending on individual noisy sites. PC clocks achieve very high technical reliability, with most technical replicates agreeing within about a year. This is a genuine methodological advance, and whether a given result came from a PC clock or a non-PC clock materially changes how much weight it can bear.

Biological reliability

Biological reliability is stability across repeated collections from the same person over hours or days - during which the person eats meals, experiences ordinary stress, moves through their circadian cycle, sleeps badly, and encounters minor exposures. Nothing that would count as an intervention. Just living.

This is where the problem is. Recent work - including a 2025 bioRxiv preprint, Biological versus Technical Reliability of Epigenetic Clocks - makes exactly this point: most clocks achieve excellent technical reliability, but their biological instability limits prognostic and interventional utility. Fixing the assay did not fix the measurement, because the remaining variability is not coming from the laboratory.

The consequence, stated plainly

If a measure moves by several years between Tuesday and Thursday with no intervention, then a several-year "improvement" measured after an intervention is not interpretable as an effect.

That sentence is the practical core of this article. It is not an argument that the clocks are meaningless - they are not. It is an argument about what a single reading, or a pair of readings, in one person can support. Within-person short-term variability sets a floor on detectable change, and if a claimed effect is smaller than that floor, the measurement cannot distinguish it from noise. This applies regardless of how impressive the underlying science is.

What each is correlated with, and how strongly

Second-generation clocks show real associations with mortality and morbidity at population level. This is well replicated and should not be dismissed. Across large cohorts, people whose clock reads older than their chronological age have, on average, worse subsequent outcomes.

The word doing the work is average.

Population-level association and individual-level prediction are different quantities, and the gap between them is one of the most consistently collapsed distinctions in consumer health marketing. A marker can show a robust, highly significant association across thousands of people and still have weak discriminative power for any single person - because the distributions of outcomes for "clock reads old" and "clock reads young" overlap heavily even when their means differ reliably.

This is standard statistics, not a criticism unique to aging clocks. The same applies to cholesterol, blood pressure and most risk markers. The difference is that clinical medicine generally handles the distinction explicitly, with absolute risk framing and decision thresholds, while biological age marketing typically presents a population-level association as though it were a personal statement.

So: the clocks are measuring something real about population aging. What they tell an individual about their own trajectory is considerably weaker than the number's precision implies. A result reported to one decimal place is not more informative for being reported to one decimal place.

What a result can and cannot tell you about an intervention

This is where the two preceding sections combine, and the answer is unambiguous.

A biological age result cannot currently serve as a validated surrogate endpoint for an individual's response to an intervention.

Epistemic-uncertainty work has specifically challenged the reliability of aging clocks for detecting rejuvenation effects - that is, the exact use case in which they are most often deployed. The problem is not that the clocks fail to move. They move. The problem is that they also move without any intervention at all, and a measurement that fluctuates on its own cannot attribute a change to a cause.

To be concrete about what this rules out and what it does not:

What a result can reasonably support: a rough population-referenced sense of where you sit; a prompt to look at inputs - particularly in phenotypic clocks, where a driving abnormal lab value is visible and clinically actionable; and research use, where large samples and control groups absorb the individual-level noise.

What a result cannot support: a claim that a specific intervention worked for you; a before-and-after comparison used as evidence; a decision to continue or stop something based on the number having moved; or a statement about your personal rate of aging read off a single measurement.

That last group covers most of what these tests are actually sold to do.

What we would want to see before treating a number as an outcome measure

It is worth being constructive rather than merely sceptical. There is a well-defined evidentiary path from "interesting marker" to "usable outcome measure", and it is the same path any surrogate endpoint has to walk. Five things would need to be in place.

Demonstrated within-person short-term stability. Repeated sampling in the same individuals across days, establishing how much the measure moves on its own. This defines the floor below which no claimed effect is detectable. Until it is characterised for a given clock, no effect size is interpretable.

Pre-registered analysis. The analysis plan specified before the data are seen. Clocks are numerous, and a study free to choose which clock to report after the fact will find one that moved.

A control group. Ageing markers drift, regress toward the mean, and respond to season, illness and life circumstance. Without a comparison group, you cannot distinguish an intervention effect from what would have happened anyway.

Demonstrated surrogate validation - not correlation. This is the demanding one, and it is the one most often skipped. It is not enough that the marker correlates with outcomes across a population. It must be shown that a change in the marker corresponds to a change in an actual clinical outcome. Correlation establishes that a marker tracks a process; surrogate validation establishes that moving the marker means the process moved. Medicine has an instructive history of surrogates that correlated beautifully with outcomes and then, when moved by drugs, failed to deliver the benefit - or delivered harm.

Independent replication. By groups without a stake in the result, on different samples, with different handling.

None of this is unreasonable, and none of it is unattainable. It is simply the standard any measurement has to meet before it is treated as an outcome, and the field is not there yet.

Frequently asked questions

Which type is most accurate? "Accurate" needs unpacking, because the tests answer different questions. For predicting the birthday you already know, first-generation methylation clocks are impressively precise and correspondingly uninformative. For association with health outcomes at population level, second-generation clocks (trained against mortality and clinical composites) outperform first-generation ones. For estimating rate rather than level, pace-of-aging measures are conceptually better suited. For technical reproducibility, principal-component versions of the major clocks are clearly superior. And for actually predicting mortality and disability, functional measures - VO2max, grip strength, gait speed - hold up as well as anything, at a fraction of the cost. No single answer wins across all four criteria.

Can I lower it? You can lower the number. That is not the same as changing your aging, and establishing the second requires the surrogate-validation chain described above - which does not currently exist for these markers. It is worth sitting with that distinction rather than moving past it, because the entire commercial proposition rests on eliding it. Worth noting too: the interventions with the strongest evidence for the underlying outcomes - not the marker - are unglamorous and thoroughly familiar. Cardiorespiratory fitness. Resistance training. Sleep. Not smoking. Controlling blood pressure and glycaemia. These have outcome evidence, not just marker evidence, and none of them require a test to justify.

Is it worth the money? That depends entirely on what you expect from it. As a curiosity, or as a prompt to look harder at inputs you were ignoring, it may be. As a measurement you intend to act on, or to re-test after an intervention to see whether the intervention worked, the reliability picture above says the money is buying a number that cannot support that use. If the budget is limited and the goal is information that changes decisions, a standard blood panel, blood pressure, and a measured VO2max or grip strength will tell you more for less - and every one of them is interpretable by a clinician.

Do doctors use these? Not as standard clinical tools. What actually informs care is routine bloods, blood pressure, established risk calculators, imaging where indicated, and functional assessment. Biological age markers currently sit in research and in consumer testing rather than in clinical decision-making, and the reason is the one this article has laid out: they have not completed the validation chain that would let a clinician act on a change in the number. That may change. It has not changed yet. Our notes on optimal vs normal lab ranges and biomarker monitoring cover the measurements that do currently inform decisions.

Related reading

• Epigenetic clocks: the molecular detail
• Can anything actually move an epigenetic clock? - what the randomised evidence shows, and how much of a change is measurement noise
• Organ-specific aging clocks: why two biological age tests disagree
• Functional measures of biological age - VO2max, grip strength and gait speed, and how they compare on evidence with the molecular tests
• Optimal vs normal lab ranges: what the reference interval actually means
• Biomarker monitoring: what to track and why
• Mitochondria and aging: mechanism versus measurement
• Telomere length testing - what a consumer telomere test can and cannot tell you
• "Fitness age" on a wearable is not a biological-age test - what the device score is computed from

Part of Testing & Biological Age: The Measurement Guide.

Explore it with verified supply. Every Vivre Labs lot is independently tested and the certificate of analysis is published, so you know exactly what is in the vial.

---

This article is educational and does not constitute medical advice. No commercial biological age test providers are named, recommended or ranked. No test described here is validated for tracking an individual's response to an intervention, and no compound is positioned as a means of altering a result. Interpretation of any laboratory measurement belongs with a qualified clinician who has your history. Claims for the Malaysian market are subject to MAB requirements.

References

• Horvath and Hannum first-generation DNA methylation clocks - trained to predict chronological age.
• PhenoAge and GrimAge second-generation clocks - trained against mortality and clinical-biomarker composites.
• DunedinPACE - third-generation pace-of-aging measure estimating rate rather than level.
• Principal-component (PC) versions of major epigenetic clocks - developed to address technical reliability; technical replicates typically agreeing within approximately one year.
• bioRxiv (2025) - Biological versus Technical Reliability of Epigenetic Clocks: most clocks achieve excellent technical reliability, but biological instability limits prognostic and interventional utility.
• References on CpG-level reliability, slide-position and DNA-extraction effects on methylation clock reproducibility, including reported technical-replicate deviations approaching a decade.
• Epistemic-uncertainty literature challenging the reliability of aging clocks for detecting rejuvenation effects.
• General references on leukocyte telomere length measurement variability and individual-level predictive performance.
• General references on VO2max, grip strength and gait speed as predictors of mortality and disability.
• General methodological references on the distinction between population-level association and individual-level prediction, and on surrogate endpoint validation.