Biological Age Tests Compared: Noise Alone Can Move Your Result 9 Years

TL;DR: When the same blood sample was run twice, six widely used epigenetic clocks disagreed with themselves by as much as 9 years (Higgins-Chen, Nature Aging, 2022). That one finding reorders the entire consumer market, because a test that cannot reproduce its own answer cannot tell you whether anything you did worked. Methylation-based pace-of-aging measures are the most defensible category, standard blood chemistry is the cheapest and most actionable, and consumer telomere tests fail on reproducibility. Our verdict: most readers should not buy any of these tests.

You spit in a tube, or a lancet takes a few drops of blood, and three weeks later a dashboard tells you that you are 42.7 years old when your passport says 49. It feels like a measurement. It is priced like a measurement. The question this article exists to answer is whether it behaves like one.

There is now a real consumer market for biological age testing, built on three genuinely different technologies that get marketed as if they were interchangeable. They are not. They differ in what they were trained to predict, in how much of their output is signal, and in whether a second test a year later can distinguish a real change from laboratory noise. Below is a side-by-side comparison built on the peer-reviewed literature behind each method, not on vendor validation pages.

Table of Contents

• What “Biological Age” Actually Means
• How We Compared These Tests
• Reliability Is the Axis That Decides Everything
• The Comparison Table
• Category 1: Epigenetic and DNA Methylation Clocks
• Category 2: Blood Biomarker Panels
• Category 3: Telomere Length Tests
• What a Result Can and Cannot Tell You
• The Verdict: Who Should Test, and Who Should Not
• If You Test Anyway: A Protocol That Reduces Noise
• FAQ
• References

What “Biological Age” Actually Means

Chronological age is a fact about the calendar. Biological age is a statistical construct: a model looks at something measurable in your body, compares it against a reference population, and reports the chronological age at which your measurement would be typical. Nothing in your cells knows its own age. The number is an output of a regression, and the honest question about any such number is what the regression was trained on.

That matters more than it sounds. A model trained to predict chronological age is, by construction, penalised for disagreeing with the calendar. A model trained to predict mortality is not. Those two models will give you different numbers from the same blood sample, and only one of them is trying to tell you something the calendar does not already say. We covered the underlying biology in more depth in our piece on how behaviour influences epigenetic aging, and the emerging organ-specific approach in methylation clocks for every organ.

How We Compared These Tests

Market snapshot date: 23 August 2026. Prices, product names and panel contents in this category change often, so treat any specific product mentioned below as an example of a category rather than a standing recommendation.

Inclusion criteria. A test qualified for this comparison if it met all four of the following: it can be bought by an individual without a physician’s order or through a consumer-facing partner lab; it returns a single biological age or pace-of-aging figure rather than only raw data; its underlying method is disclosed and named, meaning a specific published clock, a defined clinical chemistry composite, or a named telomere assay; and at least one peer-reviewed paper describes the algorithm or assay class it uses.

Exclusions. We left out research-only assays not sold to the public, wearable-derived “fitness age” or “heart age” scores that are computed from activity and heart rate rather than from a biological sample, and any product whose method is described only as proprietary. If a company will not name the clock or the assay, there is nothing to evaluate.

Comparison axes. Six, in order of weight: (1) test-retest reliability, meaning what happens when the same sample or the same person is measured twice; (2) what the underlying model was actually trained to predict; (3) the strength of the published outcome evidence linking the measure to death or disease; (4) responsiveness, meaning whether the measure can move over a realistic 6 to 24 month window; (5) actionability, meaning whether the output points at anything you can change; and (6) cost relative to the other categories.

What we did not do. We did not run any of these tests in-house. No sample was submitted, no vendor supplied us with data, no company reviewed this article before publication, and this page carries no affiliate links and no sponsored placements of any kind. That is deliberate. A comparison that earns a commission on the product it recommends is not a comparison. Vendor white papers and company blog posts are treated here as marketing claims, never as evidence. Every judgement below rests on the peer-reviewed literature about the method itself.

Reliability Is the Axis That Decides Everything

Most reviews of this category lead with accuracy, which is the wrong place to start. Accuracy asks whether the number is close to some truth we cannot observe. Reliability asks something answerable: run the same sample twice, do you get the same answer? If the answer is no, everything downstream collapses, because you can never tell whether next year’s improvement was your training block or the assay having a different day.

Albert Higgins-Chen and colleagues at Yale did exactly this test and published the result in Nature Aging in 2022. Running technical replicates through six prominent epigenetic clocks, they found that technical noise alone produced deviations of up to 9 years between replicates of the same sample. Not nine years of biological difference. Nine years of nothing, from the same DNA, processed twice. The authors’ own framing is blunt about what that does to the clocks’ usefulness (PMID 36277076).

They also published the fix. By computing principal components from the raw CpG-level data before predicting age, rather than feeding individual noisy probes into the model, their retrained versions of the same six clocks brought most replicate pairs into agreement within 1.5 years. This is the single most important practical fact in the whole category: a PC-adjusted clock and a standard clock can carry the same brand name and be different instruments.

The pace-of-aging measures were built with this problem in mind. Daniel Belsky’s DunedinPACE, published in eLife in 2022, reported an intraclass correlation coefficient of 0.96, with a 95% confidence interval of 0.92 to 0.98, in the Lehne replicate dataset (PMID 35029144). An ICC of 0.96 means roughly 96 percent of the variance between measurements is real between-person variance rather than measurement error. That is the number to ask any vendor for, and the number most of them will not put on the page.

Telomere testing has a parallel literature and it is not flattering. In an international collaborative study published in the International Journal of Epidemiology, Carmen Martin-Ruiz and colleagues had 10 laboratories using three techniques (Southern blotting, STELA and quantitative PCR) each measure 10 blinded human DNA samples across two rounds. Between laboratories, coefficients of variation averaged about 10 percent for Southern blotting and STELA, and more than 20 percent for qPCR, which is the method most consumer telomere tests use. Their conclusion was that this variation “severely limits the usefulness of data pooling and excludes sharing of reference ranges between laboratories” (PMID 25239152).

The Comparison Table

Category	What it measures	Trained to predict	Test-retest reliability	Moves in 6 to 24 months?	Actionable output?	Relative cost
First-generation methylation clocks (Horvath-type)	Methylation at hundreds of CpG sites	Chronological age	Poor unadjusted, up to 9 yr replicate deviation	Rarely, and slowly	No	High
Second-generation clocks (PhenoAge, GrimAge)	Methylation, weighted toward mortality-linked sites	Mortality and healthspan outcomes	Poor unadjusted, good in PC-adjusted versions	Sometimes, mostly via smoking and metabolic change	Partly	High
Pace-of-aging clocks (DunedinPACE)	Rate of change across organ systems	Speed of biological decline	Best in category, ICC 0.96	Yes, this is what it is built for	Partly	High
PC-adjusted clocks	Principal components of CpG data	Same targets, less noise	Most replicates within 1.5 yr	Yes	Partly	High
Blood biomarker panels (Phenotypic Age type)	9 routine clinical chemistry markers plus age	Mortality risk	Good, standard clinical assays	Yes, often within months	Yes, every input is treatable	Lowest
Consumer telomere tests (qPCR)	Average leukocyte telomere length	Age association, weakly	Worst in category, inter-lab CV over 20%	Not interpretably	No	Medium

Category 1: Epigenetic and DNA Methylation Clocks

This is the category most consumer marketing means when it says “biological age.” The technology reads how heavily methylated your DNA is at specific sites, and a trained model converts that pattern into a number. The field has three distinct generations, and conflating them is the most common error in consumer coverage.

First generation. Steve Horvath’s 2013 multi-tissue clock, published in Genome Biology, was built from 8,000 samples across 82 Illumina methylation datasets covering 51 healthy tissue and cell types (PMID 24138928). It was a genuine scientific landmark, and it is the wrong tool for a consumer. It was trained to predict chronological age, so the best it can do is agree with your birth certificate. Deviation from the calendar was an interesting residual for researchers, not a designed output.

Even so, that residual carried real signal. Riccardo Marioni and colleagues meta-analysed four longitudinal cohorts of older adults in Genome Biology in 2015 and found that a 5-year higher methylation age relative to chronological age was associated with a 21 percent higher all-cause mortality risk after adjusting for age and sex, and still a 16 percent higher risk after further adjustment for childhood IQ, education, social class, hypertension, diabetes, cardiovascular disease and APOE status (PMID 25633388). That is a population-level association, and it is the foundation the second generation was built on.

Second generation. Morgan Levine’s DNAm PhenoAge, published in Aging in 2018, changed the training target. Instead of predicting the calendar, it was trained through a two-step process against a composite clinical phenotype that captures differences in lifespan and healthspan, and it outperformed earlier measures on all-cause mortality, cancers, healthspan, physical functioning and Alzheimer’s disease (PMID 29676998). Ake Lu’s DNAm GrimAge, published in Aging in 2019, went further: it built methylation-based surrogates for seven plasma proteins plus a methylation estimate of smoking pack-years, then combined them into a lifespan predictor. Its association with time to death was extraordinarily strong, with a Cox regression p-value on the order of 10 to the minus 75, alongside significant prediction of time to coronary heart disease and time to cancer (PMID 30669119).

Third generation. DunedinPACE is not an age at all, which is why it is often misreported. It estimates a rate: how many biological years you are accumulating per calendar year, derived from longitudinal change across multiple organ systems in the Dunedin birth cohort. A result of 1.0 is average pace. Belsky’s team showed it associated with physical and cognitive decline, with morbidity, disability and mortality across cohorts, and with early-life adversity, while also showing the best reliability figures in the category (PMID 35029144). If you are trying to detect whether an intervention changed anything, a rate measure is structurally the right instrument, because a cumulative age measure has decades of unchangeable history baked into it.

The catch remains reliability. Unless a vendor explicitly says it uses principal-component-adjusted clocks or reports an ICC, assume you are buying the noisy version, and assume that a change of a few years between two of your own tests means nothing. Several of the interventions people buy these tests to evaluate have been studied directly against methylation endpoints, including omega-3 supplementation and DNA methylation age and dietary nucleotides and biological age, which is usually a better use of your attention than measuring yourself.

Category 2: Blood Biomarker Panels

The least glamorous option in this comparison is also the one with the strongest practical case. Zuyun Liu and colleagues, publishing in PLoS Medicine in 2018, showed that Phenotypic Age, calculated as a linear combination of chronological age and just nine routine multi-system clinical chemistry biomarkers, differentiated mortality and morbidity risk across diverse subpopulations in NHANES IV, including healthy and unhealthy groups, different age bands, and varied racial, ethnic and socioeconomic groups (PMID 30596641).

Nine standard markers. Assays that hospital laboratories have run for decades under external quality control, with known reference ranges and known coefficients of variation. No proprietary array, no bioinformatics pipeline you cannot inspect.

The advantage compounds when you consider actionability. If a methylation clock says you are aging quickly, it does not tell you why or what to do. If a blood panel says your fasting glucose, C-reactive protein and albumin are drifting the wrong way, each of those is a target with an evidence-based intervention behind it. The output is not just a score, it is a to-do list. And several of these markers respond within months rather than years, which makes a retest meaningful on a human planning horizon.

The honest criticism is that the “age” framing adds nothing informational. Converting nine lab values into a single year figure is a communication device, not a measurement advance. It is a good communication device, because “your metabolic profile matches a 58-year-old” changes behaviour in a way that a table of numbers does not. But you should know that you are paying for framing, and that most of these markers are already in a standard annual panel your physician can order.

Category 3: Telomere Length Tests

Telomeres are the nucleoprotein caps on the ends of chromosomes. They shorten with cell division and oxidative stress, short telomeres induce cellular senescence, and so telomere length became an intuitive candidate biomarker of aging. Intuitive is not the same as validated.

Jason Sanders and Anne Newman reviewed more than a decade of epidemiological work on leukocyte telomere length in Epidemiologic Reviews in 2013 and described the state of the field candidly: conflicting data had “generated heated debate about the value of LTL as a biomarker of overall aging” (PMID 23302541). Shorter leukocyte telomere length does track with older age and with several health variables, but the association at the individual level is loose, and prospective prediction of outcomes has been inconsistent.

Then there is the measurement problem, which for consumers is decisive. Recall the Martin-Ruiz collaborative study: inter-laboratory coefficients of variation above 20 percent for qPCR, the technique underlying most direct-to-consumer telomere kits, and no shareable reference ranges between laboratories (PMID 25239152). Practically, that means two different companies can measure the same person and place them in different percentiles, and a repeat test at the same company a year later cannot separate a real change from assay drift. We looked at how fragile the intervention literature is here in our article on vitamin D3 and telomere length.

Telomere biology is genuinely important science. Consumer telomere length testing, as currently sold, is the weakest product in this comparison and we would not spend money on it.

What a Result Can and Cannot Tell You

Three interpretation traps catch almost everyone who buys one of these tests.

Population associations are not individual predictions. Marioni’s 21 percent mortality figure describes what happens on average across thousands of people. It does not mean your personal risk rose 21 percent. Group-level hazard ratios of that size are compatible with enormous individual variation, and no published clock has been validated as an individual prognostic instrument.

Regression to the mean will fool you. If your first result is unusually bad, your second result will probably be better even if you changed nothing at all, purely because extreme measurements tend to be followed by less extreme ones. Any company that sells you a test, sells you a protocol, then sells you a retest has this statistical tailwind working in its favour whether or not the protocol does anything.

An age is a cumulative quantity, a pace is a current one. Decades of your history are already priced into a cumulative age measure, so even a genuinely excellent year moves it a little. A pace measure asks what your body is doing right now, which is the question a person changing their habits actually wants answered.

The Verdict: Who Should Test, and Who Should Not

We will state it plainly, because the whole point of an independent comparison is being able to reach an unwelcome conclusion: most readers should not buy any of these tests.

Here is the reasoning. The behaviours that reliably move biological aging markers are already known and none of them requires a test to justify: not smoking, keeping metabolic markers in range, sustained aerobic and resistance training, adequate sleep, a diet that is mostly whole foods, and managing chronic psychological stress. If you are not already doing those things, a number will not make you. If you are already doing them, the number will not tell you to do anything different. In both cases the test is priced entertainment with a biology theme.

Do not buy one if you want a diagnosis, you expect it to tell you which supplement to take, you plan to make decisions from a single measurement, or the test comes bundled with the seller’s own supplement subscription. That last one is a conflict of interest sitting in plain sight.

A test may be worth it if all of the following are true: you are running a deliberate, fixed intervention for at least 12 months; you will retest on the same platform with the same method; you are using a pace-of-aging measure or a principal-component-adjusted clock with a published ICC; you will ignore any change smaller than the published noise band; and you can afford to learn nothing. That is a narrow set of people, and if you are in it you already know why.

If you want one thing to spend money on, it is a standard comprehensive metabolic and lipid panel with an inflammatory marker, ordered through normal medical care. It is cheaper, better standardised, repeatable, interpretable by a clinician, and every abnormal value on it points at something you can actually change.

If You Test Anyway: A Protocol That Reduces Noise

• Pick the measure before the brand. Decide you want a pace-of-aging measure or a PC-adjusted clock, then find who sells it. Do not let the packaging choose for you.
• Ask for the ICC in writing. If a vendor cannot state the test-retest reliability of the exact measure they are selling, that is your answer.
• Standardise the draw. Same laboratory, same collection method, same time of day, same fasting state, same season if you can manage it. Every uncontrolled variable widens the noise band.
• Wait at least 12 months between tests. Shorter intervals mostly measure the assay.
• Set your interpretation rule before you see the result. Write down in advance the size of change you will treat as real, based on the published reliability figure, and hold to it.
• Change one thing at a time. If you start five interventions at once, a moved number tells you nothing about which one moved it.
• Never buy the upsell. A company that sells the measurement and the remedy has an incentive structure that does not serve you.

If you want the full system behind decisions like this one, including the biomarker panel I actually track on myself and why, I wrote it all down in The MVHK Protocol.

FAQ

Are biological age tests accurate?

Accuracy is unanswerable because there is no gold standard to check against. Reliability is answerable, and it is poor for most products. Technical noise alone moved six prominent epigenetic clocks by up to 9 years between replicates of the same sample (PMID 36277076).

Which biological age test is the most reliable?

On published test-retest figures, pace-of-aging measures lead. DunedinPACE reported an ICC of 0.96 (PMID 35029144), and principal-component-adjusted clocks bring most replicate pairs within 1.5 years (PMID 36277076). Standard clinical chemistry panels are also highly reproducible.

Is an epigenetic test better than a blood biomarker panel?

Not for most people. A nine-marker clinical chemistry composite predicted mortality and morbidity across diverse subpopulations (PMID 30596641), costs far less, and every input is directly treatable. Epigenetic clocks are more novel, not more useful.

Are at-home telomere tests worth it?

No. Inter-laboratory coefficients of variation exceeded 20 percent for qPCR, the method most consumer kits use, and reference ranges cannot be shared between laboratories (PMID 25239152). A repeat result cannot distinguish real change from assay drift.

How often should I retest?

No more than once every 12 months, on the same platform and method. Shorter intervals mostly capture measurement noise rather than biology, especially with cumulative age measures that change slowly by design.

My result said I am younger than my age. Does that mean I am healthy?

It means one model placed one sample below the population average. It is not a health screen, it does not detect disease, and it has not been validated as an individual prognosis. Do not let it replace ordinary medical care.

Can anything actually lower my biological age?

Several behaviours move these markers at population level, particularly smoking cessation and metabolic improvement, which drive much of GrimAge’s predictive power (PMID 30669119). Whether an individual’s number moves for individual reasons is much harder to establish.

References

• Horvath S. DNA methylation age of human tissues and cell types. Genome Biology, 2013;14:R115. PMID 24138928.
• Levine ME, et al. An epigenetic biomarker of aging for lifespan and healthspan. Aging (Albany NY), 2018. PMID 29676998.
• Lu AT, et al. DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging (Albany NY), 2019. PMID 30669119.
• Belsky DW, et al. DunedinPACE, a DNA methylation biomarker of the pace of aging. eLife, 2022;11:e73420. PMID 35029144.
• Higgins-Chen AT, et al. A computational solution for bolstering reliability of epigenetic clocks: implications for clinical trials and longitudinal tracking. Nature Aging, 2022;2. PMID 36277076.
• Marioni RE, et al. DNA methylation age of blood predicts all-cause mortality in later life. Genome Biology, 2015. PMID 25633388.
• Liu Z, et al. A new aging measure captures morbidity and mortality risk across diverse subpopulations from NHANES IV: a cohort study. PLoS Medicine, 2018. PMID 30596641.
• Martin-Ruiz CM, et al. Reproducibility of telomere length assessment: an international collaborative study. International Journal of Epidemiology, 2015. PMID 25239152.
• Sanders JL, Newman AB. Telomere length in epidemiology: a biomarker of aging, age-related disease, both, or neither? Epidemiologic Reviews, 2013. PMID 23302541.

This article is for information only and is not medical advice. It contains no affiliate links and no sponsored content. Talk to a qualified clinician before making decisions about testing or treatment.