Biological Age Tests: What They Measure

Three hundred and fifty-three numbers. That is where this whole industry starts — 353 positions on the genome where a methyl group is either present or absent, read off a chip, multiplied by coefficients, summed, and handed back to you as an age. Not measured. Computed. The multi-tissue predictor at the origin of most of what follows was built from 8,000 samples across 82 Illumina methylation array datasets spanning 51 healthy tissues and cell types, and its author described exactly what it was for: estimating the DNA methylation age of most tissues and cell types.

The hardware is dull. A tube of blood, or a saliva sample, or a card of dried blood. DNA out. An array — in the cohort studies, usually the Illumina Infinium Methylation EPIC BeadChip — reports, at each of hundreds of thousands of cytosines, roughly what proportion of the cells in the sample carry a methyl mark at that spot. Then regression. Elastic net, mostly. A model trained on somebody else's cohort takes your fractions, applies the weights it learned there, and returns one number in units of years. The units are the model's. They are not a measurement of anything physical inside you. A second-generation clock does something subtly different and much more consequential: it never tries to guess your birthday at all. It is trained on death, or on disease, or on the rate at which nineteen organ-system indicators drifted in a birth cohort tracked across two decades.

So here is the question a reader holding a credit card actually has. If you buy an epigenetic age test, what have you bought — a health measurement, a risk estimate, or a number with an interesting provenance? The most defensible answer the published evidence currently permits is narrow and worth stating plainly: epigenetic age acceleration is a reproducible population-level predictor of mortality and, less consistently, of some age-related disease, demonstrated across tens of thousands of people; the newer clocks predict better than the older ones; and almost none of that evidence establishes that an individual's single reading is precise enough, stable enough, or actionable enough to change what a specific person should do next week. The gap between those two statements is where the entire consumer market lives.

What a biological age test actually measures

Strip the branding and there are three things being sold under one label.

The first is an epigenetic clock: a weighted sum of DNA methylation levels at a specified set of CpG sites. The specification matters more than the marketing. In one head-to-head validation, the three algorithms compared were built on 513, 1,030 and 10 CpGs respectively — the same biological substrate, wildly different footprints, and they are not interchangeable. The output is a number of years, or in the case of pace-of-aging measures, a rate: how many biological years you appear to accumulate per calendar year.

The second is a clinical-chemistry composite. PhenoAge in its original clinical form is computed from chronological age plus nine routine blood biomarkers; in a cohort of 609 patients with multivessel coronary artery disease, that equation alone reached an area under the curve of 0.705 for all-cause mortality and outperformed chronological age and the individual chemistry measures. There is a methylation version trained to reproduce it, which is a slightly strange lineage: a chip predicting a blood panel predicting mortality.

The third is a telomere length assay, usually quantitative PCR on leukocyte DNA, returning a relative length rather than a length in base pairs. We will come back to why this one is the hardest of the three to read as a single score.

Note what none of them measure. None of them measure damage, or reserve, or how long you will live. They measure a pattern that correlates with the passage of time and, in the better cases, with the hazard of dying, in the populations where the pattern was learned.

Biological age vs chronological age: what the gap is claiming

The gap — usually called age acceleration — is the residual after regressing predicted age on calendar age. Positive residual, you are “older than your years.” The claim embedded in that phrasing is that the residual captures something biologically real and heterogeneous between people of the same birth year.

There is decent evidence the residual is not noise. It is heritable enough that genome-wide association studies of epigenetic aging biomarkers, drawing on methylation and SNP data from more than 40,000 individuals, identified 137 genome-wide significant loci, of which 113 were novel; a polygenic score for GrimAge acceleration was associated with adiposity-related traits, educational attainment, parental longevity and C-reactive protein. It also tracks social gradients in a way that is hard to dismiss. Across eight prospective cohorts totalling 13,021 participants, men with low education had 57 excess deaths per 10,000 person-years (95% CI 38 to 76) compared with more advantaged men, a hazard ratio of 1.33 (1.12 to 1.57); GrimAge accounted for roughly half of that inequality in men. For women the estimate was not distinguishable from zero: 4 deaths per 10,000 person-years (95% CI −11 to 19), hazard ratio 1.15 (0.96 to 1.37), and the mediated proportion was negligible. Most of the mediation ran through differences in lifestyle and morbidity, not through some pure ageing signal.

Read that carefully, because it cuts both ways. A marker that mediates half the educational mortality gradient in men is measuring something consequential. A marker whose mediation runs mainly through smoking, body composition and existing illness is, to a large extent, a sophisticated restatement of things you may already know about yourself.

The clocks, in the order of what they were built to predict

The generations are not versions of one product. They were trained on different outcomes, and that training choice determines almost everything about how they behave.

First generation — the 2013 multi-tissue predictor and its blood-based contemporaries — were trained on chronological age. That was the point: show that methylation could recover calendar age accurately across tissue types. It could. The predictor sits near zero for embryonic and induced pluripotent stem cells, correlates with cell passage number, and applies to chimpanzee tissue. Across 6,000 cancer samples from 32 datasets, all 20 cancer types examined showed significant age acceleration, averaging 36 years. That is a striking finding about tumours. It is not a finding about how fast a healthy person is ageing.

Second generation changed the target. PhenoAge was trained in two steps on a composite clinical phenotype that captures differences in lifespan and healthspan, and its authors reported that it outperformed prior measures across all-cause mortality, cancers, healthspan, physical functioning and Alzheimer's disease, with higher epigenetic age relative to chronological age associated with increased activation of pro-inflammatory and interferon pathways. GrimAge went further and stopped pretending to be an age at all: it is a composite of seven DNA methylation surrogates for plasma proteins plus a methylation-based estimator of smoking pack-years, calibrated to time-to-death. Its reported associations are large — time-to-death Cox P = 2.0E-75, time-to-coronary-heart-disease P = 6.2E-24, time-to-cancer P = 1.3E-12, comorbidity count P = 3.45E-17.

GrimAge version 2 added methylation estimators of C-reactive protein and haemoglobin A1C and was evaluated in 13,399 blood samples across nine cohorts, where it outperformed the original for mortality prediction after adjustment for age and sex (meta P = 3.6e-167 versus P = 2.6e-144). It was trained on individuals aged 40 to 92. Its methylation A1C component is strongly associated with type 2 diabetes (P = 5.8e-155); its PAI-1 component correlates with triglycerides (r = 0.34) and visceral fat (r = 0.41). Those correlations are the tell. A large part of what a mortality-trained clock knows about you is your smoking history, your inflammatory state and your metabolic health.

Third, and different in kind: DunedinPACE does not estimate an age. It estimates a rate. It was built by tracking within-individual decline in 19 indicators of organ-system integrity across four time points spanning two decades in the Dunedin 1972–1973 birth cohort — one cohort, one two-year birth window, extended to five further datasets — then distilling that two-decade trajectory into a single-timepoint blood test using elastic-net regression on a methylation dataset deliberately restricted to exclude probes with low test–retest reliability. That last design decision is the most important sentence in this article. In five additional datasets it showed high test–retest reliability, associations with morbidity, disability and mortality, effect sizes similar to GrimAge, and incremental prediction beyond GrimAge.

How accurate are biological age tests?

Accurate at what? Split the question, because the honest answers differ.

At predicting mortality across a population, the evidence is real and replicated. A meta-analysis pooling 13 cohorts and 13,089 individuals across three racial and ethnic groups found every measure of epigenetic age acceleration predicted mortality independently of chronological age (P ≤ 8.2e-9) and after adjustment for additional risk factors (P < 5.4e-4); estimates that incorporated blood cell composition produced the most statistically significant associations (P = 7.5e-43). A separate systematic review and meta-analysis covering 23 articles and 41,607 participants put a number on the effect size: each 5-year increase in DNA methylation age was associated with an 8% to 15% increase in mortality risk. The same review concluded that the association with age-related disease and longevity was inconclusive, cited heterogeneity in design and outcomes, and flagged positive publication bias as a live concern. Both of those sentences come from the same paper. Consumer marketing tends to quote only the first.

At ranking clocks against each other, GrimAge has generally won and first-generation clocks have generally lost. In the ESTHER cohort, with 17 years of follow-up and 458 deaths, hazard ratios per standard deviation were 1.32 (1.19–1.46) for PhenoAge acceleration, 1.47 (1.32–1.64) for GrimAge acceleration, and 1.73 (1.49–2.01) for a methylation mortality risk score originally built on just 10 CpGs — the smallest assay produced the largest hazard ratio, which should temper any assumption that more probes means more signal. In 490 participants in the Irish Longitudinal Study on Ageing followed up to 10 years, Horvath and Hannum age acceleration were not predictive of health at all; PhenoAge acceleration was associated with 4 of 9 outcomes in minimally adjusted models but with none after adjustment for social and lifestyle factors; GrimAge acceleration was associated with 8 of 9 outcomes, retaining walking speed, polypharmacy, frailty and mortality in fully adjusted models.

At predicting specific diseases, the picture is thinner. In a pooled analysis of three population-based cohorts — 5,600 participants, mean age 65.5, 905 incident cases of atrial fibrillation over a mean 12.9 years — a 5-year increase in GrimAge acceleration was associated with a 19% higher hazard (1.09–1.31) and PhenoAge acceleration with 15% (1.05–1.25) after multivariable adjustment. Horvath acceleration did not survive adjustment. And the Mendelian randomisation analyses in that same paper, testing whether the relationship is causal, found no significant association between genetic instruments for any of the five epigenetic age measures and atrial fibrillation. The authors' own conclusion was that a potential underlying causal relationship remains unclear.

At adding information beyond cheaper measures, the increment is modest. This is the number consumers should care most about and see least often. In ESTHER, a mortality model using age, sex and leukocyte composition achieved a Harrell's C-statistic of 0.710. Adding an 8-CpG mortality score and a frailty index raised it to 0.759. Adding PhenoAge and GrimAge acceleration on top of that raised it to 0.766. The frailty index is a questionnaire and a clinical assessment. A separate high-resolution analysis of immune-cell fractions in blood found that a predictor built only from cell-type fractions reached a C-index of 0.69 (0.67–0.72) for all-cause mortality — and 0.83 (0.80–0.86) once ordinary epidemiological risk factors and baseline comorbidity were added. For scale, a framework using large language models to read routine health examination reports across six cohorts and more than 10 million participants reported a C-index of 0.757 (0.752–0.761) for all-cause mortality, which the authors state outperformed telomere length, a frailty index, eight epigenetic ages and four machine-learning models. Ordinary clinical data is a stubbornly strong competitor.

Why two tests give you two different ages

Because the clocks disagree, and because the assay is noisier than the reported precision implies.

Take the technical problem first, since it is the one a results page is least likely to make visible. Across six prominent epigenetic clocks, technical noise alone produced deviations of up to 9 years between replicate samples — the same DNA, run twice. Nine years. A computational fix exists: instead of feeding individual CpGs into the age model, compute principal components from the CpG-level data and use those. Retrained principal-component versions of the same six clocks brought most replicate pairs into agreement within 1.5 years, improved detection of intervention effects, and produced more coherent longitudinal trajectories. Reliability is not a cosmetic property. It is a precondition for the number meaning anything about you rather than about a population.

Then the biological problem. Blood is a mixture, and the mixture moves. A functional genomics analysis of four clocks using blood methylome and transcriptome data from 3,132 individuals found that clock CpGs are associated in trans with gene expression enriched for T-cell processes, and traced a substantial part of the signal to the difference between naive and activated T and NK cells. Within the same donor, the four clocks predicted naive cells to be up to 40 years younger than activated cells. A clock reading blood is, in significant part, reading your immune cell composition — which is itself informative, but it is not a global measure of tissue ageing.

Third, the clocks were built on different targets and behave accordingly. In an 18-month randomised dietary trial that computed six different methylation clocks plus DunedinPACE on the same 256 participants, baseline correlations of the clocks with chronological age ranged from 0.83 to 0.95, and the interventions produced no difference between arms in change on the clocks as a group; the signal that did appear, tied to adherence, appeared in two specific clocks and not the others. In a preprint reanalysis, when six clocks were deliberately stress-tested across longitudinal datasets containing either an intervention or an age-accelerating event, clocks trained on mortality or rate-of-aging captured events consistently, including in high-reliability retrained versions and after multiple-testing correction. Clocks trained on chronological age produced sporadic changes that did not replicate in high-reliability versions of the same clocks and did not survive multiple-testing correction — likely false positives, and the authors noted that some of those changes had already been published and that the literature should be re-examined.

So: two vendors, two clock choices, two array runs, two immune profiles. Two ages. This is expected behaviour, not a defect in one vendor's pipeline.

Where can you get a biological age test, and what does epigenetic testing cost?

This section is a market description, not a research finding, and it should be read as such. The abstracts underlying every other claim in this article contain no pricing data whatsoever. Nothing below about cost, turnaround or vendor practice is supported by the studies cited here; it is a structural observation about how the market is organised, and prices move.

Structurally there are three routes. Research participation is how nearly all of the evidence above was generated: cohort studies and trials run methylation arrays on banked samples, and participants are not customers. Clinician-ordered testing routes a venous blood draw through a laboratory, with a physician or specialist clinic interpreting the result alongside a standard panel. Direct-to-consumer kits ship you a collection device — saliva, or a dried blood spot from a fingerstick — which you return by post; results appear in a web dashboard weeks later. As a market observation, consumer epigenetic age kits commonly sit in a band of roughly $200 to $500, with subscription retesting offered at intervals; clinician-ordered panels bundling methylation with laboratory chemistry typically sit above that. Again: not a study finding.

Two things from the literature are relevant to the delivery model. First, sample type is not cosmetically neutral but is not disqualifying either — GrimAge version 2 was reported to apply to saliva samples, where it tracked markers of metabolic syndrome. Second, the direct-to-consumer genomics market has a documented interpretation problem. In a survey of 1,513 Brazilian direct-to-consumer genetic testing customers, 87.2% bought in to learn about disease predispositions, 66.9% felt confident interpreting their results without help, and only 18.4% consulted a healthcare professional about them. In a qualitative study of 15 people who did seek genetic counselling after direct-to-consumer testing, expectations of the session were high and participants wanted their results pooled with personal and family history into a plan; the authors' recommendations included explaining how consumer testing differs from clinical genetic testing and why further testing may not be indicated. Those studies are about genotyping rather than methylation clocks, and that limit should be stated. The behavioural pattern — confident self-interpretation, little clinical contact — is the part that transfers.

What an at-home biological age test can and cannot tell you

The clock a home kit runs is the clock the vendor licensed or built. So the honest boundary is drawn by the clock and by the reliability of the run, not by whether you spat into a tube in your kitchen.

What a kit can do, fairly: give you an index that is correlated with a set of risk factors you can also measure more cheaply, and give some people a reason to pay attention. Whether attention converts into anything is an open question the literature has not answered. In the Brazilian consumer survey, 64.0% reported altering habits after results, with dietary adjustments in 38.7%, regular check-ups in 24.2% and increased physical activity in 16.5% — self-reported, uncontrolled, and about genotyping rather than clocks, but the closest the pool comes to a behavioural outcome.

Can you change the number?

Yes, in at least one randomised trial, in one specific measure, by a small amount. That sentence is doing a lot of careful work.

The strongest evidence is a post hoc analysis of the CALERIE trial, in which 220 adults without obesity were randomised to 25% caloric restriction or an ad libitum control diet for two years. Caloric restriction slowed the pace of aging as measured by DunedinPACE. It did not produce significant changes in biological age estimates from the various methylation clocks, including PhenoAge and GrimAge. The authors describe the treatment effect sizes as small and state explicitly that a conclusive test of the geroscience hypothesis will require trials with long-term follow-up establishing effects on primary healthy-ageing endpoints, including incidence of chronic disease and mortality. No trial in this pool has done that. A change in DunedinPACE is not a demonstrated gain in healthy years; it is a change in a surrogate whose relationship to clinical outcome has been shown observationally, not experimentally.

The dietary trials are mixed and instructive. In DIRECT PLUS, 256 participants with abdominal obesity or dyslipidaemia were randomised for 18 months to healthy dietary guidelines, a Mediterranean diet, or a polyphenol-rich green Mediterranean diet, the latter adding green tea at 3–4 cups a day and a 500 ml Mankai shake for an additional 800 mg of daily polyphenols on top of the 28 g of walnuts (440 mg/day) common to both Mediterranean arms. The headline result deserves quoting precisely: the interventions did not differ in change on the methylation clocks. What the authors found was an association between greater adherence to the green Mediterranean diet and lower 18-month relative change in two of the seven clocks computed — Li (β = −0.41, P = 0.004) and Hannum (β = −0.38, P = 0.03) — with green tea intake (β = −1.57, P = 0.0016) and Mankai (β = −1.8, P = 0.061) as the contributing components. Participants in either Mediterranean arm showed roughly an 8.9-month favourable difference between observed and expected Li methylation age at the end of the intervention (P = 0.02). That is an adherence analysis inside a randomised trial, in a minority of the clocks tested. It is a hypothesis, not a result.

The MACRO trial is the cold shower. In 144 participants with obesity randomised to low-carbohydrate or low-fat weight loss for 12 months, DunedinPACE at baseline was significantly associated with insulin, HOMA-IR, total cholesterol, HDL cholesterol, C-reactive protein, adiponectin and ghrelin. After the intervention, only C-reactive protein and adiponectin remained significant. Crucially, changes in the epigenetic aging measures were not significantly associated with changes in the cardiometabolic biomarkers, and did not mediate the effects of weight loss. The authors' own framing is that this raises questions about the utility of epigenetic clocks as causal targets in short-term lifestyle interventions and underscores the need for caution in treating epigenetic aging as a modifiable intervention target.

Observationally, lifestyle and the number move together. In NHANES data with mortality follow-up to the end of 2019, a high Life's Essential 8 score (80–100) was associated with a PhenoAge 3.30 years younger, and PhenoAge mediated an estimated 36% of the effect of Life's Essential 8 on all-cause mortality and 22% on cardiovascular mortality; each 1-year increment in PhenoAge carried a hazard ratio of 1.04 (1.04–1.05) for all-cause mortality independent of chronological age. Mediation analysis in a cross-sectional exposure is not an experiment. The direction is consistent, the causal claim is not established, and the largest single mediated component was diet at 30%.

One more caution about intervention studies specifically. The clocks most likely to produce a headline in a small trial are the ones least likely to be reporting a real effect: chronological-age-trained clocks generated sporadic, non-replicable movements in reanalysis, while mortality-trained and rate-of-aging clocks behaved consistently. If a supplement study reports that Horvath age fell, the appropriate first hypothesis is technical noise.

Telomere length tests, and what they actually predict

Telomere assays point in different directions depending on which outcome you ask about.

They are not worthless. In an individual-participant-data meta-analysis of 12,199 adults aged 43 to 75 from two prospective cohorts, with 2,882 deaths, the age-adjusted hazard ratio for the shortest versus longest leukocyte telomere length quintile was 1.23 (1.04–1.46) for all-cause mortality — but 1.29 (0.83–2.00) for cardiovascular mortality and 1.10 (0.88–1.37) for cancer mortality, neither of which excludes the null. In 4,174 NHANES participants without cardiovascular disease followed a median 217 months, the highest telomere length group had a hazard ratio of 0.65 (0.54–0.78) for all-cause and 0.64 (0.45–0.93) for cardiovascular death versus the lowest.

In a different population the direction reverses. In 26,540 middle-aged and older Chinese adults with 4,353 incident cancers and 7,609 deaths, participants in the longest decile of telomere length had a 26% higher risk of total cancer incidence (95% CI 11% to 44%) than the shortest decile, while longer telomeres were associated with lower non-cancer death (HR 0.81, 0.71–0.92), including death from chronic obstructive pulmonary disease and pneumonia (0.79, 0.70–0.89) and digestive disease (0.60, 0.42–0.88). All-cause mortality was not significantly reduced (HR 0.93, 0.84–1.03). A biomarker where “longer is better” for one outcome and “longer is worse” for another is not a scalar you can put on a dashboard next to a smiley face.

Causal tests have been unkind too. A two-sample Mendelian randomisation study of telomere length and pulmonary arterial hypertension found no significant causal association (OR per standard deviation 1.229, 95% CI 0.469–3.222, P = 0.676), with null findings repeated across granulocyte, lymphocyte, naive and memory T-cell, B-cell and natural killer cell telomere lengths.

Where telomere length still earns its place is as a targeted clinical test in a defined disease context, and even there the operating characteristics are sobering. Using a whole-genome-sequencing telomere length nomogram built in 9,940 community-dwelling adults and smokers, 28% of 825 patients in an idiopathic pulmonary fibrosis registry fell below the 10th percentile for age-adjusted telomere length, versus 5% and 10% in the reference populations. Age adjustment improved discrimination of telomere-related qualifying variants only modestly, from an area under the curve of 0.62 (0.56–0.69) to 0.66 (0.60–0.73). Specificity ran 0.74 to 0.99; sensitivity ran 0.10 to 0.58. Being below the threshold was associated with shorter three-year transplant-free survival (HR 1.35, 1.05–1.74) but did not improve survival discrimination over clinical variables alone. That is what a well-characterised, disease-specific telomere test looks like: useful for ruling in, poor at ruling out, and not additive to clinical variables for predicting survival.

What we would want before treating one of these numbers as a health metric

Five things, none of which are exotic.

• Published test–retest reliability for the specific product, not the clock family. Nine years of replicate deviation is the documented baseline problem; principal-component reformulation narrows most replicates to within 1.5 years and requires no replicate samples to train. Any vendor reporting a change of two or three years between annual tests, without stating their replicate variance, is reporting an unknown mixture of signal and array noise.
• Reporting of cell composition. Clock signal traces substantially to naive versus activated T and NK cell fractions — up to a 40-year predicted difference between those populations in the same donor — and estimates incorporating blood cell composition produce the most statistically significant mortality associations. A result without a cell-composition estimate is missing a large part of its own explanation.
• Demonstrated incremental value over cheap alternatives. The relevant comparison is not clock-versus-nothing but clock-versus-your-existing-blood-panel-and-a-frailty-assessment. In ESTHER that increment was 0.759 to 0.766 in C-statistic. A large language model reading routine health examination reports reached a C-index of 0.757 for all-cause mortality and, per its authors, outperformed telomere length, frailty index and eight epigenetic ages.
• Randomised evidence that moving the number moves a clinical outcome. No trial reviewed here provides it. CALERIE moved DunedinPACE over two years without moving PhenoAge or GrimAge, and its authors said outright that a conclusive test requires long-term trials with chronic disease incidence and mortality as endpoints. MACRO found that changes in three epigenetic aging measures were not associated with changes in cardiometabolic biomarkers and did not mediate weight loss effects.
• Pre-specification and multiple-testing discipline in any study a vendor cites. With six or more clocks available, the probability of one moving by chance is not small, and reanalysis has already identified published clock changes that look like false positives.

There is a broader direction of travel worth noting, because it suggests the methylation clock may not be the endpoint of this field. Adapted pace-of-aging methods built from longitudinal blood biomarkers, physical measurements and functional tests in the US Health and Retirement Study and the English Longitudinal Study of Ageing — combined n = 19,045 — showed strong and consistent prospective associations with incident morbidity, disability and mortality without requiring a methylation array at all. If a validated pace-of-aging measure can be assembled from a blood panel, physical measurements and functional tests, the case for a methylation array becomes a case about convenience and precision, and it will have to be argued on those terms.

The defensible bottom line for someone with the credit card out: an epigenetic age test is a legitimate research instrument sold as a personal metric, and the translation has not been earned. If you buy one, buy it knowing that the number's population-level meaning is well documented, its individual-level precision is documented to be poor unless the vendor has done specific work to fix it, and no trial reviewed here has shown that changing it changes a clinical outcome.

Questions people actually ask

Where can I get my biological age test? Structurally there are three routes, described above: research participation, where cohort studies run methylation arrays on banked samples and participants are not customers; clinician-ordered testing, where a venous draw goes through a laboratory and is read alongside a standard panel; and direct-to-consumer kits, which post you a collection device and return a result in a web dashboard. No study cited here evaluates a particular provider.

What is epigenetic age? A number in units of years produced by a model, not a measurement of anything physical inside you. Methylation fractions at specified CpG sites are multiplied by coefficients learned in somebody else's cohort and summed; the units belong to the model.

What is an epigenetic clock? A weighted sum of DNA methylation levels at a specified set of CpG sites. In one head-to-head validation the three algorithms compared were built on 513, 1,030 and 10 CpGs respectively, and they are not interchangeable.

How does biological age test work? DNA is extracted from blood, saliva or a dried blood spot and read on a methylation array, which reports roughly what proportion of the cells in the sample carry a methyl mark at each of hundreds of thousands of sites. Those fractions go into a regression model trained on another cohort, which returns one number. A telomere assay works differently: usually quantitative PCR on leukocyte DNA, returning a relative length rather than a length in base pairs.

Biological age vs chronological age test: what is the gap? The gap, usually called age acceleration, is the residual after regressing predicted age on calendar age. It is not simply noise — it is heritable and it tracks social gradients — but most of what it mediates runs through lifestyle and existing illness rather than through a separate ageing signal.

How much does epigenetic testing cost? None of the studies cited here report pricing, and the section above flags its cost figures explicitly as a market observation rather than a research finding. On that flagged basis only: consumer kits commonly sit in a band of roughly $200 to $500, clinician-ordered panels typically above it, and prices move.

Is there a biological age test near me? Local availability is not something the evidence reviewed here speaks to. The structural choice is the same wherever you are: a clinician-ordered blood draw processed by a laboratory, or a direct-to-consumer kit posted to you and returned by post.

Can I do a biological age test at home? Home kits collect saliva or a fingerstick dried blood spot. What the result can support is set by the clock the vendor runs and by the reliability of that run, not by where the sample was taken. GrimAge version 2 was reported to apply to saliva samples, where it tracked markers of metabolic syndrome.

Which biological age test is best? For predicting mortality across a population, mortality-trained clocks have generally outperformed first-generation ones: in ESTHER, hazard ratios per standard deviation were 1.32 for PhenoAge acceleration and 1.47 for GrimAge acceleration, while in 490 TILDA participants Horvath and Hannum acceleration were not predictive of health at all. For following one person over time the property that matters is test–retest reliability, which results pages rarely report.

Is biological age testing accurate or useful? Accurate at ranking a population by mortality risk, yes, and replicated across cohorts. Accurate for one person on one sample, much less so: technical noise alone produced deviations of up to 9 years between replicates across six prominent clocks. Useful in the sense of telling you to do something differently has not been established — no trial reviewed here has shown that changing the number changes a clinical outcome.

Related on Magellan: Why aging clocks disagree · Do GLP-1 drugs slow biological aging? · Do healthy adults need a CGM?.

Educational, not medical advice.