Biological age: what the tests measure and what your numbers actually tell you

Most commercial tests rest on epigenetic age clocks. What gets measured is DNA methylation: chemical marks at specific positions in the genome that shift across a lifetime. An algorithm weights a few hundred of these positions and converts them into a single figure in years.

The first of these clocks came from Steve Horvath in 2013 and uses 353 measurement points. It was trained to predict chronological age, which it does with a median absolute error of 3.6 years. Later generations went further: PhenoAge (Levine, 2018) incorporates clinical lab values, and GrimAge (Lu, 2019) combines seven methylation-based protein surrogates with an estimated smoking history. DunedinPACE (Belsky, 2022) is the only one that measures a rate rather than an age, derived from twenty years of longitudinal change across nineteen organ systems.

The construction already explains part of what comes out. A substantial share of GrimAge's signal is smoking behaviour and inflammatory proteins. Smokers reliably score higher, which is information a two-minute conversation also produces.

In short

• Epigenetic clocks measure methylation patterns and translate them into a figure in years via an algorithm.
• The first generation was trained on calendar age; later ones on health trajectories.
• Part of the signal reproduces known risk factors, smoking above all.

For judging any single result, a paper by Higgins-Chen and colleagues in Nature Aging (2022) is the decisive one. The group ran the same blood samples through the same clocks repeatedly. Technical noise alone produced deviations of up to nine years between replicate measurements of the same person across six widely used clocks. The authors then built principal-component-based versions that bring most replicates within 1.5 years of each other.

That fits a probe-level analysis published in Patterns (Sugden et al., 2020): across the whole array, mean reliability of individual measurement points sits at an intraclass correlation coefficient of 0.21, with a median of 0.09. Clock probes perform better than background, and each algorithm still contains many unstable points.

What this means in practice: a three-year improvement after a six-month programme sits entirely inside the documented technical noise band of the standard clocks. For tracking change over time, the principal-component variants and DunedinPACE are currently the defensible choices; DunedinPACE reports a test–retest reliability of 0.96.

In short

• Repeat measurements of the same sample can differ by up to nine years on common clocks.
• Longitudinal comparison requires methods with demonstrated reliability, because otherwise laboratory variation dominates the result.
• DunedinPACE and principal-component clocks are markedly more stable.

At group level, the clocks work. An analysis of fourteen clocks across 18,859 people and 174 incident diseases (Mavrommatis et al., Nature Communications, 2025) found 176 significant associations. The gain over conventional risk factors stayed small: the best performer, GrimAge v2, lifted all-cause mortality prediction from an AUC of 0.851 to 0.865. Only 32 of the findings improved classification by more than one percent.

A review by Bell and colleagues (Genome Biology, 2019) puts the consultation-room consequence plainly: the clocks can predict all-cause mortality "at a population, but not individual level." The same paper notes there is no evidence that the measurement points used are enriched for biological function. The clocks describe a pattern that accompanies ageing, without explaining a mechanism.

One further limitation belongs here. We found no published study validating any commercial direct-to-consumer biological age test for use in individuals, and no published head-to-head comparison of different providers on the same samples. That gap is worth knowing about before a result becomes the basis for decisions.

The same hour of diagnostics can be filled differently. The five measurements below rest on large cohorts, have clear reference points, and connect directly to things you can change.

Cardiorespiratory fitness (VO₂max)

The largest study covers 122,007 people undergoing exercise treadmill testing over a median 8.4 years (Mandsager et al., JAMA Network Open, 2018). Adjusted all-cause mortality for elite versus low fitness gave a hazard ratio of 0.20 (95% CI 0.16–0.24). Even elite versus high fitness still returned 0.77 (0.63–0.95). No upper limit of benefit was visible in this cohort.

A meta-analysis of 33 studies (Kodama et al., JAMA, 2009) quantifies the step size: each additional MET is associated with 13 percent lower all-cause mortality (RR 0.87; 0.84–0.90) and 15 percent lower risk of coronary events.

Context: both datasets are observational. People referred for treadmill testing are a selected group, and existing disease lowers fitness just as the reverse holds.

Grip strength

The PURE study followed 139,691 people across seventeen countries for a median four years (Leong et al., The Lancet, 2015). Each five-kilogram reduction in grip strength was associated with 16 percent higher all-cause mortality (HR 1.16; 1.13–1.20) and 17 percent higher cardiovascular mortality. Per standard deviation, grip strength predicted death from any cause more strongly than systolic blood pressure (1.37 versus 1.15).

Honesty requires the other side. PURE found no significant association between grip strength and incident diabetes, hospital admission for pneumonia or COPD, fall injury, or fracture. And for cardiovascular disease specifically, blood pressure carried more information than grip strength.

ApoB and lipoprotein(a)

The ESC/EAS dyslipidaemia guideline (Mach et al., European Heart Journal, 2020) recommends ApoB measurement for risk assessment at Class I, particularly with high triglycerides, diabetes, obesity, metabolic syndrome, or very low LDL-C. ApoB counts atherogenic particles while LDL-C measures the cholesterol inside them, and in exactly those constellations the two diverge.

For Lp(a), the recommendation is to measure it once in a lifetime. The 2025 focused update classifies levels above 50 mg/dL (105 nmol/L) as risk-modifying in all adults. An elevated value can move someone from moderate into high risk.

One point for calibrating expectations: the guideline states explicitly that a benefit of lowering Lp(a) on cardiovascular risk has not yet been demonstrated. Lp(a) identifies risk and is not currently a validated treatment target.

HbA1c

In the ARIC cohort of 11,092 adults without diabetes or cardiovascular disease (Selvin et al., NEJM, 2010), adjusted hazard ratios for later diagnosed diabetes rose across HbA1c categories from 1.00 (5.0–5.5%) to 1.86, 4.48 and 16.47. For coronary heart disease the figures were 1.23, 1.78 and 1.95. After adjustment for HbA1c, fasting glucose lost statistical significance.

A distinction belongs here: that is prognostic value once a measurement exists. Whether HbA1c screening in asymptomatic adults adds benefit beyond a plain glucose measurement is a separate question, and Germany's IGeL-Monitor rates it "unclear" for want of studies.

Cystatin C

The CKD Prognosis Consortium analysed 90,750 participants from eleven general-population studies (Shlipak et al., NEJM, 2013). An eGFR derived from cystatin C markedly improved risk classification over the creatinine-based estimate, with a net reclassification improvement of 0.23 (0.18–0.28) for all-cause death. Because creatinine depends on muscle mass, this matters most for very muscular and very low-muscle-mass people, where the standard formula can miss in either direction.

hsCRP appears in almost every longevity panel. The Emerging Risk Factors Collaboration quantified what it adds (NEJM, 2012): adding CRP to conventional risk factors raised the C-index by 0.0039. Translated, that is roughly one additional cardiovascular event prevented over ten years for every 400 to 500 intermediate-risk people screened. The US Preventive Services Task Force (JAMA, 2018) issued a Grade I statement for hsCRP, ankle-brachial index and coronary artery calcium score: insufficient evidence to weigh benefits against harms in asymptomatic adults.

hsCRP stays useful when a specific question sits behind it, such as interpreting an unexpected ferritin or tracking known inflammation. As a routine addition to a screening panel, the gain is small.

Our approach to longevity questions runs in four steps.

1. History before labs. Family patterns, smoking history, sleep, training volume and medication determine which measurements will have any consequence at all.
2. Function belongs in the panel. VO₂max and strength values come from device diagnostics in the practice, and they are the two measurements most directly responsive to training.
3. A base panel with consequences attached. ApoB, Lp(a) once, HbA1c, cystatin C, full blood count, ferritin alongside an inflammation marker, liver and thyroid values. Every value gets an answer in advance to the question of what changes at which result.
4. Agree targets and a re-test date. A baseline without a defined target range and without a scheduled follow-up stays a snapshot.

An epigenetic clock can sit alongside all of this if the question interests you. It should then be a variant with demonstrated reliability, and the result belongs alongside the other findings as one of several.

The age clocks are a serious research instrument with real predictive power across large groups. For the question of how fast you are ageing, they currently produce a number whose uncertainty exceeds the effects you are trying to detect.

If you want a picture of your coming decades, your fitness, your strength, your lipoproteins, your glucose metabolism and your kidney function are the measurements with the most solid data behind them. They can be measured reproducibly, compared against clear references, and shifted through training, nutrition and, where indicated, medication.

If you already have results, epigenetic ones included, bring them to your first appointment. We go through them together and decide which measurements carry a consequence in your situation.