Biological Age Tests: The Reliability Problem Nobody Advertises

The full story

Split one blood sample in two, send both halves for the same first-generation methylation clock, and the two reports can come back years apart. Not because you aged between tubes - because the assay is noisy and the clock was never built to resist that noise. A paper in Nature Aging quantified the problem across six prominent epigenetic clocks and found deviations of up to nine years between replicate samples of the same blood.

Hold that number next to the way these results are presented. A consumer report hands back a biological age with a decimal point on it, which implies a precision of months. The published test-retest figure for the older clocks is measured in years, plural. That decimal is a rhetorical flourish the measurement cannot support.

What a methylation clock actually is

The underlying biology is real and rather beautiful. The original multi-tissue clock, described in Genome Biology, was built from 8,000 samples across 82 Illumina methylation datasets covering 51 healthy tissues and cell types, and rests on 353 CpG sites - specific positions in the genome where a methyl group is either present or absent. Methylation age comes out near zero in embryonic and induced pluripotent stem cells, tracks how many times cells have been passaged in culture, and was accelerated in all 20 cancer types examined by an average of 36 years. Whatever these clocks are measuring, they are measuring something.

What they are not is a thermometer. A thermometer gives the same reading twice.

The reliability problem, and its fix

The Nature Aging paper that measured the up-to-nine-year discrepancy also solved a large part of it. Instead of training a clock directly on individual CpG measurements, the authors retrained six clocks on principal components of the CpG-level data - a statistical compression step that keeps shared signal and discards site-by-site noise. Most replicate pairs then agreed within 1.5 years, and the retrained clocks were better at detecting clock associations, intervention effects and longitudinal trajectories.

A second team applied the same approach to a multi-region brain clock, reported in Aging, and confirmed the mechanism: combining principal-components analysis with regularised regression "reduces technical noise and greatly improves test-retest reliability", and the more reliable predictor showed stronger associations with clinical and pathologic Alzheimer's disease than earlier clocks.

Notice what that implies. Reliability is not a footnote about tidiness. A noisy clock does not merely give you a fuzzy answer - it hides real biology. Improving reliability made the associations appear.

Second generation, and pace of aging

The clocks sold today mostly are not the 2013 original. Two families followed it.

• Mortality-trained clocks. DNAm PhenoAge, described in Aging, was trained on a composite clinical phenotypic-age measure rather than on chronological age, and outperformed first-generation clocks for predicting all-cause mortality, cancers, healthspan, physical functioning and Alzheimer's disease. DNAm GrimAge, also in Aging, went further from chronology still: it is a composite of seven methylation-based surrogates for plasma proteins plus a methylation estimator of smoking pack-years, and in validation across thousands of individuals it outperformed existing clocks for time to death, time to coronary heart disease and time to cancer, with its age-adjusted form strongly associated with comorbidity count.
• Pace-of-aging measures. DunedinPACE, published in eLife, is a different object entirely. It was trained on within-individual decline across 19 organ-system integrity indicators measured at four timepoints over two decades in the Dunedin 1972-73 birth cohort, and it was deliberately restricted to probes with acceptable test-retest reliability. It reports a rate rather than an age, showed high test-retest reliability, was associated with morbidity, disability and mortality, and added prediction beyond GrimAge.

At the population level, the mortality signal is not in serious doubt. A meta-analysis in Aging pooled 13 cohorts totalling 13,089 individuals across three racial and ethnic groups and found that every measure of epigenetic age acceleration examined predicted all-cause mortality independent of chronological age, and after adjustment for additional risk factors. Estimates that incorporated blood cell composition performed best.

What the evidence does not show

It does not show that a clock can tell you whether your regimen is working. The cleanest demonstration is the CALERIE randomised trial, reported in Nature Aging: 220 adults without obesity assigned to 25% caloric restriction or an ad libitum diet for two years. The intervention slowed the pace of aging as measured by DunedinPACE, and produced no significant change in the biological age estimates from PhenoAge, GrimAge or the other clocks. Treatment effect sizes were small. Which clock the trial used determined whether the intervention appeared to work at all - and that is a two-year randomised trial of a well-defined intervention, which is a far stronger design than you retesting yourself after a supplement.

Nor does the field claim individual-level validity yet. A review in Nature Medicine concludes there is no consensus on how biomarkers of aging should be validated before translation to the clinic, and documents problems of comparability and generalisability across omic aging biomarkers in population studies. A companion expert survey in Nature Aging identified six key barriers to clinical translation and recommended, among other things, linking these markers to clinically actionable insights and validating biomarkers that are "robust and responsive at the level of individuals" - which is precisely the property a single consumer report assumes and first-generation clocks lack.

The regulatory backstop is also thinner than most buyers assume. The FDA states that direct-to-consumer tests for non-medical, general wellness or low-risk medical purposes are in general not reviewed by the agency before they are offered, that its review of higher-risk direct-to-consumer tests covers both analytical validity - whether a test accurately and reliably measures what it claims - and clinical validity, and that these tests are not a substitute for a traditional health care evaluation.

What to check before you buy

• Ask whether the clock is principal-component based. This is the single highest-yield question, because it is the published fix for the reliability problem and it is not universal across products.
• Ask for the test-retest deviation, in years. Vendors should be able to state how far apart two replicates of the same sample land. Up to nine years is the published figure for six first-generation clocks; within 1.5 years is what the retraining achieved.
• Prefer a pace measure if you intend to retest. DunedinPACE was built from within-individual change and screened for reliable probes, and it was the measure that moved in the randomised caloric-restriction trial. An age estimate is a worse instrument for tracking change than a rate estimate.
• Separate the two validity questions. Analytical validity asks whether the assay measures what it says; clinical validity asks whether that measurement predicts a health state. Most marketing answers the second with population citations and quietly skips the first.
• Then ignore the decimal. Read your result as a wide band, not a point, and read a single result as a hypothesis rather than a finding.

None of this makes methylation clocks a scam. They are the most interesting quantitative handle anyone has on biological aging, they predict mortality across cohorts, and the reliability problem has a documented engineering solution that good vendors have already adopted. The gap is between what the field can do with thousands of people and what it can honestly tell one person on a Tuesday. Buy the test because the biology interests you and you want a baseline. Do not buy it expecting a scoreboard, and be suspicious of any product whose number always moves in the direction you were hoping for.