Your biological age result would have been different if you had eaten lunch

An epigenetic clock estimates how fast you are ageing from patterns of DNA methylation in a blood sample. The good ones do this with impressive technical consistency. Run the same sample twice and you get almost the same number. This is the fact the industry quotes, and it is true.

It is also the wrong reliability question.

What a buyer wants to know is not whether the machine agrees with itself, but whether the number describes them stably enough that a change in it means something changed in them. A body of repeated-measures work has now answered that question, and the answer is uncomfortable.

Two kinds of reliability, and only one is good

Take repeated samples from the same person across ordinary conditions: before and after a meal, under acute stress, after exposure to air pollution. Nothing exotic, nothing that would plausibly alter the rate at which somebody ages over a lifetime. Under those perturbations, epigenetic age estimates fluctuate substantially, and most clocks reach only moderate reliability at best.1

The finding that should reorganise how these tests are sold comes next. Technical reproducibility did not predict biological stability. The two properties came apart entirely. GrimAgeV2 and DunedinPACE, the clocks with the strongest technical credentials and the ones deliberately engineered around high-reliability CpG probes, were among the most biologically fragile.1 DunedinPACE’s own development paper documents its restriction to 173 CpG sites selected for high test-retest reliability, and it delivers exactly that.2 The engineering worked. It simply did not buy the property that matters.

SystemsAge and principal-component-based clocks came out both technically stable and comparatively more biologically reliable.1 No consumer product we are aware of puts that distinction in front of the person paying for the test.

The inference that breaks

At population level these clocks are not junk, and calling them that would be lazy. DunedinPACE is associated with morbidity, disability and mortality, with effect sizes in the same range as GrimAge.2 It predicts incident metabolic syndrome in the Berlin Aging Study II5 and tracks accelerated cognitive ageing in the Framingham cohort.4 Across thousands of people the signal is real and points the right way.

The problem lies in the step from there to the product. A cohort association tells you that a group with faster clocks does worse than a group with slower ones. It does not license the two inferences these tests are actually bought for: that this is my pace of ageing, and that my pace of ageing improved because of what I did.

The second is the expensive one. Retesting to evaluate an intervention requires measurement error to be small relative to the effect you are hunting. If the estimate moves materially depending on whether the draw happened before or after lunch, a six-month comparison cannot separate signal from sampling. You will get a number. It will be a different number. Neither you nor the company that sold it to you will know why.

These are honest instruments being asked a dishonest question. Almost nothing in how they are marketed reflects the difference.

Why grade D

Grade D means marketed ahead of its evidence, with unreliable measurement or an unsupported inference. Consumer biological age testing manages both at once, which is rare.

The scope of the grade matters, because the researchers building these tools deserve better than a blanket dismissal. The underlying science is serious and improving, and recent reviews mapping clocks from laboratory to lifestyle applications are candid about the translation gap.3 The grade attaches to the commercial product: a single-timepoint estimate sold to an individual with an implied promise that the number is theirs and that they can move it.

Where these clocks are being sold

The grade above applies to a single-timepoint estimate sold direct to an individual. It is worth being specific about where that leaves the clinical programmes, because several of them include the same instruments.

Epigenetic clocks appear in the upper tiers of the premium screening market. The €16,900 Ultimate programme at YEARS in Berlin lists DunedinPACE and GrimAge by name, which are precisely the two clocks that came out most biologically fragile in the repeated-measures work above.6 Fountain Life and other membership programmes include comparable epigenetic panels.

Two things follow, and they pull in different directions.

A clock inside a supervised programme is a different object from a mail-order result. It arrives next to 230-plus biomarkers, imaging and a physician who can say what weight to put on it, which is the context that stops a noisy number becoming a decision. YEARS also publishes a transparency section stating that it discloses sensitivity, specificity and positive predictive value for its screening methods and will tell patients where the study evidence is thin.7 A provider that includes a weakly-evidenced measure while saying out loud that the evidence is weak is in a coherent position. That is not the same as the marketing problem this article describes.

The grade does not move, though. Including DunedinPACE at the top of a price list implies the number is worth €9,300 more than not having it, and on the reliability evidence it is not. If these clocks are being sold as context rather than as findings, the honest presentation is to report them with their within-person confidence interval and to state that they are not suitable for tracking whether an intervention worked. We would regrade any programme that did.

What would change the grade

Three things, none requiring a scientific breakthrough.

Report a confidence interval. Every clock has a known within-person variance under ordinary conditions and every result could carry one. “Your pace of ageing is 1.08, 95% CI 0.94 to 1.22” is harder to sell and far more honest than “1.08”.

Standardise the draw. Fasting state, time of day and recent acute stress are all controllable. Almost nobody controls them, because controlling them is inconvenient for a mail-order business model.

Disclose and switch the clock. If SystemsAge and PC-based clocks are more biologically stable, consumer tests should use them and should state which clock produced the result. Many disclose nothing.

Until at least the first of those arrives, treat a biological age result as a conversation piece and price it accordingly. Be sceptical of any programme that uses one as a headline outcome measure, because that is a claim about sensitivity to change and the sensitivity is not there.