What Is This?
Epigenetic clocks are DNA-methylation models that turn chemical marks on DNA into an ageing-related score.
The mistake is treating every clock as the same instrument. They aren't. A clock trained to predict chronological age, a clock trained to predict mortality risk, and a clock trained to estimate the pace of ageing are measuring different targets.
That matters because a 2026 Nature Human Behaviour systematic review and meta-analysis found that social disadvantage shows up more strongly in newer epigenetic clocks than in first-generation age-prediction clocks. The review covered 140 studies, 65,919 participants, and 1,065 effect sizes linking socioeconomic status, race, and ethnicity with three generations of epigenetic clocks.
The useful model is simple:
clock output = biology + training target + population + sample type + measurement noise
A biological-age number is not a direct readout of destiny. It is a model output.
Why Does It Matter?
Biological-age testing is easy to overread.
A report says someone is older than expected or ageing faster than average, and the number feels objective. But the number only makes sense if you know what the clock was trained to predict.
The Willems review makes that problem sharper. It found that associations between socioeconomic status and biological ageing differed by clock generation:
- first-generation clocks showed the weakest SES association;
- second-generation clocks showed stronger associations;
- third-generation pace-of-ageing clocks showed the strongest associations.
That doesn't mean newer clocks are contaminated by social factors. It means newer clocks may be picking up the biological embedding of stress, environment, morbidity risk, and life-course exposure more directly than age-prediction clocks.
So the question changes from:
What is my biological age?
to:
Which clock is this, what was it trained to predict, and what kind of exposure is it sensitive to?
The Three Clock Generations
First-generation clocks: age-pattern instruments
First-generation clocks were trained mainly to predict chronological age from DNA methylation.
Steve Horvath's 2013 pan-tissue clock is the canonical example. It used thousands of samples across many tissues and selected 353 CpG sites to estimate DNA methylation age. The point was broad age prediction across tissues and cell types.
That is powerful. It also means the target is calendar age.
If a first-generation clock says a sample looks older or younger than expected, that can be interesting, but the model was not directly trained on mortality, disease, function, or pace of decline.
Second-generation clocks: risk instruments
Second-generation clocks were trained closer to health outcomes.
DNAm PhenoAge, introduced by Levine and colleagues in 2018, was trained through a phenotypic-age model built from clinical biomarkers associated with morbidity and mortality. DNAm GrimAge, introduced by Lu and colleagues in 2019, was designed as a mortality-risk estimator using DNA-methylation surrogates for smoking pack-years and plasma proteins.
These clocks still output age-like numbers, but the meaning is different. They are closer to risk estimators than literal age meters.
A GrimAge-like result is not saying "your body is literally this old." It is saying the methylation pattern resembles a profile associated with mortality and healthspan risk in the training data.
Third-generation clocks: pace instruments
Third-generation clocks try to estimate the rate of ageing.
DunedinPACE is the clean example. Belsky and colleagues trained it from longitudinal biological change in the Dunedin Study: repeated measures of organ-system integrity across two decades, then distilled into a DNA-methylation blood biomarker. It is scaled around a pace: roughly biological years per chronological year.
That makes it closer to a speedometer than a clock.
This is why the Willems review's finding matters. If pace clocks are most sensitive to SES and racialized exposure, the interpretation is not just "low status makes people older." The better interpretation is that pace-of-ageing models may capture how life-course stress and environment become biologically embedded.
What The 2026 Meta-Analysis Adds
Willems, Rezaki, Aikins, Bahl, Wu, Belsky, and Raffington reviewed studies linking social determinants of health with epigenetic clocks.
The review was preregistered on OSF, included studies from 2013 onward, and pooled effect sizes across SES, race, ethnicity, and clock generation.
The headline result was a clock-generation gradient. The abstract reports SES associations of:
- first-generation clocks: r = -0.03;
- second-generation clocks: r = -0.11;
- third-generation clocks: r = -0.13.
The authors also report that sex, tissue, and array type minimally modified the results and that publication bias was negligible.
The practical point: the clock's design controls what social exposure can show up as.
A clock trained on chronological age may miss or dampen biology that matters for disease risk. A clock trained on morbidity, mortality, or pace of ageing may be more sensitive to the pathways through which social conditions affect health.
Why Smart People Get This Wrong
They collapse all clocks into one number
"Biological age" sounds like one thing. It isn't.
Horvath-style DNAm age, PhenoAge, GrimAge, and DunedinPACE are not interchangeable. They have different training targets, different outputs, and different use cases.
They mistake population validity for personal precision
A clock can be useful across cohorts while still being noisy for one person making one decision from one test.
This is especially important for consumer tests. Lab method, tissue type, sample handling, batch effects, reference population, and algorithm choice can all affect the result.
They read risk as causation
If a clock is associated with socioeconomic exposure, mortality risk, or disease burden, that does not prove the methylation pattern caused the outcome.
It may be a marker, mediator, consequence, proxy, or mixture of all four.
They compare results across companies
Two products can both say "biological age" while using different models. Comparing those numbers as if they were the same instrument is like comparing heart rate to lactate and pretending both are one score called "fitness."
How To Use This
When reading a biological-age claim, ask five questions.
1. Which clock?
If the product or paper doesn't name the clock, treat the result as weak.
Useful names include Horvath, Hannum, PhenoAge, GrimAge, GrimAge2, DunedinPoAm, and DunedinPACE.
2. What was it trained to predict?
This is the key question.
chronological age -> age-pattern signal
clinical/mortality risk -> risk signal
longitudinal decline -> pace signal
The target decides the meaning.
3. What sample type was used?
Many clocks are blood-based. Some are designed across tissues. A saliva, blood, buccal, or tissue sample may not mean the same thing.
4. What can change the number?
A good result should come with uncertainty. Ask whether repeated measures, lifestyle changes, illness, medication, smoking, cell composition, or lab method can shift the score.
5. Does it change an action?
A biomarker is more useful when it changes what you do.
If the result just says "optimize sleep, training, nutrition, stress, smoking, alcohol, and metabolic health," it may be educational but not decisive. Those inputs were already the base layer.
Practical Takeaways For Jamie
Use epigenetic clocks as model-building tools, not authority figures.
- Treat Horvath-like clocks as age-pattern signals.
- Treat PhenoAge and GrimAge-like clocks as risk signals.
- Treat DunedinPACE-like clocks as pace signals.
- Don't compare consumer test numbers unless they use the same clock, same sample type, and similar lab process.
- Don't let a single biological-age result override training, sleep, blood markers, symptoms, or medical advice.
The real value is not the number. It is learning what kind of biological process the number is trying to summarize.
Key Terms
- DNA methylation: A chemical modification to DNA that can influence gene regulation and changes with development, ageing, environment, and disease.
- CpG site: A DNA location where methylation is often measured.
- Epigenetic clock: A statistical model that uses DNA methylation patterns to estimate an ageing-related target.
- DNA methylation age: An age estimate produced from DNA methylation patterns.
- Age acceleration: The gap between DNAm age and chronological age, or an age-adjusted residual depending on the model.
- PhenoAge: A second-generation clock trained toward a clinical phenotypic-age target related to morbidity and mortality.
- GrimAge: A second-generation mortality-risk clock using methylation surrogates for smoking and plasma proteins.
- DunedinPACE: A third-generation pace-of-ageing biomarker trained from longitudinal physiological decline.
- Social determinants of health: Social and environmental conditions, including socioeconomic status and racialized exposure, that affect health and survival.
Recall Questions
- Why is "biological age" not one measurement?
- What is the difference between a first-generation and a second-generation epigenetic clock?
- Why is DunedinPACE better described as a speedometer than a clock?
- What did the 2026 Willems meta-analysis show about SES and clock generation?
- What five questions should you ask before trusting a consumer biological-age result?
Best Resources to Learn More
- Start with Willems et al. for the social-determinants meta-analysis.
- Read Horvath 2013 to understand the first-generation clock idea.
- Read Levine 2018 and Lu 2019 to see how second-generation clocks move toward health and mortality risk.
- Read Belsky 2022 for the pace-of-ageing model behind DunedinPACE.
Sources
- Willems, Y. E., Rezaki, A. D., Aikins, M., Bahl, A., Wu, Q., Belsky, D. W., & Raffington, L. "Social determinants of health and epigenetic clocks: a systematic review and meta-analysis of 140 studies." Nature Human Behaviour. 2026. DOI: https://doi.org/10.1038/s41562-026-02477-6
- PubMed record for the earlier preprint and later peer-reviewed update. PMID 40385415; later update PMID 42286117. https://pubmed.ncbi.nlm.nih.gov/40385415/
- Horvath, S. "DNA methylation age of human tissues and cell types." Genome Biology. 2013. DOI: https://doi.org/10.1186/gb-2013-14-10-r115
- Levine, M. E., Lu, A. T., Quach, A., et al. "An epigenetic biomarker of aging for lifespan and healthspan." Aging. 2018. DOI: https://doi.org/10.18632/aging.101414
- Lu, A. T., Quach, A., Wilson, J. G., et al. "DNA methylation GrimAge strongly predicts lifespan and healthspan." Aging. 2019. DOI: https://doi.org/10.18632/aging.101684
- Belsky, D. W., Caspi, A., Corcoran, D. L., et al. "DunedinPACE, a DNA methylation biomarker of the pace of aging." eLife. 2022. DOI: https://doi.org/10.7554/eLife.73420