Do epigenetic clocks measure aging or just correlate with it

Epigenetic clocks are the most talked-about biomarkers in longevity research right now, and for good reason: they predict mortality better than your birth year does. GrimAge outperforms calendar age at forecasting who dies sooner [1], PhenoAge adds further predictive power over first-generation clocks [2], and DunedinPACE captures not biological age but the rate at which someone is aging [3]. These tools have also moved in encouraging directions when people do things we'd expect to help: caloric restriction slowed DunedinPACE in a controlled trial [4], omega-3 supplementation slowed multiple clocks [5], and partial reprogramming reset clock readings in mice [6].

None of that makes them a proven window into how long you'll live.

What a clock measures

An epigenetic clock reads patterns of methylation, chemical tags on DNA that shift predictably with age. The clock predicts outcomes well. What it doesn't do, yet, is prove that the methylation pattern causes those outcomes rather than reflecting them.

Steve Horvath, who built the original Horvath clock, frames this as the difference between "the clock face and the clockwork" [7]. The hands tell you something real about time, but moving the hands doesn't fix the mechanism. Morgan Levine is even more direct: "I would argue that we haven't proven that a change in the epigenetic clock is a change in your biological aging process" [8]. These are the inventors and leading advocates, not critics. The concession matters.

No trial has yet shown that lowering a clock score extends lifespan or reduces disease. Until one does, clock changes generated by an intervention are hypothesis-generating data, not evidence that someone will live longer.

The noise problem is larger than most people realize

First-generation clocks carry up to 8-9 years of measurement noise on replicate samples from the same blood draw [8]; formal reliability analyses confirm similar magnitudes [9]. A single consumer "biological age" reading sits inside that noise band. It is close to uninterpretable at the individual level.

The Fitzgerald study makes this concrete. A dietary and lifestyle intervention produced a roughly three-year reversal on the Horvath clock, generating considerable press coverage. When Levine and colleagues reanalyzed the data using PC-clock reliability correction, the entire effect disappeared, it was noise [8]. The original signal was an artifact of applying a first-generation clock to an underpowered study without correcting for measurement error.

Only PC-corrected clocks and DunedinPACE are considered trial-grade for intervention research. Any headline claiming "X reversed aging by N years" on a first-generation clock deserves immediate skepticism about whether reliability correction was applied.

How heritable are clock readings?

Two prominent researchers gave estimates that initially looked contradictory. Horvath cited roughly 40% heritability; Levine cited 10-20% [7, 8]. The disagreement turned out to be methodological, not factual, they were measuring different things.

Twin studies, which capture all genetic variance including rare variants, dominance, and gene-environment interactions, find heritability of 55% in Scandinavian pairs [10], 58-72% across PC-corrected clocks in a large Chinese twin sample [11], and 40-60% in Finnish female twins followed for 18 years [12]. Horvath's ~40% is a conservative lower-bound rounding of these estimates.

GWAS-based SNP heritability, which captures only common additive variants and misses much of the genetic architecture, comes in at 17% for the Horvath clock in a meta-analysis of over 40,000 participants [13], consistent with Levine's 10-20% figure. Both numbers are correct for their respective frameworks; the gap between them is the classic "missing heritability" problem.

One additional finding from the Hong 2024 twin data [11]: heritability declines between ages 55-70 for most clocks. DunedinPACE and PC-PhenoAge show the sharpest drops in the 55-65 window. Environmental factors, the ones modifiable by lifestyle intervention, become more dominant precisely in the age range most relevant to clinical longevity research.

A sharper dissent

Matt Kaeberlein takes a harder line than Levine or Horvath. Roughly two dozen clocks now exist with no consensus on which is most valid [14]. No mouse lifespan experiment has yet demonstrated that reversing a methylation clock reading extends life [14]. Consumer biological-age products, in Kaeberlein's assessment, are not clinically validated and a single reading is noise-limited and unactionable [14]. His preferred alternatives are functional markers, VO₂max, grip strength, muscle mass, zone-2 capacity, ApoB, which carry validated mortality and morbidity associations and cost far less to measure [14].

What to do with your clock reading

A one-off biological age number, including the kind generated by high-profile self-tracking programs, sits inside the noise floor of first-generation clocks [15]. Repeated measures over time, using PC-corrected clocks or DunedinPACE, can be informative in a research context. For most people, the functional markers Kaeberlein lists are cheaper, better validated, and immediately actionable.

Clocks remain the best research surrogates the field has. They are generating useful hypotheses about what slows biological aging. What they have not yet done is prove that hypothesis, the decisive study would be a randomized trial showing that lowering a clock score produces a hard outcome benefit like reduced mortality or disease. That trial has not been run.

Where the evidence is contested

The core surrogate question has no resolution yet. Clocks predict lifespan outcomes well [2, 1, 3] and respond to plausible interventions [6, 5, 4]. Horvath and Levine argue this makes them useful research tools even without proven causation [7, 8]. Kaeberlein argues the causal gap is large enough that commercial applications built on clock readings lack sufficient validation to be useful to individuals [14]. No experiment at any level, human or animal, has closed that gap by linking clock reduction to a hard longevity outcome.

On heritability, the apparent conflict between Horvath (~40%) and Levine (10-20%) is resolved: they cited twin heritability and SNP heritability respectively, both of which are accurate for their measurement framework [10, 11, 13].

On clock proliferation, the field has not converged on which clock is best suited for which purpose, and the experts quoted here disagree on how much that matters in practice [14, 7, 8].

References

1. Lu et al. 2019, Aging — DNAm GrimAge · PubMed 30669119
2. Levine et al. 2018, Aging — DNAm PhenoAge · PubMed 29676998
3. Belsky et al. 2022, eLife — DunedinPACE · PubMed 35029144
4. Kraus et al. 2019, Lancet D&E (CALERIE) — 2yr caloric restriction · PubMed 31303390
5. DO-HEALTH 2025, Nature Aging — omega-3 & DNAm aging clocks · PubMed 39730768
6. Lu et al. 2020, Nature — OSK reprogramming restores vision (mouse) · PubMed 33268865
7. raw/foundmyfitness/fmf-horvath-phenoage-grimage-clocks.txt
8. FMF — Rhonda Patrick + Morgan Levine: PhenoAge + age acceleration
9. Higgins-Chen et al. 2022, Nat Aging — clock reliability (PC-clocks) · PubMed 36277076
10. Jylhävä J et al. 2019 (eBioMedicine) — Longitudinal twin heritability of epigenetic clocks; SATSA (Swedish) + LSADT (Danish); n=104 pairs ages… · PubMed 30704927
11. Hong X, Li L et al. 2024/2025 (Aging Cell) — PC-corrected epigenetic clock heritability; Chinese National Twin Registry; n=1,084 (ages 31–70);… · PubMed 39543924
12. Schandelmaier et al. 2017 (Cochrane) — 23 RCTs; null overall; low-risk-bias subgroup: statins > niacin (p=0.05, HPS2-THRIVE dominant) · PubMed 28616955
13. McCartney DL, Marioni RE et al. 2021 (Genome Biology) — GWAS of DNAm aging biomarkers; n=40,905; 137 loci; IEAA SNP heritability h²_SNP=0.17 (17%);… · PubMed 34187551
14. raw/attia/attia207-antiaging-drugs-nad-metformin-rapamycin.txt (paywalled; only intro accessible)
15. Bryan Johnson — test your biological age for $0 (functional proxies)