An independent NHANES reanalysis of people identified by one epigenetic clock but not the other.
The Null Institute | Research Note | 17 November 2026
Our finding: In the same blood-sample dataset, two epigenetic algorithms selected substantially different people as unusually high relative to their peers. 343 participants were flagged by the Horvath clock but not GrimAge2, and another 343 by GrimAge2 but not Horvath. In the following decade, 92 of 343 (26.8%) in the first group died, versus 127 of 343 (37.0%) in the second.
That is an unweighted difference of 10.20 percentage points. Using the NHANES DNA-methylation sampling weights, the corresponding observed difference fell to 6.27 percentage points. This is a description of historical outcomes in two differently classified groups, not a measure of how fast an individual is aging.
One dataset, two different sets of people
We linked publicly released DNA-methylation biomarker estimates from the US National Health and Nutrition Examination Survey (NHANES), based on blood collected in 1999–2002, to demographic data and public mortality follow-up through 2019. The primary analysis includes 2,402 adults aged 50–84 with the required clock estimates and mortality information. 593 died within ten years of their baseline examination.
The two clocks were not developed for identical purposes. Horvath primarily estimates chronological age from methylation patterns; GrimAge2 was trained using mortality-related signals. Comparing their raw age figures as though they measured one agreed-upon biological quantity would be misleading.
Instead, we removed each clock’s average association with calendar age and sex, then selected the highest fifth of the remaining values for each clock within the same sample. This fifth is a research-defined ranking, not a clinical threshold for accelerated aging. We asked who was flagged by only one clock, both, or neither.
Four groups, ten years later
| High age-adjusted result | Participants | Deaths within 10 years | Observed mortality |
|---|---|---|---|
| Neither clock | 1,578 | 313 | 19.8% |
| Horvath only | 343 | 92 | 26.8% |
| GrimAge2 only | 343 | 127 | 37.0% |
| Both clocks | 138 | 61 | 44.2% |
| Total | 2,402 | 593 | 24.7% |
Independent calculation from linked CDC NHANES data. The percentages are unweighted, observed group proportions, not individualized predictions.
Only 138 of the 481 people flagged by Horvath were also flagged by GrimAge2. Each clock selected 343 different people who were not in the other clock’s high-residual group. The two discordant groups, 686 people in total, form the central comparison.
For those two groups, the unweighted difference is (127/343 − 92/343) × 100 = 10.20 percentage points. Applying the NHANES methylation-specific sample weights reduces the unadjusted observed difference to 6.27 points. A model adjusting for age, sex and survey cycle instead estimates a 15.0-point gap; these procedures answer related but different questions.
In the subset of 2,313 people with complete smoking-status and body-mass-index data, an additional adjustment yielded an estimated difference of 12.2 percentage points; a sample-weighted pseudo-likelihood sensitivity check gave 6.9 points. The change with weighting is consequential. These are observational comparisons, not causal effects; the weighted model did not implement the complete NHANES survey variance design.
Did either clock add useful predictive information?
We also tested whether the clock estimates helped distinguish who died within ten years, beyond a small clinical predictor set. In five-fold out-of-fold prediction on the same 2,313 complete cases, a logistic model with age, age squared, sex, survey cycle, smoking and BMI reached AUC 0.757. Adding Horvath reached 0.759; adding GrimAge2 reached 0.777; adding both did not improve on GrimAge2 alone.
This modest +0.020 AUC for GrimAge2 over that baseline is not proof of clinical utility. The comparison used one historical dataset with cross-validation, not an independent external validation cohort, and the baseline omitted other potentially important clinical predictors. AUC measures ranking performance, not the probability that any person’s biological-age score is “right.”
GrimAge2 performing better for mortality is not, in itself, a new discovery. Mortality-related signals informed its development, and prior research has documented the prognostic advantage of mortality-oriented clocks. What we independently add here is the person-matched four-group disagreement table, its subsequent death counts, and the sensitivity of their contrast to weighting and common covariates.
What a biological-age result cannot establish
There is no single verified “true biological age” against which these two models can simply be marked right or wrong. A birthday-trained model and a mortality-trained model can disagree without either being defective. Better mortality prediction does not necessarily imply a better measurement of an individual’s pace of aging.
This dataset has one baseline methylation sample per person. It cannot measure within-person changes in aging speed, show whether an intervention slows aging, or establish whether acting on a test result improves health.
The high-residual quintile is an analytical threshold selected for this reanalysis, not a validated clinical cutoff. Public-use CDC linked mortality files substitute synthetic follow-up times for some records for confidentiality, which could affect a ten-year boundary. Checks at 15 years and with death status through 2019 retained the direction of the observed difference. The sample was restricted to ages 50–84 because public-use age data are top-coded at older ages.
Practical implication: Before treating a biological-age result as meaningful, ask which outcome the model was designed to predict, whether it adds information beyond available clinical measures, and whether any health benefit has been demonstrated from decisions based on that result. This reanalysis addresses the first two questions partially, not the third.
Reproduce the independent result
The calculation behind the central finding is visible in the table above. In Python, the unweighted discordant-group comparison is:
horvath_only = (92, 343)
grimage2_only = (127, 343)
a = horvath_only[0] / horvath_only[1]
b = grimage2_only[0] / grimage2_only[1]
print(round(100*a, 2), round(100*b, 2),
round(100*(b-a), 2))
# 26.82 37.03 10.20
The full analysis merges CDC methylation, demographics, smoking/BMI and mortality files on respondent ID SEQN. It fits age- and sex-adjusted clock residuals on the matched sample, defines the highest 20% of each, and reports each mutually exclusive group. Sampling-weight sensitivity uses WTDN4YR; the secondary AUC comparison uses the same five held-out folds across models. Source-file SHA-256 values, the full Python script, aggregate outputs and a verification manifest are available in the public reproducibility package (ZIP; 64,214 bytes). SHA-256: 35c283cc2400657b7efa4d3cd3bc347283bd6f1fc96a4fe50f17b5c2a90a4665. The package contains no participant-level NHANES records.
Limitations: one historical cohort; different model training targets; single-time-point methylation; some synthetic public follow-up durations; no full NHANES design-based confidence intervals; no external validation of our predictive comparison; and no evidence here that ordering or acting on these tests improves patient outcomes. The model contrasts are descriptive and cannot identify causal effects or an actual biological aging rate.
Data and prior work
- CDC NHANES: DNA methylation and epigenetic biomarkers, publicly released data and documentation.
- CDC NCHS: public-use linked mortality files (follow-up through 2019).
- Zou et al. (2025), Clinical Epigenetics, epigenetic clocks and mortality in NHANES.
- Moqri et al. (2025), Nature Aging, benchmarking aging biomarkers.
Transparency: This is an independent secondary analysis of public data, assisted by AI tools for coding and drafting. It has not undergone external peer review. The Institute checked source files, results and the reproducibility package; changes to source documentation are distinguished from original computation. No clinical recommendations or DOI are implied.