
A UK study found four epigenetic clocks produced different age gaps and socioeconomic patterns. Learn how clock design shapes interpretation.

On November 16, 2026, Tobias Rähse and colleagues published “Epigenetic clock choice matters: systematic variation in age estimation and socioeconomic patterning of biological aging” in GeroScience. Their analysis compared four epigenetic measures using Understanding Society, the UK Household Longitudinal Study.
The researchers found that different clocks produced substantially different age estimates and socioeconomic patterns within the same cohort. Some measures suggested clearer relationships between socioeconomic advantage and younger age estimates than others. For healthspan research, the central finding concerns interpretation: the algorithm selected can change the apparent relationship between social conditions and biological aging. The paper therefore makes clock design part of the scientific question rather than a technical detail to overlook.
The study was conducted in humans using observational whole-blood DNA-methylation data.
According to the GeroScience paper, the blood data were collected from 2010 to 2012. After excluding outliers, the main analysis included 3,597 people contributing 14,388 clock-specific observations. That larger observation count represents readings from the different clocks rather than a separate population of participants.
The researchers examined both estimated age differences and their relationships with education, occupation, and household income. This design allowed them to ask whether the choice of clock changed the social pattern visible in the results.
The paper compared Horvath 2013 and Horvath Skin & Blood with PhenoAge and DunedinPoAm. The authors classify the Horvath measures as first-generation clocks oriented toward chronological age. PhenoAge is a second-generation measure designed around health-related phenotypes. DunedinPoAm measures the pace of aging rather than an absolute age level.
Those distinctions matter before any numerical comparison begins. An estimate designed around calendar age addresses a different question from a health-oriented measure or a pace measure. Treating their outputs as interchangeable would overlook the purposes described by the authors.
For the main comparisons, the researchers transformed DunedinPoAm onto a clock-equivalent scale. They defined “delta age” as DNA-methylation-derived age minus chronological age, with an analogous calculation for the transformed pace measure. All DunedinPoAm age-gap values discussed below refer to that transformed scale.
For readers evaluating how a biological-age model is constructed, the distinction between the model’s target and its displayed output is a useful starting point. A result expressed in years still needs to be understood through the method that produced it.
Rähse and colleagues reported an average delta age of +4.69 years for Horvath 2013. PhenoAge averaged −7.15 years. Transformed DunedinPoAm averaged +1.10 years.
These values describe differences between clock-derived estimates and chronological age under the study’s comparison method. They do not mean that participants gained or lost those amounts of life. Reading the negative PhenoAge average as years of rejuvenation would mistake a measurement difference for an observed biological change.
The age-group patterns also differed. Horvath 2013 averaged +14.17 years among participants aged 18 to 29 and −9.32 years among those aged 80 and older. PhenoAge remained negative across age groups. DunedinPoAm showed comparatively small fluctuations.
The descriptive result is therefore broader than disagreement about a cohort average. The clocks also differed in how their estimates were distributed across age groups, making the age profile of the sample relevant to interpreting the comparison.
For education, the Horvath 2013 results did not follow a simple pattern of higher qualifications accompanying younger clock estimates. Participants with no or other qualifications averaged +3.48 years, compared with +5.34 years among those with tertiary education.
DunedinPoAm showed the opposite direction in the authors’ comparison. Its transformed estimates ranged from +4.35 years among people with no or other qualifications to −0.66 years among those with tertiary education. The educational categories were the same, but the apparent relationship changed with the measure.
Income comparisons showed a similar dependence on clock choice. Horvath 2013 estimates rose from +4.29 years in the lowest household-income quintile to +5.05 years in the highest. PhenoAge estimates were −6.48 years in the lowest quintile and −7.79 years in the highest.
For DunedinPoAm, estimates fell from +2.85 years in the lowest income quintile to −0.12 years in the highest. The highest-quintile estimate was not statistically significant.
The authors also reported an important distinction in the statistical models. Education, occupation, and income did not show significant overall main effects in the relevant analyses. Clock-by-socioeconomic-status interactions nevertheless revealed clock-specific patterns. In plain language, the relationship depended on which measure was used rather than appearing as a single consistent pattern across the clocks.
The researchers examined whether health-related characteristics accounted for some of the variation. Adjusting for body mass index, self-rated health, and medication use reduced the clock-by-socioeconomic-status interactions without eliminating them.
In a separate subset with wave-two methylation data, adjustment for smoking, alcohol consumption, and physical activity also reduced differences between clocks. Some heterogeneity remained in those analyses.
The authors interpreted the clearer socioeconomic patterns in PhenoAge and DunedinPoAm as potentially reflecting their design. That is an interpretation of the comparison: measures developed around different targets may register social patterning differently. The adjustment results add context rather than reducing the findings to a single explanation based on health behaviors.
A separate 2026 study in Development and Psychopathology reported clock-specific associations in a US childhood cohort. Relationships involving early adversity and poverty varied across epigenetic-aging measures. Some associations pointed in different directions depending on the measure.
This adds another example of why naming the algorithm matters in social and developmental aging research. The shared methodological issue is whether a conclusion belongs to epigenetic aging broadly or to the particular measure used in an analysis.
The authors identify weaknesses that matter when moving from these comparisons to claims about aging and health:
The practical reading strategy is methodological: identify the algorithm, check its intended target, and distinguish an age-level estimate from a pace measure. Then ask whether a reported relationship applies across clocks or only within one model. When following biological-age research, keep those details attached to the headline result rather than treating “biological age” as a self-explanatory label.
The next research priority is longitudinal validation across populations and life stages to determine whether these clock-specific differences correspond to later physiological deterioration or health decline.
Conflicting epigenetic-clock results can become overstated biological-age claims when the algorithm disappears from the story. AgeAmaze explains study design and evidence quality for readers interested in healthy aging, helping them assess biological-age claims without treating a biomarker result as support for a supplement or emerging therapy.
Stay current with research on aging biology, biomarkers, nutrition, therapeutics, peptides and longevity technology. AgeAmaze reports what the evidence shows, where uncertainty remains and which claims still need stronger data.



Follow AgeAmaze for careful reporting on what longevity science can show today and what still needs stronger evidence.
read the Blog