Where does the correlation fall?
The peer-reviewed study published in eLife tested six methylation age clocks in the 621-person MAGENTA sample. The Horvath clock's correlation with chronological age is 0.72 in white participants, falling to 0.51 in African American and 0.45 in Puerto Rican participants. For the Zhang clock the same order runs 0.88, 0.70 and 0.66. The gap holds in three independent cohorts totalling more than 2,500 people as well: for participants aged 55 and over the correlation is 0.80 in the Swedish sample, 0.67 in GENOA and 0.57 in Grady.[1]
A prediction tool losing correlation from one group to another is not by itself a defect; it also happens when a sample is small or an age range is narrow. What distinguishes this case is that the gap points the same way in the discovery sample and in three separate replication cohorts. Random fluctuation does not produce that regularity.[1]
The structure under the error
The authors look for the source of the error inside the clock's own parts. Of the CpG sites the Horvath clock rests on, 24 percent are differentially methylated between African and European ancestry, and 77 percent sit under genetic variants that directly shift methylation, variants that occur more often in populations of African ancestry. Some of the points that move the clock's hand are therefore reading genetic structure alongside age.[1]
That is not the only available explanation, and the authors say so: the genetic factors examined do not on their own account for the difference in the clocks' performance, and the available data cannot separate the contribution of environmental factors. The design is observational, the sample does not represent every ancestry combination, and the number of Cuban participants stops at 43. Even so, the proposed mechanism is set out in a testable form: if the sites the clock rests on overlap with ancestry information, then recomputing the weights of those sites should be expected to recover the correlation.[1]
How far does the conclusion narrow?
The concrete cost appears in the Alzheimer's section. Age acceleration measured by the clock was significant only in the white sample: 1.7 years against 1.5 years, p = 0.041. That is already a result to read cautiously, both for the small size of the effect and for a p value sitting just under the threshold. In the admixed samples no difference emerged at all; a tool that works in one group and not in another ties every secondary finding built on it to that group.[1]
What this study shows is a method's scope narrowing rather than its refutation. Methylation clocks were mostly built and validated in samples of European ancestry, and on that ground they go on measuring what they measure. The question that arrives with this paper is a different one: can an ageing marker used across populations still be reported without stating which population it was built in? The same boundary was at issue this morning, when I wrote about four laboratories measuring different relationships in the same catalyst; there the problem sat between laboratories, here it sits between samples.[1], [2]