- Rebuilt: the comparison of two PhenoAge scores from a 2022 study of biological aging, in a second national sample (NHANES 1999 to 2002, 1,556 adults aged 50 to 85), from public files.
- Held: the disagreement. The blood version puts Black participants 4.3 years further ahead of their age than White participants, and the methylation version finds no clear gap.
- Did not hold: an exact match, as expected from a different survey. The blood gap is wider here: 4.3 years, against about 2.8 in the paper.
A 2022 paper in the American Journal of Epidemiology measured biological age eight ways in the same older adults. Two of the eight share a name: PhenoAge, from Morgan Levine’s lab. One version reads nine routine blood tests. The other reads chemical marks on DNA, called methylation, and was trained to reproduce the first. Buried in the supplement, the blood version puts Black participants about 2.8 years further ahead of their chronological age than White participants, and the methylation version finds essentially no difference. The paper never compares the two.
BloodVsMethylation rebuilds that comparison in a second national sample, entirely from public files, and then takes the disagreement apart.
nix run github:gauranii/BloodVsMethylation
# copies the code into ./BloodVsMethylation, downloads ~85MB of public
# NHANES and CDC files, and rebuilds every table and figure
How it is built
Eleven numbered R scripts, run in order by run_all.R, each in its own environment so no script leans on another’s leftovers.
- Pull NHANES 1999–2002 labs, methylation clocks, NCHS mortality, DXA, cystatin C R/01
- Build 1,556 adults 50+, survey design with methylation weights R/02
- Measure Black-White gap on every clock, design-based effect sizes R/03
- Decompose the blood PhenoAge gap, split exactly across nine markers R/04
- Explain white cells, RDW, creatinine, one marker at a time R/05 · R/10 · R/11
- Stress-test mortality models, level-shift simulation, figures, sensitivity R/06 – R/09
Reproducible. R and all 8 packages are pinned in flake.nix and flake.lock, and each run writes output/session_info.txt. A fresh clone and nix run on 11 October 2026 reproduced every committed table and figure byte for byte. Every estimate uses the NHANES survey design: strata, PSUs, and the methylation subsample’s own weights, through the survey package.
Tested. 17 testthat tests cover the logic this repo adds: the PhenoAge formula, the decomposition, the download checks, the pooling of imputed estimates, and unit invariants on the processed data. GitHub Actions runs nix flake check and the suite on every push.
Three walls, and the way around each
The published formula overflows. PhenoAge is published as two steps: a 10-year mortality score, then the age at which that score is average. For the seven highest-risk participants here, the mortality score rounds to exactly 1 in floating point and the second step returns infinity. Substituting one step into the other cancels both logarithms, and what is left is a straight line in the markers. From R/utils_phenoage.R:
# Substituting the mortality score into the age formula cancels both logs,
# so PhenoAge is exactly linear in xb: one unit of xb is 1/0.090165 = 11.09
# years. This closed form is what the repo uses (it never overflows), and the
# linearity is what makes the per-marker decomposition below exact.
YEARS_PER_XB <- 1 / 0.090165
phenoage_blood <- function(d) {
141.50225 + (log(0.00553) + log((exp(120 * GOMPERTZ_GAMMA) - 1) / GOMPERTZ_GAMMA) +
phenoage_xb(d)) * YEARS_PER_XB
}
The tests hold the closed form to the published one wherever the published one is finite, and check that it stays finite where the published one does not:
test_that("closed form matches the published two-step formula", {
people <- typical[rep(1, 5), ]
people$age <- c(50, 60, 70, 80, 85)
people$rdw <- c(12, 13, 14, 15, 16)
expect_equal(phenoage_blood(people), phenoage_blood_two_step(people), tolerance = 1e-8)
})
test_that("closed form stays finite where the two-step form overflows", {
sick <- typical
sick$creatinine <- 1200; sick$rdw <- 25; sick$glucose <- 25; sick$age <- 85
expect_false(is.finite(phenoage_blood_two_step(sick)))
expect_true(is.finite(phenoage_blood(sick)))
})
A failed download looks like a success. When a CDC file is missing, the server answers with a small HTML error page and a 200 status, which a naive script would cache and try to parse as data. R/utils_download.R trusts a file only if it is large enough and does not begin with <, and deletes it otherwise.
Muscle mass comes in five copies. NHANES publishes body-scan lean mass as five multiply imputed datasets, not one. The creatinine check runs on all five and pools them with Rubin’s rules, in R/utils_mi.R, so the uncertainty from the imputation stays in the answer.
Paper against rebuild
| Graf et al. (HRS 2016) | Rebuild (NHANES 1999–2002) | |
|---|---|---|
| Sample | adults 50–90 | 1,556 adults 50–85, deaths to 2019 |
| Blood PhenoAge, Black-White gap | about 2.8 years (d = 0.39) | 4.3 years (d = 0.45) |
| Methylation PhenoAge gap | essentially none (d = 0.08) | −0.7 years (d = −0.10) |
| Other clocks (GrimAge, DunedinPoAm, Horvath, Hannum) | same direction as the paper, every one |

The disagreement reproduces, and it is wider here.
Taking the gap apart
Because blood PhenoAge is linear in its markers, the 4.3-year gap splits exactly into nine pieces, one per marker, with nothing left over.

Red cell distribution width is the largest piece, at 2.5 years. Inherited hemoglobin traits more common in people of African ancestry, including a form of alpha-thalassemia and sickle cell trait, raise it without having anything to do with aging. Using published effect sizes, they plausibly account for about a third of the difference, and iron status for a little more. The rest I could not explain.
Creatinine adds 1.7 years. I expected muscle mass. It was mostly 40 participants with advanced kidney disease, confirmed by a second kidney marker, cystatin C. Muscle mass accounts for about 0.3 years.
The two white blood cell markers subtract 1.3 years. Duffy-null, a common and harmless variant in people of African ancestry, lowers neutrophil counts for life, and the formula reads that as younger. The pattern in the data matches it closely.
So the 4.3 years is a sum of parts pulling in different directions: inherited traits that are not aging, real disease, and a remainder still unexplained.
A check that could not have caught this
The paper argued its measures work equally well across groups because they predict health outcomes about equally well in Black and White participants. R/07_level_shift_simulation.R copies the White participants into a synthetic second group, identical in every way, and shifts only their blood markers by the observed differences. That makes a 4.3-year gap that is entirely artificial. The check passes it perfectly: identical spreads, identical hazard ratios, an interaction of zero. It can confirm that a measure predicts equally well. It cannot tell whether the measure is shifted.
What this does and does not say
It does not show that Black Americans are not aging faster. The methylation clocks built to predict death still find gaps of one to two years in the same data. Nor does it show that the methylation version is the unbiased one; there is published evidence it carries its own ancestry signal. What it shows is that a biological age score can find a racial difference that another score, built toward the same target, does not, and that before a gap in such a score is read as a gap in aging, it is worth asking which of its parts is doing the work. The README lists what is still open.