The multivitamin that slowed a clock

This one is different from most supplement reports I write, and I want to say that up front so the rest reads correctly.

Usually the problem is the study: nine people per arm, no comparator, an endpoint swapped after the fact. Not here. This is a prespecified ancillary analysis inside COSMOS, a randomized, double-blind, placebo-controlled, factorial trial run out of Brigham and Women's Hospital with 21,442 participants in the parent study. The methodology is good. The reporting is careful. The authors hedge their own result in the abstract.

The problem is entirely in the translation layer between the paper and the headline. So this report is not about a bad study. It is about what happens to a good one on the way to a product page.

What the trial did

COSMOS gave older adults a daily multivitamin-multimineral (Centrum Silver), a cocoa extract (500 mg cocoa flavanols including 80 mg (-)-epicatechin), both, or neither, in a two-by-two factorial design. The parent trial ran from June 2015 to June 2023.

This particular paper is a substudy. It took blood from 958 participants, 482 women and 476 men, and measured DNA methylation at baseline and after two years. It then computed five different "epigenetic clocks" from that methylation data and asked whether the supplements changed how fast those clocks ticked.

The five were PCHannum, PCHorvath, PCPhenoAge, PCGrimAge and DunedinPACE. Those names matter, and I will come back to why.

What an epigenetic clock actually is

This is the part that almost never survives into coverage, and it is the part that determines what the result means.

An epigenetic clock is not a measurement of aging. It is a regression model. Someone took a large cohort, measured methylation at a few hundred or a few thousand sites across the genome, and fit a set of weights so that the weighted sum best predicts some target. The clock is those weights.

What the target was is the whole story, and the five clocks in this study do not share one:

PCHannum and PCHorvath are first-generation clocks. They were trained to predict chronological age. Their target was the number on your driver's license. A perfect first-generation clock tells you nothing about health, because a perfect one just recovers a number you already knew.

PCPhenoAge and PCGrimAge are second-generation. They were trained against mortality and clinical-biomarker composites, so they are built to predict how long you have left rather than how long you have been here. DunedinPACE is different again: it estimates a current rate of change rather than an accumulated age.

So "biological age" is not one quantity that five instruments measure with varying precision. It is five different constructs with five different training targets. That is why the paper's own framing matters: the two clocks that moved were the two second-generation ones. That is the coherent subset, and the authors say so. It is a genuinely better result than two random clocks moving.

But here is the thing to hold onto. Each of these is a predictor trained on an outcome. Moving the predictor is not the same as moving the outcome. If I train a model to predict house fires from smoke-detector readings, and then I find an intervention that lowers the readings, I have not established that it prevents fires. I have established that it lowers the readings. Whether the fires follow is a separate empirical question, and it is the one everybody actually cares about.

The authors know this. It is why their concluding sentence reads the way it does.

The numbers, in their real units

Here is what the abstract reports for the multivitamin versus placebo, as a between-group difference in yearly change:

PCGrimAge: −0.113 years (95% CI −0.205 to −0.020), P = 0.017.
PCPhenoAge: −0.214 years (95% CI −0.410 to −0.019), P = 0.032.

Read the unit carefully. It is years per year. It is a rate, not a quantity. The multivitamin group's clocks went up more slowly, by 0.113 and 0.214 of a year for every year that passed.

Convert to months and the arithmetic behind the headlines appears. 0.113 years is about 1.4 months per year. 0.214 years is about 2.6 months per year. The trial ran two years. Multiply, and you get roughly 2.7 months and 5.1 months of accumulated difference, which is where "a multivitamin makes you four months younger" comes from. Four months is a rounded midpoint of two numbers, each of which is a rate multiplied by the trial's duration.

That is not a trick. It is a legitimate way to express a rate difference. But it changes how you should think about it. The result is not "you are four months younger." It is "one of these predictors rose about two and a half months per year more slowly, and we watched for two years."

Now do the confidence intervals, which nobody does and everybody should. Take the PCPhenoAge interval to its ends and multiply by two years. The lower bound, −0.410 per year, becomes about ten months over the trial. The upper bound, −0.019 per year, becomes about two weeks.

Two weeks. Over two years. That is what the data are also consistent with, at conventional confidence. The point estimate is the middle of a range whose optimistic end is a school year and whose pessimistic end is a fortnight. PCGrimAge is similar: roughly five months at one end, about two weeks at the other.

When you see a single number reported without its interval, you are being shown the most flattering point in a range the study could not narrow further.

Five clocks, two results

Three of the five clocks showed no significant effect. The cocoa extract showed no effect on any of the five.

The lead author's framing, quoted in trade coverage, is this:

The statistical significance for two of five clocks — coupled with consistent improvements across all five epigenetic clocks — is very encouraging.

That is Howard D. Sesso, and I think it is a defensible thing to say. Directional consistency across correlated measures is real evidence, and these clocks are correlated because they are computed from overlapping methylation sites in the same blood samples. They are not five independent experiments.

But there is an arithmetic point worth making plainly. Five clocks were tested. The two significant p-values are 0.017 and 0.032. A conventional Bonferroni correction for five comparisons puts the threshold at 0.01. Neither result clears it.

I want to scope that honestly, because it is the kind of statement that gets overread. Bonferroni is conservative, and it is arguably the wrong correction for five correlated measures of a related construct. I read the abstract, not the full text, so I do not know what multiplicity handling the paper itself describes. What I can say is narrow and checkable: the p-values as reported in the abstract are on the near side of 0.05, not the far side, and there were five tests. That is a fact about the numbers, not an accusation about the analysis.

The subgroup, and why it cuts both ways

The abstract reports a stronger PCGrimAge effect among participants who already had accelerated biological aging at baseline (−0.236, 95% CI −0.380 to −0.091) than among those with normal or decelerated aging (−0.013, 95% CI −0.130 to 0.104), with P = 0.018 for the interaction.

Two honest readings coexist here.

The generous one: this is mechanistically plausible. If a multivitamin corrects micronutrient shortfalls, you would expect the benefit to land on people who were deficient, and to be near zero in people who were replete. The subgroup effect has the shape the underlying theory predicts. That is not nothing.

The skeptical one: subgroup analyses are where noise goes to look like biology, and note that the "normal or decelerated" estimate of −0.013 has a confidence interval crossing zero comfortably. Read the second number rather than the first and the honest summary of this trial for a well-nourished person is: no detectable effect.

Both readings are available from the same line of the abstract. Which one you lead with is an editorial choice, and a supplement label will not make it the way I just did.

Who paid for what

This deserves stating because it is unusually tangled, and because the authors disclosed it properly.

The parent trial's registry lists Brigham and Women's Hospital as lead sponsor, with Fred Hutchinson Cancer Center, Mars, Inc. and Pfizer as collaborators. The publisher-deposited competing-interests statement records that two authors received investigator-initiated grants from Mars Edge, that Pfizer Consumer Healthcare (now Haleon) donated the study pills and packaging, and that FOXO Technologies made in-kind donations for the DNA methylation data generation. Other authors declared no competing interests.

Take the second and third of those together. The company that makes the tested product supplied the product, and a company in the epigenetic-testing business supplied the measurement. Neither of those facts makes the result wrong. In-kind supply is completely normal in nutrition research, disclosure is exactly what is supposed to happen, and randomization protects against a great deal.

But it does mean the commercial interest is present on both sides of the comparison: in what was tested, and in what it was tested with. If epigenetic clocks become an accepted endpoint for supplement claims, the companies selling clock measurements have a market. That is worth noticing while the endpoint is still being established rather than after.

What I am not claiming

Not that the trial is flawed. It is a prespecified ancillary analysis inside a large, properly randomized, double-blind trial. That is a strong design and it is much better than the evidence base most supplement claims rest on. I would rather have this study than not have it.

Not that multivitamins do nothing. COSMOS has produced separate findings on cognition and other outcomes that are their own conversation. This report is about one substudy and one class of endpoint.

Not that epigenetic clocks are junk. Second-generation clocks predict mortality in cohort data, which is a real and useful property. My argument is about what a change in a predictor establishes, not about whether the predictor works.

Not that I read the full paper. I read the complete published abstract verbatim through the PubMed record, the ClinicalTrials.gov registration for the parent trial through the API, and the publisher-deposited metadata including the competing-interests statement. The full text is paywalled and I did not read it. The paper may well address multiplicity, the clock-construct distinction, and more, inside. Every claim above is scoped to the documents I actually opened, and I have tried to flag that wherever it matters.

The signal

The authors' own final sentence is the most useful thing in the abstract, and it is the sentence least likely to appear on a bottle:

Although the statistically significant but small effects of daily MVM supplementation on slowing biological aging are encouraging, additional studies are needed to determine the clinical relevance of daily MVM supplementation on epigenetic clocks and whether such effects can help explain the beneficial effects of MVM supplementation on aging-related chronic conditions.

"Additional studies are needed to determine the clinical relevance" is not boilerplate here. It is the precise statement of the gap: they moved a predictor, and whether that matters clinically is not yet established. The researchers said the true thing. It just does not fit on a label.

Three habits to carry out of this, all portable to the next aging-supplement headline:

Ask what the endpoint was trained to predict. Any time a study reports a "biological age," a "score," or an "index," someone fit that quantity to a target. Ask what the target was. A clock trained on chronological age and a clock trained on mortality are different instruments wearing the same word. This is the same reflex as asking what "clinically proven" was proven against.

Check whether the number is a rate or a quantity. "Four months younger" and "0.2 years per year" describe the same finding, and one of them sounds like a result you can feel. Multiplying a rate by a duration is legitimate. Presenting the product without the rate is where the exaggeration enters.

Find the confidence interval and read its unflattering end. If the honest range runs from two weeks to ten months, then "four months" is a choice, not a measurement. This single habit will deflate more supplement marketing than any amount of chemistry knowledge.

My own read, for what it is worth: a cheap multivitamin is a reasonable thing for an older adult with a marginal diet to take, and this trial is mild supporting evidence for that entirely unglamorous position. What it is not is evidence that a pill slows aging. It is evidence that a pill nudged two statistical predictors of aging, by an amount whose confidence interval reaches down to a fortnight, in people over 60, over two years. That is a real finding. It is just much smaller and much stranger than the sentence it turns into.

Not medical advice. This is educational analysis, not a recommendation — a study is not a prescription. Talk to a qualified clinician before acting on anything you read here. Full disclaimer →

Disclosure, plainly: I founded and run Shroombiosis (a company I run), which formulates and sells functional-mushroom supplements. That is a direct commercial stake in the industry this report examines, and the specific temptation described here, reporting a moved surrogate as a delivered outcome, is one my own category is fully capable of. Apply the checks above to anything I sell: ask what the endpoint was trained to predict, check whether the number is a rate or a quantity, and read the unflattering end of the interval. Nothing here is sponsored and no link earns a commission; here's the full policy. A recommendation with no stake at all: for performance nutrition, Die Tryin Co. is a fellow combat-veteran-owned brand I'm glad to point people to. I don't own it and earn nothing from the link.