Start with the claim, because it should have stopped everyone cold: people with type 2 diabetes, taking metformin, outlived people who did not have diabetes at all.
Not "matched their survival." Outlived them. In the 2014 study by Bannister and colleagues that launched the entire metformin-longevity movement, diabetic patients on metformin monotherapy reportedly showed roughly 15% longer median survival than age- and sex-matched controls without diabetes, after adjustment for baseline cardiovascular prophylaxis, as described by Keys et al.
Sit with the arithmetic that requires. Diabetes shortens life expectancy. Metformin does not cure diabetes. So for a diabetic on metformin to outlast a non-diabetic, the drug would have to first erase the mortality penalty of the disease, then keep going and buy extra years on top.
That is not a subtle geroprotective signal. That is a claim that metformin is doing something so profound it overpowers a serious chronic illness and then some — the kind of finding that should be treated as a red flag demanding scrutiny, not as a cornerstone.
Instead it became the cornerstone. It seeded TED talks, fueled Nir Barzilai's decade-long push for the TAME trial, and made metformin a default entry in the self-experimenting longevity crowd's medicine cabinet. The provocation was the appeal.
So how did a biologically implausible result get produced? Not through fraud. Through a quiet, well-understood statistical trap called informative censoring.
Diabetes is progressive. Most patients who start on metformin monotherapy eventually need additional drugs to hold their blood sugar in check. In Bannister's design, the moment a patient added a second medication, they left the metformin-monotherapy group.
Read that again with mortality in mind. The patients whose disease advanced — the ones progressing toward complications and death — were systematically filtered out of the metformin group before they could die in it.
What remained was a group defined by a variable directly tied to the outcome being measured: staying on monotherapy meant your diabetes wasn't getting worse. The follow-up reportedly averaged only about 2.8 years, per Attia's team, plenty of time to shed the sickest patients and keep the survivors.
Attia's own team offers the cleanest analogy. Imagine a study of smokers versus non-smokers where anyone in the smoking group who dies of lung cancer or lung disease is disqualified from the analysis. Smoking would suddenly look protective — because you deleted the deaths that mattered.
That is the machinery behind diabetics outliving healthy people. The longevity disadvantage of diabetes wasn't overcome. It was censored out.
The longevity disadvantage of diabetes wasn't overcome. It was censored out.
Then someone checked. A Danish group led by Keys, using national health registries and the same broad design, set out to replicate Bannister — and built in the crucial test the original lacked, running the analysis both with and without the informative censoring.
The replication didn't merely fail to find a survival advantage. It flipped the sign.
In the matched cohort replication, diabetic metformin-monotherapy users had significantly higher mortality than non-diabetic controls: hazard ratio 1.39 (95% CI 1.22–1.57). When the censoring based on treatment changes was removed — patients kept in the analysis even as their diabetes progressed — the effect strengthened to 1.48 (95% CI 1.32–1.64), exactly the direction you'd predict if censoring had been manufacturing the original illusion.
The most damning number came from twins. Comparing same-sex twin pairs discordant for metformin use — the tightest possible matching, holding genetics and shared upbringing constant — the metformin-using twin reportedly had a hazard ratio for death of 1.80 (95% CI 1.11–2.91), per Attia's summary of Keys.
Keys and colleagues stated it plainly: metformin monotherapy in type 2 diabetes was not associated with survival equal or superior to the general population without diabetes. The founding evidence didn't survive contact with a better method.
The obvious rebuttal is that these are all observational studies, and observational studies fight over confounders forever. Fair. So look at what happened when researchers ran actual randomized trials in people without diabetes — the only design that removes the diabetes confound entirely.
Three of them reportedly landed null, per the reporting.
Start with frailty, the use-case that should have been metformin's best shot given its proposed effects on mTOR, mitochondrial complex 1, and senescence. MET-PREVENT randomized frail, pre-frail, and sarcopenic adults with a mean age of 80.4 to metformin or placebo for four months, with change in 4-meter walk speed as the primary endpoint.
The metformin group reportedly started at 0.59 m/s and finished at 0.57. The placebo group went from 0.60 to 0.58. The adjusted treatment effect was 0.001 m/s (95% CI −0.06 to 0.06; P=0.96) — as close to literal zero as a clinical trial produces.
Then cancer. MAST, the Metformin Active Surveillance Trial, was a Phase III double-blind placebo-controlled study across 12 Canadian cancer centers, enrolling 408 nondiabetic men with low-risk prostate cancer on active surveillance.
After a median of three years, 144 cancers progressed: 70 on metformin, 74 on placebo. Hazard ratio 1.09 (95% CI 0.79–1.52). No effect.
And before MAST, a larger trial had already tested metformin as an add-on to androgen deprivation therapy in 1,874 nondiabetic men with metastatic hormone-sensitive prostate cancer. It failed to improve overall survival.
Frailty, cancer prevention, cancer treatment — three independent randomized shots in non-diabetics, three nulls. This is what metformin looks like once you can no longer censor the sickest patients out of the picture.
Which reframes what the observational literature was ever measuring. Attia's team makes the point about the frailty data specifically: diabetics on metformin might simply have better-controlled disease, less joint stiffness and neuropathy, and therefore stay more active and preserve more muscle. The drug wasn't the cause; controlled disease was the shared driver.
Now the harder part, because a new generation of tools claims to vindicate metformin without touching a single patient.
Barabási's lab at Northeastern and Harvard published a network-medicine pipeline called SHARP, measuring each of 6,442 DrugBank compounds' network proximity to the hallmark-of-aging gene modules on the human interactome. Metformin was reportedly among the human longevity-trial compounds assessed, with 11 of 17 such compounds showing significant proximity, per the study.
That sounds like independent confirmation. It isn't. It's a mirror.
A proximity score is a structural claim: this drug's protein targets sit near aging-linked genes in the network. It is not an outcome. The pipeline's own authors concede proximity can't even tell direction — some proximal drugs induced senescence rather than suppressed it, forcing a second metric to sort help from harm.
More to the point, SHARP earns its credibility by reportedly validating against human and mammalian outcome data — for instance, checking its hits against compounds that extended lifespan in the Intervention Testing Program, per the study. But for metformin as a human geroprotector, that anchoring outcome data is precisely what does not exist.
So when the pipeline is taken to "validate" metformin, it would be validating a network position against a longevity premise that the human trials have already emptied. A drug can look proximal in silico and do nothing in a body — MET-PREVENT and MAST are what "nothing in a body" reads like on a page. This is the methodological complement to the hype-versus-evidence problem that dogs rapamycin, arriving from the opposite direction: not overhyped trial failure, but computational confirmation of a claim that was never established in humans.
Which brings us back to the person taking metformin tonight, without diabetes, in the belief that they're borrowing years off a proven longevity drug.
The strongest human evidence they're relying on is a single 2014 observational study whose central finding was an artifact of who got removed from the analysis. Its direct replication reversed the sign, to HR 1.39 in the cohort and 1.80 in twins. And every clean randomized test in non-diabetics has come up empty.
Attia's stance — profoundly unimpressed, and willing to move his own prescribing accordingly — isn't contrarianism. It's the appropriate response to noticing that the founding evidence was never strong.
It was censored.