How likely is 'we reversed aging in mice'? The calibrated prior
Every few weeks a headline announces that scientists reversed aging or extended lifespan. Before believing one, anchor on the base rate. In the most rigorous public dataset, about 1 in 4 compounds extend mouse lifespan - but the calibrated prior that any single mouse headline becomes a proven human benefit is only low single digits: the ~90% drug-failure base rate, times the low odds any given compound is ever fairly tested in humans.
Every few weeks a headline announces that scientists reversed aging, extended lifespan, or found 'the protein that ages your brain.' Before believing any single one, it helps to know the base rate. Here is the arithmetic from the most rigorous public dataset in the field.
The NIA Interventions Testing Program (ITP) tests candidate longevity compounds in genetically diverse mice, in parallel at three independent labs, specifically to weed out single-lab and single-strain flukes. Two numbers matter:
- Of the ~40 compounds it has tested, only about 8 to 12 significantly extended lifespan in at least one sex - roughly 1 in 4. And most of those wins are small and sex-specific: acarbose added ~22% of male median lifespan but only ~5% of female (Harrison et al., Aging Cell 2014); and where a male signal looked large (NDGA), much of it traced to unusually short-lived male controls at two of the three sites.
- Of those mouse winners, the number with a proven human lifespan or healthspan benefit on a hard clinical endpoint is zero - but that zero is mostly empty space, not failure. Only one (rapamycin, the PEARL trial) has had a dedicated human longevity trial at all, and it was powered for safety, not lifespan.
| Funnel stage | Count | Rate |
|---|---|---|
| Compounds tested (ITP) | ~40 | - |
| Extended mouse lifespan (≥1 sex) | ~8-12 | ~1 in 4 (~20-25%) |
| Fairly tested at a hard human endpoint | ~1 | rapamycin / PEARL (safety-powered) |
| Proven human lifespan/healthspan benefit | 0 | not yet measurable |
So what is the calibrated prior? Not a measured 0%. A rate needs a denominator of completed human trials, and here that denominator is essentially one - so '0 of ~8' is censoring (the trials have not been run, and a decades-long human-lifespan trial mostly cannot be), not a measured failure rate. To get an honest number you have to combine what we do know, and it has two factors. First, the in-trial base rate: about 90% of drugs that reach human trials fail to win approval, so a compound that gets a fair human trial has only ~10% odds of clearing a hard endpoint - and that all-disease ~10% is, if anything, generous for longevity, where the endpoint takes decades and has no validated surrogate. Second, the access rate: almost none of these mouse winners is ever fairly tested - by the table above, about 1 in 8. Multiply them - ~10% conditional success times a small chance of ever being fairly tested - and the unconditional prior that a given mouse headline becomes a proven human benefit lands in the low single digits, with a wide interval. The sparse human data barely moves it; the number is driven by the base rate and by how rarely these compounds are ever fairly tested. That is the lesson: low, but not zero, and your uncertainty comes from the structure, not from the headline.
This is not a claim that aging biology is fake - tissue-level mechanisms are real - nor pessimism about the field. It is calibration: treat any single mouse headline as a low-single-digit bet on a proven human benefit within a decade, and expect the eventual human effect to be far smaller than the mouse result. And note that the headlines that most need this discipline - 'we reversed aging,' partial reprogramming, young-blood parabiosis - are usually not even ITP compounds; they are a newer, less-validated reference class, so if anything the prior there is lower, not higher.
Why so low? Three measurable reasons.
- The verdict can depend on the statistic you choose - in one specific regime. When a benefit is age-localized (it rescues early-dying animals and fades with age, so the hazards are non-proportional), the standard log-rank test and the Gehan-Wilcoxon test disagree on whether the compound 'worked' on about 40% of identical simulated datasets, and reporting the best of three tests inflates the false-positive rate from 5% to about 8%. This is a textbook property of weighted survival tests (Gehan 1965; Tarone-Ware 1977), not our discovery - and Jiang et al. (GeroScience 2024) showed it on the real ITP data, where the Gehan test flags compounds the log-rank missed. The caveat: that discordance is specific to the non-proportional regime, and 'best of three' is a worst-case analyst behavior, not the ITP's pre-specified protocol. Pre-registration is exactly what protects the ITP's own winners - but the single-lab results that make 'reverse aging' headlines rarely pre-register their test, so this is one reason those mouse claims are the most fragile of all.
- The biomarker can move while the outcome doesn't. The clearest 2026 example: a supplement (NMN) sharply raises a blood marker (NAD+, roughly 1.5 to 2x), yet meta-analyses of randomized trials show no benefit to muscle or metabolism - and new work finds that blood marker doesn't even decline with age in humans. The gauge moved; the outcome didn't. (Tissue NAD+ may still decline - the stable measure is whole blood.)
- Effects shrink and fragment across sex, strain, and lab.
Method, in two sentences: we took the ITP's published results and the human-trial record as the population and computed the pass rate at each stage; separately, we simulated weighted survival tests (log-rank, Gehan-Wilcoxon, Tarone-Ware) on age-localized effects to measure how often the 'works/null' verdict flips with the test choice. The simulation is illustrative of a known statistical fact, not a new discovery.
What would change our mind: a hard-endpoint human result from any current mouse-stage longevity hit (a positive successor to PEARL, a positive metformin TAME trial, or a human signal from a newer target) would update the likelihood and pull the posterior up - and we would say so.
This is not pessimism about longevity science. It is a calibrated prior - the number you should carry into the next headline.
FAQ
What's the right prior for a “we reversed aging in mice” headline? Low single digits for a proven human benefit, not zero. About 1 in 4 ITP compounds (~8-12 of ~40) extend mouse lifespan in at least one sex, but the prior of a proven human benefit is the ~10% in-trial success rate further discounted by how rarely any of these compounds is ever fairly tested - which lands it in the low single digits, and the sparse longevity human data barely updates it.
Why not just say the human conversion rate is 0%? Because 0 of ~8 is censoring, not a measured failure rate: only one mouse winner (rapamycin, in the PEARL trial) has had a dedicated human longevity trial at all, and it was powered for safety. A rate needs a denominator of completed trials, and that denominator is essentially one.
How big are the mouse wins themselves? Usually small and sex-specific - acarbose added ~22% of male median lifespan but only ~5% of female. A statistically significant mouse result is rarely a large or general one, and the verdict can even depend on which survival test you run.
How should I read the next aging headline? Treat a single mouse result as weak evidence and anchor on the base rate: the prior of a proven human benefit within a decade is low single digits, set by the base rate and by how rarely these compounds are ever fairly tested. Demand a hard human endpoint before updating - and note that flashy 'reverse aging' headlines are often not even ITP compounds, a less-validated class still.