How to Read a Longevity Study

Longevity is the worst possible field for evidence. The outcome you care about takes decades to appear, the interventions are behaviors people cannot be blinded to, and the population that volunteers for health studies is systematically different from the population that does not.

So the field runs mostly on association. And association, in this field, has a long history of being confidently, expensively wrong.

What follows is not a methods lecture. It is four documented cases where the evidence looked strong and was not, and the specific question that would have caught each one.

Case 1: HDL cholesterol

For decades, “good cholesterol” was standard advice. The observational evidence was excellent. Pooled conventional epidemiology found that a 1 standard deviation rise in usual HDL cholesterol was associated with substantially lower myocardial infarction risk — an odds ratio of 0.62 (95% CI 0.58–0.66) (Voight et al., 2012). A 38% lower risk. Consistent, large, biologically plausible.

Then researchers ran a Mendelian randomization study. The logic: people inherit genetic variants that raise HDL more or less at random, independent of lifestyle, from birth. If high HDL causes lower heart attack risk, people who inherited HDL-raising variants should have fewer heart attacks.

They did not.

Carriers of the LIPG Asn396Ser variant had higher HDL and showed no association with myocardial infarction (OR 0.99, 95% CI 0.88–1.11, p=0.85). A genetic score of 14 HDL-raising variants: also nothing (OR 0.93, 95% CI 0.68–1.26, p=0.63).

The same paper ran the same method on LDL cholesterol as a control. There, a 1 SD genetic increase in LDL was strongly associated with myocardial infarction (OR 2.13, 95% CI 1.69–2.69, p=2×10⁻¹⁰). The method works. It just does not validate HDL.

The interpretation: high HDL is a marker of something protective, not the protective thing itself. Several pharmaceutical programs to raise HDL failed after this.

The question to ask: is there genetic or randomized evidence, or only observational? For any biomarker you are trying to move, this is the first question, not the last.

Case 2: Beta-carotene

Observational studies repeatedly found that people with higher beta-carotene intake and blood levels had less lung cancer. The obvious inference was to supplement.

Two large randomized trials tested it in high-risk populations. The Beta-Carotene and Retinol Efficacy Trial randomized 18,314 smokers and asbestos-exposed workers to daily beta-carotene plus retinyl palmitate or placebo. It was stopped early, in January 1996, because the treatment arm showed 28% more lung cancer and 17% higher mortality than placebo (Goodman et al., 2004).

The supplement did not fail to help. It killed people.

Follow-up over six years after supplements stopped found relative risks for lung cancer and all-cause mortality remained above 1.0 throughout, though no longer statistically significant.

The likely explanation is that beta-carotene in food travels with everything else in vegetables and with the kind of person who eats them. Isolated, at pharmacological dose, in damaged lung tissue, it behaves differently.

The question to ask: is the exposure being studied the same as the intervention being sold? “People who eat X live longer” and “taking X in a capsule makes you live longer” are different claims requiring different evidence.

Case 3: PREDIMED

This one is subtler, because the conclusion survived.

PREDIMED was a landmark trial of a Mediterranean diet for cardiovascular prevention, published in the New England Journal of Medicine in 2013. In 2018 it was retracted and republished after the authors’ own review found protocol deviations: household members enrolled without randomization, one of eleven sites allocating whole clinics rather than individual patients, and apparent inconsistent use of randomization tables at another site (Harvard Nutrition Source, 2018).

The reanalysis adjusted for family and clinic clustering and excluded 1,588 participants — roughly 21% of the sample — whose assignment deviated from protocol. The headline finding held: cardiovascular disease incidence roughly 30% lower in the Mediterranean diet groups.

But something did change. The republished study is no longer a clean individual-level randomized trial. It is better described as quasi-experimental, with cluster-randomization at some sites and household contamination at others.

The result survived. The strength of the claim did not.

The question to ask: what is the study’s actual design, as executed, not as described in the abstract? A trial that was partly cluster-randomized supports a weaker inference than one that was not, even when the point estimate is identical.

Case 4: Sleep and dementia

Short sleep in midlife is associated with later dementia. The Whitehall II cohort followed 7,959 people for a mean of 24.6 years and found that sleeping six hours or less at age 50 carried a hazard ratio of 1.22 (95% CI 1.01–1.48) for dementia, and 1.37 (1.10–1.72) at age 60 (Sabia et al., 2021).

The problem is that dementia has a preclinical phase lasting many years, and disrupted sleep is one of its early features. So does short sleep cause dementia, or does incipient dementia cause short sleep?

The authors name this directly: reverse causation bias is a central challenge, and their design — measuring sleep at specific ages decades before diagnosis rather than shortly before — is an attempt to mitigate it, not to eliminate it. The age-70 estimate, closest to diagnosis, was the one that lost statistical significance (HR 1.24, 95% CI 0.98–1.57).

The question to ask: could the outcome be causing the exposure? In any study where the disease has a long silent phase, the answer is often yes, and the honest papers say so.

What to do with this

• Rank evidence by design before you rank it by result. Randomized trial beats Mendelian randomization beats prospective cohort beats cross-sectional beats mechanism. A dramatic finding from a weak design is still a weak finding.
• Ask what a null result would have looked like. If a study could not have come out negative — because it measured a proxy, or an unblinded self-report, or a population that self-selected — its positive result carries little information.
• Check whether the intervention matches the exposure. Food is not a supplement. Fitness is not exercise. HDL level is not HDL function.
• Look for the caveat paragraph. Good papers name their own weaknesses, as Sabia’s team named reverse causation. If a paper claims no meaningful limitations, that is itself a limitation.
• Give special weight to disconfirmation. The HDL and beta-carotene stories both came from researchers testing a belief they expected to confirm. That is the most valuable kind of study and the least commonly promoted.

None of this means observational evidence is worthless. It means it is a hypothesis generator. The gap between “associated with” and “causes” has cost this field decades and, in the beta-carotene case, lives. Read every longevity claim as sitting somewhere on that gap, and make the author show you where.