PREDIMED, dissected: the diet trial that got retracted and came back

In 2018 the New England Journal of Medicine did something it almost never does. It retracted one of the most-cited nutrition papers of the decade — and then republished a re-analysed version of the same trial, with almost identical numbers in it.

The trial is PREDIMED, and if you have ever read anything about the Mediterranean diet, you have read about PREDIMED. It is the study underneath the olive oil evidence, underneath the case for nuts, and underneath the confident claim in most longevity diet writing — including mine — that the Mediterranean pattern is the one dietary approach with proper randomised backing rather than just epidemiology and vibes.

So a retraction is not a small thing. It is worth understanding precisely what went wrong, what the correction did to the numbers, and — the part almost nobody covers — what the surviving result actually claims, which is considerably narrower than what it gets used for.

Bit of a spoiler: the trial comes out of this better than you’d expect and the headline comes out of it much worse.

Why this trial carries so much weight

Nutrition research has a structural problem. You cannot blind people to what they are eating, you cannot keep them on an assigned diet for decades, and the outcomes you care about — heart attacks, strokes, dying — take years to accumulate. So the field runs mostly on observational cohorts, where people who eat olive oil are compared with people who don’t, and the people who eat olive oil also tend to be wealthier, more educated, less likely to smoke and more likely to walk places. The confounding is severe and everybody knows it.

PREDIMED is the rare exception: someone actually assigned the diet, in a large population, and waited for hard clinical events. That is why it does so much work in the literature. Take PREDIMED out and the Mediterranean diet’s evidence base drops back to correlation.

Which is exactly why the retraction mattered enough for the journal to handle it in public rather than quietly.

The design: who, how many, and what they actually ate

Ramón Estruch and colleagues recruited 7,447 adults aged 55 to 80 across multiple centres in Spain. Fifty-seven per cent were women. All were at high cardiovascular risk — type 2 diabetes, or at least three major risk factors — but crucially, none had cardiovascular disease at enrolment. This was primary prevention: people who hadn’t had the event yet.

They were put into one of three groups:

• Mediterranean diet + extra-virgin olive oil — around a litre a week, provided free
• Mediterranean diet + mixed nuts — 30g a day of walnuts, almonds and hazelnuts, provided free
• Control — advice to reduce dietary fat, plus small non-food gifts

Note that those are three arms, not two-things-plus-a-control: nobody was assigned both the olive oil and the nuts. Any figure quoted for “the Mediterranean diet arms” is pooling two separate comparisons, not measuring what happens if you do both.

Nobody was told to lose weight or exercise more. The intervention was the food — though not only the food, and this is the design detail I’d push back on hardest. The two Mediterranean arms got free groceries and quarterly educational sessions; the control arm got advice to cut fat and small non-food gifts. That is not a symmetric amount of attention, and how symmetric it was early in the trial is itself contested — note that the 2013 paper described “quarterly individual and group educational sessions” while the 2018 republication trimmed that to “quarterly educational sessions”. Some of the gap between the arms is a diet. Some of it is being handed free olive oil by someone who keeps checking in on you.

The primary endpoint was a composite: myocardial infarction, stroke, or death from cardiovascular causes. Hold onto the word composite — it becomes important later, because a composite endpoint reports one number for three quite different things, and those three things did not behave the same way.

The trial was stopped early, after a median 4.8 years of follow-up, on the basis of a prespecified interim analysis. Also hold onto that. (According to PubMed: Estruch et al., N Engl J Med 2018, DOI.)

What they found

A primary endpoint event occurred in 288 participants out of 7,447. Broken down:

In the 2018 intention-to-treat analysis, adjusted for baseline characteristics and propensity scores, that came out as a hazard ratio of 0.69 (95% CI 0.53–0.91) for olive oil and 0.72 (95% CI 0.54–0.95) for nuts, versus control.

Which is where “a 30% reduction in cardiovascular events, from a diet” comes from. And as a relative statement it is entirely correct.

The bit that went wrong

The unravelling started somewhere unexpected: an anaesthetist in Torbay.

In 2017 John Carlisle published a screen of 72,261 baseline means from 29,789 variables across 5,087 randomised trials in eight journals, checking whether the distribution of baseline data looked like something randomisation would actually produce. Properly randomised groups differ from each other by chance in a predictable way — too similar is as suspicious as too different. He found that 15.2% of non-retracted trials had baseline p values within 0.05 of 0 or 1, against an expected 10%. (According to PubMed: Carlisle, Anaesthesia 2017, DOI.)

PREDIMED was one of the trials the screen flagged. The investigation that followed found three specific problems, which the 2018 paper states plainly:

1. Household members were enrolled without randomisation. If one member of a household was in the trial, their spouse could be signed up and given the same diet — sensible for adherence, fatal for randomisation.
2. At one of the 11 sites, participants were assigned by clinic rather than individually — a cluster design smuggled into an individually randomised trial.
3. At another site, the randomisation tables appear to have been used inconsistently.

Total affected: 1,588 of 7,447 participants. Twenty-one per cent. One in five people in the flagship randomised trial of the Mediterranean diet was not, in the sense the word usually means, randomised.

None of this looks like fabrication. It looks like a large, distributed, decade-long trial where individual sites did something locally reasonable and methodologically ruinous. Which is, honestly, more unsettling than fraud — fraud is rare and detectable, procedural drift across 11 sites is neither.

What the re-analysis changed, and what it didn’t

Here’s the part that surprises people who only saw the word “retracted.”

The authors re-ran everything without relying on the assumption that all participants were randomly assigned: intention-to-treat with adjustment for baseline characteristics and propensity scores, plus a sensitivity analysis that deleted all 1,588 questionable participants outright.

The numbers barely moved.

The results were also similar after omitting the 1,588. That is a genuinely good outcome for the trial. A result that survives having a fifth of its participants thrown out is a robust result.

But something did change, and it is not in the table. It’s in the prose — and it’s the single most instructive thing about this whole episode.

The 2013 abstract began: “we randomly assigned participants who were at high cardiovascular risk…” The 2018 version begins: “we assigned 7447 participants… who were at high cardiovascular risk.” One word, quietly removed.

And the conclusions:

2013: “a Mediterranean diet supplemented with extra-virgin olive oil or nuts reduced the incidence of major cardiovascular events.”

2018: “the incidence of major cardiovascular events was lower among those assigned to a Mediterranean diet supplemented with extra-virgin olive oil or nuts than among those assigned to a reduced-fat diet.”

The first sentence is a causal claim. The second is a description of what happened in two groups. That is the retraction’s real content: the effect size survived, and the entitlement to say “caused” did not. Everything now rests partly on statistical adjustment, and statistical adjustment can only correct for confounders you thought to measure — which is precisely the weakness randomisation exists to fix.

The caveats the headlines skipped

Right, so the retraction is the famous problem. These three are the ones that actually change how you should use the result, and none of them has anything to do with it.

The 30% is relative, and the absolute number is small

Work it through on the raw event rates. Control 4.4%, olive oil 3.8%: an absolute difference of 0.6 percentage points over 4.8 years. Nuts 3.4%: 1.0 percentage point.

Turn that into a number needed to treat and you get roughly 165 people fed extra olive oil for nearly five years to prevent one event, or about 100 for nuts. (Those are the crude rates, not the adjusted hazard ratios, so treat them as the right order of magnitude rather than precise figures.)

That is not a bad result — for a food, in older adults at high risk, it’s rather good. But “30% reduction” and “one event prevented per 100-odd people over five years” are the same fact wearing very different clothes, and only one of them ever makes it into a headline.

It is a stroke result, not a mortality result

This is the one I’d most want people to know, and it’s the reason a Mediterranean-diet post shouldn’t be filed under “living longer” without a caveat.

The primary endpoint was a composite of heart attack, stroke and cardiovascular death. The 2019 Cochrane review of Mediterranean-style diets pulled the components apart, and they did not move together:

• Stroke: HR 0.60 (95% CI 0.45–0.80) — a fall from about 24 per 1,000 to 14 per 1,000. Cochrane graded this moderate-quality evidence. (This one pools both Mediterranean arms against the low-fat control, so read it as “the diet plus one of the two supplements”, not as either supplement individually.)
• Cardiovascular mortality: HR 0.81 (95% CI 0.50–1.32) — not significant.
• Total mortality: HR 1.00 (95% CI 0.81–1.24). Exactly one. Flat.

(According to PubMed: Rees et al., Cochrane Database of Systematic Reviews 2019, DOI.)

So the composite was carried by stroke. Over 4.8 years, in this population, PREDIMED found no detectable difference in how many people died. The trial wasn’t powered for mortality and wasn’t designed to answer it — but that’s precisely the point. “The Mediterranean diet is the best-proven longevity diet” is a claim the trial never tested, resting on a mortality hazard ratio of 1.00.

It was stopped early, with fewer than 500 events

Trials halted early for benefit are stopped at the moment the result looks best. That is not a conspiracy; it’s arithmetic. Random fluctuation is what triggers the stopping rule, and you stop on the upswing.

Bassler and colleagues quantified this across 91 truncated trials matched against 424 comparable trials that ran to completion: the pooled ratio of relative risks was 0.71 (95% CI 0.65–0.77) — truncated trials produced systematically larger effects. And the effect was worst in trials with fewer than 500 events. (According to PubMed: Bassler et al., JAMA 2010, DOI.)

PREDIMED stopped early, at 288 events. It sits squarely in the zone where early stopping inflates effect estimates most.

Stack that on top of the loss of clean randomisation and the honest reading is that 0.69 is probably the generous end of the true effect, not the middle of it.

What it actually means for you

The thesis, stated plainly: PREDIMED’s effect estimate survived a serious audit almost unchanged, which makes the Mediterranean diet the best-evidenced dietary pattern going — but what survived is a modest stroke-prevention effect in older adults at high cardiovascular risk, with a total mortality hazard ratio of 1.00, not proof that the diet makes you live longer.

Both halves matter. People tend to grab one and drop the other.

Four things I’d actually take from it:

• Keep eating this way. The extra-virgin olive oil and the daily handful of nuts are cheap, pleasant, and each separately beat the control arm in the strongest dietary trial we have. A trial that keeps its effect after losing 21% of its participants has earned that. Doing both is a perfectly sensible thing to do — just don’t imagine the trial tested it.
• Downgrade the claim, not the diet. Say “reduces stroke risk in older adults at high cardiovascular risk” instead of “makes you live longer.” The first is what was measured. The second is marketing wearing the first one’s credentials.
• Notice who was in the room. Participants were 55 to 80, Spanish, at high cardiovascular risk, and already eating fairly Mediterranean-ish before the trial started — the “control” arm was low-fat advice, not a bad diet. What PREDIMED tested was adding olive oil and nuts on top of a decent baseline in people already at elevated risk. Whether it transfers to a healthy 35-year-old is an assumption, not a finding. That’s the same headroom problem that showed up in the DO-HEALTH trial: the size of a benefit is capped by the size of the deficit it corrects.
• Composite endpoints deserve suspicion by default. Whenever you see one, ask which component moved. Here it was stroke, entirely — and a composite that reports one number for three outcomes will always let the strongest one speak for the group.

The thing I’ve come round to, genuinely, is that the retraction is the most reassuring part of the story. A statistical screen written by an anaesthetist with no stake in nutrition flagged a problem in a famous trial; the authors opened their data, re-analysed it three ways, published the corrections, and weakened their own conclusion in public. The result held. That is the system working, slowly and embarrassingly, exactly as it’s meant to.

We just never updated the headline afterwards.