The PREDIMED Trial, Read Honestly

PREDIMED is the most famous randomized trial in nutrition science — and the most instructive one to read carefully. Its headline number, a roughly 30% reduction in major cardiovascular events, was retracted, re-analyzed, and republished. This page walks through what happened, what changed, and what still stands.

The Claim: 30% in One Number

PREDIMED — PREvención con DIeta MEDiterránea — enrolled 7,447 adults in Spain, aged 55–80, each with either type 2 diabetes or at least three cardiovascular risk factors, and none with cardiovascular disease at baseline. They were assigned to one of three arms: a Mediterranean diet supplemented with extra-virgin olive oil (roughly a liter a week for the household), a Mediterranean diet supplemented with mixed nuts (about 30 grams a day), or a control group advised to reduce all dietary fat. The primary endpoint was major cardiovascular events — myocardial infarction, stroke, or cardiovascular death. After a median 4.8 years, the event counts were 96 in the olive oil arm (3.8%), 83 in the nuts arm (3.4%), and 109 in the control arm (4.4%). Translated to rates: 8.1, 8.0, and 11.2 events per 1,000 person-years. The hazard ratios were 0.70 (95% CI, 0.54–0.92) for the olive oil arm and 0.72 (0.54–0.97) for the nuts arm (NEJM, 2013). That is the entire origin of "30%" — a relative reduction of about three events per 1,000 person-years of adherence.

What Happened in 2018

In June 2018 the New England Journal of Medicine retracted the paper and republished a corrected version on the same day. The trigger was a 2017 analysis by the British anaesthetist John Carlisle, who had built a statistical tool for spotting suspect randomization in published trials. Scanning PREDIMED's baseline tables, he found the between-arm distributions of variables like height and waist circumference looked improbably tidy — more alike than individual randomization should produce. The authors' own audit then found the problem: at one of the eleven recruiting sites, allocation had been carried out by household or clinic rather than person by person, meaning people sharing a kitchen could land in different arms. The republication re-analyzed the data with statistical methods that respect that cluster structure; the hazard ratios landed at 0.69 and 0.72 with slightly wider confidence intervals, and the conclusion was unchanged (NEJM, 2018). Two things deserve emphasis. First, this was a procedural error, not fraud — no fabrication was alleged by anyone, and the journal's notice says exactly that. Second, the sequence itself is the system working: an outsider's tool flagged a problem, the journal acted, the authors re-analyzed, and the record was corrected within a day.

What the Retraction Did Not Touch

Three things were never in question. The diets were actually delivered: adherence was verified with objective biomarkers — urinary hydroxytyrosol in the olive oil arm and plasma alpha-linolenic acid in the nuts arm — which is rare and valuable in a free- living diet trial. The outcomes were adjudicated by a blinded events committee, so the endpoint data were sound. And, most importantly, the Mediterranean pattern's case never rested on PREDIMED alone. The Lyon Diet Heart Study randomized 605 post-heart-attack patients to a Mediterranean-style diet or a conventional "prudent" diet and found 50–70% fewer recurrent events in the Mediterranean group (Circulation, 1999). CORDIOPREV randomized 1,002 patients with established coronary disease and, over seven years, found fewer major cardiovascular events on a Mediterranean diet than on a low-fat diet, with the difference concentrated in men (Lancet, 2022). A Cochrane review concluded Mediterranean-style diets probably reduce cardiovascular events in both primary and secondary prevention (2019). A retracted-and-republished trial that agrees with its neighbors is more credible, not less.

Reading the Fine Print

Four details calibrate the headline. First, the control arm was not actually low-fat: its target was 30% of energy from fat or less, but measured intake stayed around 37% — so PREDIMED compared a supplemented Mediterranean pattern with fairly standard advice that never fully took. Second, the Mediterranean arms received free olive oil and nuts plus quarterly dietitian sessions, while the control group got written material and occasional small non-food gifts; attention was unevenly distributed. Third, the trial was stopped early at the interim analysis once the benefit crossed the prespecified boundary, and early-stopped trials tend to overestimate effect sizes. Fourth, the population was older, high-risk, and Spanish. Extrapolating "30%" to a healthy 35-year-old goes beyond what the data license. None of this overturns the result — it calibrates it.

The Sober Arithmetic

Relative risk reductions are the headline; absolute rates are the story. Thirty percent relative translates to about three fewer major events per 1,000 person-years, and the commonly cited number needed to treat is roughly 61 high-risk people treated for about five years to prevent one event. For someone with multiple risk factors, that is meaningful — the same order of magnitude as what intensive risk-factor care buys. For everyone else, the honest framing is that the pattern's value was never one number: it is the downstream stack of weight stability, glucose control, and cognitive outcomes, which the disease states page unpacks. And PREDIMED's most durable finding may be the least quoted one: thousands of people stayed on the diet for years and liked it, which is what makes any dietary pattern work in the long run.

Lessons for Reading Any Diet Trial

• 📋 Ask how randomization was actually done. Individual, household, or clinic? Cluster allocation is common and fixable, but it should be disclosed — and analyzed for.
• 🧾 Ask what the control arm actually ate. "Low-fat control" describes intent, not dinner. What people were measured eating is the honest label.
• ⚖️ Ask for absolute rates. Hazard ratios give direction; events per 1,000 person-years give size. Both matter.
• 🛑 Check for early stopping. Trials halted at interim analyses systematically overstate benefits. Budget for it.
• 🔁 One trial is never the case. Look for the neighbors — Lyon, CORDIOPREV, the cohorts. Converging independent evidence is the real verdict.
• 🧯 A retraction is not a verdict of fraud. Corrected science is science working. The question is whether the conclusion survives the correction.

Questions, Answered Briefly

• 🚫 Was the retraction because of fraud? No. The problem was procedural — allocation at one site was not individual — and no fabrication was alleged. The journal retracted and republished with corrections in the same issue.
• 🔢 Are the numbers different now? Barely. The reanalysis produced hazard ratios of 0.69 and 0.72 where the original had 0.70 and 0.72. Direction and rough size are unchanged.
• 🏛️ Is PREDIMED the whole case for the Mediterranean diet? No. The pattern's support comes from converging evidence — Lyon, CORDIOPREV, Cochrane, and decades of cohorts. PREDIMED is the largest single brick, and it held.
• 🗣️ Can I still say "30%"? With the fine print: roughly 30% relative reduction in major cardiovascular events in high-risk older adults, surviving reanalysis, against a control that never became low-fat — about three fewer events per 1,000 person-years.

Related Topics