In February 2013, a Spanish trial called PREDIMED gave the Mediterranean diet its scientific crown. Published in the New England Journal of Medicine, it reported that a Mediterranean pattern supplemented with extra-virgin olive oil or nuts cut major cardiovascular events by about 30% in high-risk adults. ⁴ That single number — a 30% reduction from a randomized experiment rather than a mere observation — became the evidentiary backbone for a decade of dietary guidance.

The "Mediterranean diet" itself is not one recipe but a pattern: heavy on vegetables, fruit, whole grains, legumes, nuts, fish and olive oil, light on red meat and processed food. Researchers measure how closely a person follows it using adherence scores, and here lies a recurring complication — there is no single standard. Studies variously use the rMED score, the PREDIMED 14-point test, the MEDLIFE index and others, each weighting different foods and setting different thresholds for "high" adherence. ⁴ A reader comparing two studies is often comparing two different rulers.

What follows examines what the research literature actually documents — the confirmed benefits, the null results, the quality problems that forced a retraction, and how all of it changes depending on who is eating and how it is measured.

The clearest experimental signal concerns the heart. In PREDIMED, 288 major cardiovascular events occurred over 4.8 years, unevenly distributed: 96 in the olive-oil group (3.8%), 83 in the nut group (3.4%), and 109 among controls (4.5%). ¹ Those raw counts translate into unadjusted hazard ratios of 0.70 for each intervention arm — roughly a 30% lower rate of events, the figure that made the trial famous. ⁴ The arithmetic holds up: the reported percentages match the counts once the trial's three arms of about 2,500 people each are accounted for. ¹

Independent syntheses point the same way, though not identically. A meta-analysis of four randomized trials totalling 10,054 participants found the Mediterranean diet reduced major adverse cardiac events with an odds ratio of 0.52, and cut cardiovascular death with an odds ratio of 0.54. ¹ A separate study of nearly 26,000 women reported 25% lower cardiovascular disease risk over 12 years, and a pooled analysis of 16 cohorts covering more than 22,000 women found about 24% lower heart disease risk. ² Those two women's-health figures are close enough — 24% versus 25% — that they may describe the same underlying meta-analysis reported through different outlets, so they corroborate each other only weakly. ² A caution also applies to combining these numbers: PREDIMED reports hazard ratios that account for time-to-event, while the meta-analyses report odds ratios that do not, and the two metrics are not directly comparable in magnitude.

The metabolic picture is genuinely split. In PREDIMED, the olive-oil arm saw a 40% lower diabetes rate (hazard ratio 0.60). ² But a large meta-analysis of 58 studies covering 74,058 adults found the diet improved waist circumference, triglycerides and HDL cholesterol while showing no significant effect on fasting blood glucose (SMD −0.21) or systolic blood pressure (SMD −0.15). ¹ In other words, the diet moved some metabolic markers and not others.

The most important limit sits at the top of the outcome ladder. PREDIMED showed no effect on all-cause mortality, a finding echoed by a Spanish review that found no cause-and-effect relationship with total mortality and by the four-trial meta-analysis, where all-cause mortality was not significantly reduced (OR 0.77, p=0.20). ³ The diet appears to shift cardiovascular events without, in trial settings, measurably extending life.

Open: Whether the ~26,000-women study and the 16-cohort pooled analysis are the same underlying dataset reported twice or two independent bodies of evidence.; Which specific cardiovascular components (heart attack, stroke, cardiovascular death) drive the composite reductions in the four-trial meta-analysis, whose MACE definition is unspecified.

Beyond the heart, the strongest long-term claims come not from trials but from large observational cohorts — and that distinction matters. A UK Biobank study of 110,799 people using the MEDLIFE lifestyle index reported 29% lower all-cause mortality and 28% lower cancer mortality over nine years. ¹ A systematic review of 14 cohort studies found the highest Mediterranean adherence associated with 14% lower cancer mortality (RR 0.86). ² Those two findings, drawn from different populations and different scoring systems, land in broadly the same range and reinforce each other. An official synthesis agrees on longevity: the Italian national health institute reports a relative risk of 0.96 for all-cause mortality per one-point increase in adherence — about 4% lower per point — at moderate certainty, consistent with the larger effect seen for multi-point differences in UK Biobank. ²

The cancer story carries a sharp caveat. The same review that found lower cancer mortality overall found no significant benefit among people who already had cancer — the association between highest adherence and cancer mortality or recurrence in survivors was null (RR 0.95, based on just four studies). ² And the review's authors found no significant associations for individual food components — cereals, dairy, fish, legumes, meat, nuts or olive oil taken alone — while noting funnel-plot asymmetry that means publication bias cannot be ruled out. ¹ The benefit, if real, appears to live in the whole pattern rather than any one food.

Dementia is where the evidence thins most. The EPIC-Spain cohort of 16,160 people followed for 21.6 years found high versus low adherence linked to 20% lower dementia risk — but the confidence interval crossed one (HR 0.80, 0.60–1.06), so the result was not statistically significant. ¹ Only when adherence was treated as a continuous score did a significant effect appear: 8% lower risk per two-point increment (HR 0.92, p=0.021). ¹ Against this sits a US cohort of 13,630 people that found no significant dementia associations for the Mediterranean or two other diet scores after 27 years of follow-up. ¹ These studies used different populations, different adherence scores and different follow-up lengths, so they cannot be treated as measuring the same thing — but the divergence is real, and it means the dementia claim rests on a single cohort showing a borderline effect.

Open: Whether the observational cancer- and all-cause-mortality benefits reflect the diet itself or the health-conscious behaviours that cluster with it, absent any randomized cancer-outcome trial.; Whether the EPIC-Spain continuous dose-response would survive replication in a non-Mediterranean population using the same adherence instrument and dementia ascertainment.

The Mediterranean diet's flagship result did not simply age — it was formally corrected. In June 2018 the New England Journal of Medicine retracted the 2013 PREDIMED paper and republished it, after the discovery that a share of the more than 7,400 participants had not been randomly assigned to their diet groups. ³ Randomization is the mechanism that makes a trial trustworthy: without it, the groups being compared may differ in ways that mimic or mask a treatment effect. One documented failure was stark — at one of the trial's 11 sites, the lead investigator assigned the same diet to an entire village without telling the other investigators. ¹

How much of the trial was affected is itself disputed, and the disagreement is instructive rather than damning. NPR, relaying the retraction, reported that about 14% of participants were not randomly assigned. ³ A PLOS blog by methodologist Hilda Bastian put the figure higher: 21%, or 1,588 of 7,447, attributing it to household-level assignment, a misused randomization table and clinic-level randomization at one site. ¹ Medscape's coverage implied a figure closer to 10%. ¹ These are not three estimates of one quantity so much as three different scopes — the narrow count of directly non-randomized individuals versus the broader count of everyone touched by any protocol deviation. ³ They do not contradict; they measure different things.

The decisive question is whether the correction changed the answer. It barely did. In the republished analysis, the adjusted hazard ratio for cardiovascular events moved from 0.70 to 0.69 for the olive-oil arm and from 0.70 to 0.72 for the nut arm — shifts of a hundredth or two, well within the original confidence intervals. ¹ The headline benefit, in other words, survived re-analysis intact. ⁴

But critics documented deeper problems that predate and outlast the retraction. According to Bastian, the UK's NICE had assessed PREDIMED at "serious" risk of bias with low or very low quality data and considered it underpowered for mortality, and two Cochrane reviews had excluded the trial before the retraction, chiefly because its control group was not a genuine minimal-intervention comparator. ¹ Bastian also reported that even for the full trial group there was no statistically significant benefit on heart attack or cardiovascular death — only on stroke. ¹ These are single-source, reported-grade claims relaying the assessments of others, but they are consistent with the null all-cause-mortality findings established elsewhere. ³

The quality concerns are not confined to PREDIMED. A 2026 systematic review found that of 11 randomized trials comparing Mediterranean and low-fat diets, only 2 measured hard cardiovascular outcomes at all, and one of those (Tuttle et al.) found no significant difference between the Mediterranean and low-fat groups, though both beat usual care. ¹ Meta-analyses report severe statistical heterogeneity — I² of 87% across the four-trial cardiovascular analysis and 95–98% across metabolic-syndrome parameters, with evidence of publication bias for triglycerides. ² And in 2025 the Italian national health institute issued a strong recommendation against using a nut-enriched Mediterranean diet for the sole purpose of reducing atrial-fibrillation risk in high-cardiovascular-risk people — a rare formal "do not do this" from an official body. ¹

Open: Whether a PREDIMED analysis restricted to verified randomly-assigned participants would reproduce the corrected effect sizes, which no source in this record reports.; The exact content of the NICE risk-of-bias assessment and the two Cochrane exclusion decisions, available here only via a single relaying blog.

Even granting the trial results, a practical gap separates the laboratory from the dinner table. PREDIMED's own achieved difference in adherence between intervention and control groups was modest — just 2 points out of 14. ¹ If a tightly counselled trial with free food could open only a small gap, the real-world gap is likely smaller still, which matters because effect sizes scale with how much the diet actually changes.

Real-world adherence bears this out, and does so consistently across independent surveys. A Portuguese survey of 1,000 people found only 26% scored high on the PREDIMED test. ¹ A survey spanning Spain, Italy, Türkiye and Lebanon found roughly 22% strongly adhered. ¹ Among 250 healthy Greek adults, 82% showed only moderate adherence and just 14% high. ¹ Despite using three different scoring systems in three different settings, all three converge on the same picture: even in the Mediterranean's home countries, only about one in five to one in four adults eats this way. ³ That same Greek study did, however, find that higher adherence tracked with lower BMI, waist circumference and waist-to-hip ratio. ¹

Benefits also appear to vary by age. In the UK Women's Cohort Study of 24,703 women followed for nearly 20 years, the diet's protective association with multimorbidity was attenuated and non-significant among women aged 60 and over (HR 0.90, confidence interval crossing one) — suggesting the pattern may do more when adopted earlier. ¹ And the dementia evidence discussed above underlines the timing problem in a different way: the EPIC-Spain association surfaced only after excluding people who misreported their energy intake and only after roughly 18 years of follow-up. ¹ An effect that requires near-perfect dietary reporting and two decades to appear is, whatever its biological reality, a demanding one to deliver at population scale.

Open: Whether population-level promotion (workplace or municipal programmes) can achieve adherence gaps comparable to intensively counselled trials, verified by objective biomarkers.; Why the multimorbidity protection attenuates after age 60 — whether from competing risks, measurement, or a genuine window of effectiveness earlier in life.

The evidence supports several explanations at once, and the honest task is to say which the data force and which they merely permit.

The leading reading is that the Mediterranean diet genuinely reduces cardiovascular events in higher-risk people by roughly 30%. This is supported by PREDIMED's corrected hazard ratios, the four-trial meta-analysis, and the women's cohort findings of about 24–25% lower risk. ⁴ The chief thing weighing against it is that even in the full PREDIMED group the benefit reached significance only for stroke, not heart attack or cardiovascular death. ¹ The finding is plausible and reasonably well-supported, but a clean replication — an independent trial in a non-Mediterranean population with verified randomization and a composite endpoint powered beyond stroke alone — would settle it.

A second, well-established reading concerns what the diet does not do: it does not measurably reduce all-cause mortality in trial settings. Three independent lines agree — PREDIMED, a Spanish review, and the four-trial meta-analysis (OR 0.77, not significant). ³ The observational longevity signals are the counterweight, but they carry the confounding risk inherent to cohort studies. ³ The most economical explanation is that trials have been too short and too underpowered for deaths to detect a mortality effect that observational data hint at — a genuine open tension rather than a resolved contradiction.

A third explanation holds that PREDIMED's benefit was inflated by its randomization failures. The retraction, the village episode, the NICE and Cochrane assessments all support the worry. ³ But the strongest single piece of contradicting evidence is decisive on its own terms: the corrected re-analysis moved the hazard ratios by no more than two hundredths. ¹ Unless a subgroup analysis restricted to verifiably randomized participants shows otherwise, the bias hypothesis explains the scandal but not a collapse of the effect.

A fourth reading treats the benefit as real but heavily context-dependent — strong for lipids and waist circumference, absent for fasting glucose and blood pressure, attenuated after age 60, detectable for dementia only after 18 years and only in one cohort, and null in cancer survivors. ⁴ This is the explanation the fine-grained data most strongly support: the pattern behaves less like a uniform drug and more like a modifier whose size depends on outcome, population, timing and measurement.

Finally, a practical hypothesis: whatever the efficacy, low real-world adherence blunts the impact. The convergent surveys showing 14–26% high adherence and PREDIMED's own 2-point achieved gap make this the best-supported claim of all — it does not require the diet to work in order to be true, only that few people follow it closely. ⁴

The evidence forces a qualified conclusion. On cardiovascular events, the direction of benefit is robust: it appears in the corrected landmark trial, in an independent trial meta-analysis, and in large cohorts. ⁴ The retraction damaged PREDIMED's reputation but not, on the record here, its central number. ¹ A reader is entitled to conclude that closer adherence to the pattern is associated with fewer cardiovascular events and lower cardiovascular death. ¹

The evidence equally forces caution in the other direction. There is no trial-based case that the diet extends overall life; its metabolic effects are selective, not comprehensive; its dementia benefit is borderline in one long cohort and absent in another; and it offers no documented benefit to people who already have cancer. ⁴ The meta-analytic heterogeneity is severe, and at least one official body has formally recommended against one specific use. ³ These are not fatal to the diet, but they are fatal to any claim of a universal, large, causally certain effect.

Weighing the two steelman cases, neither wins outright. The affirmative case is strongest on cardiovascular events and weakest on mortality and food-component specificity; the skeptical case is strongest on methodology and real-world applicability and weakest against the fact that the corrected effect held. ¹ What the record best supports is a middle position: a plausible, moderate, pattern-level cardiovascular and longevity association, established with more confidence in observation than in experiment, and delivered to a population where — speculatively, given adherence figures of one in five to one in four and a trial gap of only 2 of 14 points — the achievable real-world effect is likely smaller than the trial figures imply. ⁴ That last inference is labelled speculative: it follows the logic that effect size scales with adherence, but no study in this record measures the population-level clinical outcome of a real-world adherence campaign.