Domain-opener / nucleus for fruit-and-vegetable intake (single gold source: Aune 2017, a dose-response SR+MA of 95 prospective cohorts / 142 publications, up to ~2.1 million participants and up to 43k CHD, 47k stroke, 82k CVD, 112k cancer cases and 71k-94k deaths). All evidence here is observational — no whole-food RCT of F&V vs a hard outcome exists (PREDIMED tests a whole Mediterranean pattern, not F&V alone) — so the ceiling on certainty is confounding, not sampling. The exposure is whole fruit + vegetables combined (80 g = 1 serving); potatoes/legumes handled as subtypes, juices and tinned forms separated out below.
The per-200 g/day effect estimates (linear dose-response, random-effects)
| Outcome | RR per 200 g/day F&V (95% CI) | I2 | n | Fruit alone | Veg alone |
|---|---|---|---|---|---|
| Coronary heart disease | 0.92 (0.90-0.94) | 0% | 15 | 0.90 (0.86-0.94) | 0.84 (0.79-0.90) |
| Stroke (total) | 0.84 (0.76-0.92) | 73% | 10 | 0.82 (0.74-0.90) | 0.87 (0.79-0.96) |
| Cardiovascular disease | 0.92 (0.90-0.95) | 31% | 13 | 0.87 (0.82-0.92) | 0.90 (0.87-0.93) |
| Total cancer | 0.97 (0.95-0.99) | 49% | 12 | 0.96 (0.94-0.99) | 0.96 (0.93-0.99) |
| All-cause mortality | 0.90 (0.87-0.93) | 83% | 15 | 0.85 (0.80-0.91) | 0.87 (0.82-0.92) |
- The gradient is shallow per serving but real. ~8-10% lower risk per 200 g/day for the CVD/death endpoints; cancer is the weak arm (~3% per 200 g/day). CHD is the tightest estimate in the corpus (I2 0%, CI 0.90-0.94); stroke and all-cause carry high heterogeneity (I2 73-83%) — driven by size of association, not its presence (most cohorts inverse).
- Stroke subtype split: ischaemic 0.92 (0.87-0.97, I2 9%) vs haemorrhagic 0.88 (0.78-0.99, I2 0%) per 200 g/day — both inverse, ischaemic tighter.
- F&V is nutritionally not one thing. Aune attributes the effect to «a myriad of nutrients and phytochemicals, including fibre, vitamin C, carotenoids, antioxidants, potassium, flavonoids and other unidentified compounds which are likely to act synergistically» (Aune et al., 2017) — so F&V benefit is not clean evidence for any one component (the observed-healthy-population / component-not-category trap) -> Dietary Fibre and Health, Potassium Intake and Blood Pressure.
Dose-response SHAPE — outcome-specific, and the “lowest risk” is the study-edge
Lowest observed risk (nonlinear splines): cancer plateaus at 600 g/day (7.5 servings) with «little evidence of further reductions» above; CHD, stroke, CVD and all-cause reach their lowest at 800 g/day (10 servings) — «a level of intake that is double the five servings per day (400 g/day) currently recommended» (Aune et al., 2017). Reductions at 800 g/day: CHD 24%, stroke 33%, CVD 28%, all-cause 31% (vs no intake); cancer 14% at 550-600 g/day.
- 800 g/day is the sampling edge, not a demonstrated optimum. It is where the studied range thins out, so read it as a floor-for-most-benefit region, not a point-target — the The Underivable Optimum rule. The curve is steepest at the low end (all-cause: steepest up to ~400 g/day), so the first servings carry most of the effect; the marginal value of the 9th-10th serving is small and uncertain.
- Shape is outcome-specific, and CHD-combined is the anomaly: F&V-combined -> CHD showed no nonlinearity (Pnonlinearity 0.30, approximately linear over the range), while fruit-alone and veg-alone -> CHD were nonlinear (Pnonlin < 0.0001) — a per-food-group display artifact as much as a biology claim. Stroke, CVD, cancer and all-cause were all nonlinear (steeper at low intake).
- The high-intake flattening may be a measurement artifact, not a true plateau. Aune: «we can also not entirely exclude the possibility that the weaker dose-response curve at higher compared with lower intakes could partly be due to measurement errors» — a hidden knee, not a real one. Measurement can HIDE a knee but not manufacture one, so the «every reduction pays» default holds -> Measurement Error in Dietary Assessment, The U-Shaped Association Artifact.
Specific types — the pooled benefit is NOT uniform, and the “superfoods” are the thin cells
Per 100 g/day dose-response (or high-vs-low where noted). The tight, steep inverse signals sit with leafy greens and citrus, NOT the berry/grape end that carries most marketing:
| Type | CHD | Stroke | Read |
|---|---|---|---|
| Green leafy vegetables | 0.72 (0.64-0.82) | 0.73 (0.57-0.94) | steepest + tightest, both endpoints |
| Citrus fruits | HvL 0.91 (0.86-0.96) | 0.78 (0.69-0.90) | robust, esp. stroke |
| Beta-carotene-rich F&V | 0.77 (0.65-0.91) | — | steep |
| Apples/pears | HvL 0.85 (0.79-0.93) | HvL 0.88 (0.81-0.96) | HvL sig, dose-resp NS |
| Pickled vegetables | — | 0.57 (0.43-0.74) | steep (2 studies) |
| Grapes | 0.87 (0.56-1.37) NS | 0.57 (0.34-0.97) | CHD null; stroke sig on 2 studies |
| Berries | 1.13 (0.90-1.43) NS | 1.07 (0.79-1.45) NS | point estimate >1 both; haemorrhagic 1.66 NS |
| Strawberries | 4.66 (1.14-19.03) | — | single study, absurd CI — discard |
| Cruciferous (CHD/stroke) | 0.99 NS | 1.04 NS | NS for CVD; sig for total cancer |
- Answering the grapes-vs-berries question directly: neither grapes nor berries shows a robust CHD benefit (both NS, berries point above 1.0); grapes’ one significant signal (stroke 0.57) rests on 2 studies with a wide range; the strawberry 4.66 is a one-study artifact. The subtype cells are thin (n=2-6, wide CIs) — the data cannot individuate most single fruits, and where they can, the winners are leafy greens/citrus, not berries/grapes. Do not read the pooled F&V effect as distributing evenly onto a favourite fruit.
- FORM matters, and processing is the harm edge: fruit juice was inverse for stroke (HvL 0.67 [0.60-0.76]; per-100 g 0.72 [0.63-0.83]) and CHD (HvL 0.79 [0.63-0.98]) in THIS body of evidence, whereas tinned fruit was positively associated with cardiovascular disease (a harm signal) — so the whole-vs-processed axis, not a blanket fruit-vs-juice rule, is what these data mark. Note the outcome scope: this is CVD/stroke, not the metabolic/hepatic channel where free-sugar juice concern lives -> Free Sugars Intake.
Measurement error and regression dilution — the effect is if anything UNDER-stated
Only three cohorts corrected for measurement error / regression-dilution, and each correction made the association stronger, not weaker:
- EPIC: all-cause HR 0.97 (0.96-0.98) uncorrected -> 0.94 (0.91-0.96) corrected per 200 g/day; IHD mortality 0.97 (0.95-0.99) -> 0.95 (0.91-0.99) per 80 g/day.
- China Kadoorie: CV death per one daily portion of fresh fruit 0.77 (0.72-0.83) before -> 0.63 (0.56-0.72) after regression-dilution correction — a large attenuation of the true effect by self-report noise. (Aune et al., 2017)
- Aune’s own reading: «both measurement error and regression dilution bias may have attenuated the observed risk estimates» and «Confounding might exaggerate the observed associations but measurement errors would most likely tend to attenuate» — the two large biases push in opposite directions, so the reported RRs are a floor on the causal signal if the association is causal at all -> Measurement Error in Dietary Assessment.
Why the certainty stays low — the confounding ceiling
(inferred from Aune et al., 2017) The consistency is strong (95 cohorts, robust to one-study-out and to region stratification, modest publication-bias signal that did not move the summaries), but the causal reading is capped:
- Healthy-user confounding, the textbook profile. F&V intake is «often associated with other lifestyle factors such as lower prevalence of smoking, less overweight and obesity, higher physical activity and lower intakes of alcohol and red and processed meat» (Aune et al., 2017). Associations persisted after adjustment, but adjustment is the weak check; there is no whole-food RCT on a hard outcome, and the one MR hint (Kobylecki, genetic plasma vitamin C) is a single embedded study, not the design’s verdict -> The U-Shaped Association Artifact (adjudication-route framing).
- Bradford-Hill invoked but not decisive. Aune argues consistency, temporality, a biological gradient, plausibility and coherence — a moderate-strength association that «should not be dismissed as non-causal», while granting residual confounding cannot be entirely ruled out. This is an argument for direction, not proof of magnitude.
- Reverse-causation / access route. Persons with high F&V intake «may be more likely to undergo screening or have better access to or compliance with treatment» — a mortality-specific bias the design cannot remove.
The population-absolute expression — attributable deaths (causal assumption flagged)
The relative effects become an absolute quantity only via baseline risk. Aune’s population layer, assuming causality: an estimated 5.6 million (intake below 500 g/day) and 7.8 million (below 800 g/day) premature deaths worldwide in 2013 attributable to low F&V intake — of the latter, ~1.34M CHD, ~2.68M stroke, ~660k cancer. This is a modelled population burden (Miettinen PAF, World Health Survey prevalence x GBD 2013 mortality), inheriting the causal, no-confounding and generalizability assumptions Aune names -> Baseline Risk and the Relative-Absolute Split. It is not a per-person effect and not a validated projection.
Decision relevance
- F&V is a moderate, observational-grade lever, best framed as addition/substitution, with most of the associated benefit reached in the first few servings and the curve steepest at low intake — so moving from ~0-2 to ~5 servings is where the decision lives; the 5->10 servings gap buys progressively less and rests on the thin high-intake edge. Rank below the big rocks and below RCT/MR-backed levers.
- Prefer whole forms and leafy greens/citrus where a within-F&V choice is being made — the subtype data that survive are there, and tinned fruit carries a harm signal. But do not over-personalize onto a single fruit: the subtype cells are too thin to individuate (route-b effect-modification evidence is absent).
- The ceiling is a finding. For someone already at ~5 servings of varied whole F&V, the remaining lever (more servings, or swapping fruit for veg) is small and uncertain — a licence to stop optimizing here, not a failure to find more.
- Open loop: nothing grades F&V intake against a realized outcome in a randomized design; the causal step rests on consistency + mechanism + the toward-null measurement argument, not adjudication.
(inferred from Aune et al., 2017) Independence note for later weaves: Aune’s team (Imperial/NTNU; Aune-Norat-Riboli-Giovannucci) also authored the parallel whole-grain (ref 192) and nut (ref 193) dose-response MAs cited in this paper, using the same methods and overlapping cohort infrastructure. Any future consistency across Aune’s plant-food MAs is same-lab type-F / shared lineage, NOT independent type-E backing -> Nut Consumption and Mortality, Whole Grains Refined Grains and Pulses.
A guidance family sets a recommended range and reports the same shape — NNR2023 [2026-08-27, NNR revisit]
NNR2023 converts this observational evidence into an actionable range and, in doing so, reports the same dose-response shape this page holds from Aune — an F-refinement that turns a single-source dose-response page into a guidance-anchored recommendation.
- The recommended intake: «It is recommended to consume 500-800 grams, or more, per day of vegetables, fruits and berries in total. A variety of different types of both vegetables and fruits (including berries) should be consumed, with emphasis on dietary fibre contribution (potatoes and pulses are not included).» (Nordic Council of Ministers, 2023) The potatoes-and-pulses exclusion matches Aune’s exposure definition (F&V excluding potatoes/legumes).
- The shape NNR reports mirrors Aune’s: «The largest reductions in risk are generally seen at the lower intake ranges, but for cardiovascular disease, reductions have been observed up to 800 g of fruits and vegetables per day.» (Nordic Council of Ministers, 2023) — steepest at low intake, CVD benefit extending to ~800 g, the same two features Aune’s splines show.
One mild divergence on where all-cause mortality plateaus. Aune places the all-cause nadir at 800 g/day (10 servings); NNR (citing Rosell & Fadnes 2023) reports all-cause mortality «inverse associations with all-cause mortality, levelling off at 5-6 servings of fruits and vegetables per day» (~400-500 g). (Nordic Council of Ministers, 2023) Both are drawn off observational splines at the thinning edge of the data — a difference in which underlying MA set the plateau point, not a contradiction about direction (more F&V, lower risk, flattening somewhere in the 400-800 g region). It sharpens the page’s existing caveat that the high-intake flattening may be a measurement artifact rather than a located knee.
Not independent backing, and confidence stays low. NNR2023 is a guideline synthesis resting on the
same observational F&V literature (Rosell & Fadnes 2023, WCRF/AICR 2018, Fogelholm 2012, Stanaway 2022 —
the field Aune’s cohorts sit inside), so it does not lift the confounding ceiling that caps this
page: there is still no whole-food F&V RCT on a hard outcome. It confirms direction and shape by a
second body, not by an independent route — no [E-independent]; confidence: low is unchanged.
(inferred from Nordic Council of Ministers, 2023)
Refinement — the DIfE/Boeing 12-food-group series (2026-08-28)
The series separates fruit from vegetables and shows their effects diverge. Fruit is modestly protective across all five outcome families (mortality 0.94, T2D 0.98, CHD 0.94, stroke 0.90 per 100 g/d). Vegetables are protective for mortality/T2D/CVD but read null for hypertension (1.00, 0.98-1.01) and null for adiposity — the source attributes the HTN null partly to BMI over-adjustment. Both plateau at ~200-400 g/d. (Schwingshackl et al., 2017) (Bechthold et al., 2017) (Schwingshackl, Schwedhelm, Hoffmann, Knüppel, et al., 2017) Full cross-outcome grid -> Food Groups and Health Outcomes - A Dose-Response Matrix.
Cognition / dementia outcome arm — a second observational lever, weaker than the CVD arm (2026-09-04)
Zhou 2022 (MA of 16 observational studies, 64,348 older adults >=60, 9,879 cases; 9 cohort / 6 cross-sectional / 1 case-control) extends F&V from the CVD/cancer/mortality arms above to a cognitive-disorder outcome — the first cognition evidence on this nucleus. It is observational only, FFQ-dominated, and confidence stays low; treat it as a weaker sibling of the CVD arm, not a new hard finding.
- Pooled highest-vs-lowest: OR 0.82 (95% CI 0.75-0.90), I2 35.3% for cognitive disorders overall; fruit alone 0.83 (0.77-0.89), vegetables alone 0.75 (0.70-0.80) — the same veg-slightly-stronger ordering the CVD arm shows. (Zhou et al., 2022) (The paper’s own abstract headlines 0.79 [0.76-0.83], which is actually its mixed-gender subgroup, not the Figure-2 pool — carry 0.82.)
- Outcome-specific, and Alzheimer’s is NULL. Cognitive impairment/MCI 0.76 (0.72-0.80) and dementia 0.84 (0.78-0.91) are significant, but Alzheimer’s disease is null (0.88, 0.76-1.01) (Zhou et al., 2022) — consistent with a vascular/confounding-mediated route rather than a direct anti-neurodegenerative effect, i.e. the cognition signal likely rides the same cardiometabolic pathway as the CVD arm, not an independent one.
- The design gradient is a reverse-causation tell. The prospective cohort subgroup shows the WEAKEST effect (0.83, I2 0%); cross-sectional (0.70) and case-control (0.68) — the designs where prevalent cognitive disease can depress recalled/current intake — show the strongest. The association shrinks as the design gets less vulnerable to reverse causation, so the pooled 0.82 is an upper bound on any causal reading -> The U-Shaped Association Artifact.
- Dose-response reported linear (P=0.03) but weak. Rests on only 4 studies (the only ones with
=3 exposure categories) with no stated studied-range and no located knee/plateau, so it is “linear over an unstated range on 4 studies” — weak evidence of a true linear curve, and the authors concede it «may have lacked sufficient effectiveness in quantifying the relevant dose-response model» (Zhou et al., 2022). Not comparable to Aune’s 95-cohort splines above -> The Underivable Optimum.
- Same measurement-error and confounding ceilings as the CVD arm, plus two Zhou-specific ones: FFQ/self-report recall bias (23 of 31 effect groups FFQ), and an OR-as-RR approximation the paper flags as inflating the estimate — «using the OR as an approximation of the RR produces progressively larger errors as the outcome rate rises above 1%» (Zhou et al., 2022), and the case fractions here sit well above 1%, so the OR overstates the RR -> Measurement Error in Dietary Assessment. No causal claim: «no casual relationships could be established due to the observational nature of the studies» (Zhou et al., 2022).
Decision read: F&V is a candidate dietary lever for cognitive-disorder risk in older adults, directionally the same as its CVD effect and plausibly mediated by it — not an additive, independently-established brain benefit. It ranks below the established dementia big rocks and is carried as a low-confidence food-group lever on Dementia Prevention and Modifiable Risk Factors.