Domain-opener for nuts (single gold source: Aune 2016, a dose-response SR+MA of 20 prospective cohorts / 29 publications, up to 819,448 participants and 85,870 deaths). The exposure is any edible nut — tree nuts, peanuts (a legume), and peanut butter, pooled for their similar nutrient profile. All evidence here is observational (no RCT of nuts-vs-outcome; PREDIMED tests a whole pattern, not nuts alone) — so the ceiling on certainty is confounding, not sampling.
The per-serving effect estimates (dose-response, per 28 g/day = 1 serving/day)
Summary RRs per one serving/day increase in total-nut intake, random-effects:
| Outcome | RR per 28 g/day (95% CI) | I2 | n studies | Note |
|---|---|---|---|---|
| Coronary heart disease | 0.71 (0.63-0.80) | 47% | 11 | the steepest arm |
| Stroke | 0.93 (0.83-1.05) | 14% | 11 | not significant |
| Cardiovascular disease | 0.79 (0.70-0.88) | 60% | 12 | |
| Total cancer | 0.85 (0.76-0.94) | 42% | 8 | |
| All-cause mortality | 0.78 (0.72-0.84) | 66% | 15 | the headline |
| Respiratory-disease mortality | 0.48 (0.26-0.89) | 61% | 3 | few studies |
| Diabetes mortality | 0.61 (0.43-0.88) | 0% | 4 | |
| Neurodegenerative mortality | 0.65 (0.40-1.08) | 6% | 3 | ns |
| Infectious-disease mortality | 0.25 (0.07-0.85) | 54% | 2 | 2 studies only |
| Kidney-disease mortality | 0.27 (0.04-1.91) | 61% | 2 | ns, very wide |
The headline: a ~22% lower all-cause mortality per daily serving. Stroke is the one primary CVD outcome with no significant linear association. The rarer causes of death (respiratory, diabetes, infection) carry the largest point estimates but rest on 2-4 studies — treat as insufficient-evidence-leaning, not banked.
Dose-response SHAPE — a plateau at 15-20 g/day, not a monotone-more-is-better curve
Restricted-cubic-spline analyses found nonlinear associations for CHD, stroke, and all-cause mortality (Pnonlinearity < 0.0001) and for CVD (Pnonlinearity = 0.001), «with most of the reduction in risk observed up to an intake of approximately 15-20 grams per day or 5-6 servings per week for most of the outcomes.» (Aune et al., 2016) Above ~15-20 g/day the curve flattens. Total cancer is the exception — no nonlinearity detected (Pnonlinearity = 0.11), i.e. roughly linear over the studied range. This is a rare observed plateau in this corpus (most decision curves the wiki examined were monotone-or-not-estimable) -> The Underivable Optimum.
- The 20 g/day optimal is a study-edge, not a derived optimum. Aune sets 20 g/day as the PAF reference «because there was little evidence of further reductions in risk above this level of intake» (Aune et al., 2016) — i.e. the number marks where the spline flattens within the sampled range, not a demonstrated point-optimum. Read it as a knee-region (a floor for most of the benefit), not a target -> The Underivable Optimum.
- Measurement caveat on the plateau. Nut intake is FFQ self-report; dietary measurement error compresses slopes and can erase a knee but never manufacture one, so a measured plateau is weak evidence of a true one -> Measurement Error in Dietary Assessment.
Nut subtype — the pooled benefit is not uniform across types
Tree nuts and peanuts both associate with lower CHD, CVD, and all-cause mortality (per 10 g/day, peanuts often steeper, e.g. CVD 0.64 [0.50-0.81]). But the outcome pattern splits: only peanuts reach significance for stroke; only tree nuts for total cancer (tree nuts 0.80 [0.72-0.89]; peanuts null 0.92 [0.82-1.03]). Peanut butter is inverse for mortality in the high-vs-low analysis (0.89 [0.80-0.99]) but null in dose-response (0.94 [0.86-1.02], n=2) — the authors hypothesize added sugar or salt may offset plain-peanut benefit, but 2 studies cannot settle it. The whole-nut pooling hides these seams; the subtype estimates are thin (n=2-5).
The population-absolute expression — PAF (causal assumption flagged)
The relative effects above become an absolute quantity only via baseline risk. Aune’s population layer: assuming causality, ~4.4 million premature deaths in 2013 (the Americas, Europe, SE Asia, W Pacific) would be attributable to nut intake below 20 g/day — 1.19M from CHD, 469k cancer, 1.07M respiratory, 139k diabetes. This is a modelled population burden (PAF = p(rr-1)/(1+p(rr-1)), Miettinen), not a per-person effect, and it inherits every assumption below -> Baseline Risk and the Relative-Absolute Split.
Why the certainty stays low — the confounding ceiling
(inferred from Aune et al., 2016) The consistency is real (20 cohorts, robust to one-study-out, publication bias only in the all-cause analysis and gone after excluding <500-death studies), but three things cap the causal reading:
- Healthy-user confounding. «Subjects with a high intake of nuts tend to be less likely to smoke, to be slimmer and more physically active, and to have a lower intake of red and processed meat and a higher intake of fruits and vegetables» — the textbook profile. (Aune et al., 2016) Associations persisted after adjustment for smoking, alcohol, PA, BMI, and diet, but adjustment is the weak check; there is no MR / genetic instrument and no whole-food RCT, so the effect is not adjudicated causal -> The U-Shaped Association Artifact (adjudication-route framing).
- Measurement error runs toward the null here. No included study corrected for it, but «because of the prospective design … such errors would most likely attenuate the strength of the observed associations» (Aune et al., 2016) — so the true effect is, if anything, larger. (The toward-null direction holds for non-differential error in the univariate case; with mismeasured covariates the bias can run either way -> Measurement Error in Dietary Assessment.)
- PREDIMED cannot isolate nuts. The one RCT (a Mediterranean diet with nuts vs control) cut CVD, but «it is not clear if this association is due to the Mediterranean diet component, nuts, or a combination of the two.» (Aune et al., 2016) -> Mediterranean Diet and Cardiovascular Events.
Decision relevance
- Nuts are a plausible moderate lever, best framed as an addition/substitution, with most of the associated benefit reached by ~15-20 g/day (a small handful, 5-6 servings/week) — chasing higher intake buys little in the data. The lever is observational-grade; rank it below the big rocks and below interventions with RCT/MR backing.
- A regional contraindication (route c): in areas where nuts are a major aflatoxin source, «increasing nut intake should only be recommended as long as aflatoxin contamination is avoided.» (Aune et al., 2016)
- Open loop: nothing here grades nut 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., 2016) Independence note for later weaves: Aune’s team also authored the parallel whole-grain dose-response MA (ref 47) using the same methods and overlapping cohort infrastructure — so any future consistency across Aune’s plant-food MAs is same-lab type-F, NOT independent type-E backing.
Sibling plant-food MA — F&V (Aune 2017) confirms the pattern, not independently [2026-08-13]
The parallel F&V dose-response MA from the same team recurs on every structural feature this page notes, which is a same-lab type-F consistency, not type-E backing (F&V explicitly cites this nut MA as ref 193) (Aune et al., 2017):
- Subtype non-uniformity, again. As nut benefit split by type (tree vs peanut; peanut butter null), F&V benefit does not distribute evenly onto favourite fruits — grapes and berries are NS (berries point >1), while leafy greens/citrus carry the tight signals. The whole-category pooling hides the seams in both foods -> Fruit and Vegetable Intake and Health.
- The plateau/edge reads the same way. Nut benefit flattened by ~15-20 g/day; F&V’s lowest risk sits at the 800 g/day sampling edge — in both, the guideline-relevant number marks study density, not a demonstrated optimum -> The Underivable Optimum.
- Both are observational, both cap at the confounding ceiling, both argue measurement error attenuates
toward the null. Convergence across the two is shared-lineage, so it does not raise
confidence:.
Refinement — the DIfE/Boeing 12-food-group series (2026-08-28)
Across the five-outcome DIfE/Boeing dose-response series, nuts carry the largest single protective point estimates in the whole matrix (all-cause mortality RR 0.76, 95% CI 0.69-0.84 per 28 g/d, MODERATE; CHD 0.67, 0.43-1.05), but the CIs are wide and the grade is low outside mortality — a big-but-uncertain lever whose benefit plateaus by ~10-20 g/d (a small handful captures most of it). (Schwingshackl et al., 2017) (Bechthold et al., 2017) Full cross-outcome placement -> Food Groups and Health Outcomes - A Dose-Response Matrix.