The fish-seafood nucleus. Fish is the rare exposure a guidance body priced on both sides with
numbers — a benefit (EPA/DHA -> neurodevelopment and coronary mortality) weighed against a risk
(methylmercury -> fetal neurotoxicity) — and found the balance flips by stratum and by species.
EFSA 2014 is the benefit arm of a two-part benefit-risk analysis (the methylmercury toxicity arm is
the separate CONTAM 2012 opinion, cited here for the tolerable intake).
(inferred from European Food Safety Authority, n.d.)
The bottom line — where the benefit sits
- About 1-2 servings/week, and up to 3-4/week in pregnancy, is associated with better outcomes vs no seafood; the benefit is already present at ~1 serving/week, and no additional benefit is expected above ~4-5 servings/week — on both neurodevelopment and CHD mortality. (European Food Safety Authority, n.d.)
- This is a plateau, not a monotone curve — a located knee for a food exposure (contrast the monotone fibre and sodium curves -> Dietary Fibre and Health, Sodium Intake and Blood Pressure). It is outcome-robust: the same «no benefit might be expected at higher intakes» holds for the child-IQ and the adult-CHD arms.
- A serving is ~100-150 g in the European guidelines surveyed; most national FBDGs recommend «(a minimum of) two servings of about 150 g per week». (European Food Safety Authority, n.d.)
The benefit side — three outcomes, all observational
Adult cardiovascular: CHD mortality specifically
The evidence is outcome-narrow: strong for CHD mortality, inconsistent or null for stroke, atrial fibrillation, total CVD events. «there is strong evidence for an effect of n-3 LCPUFA from seafood on the reduction of CHD mortality», while arrhythmia/stroke/total-CVD signals are not supported by the RCTs. (European Food Safety Authority, n.d.)
Four meta-analyses of cohorts, magnitudes (all vs little/no seafood):
| Meta-analysis | Dose-response magnitude |
|---|---|
| He 2004 | per 20 g/day, pooled RR 0.93 (0.87-0.99) = a 7% lower CHD-mortality risk; benefit from 1 serving/week |
| Whelton 2004 | any vs little/no seafood RR 0.83 (0.76-0.90); no benefit >4 servings/week |
| König 2005 | ~17% lower at 0.5 servings/week, +3.9% per added serving (5.5% per 20 g/day) |
| Zheng 2012 | non-linear; ~6% per added 15 g/day up to 4 servings/week; benefit may be lost at >=5/week |
(European Food Safety Authority, n.d.)
- Attributed to EPA+DHA. Per 100 mg/day EPA+DHA, CHD mortality fell 14.6% (8-21%) up to 250 mg/day, overall 36% (20-50%) vs none (Mozaffarian & Rimm 2006) — and «RCTs conducted with supplemental, higher doses of n-3 LCPUFAs did not show higher benefits than cohort studies». (European Food Safety Authority, n.d.) EFSA set an Adequate Intake of 250 mg/day EPA+DHA for CV health. (European Food Safety Authority, n.d.)
Fetal neurodevelopment (maternal pregnancy consumption)
Large cohorts (ALSPAC ~12k, Danish National Birth Cohort ~26k) found maternal fish consumption associated with higher child IQ/developmental scores. Modelled by FDA/FAO-WHO: «an average of 4.0 IQ points could be gained from an intake of 100 mg DHA per day, with a maximum attainable IQ gain of 5.8 points». (European Food Safety Authority, n.d.)
But this is repletion of a deficiency, not enhancement — three facts converge:
- DHA supplementation RCTs in pregnancy were NULL on child neurodevelopment: «there is no evidence for an effect of n-3 LCPUFA supplementation (mostly DHA) during pregnancy on any functional outcome of children’s neurodevelopment». (European Food Safety Authority, n.d.)
- The association was confined to the low-DHA tail: maternal DHA predicted IQ «within the lowest quartile of maternal DHA» but «no significant associations… within the highest quartile». (European Food Safety Authority, n.d.)
- EFSA’s own conclusion: benefit «may depend on the maternal status» and «No effect of these nutrients… is expected when maternal requirements are met.» (European Food Safety Authority, n.d.)
So the neurodevelopment lever is a deficiency-repletion lever (DHA and iodine), and it does nothing once maternal status is adequate -> Deficiency Repletion vs Enhancement. The observational food signal survives while the isolate RCT is null — a design story, not proof the food beats the nutrient.
Cognitive decline / dementia (older adults) — Wei 2023
A gold SR+MA (Wei 2023; 48 cohorts, 31 pooled, 103,651 participants) adds a third observational benefit outcome, and it repeats this page’s central lesson — the food/dietary signal is protective while the isolate-supplement RCTs are null. Two exposure-form facts carry the decision, and they dissociate:
- Dietary DHA is the strongest signal: RR 0.82 (0.72, 0.93; Level H), 13 studies — Wei’s own highest-credibility dietary estimate, reported as a 27% lower dementia risk / 24% lower AD risk. Dietary total omega-3 is weaker and borderline (RR 0.91, 0.82-1.00; Level M); dietary EPA and ALA are null. (Wei et al., 2023) Fish is the dominant food source of DHA, so this lands on the fish nucleus — but the exposure Wei measured is dietary DHA by FFQ, not fish servings, so read it as the n-3 mechanism, not a fish-specific dose.
- The biomarker compartments split, and the split is diagnostic. Erythrocyte DHA (the long-term, ~60-90 d intake marker) is protective — RR 0.94 (0.89, 0.98; Level H) — tracking the dietary signal. But plasma DHA (the short-term, ~7-14 d marker) is null and publication-bias-driven: «No significant association was found between higher levels of plasma DHA and a lower risk of cognitive decline (RR: 0.88, 95% CI: 0.76, 1.03; I2 ¼ 63.6%, Level L), with publication bias (Egger’s P ¼ 0.007, corrected RR: 0.99, 95% CI: 0.85, 1.14; I2 ¼ 69%)» (Wei et al., 2023) — the bias correction moves it to a flat null.
The diet-vs-supplement / marker-vs-lever pattern, stated by Wei itself. Wei’s ADNI arm found supplement use protective (long-term users HR 0.36) but blood omega-3 levels null in the same cohort, and concedes «residual confounding from unmeasured confounders (such as dietary intake and physical activity) remains an issue» (Wei et al., 2023) — the healthy-user tell. Wei’s intro concedes the design split outright: «In contrast, randomized clinical trials have shown limited efficacy of omega-3 fatty acid sup-plementation in reducing cognitive decline and probable AD» (Wei et al., 2023). So the observational dietary-DHA benefit does NOT license an omega-3 supplement recommendation for cognition — the held RCTs (MAPT 800 mg DHA + 225 mg EPA/d, cognition-null; VITAL 1 g/d marine n-3) find no cognitive/CV benefit -> Multidomain Lifestyle Intervention and Cognitive Decline, Vitamin and Mineral Supplements for Disease Prevention.
Parameter table — Wei (dietary DHA, observational) vs MAPT (DHA+EPA supplement, RCT):
| Parameter | Wei 2023 dietary DHA | MAPT / Andrieu (RCT) | Same quantity? |
|---|---|---|---|
| Exposure | habitual dietary DHA (FFQ, years) | 800 mg DHA + 225 mg EPA/d capsule, 3 y | NO — food-borne vs isolated supplement |
| Design | pooled prospective cohorts (confounded) | double-blind RCT (unconfounded) | NO |
| Endpoint | incident cognitive decline / dementia / AD | change on a cognitive composite over 3 y | NO — incidence vs surrogate slope |
| Result | RR 0.82 (0.72, 0.93) protective | null | direction differs |
Verdict: the two are not the same quantity (different exposure form, design, and endpoint; not-joined check (ii)), so this is a type-F refinement / diet-vs-supplement distinction, NOT a filed tension — the composite lesson (type A) is that the observed dietary-DHA benefit does not transport to supplementation, and the marker that would carry the supplement mechanism (plasma DHA) is itself null. (inferred from Wei et al., 2023)
The 1.0 g/d threshold is an authors’ proposal at the edge of the data, not a located knee. Wei «propose[s] that 1.0 g/d may be the threshold» dosage (Wei et al., 2023), but the same sentence reports the total-omega-3 dose-response was «not significantly linear», and the number sits at the upper edge of the observed dietary range. Per the dose-response discipline, treat it as the edge of the evidence, not a curve feature -> The Underivable Optimum; do not state it bare.
The key structural move — the benefit is measured NET of the mercury
The load-bearing sentence: the observed associations «refer to seafood per se and include beneficial and adverse effects of nutrients and non-nutrients (i.e. including contaminants such as methylmercury)». (European Food Safety Authority, n.d.)
A cohort eating real fish already ate the mercury inside it. So the positive net association is the benefit-minus-risk balance, at the species and intake European populations actually consume — the weighing is not a modelled subtraction of two separate estimates but a directly observed net effect. This is why the finding can be stated as a serving recommendation rather than an exchange rate: at typical European species/intakes, the whole-food net is already positive.
(inferred from European Food Safety Authority, n.d.)
The risk side — methylmercury, and why species is the decision variable
- Tolerable Weekly Intake = 1.3 µg/kg body weight/week (EFSA CONTAM 2012), «expressed as mercury… based on neurodevelopmental endpoints», point of departure 11.5 mg/kg maternal hair. (European Food Safety Authority, n.d.)
- Methylmercury biomagnifies up the food chain, so apex predators carry the most. The pregnancy FBDG guidance is a species rule, not an intake rule: «prefer seafood low in pollutants (e.g. trout, ocean perch, cisco, sardine, white halibut, salmon, mackerel, herring, sprats, anchovies, carp and prawns); not to eat swordfish, dogfish, marlin, shark, and ray; and to consume at most one serving of fresh tuna or pike per week». (European Food Safety Authority, n.d.)
- The harm signal itself comes from extreme exposure. The Faroese cohort (the classic methylmercury-neurotoxicity source) eats pilot whale, and the Seychelles cohort eats far more fish than any European population — EFSA discounts both as «much higher than current intakes (and current recommendations) in the majority of European countries». (European Food Safety Authority, n.d.)
Benefit and risk are largely DECOUPLED across species — that is the escape. The species richest in EPA/DHA (herring, salmon, mackerel, sardine, trout: n-3 ~600-2500 mg/100 g) are mostly low-mercury; the high-mercury apex predators (swordfish, shark, marlin) carry the risk without an offsetting n-3 advantage. Tuna is the one species that sits on both axes (n-3-rich and can be high-mercury) — which is exactly the species the guidance singles out to limit. (European Food Safety Authority, n.d.)
(inferred from European Food Safety Authority, n.d.)
The stratum flip — the crown-jewel finding
The balance is not the same for everyone, and the flip is stratum × species, not a blanket verdict:
- General adults: benefit dominates (CHD mortality); methylmercury CV harm is «inconclusive». Eat 1-2 servings/week of any commonly-consumed species.
- The mercury-sensitive stratum is the developing fetus/young child — hence pregnancy and women of childbearing age. Here the risk can bind, but the FAO/WHO weighing lands against the intuitive prediction: for women of childbearing age, in most circumstances the «risk of suboptimal neurodevelopment in offspring of women consuming seafood is lower than in the offspring of women not eating seafood». (European Food Safety Authority, n.d.)
So the flip is not “pregnant women should avoid fish” — it is “the species constraint binds in the sensitive stratum.” Low-mercury fish still nets positive even for the fetus (the lost DHA/iodine of not eating fish outweighs the mercury of low-mercury fish); the balance only tips negative for apex-predator species / high intake. The decision-change is a species swap, not abstention — and abstaining is itself the harmful option for the outcome the sensitive stratum most cares about.
(inferred from European Food Safety Authority, n.d.)
Where this sits in the Layer-1 ranking
Fish is a small-to-moderate lever, not a big rock:
- The adult CHD-mortality benefit is observational (loop open), and the isolated-nutrient RCTs split by dose: near-null in the general replete population at ~1 g/day (VITAL, and the low-dose stratum across 14 pooled RCTs) and positive only above 1 g/day, concentrated in high-risk statin-treated patients (REDUCE-IT) — see the Lombardi 2020 dose-response section below -> Vitamin and Mineral Supplements for Disease Prevention, Is the Food Category Doing Any Work. So a replete general-population adult should not expect the cohort magnitude, nor a benefit from an OTC 1 g/day capsule.
- The neurodevelopment benefit is repletion only — it matters for a mother with low DHA/iodine status, and is a flat zero once requirements are met.
(inferred from European Food Safety Authority, n.d.)
The supplement dose-response — Lombardi 2020 NMA firms the VITAL-vs-REDUCE-IT split (type-F, 2026-09-01)
The held omega-3 supplement evidence was two single trials pointing opposite ways — VITAL (1 g/d EPA+DHA, CV-null) and REDUCE-IT (4 g/d EPA, CV-positive) — already resolved on Vitamin and Mineral Supplements for Disease Prevention as a distinction, not a tension (they test different compounds at a 4x dose gap in different strata). Lombardi 2020 is the first pooled RCT evidence woven directly onto this nucleus: a pairwise + network meta-analysis of 14 RCTs (125,763 patients, median 4.6 y follow-up, 78.6% statin-treated) that generalizes the single-trial split into a dose-stratified finding across the whole trial set. (Lombardi et al., 2020)
The dose finding. Splitting trials at a prespecified 1 g/day cutoff (LD ≤ 1 g, HD > 1 g), high-dose vs control lowered the ischemic endpoints — major vascular events IRR 0.78 (0.71-0.85), MI 0.71 (0.62-0.82), coronary revascularization 0.74 (0.66-0.83), unstable angina 0.73 (0.62-0.86), cardiac death 0.79 (0.65-0.96) — and HD beat LD on the same endpoints (e.g. MI HD vs LD 0.79, 0.67-0.92). Low-dose vs control was itself weakly protective on cardiac death (0.92, 0.87-0.98) and MI (0.91, 0.84-0.98) but «no benefit of LD treatment compared to HD in any of the efficacy endpoints». Total death, sudden cardiac death, and stroke were null even at high dose (total death 0.95, 0.85-1.06). (Lombardi et al., 2020)
This firms VITAL from a lone null into the low-dose pattern: pooling every trial regardless of dose (the all-dose pairwise analysis) washes the ischemic signal toward null — total death 0.98 (0.94-1.02), revascularization 0.93 (0.83-1.02), stroke 1.02 both NS — because the low-dose trials dilute the high-dose signal. (Lombardi et al., 2020) The benefit lives in the > 1 g/day trials.
But dose is confounded with form and stratum — the NMA cannot separate them, and says so. The high-dose arm is dominated by EPA-only trials (REDUCE-IT 4 g, JELIS 1.8 g) in higher-risk populations, so «it remains unknown whether it was the dose or formulation that drove the observed CV benefit», and the closed-loop needed to compare doses directly does not exist: «there is no direct comparison be-tween HD and LD O3FA», on «an arbitrary cut-off of 1 g/day». (Lombardi et al., 2020) So Lombardi refines the supplements-page distinction (it adds that the dose axis tracks benefit across many trials, not just the one VITAL-vs-REDUCE-IT contrast) without upgrading it to a causal dose law — dose, EPA-only form, and high-risk stratum remain entangled. It is not a filed tension (the split was never joined at the same quantity) and not independent type-E (it re-pools the same trials the wiki already holds).
Net-effect and comparator caveats. The high-dose benefit is bought with increased bleeding (1.49, 1.2-1.84 vs control) and atrial fibrillation (1.35, 1.1-1.66 vs control) — a net-effect tradeoff, not a free lever. And a comparator threat runs across the whole literature: the n-3-vs-control difference was «smaller, and not sig-nificant in those studies using an olive oil-based control», suggesting «olive oil is not an inert control» — the mirror image of the REDUCE-IT mineral-oil controversy, here biasing toward the null. (Lombardi et al., 2020) The AF harm is now held as a page in its own right: the dedicated SR+MA (Gencer 2021, 7 trials, 81,210 pts) puts overall incident-AF HR at 1.25 (1.07-1.46) and a dose-gradient (>1 g/d 1.49 vs <=1 g/d 1.12, P-interaction <0.001) -> Omega-3 Supplementation and Atrial Fibrillation. (inferred from Gencer et al., 2021)
Decision-change. For a general, replete adult, an OTC ~1 g/day fish-oil capsule is not expected to reduce CV events — the low-dose stratum is near-null across the pooled RCT evidence, firming VITAL beyond a single trial. A CV benefit appears only above 1 g/day and concentrates in high-risk (statin-treated, high-TG, established-CVD/diabetic) strata, at the cost of bleeding and AFib, and even there cannot be cleanly credited to dose rather than the EPA-only formulation. This does not touch the fish-as-food recommendation (1-2 servings/week, observational), which turns on the whole food matrix, not a high-dose isolate. (Bhatt et al., 2019; inferred from Lombardi et al., 2020; Manson et al., 2019)
Is the food category doing any work?
Fish fails the category test badly. n-3 content ranges ~200 mg/100 g (cod, whiting) to ~2500 mg/100 g (herring, tuna) — an order of magnitude — and mercury ranges from negligible (sardine) to guideline-limiting (shark, swordfish). (European Food Safety Authority, n.d.) The two decision variables are (a) oily vs white (the n-3 benefit) and (b) trophic level / low vs high mercury (the risk) — and eat fish twice a week names neither. EFSA itself avoids the fatty/oily/lean/white labels because their use has been «inconsistent and their meaning ill defined». (European Food Safety Authority, n.d.) Replace the category with the species properties in any decision -> Is the Food Category Doing Any Work (Test 3).
The pooled magnitudes EFSA could not supply — Jayedi 2020 umbrella (type-F refinement)
EFSA’s benefit arm was not quantified with certainty — its CHD/IQ numbers were FDA/FAO-WHO models it merely summarized, and it called its own quantitative synthesis «generally hampered by the heterogeneity of the studies» (European Food Safety Authority, n.d.). The Jayedi & Shab-Bidar 2020 umbrella review (34 meta-analyses of cohorts, 48 SRRs, 298 primary cohorts, graded by NutriGrade) is the first source in the vault to supply pooled per-100-g/d magnitudes with a per-outcome certainty grade. It does not overturn EFSA — it firms and bounds it. (Jayedi & Shab-Bidar, 2020)
The pooled benefit, by NutriGrade certainty (all per 100 g/d unless noted):
| Certainty | Outcome | SRR (95% CI) | Reading (relative, per dose) |
|---|---|---|---|
| Moderate | CVD mortality | 0.75 (0.65, 0.87) | ~25% lower per +100 g/d |
| Moderate | Myocardial infarction | 0.75 (0.65, 0.93) | ~25% lower per +100 g/d (highest heterogeneity, I2>75%) |
| Moderate | Heart failure | 0.80 (0.67, 0.95) | ~20% lower per +100 g/d |
| Moderate | Stroke | 0.86 (0.75, 0.99) | ~14% lower per +100 g/d |
| Moderate | CHD | 0.88 (0.79, 0.99) | ~12% lower per +100 g/d |
| Moderate | All-cause mortality | 0.92 (0.87, 0.97) | ~8% lower per +100 g/d; 38 studies, 153,998 cases |
| Moderate | Depression | 0.88 (0.79, 0.98) high vs low | new outcome (not on EFSA’s arm) |
| Moderate | Liver cancer | 0.65 (0.48, 0.87) | the one cancer that reaches moderate |
| Low | CHD mortality | 0.65 (0.48, 0.87) | largest effect, but low certainty + nonlinear |
The RRs are relative, per +100 g/d (no baseline risk is given in the umbrella, so a true absolute risk reduction cannot be computed here — the streetlight caveat -> Baseline Risk and the Relative-Absolute Split).
The certainty ceiling is the headline finding, not a footnote: «Overall, there was no high-quality evidence for the relation of fish consumption and the risk of chronic disease» — 8/48 SRRs moderate (17%), the rest low/very low. (Jayedi & Shab-Bidar, 2020) The ceiling is observational (no fish RCTs; 91% of the pooled cohorts adjusted age but only 33% adjusted family history), so moderate is as high as the fish->mortality evidence goes.
Nulls and harms that sharpen the food-category point:
- Total fish -> T2D, hypertension, atrial fibrillation are all NULL. But «higher intake of fatty fish, but not lean fish, was associated with a lower risk of T2D» (0.89 vs 0.96) — the oily/white split doing decision work again -> Is the Food Category Doing Any Work. (Jayedi & Shab-Bidar, 2020)
- Preparation flips the sign: «higher intake of fried fish was associated with a higher risk» of heart failure (RR 1.40, 1.22-1.61) while nonfried fish is protective (0.69). Fried fish is a different exposure than fish. (Jayedi & Shab-Bidar, 2020)
- Two low-certainty cancer harm signals (gastric 1.16, myeloid leukemia 1.60), few studies each.
- Mercury corroboration: cohort evidence found no association between toenail mercury (the best long-term intake marker) and CVD, and the 2018 AHA statement holds that «the benefits of 1-2 servings/wk… outweigh the potential risks associated with mercury content of fish» — consistent with EFSA’s net-benefit weighing above. (Jayedi & Shab-Bidar, 2020)
Cross-source parameter table (EFSA 2014 vs Jayedi 2020) — the BLOCKING commensurability check
| Parameter | EFSA 2014 (quoted) | Jayedi 2020 (quoted) | Same quantity? |
|---|---|---|---|
| Quantifiability of the benefit | «generally hampered by the heterogeneity of the studies» — no endorsed effect size (chunk 03) | 48 pooled SRRs, NutriGrade-graded; «no high-quality evidence» but 8 moderate (chunk 01) | Yes — Jayedi supplies exactly the graded pooled estimate EFSA said it could not; both converge on an imperfect ceiling (F-refinement) |
| CHD-mortality dose-response | He 2004 per 20 g/day pooled RR 0.93, cohorts (chunk 03) | CHD mortality 0.65 (0.48, 0.87) per 100 g/d, 17 studies (chunk 02) | Yes, same outcome+design; increments differ (0.93/20 g ≈ 0.70/100 g vs 0.65/100 g -> directionally consistent). NOT independent — Jayedi’s pool includes He 2004 |
| Dose-response SHAPE, CHD mortality | Zheng 2012 non-linear, benefit may be lost at high intake (chunk 03) | «nonlinear dose-response relations for CHD mortality» (chunk 02) | Yes — both nonlinear for CHD mortality; consistent |
| Dose-response SHAPE, all-cause / total CVD mortality | plateau framing (no benefit above ~4-5 servings/wk) — but EFSA’s plateau evidence is CHD-mortality-specific | «inverse linear associations for all-cause and CVD mortality»; «no evidence of a U- or J-shaped association» (chunk 02) | No — different outcomes. The plateau EFSA located is CHD-mortality-specific; for all-cause + total CVD mortality Jayedi finds linear-inverse. -> a refinement (shape is outcome-specific), not a contradiction |
Verdict: type-F refinement + shared-data corroboration, not independent type-E — Jayedi re-pools the same cohort literature EFSA’s four CHD MAs drew on and rests on the identical EPA/DHA mechanism story, so its agreement raises detail and certainty, not independence. (European Food Safety Authority, n.d.; inferred from Jayedi & Shab-Bidar, 2020)
The dose-response shape reconciliation (decision-relevant): EFSA reported a plateau (benefit gone above ~4-5 servings/week); Jayedi reports linear-inverse for all-cause and CVD mortality with no U/J shape. These are not in conflict — EFSA’s plateau evidence was CHD-mortality-specific (where Jayedi also finds nonlinearity), while for the broader mortality outcomes Jayedi finds every increment still pays. The shape is outcome-specific, matching the corpus’s gate-6 finding that a curve can be monotone on one outcome and flat on another -> The U-Shaped Association Artifact. The conservative decision default (keep the ~1-2 servings/week target; more buys little on CHD mortality but may still help all-cause) is unchanged.
Limits
- EFSA could not quantify the benefit with certainty; Jayedi 2020 now can — but only to moderate. EFSA’s quantitative analyses were «generally hampered by the heterogeneity of the studies» — different intake tools, outcome tools, and confounder adjustment. (European Food Safety Authority, n.d.) The IQ and CHD magnitudes in the sections above are FDA/FAO-WHO models EFSA summarizes, not effect sizes EFSA endorses. Jayedi’s pooled SRRs supply endorsed magnitudes, but at a moderate certainty ceiling and still observational.
- All the human outcome evidence is observational (cohorts); the DHA-supplementation RCTs were null. Confounding by overall diet quality and socioeconomic position is uncontrolled (the observed-healthy-population problem).
- Intake is FFQ-measured — the binding constraint -> Measurement Error in Dietary Assessment.
- The risk arm is borrowed. This opinion is the benefit side; the methylmercury toxicity assessment and the TWI derivation are the separate CONTAM 2012 opinion.
- Three cohort sources, still not independent on the food signal. EFSA (benefit-risk guidance) + Jayedi 2020 (cohort umbrella, CV/mortality) + Wei 2023 (cohort SR+MA, cognition) — but all three share the underlying observational cohort literature and the EPA/DHA mechanism, so confidence stays medium, not high (the observational ceiling binds all three). Wei adds the cognitive-decline / dementia outcome and a partial independence on design reading (it explicitly contrasts the null supplement RCTs), but its own effect estimates are observational and FFQ-based like the others.
- Lombardi 2020 now brings pooled-RCT evidence directly onto the page — but on the supplement question, not the food. It is design-independent of the cohort literature above (14 RCTs, not cohorts), which partially cashes the previously-owed omega-3 RCT woven directly here gap; but it addresses supplemental O3FA dose (LD vs HD vs control), carries the olive-oil comparator caveat, and cannot separate dose from EPA-only form or high-risk stratum. So it firms the supplement leg (OTC 1 g/day near-null; benefit only > 1 g/day in high-risk strata) without lifting the food signal’s observational ceiling. Still owed: an RCT of food-form fish (structurally near-impossible), the mercury-toxicity CONTAM 2012 opinion, and ASCEND (not held).
(inferred from European Food Safety Authority, n.d.)
Self-critique [run 2026-08-23, before commit — cognition section added]
- Not overclaimed. The Wei cognition outcome is stated observational throughout; dietary DHA RR 0.82 is the strongest dietary signal in Wei’s own paper (source-scoped, not a wiki-holdings superlative), and the section’s load-bearing conclusion is the negative one — the observational dietary signal does NOT license an omega-3 supplement recommendation, which the held RCTs (MAPT, VITAL) refute. No “proven supplement effect” reading survives.
- Not a fake tension. Wei-vs-MAPT is filed as an F-refinement / diet-vs-supplement distinction via a parameter table whose same-quantity? column is NO on every row (food vs isolate; cohort vs RCT; incidence vs surrogate slope) — the not-joined guard (ii). The exposure-form dissociation (dietary + erythrocyte DHA protective; plasma DHA null after bias correction) is the type-A composite, marked as this page’s own reading (INFERRED), not a source finding.
- Not laundered-E. Wei is NOT claimed independent of EFSA/Jayedi — all three rest on the same
observational cohort literature and the EPA/DHA mechanism, so confidence stays
mediumand the Limits section says so. The still-owed omega-3 RCT woven directly here remains open. - Coherence, not validity (R1): the loop is open — no operation grades the fish->cognition association against a realized dementia outcome.
Refinement — the DIfE/Boeing 12-food-group series (2026-08-28)
Fish is the series’ clearest outcome-specific food: inverse for all-cause mortality (0.93, 0.88-0.98 per 100 g/d) and for all three CVD subtypes (CHD 0.88, stroke 0.86, heart failure 0.80, all MODERATE), yet null for T2D (1.09, 0.93-1.28 — even positive in American cohorts) and null/slightly positive for hypertension (1.07, 0.98-1.16). Same food, opposite glycaemic/pressor signal — a caution against reading a mortality/CVD benefit as a whole-metabolism benefit. (Schwingshackl et al., 2017) (Bechthold et al., 2017) (Schwingshackl, Hoffmann, et al., 2017) (Schwingshackl, Schwedhelm, Hoffmann, Knüppel, et al., 2017) Full grid -> Food Groups and Health Outcomes - A Dose-Response Matrix.
Self-critique [run 2026-09-01, before commit — Lombardi 2020 dose-response NMA added]
- Not overclaimed. The dose finding is stated as an association of HD (>1 g/day) with the ischemic reductions, never as a proven dose law: the section says twice that dose is confounded with EPA-only form and high-risk stratum and that the NMA has no closed loop and an arbitrary cutoff — the causal reading is explicitly withheld. The decision-change is the negative one (OTC 1 g/day near-null), which the pooled low-dose stratum and VITAL both support. No fish oil prevents heart disease reading survives; the food-vs-supplement firewall is stated (Lombardi does not touch the food arm).
- Not a fake tension. VITAL-vs-REDUCE-IT is carried as a distinction (already resolved on the
supplements page via a same-quantity-NO parameter table), and Lombardi is filed as a type-F
refinement of it, not a filed
[[tension]]— the two trials were never joined at the same quantity (dose, form, stratum, composite all differ), so the not-joined guard (ii) holds. - Not laundered-E. Lombardi is explicitly marked not independent — it re-pools VITAL,
REDUCE-IT, JELIS, GISSI, ASCEND et al. that the wiki already holds, so it raises detail/certainty on
the supplement leg, not independence; confidence stays
medium. Its one genuinely new axis is design (pooled RCT vs the page’s cohort sources), noted in Limits as partial and supplement-scoped. - Coherence, not validity (R1): the loop is open — no operation here grades the omega-3-dose -> CV-event association against a realized outcome; the olive-oil/mineral-oil comparator threat is flagged as an unresolved bias in the underlying trials.