The de-escalator verdict. Egg consumption is not a meaningful cardiovascular lever for the general population — neither the harm the dietary-cholesterol scare predicts nor a real protection. Godos 2020, the largest dose-response MA to date (39 prospective cohort studies over 38 cohorts, ~2 million people: 85,053 CHD / 25,103 stroke / 7,536 heart-failure / 147,124 CVD cases), lands on «There is no conclusive evidence on the role of egg in CVD risk» (Godos et al., 2020). Every general-population point estimate sits within a few percent of 1.00 with confidence intervals kissing the null, and the whole body is GRADE low certainty. This is a big-worry, small-effect case — a Layer-1 finding that the egg question ranks LOW and that the attention it attracts is an anti-signal, not a magnitude.
Refinement from a second gold MA (Ma 2022, mortality endpoints). The de-escalator holds for CVD / CHD / stroke incidence, but Ma 2022 — a dose-response MA restricted to mortality — finds a small but statistically-significant all-cause-mortality slope (per 1-egg/d RR 1.06, 1.02-1.10 — a general-population continuous estimate) that Godos’s incidence-weighted read did not surface, though the categorical high-vs-low contrast is null (1.04, 0.99-1.09) and subgroup significance is restricted to women · Americans · hyperlipidemia-adjusted studies · age >60 · long follow-up. Ma lands on a more cautionary author verdict. It shares most of Godos’s cohorts, so it is not an independent witness — a type-F refinement plus a type-D headline friction, not a confidence-lifting corroboration. Details, parameter table, and pool-overlap check in Second gold MA — mortality endpoints below.
The effect estimates — near-null, outcome-specific shape [@godos2020]
Summary relative risks (SRR) vs no consumption, from the two-stage random-effects dose-response model (restricted cubic splines, 3 knots). Two «arbitrarily defined» reference doses are reported for communicability: moderate = 4 eggs/week, habitual = 1 egg/day (7/week).
| Outcome (datasets/studies) | Moderate (4/wk) | Habitual (1/day) | Shape over 0-9 eggs/wk | Pnonlin |
|---|---|---|---|---|
| CVD incidence+mortality (17/14) | 0.95 (0.90-1.00) | 0.96 (0.92-1.01) | shallow dip: nadir ~0.95 at 3-5/wk, back toward null at high intake (nonlinearity NS) | 0.167 |
| CVD incidence only (12/9) | 0.93 (0.87-0.99) | 0.94 (0.89-0.99) | dips to ~0.93, returns toward null at high intake | 0.13 |
| CHD incidence+mortality (24/16) | 0.95 (0.88-1.02) | ~0.98 (0.91-1.07) | U/J: nadir ~0.95-0.96 at 2-4/wk, rises to ~1.01 at 9/wk | 0.042 |
| Stroke incidence+mortality (22/16) | 0.98 (0.93-1.02) | 0.95 (0.88-1.01) | flat/null across the whole range | 0.842 |
| Heart failure (6/4) | 1.04 (0.92-1.19) | 1.15 (1.02-1.30) | monotone increasing: 1.01 -> 1.23 at 9/wk | 0.409 |
- CVD/CHD: a shallow protective dip at moderate intake, attenuating at high — the protection is at most a ~5% relative reduction and the CIs touch 1.00 at every dose. The abstract’s headline (4/wk SRR 0.95; up to 1 egg/day CVD-incidence 0.94; CHD decreased up to 2/wk 0.96) is this dip. Only the CHD curve has statistically-supported nonlinearity (Pnonlin 0.042); the CVD nonlinearity is NS (0.167), so the CVD dip is a near-flat band around the null rather than a firm U.
- Stroke: no association at any dose (the one outcome GRADE-rated moderate, and it is «referring to no risk») (Godos et al., 2020).
- Heart failure: the single harm arm — monotone-rising, significant from 1 egg/day (1.15) upward, «especially in men from US cohorts». The mechanism is unknown; the authors flag sex as a possible effect modifier (risk up in men, down in women) and note the signal «suggests that alternative mechanisms could be responsible» (i.e. not the dietary-cholesterol pathway) (Godos et al., 2020). Low certainty, 4 studies, no low-risk-of-bias study available for it.
The diabetic stratum — direction reverses (route-(b) effect modification, evidenced but low-certainty)
The one stratified result that changes a recommendation. In cohorts restricted to diabetic individuals the direction of risk «was substantially inverted for all outcomes», peaking for CVD incidence/mortality at 1 egg/day SRR = 1.22 (95% CI 1.08-1.39, I2 = 63%) — a ~22% increase, opposite to the general-population near-null/protective direction (Godos et al., 2020). In the stratified analyses, egg’s protective CVD signal appeared in studies not adjusting for diabetic status, «which on the contrary was reported to potentially act as effect modifier toward the opposite direction» (Godos et al., 2020).
- This is genuine subgroup evidence, not an asserted modifier — a pre-specified diabetic-status subgroup with a significant reversed estimate. But it is an observational subgroup at GRADE-low certainty (the false-positive-prone route (b) — transportability / effect-modification), so it is a credible flag, not an established causal interaction. The authors’ mechanism is a hypothesis: uncontrolled T2D shows «higher cholesterol synthesis and plasma lipid concentrations», so egg cholesterol may load an already-dysregulated lipid system (Godos et al., 2020).
- Decision-change: for someone with type 2 diabetes the near-null general-population read does not transport; habitual (daily) egg intake carries a low-certainty ~22% relative CVD-risk flag, and the conservative move is to treat eggs as a non-priority-but-watch item rather than a free food. For the non-diabetic general population, the reverse — no reason to restrict.
Second gold MA — mortality endpoints (Ma 2022) [@ma2022egg]
Ma 2022 is a dose-response SR+MA of 24 observational cohort studies / 48 reports (~11.9 M participants) restricted to mortality outcomes (all-cause · CVD · IHD · stroke), modelled with restricted cubic splines. It was acquired as the second gold witness on this page. What it adds — and does not — turns entirely on the same-quantity and pool-overlap checks below.
- All-cause mortality (an outcome Godos’s page did not hold): highest-vs-lowest RR 1.04 (0.99-1.09, NS), but the per-1-egg/d slope is a significant 1.06 (1.02-1.10), «each 1-egg/d increment was associated with a 5.6% higher risk of all-cause mortality», modelled linear (the P=0.008 is the significance of the linear trend, a different test from the CVD «linear trend» P=0.867). The continuous slope is significant overall; the categorical high-vs-low contrast is null (1.04, 0.99-1.09); subgroup significance is restricted to women, Americans, and hyperlipidemia-adjusted studies (dose-response), and in the categorical analysis to age >60 (1.098), Americans (1.115), follow-up >=15 y (1.067), hyperlipidemia-adjusted (1.088).
- CVD mortality: per 1-egg/d RR 1.00 (0.90-1.11), linear P=0.867 — null overall, but the same high-risk strata turn positive (>60 y 1.143; Americans 1.149; FU >=15 y 1.137; hyperlipidemia-adjusted 1.128). IHD mortality null (per 1-egg/d 1.01, 0.76-1.33). Stroke mortality null (0.90, 0.80-1.01, P=0.080).
- Studied range / shape (carry both): the exposure spans from ~0.07/d (lowest-category median) up to a highest-category median that is outcome-dependent — ~1.1/d (CVD) to ~1.5/d (stroke), with all-cause ~1.23/d; every curve is modelled linear over that range — no U/J and no protective lower arm on any mortality endpoint. The author verdict «it may be prudent to avoid high egg consumption» extrapolates above the studied top (~1.1-1.5/d depending on outcome), where there is little data — a studied-range caution, not a located threshold.
- The hyperlipidemia-adjustment unmasking — a DIFFERENT structure from the diabetic reversal, not the same one. Ma found the all-cause signal appeared after adjusting for hyperlipidemia: «people who have hyperlipidemia are likely to reduce their daily cholesterol consumption but still have a higher risk of death» (Ma et al., 2022) — i.e. dyslipidemic people cut eggs, which confounds the crude association toward the null; the adjustment removes that downward confounding and unmasks a general-population association (Ma’s own reading is reverse causation / confounding, «a diagnosis of hyperlipidemia is likely to confound the association» (Ma et al., 2022)). This is NOT the same object as Godos’s diabetic reversal. Godos’s is a genuine individual-level subgroup (cohorts of diabetic people, estimate reverses) — route-(b) effect modification. Ma’s is a study-level covariate marker (studies that statistically adjusted for hyperlipidemia), and Ma ran no individual-level hyperlipidemia stratification — so its inference runs the opposite way (adjustment reveals a signal for everyone, it does not locate a more-harmed stratum). They converge only loosely as a dysmetabolic theme, on shared cohorts — do not read them as one mechanism or as mutually reinforcing effect-modification evidence.
Parameter table — Godos vs Ma (the same-quantity check)
| Parameter | Godos 2020 | Ma 2022 | Same quantity? |
|---|---|---|---|
| Design | dose-response MA, 39 studies / 38 cohorts, ~2 M | dose-response MA, 24 studies / 48 reports, ~11.9 M | both obs-FFQ dose-response MA — yes (class) |
| Included cohorts | NIH-AARP, PURE, CKB, Zhong’s ARIC/MESA/FHS…, WHI, Moli-sani, EPIC-Spain, Guangzhou, NLCS, NIPPON… | substantially overlapping (NIH-AARP, PURE, CKB, Zhong 2019, WHI, Moli-sani, EPIC-Spain, Guangzhou, NLCS, NIPPON, Golestan, PHS…; Ma is mortality-only so carries some mortality-specific cohorts) | NO — large overlap, not independent |
| Exposure contrast | per-egg vs no consumption; 4/wk, 1/day | per 1-egg/d; highest (~1.1-1.5/d) vs lowest (~0.07/d) | comparable per-egg — yes |
| Primary outcome | CVD/CHD/stroke/HF incidence + mortality | all-cause / CVD / IHD / stroke mortality only | NO — Godos pools incidence+mortality; Ma is mortality-only + adds all-cause |
| CVD estimate | CVD inc+mort SRR 0.95 (0.90-1.00) at 4/wk (slight dip) | CVD mortality per 1-egg/d 1.00 (0.90-1.11) (flat) | different endpoint; both near-null, opposite tiny tilt |
| Dose-response shape | shallow U/J (CHD Pnonlin 0.042) | linear, no U/J (all outcomes) | NO — nonlinear dip vs linear |
| Harm stratum | diabetic subgroup reverses (CVD 1.22) | hyperlipidemia-adjusted / American / women / >60 / >=15 y turn positive | convergent theme (dysmetabolic/high-risk), different modifier variables |
| Certainty tool | GRADE (low; stroke moderate) | NUQUEST (67% neutral / 33% good); no GRADE | different instrument — not directly comparable |
(inferred from Godos et al., 2020; Ma et al., 2022)
Verdict: type-F (+ embedded type-D), NOT independent-E. The two MAs draw on substantially the
same cohorts — Ma’s Table 1 includes Zhong 2019, NIH-AARP, PURE, CKB, WHI, etc., which overlap Godos’s
pool (the overlap is this page’s cross-source comparison, not a statement Ma makes about a prior MA) —
so two gold MAs agree earns no [E-independent] token and no confidence lift — a RAG over one reproduces the
other’s near-null. Different bylines (Godos = Italian Grosso/Godos group; Ma = Shenzhen group) do not
rescue independence when the underlying data are shared. What Ma legitimately contributes is
type-F refinement: (i) a new outcome (all-cause mortality) with a small significant linear slope;
(ii) a mortality-only read where Godos combined incidence+mortality; (iii) a linear shape that
undercuts Godos’s shallow protective dip (see The U-shaped-arm read); (iv) a loosely convergent
dysmetabolic theme — though Ma’s hyperlipidemia signal is a study-level confounding-adjustment marker,
a different object from Godos’s individual-level diabetic subgroup (see the unmasking note above), not a
second effect-modification witness. Embedded in it is a type-D friction: Ma’s
«prudent to avoid high egg consumption» / «moderate egg intake» is more cautionary than Godos’s «no
conclusive evidence» and this page’s de-escalator read — a friction driven mainly by the all-cause
endpoint and stratum-concentrated signals, not a contradiction on general-population CVD (there the
two agree: near-null). Held as a refinement, not filed as a standalone tension page.
Why the eggs are bad scare is largely unsupported — dietary cholesterol is a weak serum-cholesterol lever
(inferred from Godos et al., 2020) The scare rests on a broken two-step syllogism: (1) eggs are high in cholesterol (true — ~180 mg per medium egg (Godos et al., 2020)), and (2) therefore eggs raise blood cholesterol and CVD. Step 2 is where it fails. This is a The Descriptive-Normative Category Error: a food’s cholesterol content (a compositional descriptor) is silently promoted to a serum-cholesterol / CVD effect (a causal claim), when the dietary->serum transmission is weak for most people.
- Only a subset are hyper-responders. Godos notes that a minority of individuals «suffer of an abnormal rise in circulating LDL cholesterol (thus described as “hyper responders”) as a result of an increase[d] fractional absorption and/or endogenous cholesterol synthesis in response to dietary cholesterol intake» (Godos et al., 2020) (the exact fraction is OCR-lost at the chunk boundary — presented qualitatively). For most people dietary cholesterol barely moves serum LDL, because absorption and endogenous synthesis homeostatically compensate — a net-effect-not-intended / whole-organism-compensation case.
- The causal lever for ASCVD is apoB particle number, reached mainly by OTHER routes. The disease is caused by cumulative apoB-particle exposure, not by cholesterol mass per se, and the modifiable inputs to serum apoB/LDL are dominated by genetics and by saturated/trans fat, not by ingested cholesterol -> LDL ApoB and Cumulative Exposure, Saturated Fat Intake and Replacement. So even granting LDL/apoB causes ASCVD, eggs are a weak input to that lever for the non-hyper-responder — which is exactly what the near-null population dose-response shows. The corrective is constructive: judge eggs on their evidenced outcome dose-response (near-null), not on their cholesterol content.
- Guidance already moved this way. The AHA advised «up to one egg per day» in 2000 and «nearly 10 years later the US Dietary Guidelines Advisory Committee eliminated cholesterol restrictions from the latest US dietary guidelines» (Godos et al., 2020) — the guidance-null agrees with the de-escalator read here.
The egg is a whole food, not a cholesterol delivery vehicle
(inferred from Godos et al., 2020) -> Is the Food Category Doing Any Work. Isolating egg cholesterol from the egg is the food-category error in miniature: the whole egg also carries choline (a TMAO precursor via gut microbiota — a candidate competing harm mechanism, human-evidenced only at the surrogate TMAO level, not outcomes), phospholipids, lutein/zeaxanthin, and high-quality protein. Godos lists these as the reason the net cardiovascular signal is a wash of opposing mechanisms rather than a clean cholesterol effect. The decision-relevant exposure is the specified food and its preparation (a boiled egg is not egg-plus-bacon — the authors flag «uncontrolled risk factors associated with egg consumption (i.e., bacon)» as a confounder in the male heart-failure signal) (Godos et al., 2020).
Certainty, design, and what would move it
- GRADE low for all outcomes but stroke (moderate, at no risk). All 39 studies scored moderate or serious risk of bias (ROBINS-I); heterogeneity was high and «rather unexplained» (Godos et al., 2020).
- Both MAs are entirely observational FFQ cohorts, and largely the SAME cohorts — so they estimate habitual self-reported egg intake, carrying the domain’s binding constraint, dietary measurement error -> Measurement Error in Dietary Assessment (Ma: «all egg consumption was self-reported via questionnaires or interviews» (Ma et al., 2022); 21/24 studies used a single baseline FFQ) — and the near-null/flat dose-response inherits it (a real gradient could be attenuated toward the null; a flat curve is weak evidence of no gradient). Ma graded with NUQUEST (67% neutral, 33% good; no study poor), not GRADE, so no cross-MA certainty comparison is clean. Reverse causation is unaddressed: the authors concede «potential reverse causation (i.e., change in dietary intake due to diagnosed medical condition or disease) … [has] been not investigated» (Godos et al., 2020).
- The authors’ own GRADE caveat (a method-not-domain point): «the GRADE system may not be the best suit for assessing evidence in nutritional epidemiology, as by definition it tends to underestimate the strength of the evidence due to the observational nature of the studies» (Godos et al., 2020) — a self-interested read (it argues their own low grades understate the truth), noted, not adopted.
The contested headline (Zhong 2019) — reconciled with the pool, not a contradiction
Zhong 2019 (JAMA) is the study behind the eggs are bad again headlines: an IPD pooled analysis of 6 US cohorts (29 615 people, median 17.5 y, max 31.3 y) reporting that dietary cholesterol and egg intake are associated with higher incident CVD and all-cause mortality. Its headline is a dietary-cholesterol claim, per +300 mg/d: «Each additional 300 mg of dietary cholesterol consumed per day was significantly associated with higher risk of incident CVD (adjusted HR, 1.17 [95% CI, 1.09-1.26]; adjusted ARD, 3.24% [95% CI, 1.39%-5.08%]) and all-cause mortality (adjusted HR, 1.18 [95% CI, 1.10-1.26]; adjusted ARD, 4.43% [95% CI, 2.51%-6.36%]).» (Zhong et al., 2019). Godos pools Zhong’s ARIC/MESA/FHS cohorts and still lands near-null — so the divergence has to be explained at the parameter level, not waved at. The parameter table does that; the answer is that the two are mostly not measuring the same thing.
Parameter table — Zhong 2019 vs the pooled MAs (the same-quantity check)
| Parameter | Zhong 2019 (JAMA) | Godos 2020 / Ma 2022 (pooled MAs) | Same quantity? |
|---|---|---|---|
| Design | IPD pool of 6 US cohorts, N=29 615, single baseline FFQ | dose-response MAs, ~2 M / ~11.9 M, obs-FFQ | both obs-FFQ, but Zhong’s cohorts are INSIDE the MA pools — not independent |
| Headline exposure | dietary cholesterol, per +300 mg/d (CVD 1.17; mortality 1.18) | eggs, per-egg / categorical | NO — nutrient (DC) vs food (egg) |
| Egg estimate (matchable) | per +0.5 egg/d: CVD 1.06 (1.03-1.10); all-cause mort 1.08 (1.04-1.11) | Godos 1/d CVD 0.96 (0.92-1.01); Ma 1/d all-cause 1.06 (1.02-1.10) | NO exact — 0.5-egg vs 1-egg vs categorical; directionally Zhong-US runs above the pooled near-null |
| Egg after adjusting for dietary cholesterol | CVD 0.99 (0.93-1.05); all-cause 1.03 (0.97-1.09) — NS | MAs do not adjust for DC | Zhong’s own egg signal vanishes once DC is held — the egg HR was a cholesterol proxy |
| Population | US-only, mean DC ~285 mg/d (US ~290 vs global 228), overnutrition context | pool non-US (CKB China, PURE ~50 countries) that are neutral/protective | NO — Zhong is a high-intake US subset of the MA pool |
| Framing | absolute ARD over max 31.3 y (+300 mg DC: CVD 3.24%, mort 4.43%); «modest» | relative RR/SRR only | Zhong adds an absolute layer; magnitudes «modest but consistent» |
| Dose-response shape | monotone, no U/J (all Pnonlin 0.19-0.83) | Godos shallow U/J (CHD 0.042); Ma linear | Zhong agrees with Ma (linear), against Godos’s protective dip |
| Confounding fragility | DC->CVD falls to 1.13 (0.97-1.31), NS after adjusting eggs+red meat; DC->mortality 1.13 (0.99-1.27) NS after eggs | high, «rather unexplained» heterogeneity | both confounding-fragile; Zhong’s own DC signal is co-adjustment-unstable |
(Godos et al., 2020; Ma et al., 2022; inferred from Zhong et al., 2019) Verdict — the type-D friction largely DISSOLVES; it is not a headline contradiction on general-population eggs. Four gaps, not one disagreement, separate Zhong from the pool:
- Different exposure. Zhong’s causal claim is about dietary cholesterol the nutrient; its egg HR attenuates to non-significance (0.99 / 1.03) once DC is adjusted (Zhong et al., 2019). The MAs test eggs the food. «Eggs are bad» is a mis-headline of a dietary-cholesterol finding.
- Different population, diluted in the pool. Zhong is US-only at high DC intake, in «the context of the current US diet, in which overnutrition and overweight/obesity are more common than malnutrition» (Zhong et al., 2019), and «generalizing our results to non-US populations requires caution» (Zhong et al., 2019). When Godos pools those 29 615 US participants into ~2 M alongside neutral/protective non-US cohorts (CKB, PURE), the US signal is washed out — the near-null pool is a between-population average, and on eggs specifically Zhong-US runs a touch higher than that average (the same dysmetabolic/high-risk direction as Godos’s diabetic reversal), not against it.
- Different framing. JAMA led with an absolute risk difference accrued over up to 31.3 years at a large +300 mg/d contrast; the authors call the relative magnitudes «modest but consistent» (Zhong et al., 2019). A «3-4%» ARD headline and a «near-null RR» are compatible descriptions of one modest effect.
- Co-adjustment fragility cuts toward the de-escalator. Zhong’s own DC->CVD signal loses significance once its carrier foods are adjusted (1.13, 0.97-1.31) (Zhong et al., 2019) — DC and eggs/meat are too collinear to identify cleanly, and both remain observational, single-FFQ, and unable to «establish causality» (Zhong et al., 2019).
NOT type-E. Zhong’s 6 cohorts sit inside the Godos/Ma pools, so this is a contained primary study vs the meta-analytic average of a set it belongs to — a type-B exposure disambiguation (nutrient vs food) plus a type-F refinement, never an independent witness. The de-escalator verdict stands for the general population; Zhong sharpens why the headline reads harsher than the pool without overturning it.
The U-shaped-arm read
The CVD/CHD moderate-intake protection is an unadjudicated protective lower arm in the sense of The U-Shaped Association Artifact: a shallow observational dip that has survived only covariate adjustment (the weak check), with no referent-correction or genetic/MR instrument isolating it. By that concept’s decision rule it is not established as a real benefit — which coincides with the de-escalator read (there is no protective claim to bank). Notably the outcome with a plausible mechanism (heart failure) shows a monotone curve and the diabetic stratum reverses — the shape is outcome- and stratum-specific, exactly the pattern the concept predicts.
Ma 2022 strengthens the “not established” read: the protective dip did not replicate on mortality endpoints (Ma et al., 2022). Where Godos found a shallow protective U/J for CVD/CHD incidence+mortality, Ma’s four mortality curves (all-cause · CVD · IHD · stroke) are each linear over the studied range with no protective lower arm — if anything the all-cause slope tilts to harm. A protective arm that appears on one MA’s incidence-weighted endpoint and vanishes on a second (overlapping) MA’s mortality endpoint is the opposite of a robust benefit -> The U-Shaped Association Artifact.
Decision relevance
(inferred from Godos et al., 2020; Ma et al., 2022) The wiki’s own layer-3 synthesis over the two gold MAs — the per-stratum reads below are reasoned from their estimates, not lifted from either source’s own recommendation.
- General population: eggs are near a non-lever for cardiovascular events — do not restrict on CVD grounds, do not count as protective. Two low-certainty caveats sit at high habitual intake (>=1 egg/day): heart failure (Godos, monotone) and a small all-cause-mortality slope (Ma, +5.6% per egg/d, linear, but stratum-concentrated and NS as a categorical high-vs-low contrast, 1.04, 0.99-1.09). The question ranks LOW; spend attention on the big rocks (SFA/apoB, adiposity, smoking, BP), not on egg count.
- Diabetic stratum (Godos — individual-level effect modification): Godos’s diabetic subgroup reverses (~22% relative CVD-risk increase at ~1 egg/day) — a genuine route-(b) subgroup (cohorts of diabetic people). Treat habitual daily egg intake as a watch-item for this stratum, not free; the general-population near-null does not transport here. (Low certainty, observational.)
- Hyperlipidemia (Ma — a study-level marker, NOT an individual stratum): Ma’s all-cause/CVD-mortality signals strengthen in studies that adjusted for hyperlipidemia — a study-level covariate, not an individual-level subgroup, and Ma reads it as reverse-causation/confounding (dyslipidemics cut eggs) that unmasks a general-population slope. This is NOT evidence that eggs are more harmful in hyperlipidemic people, and it does not combine with Godos’s diabetic subgroup into a single evidenced T2D-OR-hyperlipidemia contraindication. Do not treat it as an individual contraindication.
- Hyper-responders: a subset whose serum LDL rises materially to dietary cholesterol; if identified (e.g. by measuring the response), dietary cholesterol becomes a live lever for them specifically.
- Frame as substitution: the realistic alternative to eggs at breakfast matters (eggs vs refined-carb cereal vs processed meat) more than the eggs themselves; judge the meal, not the egg.
Open loop
Coherence, not validity (R1): this page says what the evidence supports about egg->CVD for identified
strata; the loop is open — no operation here grades the recommendation against a realized outcome.
Now two gold MAs (Godos 2020 · Ma 2022), but they share most cohorts — so confidence: low holds:
the second MA is a type-F refinement on an overlapping observational base, not an independent witness,
and the two now diverge on the headline caution (a friction, not a resolved question). Both are
FFQ-cohort MAs bound by dietary measurement error. confidence is NOT lifted by two MAs agree because
their agreement on the near-null is laundered through shared data. The contested-headline source is now
held: Zhong 2019 (JAMA) is reconciled above as a different-exposure / different-population / different-
framing study whose cohorts are already in the pool — it sharpens the de-escalator rather than
overturning it, and did not lift confidence (it is a contained subset of the MA pools, not an
independent witness). Still AWAITS a dietary-cholesterol / egg RCT-on-lipids source to harden the
mechanism leg (the dietary-cholesterol -> serum-LDL transmission), which no held source supplies.
Refinement — the DIfE/Boeing 12-food-group series (2026-08-28)
The DIfE/Boeing series reproduces the near-null CVD reading and adds an outcome-specific divergence: eggs are null for CHD (RR 1.00, 0.95-1.06) and stroke (0.99, 0.93-1.05) but carry a positive heart-failure association (RR 1.16, 1.03-1.31, per 50 g/d, MODERATE) — the within-CVD split this page’s CHD/stroke focus would otherwise miss. (Bechthold et al., 2017) The all-cause mortality cell is directionally positive but very-low grade and non-significant (RR 1.15, 0.99-1.34). (Schwingshackl et al., 2017) Full cross-outcome placement -> Food Groups and Health Outcomes - A Dose-Response Matrix.