What the analysis says
The cluster asks the net causal effect of moderate habitual egg intake (up to ~1/day) on hard CVD and all-cause-mortality endpoints, read as the typical population dose-response. Five incompatible shapes compete: H-6 - Moderate egg consumption causally lowers CVD risk in Chinese adults protective, H-21 - Moderate egg intake has no causal effect on CVD risk overall null, H-27 - Higher dietary cholesterol - egg intake causally raises CVD and mortality risk (dose-response) monotonic harm, H-44 - An intermediate egg intake of about 3 to 6 per week minimizes CVD and mortality risk with harm at both lower and higher intake U-shape, and H-49 - Some other overall egg-CVD relationship residual. The prior [0.127, 0.424, 0.229, 0.093, 0.127] already anchored on the null and deflated the two large-causal-effect members by 0.60 via the Ioannidis reference-class reductio (O-70, used_for_prior). The posterior is [0.074, 0.748, 0.012, 0.037, 0.128]: null dominant, monotonic harm essentially eliminated.
Two witnesses did most of the work, both anchored on H-21=1. CG-2 - HC-1 joint over O-44+O-43 (Harvard NHS/NHSII/HPFS, trust 0.82): a crude positive HR of 1.10 that reverses to a robust null (0.93) on full adjustment, E-value 1.43 — and it scores H-27 at only 0.3. CG-5 - HC-1 joint over O-86+O-87+O-89+O-85+O-88 (PURE/ONTARGET, trust 0.78): a precise 50-country null with no adverse lipid shift, scoring H-27 at 0.2. Monotonic harm’s one strong supporter, CG-4 - HC-1 joint over O-59+O-58 (Zhong US pool, anchored H-27=1), carries the lowest trust of the set at 0.57 and is a single witness, so it could not hold the member up against two higher-trust nulls plus the prior deflation. The 33-cohort meta CG-3 - HC-1 joint over O-47+O-48 pooled to null with region-structured heterogeneity, scored best under the residual (H-49=1), keeping H-49 at 0.128. Kadoorie CG-1 - HC-1 joint over O-9+O-10+O-11+O-12+O-13 and China-PAR CG-6 - HC-1 joint over O-108+O-109 supported H-6 and H-44 but were single Chinese cohorts outweighed by the null evidence.
What the model may not capture
The deepest doubt is that “the typical population dose-response” may not be the quantity that matters. An overall null is compatible with real harm concentrated in a susceptible subgroup: KoGES’s whole-cohort null (1.14) masks a diabetic-stratum HR of 2.81 (E-38 - O-72 × HC-1 — KoGES whole-cohort egg-CVD null, A-24 approved) — HC-4 territory. If the effect is genuinely subgroup-concentrated, the honest answer here is H-49 rather than H-21, yet the model routes PURE’s cross-region consistency into confirming H-21 and denting H-49. So H-21 at 0.748 partly borrows probability that on a subgroup reading belongs to the residual; the two are less cleanly separable than the numbers imply.
Is the true answer on the list? All five members presuppose a stable causal sign estimable from cohorts. No RCT on hard endpoints exists for eggs (RCTs reach only lipids — HC-6), so the whole cluster is priced off confounding-prone associations that O-70 says are usually bias-driven. An unlisted answer — “the effect is too small to distinguish from zero and too confounded to sign, so the question is under-determined” — sits close to H-21 but implies a weaker “at what level” verdict; its consequences are no worse than the null’s, so the comfortable arithmetic (H-27 within ~0.01 of 0) hides no large risk. But H-27’s near-elimination rests on trusting the repeated-measures adjustment of CG-2/CG-5 over the single-baseline Zhong pool; if that adjustment over-controlled mediators (BMI, T2D — A-16 flags this), a real modest harm could be masked. Monotonic harm is being shipped near-zero on a contested adjustment choice, not settled biology.
What would help
- Individual-participant hard-endpoint data stratified by diabetes/metabolic status across the pooled cohorts, to separate an overall null from a subgroup-masked harm — exists, inaccessible.
- China-PAR’s intermediate intake categories (1-<3, 6-<10/week), paywalled, which would confirm or dissolve the U keeping H-44 alive — exists, inaccessible.
- Any long-term RCT of egg intake on hard endpoints, breaking the reliance on confounded cohorts — does not exist.
- Whether CG-2’s adjustment set includes partial mediators (BMI, incident T2D), i.e. whether H-27 was over-adjusted to null — exists, unread.
Confusions and contradictions
The core irreducible conflict is directional: Chinese cohorts (Kadoorie graded inverse; China-PAR U with a protective descending limb) versus US single-baseline cohorts (Zhong’s cholesterol-driven positive), with the global meta pooling to null at high, region-structured heterogeneity (I2=62.3%). The analysis reads this as null-with-heterogeneity and pushes the sign-flip into H-49 and HC-2 — defensible, but it means 0.748 on H-21 is a global average over a pattern that does not pool cleanly, not a finding that eggs are inert everywhere. Whether that averaging is legitimate is exactly HC-2’s question, so this headline should be read as conditional on HC-2 not being genuine modification.
External consensus. The mainstream (2020 meta-analyses; the 2019-2020 guideline reversals dropping the 300 mg/day cholesterol cap) lands close to the posterior: up to ~1 egg/day broadly neutral for CVD, with lingering caution for diabetics. The posterior’s near-elimination of monotonic harm is slightly stronger than guideline bodies, which still hedge on dietary cholesterol — a small divergence on the side of the analysis weighting the CG-2 reversal more decisively than consensus has.