Uses the pooled ~29,700-participant Lifetime Risk Pooling Project (6 US cohorts) to model isocaloric substitution of one serving of one animal-protein food for another, estimating the resulting change in incident CVD and all-cause mortality risk. The same substitution matrix compares eggs against every other protein food as reference, so the sign and size of the “egg effect” is shown, within one paper, to depend on the comparator.
Methodology
Pooled analysis of 29,682 US adults from six prospective cohorts (ARIC, CARDIA, CHS, Framingham Heart + Offspring, MESA; baseline visits 1985-2002), individual-participant-data harmonized. 6,963 incident CVD events and 8,875 deaths over median 19.0 y (max 31.3 y) follow-up to 31 Aug 2016. Excluded: baseline CVD, extreme energy (<500 or >6000 kcal/day), missing data. Baseline diet only, by validated diet history/FFQ harmonized across cohorts. One serving = one whole egg, 85 g other meats, 28 g nuts, half cup legumes, one slice whole-grain bread. Substitution model Y = aF1 + bF2 + covariates with all eight protein foods entered together and total constrained, so the substitution effect is the coefficient contrast exp(b-a); a quadratic term added for processed meat (non-monotonic). Cohort-stratified cause-specific hazard models (incident CVD, competing risks) and proportional hazards (mortality). Covariates: age, sex, race/ethnicity, education, total energy, smoking status + pack-years, physical-activity z-score, alcohol, hormone therapy, modified aHEI-2010 (protein components removed); blood pressure, cholesterol, glucose, BMI excluded as mediators. 30-year absolute risk differences via bootstrap (500 samples). No subgroup analyses (prior work found no consistent interactions).
Results
O-39 - Replacing one serving per day of eggs with nuts, legumes, whole grains or fish lowers incident CVD by 15-21 percent
Cohort-stratified cause-specific hazard models; all eight protein foods plus demographic/lifestyle covariates in one model (total food amount constrained). 1-serving/week HRs for eggs→fish/nuts/legumes/whole grains all ~0.97-0.99. Poultry as comparator: CVD 0.92 (0.84-1.01), non-significant.
Link to original
O-40 - Replacing one serving per day of eggs with nuts, fish, poultry or whole grains lowers all-cause mortality by 11-22 percent
Cohort-stratified proportional hazards models, same covariate set. The mortality benefit of swapping eggs for a healthier source is broadly parallel to the incident-CVD benefit.
Link to original
O-41 - Sign and size of the egg-removal effect depend on the comparator food; eggs rank near processed meat
The authors rank nuts and whole grains as healthiest and “eggs and processed meat the unhealthiest” protein sources; fish the “single not unhealthy” animal source. That eggs→unprocessed red meat lowers risk shows eggs sit at the bottom of the ranking together with processed meat.
Link to original
O-42 - Only 8 percent of participants ate at least one egg serving per day in the pooled US cohorts
Only 0.2% consumed >=1 serving/day of all four less-healthy foods (eggs, processed meat, unprocessed red meat, poultry) simultaneously. The steepest substitution benefits (down to HR 0.46) are for these small high-intake subsets with “greater room for reducing intake.”
Link to original
Interpretation
H-17 - Eggs are a relatively unhealthy protein source; benefit comes from replacing eggs with plant or fish protein
The authors’ overall ranking: nuts and whole grains healthiest; eggs and processed meat unhealthiest; fish the single not-unhealthy animal source.
Link to original
A-13 - Isocaloric substitution algebra makes the egg effect the difference of two food coefficients, so no unconditional egg effect is identified
In the substitution model Y = aF1 + bF2 + covariates, with the total amount of protein foods held fixed, decreasing F1 by one unit while increasing F2 by one unit changes the linear predictor by (b - a). The reported substitution hazard ratio is therefore exp(b - a): the contrast between the two foods’ coefficients, not an absolute coefficient for either food.
Consequently the estimated “effect of removing eggs” equals (beta_comparator - beta_eggs) and takes a different value for every comparator - near-null when the comparator is itself unhealthy (processed meat, whose coefficient is close to eggs’) and strongly protective when the comparator is healthy (nuts or whole grains, with much lower coefficients). The design never contrasts eggs against a fixed baseline (e.g. total calories or “no egg”), so it cannot estimate an unconditional egg coefficient at all. The authors make the same point concretely: substituting eggs with unprocessed red meat lowers risk yet is not a healthy choice, because red meat is itself harmful.
This is an identity of the linear model, valid regardless of confounding or measurement error, and it converts the ill-posed question “are eggs good or bad” into the well-posed “eggs relative to what.”
Validity (step 6)
status: approved, reason_if_not_false: checked.
Traced the algebra directly. In the constrained model with total protein-food amount held fixed, moving one unit from food F1 to F2 changes the linear predictor by (b − a), so the substitution HR is exp(b − a) — a contrast of two coefficients, not an absolute coefficient for either food. This is a genuine mathematical identity of the compositional/constrained linear model, so the “effect of removing eggs” is comparator-relative by construction, and the reported effect legitimately changes sign and size with the comparator (near-null vs processed meat, protective vs nuts/whole grains).
The one place to check for over-reach is the leap to “this design identifies no unconditional egg effect.” The statement is scoped to this design, and that scope is correct: a food-substitution model with total held fixed contains only among-food contrasts (all coefficients are relative to the omitted reference food), so no comparator-independent egg coefficient is estimable within it. The potential defeater — that a differently specified model (egg adjusted for total energy, i.e. egg vs equivalent energy from the rest of diet) could identify a quasi-unconditional effect — does not touch this argument, because the statement only claims non-identification for the substitution design, not in principle. So the step holds as stated. Approved.
Link to original
H-18 - The egg-health effect is a property of the comparator not of eggs alone, so eggs good-or-bad is ill-posed
Motivated by the observation that the egg-removal effect ranges from null (vs processed meat) to strongly protective (vs nuts/whole grains) within a single dataset, and the authors’ caveat that an inverse substitution association does not by itself guarantee a healthy choice.
Link to original
A-14 - Full-serving-per-day substitution benefits apply only to the small high-intake minority, so population impact of cutting eggs is smaller
A substitution analysis can only lower a person’s intake of a food by up to their current intake, so the modelled “one serving/day” reduction of eggs is counterfactually available only to participants who already consume >=1 serving/day of eggs. Here that is 8.0% of the sample, versus 58.5% at >=1 serving/week (for whom the modelled reduction is only ~2-3%, HRs ~0.97-0.98).
The steep per-day hazard ratios (down to 0.79 for eggs→nuts, incident CVD) therefore describe a minority with high intake and “greater room for reducing intake.” Averaged over a population most of whom eat far fewer eggs, the attainable reduction from replacing eggs is closer to the small one-serving/week estimates. This is an applicability/weighting inference: it leaves every hazard ratio unchanged but caps the population-level effect implied by the large per-day numbers, and it locates the meaningful egg-reduction benefit at high habitual intakes rather than across the whole intake range.
Validity assessment (step 6)
Traced step: a one-serving/day substitution is counterfactually available only to someone whose current intake is >=1 serving/day; at 8% of the sample, the large per-day HRs describe that minority, and averaged over an intake distribution dominated by weekly-or-less eaters (for whom the attainable reduction, and thus the risk delta, is a small fraction of a serving/day) the population-average benefit is necessarily much smaller. No undercutting defeater survives: even under a fully linear per-serving dose-response the low-intake majority can only realise a proportionally tiny reduction, which is exactly the conclusion. The inference holds as stated (an applicability/weighting move, HRs untouched). approved / checked.
Link to original
relevance_note: Direct primary demonstrating the comparator-dependence of the egg-health association — the core epistemic claim of this slice’s substitution thread.