The popular framing is about how much protein — high-protein for longevity, or high-protein as a kidney/longevity risk. This meta-analysis (observational) associates lower mortality with protein SOURCE, not amount. Total protein is weakly inverse with all-cause mortality, animal protein is flatly null on every outcome, and the signal tracks plant protein. The decision this points to is not a dose — it is a substitution: shift protein-bearing foods from animal toward plant sources. (Whether the lever is the protein or the plant foods it rides in is unresolved here — see the fibre-collinearity gap at the foot of the page.)
This is the mortality (patient-important, observational) half of the protein story; the muscle/
strength half (surrogate, RCT-grade, the ~1.6 g/kg dose) lives on
Protein and Resistance Training for Muscle and Strength — a genuinely distinct decision-question
(what to eat vs how much), not two facets of one. confidence: low — and it stays low after
Budhathoki 2019 (the JPHC Japanese cohort) was woven in, because Budhathoki turns out to be a
constituent study pooled inside Naghshi, so it is refinement (type-F), not the independent
replication (type-E) that would lift confidence. A genuinely independent cohort — one NOT in Naghshi’s
32 — is still owed.
(inferred from Naghshi et al., 2020)
The effect estimates — source is the axis, not amount
Prospective cohorts only (32 studies, 715 128 participants, 113 039 deaths); highest-vs-lowest, fully-adjusted, random-effects. All observational — associations, not causal effects.
| Exposure | All-cause | CVD | Cancer |
|---|---|---|---|
| Total protein | «0.94 … 0.89 to 0.99» (inverse) | «0.98 … 0.94 to 1.03» null | «0.98 … 0.92 to 1.05» null |
| Animal protein | «1.00 … 0.94 to 1.05» null | «1.02 … 0.94 to 1.11» null | «1.00 … 0.98 to 1.02» null |
| Plant protein | «0.92 … 0.87 to 0.97» inverse | «0.88 … 0.80 to 0.96» inverse | «0.99 … 0.94 to 1.05» null |
- Animal protein is null on all three outcomes — the clean finding at the aggregate. But the animal bucket pools red meat with dairy, eggs and poultry, so this null does not clear the animal-protein nutrient: a red-meat-specific protein effect diluted to null within the mixed bucket cannot be ruled out here (see the food-category caveat below). What it does show is that aggregate animal protein is not a mortality signal.
- Plant protein carries the benefit, on all-cause and CVD mortality (not cancer).
- Total protein’s inverse arm is the plant component showing through. «Given that plant protein is part of total protein, the observed inverse association for intake of total protein seems to be related to its plant protein component.» (Naghshi et al., 2020)
The magnitude is small and the shape is inverse-monotone, no U. The one significant dose-response: «an additional 3% of energy from plant proteins a day was associated with a 5% lower risk of death from all causes (pooled effect size 0.95, 95 0.93 to 0.98, P<0.001)» — with significant non-linearity (P=0.05, a steepening of the inverse relation). Total and animal protein dose-response were both flat (0.99, NS). (Naghshi et al., 2020)
No protein type shows an upper-arm harm, so the The U-Shaped Association Artifact hazard does not arise — there is no protective lower arm to adjudicate. The live question is whether the inverse arm is causal, and being observational it is unchecked by Mendelian randomization. (inferred from Naghshi et al., 2020)
Why animal protein is the wrong exposure — the food-category caveat
The animal-protein null does not clear animal-source foods. Naghshi’s own reading is a textbook nutrient-vs-food case Is the Food Category Doing Any Work: «the exposure variable was meat as a food group, whereas our exposure variable was protein as a nutrient. Animal meat contains fat, sodium, iron, and B vitamins in addition to protein» — so «findings for animal meat and animal protein could be different.» And the animal-protein bucket «combin[es] protein from different animal sources, including poultry, eggs, and dairy foods» — a heterogeneous mix whose average may describe no single food. (Naghshi et al., 2020)
So the null is a null on the nutrient, not a green light for the foods — red/processed meat harm (a food-level finding elsewhere) is untouched by it.
How robust is the plant-protein signal
- Survives macronutrient adjustment; total protein does not. «When we confined the analysis to studies that had made these adjustments [fat/carbohydrate], the inverse association of plant protein with all cause and cardiovascular disease mortality changed little, whereas the inverse association between intake of total protein and all cause mortality became non-significant.» (Naghshi et al., 2020) The plant-protein result is the sturdier one; the total-protein result is fragile.
- But confounded by diet pattern and social class (author’s own caveat). «Consumption of animal and plant proteins could be a marker of broader dietary intake patterns—or even of social class, an important independent predictor of many health outcomes.» (Naghshi et al., 2020)
- Measurement error attenuates toward null. «Measurement errors in dietary assessment are inevitable and would have tended to underestimate the associations with protein intake» (Naghshi et al., 2020) -> the true gradient may be steeper than measured Measurement Error in Dietary Assessment.
- Publication bias possible (Egger’s positive for total-all-cause, total-CVD, plant-CVD) but trim-and-fill left the estimates unchanged.
(inferred from Naghshi et al., 2020)
The proposed mechanisms — directional, not outcome evidence
Naghshi proposes: animal protein raises IGF-1 (linked to cancer/age-related disease) where plant protein does not; plant protein associates with favourable BP, waist, body composition and lower plasma cholesterol; gut fermentation of plant protein lowers toxic/carcinogenic metabolites; and plant protein’s amino-acid profile (lower lysine/histidine; more arginine) plausibly lowers apoB-lipoprotein secretion and shifts the glucagon/insulin balance. These are mechanism-grade with mixed human corroboration — they inform direction, and must not be read as outcome findings.
Corroborated by Seidelmann 2018 (type-F, shared-school — NOT independent-E)
Seidelmann’s carbohydrate->mortality study (The U-Shaped Association Artifact,
Low-Carbohydrate vs Balanced-Carbohydrate Diets) reaches the plant-favourable / animal-unfavourable
substitution verdict from a carbohydrate frame rather than a protein one: substituting plant fat
and protein for carbohydrate was associated with lower all-cause mortality, animal fat and protein
with higher. The direction agrees with Naghshi’s source-axis on a second dietary framing — but this
is F/shared, not independent-E, so it earns no [E-independent] token and no confidence lift
(the page stays confidence: low): Willett is a co-author on BOTH papers, and Seidelmann’s
meta-analysis pools the same NHS/HPFS (Fung) cohorts Naghshi’s protein MA draws on — a shared
confounding structure would move both. It is listed in sources: because the parameter table below
extracts its distinct values (not a corroboration-only line), but the second listing does not make
this two independent routes.
Parameter table (op-weave 2a — the apparent animal discrepancy is a unit difference, not a tension):
| Parameter | Naghshi 2020 (protein source) | Seidelmann 2018 (carb substitution) | Same quantity? |
|---|---|---|---|
| Plant signal, all-cause | «0.92 … 0.87 to 0.97» (highest-vs-lowest plant protein) | «0·82, 0·78-0·87» (plant fat+protein substituted for carb) | no — same inverse DIRECTION, different estimand |
| Animal signal, all-cause | «1.00 … 0.94 to 1.05» null (animal protein nutrient) | «1·18, 1·08-1·29» harm (animal fat+protein-for-carb pattern) | no — nutrient-null vs pattern-harm |
| Exposure unit | protein as a nutrient, energy-adjusted | a low-carb dietary pattern (protein AND fat) replacing carbohydrate | NO |
| Design / independence | 32-cohort FFQ MA; senior author Willett | ARIC + 8-cohort FFQ MA; senior author Willett; PURE/NHS/HPFS overlap | shared school + overlapping cohorts — NOT independent |
The animal null (Naghshi) and animal-substitution harm (Seidelmann) do NOT contradict — not-joined check (ii) fires on the unit. Naghshi isolates the protein nutrient (null); Seidelmann measures a whole low-carb pattern that carries animal fat with the protein. Naghshi’s own nutrient-vs-food caveat predicts exactly this — «animal meat contains fat, sodium, iron, and B vitamins in addition to protein» — so the harm in Seidelmann’s animal score is substantially the accompanying fat and pattern, which is consistent with animal protein-the-nutrient being null. The two reinforce the nutrient-vs-food distinction rather than clashing -> Is the Food Category Doing Any Work. (inferred from Naghshi et al., 2020; Seidelmann et al., 2018)
A larger pooled carbohydrate MA does NOT advance the source axis — Qin 2023 (G-gap) [2026-08-19].
The 41-cohort SR+MA that re-pools Seidelmann (see The U-Shaped Association Artifact,
Low-Carbohydrate vs Balanced-Carbohydrate Diets) measures total carbohydrate quantity only and runs
no substitution decomposition — it cannot say what replaces the carbohydrate, so it adds pooled
magnitude to the carb-quantity arm but leaves the animal-vs-plant source question — the axis this page
turns on — exactly where Seidelmann left it. Its own reframing points the same way qualitatively («shifting
the focus … from carbohydrate quantity to carbohydrate quality»), but quality/source is a named gap in
Qin (3 unpoolable studies), not a decomposed estimate.
(inferred from Qin et al., 2023)
Budhathoki 2019 (JPHC Japan) — same verdict, and it is a CONSTITUENT of Naghshi (type-F, NOT independent-E)
Budhathoki’s Japanese cohort (JPHC; 70 696 adults, 18-y follow-up, 12 381 deaths) reaches the identical plant-favourable / animal-null pattern: «Higher total and animal protein intake was not associated with risk of overall mortality or cause-specific mortality», while «higher plant protein intake was associated with lower total and CVD-related mortality» (plant all-cause Q5 HR 0.87 [0.78-0.96], P=.01; CVD Q5 0.73 [0.59-0.91], P=.002; cancer null). (Budhathoki et al., 2019)
It was predicted as the independent-E lift this page lacked, and the prediction FAILED. The author
lists are cleanly disjoint (Budhathoki/Sawada/Iwasaki/Tsugane, National Cancer Center Japan — no Willett,
no NHS-HPFS name), and the dataset is a distinct non-Western population. But type-E requires
independence of data, and that fails: Naghshi’s meta-analysis POOLED Budhathoki as one of its ~32
constituent cohorts (Naghshi’s included-studies table lists «Budhathoki 2019, Japan» with the exact
JPHC counts — M 32 201, W 38 495, 12 381 deaths — and Naghshi ref 18 is «Budhathoki S, Sawada N,
Iwasaki M, et al»). So Budhathoki’s data already sits inside Naghshi’s pooled estimate; reading it as
independent corroboration would double-count the JPHC data — the laundered-E trap. Verdict: F,
no [E-independent] token, no confidence lift — the page stays confidence: low.
What Budhathoki genuinely ADDS (why it is a refinement, not an echo). Naghshi reports pooled highest-vs-lowest and a per-3%-energy dose-response; Budhathoki supplies two layers Naghshi’s pooling blurs — an isocaloric food-source substitution model and absolute risk:
| Parameter | Naghshi 2020 (pooled MA) | Budhathoki 2019 (JPHC single cohort) | Same quantity? |
|---|---|---|---|
| Plant vs animal signal | plant inverse (all-cause 0.92), animal null (1.00) | plant inverse (all-cause Q5 0.87), animal null (Q5 0.98) | yes — same pattern (but Budhathoki ⊂ Naghshi) |
| Plant dose-response | «0.95 … 0.93 to 0.98» per +3%E plant-for-carbohydrate | — | n/a |
| Food-source swap | not reported | plant-for-red-meat protein «0.66 … 0.55-0.80»; plant-for-processed-meat «0.54 … 0.38-0.75»; fish-for-red-meat 0.75 | NO — food-for-food swap, larger than the plant-for-carb estimand |
| Absolute risk | not derivable (highest-vs-lowest) | 15-y ARR plant-for-red-meat 3.60% (2.10-4.86) total; processed-meat 4.95% | NO — Budhathoki adds the absolute layer |
| Animal-protein food mix | pooled (US-weighted, red-meat-heavy) | fish 47.1%, red meat 19.4%, dairy 16.7%, eggs 9.5% | NO — fish-dominated, the transportability hinge |
Note the substitution HRs are a DIFFERENT estimand from Naghshi’s dose-response, not a bigger version of it. Naghshi’s 0.95 is +3%E plant protein replacing carbohydrate; Budhathoki’s 0.66 is 3%E plant protein replacing red-meat protein — a food-for-food swap that captures the plant benefit AND the red-meat-removal, so it is necessarily larger. Do not read 0.66 as a stronger plant effect than 0.95.
The Japanese fish-dominated profile SHARPENS the nutrient-vs-food point -> Is the Food Category Doing Any Work. WITHIN “animal protein”, the sources diverge: fish-for-red-meat substitution is itself protective (0.75) while plant-for-dairy and plant-for-fish are null — so aggregate animal protein is null precisely because it pools protective fish with harmful red/processed meat. Budhathoki’s own reconciliation of its animal-null vs the US animal-positive result: «This discrepancy … may be attributable to … a difference in the main dietary source of animal protein, which was red and processed meat in the US study vs fish intake in the present study.» (Budhathoki et al., 2019) (inferred from Budhathoki et al., 2019)
Decision relevance
- The lever is source-substitution, not a protein target. Replacing animal-protein foods with plant-protein foods (legumes, nuts, whole grains, soy) is associated with lower all-cause and CVD mortality; raising or lowering total protein does little on its own. «Replacement of foods high in animal protein with plant protein sources could be associated with longevity.» (Naghshi et al., 2020)
- The effect is modest (~5% lower all-cause mortality per +3% energy from plant protein, relative, observational) — a real but small lever, ranked below the big rocks and delivered on associational evidence, not a causal RCT. On Naghshi’s highest-vs-lowest contrast the absolute reduction is not derivable; Budhathoki’s JPHC model does supply one for the food-source swap — a 15-year absolute risk reduction of 3.60% (2.10-4.86) in total mortality for replacing 3%E of red-meat protein with plant protein — but it is model-derived from a single observational cohort, so read it as scale, not a treatment effect. (Budhathoki et al., 2019)
- This does not conflict with the muscle-protein target. Hitting ~1.6 g/kg for muscle (Protein and Resistance Training for Muscle and Strength) is a quantity decision on a surrogate; this is a source decision on mortality. You can satisfy both: reach the amount, bias the sources plant-ward. Quality (DIAAS) is the bridge — plant sources are lower-DIAAS, so a plant-shifted diet needs more grams or complementation to hold the muscle target -> Protein Quality and the DIAAS Score.
Limits
- All observational; no causal claim. Residual confounding (diet pattern, social class) is the author’s own leading caveat.
- The animal protein and plant protein labels are heterogeneous buckets — the animal bucket pools poultry/eggs/dairy/red meat; the plant bucket pools legumes/grains/nuts. Category-level estimates over a wide mix (the food-category trap).
- Western-dominated cohorts — animal-protein generalisability to low/middle-income (carb-rich, low animal-source) diets is limited.
- Still zero genuinely independent sources —
confidence: low. Naghshi is the gold opener; Seidelmann is not independent (shared Willett lineage + NHS/HPFS overlap); Budhathoki is not independent either (its JPHC data is pooled inside Naghshi’s MA). So all three backings share a data or school lineage. A genuine independent-E lift AWAITS a cohort not among Naghshi’s ~32 constituents — and note Song 2016 (the staged pairing candidate) is NHS-HPFS and near-certainly pooled in Naghshi too, so it will not supply independence. The owed source is a large non-Western / non-Naghshi-pooled cohort that could also test whether the plant signal is independent of the fibre/whole-grain lever it may run through.
The plant-protein signal may be the fibre/pulse lever under another name (a gap)
Plant-protein foods are largely the legume/whole-grain/nut foods the wiki already credits with lower mortality via fibre Dietary Fibre and Health and the pulse/whole-grain evidence Whole Grains Refined Grains and Pulses. Naghshi cannot separate the plant-protein nutrient from the food matrix carrying it, so whether plant protein is an independent lever or a proxy for those foods is unresolvable here — a G-gap, and a caution against double-counting it as a separate big rock. (inferred from Naghshi et al., 2020)