One coordinated meta-analysis series (the DIfE/Boeing group: Schwingshackl, Bechthold, Schlesinger, Boeing) applied the same 12-food-group dose-response framework to five outcome families — all-cause mortality, type 2 diabetes (T2D), hypertension, cardiovascular disease (CHD/stroke/heart failure), and adiposity (overweight/obesity, abdominal obesity, weight gain). Laid side by side, the five papers form a grid no single paper contains: which food groups move which outcomes, and — more informatively — where a food’s effect diverges across outcomes. That cross-outcome grid is the emergent object here (type-A), and the divergences (fish, dairy, eggs, vegetables) are the payoff, not the uniform effects.
Load-bearing caveat: this is ONE evidence base sliced five ways, NOT five independent confirmations
The five meta-analyses share a research team, the same 12 a-priori food-group definitions, one registered protocol (PROSPERO CRD42016037069), overlapping PubMed/Scopus/Web-of-Science searches, and a heavily overlapping cohort pool (the same prospective cohorts report multiple outcomes, so one cohort feeds several columns). (Schwingshackl et al., 2017) (Schlesinger et al., 2019)
Consequently, cross-outcome consistency is NOT type-E independent corroboration. That processed meat reads harmful in all five columns is one program applying one method to one literature five times, not five independent tests converging — the agreement is partly mechanical (shared cohorts, shared confounding structure, shared dietary-measurement instrument). The confidence attached to any single cell therefore rests on that cell’s own NutriGrade rating (H/M/L/VL below), never on how many other columns agree. Do not let the visual coherence of the grid launder into raised certainty. (The one place this caveat is partially relieved: the meat -> coronary cells now carry an independent-team robustness check -> Independent-team cross-check on the meat -> coronary cells below.)
The increment key (constant per food across all five outcomes)
Each cell is the pooled linear dose-response RR per the food’s serving increment, with 95% CI. Increments (identical across the series): whole grains 30 g/d, refined grains 30 g/d, vegetables 100 g/d, fruit 100 g/d, nuts 28 g/d, legumes 50 g/d, eggs 50 g/d, dairy 200 g/d, fish 100 g/d, red meat 100 g/d, processed meat 50 g/d, SSB 250 mL/d. (Schwingshackl, Hoffmann, et al., 2017)
Grade codes = NutriGrade meta-evidence: H high, M moderate, L low, VL very
low. A CI crossing 1.00 = not statistically significant (the estimate can still inform
direction under measurement-error attenuation). * = source reports significant non-linearity
(plateau/threshold — see Shape honesty). ins. = single-study or extreme-imprecision cell,
held as insufficient evidence, not as effect.
The matrix (RR per increment, 95% CI, grade)
The CVD column shows CHD as representative; stroke and heart failure diverge from CHD and are broken out in the next table.
| Food group (increment) | Mortality | T2D | Hypertension | CVD (CHD) | Adiposity (OW/OB) |
|---|---|---|---|---|---|
| Whole grains (30 g) | 0.92 (0.89-0.95) H* | 0.87 (0.82-0.93) H* | 0.92 (0.87-0.98) L* | 0.95 (0.92-0.98) M* | 0.93 (0.89-0.96) L |
| Refined grains (30 g) | 0.99 (0.97-1.01) L | 1.01 (0.99-1.03) M | 0.99 (0.96-1.02) VL | 1.01 (0.99-1.04) L | 1.05 (1.00-1.10) VL* |
| Vegetables (100 g) | 0.96 (0.95-0.98) L* | 0.98 (0.96-1.00) M* | 1.00 (0.98-1.01) VL | 0.97 (0.96-0.99) M | 0.98 (0.93-1.03) L |
| Fruit (100 g) | 0.94 (0.92-0.97) L* | 0.98 (0.97-1.00) M* | 0.97 (0.96-0.99) L* | 0.94 (0.90-0.97) M* | 0.93 (0.86-1.00) L |
| Nuts (28 g) | 0.76 (0.69-0.84) M* | 0.89 (0.71-1.12) L | 0.70 (0.45-1.08) L | 0.67 (0.43-1.05) M* | 0.78 (0.58-1.06) L* |
| Legumes (50 g) | 0.96 (0.90-1.01) M | 1.00 (0.92-1.09) L | 0.98 (0.95-1.01) VL | 0.96 (0.92-1.01) L* | 0.88 (0.84-0.93) L ins. |
| Eggs (50 g) | 1.15 (0.99-1.34) VL | 1.08 (0.95-1.22) M | ins. | 1.00 (0.95-1.06) L | ins. |
| Dairy (200 g) | 0.98 (0.93-1.03) M | 0.97 (0.94-0.99) M | 0.95 (0.94-0.97) L | 0.99 (0.96-1.02) M | 0.97 (0.93-1.01) L |
| Fish (100 g) | 0.93 (0.88-0.98) M | 1.09 (0.93-1.28) M | 1.07 (0.98-1.16) L* | 0.88 (0.79-0.99) M | ins. |
| Red meat (100 g) | 1.10 (1.04-1.18) M | 1.17 (1.08-1.26) H | 1.14 (1.02-1.28) L | 1.15 (1.08-1.23) M* | 1.10 (1.04-1.16) VL |
| Processed meat (50 g) | 1.23 (1.12-1.36) M* | 1.37 (1.22-1.55) H* | 1.12 (1.00-1.26) L* | 1.27 (1.09-1.49) M | 1.18 (1.02-1.36) VL ins. |
| SSB (250 mL) | 1.03 (0.91-1.18) L* | 1.21 (1.12-1.31) H* | 1.07 (1.04-1.10) L | 1.17 (1.11-1.23) M | 1.05 (1.00-1.11) VL |
Cell provenance, one tag per source (each on its own line so the slug stays intact):
- Mortality column: (Schwingshackl, Schwedhelm, Hoffmann, Lampousi, et al., 2017)
- T2D column: (Schwingshackl, Hoffmann, et al., 2017)
- Hypertension column: (Schwingshackl, Schwedhelm, Hoffmann, Knüppel, et al., 2017)
- CVD column: (Bechthold et al., 2017)
- Adiposity column: (Schlesinger et al., 2019)
Cell notes: Eggs T2D RR is reported per 30 g (not 50 g) in the source — read as directional
only. Eggs hypertension rests on one study (144 cases, RR 0.25) — insufficient, not a
protective finding. Legumes / processed meat / fish adiposity cells rest on single studies
or extreme imprecision (processed-meat abdominal RR 8.80, CI 1.20-64.28) — ins. The
0.88 legumes-adiposity point is one study. Fish adiposity is inverse only for abdominal
obesity (0.83, 0.71-0.97), null for overweight/obesity.
CVD is not one outcome — the within-CVD divergence (Bechthold)
| Food group (increment) | CHD | Stroke | Heart failure |
|---|---|---|---|
| Whole grains (30 g) | 0.95 (0.92-0.98) M | 0.99 (0.95-1.03) L | 0.96 (0.95-0.97) L |
| Vegetables (100 g) | 0.97 (0.96-0.99) M | 0.92 (0.86-0.98) M | 0.96 (0.94-0.98) L |
| Fruit (100 g) | 0.94 (0.90-0.97) M | 0.90 (0.84-0.97) M | 0.98 (0.94-1.01) L |
| Nuts (28 g) | 0.67 (0.43-1.05) M | 0.99 (0.84-1.17) L | 1.09 (0.97-1.22) L |
| Eggs (50 g) | 1.00 (0.95-1.06) L | 0.99 (0.93-1.05) M | 1.16 (1.03-1.31) M |
| Dairy (200 g) | 0.99 (0.96-1.02) M | 0.98 (0.96-1.00) M | 1.08 (1.01-1.15) L |
| Fish (100 g) | 0.88 (0.79-0.99) M | 0.86 (0.75-0.99) M | 0.80 (0.67-0.95) M |
| Red meat (100 g) | 1.15 (1.08-1.23) M | 1.12 (1.06-1.17) M | 1.08 (1.02-1.14) M |
| Processed meat (50 g) | 1.27 (1.09-1.49) M | 1.17 (1.02-1.34) M | 1.12 (1.05-1.19) M |
| SSB (250 mL) | 1.17 (1.11-1.23) M | 1.07 (1.02-1.12) M | 1.08 (1.05-1.12) L |
Whole grains and nuts protect CHD but are null for stroke; eggs and dairy are null for CHD/stroke but positive for heart failure (the egg->HF signal is the strongest divergence in the series). Fish is the only food inverse across all three CVD subtypes. Reading “CVD” as a single endpoint hides these — a stratum whose dominant risk is stroke ranks vegetables/fruit above whole grains/nuts.
Independent-team cross-check on the meat -> coronary cells (Papier 2021) [2026-09-02]
The load-bearing caveat above says cross-outcome agreement inside the DIfE/Boeing series is NOT independent corroboration. A separate gold SR+MA from an independent team (Papier/Key, Oxford; no Schwingshackl/Boeing/DIfE overlap) re-estimates the red-meat, processed-meat, and poultry -> ischemic-heart-disease (IHD) cells and lands concordant with the Bechthold CHD column: over 1.4 million adults, 32,630 cases; unprocessed red meat RR 1.09 (95% CI 1.06-1.12) and processed meat 1.18 (1.12-1.25) per 50 g/day; no association for poultry (1.02, 0.97-1.07 per 50 g/day). (Papier et al., 2021)
Parameter table — matched, dose-harmonized, same-quantity checked
| Parameter | Bechthold 2019 (matrix CHD) | Papier 2021 (IHD) | Same quantity? |
|---|---|---|---|
| Team | Schwingshackl / Boeing / DIfE | Key / Oxford | NO — genuinely independent |
| Endpoint | CHD; excludes fatal-only cohorts | IHD, incidence and/or death; includes fatal-only | related, not identical |
| Pooling model | random-effects | fixed-effects (RE sensitivity concordant) | different primary model |
| Unprocessed red-meat dose | per 100 g/day | per 50 g/day | NO — must harmonize |
| Red-meat RR | 1.15 (1.08-1.23) /100 g | 1.09 (1.06-1.12) /50 g = 1.19 (1.12-1.25) /100 g | YES after log-linear conversion; concordant |
| Processed-meat dose | per 50 g/day | per 50 g/day | YES — same unit, no conversion |
| Processed-meat RR | 1.27 (1.09-1.49) /50 g | 1.18 (1.12-1.25) /50 g | YES — CIs overlap; concordant |
| Poultry | not among the 12 groups | 1.02 (0.97-1.07) /50 g, null | new food group |
| Primary cohort pool | older cohorts (NHS, EPIC, ARIC, Adventist, HPFS-lineage) | same older cohorts + >1M new (UK Biobank, PURE, JPHC, DNSDP) | PARTIAL overlap |
Dose harmonization is log-linear: RR per 100 g = (RR per 50 g)^2, so Papier’s red-meat 1.09/50 g = 1.19/100 g (CI 1.06^2 to 1.12^2). The processed-meat comparison needs no conversion — both report per 50 g/day. On both meats the harmonized estimates are the same quantity (modulo the CHD-vs-IHD endpoint difference) and agree.
Independence verdict — type-F, NOT type-E (do not stamp [E-independent])
Papier and Bechthold share a substantial fraction of primary cohorts (NHS/Bernstein, EPIC, ARIC/Haring, Adventist/Fraser, HPFS-lineage all recur), and Papier cites Bechthold as an antecedent MA it updates. Two MAs pooling overlapping primary data are not independent tests, so this fails the strict-E bar (neither shared-lineage nor citing-each-other). It is type-F: an independent team, using a different endpoint definition (IHD incl. fatal-only vs CHD excl. fatal-only), a different pooling model, and >1 million participants of new cohort data absent from Bechthold, reaches a concordant estimate. That partially relieves the “one team, one pool, one method” caveat for the meat -> coronary cells specifically — a team-and-method robustness check — without being the fully independent corroboration a raised certainty would need.
Why red-meat -> IHD meta-analyses disagreed — a power story, not an effect-absence (type-C)
Papier states that «previous meta-analyses on unprocessed red meat and IHD were based on few studies» — two early ones (Micha 2010, Abete 2014) finding no association for incident or fatal IHD, two more recent (Bechthold 2019, Zeraatkar 2019) reporting a positive association. (Papier et al., 2021) The reconciliation is precision, not contradiction: Papier’s red-meat pool carries «34,949 cases from 12 studies», against the «6,659 cases» from five studies that Bechthold pooled — over four times the cases. (Papier et al., 2021) So the earlier null MAs are insufficient-evidence (underpowered), not demonstrated no-effect — the four evidence-states distinction, and the expectancy test applied to a growing case count.
Absolute-risk anchor for the meat rows — the named G-gap, now cashed for meat (Zhong 2020) [2026-09-08]
The Gaps section flagged that this series «computes no absolute risk», so the meat RRs could not be sized against a drug comparator (Layer-1). A pooled analysis of 6 US prospective cohorts (Zhong 2020; 29,682 adults, 6,963 incident CVD events, 8,875 deaths, median 19-y follow-up) supplies that anchor for the meat rows — a US-adult baseline and a 30-year absolute risk difference (ARD) per a realistic +2 servings/week increment, from the fully food-adjusted model (model 3).
Baseline event rates in the pool: incident CVD 13.2 / 1000 person-years, all-cause mortality 15.8 / 1000 person-years. (Zhong et al., 2020)
| Exposure (+2 svg/wk) | Incident-CVD HR | CVD 30-y ARD | All-cause HR | ACM 30-y ARD |
|---|---|---|---|---|
| Processed meat | 1.07 (1.04-1.11) | +1.74% (0.85-2.63) | 1.03 (1.02-1.05) | +0.90% (0.43-1.38) |
| Unprocessed red meat | 1.03 (1.01-1.06) | +0.62% (0.07-1.16) | 1.03 (1.01-1.05) | +0.76% (0.19-1.33) |
Across all four exposures Zhong found «approximately 3% to 7% higher relative risks and less than 2% higher absolute risks of incident CVD and all-cause mortality over the 30 years of follow-up». (Zhong et al., 2020) So the absolute meat rock is small: even processed meat — the matrix’s strongest harmful lever — moves 30-year CVD risk by under 2 percentage points at a realistic intake change. That is the Layer-1 rock-sizing the relative matrix could not compute on its own, converting the RR columns into a rankable magnitude for the meat rows (net of a drug comparator -> Layer 1 - Ranking Interventions for a Stratum).
Parameter table — Zhong vs the matrix CVD column (same quantity?)
| Parameter | Bechthold 2019 (matrix CVD) | Zhong 2020 | Same quantity? |
|---|---|---|---|
| Endpoint | CHD (stroke / HF broken out separately) | COMPOSITE incident CVD (CHD + stroke + HF + CVD death) | NO — Zhong pools the subtypes Bechthold splits |
| Estimand | relative RR per increment | HR + 30-y absolute ARD | NO — Zhong adds the absolute layer the matrix lacked |
| Increment | per 100 g/d (red), 50 g/d (processed) | per +2 svg/wk (~32 g/d red; ~14 g/d processed) | NO — different, and processed->CVD is nonmonotonic so log-linear harmonization is invalid |
| Cohort pool | international cohorts incl. US | 6 US cohorts (ARIC, CARDIA, CHS, FHS, FOS, MESA) | PARTIAL overlap (shared US cohorts) |
| Direction | meat harmful | meat harmful | YES on sign only |
Verdict: type-F (absolute anchor + partial robustness), NOT type-E. Zhong shares US cohorts with the matrix pool and only the direction agrees, so this is not independent corroboration of the relative estimate — it does not raise the meat cells’ certainty. Its value is the absolute layer the matrix’s G-gap named as missing, plus a hard-endpoint composite read; the relative concordance is a bonus, not a certainty upgrade.
Shape note. All four Zhong exposures were monotonic over the studied range except processed meat -> incident CVD, nonmonotonic at P for nonlinearity = .006 (modelled with a quadratic term; its ARD is the 2-vs-0 servings/week contrast, not a linear per-serving slope). Zhong does not characterize the curve’s shape verbally, so no knee or plateau is located — a harmful-food nonlinearity, not a protective lower arm, so the every-reduction-pays default holds for the harmful rows. (Zhong et al., 2020) -> The U-Shaped Association Artifact
Genetic-instrument check on the meat rows (Li 2024 MR) [2026-10-06]
A two-sample Mendelian-randomization study instrumented UK Biobank intake of processed meat, pork, beef and lamb and tested CAD, hypertension, stroke and T2D. Every IVW estimate was null: «Inverse-variance weighted analyses reveal no causal association between the consumption of processed and red meat and the incidence of CAD, hypertension, stroke, and T2DM (P > 0.05).» (Li et al., 2024). The question for this page is whether that null bears on the matrix cells. It mostly does not, for the reasons below.
Parameter table — matrix cell vs MR estimate (same quantity?)
| Parameter | Matrix cell (observational dose-response MA) | Li 2024 MR (IVW) | Same quantity? |
|---|---|---|---|
| Exposure contrast | per 100 g/d red meat, per 50 g/d processed meat | per unit of genetically predicted UK Biobank intake answer; unit never stated | No |
| Meat definition | red meat pooled; processed meat pooled | pork, beef, lamb separately; processed meat | No for red meat |
| T2D estimate | red 1.17 (1.08-1.26); processed 1.37 (1.22-1.55) | beef 1.803 (0.867-3.750); pork 1.235 (0.362-4.223); lamb 0.929 (0.498-1.733); processed 1.090 (0.714-1.664) | No (RR per gram vs OR per unknown unit) |
| CHD/CAD estimate | red 1.15 (1.08-1.23); processed 1.27 (1.09-1.49) | beef 1.182 (0.702-1.989); processed 0.935 (0.630-1.386) | No (same reason) |
| Confounding control | multivariable adjustment in cohorts | genetic instrument; PhenoScanner screen for SNPs tied to the outcome diseases (Methods) or unnamed confounding factors (Results), MR-Egger intercept | different design, not a parameter |
Matrix T2D values (Schwingshackl, Hoffmann, et al., 2017); CHD values (Bechthold et al., 2017); MR values (Li et al., 2024). Because no row reaches yes, no direct contradiction or confirmation of a matrix cell is claimed here.
Expectancy test — could this MR have seen the effect?
- On CAD and T2D, apparently not. Read heuristically, and only if one genetically predicted unit is roughly one serving (unverifiable, since the unit is unstated): each CAD/T2D interval contains both OR 1 and the observational 1.15-1.37 (e.g. processed->T2D 0.71-1.66, beef->T2D 0.87-3.75, beef->CAD 0.70-1.99). That is the insufficient-evidence state, not no meaningful effect.
- On hypertension and stroke, the estimates are not interpretable yet. Their intervals are far narrower than CAD’s despite similar or much smaller case counts, which suggests linear-probability outcome betas rather than log-odds; the paper does not state the scale (detail on the source page). The hypothesis cuts both ways: if these betas are in fact log-odds, the hypertension and stroke rows are precise nulls the matrix’s HTN and stroke cells would have to answer. It is an open check, not a free pass.
- Missing features that would make a null informative: no stated exposure unit, no power calculation, no F values in the main text (only the F>10 cut-off; supplement unheld), no sample-overlap assessment, no named BMI or education pleiotropy screen for instruments the authors themselves describe as «SNPs related to the dietary preference of red and processed meat» (Li et al., 2024). Weak instruments bias toward the null only when exposure and outcome samples do not overlap; with possible UK Biobank overlap (unassessed) the bias runs toward the confounded observational association, so the direction of weak-instrument bias in this study is unknown.
- Symmetric standard. A harm finding from an MR with the same instruments (preference SNPs, no named BMI/education screen, no overlap handling) would earn the same validity discount. The authors’ line that their findings «challenge the prevailing conventional perspective in the field» (Li et al., 2024) outruns their own intervals.
Net effect on the matrix: the meat cells keep their grades and their association-not- causation caveat. What changes is the description of the gap: a genetic test has been attempted once, at a precision that cannot separate the observational effect size from zero. Value: F (corrects no MR in either direction) + G (below). Not D: the issue is not joined, since the quantities differ and the intervals overlap. Not E: not concordant, and Li shares its exposure GWAS with the only earlier meat->CVD MR it cites (Yang 2023, not held).
Layer-1 across-food-group ranking (which levers move the most outcomes)
This is the Layer-1 input the series was acquired for -> Layer 1 - Ranking Interventions for a Stratum. Ranking by direction-consistency x magnitude x certainty, net of each cell’s grade:
Largest, most consistent PROTECTIVE levers. Whole grains lead — protective in every outcome family and the only protective food carrying two HIGH-grade cells (mortality, T2D). Fruit is the consistent runner-up (modest inverse across all five, mostly moderate/low grade). Nuts carry the largest point estimates (mortality 0.76, CHD 0.67) but wide CIs and low grade outside mortality — a big-but-uncertain lever. Fish is high-value but outcome-specific (see below).
Largest, most consistent HARMFUL levers. Processed meat leads — harmful in all five, the only harmful food with a HIGH-grade cell (T2D, RR 1.37 per 50 g — the largest single effect in the matrix). Red meat mirrors it one notch smaller (harmful all five). SSB is a strong cardiometabolic-harm lever (T2D/CVD/adiposity/HTN) that is null for all-cause mortality.
Causal-attribution caveat — the red-meat -> T2D cell is a robust ASSOCIATION, not established causation. The grade does less work than the word “harmful” implies, and the fabric holds this cell OPEN in both directions (untested for causation), not as a demonstrated causal harm:
- NutriGrade “high” grades association-robustness, not confounding-exclusion. The source grades meta-evidence “high for processed meat, red meat, whole grains, and SSB” (Schwingshackl, Hoffmann, et al., 2017) on study count, dose-response, precision and low heterogeneity, and its prospective design “effectively avoided recall bias and reduced the potential for selection bias” (Schwingshackl, Hoffmann, et al., 2017). Neither move touches lifestyle (healthy-user) confounding, which prospective sampling does not remove — so a “high” grade certifies the association is real and consistent, not that red meat causes it.
- The coded exposure is a decontextualized quantity, so red meat may be a PATTERN marker. The series pools red meat only as a per-100 g/day increment and reports servings, not questionnaire items (a
food frequencysearch of the source returns a true corpus-zero); it carries no information on the meal package (refined-carb bun, fried sides, SSB) or the Western dietary pattern and lower health-consciousness that higher red-meat intake co-occurs with. The RR can attach to that package/pattern rather than to the meat -> the substitution/pattern G-gap below. - Guideline-adherence confounding is a self-fulfilling pathway the design cannot exclude. Because guidance itself instructs the health-conscious to cut red meat, red-meat avoidance is collinear with the whole guideline-adherence bundle (not smoking, exercising, screening, taking prescribed medication). In a population that is heavily prediabetic/dysglycemic, avoiders may develop less T2D partly because they are adherent — so the guideline, not the meat, can be the operative cause. This is a sharpened form of healthy-user confounding that is especially hard to adjust away, precisely because the exposure is itself the subject of the guidance. The bundle’s manufacturable magnitude is not small: adherence to a placebo alone carried «lower mortality (0.56, 0.43 to 0.74)» (Simpson et al., 2006) — an all-cause-mortality OR the health-conscious bundle produces with zero causal input -> The Observational-Trial Discordance.
- The signal is not stratum-stable. The red-meat association “could not be confirmed pooling two Asian studies,” and the Shanghai Women’s Health Study showed “an inverse association between red meat and T2D among normal weight women and an increased risk among obese women” (Schwingshackl, Hoffmann, et al., 2017) — a within-population flip that tracks adiposity, not red meat per se.
- No natural experiment either way. No Mendelian-randomization or feeding trial isolates red-meat (or heme-iron) -> T2D in either direction. So — symmetric with coffee -> T2D, which the fabric downgrades to insufficient-for-causation on a null MR — red meat is LESS resolved than coffee (untested, not disconfirmed): held open, neither established-harmful nor shown-benign -> The Observational-Trial Discordance.
- Superseded in part
[2026-10-06, Li 2024]: one intake-instrumented MR is now held (processed meat, pork, beef, lamb -> T2D, all IVW null), so no Mendelian-randomization no longer holds for red meat. It does not resolve the cell: its intervals contain the observational estimates and its exposure unit is unstated, so the cell moves from untested to tested once, insufficient and stays open both ways. Still no MR isolates heme iron, and no feeding trial exists -> Genetic-instrument check on the meat rows above.
- Superseded in part
Weak or null levers (the ceiling-is-a-finding cases). Refined grains are essentially null everywhere except mild adiposity harm at high intake; legumes are null in most linear cells (protective only non-linearly); vegetables are weak and null for hypertension and adiposity; eggs are mostly null/insufficient. For a stratum already eating these, the marginal lever is small by construction.
Divergence findings — outcome-specificity is the informative signal
Uniform effects (meat harmful, whole grains protective) mostly restate what each food-group page already holds. The divergences are where the matrix earns its place:
- Fish — protective for mortality and all three CVD subtypes, but null for T2D (even positive in American cohorts) and null/slightly positive for hypertension. Same food, opposite glycaemic/pressor signal.
- Dairy — protective for T2D and hypertension only; null for mortality/CHD/stroke and positive for heart failure. A cardiometabolic-marker lever, not a mortality lever.
- Eggs — null across mortality/T2D/CHD/stroke but positive for heart failure specifically (HF 1.16-1.25). The one place eggs move a hard outcome.
- Vegetables — protective for mortality/T2D/CVD but null for hypertension and adiposity, against DASH-era expectation; the source attributes the HTN null partly to BMI over-adjustment.
- SSB — strong cardiometabolic harm but null for all-cause mortality (the mortality CI is wide, 0.91-1.18 — an unmeasured, not a demonstrated-absent, effect).
Shape honesty — where the curve has a knee or plateau
Most linear cells are monotone over the studied range only — do not read more-is-always-better past the data edge. Reported non-linear features:
- Nuts plateau early: most of the mortality/CHD benefit is captured by ~10-20 g/d (a small handful); little added benefit above. (Bechthold et al., 2017)
- Vegetables and fruit plateau at ~200-400 g/d for mortality/CVD — the rising arm is at low intake. (Schwingshackl, Schwedhelm, Hoffmann, Lampousi, et al., 2017)
- Whole grains most benefit by ~50 g/d (T2D) to ~100 g/d (CHD).
- SSB->T2D is monotone increasing over the studied range — the included studies’ overall intake range is 0-748 mL/d — non-linear in shape but with risk rising «throughout all the range of SSB investigated»; no threshold is shown within that range, and nothing is said above it (corrected 2026-10-06: across the entire range, every increment adds risk -> monotone over 0-748 mL/d only). (Schwingshackl, Hoffmann, et al., 2017)
- Refined grains->adiposity is J-shaped, harm emerging above ~90 g/d.
The plateaus mean over-shooting a protective food merely fails to add benefit (rarely harms) — the every-reduction-pays default holds for the harmful foods, whose curves show no protective lower arm. -> The U-Shaped Association Artifact
Cross-source check — the legume row against a dedicated legume SR+MA [2026-08-28, Thorisdottir]
A purpose-built legume SR+MA (Thorisdottir 2023, NNR2023 — 47 studies, 31 cohorts) reaches the same near-null cohort verdict this row shows: high-vs-low CHD RR 1.00 (0.95, 1.05), stroke 0.98 (0.91, 1.05), T2D 0.90 (0.77, 1.06), CVD 0.95 (0.86, 1.06), with «No clear dose-response association was found for any of the outcomes». (Thorisdottir et al., 2023) Two caveats keep this from being independent corroboration. (i) Different estimand — this row’s cells are per-50 g linear slopes; Thorisdottir’s are high-vs-low category contrasts across a low-intake Nordic range (~12 g/day mean), so the numbers agree in direction but are not the same quantity. (ii) Shared cohort base (EPIC, ARIC, NHS/HPFS recur) and shared NNR SR team -> this is type-F / shared-source, NOT type-E independence. What Thorisdottir adds that this cohort-only series structurally cannot: a pooled RCT risk-factor arm — LDL-C -0.19 mmol/L (95% CI -0.27, -0.11) at ~120-150 g/day, robust to excluding soynut trials — i.e. legumes move a surrogate at RCT doses while the hard-endpoint cohort cells stay null. Full treatment -> Whole Grains Refined Grains and Pulses.
The pattern level sits one step up — the composite-score complement
This matrix operates at the food-group level — one row per food, a per-serving increment. A diet-quality score (Diet Quality Scores and Cardiovascular Risk, Mente’s PURE Healthy Diet Score) operates one level up: it collapses several protective foods into a single 0-6 composite and reads that against CVD/mortality. The two are different quantities (a per-50 g linear slope for one food vs an ordinal composite of six), so they cannot contradict — but laid together they pose the level-of-analysis question this fabric already owns.
- The composite carries a located knee the component rows do not. The pattern-score curve is steeper below the median and reaches diminishing returns near score ~4/6; the food-group rows here are mostly weak, heterogeneous, and monotone-over-range, with many null cells (the legume events-null the sharpest case). A clean knee at the pattern level over mostly-flat component rows is the empirical shape that makes the whole-vs-component question bite.
- Whether the composite adds signal beyond the sum of its rows is NOT adjudicated here — it is exactly Is the Food Category Doing Any Work’s question. Three readings stay live and are that page’s to weigh, not this one’s: genuine whole-diet synergy; mere aggregation that averages out per-food measurement noise (so the composite looks cleaner without being more causal); or healthy-user confounding loading onto the composite (the DQS page flags this and the ordinal-composite caveat itself). This matrix contributes the component-side evidence to that adjudication — that no single row reproduces the composite’s knee — and routes the verdict there rather than asserting one.
Gaps (type-G)
- All cells are observational (prospective cohorts) with self-reported intake; dietary
measurement error attenuates every gradient toward null, so the many
NS/VLcells are weak evidence of no effect, not evidence of absence -> Measurement Error in Dietary Assessment. - No cell isolates a substitution (what replaces the food) — the RRs are addition-to-diet
contrasts, so the Layer-3 replacement question (whole grains instead of refined) is unanswered
here.
G (needs a substitution/network meta-analysis). - No genetic (Mendelian-randomization) or feeding-trial evidence (superseded in part
[2026-10-06], see the sub-bullet) isolates red-meat or heme-iron -> T2D in either direction; the red/processed-meat T2D cells rest entirely on observational association. Heme iron is now held as an attributed observational channel for cancer, all-cause / cause-specific mortality (the latter via one large cohort’s mediation model), AND T2D — the T2D leg is a gold SR+MA (dietary heme-iron intake highest-vs-lowest RR 1.31, 95% CI 1.21-1.43, 4 prospective cohorts, I2=0%) (Zhao et al., 2012) -> Heme Iron and Cardiometabolic Risk (the cross-outcome nucleus). What is still missing for T2D is a natural experiment: a heme-iron -> T2D Mendelian-randomization or adherence-controlled feeding design remains the named decider for whether this cell is causal rather than confounded by red meat.G (needs a genetic/MR or feeding-trial test).- Refined
[2026-10-06]: a red/processed-meat intake MR is now held (Li 2024) and is uninformative at the needed precision (see the genetic-instrument section). The gap is now specific: a meat-intake MR that states its exposure unit, reports instrument F values and power against the observational effect size, handles sample overlap, and screens instruments for BMI/education pleiotropy, ideally pooled across MRs — what it would unlock: whether the red/processed-meat T2D and CHD cells are causal.
- Refined
- CASHED for the meat rows
[2026-09-08]— the series computes no absolute risk, so its RRs could not be sized against a drug comparator (Layer-1). Zhong 2020 now supplies a US-adult baseline (incident CVD 13.2 / 1000 py, ACM 15.8 / 1000 py) and a 30-year ARD for processed and unprocessed red meat -> Absolute-risk anchor for the meat rows above, closing the gap for those rows. It stays OPEN for the non-meat rows (whole grains, fruit, vegetables, dairy, nuts, SSB), which have no absolute anchor held; the mortality paper’s optimal-combined-intake -> 56% relative reduction figure is still relative-only.G (needs a baseline-risk source per stratum — non-meat rows).