Nucleus of the dairy cluster — the canonical owner of what dairy does to cardiometabolic
outcomes. The fermented-dairy CVD leg lives on Fermented Foods and Health and the
bone/fracture leg on the facet Dairy and Bone Health (do not duplicate either); this page owns
the non-fermented cardiometabolic axes — total dairy, milk, fat content, the milk-mortality controversy,
and dairy -> type-2 diabetes.
The through-line (two overlapping gold observational MAs): across ~940k people, dairy is broadly neutral for CVD and all-cause mortality (Guo, robust across its category decomposition), and modestly inverse for type-2 diabetes (Mishali alone — a single, Israel-Dairy-Board-funded MA that does not stratify by fat content, so the T2D leg is the weaker, discounted one) — with the effects small, heterogeneous, FFQ-confounded, and category-dependent. The one scary signal in public discourse — high milk drinking -> higher mortality — is a single-cohort confounding artifact, not a finding. No hard-outcome RCT exists; every magnitude here is observational. Symmetric-standards guard: this does not license full-fat dairy is health food (a dairy-industry-funded source pushes that framing) nor dairy fat is dangerous (the pooled signals resist it). Earn each verdict by category and endpoint.
Dairy is a type-B category — decompose before you rank
Milk, cheese, butter, yogurt and cream are almost different exposures at similar saturated-fat content, so a whole-milk finding must never read as a yogurt finding, nor a butter finding as a cheese finding -> Is the Food Category Doing Any Work. Guo 2017 is the worked case: it splits dairy five ways (total / high-fat / low-fat / milk / fermented+cheese+yogurt) and the answers differ by cell, while the aggregate «total dairy» hides them. The decision-relevant axes:
- Fat content — high-fat vs low-fat dairy, the guidance fault-line (guidance favours low-fat on SFA logic). Both were null in Guo — see below.
- Fermented vs unfermented — the only inverse CVD signals sit in fermented dairy/cheese (and even those are artifact-fragile) -> Fermented Foods and Health.
- Milk specifically — the controversy leg (Michaelsson) lives here.
Guo 2017 — dose-response: dairy is NEUTRAL for CVD and mortality
29 cohorts, 938,465 participants, dose-response (per-increment) meta-regression:
- «Total dairy intake (per 200 g/day) was not associated with the risk of all-cause mortality (Supplemental Figure 1; RR 0.99, 95% CI 0.96–1.03 …) … or CVD (Supplemental Figure 3; RR 0.97, 95% CI 0.91–1.02)» (Guo et al., 2017).
- High-fat AND low-fat dairy both null across mortality/CHD/CVD (high-fat CVD 0.93 [0.84-1.03]; low-fat CVD 0.98 [0.95-1.01]) — Guo ran no low-for-high substitution model, so this does not endorse the low-fat guidance default; it finds no difference in the observed contrast.
- Milk (per 244 g/d): null for mortality 1.00 (0.93-1.07), CHD 1.01, CVD 1.01 — but with extreme heterogeneity (mortality I2 = 97.4%).
- Overall: «neutral associations between dairy products and cardiovascular and all-cause mortality» (Guo et al., 2017).
The only inverse signals — fermented dairy (0.98/20 g mortality + CVD) and cheese (0.98/10 g CVD) — are artifact-fragile and cross-linked, not owned here -> Fermented Foods and Health. Yogurt was null (3 populations).
Butter is not one of Guo’s own pooled cells, but Guo reports an external butter MA (ref [10]): per 14 g/d, mortality RR 1.01 (1.00-1.03), no significant CHD/CVD/stroke association, and an inverse association with diabetes RR 0.96 (0.93-0.99) (Guo et al., 2017). So even the near- pure-fat, matrix-stripped category is at worst weakly-positive for mortality and neutral for CVD — the SFA-in-butter worry is not borne out at the whole-food level (a cited MA, secondhand through Guo).
Zhang 2025 quantifies butter directly, and finds a positive mortality signal — but on a different contrast [2026-09-02]
Zhang 2025 (NHS/NHSII/HPFS, 221,054 adults, up to 33 y, 50,932 deaths) is the direct butter->mortality study Guo’s cell lacked. Highest-vs-lowest total-butter intake carried a 15% higher total mortality (HR 1.15; 95% CI, 1.08-1.22) and higher cancer mortality (per 10 g/d HR 1.12; 95% CI, 1.04-1.20), with no statistically significant CVD-mortality association (Zhang et al., 2025). The signal is culinary-use split: butter as a spread (per 5 g/d HR 1.04; 95% CI, 1.02-1.05) drives it, while butter used for baking and frying showed no significant association — plausibly small quantities and misclassification, not safety (Zhang et al., 2025). Substituting 10 g/d of butter with plant-based oils was associated with a 17% lower total mortality (HR 0.83; 95% CI, 0.79-0.86).
This is an F-refinement of the Guo cell, NOT a contradiction of it — the comparators differ. Zhang himself locates the gap: the neutral butter MA «did not explicitly compare butter with alternative foods, instead implying a comparison to the typical Western diet, which is replete with other unhealthy components like refined grains, sugars, starches, and red meat» (Zhang et al., 2025). Matched parameters:
| Parameter | Guo’s cited butter MA (Pimpin) | Zhang 2025 | Same quantity? |
|---|---|---|---|
| butter -> total mortality | per 14 g/d, RR 1.01 (1.00-1.03) | Q4-vs-Q1, HR 1.15 (1.08-1.22) | NO — per-14g increment vs extreme-quartile contrast |
| implicit comparator | butter vs typical Western diet | within-cohort intake contrast (and a modelled butter->oil swap) | NO — Western-diet average vs a specified alternative food |
The two are consistent once matched: a near-flat per-gram slope against a junk-food-laden reference (Guo/Pimpin) and a positive extreme-quartile contrast that resolves into benefit only when butter is replaced by a named better food (Zhang’s substitution model) answer different questions. NOT-E (no confidence upgrade): Zhang shares the NHS/NHSII/HPFS cohorts, the Willett FFQ, and the Wang/Hu lab with the Harvard dairy-fat evidence this page already holds — agreement here is shared backing, not independent corroboration. And it stays observational: self-reported FFQ intake, residual confounding, and a statistical substitution model, not a feeding trial -> Saturated Fat Intake and Replacement (Butter vs plant oils at the food level).
Mishali 2019 — high-vs-low: modest inverse for T2D and CVD, concentrated in women
16 cohorts (T2D, 545,677) + 13 cohorts (CVD, 460,798), highest-vs-lowest intake:
- T2D: «Pooled results indicated an inverse association between the two (RR ¼ 0.897; 95%CI, 0.834–0.963; P < 0.01)» (Mishali et al., 2019) — ~10% lower.
- CVD: pooled RR 0.942 (0.892-0.994, P<0.05) (Mishali et al., 2019) — ~5% lower.
- The sex split is the paper’s distinctive move: «The pooled RR for men was not significant (RR ¼ 1.023; 95%CI, 0.91–1.15 …). The pooled RR for men and women to- gether was significant and showed a moderate protec- tive effect (RR ¼ 0.930; 95%CI, 0.87–0.99 …)»; women RR 0.864 (0.82-0.98) (Mishali et al., 2019). Per-outcome: women T2D 0.868, women CVD 0.837; men NS for both; between-group heterogeneity I2 = 86%.
Two guards on Mishali (symmetric standards): (1) it was financed by the Israel Dairy Board and all four authors are tied to it (two consultants, one paid by the Board for her work, one its Chief Health Officer) (Mishali et al., 2019), and it does not stratify by fat content while leaning on an «exonerates fat» framing — a directional-bias tell, so the effect estimate is discounted, not the finding deleted. (2) The women-only effect is a subgroup claim (route-b effect modification) with no identified mechanism (menopause-age and region moderators all came back null) — plausible but the false-positive-prone route; hold it as hypothesis-generating, not established.
Guo (neutral) vs Mishali (inverse): a distinction, NOT a tension
The two reach opposite headlines on largely the same cohorts. The parameter table dissolves the apparent clash:
| Parameter | Guo 2017 | Mishali 2019 | Same quantity? |
|---|---|---|---|
| Exposure contrast | per-increment dose-response slope (200 g/d) | highest-vs-lowest intake category | NO |
| Total dairy -> CVD | 0.97 (0.91-1.02), NS | 0.942 (0.892-0.994), sig | point estimates compatible; only significance differs |
| Fat stratification | high/low-fat analysed separately (both null) | none (total dairy/milk only) | NO |
| Sex | pooled + subgroups | central moderator; effect women-only | NO |
| Cohort set | 29 cohorts (938k) | 13 CVD / 16 T2D (461k/546k), overlapping subset | NOT independent |
| Funding | partly Global Dairy Platform / Dairy Research Institute / Dairy Australia (funders had no role) | Israel Dairy Board | — |
The clash is not joined — by not-joined check (ii) (different unit/contrast) plus the shared data.
The two estimates are not commensurable: Guo’s 0.97 is a per-200 g/d dose-response slope, Mishali’s
0.942 is a highest-vs-lowest category contrast (the parameter-table row marks these NO) — so their numeric
closeness is not a same-observable argument, it is a coincidence of scale. What actually makes this not a
tension is that the two run on largely the same cohorts: they cannot be two conflicting findings
when they are two analyses of one body of data. The divergence is therefore a framing / analytic-choice
difference, not an empirical one: Guo tests a per-unit slope (CI crosses 1) and decomposes to find the
aggregate hides null cells; Mishali contrasts extremes, aggregates across fat content, and — with an
industry sponsor — reads a ~5% inverse CVD estimate as «beneficial». So no tension is filed;
this is a structured distinction. It is also not type-E independent backing — the shared cohorts make
the agreement a re-pooling, not two independent routes.
The milk-mortality controversy is a confounding artifact
Public discourse’s scariest dairy claim — high milk -> higher death — comes from the Swedish cohorts: «higher milk consumption was associated with a doubling of mortality risk including CVD mortality in the cohort of women» (Guo et al., 2017). Run the artifact diagnostics -> The U-Shaped Association Artifact:
- The pooled milk -> mortality estimate is null (1.00); the only thing the Swedish cohort adds is heterogeneity (I2 = 97.4%). Removing it: I2 -> 70.1%, RR 0.99 (0.96-1.01).
- Confounding by lifestyle is named in-source: the high-milk Swedish drinkers had low education and «the highest milk drinkers had highest percentage of smokers and those living alone» (Guo et al., 2017).
- Same cohort drives the opposite (protective) fermented/cheese signal: «the inverse associations of fermented dairy and cheese with all-cause mortality or CVD disappeared after removing the study of Michaelsson et al.» (Guo et al., 2017). One confounded outlier cohort manufactures both poles of the debate — the cleanest demonstration that the milk-mortality signal has no diagnostic value.
Verdict: the milk-mortality signal does not survive the artifact check — it is single-cohort, confounder-explained, and adjudicated only by leave-one-out sensitivity (a weak check; no MR or genetic instrument in-source). Note Guo does not invoke Michaelsson’s own D-galactose hypothesis, so neither do we.
Mechanism (candidate, not demonstrated) and the matrix hinge
Both MAs propose the same channels without RCT confirmation: dairy minerals (Ca, K, Mg) lower total/LDL-C in short-term feeding studies (a neutral-total-dairy explanation); casein/whey lactotripeptides (IPP/VPP) inhibit ACE -> lower BP; milk fat raises HDL-C; the fermented-dairy signal may run through «the food matrix reducing lipid absorption and short chain fatty acids produced by the bacteria in the large intestine» (Guo et al., 2017). This is the dairy-matrix hypothesis — the same SFA behaving differently inside cheese vs butter — held as a mechanism to weigh, not a finding; the SFA verdict itself is deferred to Saturated Fat Intake and Replacement and Does Reducing Saturated Fat Reduce Cardiovascular Events.
controlled-feeding matrix mechanism (cheese vs butter on LDL) that adds a surrogate-LDL directional
leg; the hard-outcome matrix claim stays a candidate (below).
### The food-matrix concept and the cheese-vs-butter LDL leg (Thorning 2017 — scoping review, directional-only)
Thorning et al. 2017 names the mechanism the hypothesis above was invoking without a home: `«The nature
of the food structure and the nutrients therein (i.e., the food matrix) will determine the nutrient
digestion and absorption, thereby altering the overall nutritional properties of the food.»`
(Thorning et al., 2017). The **directional** leg the matrix claim was
missing is a controlled-feeding contrast: in Tholstrup's whole-diet trial, the diets were balanced for fat,
protein and lactose «but not of calcium» (Thorning et al., 2017), and `«butter caused significantly higher LDL-cholesterol concentrations ... than cheese did,
whereas the effect of milk was intermediate»` (Thorning et al., 2017)
— so the same milk fat raised LDL more as butter than as cheese. Thorning reads this as «These studies
suggested that protein and lactose were not the determinants of the difference between cheese and butter
on blood lipids.» (Thorning et al., 2017), pointing instead to dairy calcium, which the diets did not
balance.
This is the mechanism-directional rule at work: a **surrogate** endpoint (LDL-C, not a hard outcome),
admitted to inform DIRECTION only, discounted, and marked as mechanism.
Two guards hold the weight down and are why `confidence:` stays **medium**, governed by the gold sources
above and NOT lifted by Thorning. First, the source states its own limit — `«It has to be clarified
whether the effects of the various mechanisms have a magnitude to impose an effect in the whole diet
setting.»` (Thorning et al., 2017) — so whether the matrix moves a
patient-important outcome in a real diet is explicitly open. Second, independence and sponsorship. Thorning shares
authors with Guo 2017 (Astrup and Soedamah-Muthu; Givens very likely the same person), so it cannot back
Guo independently. The underlying expert workshop was funded by the European Milk Forum, and authors
disclose dairy-industry funding or consultancy (Astrup: Arla Foods, the Danish Dairy Research Foundation;
Givens: the Dairy Council). The source states «The study sponsor had no influence on the meeting program
or the selection of speakers, was not present at the meeting, and was not involved in the production of
the manuscript.» (Thorning et al., 2017) A milk-industry-funded scoping review (of a consensus
workshop), sharing authors with the gold anchor, is **not** independent robustness backing (NOT-E — no `[E-independent]` with any gold
source here); it is a mechanistic framing, credited and labelled at point of use.
</div>
## Off-axis endpoint: a minor protective dental-erosion association (cashes HELD-DAIRY-1)
SACN 2015 reports, in its dental-erosion section, that «Consumption of yoghurt and milk products was
negatively associated with the incidence of erosive wear»
(Scientific Advisory Committee on Nutrition, 2015) (cashes `HELD-DAIRY-1`). **Two guards apply
and are kept:** (1) erosive wear is a **dental** endpoint in a carbohydrates report — it is off-axis to
the cardiometabolic outcomes this page owns and must **not** corroborate any matrix or CVD claim; (2)
the same dataset found **no** association for carbonated soft drinks and erosion, which is
counter-intuitive enough to suggest the confounding structure is not simple — read the dairy result with
equal scepticism. A minor, low-weight note, not a decision-change.
## Whole-fat dairy as a protective PATTERN component — PURE (Mente 2023), and why it does not overturn the neutral verdict
PURE's healthy-diet-score analysis (Mente 2023; 147,642 + 5 replication studies) is the one large source
that reads dairy as **actively protective** rather than neutral — and it is the score's distinctive
element, since «dairy (mainly whole-fat)» is one of only six protective foods it counts, where other
diet scores favour low-fat dairy or exclude it -> [[Diet Quality Scores and Cardiovascular Risk]]. Mente
states «our findings in PURE showed that dairy foods, especially whole-fat dairy, may be protective
against risk of hypertension and metabolic syndrome», and «intakes of dairy (up to 185 g/day; or \~two
servings/ day, mainly from whole-fat dairy) can be included with other beneficial foods as part of a
healthy diet» (Mente et al., 2023).
**Three guards keep this from upgrading the page's verdict past neutral, and they are the point:**
- **Different, weaker contrast — not commensurable with Guo (no tension filed).** Mente's dairy signal
is dairy *as one component of a highest-vs-lowest whole-diet score*, hopelessly confounded with the
other five protective foods and with the healthy-user gradient; Guo is a dairy-specific per-200 g/day
dose-response. These are **not the same quantity** (parameter-table row: NO), so PURE's protective
reading and Guo's neutral slope are a **distinction, not a joined tension** — and PURE lands
*compatibly* with Guo where it matters: dairy is not a food to avoid.
- **Not independent (not type-E).** PURE shares authors (Mente, Dehghan) and cohort infrastructure with
the Dehghan PURE macronutrient work; a shared confounding/measurement structure moves them together.
So this is not a second independent witness for a dairy benefit — it is the same research programme.
- **Directional-sponsor tell (symmetric standards, same as Mishali).** PURE is partly funded by «an
unrestricted grant from Dairy Farmers of Canada and the National Dairy Council (U.S.)»
(Mente et al., 2023), and the whole-fat-dairy-is-protective
reading is exactly where such a sponsor would push. The estimate is discounted, not deleted.
**Net:** PURE strengthens the *neutral-to-mildly-favourable* end of this page's range for whole-fat
dairy within a whole-diet pattern; it does **not** license a standalone *full-fat dairy is
cardioprotective* claim. The hypertension/metabolic-syndrome mechanism it invokes remains a candidate
channel, not a demonstrated one.
## Confidence, decision-relevance, and gaps
- **`confidence: medium`** — two gold observational MAs converging on *small effects* (neutral CVD/
mortality; modest inverse T2D), but the backing is **not independent** (shared cohorts), FFQ-confounded,
absolute risks unrecoverable ([[Relative vs Absolute Risk]] — every figure is RR/OR with no baseline
rate), and no hard-outcome RCT exists. The `medium` rests on the *neutrality* verdict (robust across
both MAs and Guo's category decomposition), not on any protective claim.
- **Decision-change (per stratum):** for someone with the big rocks already handled
([[Layer 1 - Ranking Interventions for a Stratum]]), **dairy is not a lever to worry about** — the
milk-mortality scare is an artifact, and total/full-fat vs low-fat dairy is a wash for CVD in the
observed data. Fermented dairy carries the (weak, artifact-fragile) favourable signal; butter is the
near-pure-fat category with an inverse *diabetes* signal but a weak positive mortality one. This is a
*small lever*; *attention is an anti-signal* applies (dairy is discussed far out of proportion to its
established effect).
- **Gaps (G):**
- ~~**Milk -> fracture / bone** — the milk-fracture paradox is unheld.~~ **CASHED 2026-08-06 by Malmir
2019** — the bone/fracture endpoint now lives on the cluster facet [[Dairy and Bone Health]]: milk/
dairy is **null** for osteoporosis and hip fracture in cohorts (protective only in reverse-causation-
prone cross-sectional/case-control designs; milk trends to +9%/200 g harm in cohorts, Michaelsson-
driven — the same artifact cohort as this page's milk-mortality scare). Dairy is not a fracture lever.
- ~~**Dairy -> cognition / dementia** — unheld.~~ **CASHED 2026-09-05 by Villoz 2024** — the
cognition/dementia endpoint now lives on the cluster facet [[Dairy and Cognitive Decline]]: dairy is
**null** for cognitive decline/dementia in the reasonably-healthy Western stratum (highest-vs-lowest RR
0.94, 0.82-1.07), the one dose-response "nadir" (\~150 g/d) is an Asia/Europe intake-range pooling
artifact, and milk-alone/cheese-alone are null. Same type-B decomposition + FFQ-confounding as this
page. Dairy is not a cognition lever.
- **Dairy -> cancer, opposing directions** (probable-protective colorectal, probable-increased prostate)
— direction held on [[Red and Processed Meat and Cancer]]/WCRF; magnitudes unheld here. Name both,
never net.
- **Sex as effect-modifier (Mishali)** — a single-MA subgroup claim, no mechanism; needs replication
before it drives a sex-specific recommendation (route-b bar).
- **`G (needs aggregation)`:** a pooled dairy-fat -> CVD-events effect that adjusts for the Michaelsson
artifact across designs — a magnitude the fabric cannot compute from these two non-commensurable MAs.
## Self-critique `[run 2026-08-06, before commit]`
- **No dairy halo, no dairy scare — both checked.** The neutral verdict is stated as neutral (not
*safe/beneficial*); the one protective source (Mishali) is discounted for its Israel-Dairy-Board
sponsorship and no-fat-split design; the milk-mortality scare is dismantled as a single-cohort artifact
rather than either endorsed or ignored.
- **Independence not laundered.** Guo and Mishali share cohorts, so their agreement is filed as
re-pooling, not type-E; the parameter table carries the shared-cohort row explicitly. No `[E-independent]`
claimed.
- **Not-joined check ran.** Guo-neutral vs Mishali-inverse resolves to a distinction (near-identical
point estimates, different contrast + framing) — no tension manufactured to fill a domain-opener quota.
- **Michaelsson artifact stays within what Guo states** — leave-one-out sensitivity + lifestyle
confounding are Guo's; the D-galactose mechanism is Michaelsson's own and is explicitly NOT imported.
- **Sex subgroup held to route-b.** The women-only effect is marked a false-positive-prone effect-
modification claim, not an established stratum rule.
- **HELD-DAIRY-1 cashed with both guards** — dental endpoint quarantined as off-axis, soft-drink-null
scepticism carried.
- **Blind self-critique catches applied `[2026-08-06]`:** (a) the butter direction (inverse-diabetes /
weak-positive-mortality) was asserted in the verdict with no on-page support — now sourced in the Guo
section (Guo's cited external butter MA); (b) the not-joined justification wrongly leaned on the
non-commensurable numeric closeness (0.97 vs 0.94) — re-anchored on check (ii) different-contrast +
shared-cohorts; (c) the T2D-inverse headline now carries its single-source / industry-funding caveat
inline rather than three sections later.
- **PURE whole-fat-dairy section added `[2026-08-25]`, three guards checked.** The protective-dairy
reading is (i) filed as a *distinction* from Guo not a tension (different contrast — score component vs
dose-response, parameter-table NO); (ii) marked non-independent (shared PURE authors/infrastructure with
Dehghan — no `[E-independent]`); (iii) discounted for the Dairy-Farmers-of-Canada / National-Dairy-
Council funding tell, same treatment as Mishali's Israel-Dairy-Board sponsorship. The neutral verdict is
not upgraded — no full-fat-dairy halo introduced.
<div class="recent-update" data-last-updated="2026-10-06">
## Refinement — the DIfE/Boeing 12-food-group series (2026-08-28)
The series places dairy as a **cardiometabolic-marker lever, not a mortality lever**: inverse for
T2D (0.97, 0.94-0.99 per 200 g/d, MODERATE) and hypertension (0.95, 0.94-0.97, LOW), but **null for
all-cause mortality** (0.98, 0.93-1.03), null for CHD/stroke, and **positive for heart failure**
(1.08, 1.01-1.15, but from a single study, n = 1). Low- vs high-fat dairy showed no significant difference for these endpoints. The
T2D/HTN benefit not carrying through to mortality is the divergence to hold onto.
(Schwingshackl et al., 2017)
(Schwingshackl, Schwedhelm, et al., 2017)
(Schwingshackl, Schwedhelm, Hoffmann, Lampousi, et al., 2017)
(Bechthold et al., 2017) Full grid
-> [[Food Groups and Health Outcomes - A Dose-Response Matrix]].
</div>
<div class="recent-update" data-last-updated="2026-10-07">
## Gijsbers 2016 — the dairy -> T2D DOSE-RESPONSE by subtype: the inverse signal is yogurt/low-fat, milk and cheese are null `[2026-09-05]`
Gijsbers is the dedicated dairy -> T2D **dose-response** MA (22 prospective cohorts, 579,832 individuals,
43,118 incident T2D cases) — the first held source that runs the *per-serving shape* by dairy subtype
rather than a whole-dairy point estimate. Its verdict is a clean [[Is the Food Category Doing Any Work]]
case: *dairy* is not one exposure for T2D, and the modest total-dairy inverse is carried by two
subtypes while the others are flat.
- **Total dairy: `«RR: 0.97 per 200-g/d increment; 95% CI: 0.95, 1.00; P = 0.04; I2 = 66%»`, linear
over the studied range** (median total-dairy intake 111-400 g/d across cohorts)
(Gijsbers et al., 2016) — a 3% lower risk per
200 g/d, borderline, heterogeneous.
- **Low-fat dairy: `«RR: 0.96 per 200 g/d; 95% CI: 0.92, 1.00; P = 0.072»`** — a *suggestive but
similar* linear inverse, NOT significant at 0.05
(Gijsbers et al., 2016).
- **High-fat dairy: null** (RR 0.98 per 200 g/d; 95% CI 0.93, 1.04; P = 0.52)
(Gijsbers et al., 2016).
- **Milk: null across the board.** `«Total milk intake ... was not associated with T2D risk (RR: 0.97
per 200 g/d; 95% CI: 0.93, 1.02; P = 0.25)»`; low-fat milk RR 1.01 (0.97-1.05) and high-fat milk
RR 0.99 (0.88-1.11) both null too
(Gijsbers et al., 2016).
- **Cheese: null** (RR 1.00 per 10 g/d; I2 = 62%) — the only cheese signal is a 2-study men-only +5%
per 10 g/d, not stratum-stable (Gijsbers et al., 2016).
- **Yogurt: a NONLINEAR inverse with a located plateau.** `«Yogurt ... was non-linearly inversely
related to T2D, showing a 14% lower risk for an intake of 80 g/d (RR: 0.86 compared with 0 g/d;
95% CI: 0.83, 0.90; P < 0.001)»`, and `«The risk did not further decrease at higher intake amounts
of yogurt >80 g/d»` (Gijsbers et al., 2016) — the
benefit is acquired by \~80 g/d (about one small pot) and plateaus; more buys nothing.
**This is the type-B decomposition with T2D-specific numbers:** the whole-dairy 0.97/200 g/d hides a
yogurt-and-low-fat inverse against a milk/cheese/high-fat null. A whole-milk or cheese eater gets no
T2D signal from this data; a yogurt eater gets a small one that saturates early -> the yogurt leg is
cross-linked to [[Fermented Foods and Health]] (it cashes that page's yogurt -> T2D dose-response gap).
### The total-dairy leg is ECHO, not independent corroboration — the subtype shape is the actual add `[2026-09-05]`
Gijsbers's total-dairy slope is numerically identical to the one the fabric already held from the DIfE
12-food-group series, and the two re-pool the **same** canonical dairy-T2D cohorts (Chen's US cohorts,
EPIC-InterAct, Ericson/Malmo, Whitehall II). Matched:
| Parameter | Gijsbers 2016 | Schwingshackl 2017 (held) | Same quantity? |
|---|---|---|---|
| Exposure | total dairy, per 200 g/d | total dairy, per 200 g/d | **YES** |
| Outcome | incident T2D | incident T2D | **YES** |
| Effect form | linear dose-response RR | linear dose-response RR (NutriGrade) | **YES** |
| Estimate | 0.97 (0.95-1.00) | 0.97 (0.94-0.99) | **YES — near-identical** |
| Constituent cohorts | 22 dairy-T2D cohorts | same dairy-T2D cohort pool, re-pooled | **overlapping — NOT independent** |
**So on total dairy this is type-F/echo (re-pooling), NOT type-E:** the agreement raises no confidence
beyond Schwingshackl because it is the same data pooled by a second team — no `[E-independent]` claimed.
It is **also not independent of this page's own [[Fermented Foods and Health]] / CVD leg**: Lieke Gijsbers
is a co-author of the held Guo 2017 dairy-CVD MA, and the senior author (Soedamah-Muthu) ran the prior
dairy-CVD and dairy-hypertension dose-response MAs this project explicitly re-used the methods of. The
genuine beyond-summary add is the **subtype dose-response** (yogurt nonlinear knee, milk/cheese/high-fat
null, low-fat borderline) — which no held source carried.
**Directional-sponsor lineage tell (symmetric standards, same treatment as Mishali/PURE).** The study
itself was Wageningen-University-funded with no sponsor role, but three senior authors disclose prior
dairy-industry funding (Global Dairy Platform, Dairy Research Institute, Dairy Australia, Dutch Dairy
Association) (Gijsbers et al., 2016). The whole
dairy-T2D-protective literature clusters in industry-adjacent groups (Mishali = Israel Dairy Board;
PURE = Dairy Farmers of Canada; the Soedamah-Muthu lineage here; Guo 2017 itself, partly funded by the
Global Dairy Platform, Dairy Research Institute and Dairy Australia (Guo et al., 2017)) — a **uniformity tell** about the
field, not a refutation of any one estimate. The effect is discounted, not deleted.
**The genetic-instrument leg is now HELD — see the Vissers MR section below.** Gijsbers noted secondhand
a Mendelian-randomization study finding `«no evidence of an association between milk and T2D»`
(Gijsbers et al., 2016). That gap is **cashed by
Vissers 2019** (EPIC-InterAct MR, held 2026-09-15) — a genetic instrument reaching the same milk-null.
~~~~ **CASHED 2026-09-15 by Vissers 2019**: the
gap named "genetic-instrument corroboration of the milk -> T2D null" is filled by the held Vissers MR.
Bergholdt 2015 remains a *distinct* Danish MR (Vissers cites it secondhand, OR 0.99 per 250 g/wk), so it
is not an independent second witness and is not separately acquired.
**Net effect on this page's T2D leg:** the through-line's *modestly-inverse-for-T2D (Mishali alone)*
reading is upgraded — the inverse total-dairy signal is now convergent across Gijsbers
(dose-response), Schwingshackl (food-groups) and Mishali (high-vs-low), but all three re-pool one
cohort set, so the convergence is robustness-of-pooling, not independence. And the signal is **carried
by yogurt and low-fat dairy**, with milk, cheese and high-fat dairy null and the milk-null now
**MR-corroborated by the held Vissers 2019 (section below)**. Still a small lever; still observational for
the yogurt/low-fat arm; still no hard-outcome RCT.
### Self-critique — Gijsbers append `[run 2026-09-05, before commit]`
- **Overclaim check.** The total-dairy 0.97 is written as *borderline* (CI to 1.00, P=0.04) and low-fat
0.96 as *suggestive-NS* (P=0.072) — neither dressed as a firm effect. The yogurt knee is named an
*observational-spline* soft knee, not an intervention target. The whole section is prefaced
observational/FFQ; no causal language on any RR.
- **Laundered-independence check — the load-bearing risk here.** The total-dairy leg is explicitly a
parameter table marked ECHO/NOT-E: shared cohorts with Schwingshackl (identical 0.97/200g), Gijsbers
co-authored held Guo, Soedamah-Muthu lineage. No `[E-independent]` token minted. The only claimed
independence signal (Bergholdt MR) is held secondhand and routed to `[AWAITS]`, not asserted.
- **Fake-tension check.** No tension filed — this refines the existing T2D leg (F), it does not clash
with Guo/Mishali. The Guo(neutral-CVD) vs Mishali(inverse-T2D) distinction upstream is untouched.
- **Halo / directional-sponsor.** Prior dairy-industry funding of three senior authors is flagged as a
field-wide uniformity tell (with Mishali/PURE), discounted symmetrically, not used to delete the effect.
- **Open loop.** All magnitudes are observational FFQ, never graded against a realized T2D outcome; the
live-culture and measurement-error gaps are named, not closed.
</div>
## Vissers 2019 — the genetic-instrument (MR) leg: milk -> T2D is causally null, but only for MILK `[2026-09-15]`
The observational milk-null (above, and the milk-mortality section) always carried an alternative
reading: residual confounding could hide a real protective *or* harmful milk effect. A Mendelian
randomization (MR) study breaks that ambiguity, because a birth-fixed genotype cannot be confounded by
the lifestyle that tracks milk drinking. Vissers 2019 is that leg — EPIC-InterAct, the
lactase-persistence SNP **rs4988235 (LCT-12910C>T)** as an instrumental variable for milk intake,
21,820 participants including 9,686 incident T2D cases.
- **The causal estimate is null.** Genetically-predicted milk intake, 2SLS IV analysis: `HR 0.99 per
15 g/day (95% CI 0.93, 1.05, I2 = 44%)`; reduced-form per LP allele `HR 0.99 (95% CI 0.94, 1.04)`
(Vissers et al., 2019). The
authors conclude «In this European case-cohort study including 9,686 incident type 2 diabetes cases,
we did not find evidence for a causal relationship between milk intake and diabetes.»
(Vissers et al., 2019) Bootstrap,
LP-dominant, and HbA1c-exclusion sensitivity analyses all left the null unchanged.
- **This is what breaks the confounding alternative for the milk arm.** Vissers reads the observational
inverse milk signal directly: «These findings combined support the notion that the inverse association
between milk intake and diabetes as seen in some studies (2) is likely caused by residual confounding»
(ref 2 = the held Gijsbers 2016 MA) (Vissers et al., 2019). So the composite (observational null-to-weak-inverse + MR null) is
stronger than either alone: the residual-confounding story that would rescue a hidden milk effect is
now the *less* supported explanation.
### The instrument tests MILK ONLY — the boundary that stops the null being over-read
rs4988235 raises milk intake and essentially nothing else in the dairy group. Per LP allele: milk
`β 17.1 g/day (95% CI 10.6, 23.6)`, «but not with intake of other dairy products» — yogurt (β 2.2,
95% CI -0.5, 4.9; p=0.11) and cheese (β -0.09; p=0.85) both individually null
(Vissers et al., 2019). Instrument
strength is not the issue: `F statistic of the model predicting milk intake was 74.0`
(Vissers et al., 2019). So the MR is a
clean test of the milk arm and **silent on the yogurt/low-fat arm** — Vissers says so: «No conclusion
can be drawn regarding causality of the relationship of dairy products other than milk with diabetes.»
(Vissers et al., 2019) This matters for
the decision: the yogurt/low-fat inverse signal that Gijsbers isolates is exactly where the
observational protective arm lives, and this MR *cannot* refute it. The milk-MR-null and the
yogurt-observational-signal are about different exposures; reading the first as sinking the second
would be the category error the page's type-B decomposition exists to prevent.
### Why this is type-F refinement, NOT type-E independent corroboration
The tempting move is to call MR-agrees-with-observational an independent-route (type-E) convergence
that should raise confidence. It fails the strict-E test on two counts, so it is scored **F** (the MR
refines the milk claim by removing the confounding alternative), not E:
| Parameter | Observational milk-null (Gijsbers 2016 MA) | Vissers 2019 MR | Same quantity / independent? |
|---|---|---|---|
| Exposure | reported milk intake (FFQ) | genetically-predicted milk intake (rs4988235 IV) | related, different identification |
| Outcome | incident T2D | incident T2D | yes — same outcome |
| Participant base | 22-cohort observational pool incl. EPIC cohorts | EPIC-InterAct (8 countries) | **overlapping, NOT disjoint** |
| Citation relation | — | cites Gijsbers (ref 2) as the MA it is testing | **cite-antecedent — not independent** |
Two disqualifiers, either alone fatal to E: (i) **shared sample** — Vissers runs on EPIC-InterAct,
whose own observational dairy-T2D analysis (Sluijs 2012) sits inside the observational pool Gijsbers
meta-analysed, so the two partly rest on the same participants; (ii) **cite-antecedent** — Vissers
names it as the finding it is testing. Independent routes must be disjoint *and* not cite each
other; this pair is neither. No `[E-independent]` token is minted. The genuine value is F: the MR adds
a confounding-proof identification the observational leg structurally lacks. The E-vs-F call is the
wiki's own strict-E test on the two held sources, not a claim either source makes.
### The MR's own limits (symmetric-standards read)
- **Exclusion restriction is uncheckable.** «The third IV assumption is that the IV solely influences
the outcome via a causal pathway that includes the exposure of interest, and cannot be checked»
(Vissers et al., 2019). rs4988235 tracks
a broader LP-associated dietary pattern (lower fruit, wine, cereal; higher potatoes) plus larger hip
circumference and lower LDL — the authors judge these «modest» and unlikely to have caused the null,
uncorrected for multiple testing.
- **The null does not exclude a small effect.** «Due to aforementioned limitations of this study, we
cannot exclude the possibility that a small effect of milk intake on diabetes risk is present, despite
our null finding.» (Vissers et al., 2019)
- **G-gap (named, not silent):** the yogurt / low-fat-dairy causal leg — where the observational
protective signal actually concentrates — is **untestable by rs4988235** (the SNP does not move
yogurt or cheese intake). Closing it would need a genetic instrument specific to fermented-dairy
intake, or a hard-outcome RCT; neither is held. — no fermented-dairy MR identified.
### Self-critique — Vissers append `[run 2026-09-15, before commit]`
- **Overclaim check.** The MR is stated as a null on the *milk* arm only, with the authors' own
"cannot exclude a small effect" caveat carried inline. No causal language extended to yogurt/cheese;
the boundary is stated three times (instrument specificity, the authors' scope wall, the G-gap).
- **Laundered-independence check — the load-bearing risk here.** The E-vs-F call is the whole point of
the section: a parameter table marks the shared EPIC-InterAct sample and the Gijsbers cite-antecedent,
and NO `[E-independent]` token is minted. The convergence is scored F (identification added), not E
(independent witness), exactly to avoid laundering shared-cohort agreement into false robustness.
- **Fake-tension check.** No tension filed — the MR-null and the observational-null agree on the milk
arm (F-refinement), and are silent-not-opposed on the yogurt arm (different exposure, not a clash).
- **Bergholdt handling.** The gap is cashed by the held Vissers, and Bergholdt
is explicitly kept as a distinct secondhand-cited Danish MR, not promoted to a held independent
witness — so no phantom second source is counted.
- **Open loop.** MR identifies a causal null for milk; it grades nothing against a realized outcome for
the yogurt/low-fat arm, which stays observational. The gap is named, not closed.
## ReferencesBechthold, A., Boeing, H., Schwedhelm, C., Hoffmann, G., Knüppel, S., Iqbal, K., De Henauw, S., Michels, N., Devleesschauwer, B., Schlesinger, S., & Schwingshackl, L. (2017). Food groups and risk of coronary heart disease, stroke and heart failure: A systematic review and dose-response meta-analysis of prospective studies. Critical Reviews in Food Science and Nutrition, 59(7), 1071–1090. https://doi.org/10.1080/10408398.2017.1392288
Gijsbers, L., Ding, E. L., Malik, V. S., de Goede, J., Geleijnse, J. M., & Soedamah-Muthu, S. S. (2016). Consumption of dairy foods and diabetes incidence: a dose-response meta-analysis of observational studies. The American Journal of Clinical Nutrition, 103(4), 1111–1124. https://doi.org/10.3945/ajcn.115.123216
Guo, J., Astrup, A., Lovegrove, J. A., Gijsbers, L., Givens, D. I., & Soedamah-Muthu, S. S. (2017). Milk and dairy consumption and risk of cardiovascular diseases and all-cause mortality: dose–response meta-analysis of prospective cohort studies. European Journal of Epidemiology, 32(4), 269–287. https://doi.org/10.1007/s10654-017-0243-1
Mente, A., Dehghan, M., Rangarajan, S., O’Donnell, M., Hu, W., Dagenais, G., Wielgosz, A., A. Lear, S., Wei, L., Diaz, R., Avezum, A., Lopez-Jaramillo, P., Lanas, F., Swaminathan, S., Kaur, M., Vijayakumar, K., Mohan, V., Gupta, R., Szuba, A., … Yusuf, S. (2023). Diet, cardiovascular disease, and mortality in 80 countries. European Heart Journal, 44(28), 2560–2579. https://doi.org/10.1093/eurheartj/ehad269
Mishali, M., Prizant-Passal, S., Avrech, T., & Shoenfeld, Y. (2019). Association between dairy intake and the risk of contracting type 2 diabetes and cardiovascular diseases: a systematic review and meta-analysis with subgroup analysis of men versus women. Nutrition Reviews, 77(6), 417–429. https://doi.org/10.1093/nutrit/nuz006
Schwingshackl, L., Hoffmann, G., Lampousi, A.-M., Knüppel, S., Iqbal, K., Schwedhelm, C., Bechthold, A., Schlesinger, S., & Boeing, H. (2017). Food groups and risk of type 2 diabetes mellitus: a systematic review and meta-analysis of prospective studies. European Journal of Epidemiology, 32(5), 363–375. https://doi.org/10.1007/s10654-017-0246-y
Schwingshackl, L., Schwedhelm, C., Hoffmann, G., Knüppel, S., Iqbal, K., Andriolo, V., Bechthold, A., Schlesinger, S., & Boeing, H. (2017). Food Groups and Risk of Hypertension: A Systematic Review and Dose-Response Meta-Analysis of Prospective Studies. Advances in Nutrition, 8(6), 793–803. https://doi.org/10.3945/an.117.017178
Schwingshackl, L., Schwedhelm, C., Hoffmann, G., Lampousi, A.-M., Knüppel, S., Iqbal, K., Bechthold, A., Schlesinger, S., & Boeing, H. (2017). Food groups and risk of all-cause mortality: a systematic review and meta-analysis of prospective studies ,. The American Journal of Clinical Nutrition, 105(6), 1462–1473. https://doi.org/10.3945/ajcn.117.153148
Scientific Advisory Committee on Nutrition. (2015). Carbohydrates and Health. https://www.gov.uk/government/publications/sacn-carbohydrates-and-health-report
Thorning, T. K., Bertram, H. C., Bonjour, J.-P., de Groot, L., Dupont, D., Feeney, E., Ipsen, R., Lecerf, J. M., Mackie, A., McKinley, M. C., Michalski, M.-C., Rémond, D., Risérus, U., Soedamah-Muthu, S. S., Tholstrup, T., Weaver, C., Astrup, A., & Givens, I. (2017). Whole dairy matrix or single nutrients in assessment of health effects: current evidence and knowledge gaps ,. The American Journal of Clinical Nutrition, 105(5), 1033–1045. https://doi.org/10.3945/ajcn.116.151548
Vissers, L. E. T., Sluijs, I., van der Schouw, Y. T., Forouhi, N. G., Imamura, F., Burgess, S., Barricarte, A., Boeing, H., Bonet, C., Chirlaque, M.-D., Fagherazzi, G., Franks, P. W., Freisling, H., Gunter, M. J., Quirós, J. R., Ibsen, D. B., Kaaks, R., Key, T., Khaw, K. T., … Wareham, N. J. (2019). Dairy Product Intake and Risk of Type 2 Diabetes in EPIC-InterAct: A Mendelian Randomization Study. Diabetes Care, 42(4), 568–575. https://doi.org/10.2337/dc18-2034
Zhang, Y., Chadaideh, K. S., Li, Y., Li, Y., Gu, X., Liu, Y., Guasch-Ferré, M., Rimm, E. B., Hu, F. B., Willett, W. C., Stampfer, M. J., & Wang, D. D. (2025). Butter and Plant-Based Oils Intake and Mortality. JAMA Internal Medicine, 185(5), 549. https://doi.org/10.1001/jamainternmed.2025.0205