Milk, cheese, butter and yoghurt carry similar saturated fat, yet the outcomes they touch — heart disease, early death, diabetes, bone fracture, cancer, dementia — point in different directions. Read one food and one outcome at a time, and almost every effect turns out small.
For heart disease, stroke and early death, dairy is close to a wash. Eating more, or less, barely moves the needle, and full-fat versus low-fat is not the fault-line that guidance implies. The one diabetes signal is real but small, and it belongs to yoghurt and low-fat dairy — milk, cheese and high-fat dairy show nothing, and a genetic study confirms milk has no causal effect. The scary milk shortens your life result is a statistical artifact, traceable to a single Swedish cohort that takes the association with it when removed. Dairy does not build bone or protect the brain either. Cancer is the one endpoint this evidence cannot settle. None of this makes dairy a lever worth much worry — the noise around it runs far ahead of the effect.
Dairy is several foods, not one — judge each category and endpoint on its own
The single word dairy hides categories that behave almost like different foods. Split them three ways before ranking anything (Is the Food Category Doing Any Work):
- Matrix-intact vs matrix-stripped — cheese and yoghurt keep the milk’s structure; butter and cream strip it down toward the fat. The saturated fat is similar; the food around it is not.
- Fermented vs unfermented — yoghurt and cheese carry live or once-live cultures; milk does not.
- Full-fat vs low-fat — the axis guidance leans on, and (as below) the one that turns out to matter least.
Guo’s meta-analysis is the worked case: it splits dairy five ways and the answers differ by cell, while the aggregate total dairy hides them (Guo et al., 2017). So a whole-milk finding must never be read as a yoghurt finding, nor a butter finding as a cheese finding. Every claim below attaches to a named category and a single endpoint.
For heart disease and early death, dairy barely moves the needle
Across roughly 940,000 people, Guo found total dairy neutral for both death and cardiovascular disease: per 200 g/day, all-cause mortality RR 0.99 (95% CI 0.96-1.03) and CVD RR 0.97 (0.91-1.02) (Guo et al., 2017). Milk alone was null too (per 244 g/day, RR 1.00 for mortality), though with extreme study-to-study heterogeneity (I2 = 97.4%) that the next section explains.
Full-fat and low-fat dairy were both null — high-fat CVD RR 0.93 (0.84-1.03), low-fat 0.98 (0.95-1.01) (Guo et al., 2017). Guo ran no direct swap of one for the other, so this does not endorse the low-fat guidance default; it simply finds no difference in the observed contrast. The large food-group meta-analysis series agrees where it overlaps: dairy was null for all-cause mortality (RR 0.98, 0.93-1.03) and for coronary heart disease and stroke (Schwingshackl et al., 2017). The one exception is a small heart-failure signal: RR 1.08 (1.01-1.15) per 200 g/day from a single study, while the highest-vs-lowest comparison across three studies was null (RR 1.00, 0.90-1.10) (Bechthold et al., 2017). It is low certainty and observational, and named here rather than inflated into a headline.
Butter, the near-pure-fat category, is at worst weakly positive
Butter is the case where the saturated-fat worry should bite hardest. It does not, much. A butter meta-analysis reported through Guo found per 14 g/day only RR 1.01 (1.00-1.03) for mortality, no significant heart-disease association, and an inverse signal for diabetes (RR 0.96, 0.93-0.99) (Guo et al., 2017).
A direct study (Zhang 2025, 221,054 adults) later found highest-versus-lowest butter intake carried 15% higher total mortality (HR 1.15, 1.08-1.22) with no significant CVD-mortality link (Zhang et al., 2025). The two are consistent once you match the comparator. The near-flat per-gram slope measures butter against a junk-food-heavy Western diet; the positive contrast turns into benefit only when butter is replaced by a named better food — swapping 10 g/day of butter for plant oils tracked 17% lower mortality (HR 0.83, 0.79-0.86) (Zhang et al., 2025). The lesson is a substitution one, not a verdict on butter in isolation (Saturated Fat Intake and Replacement).
The one source that reads dairy as actively protective — and why it does not overturn neutral
PURE’s healthy-diet analysis (Mente 2023) counts whole-fat dairy among only six protective foods and suggests up to ~185 g/day «can be included with other beneficial foods as part of a healthy diet» (Mente et al., 2023). Read it with three guards. It measures dairy as one component of a whole-diet score, tangled with the other five protective foods and the healthy-eater gradient — not the same quantity as Guo’s dairy-specific dose-response. It shares authors and cohort machinery with the wider PURE programme, so it is not an independent second witness. And it is part-funded by dairy-industry councils, exactly where such a sponsor would push (Mente et al., 2023). It nudges the range toward neutral-to-mildly-favourable for whole-fat dairy; it does not license a full-fat dairy is cardioprotective claim.
The milk shortens your life scare is a single-cohort artifact
Public discourse’s scariest dairy claim — high milk intake, higher death — comes from Swedish cohorts where «higher milk consumption was associated with a doubling of mortality risk» (Guo et al., 2017). Run the artifact checks (The U-Shaped Association Artifact) and it collapses:
- The pooled milk-mortality estimate is null (1.00). The Swedish cohort adds only heterogeneity; removing it drops I2 from 97.4% to 70.1% and leaves RR 0.99 (0.96-1.01) (Guo et al., 2017).
- The high-milk drinkers were confounded, being the group with the most smokers and the most people living alone (Guo et al., 2017).
- The same outlier cohort manufactures the opposite pole too — the protective fermented-dairy and cheese signals also vanished once it was removed (Guo et al., 2017).
One confounded cohort generates both the milk scare and the cheese halo. That is the cleanest sign the signal carries no information. The verdict rests on leave-one-out sensitivity, a weak check, with no genetic instrument in this source — but for milk that gap is closed below.
Fermented dairy carries a faint, fragile heart signal
The only inverse cardiovascular signals in dairy sit in the fermented category. A fermented-dairy meta-analysis (Zhang 2019, 385,122 people) found highest-versus-lowest intake associated with lower CVD (OR 0.83, 0.76-0.91), with cheese 0.87 (0.80-0.94) and yoghurt 0.78 (0.67-0.89) (K. Zhang et al., 2019). Two facts shrink it. Heterogeneity was extreme (I2 = 94%), so the pooled figure is a washed-out average across very different studies. And Guo’s per-unit slope makes the effect marginal — about 2% per 20 g/day — and traces it to that same Swedish cohort, without which the fermented and cheese signals disappear (Guo et al., 2017). The honest reading: fermented dairy is at best weakly and fragilely favourable for the heart, and it is no reason to start eating cheese for your arteries (Fermented Foods and Health).
Dairy nudges type-2 diabetes down — and it’s yoghurt and low-fat doing the work
Diabetes is the one endpoint where dairy shows a consistent, if small, benefit — and it disaggregates cleanly. Gijsbers’ dose-response meta-analysis (22 cohorts, 43,118 cases) found total dairy at RR 0.97 per 200 g/day (0.95-1.00), a 3% lower risk, borderline and heterogeneous (Gijsbers et al., 2016). But the whole-dairy number hides where the signal lives:
- Yoghurt carries it, and it saturates early — a 14% lower risk at 80 g/day (about one small pot; RR 0.86, 0.83-0.90), with no further drop above that (Gijsbers et al., 2016).
- Low-fat dairy is suggestive (RR 0.96 per 200 g/day, 0.92-1.00) but not significant.
- Milk, cheese and high-fat dairy are null — milk RR 0.97 (0.93-1.02), cheese 1.00 per 10 g/day, high-fat dairy 0.98 (0.93-1.04) (Gijsbers et al., 2016).
The high-versus-low and food-group analyses land in the same place — Mishali at ~10% lower T2D (Mishali et al., 2019), Schwingshackl at RR 0.97 per 200 g/day (0.94-0.99) (Schwingshackl, Hoffmann, et al., 2017) — but they re-pool the same cohorts, so the agreement is robustness of pooling, not independent confirmation. Two cautions travel with the signal. Mishali’s effect ran only in women (men were null), a subgroup claim with no identified mechanism, so treat it as a hypothesis rather than a sex-specific rule. And nearly every dairy-diabetes analysis clusters in industry-adjacent research groups — a uniformity tell about the field, discounted symmetrically rather than used to delete the finding.
A genetic study confirms milk has no causal effect — but tests milk only
The observational milk-null always carried an escape hatch: residual confounding could hide a real effect either way. A Mendelian-randomization study breaks that ambiguity, because a birth-fixed genotype cannot be confounded by the lifestyle that tracks milk drinking. Vissers used the lactase-persistence gene as an instrument for milk intake in 21,820 people (9,686 diabetes cases) and found genetically predicted milk causally null for diabetes: HR 0.99 per 15 g/day (0.93-1.05) (Vissers et al., 2019). The authors read the observational inverse signal as «likely caused by residual confounding» (Vissers et al., 2019).
The boundary matters as much as the result. The gene raises milk intake and essentially nothing else in the dairy group — yoghurt and cheese are individually null on the instrument (Vissers et al., 2019) — so the study is a clean test of the milk arm and silent on yoghurt. That is exactly where the observational protective signal lives, and this genetic test cannot refute it. Reading the milk-null as sinking the yoghurt signal would be the category error the whole decomposition exists to prevent. So the composite is: milk has no causal diabetes effect, yoghurt keeps a small observational one, and neither is a big lever.
Milk is not a bone-fracture lever
The intuitive story — dairy is rich in calcium and protein, so more dairy means stronger bones — does not survive the strongest study designs. Malmir pooled 34 studies (616,000 people for fracture) and found the protective associations only in cross-sectional and case-control studies, the designs most prone to reverse causation; in prospective cohorts, dairy and milk were null for both osteoporosis and hip fracture (Malmir et al., 2019). The milk-fracture paradox is stark: weaker designs say high milk drinkers have 25% less hip fracture, while cohorts show a 9% higher risk per 200 g/day (RR 1.09, 1.07-1.11) (Malmir et al., 2019). The same Michaelsson Swedish cohort behind the milk-mortality scare largely drives that cohort harm signal, so read it as confounded and unadjudicated, not as evidence milk breaks bones (Dairy and Bone Health).
Decomposing dairy by product does not rescue a bone benefit. A second gold meta-analysis (Bian 2018, 381,987 people across 10 cohorts and 8 case-control studies) reaches Malmir’s cohort-null milk result and adds the by-product cells Malmir left empty: in cohorts, yogurt (RR 0.75, 0.66-0.86) and cheese (RR 0.68, 0.61-0.77) both associate with fewer hip fractures (Bian et al., 2018).
Read alone, those numbers look like a fermented-dairy bone lever. They are not: roughly 98% of the pooled yogurt and cheese participants come from a single Swedish cohort (Michaelsson), the same confounded high-dairy population behind the milk-harm scare, and the individual non-Swedish arms are null (inferred from Bian et al., 2018). Within that one cohort milk tracks with harm while yogurt and cheese track with protection, which rules out generic dairy confounding but fits product-specific dietary-pattern confounding just as well as a real fermented effect. So the by-product protection is a candidate distinction to test, not a reason to eat yogurt or cheese for your bones -> Dairy and Bone Health.
Two loose ends stay honest. A dose-response J-shape put dairy above 250 g/day at higher osteoporosis risk, but it rests on the same reverse-causation-prone designs and is unadjudicated. And the food route lands where the supplement route does. Calcium plus vitamin D is also null for fracture in replete adults, so neither the glass of milk nor the pill moves fracture risk in someone already well-nourished (Deficiency Repletion vs Enhancement). Exercise prevents falls and fractures in older adults; dairy intake does not.
Dairy is not a brain-health lever either
Villoz pooled the only dose-response evidence on dairy and the brain (15 cohorts, 312,580 people, no trials), and the headline is null: highest-versus-lowest intake RR 0.94 (0.82-1.07). The prediction interval (0.61-1.45) runs so wide that a future study could as easily find harm as benefit (Villoz et al., 2024).
The one apparent signal — a dose-response nadir around 150 g/day — is a pooling artifact. Below 150 g/day, low-intake Asian cohorts draw a protective arm (RR 0.83); above it, high-intake European cohorts draw a flat one (RR 1.01, null). The nadir sits exactly on the seam between two populations sampled at non-overlapping doses (Villoz et al., 2024). Milk alone and cheese alone were both null, and the whole association vanished in the studies that adjusted for overall dietary pattern — dairy reads as a marker of a healthier diet, not a cause (Villoz et al., 2024). For a reasonably healthy adult in a Western intake range, there is nothing here to worry about or optimize (Dairy and Cognitive Decline).
Dairy and cancer is the one endpoint this evidence can’t answer
This is a genuine gap, not a null. The dairy chapters of the major cancer report were never extracted into this fabric, so no magnitude for either cancer is held here. What is known at the grading level runs in two opposite directions, and they must stay un-netted. The World Cancer Research Fund grades dairy and dietary calcium as probably protective against colorectal cancer, and dairy as probably associated with higher prostate cancer risk (Red and Processed Meat and Cancer). No dose-response magnitude is held for either, and the weighing of a colorectal benefit against a prostate risk is the person’s, not the evidence’s (The Weighting Problem - Why Population Guidance Is Ill-Posed and Individual Advice Is Not). Anyone using dairy to weigh a personal cancer decision needs the underlying reviews, which this cut does not yet hold.
Does the food matrix change what the saturated fat does?
The same saturated fat might behave differently inside cheese than inside butter — the recurring idea that a food’s structure, not just its fat, drives the outcome. A controlled-feeding contrast supports the direction. 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 raised LDL-cholesterol more than cheese did, with milk in between (Thorning et al., 2017). Thorning reads these trials as showing protein and lactose were not the determinants, and cites a second trial, with milk and cheese matched for calcium against a butter control, as «support of an effect of dairy calcium on the LDL-cholesterol response, whereas fermentation did not seem to be involved» (Thorning et al., 2017). So the matrix plausibly softens what the fat does.
Three limits keep this a framing, not a finding. The endpoint is LDL, a surrogate, not a heart attack, and the review itself asks whether the effect is even large enough to matter in a whole diet (Thorning et al., 2017). It also shares authors with the Guo meta-analysis (Astrup, Givens and Soedamah-Muthu), so it cannot back Guo independently.
Thorning’s underlying workshop was funded by an unrestricted European Milk Forum grant, and several authors disclose dairy-industry funding or consultancy, Astrup among them (Arla Foods, the Danish Dairy Research Foundation); the paper states the sponsor had no role in the manuscript (Thorning et al., 2017). Guo was partly funded by an unrestricted grant from the Global Dairy Platform, Dairy Research Institute and Dairy Australia, which it states had no role in design, analysis or writing (Guo et al., 2017). Thorning is credited as a mechanism to weigh, not as independent corroboration, and it does not lift the confidence of anything above.
The broader reassessment of saturated fat argues the same thing — the health effect depends on the whole food, not the isolated fat (Astrup et al., 2020) — and it carries its own industry and design caveats. The saturated-fat verdict itself belongs to Saturated Fat Intake and Replacement and Does Reducing Saturated Fat Reduce Cardiovascular Events. This cut carries only the dairy-specific matrix nuance, which for now points in a plausible direction on a surrogate endpoint and no further.
What to actually do about dairy
Dairy is a small lever, and the attention it gets runs far ahead of the effect. If the big rocks — smoking, excess weight, inactivity, poor sleep, heavy alcohol — are still in play, dairy is not where the gains are (Layer 1 - Ranking Interventions for a Stratum). Within dairy, the realistic moves net small:
- Keep or add, without worry: milk (causally null for diabetes, artifact-only for mortality), cheese, and full-fat dairy. The full-fat versus low-fat distinction guidance emphasizes is not supported by the hard-outcome evidence.
- A mild nudge, if you want one: yoghurt and low-fat dairy carry a small diabetes benefit that saturates around one pot of yoghurt a day. It is real but modest, and observational.
- Swap for a better food where it competes: replacing butter with plant oils tracks lower mortality — a substitution worth making, though it is the oil earning the credit as much as the butter losing it.
- Do not eat dairy to build bone or protect the brain: neither endpoint responds to it.
Cancer is the open question — two graded directions, no magnitude held here, and the trade-off is personal. For everything else, the number worth watching is not your dairy intake. Every estimate above is observational apart from the milk genetic study; no long-term trial of whole dairy on hard outcomes exists, and none is likely, so these verdicts are calibrated small effects, not settled certainties.
Evidence box
Question ’What does the evidence show about dairy”s effect on each patient-important outcome (CV events, all-cause mortality, type-2 diabetes, bone/fracture, cancer, cognition/dementia) — in which direction, how large, for whom, how certain — once “dairy” is decomposed by category (fermented vs unfermented; cheese vs butter vs cream; full-fat vs low-fat)? Does the food matrix change what the saturated fat does, and how do the endpoints and categories vary, so the realistic options — add, keep, swap, or cut a dairy food — can be weighed against the big rocks?‘ Evidence included 16 sources — 11 gold, 3 high, 1 moderate, 1 weak Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 2 source(s) below the gold/high evidence bar feed this page: Astrup (narrative review, weak); Thorning (narrative review, moderate). Their roles and limitations are described in the text. Last updated 2026-10-07 · Independently reviewed: No · Full edit history