Dairy covers foods that behave very differently in the body. Milk, cheese, butter and yoghurt carry similar saturated fat, yet the outcomes they touch — heart disease, early death, diabetes, bone fracture, cancer — point in different directions. Judge dairy one food and one outcome at a time, and the effects turn out to be small.
For coronary heart disease, stroke and early death, dairy is a wash — more of it, or less, barely moves the needle. Full-fat versus low-fat is not the fault-line guidance implies. The one exception is a small, low-certainty heart-failure harm signal in the food-group evidence, named below and not inflated into a headline.
The milk shortens your life result is a statistical artifact, traceable to one Swedish study whose outlier women’s cohort, once removed, takes the association with it.
Fermented dairy — cheese and yoghurt — carries a faint favourable heart signal, but it is small, fragile and observational, not a reason to start eating cheese for your arteries.
Milk is not a bone lever. The old promise that milk prevents fractures does not survive the better studies; if anything the strongest designs lean the other way.
Cheese and butter are near-opposite foods at the same saturated fat, which is the single most useful thing to know about dairy — the food, not the fat number, is doing the work.
Cancer is the honest gap: dairy leans protective for one cancer and adverse for another, the two must not be netted against each other, and we cannot yet put a number on either.
Bottom line for someone who has already handled the big rocks — not smoking, a healthy weight, active, sleeping enough — dairy is a small lever, and the amount of noise around it is inversely proportional to how much it matters. Keep the dairy you enjoy; do not adopt or avoid it for your heart or your bones on the current evidence.
Split dairy into its foods before you judge it — the label hides opposites
Every dairy claim attaches to a specific food, never to dairy as a monolith. When one large dose-response meta-analysis split dairy five ways (total, high-fat, low-fat, milk, and fermented dairy/cheese/yoghurt), the answers differed cell by cell while the aggregate total dairy hid them (Guo et al., 2017). The three axes that carry a decision:
- Matrix intact vs stripped — the fat inside cheese or yoghurt sits in an intact food structure; butter and cream are largely that fat pulled out of the matrix.
- Fermented vs unfermented — the only inverse heart signals sit in fermented dairy (cheese, yoghurt), not in milk.
- Full-fat vs low-fat — the guidance fault-line, which turns out not to be a fault-line at all in the observed data.
Dairy is a classic case of a food category whose boundary may carry no information -> Is the Food Category Doing Any Work. Never read a whole-milk finding as a yoghurt finding, nor a butter finding as a cheese finding. Everything below is organised one food and one outcome at a time.
For coronary heart disease, stroke and mortality, dairy is a wash — with a small heart-failure harm signal
The largest dose-response evidence — 29 cohorts, 938,465 people (Guo 2017, gold-tier observational meta-analysis) — finds no association between how much dairy people eat and whether they die or develop cardiovascular disease:
- «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 …), CHD (… RR 0.99, 95% CI 0.96–1.02 …) or CVD (… RR 0.97, 95% CI 0.91–1.02)» (Guo et al., 2017). Every interval straddles 1.0 — no effect either way.
- Full-fat and low-fat dairy were analysed separately, and both were null — high-fat dairy CVD RR 0.93 (0.84–1.03), low-fat dairy CVD RR 0.98 (0.95–1.01), with mortality and CHD null in each (Guo et al., 2017). High-fat’s point estimate sits slightly lower on a wider interval, but both straddle 1.0. Guo ran no model that substitutes low-fat for high-fat, so this does not endorse the low-fat guidance default — it simply finds no difference in the observed contrast.
- Milk on its own was null too (per 244 g/day, no association with mortality, CHD or CVD) (Guo et al., 2017).
Two honest limits sit on top of these numbers. Every estimate is a relative risk with no baseline event rate attached, so you cannot recover the decision-relevant absolute effect -> Relative vs Absolute Risk. And the evidence is all observational and self-reported: dietary intake measured by questionnaire carries error large enough to flatten a real gradient, and dairy eaters differ systematically from non-eaters -> Measurement Error in Dietary Assessment. A null here is best read as no signal survives the noise, not as a guarantee of exactly zero effect.
The food-group evidence sharpens this by splitting cardiovascular disease into its subtypes, and one of them is not a wash. A coordinated 12-food-group dose-response series (the DIfE/Boeing team; PROSPERO CRD42016037069) finds dairy null for the two large CVD endpoints but a modest harm for heart failure: «Each additional daily 200 g of dairy were not associated with risk of CHD (RR: 0.99; 95% CI 0.96 to 1.02 …) or stroke (RR: 0.98; 95% CI 0.96 to 1.00 …), but were positively associated with risk of HF (RR: 1.08; 95% CI 1.01 to 1.15 …)» (Bechthold et al., 2017).
- CHD and stroke are null — both intervals sit on 1.0, consistent with Guo’s aggregate wash (certainty MODERATE for each).
- Heart failure is a small harm signal (RR 1.08 per 200 g/day), and its certainty is LOW — it rests on a single dose-response study within the meta-analysis, one CVD subtype among three, set against CHD, stroke and mortality nulls. Name it, do not headline it: a low-certainty, single-subtype harm is a divergence to weigh, not a reason to drop dairy. Guo and Mishali never split heart failure out, so this is a signal the earlier evidence could not see, not a contradiction of it.
- Same series, all-cause mortality is null too — the DIfE/Boeing food-group mortality analysis
finds dairy null for all-cause mortality (RR 0.98, 95% CI 0.93–1.03 per 200 g/day, MODERATE;
Schwingshackl food-group mortality series, Supplemental Figure 20). This finding merely echoes Guo’s
mortality null from within the same series — it adds no mortality claim Guo does not already carry.
So it is held as a corroboration-only body line, not a separate
sources:entry. Bechthold and the Schwingshackl members are one team sharing food-group definitions and an overlapping cohort pool, and they overlap Guo’s cohorts too, so agreement among them re-pools shared data rather than confirming it from a second route.
One net-new arm the earlier cut never covered: dairy tracks with slightly lower incident hypertension — «200 g dairy/d (RR: 0.95; 95% CI: 0.94, 0.97)» (Schwingshackl et al., 2017), LOW certainty. It is a small inverse association on a surrogate-adjacent endpoint from the same non-independent series; proportionate to its low certainty, note it and move on.
The milk shortens your life scare is one confounded cohort, not a finding
The most alarming dairy claim in public discourse — that heavy milk drinkers die younger — comes from a single Swedish study (the Michaelsson cohorts of women and men) in which «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 checks -> The U-Shaped Association Artifact and it dissolves:
- The pooled milk-mortality estimate is flatly null, and the Swedish cohort is doing almost nothing except injecting heterogeneity (I2 = 97.4%). Remove it and heterogeneity collapses to 70.1% with RR 0.99 (95% CI 0.96–1.01) (Guo et al., 2017).
- The confounding is named in the source itself: the heaviest milk drinkers in that cohort were the least educated (~80% of women, 70% of men schooled nine years or less), and «the highest milk drinkers had highest percentage of smokers and those living alone» (Guo et al., 2017) — a profile that predicts higher mortality regardless of what is in the glass.
- The same outlier cohort manufactures the opposite pole of the debate: «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 study drives both the milk harm and the cheese benefit — the cleanest sign that neither has diagnostic value.
Verdict: the milk-mortality signal does not survive the check. The caveat on the check itself: it rests on leave-one-out sensitivity, a weak instrument, with no genetic or randomised evidence in the source to settle causation. The milk-mortality signal is not a finding of harm, and equally not a clean bill of health — it is an artifact removed.
Fermented dairy carries a weak, fragile favourable heart signal
Fermented dairy — cheese and yoghurt — is the one dairy category with a favourable cardiovascular association, and it is worth keeping the confidence honestly low. A meta-analysis of 10 cohorts (385,122 people; Zhang 2019, gold-tier) reported that «statistical evidence of significantly decreased CVD risk was found to be associated with fermented dairy foods intake (OR 0.83, 95% CI 0.76–0.91)» (Zhang et al., 2019). But the aggregate oversells, and the subgroups say why:
| Endpoint | OR (95% CI) | Significant? |
|---|---|---|
| Overall CVD | 0.83 (0.76-0.91) | yes |
| CVD incidence | 0.80 (0.72-0.89) | yes |
| CVD mortality | 0.94 (0.80-1.11) | no — crosses 1 |
| Cheese | 0.87 (0.80-0.94) | yes |
| Yoghurt | 0.78 (0.67-0.89) | yes |
| Stroke | 0.87 (0.75-1.01) | no |
| CHD | 0.85 (0.67-1.08) | no |
The signal is on incidence, not mortality, and vanishes for stroke and CHD taken alone. Heterogeneity is extreme (I2 = 94.0%), which means the pooled central estimate describes no single population (Zhang et al., 2019).
The dose-response meta-analysis (Guo 2017) shrinks it further. Per-unit, fermented dairy is a marginal 2% (RR 0.98, 95% CI 0.97–0.99 per 20 g/day; cheese RR 0.98, 95% CI 0.95–1.00 per 10 g/day for CVD; yoghurt null), and — as above — the inverse association disappears when the one Swedish cohort is removed (Guo et al., 2017). Guo does not confirm Zhang so much as bound it: the two share cohorts and measure different contrasts, so this is refinement-downward, not independent corroboration. The composite reading — fermented dairy is at best weakly and fragilely favourable for cardiovascular events, driven by cheese and by incidence, and artifact-sensitive on mortality — lives on Fermented Foods and Health and Dairy and Cardiometabolic Health. Kefir specifically has essentially no hard-outcome data; hold it at insufficient evidence, not at null.
More dairy tracks with slightly less type-2 diabetes — two estimates, one non-industry, moderate certainty
For type-2 diabetes the signal points modestly favourable, and it now rests on two gold estimates, not one. Mishali 2019 (16 cohorts, 545,677 people) found that «Pooled results indicated an inverse association between the two (RR 0.897; 95% CI, 0.834–0.963; P < 0.01)» (Mishali et al., 2019) — roughly 10% lower diabetes risk in the highest-versus-lowest dairy eaters, with a parallel ~5% lower cardiovascular signal (CVD RR 0.942, 95% CI 0.892–0.994) (Mishali et al., 2019).
The DIfE/Boeing food-group series adds a second, non-industry estimate at a compatible magnitude: «Each additional daily 200 g of dairy products was inversely associated with diabetes risk (RR: 0.97; 95% CI 0.94–0.99 …)» (Schwingshackl, Hoffmann, et al., 2017), MODERATE certainty, from a research-institute team with no dairy-industry sponsor. This estimate refines the inverse lean — a per-200-g dose-response of ~3% sitting comfortably inside Mishali’s highest-vs-lowest 10%. But it is not independent convergence: Schwingshackl shares the cohort pool and much of the observational machinery, so it re-pools overlapping data rather than confirming it by a second route. Two caveats keep it honest — the inverse held «only in Asian and Australian studies, but not for American and European studies» (Schwingshackl, Hoffmann, et al., 2017), and it is total dairy, not a fermented-specific curve.
That cardiovascular number from Mishali does not independently corroborate Guo’s cardiovascular null: the two meta-analyses draw on overlapping cohorts and measure a different contrast (Mishali’s is a highest-vs-lowest categorical estimate, Guo’s a per-200-g dose-response), so the two amount to a distinction, not type-E independent backing. Do not read Guo-plus-Mishali as two separate confirmations on cardiovascular disease.
Two guards, applied symmetrically:
- The distinctive result is a sex split, and it is a subgroup claim. The protective effect sat in women; «The pooled RR for men was not significant (RR 1.023; 95% CI, 0.91–1.15 …)» (Mishali et al., 2019). No mechanism was identified (menopause-age and region moderators came back null), so this is the false-positive-prone effect-modification route -> route (b). Hold it as hypothesis-generating; do not build a sex-specific recommendation on it.
- Mishali’s funding is a scrutiny flag on that estimate, no longer the whole leg. «This work was financed by the Israel Dairy Board» (Mishali et al., 2019), and the review does not stratify by fat content while leaning on an exonerates fat framing — a directional tell that discounts Mishali’s estimate. But it no longer carries the inverse on its own: Schwingshackl’s non-industry 0.97 lands at compatible magnitude, so the finding survives the discount rather than depending on the sponsored source.
So the diabetes leg is now firmer than a single industry-funded reading: a modest inverse lean, two estimates (one non-industry) at compatible magnitude, moderate certainty. It is still total dairy rather than fermented-specific, and not a big rock. Treat it as a small favourable lean, not a reason to prescribe dairy.
The dose-response by subtype is now held, and it turns the leg into a clean split-the-food case. Gijsbers’ dedicated dairy -> T2D dose-response meta-analysis (22 cohorts, 579,832 people, 43,118 incident cases) runs the per-serving shape by subtype instead of one whole-dairy point estimate, and the modest total inverse (RR 0.97 per 200 g/d, 0.95-1.00) turns out to be carried by two subtypes while the rest sit flat (Gijsbers et al., 2016):
- Yogurt is the one real signal, and it saturates early. A non-linear inverse — 14% lower risk at 80 g/d (RR 0.86 vs 0 g/d, 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 bought by about one small pot a day and then plateaus; more buys nothing.
- Low-fat dairy: a suggestive inverse that misses significance (RR 0.96 per 200 g/d, 0.92-1.00, P=0.072); milk, cheese and high-fat dairy are null across the board (milk 0.97 per 200 g/d, 0.93-1.02; cheese 1.00 per 10 g/d; high-fat 0.98, 0.93-1.04) (Gijsbers et al., 2016).
Two honesty checks keep this in proportion. Gijsbers’ total-dairy slope (0.97 per 200 g/d) is the same figure the leg already holds from Schwingshackl, re-pooling the same canonical cohorts (EPIC-InterAct, Whitehall II, the US and Malmo cohorts) — echo, not a second independent route; the subtype shape, not the total, is what this source adds. And it is still not a big rock: a yogurt eater gets a small, early-saturating signal, a milk or cheese eater none, and none of it is a reason to prescribe dairy -> Fermented Foods and Health (the yogurt leg cashes that page’s yogurt -> T2D dose-response gap).
Milk is not a bone-fracture lever
The intuitive story — dairy is rich in calcium and protein, so more dairy means stronger bones and fewer fractures — does not survive the better study designs. In a meta-analysis of 34 studies (616,000 people for fracture; Malmir 2019, gold-tier), the protective association appears only in the weaker designs and vanishes in the strong ones:
| Outcome / exposure | Cross-sectional + case-control | Prospective cohort |
|---|---|---|
| Total dairy -> osteoporosis | 0.63 (0.55-0.73), significant | 0.82 (0.56-1.18), NS |
| Milk -> osteoporosis | 0.68 (0.50-0.94), significant | 1.08 (0.52-2.24), NS |
| Total dairy -> hip fracture | 0.86 (0.53-1.37), NS | 0.90 (0.73-1.11), NS |
| Milk -> hip fracture | 0.75 (0.57-0.99), significant | 0.93 (0.75-1.15), NS |
The author’s own verdict: «a greater intake of milk and dairy products was not associated with a lower risk of osteoporosis and hip fracture» (Malmir et al., 2019), on the reasoning that «findings from cohort studies are closer to the causal associations than those from cross-sectional and case-control studies» (Malmir et al., 2019). The weaker designs are prone to reverse causation — a fracture or osteoporosis diagnosis can change what someone reports eating -> The U-Shaped Association Artifact.
The milk-fracture paradox is worth stating precisely, because it looks alarming and is not. In cohorts, «every additional 200-gram increase in milk intake was associated with a 9% greater risk of hip fracture (RR 1.09; 95% CI: 1.07–1.11)» (Malmir et al., 2019) — but that per-gram positive slope is substantially driven by the same Michaelsson Swedish cohort family that manufactures the milk-mortality scare above, the pooled milk-cohort estimate is itself null (0.93, NS), and Malmir ran no leave-one-out to confirm it. Read it as confounded and unadjudicated, not as evidence that milk breaks bones -> Dairy and Bone Health.
The food route and the supplement route land in the same place. Dairy food is null for fracture here, and calcium-plus-vitamin-D supplements are also null for fracture in replete, community-dwelling adults — the benefit appears only in the deficient or institutionalised -> Deficiency Repletion vs Enhancement, Vitamin D and Calcium Supplementation for Fracture Prevention. For someone who is not calcium/vitamin-D deficient, neither the glass of milk nor the pill moves fracture risk; the better-evidenced lever is exercise and fall prevention -> Exercise for Preventing Falls in Older Adults. The deficient person is a genuinely different stratum, and dairy/calcium may still matter there — Malmir did not screen for deficiency, so that arm stays a separate, open question.
Cheese and butter part ways at the same saturated fat — the food does the work, not the fat number
The cheese-butter divergence is the mechanistic heart of the dairy question, and it needs stating carefully: it is a claim about why the categories diverge, corroborated in humans but not proven by a hard-outcome trial — it does not override the outcome evidence above.
The observed pattern is real. Even butter — near-pure fat, matrix stripped away — is at worst weakly positive for mortality (per 14 g/day, RR 1.01, 95% CI 1.00–1.03), with no significant association for CHD, CVD or stroke, and an inverse association with diabetes (RR 0.96, 95% CI 0.93–0.99) (Guo et al., 2017). Cheese, meanwhile, is the category with the (fragile) inverse CVD signal. Same saturated fat, different food, different association.
The candidate explanation is the food matrix: saturated fat behaves differently depending on the structure it sits in. As Astrup’s reassessment puts it, «the healthfulness of fats is not a simple function of their SFA content, but rather is a result of the various components in the food, often referred to as the “food matrix.”» (Astrup et al., 2020), and specifically for dairy, «The complex matrix and components of dairy may explain why the effect of dairy food consumption on CVD cannot be explained and predicted by its content in SFAs» (Astrup et al., 2020).
Three cautions keep this from becoming a dairy halo:
- The matrix claim is a mechanism with human corroboration, not an outcome finding. No trial has fed cheese against butter over decades and counted heart attacks — that study cannot be blinded and does not exist. The matrix claim is admitted directionally and marked as such, discounted accordingly.
- The source is dairy-adjacent. Astrup’s review is a narrative (not systematic) review with food-industry funding and dairy ties across the author line — a provenance flag to scrutinise under symmetric standards, not to dismiss, and a reason the matrix point stays a distinction (the fat number does not predict the food) rather than a positive claim that these foods protect.
- The saturated-fat verdict itself is settled elsewhere. Whether reducing saturated fat reduces cardiovascular events is a separate question -> Saturated Fat Intake and Replacement, Does Reducing Saturated Fat Reduce Cardiovascular Events; this page carries only the dairy-specific matrix nuance.
The decision consequence is Test 3 of the food-category diagnostic: when the mechanism lives in the food and not the nutrient, decide on the food, not the saturated fat label -> Is the Food Category Doing Any Work.
Dairy and cancer point in two directions, and neither magnitude is in yet
Cancer is the honest gap in this cut. The direction of the two best-characterised dairy-cancer associations is held on the cancer pages — dairy leans probably protective for colorectal cancer and probably adverse for prostate cancer -> Red and Processed Meat and Cancer — but the magnitude-quantifying evidence has not been extracted into the fabric (the WCRF dairy chapters were ingested only for the red-and-processed-meat colorectal analysis). So:
- The two directions are named and stay un-netted. They move different cancers, and collapsing a colorectal benefit against a prostate risk into a single dairy and cancer verdict would be false arithmetic. Which one weighs more is the person’s call, made against their own risk profile — name the two axes and stop.
- The numbers are a stated gap, not a null. This is insufficient evidence held here, distinct from no effect; do not read the absence of a magnitude as an absence of an effect. Extracting the WCRF dairy chapters is the follow-up that would close it.
The bottom line
For someone who has already handled the big rocks -> Layer 1 - Ranking Interventions for a Stratum:
- Do not worry about dairy for your heart or your lifespan. Total, full-fat and low-fat dairy are a wash for coronary heart disease, stroke and mortality in the best available (observational) evidence. The one asterisk is a small, low-certainty heart-failure harm signal (RR 1.08 per 200 g/day, one CVD subtype) — worth naming, not worth cutting dairy over.
- Ignore the milk kills headline. It is a single-cohort confounding artifact that removes itself when the outlier study is dropped.
- Do not eat dairy to protect your bones, and do not fear milk for them either — dairy is null for fracture in the strong designs, and so are calcium/vitamin-D supplements in people who are not deficient. Exercise and fall prevention are the better fracture lever.
- If you want the (small, fragile) favourable cardiovascular lean, it is in fermented dairy — cheese and yoghurt — not milk. It is a low-cost swap, not a therapy.
- Decide on the food, not the saturated fat number. Cheese and butter diverge at the same fat content; choose the specific dairy food you will actually eat and enjoy.
- Keep the cancer trade-off explicit and personal — a probable colorectal benefit and a probable prostate risk, un-netted, weighed against your own risks.
- Swap sweetened or flavoured dairy to plain. A flavoured yoghurt or a sweetened milk drink routes to the added-sugar question, not to anything dairy is doing; the sugar is the exposure. Swap to plain and the two questions separate cleanly.
- There is no health mandate to eat dairy at all. No endpoint in the held evidence makes it a required food; if it is unwanted or poorly tolerated, nothing is lost by leaving it out and meeting protein, calcium and calories elsewhere (fermented dairy and hard cheeses are the lower-lactose options for those who want some).
Net: dairy is a small lever, and the volume of attention it attracts is itself a signal that the established effects are small -> attention is an anti-signal. The realistic decision is a handful of category-level swaps, not a dairy: yes or no.
The necessary caveats:
- This wiki appraises evidence; it does not prescribe. It reports what dairy does to each outcome, in which direction and how certainly — selecting, dosing and screening for your situation is a clinician’s job with materials this page does not hold.
- The loop is open. Nothing here has been graded against a realized outcome in a real person; the wiki grades coherence and fidelity to its sources, not truth about the world. A clean appraisal is not a validated recommendation.
- These estimates are population-level by default; stratify per person. Personalise beyond the population estimate only on positive evidence of effect modification — the sex-by-diabetes and matrix-sensitivity claims here are hypotheses, not established strata. Absolute benefit scales with your baseline risk even when the relative effect does not.
- One health axis only. This page weighs longevity, cardiometabolic and bone outcomes and cancer. It holds no environmental, animal-welfare, ethical or cost data and does not price them; where a dairy choice carries load on those axes, that trade-off exists and is yours to weigh — it is not netted into anything here.
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) — 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 (which dairy to keep, which to swap) can be weighed against the big rocks?‘ Evidence included 9 sources — 8 gold, 1 weak Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 1 source(s) below the gold evidence bar feed this page: Astrup (narrative review, weak). Each labelled by tier; none load-bearing for the core claims. Last updated 2026-09-05 · Independently reviewed: No · Full edit history