Is a drink good for me? has no single answer, because alcohol affects different organs differently, and an honest reply has to take them one at a time. But the headline is firm: the reassuring idea that a moderate drinker outlives a teetotaller does not hold up. Many of the “non-drinkers” it rests on are former drinkers who stopped because they were already ill; correct for that, then check it against genetics, and the apparent benefit of a daily glass disappears.
Taken organ by organ, what remains is mostly harm that begins early. Alcohol raises blood pressure and stroke risk with no safe lower dose, and it raises the risk of several cancers from the very first drink — moderate drinking included. Past a few drinks a day it measurably shortens life, and the level at which that shows is lower for women than for men. Only the heart is genuinely unresolved: for coronary disease a small real benefit cannot yet be ruled out, so the harm alcohol does to the vessels of the brain does not simply extend to it.
What counts is the total amount of ethanol, not the form it comes in. Saving a week’s drinks for two heavy nights is worse than spreading them out; the wine-versus-spirits question — and the red-wine “resveratrol” story in particular — buys no measurable health. Across a whole population the risk is lowest at zero, which sits uneasily beside guidelines that permit a daily drink or two; this page names that tension without pricing it. Confidence here is medium, not high, and the loop stays open: every judgment is graded for coherence, never against what later became of anyone who drank.
Does a little drinking protect? The low-dose arm across mortality, stroke, and brain
One belief about alcohol outranks every other: that a little beats none. Test it first, because settling it changes how every outcome below reads. The belief rests on one shape in the observational data: risk lowest at a low-but-non-zero dose, higher at zero and higher again above. The low-dose segment — the lower arm — is the fragile part, and whether it is real decides how every outcome below reads.
The appearance, before adjudication
Two large observational analyses show the arm. In Stockwell’s meta-analysis (87 studies), low-volume drinkers (1.3-24.9 g/day) appear to outlive abstainers — RR 0.86, 95% CI [0.83, 0.90]. (Stockwell et al., 2016) In Millwood’s half-million-person China Kadoorie cohort, self-reported intake traces a U-shape for ischaemic stroke, intracerebral haemorrhage, and myocardial infarction, with the lowest risk at about 100 g/week. (Millwood et al., 2019) Both are the pre-adjudication appearance — the shape the data show before the referent is checked — not a finding. This page does not carry the 0.86 forward as the effect of low-volume drinking.
Two independent methods dissolve the arm [E-independent]
The lower arm does not survive adjudication, and it fails by two routes that share no data, method, or lineage — the convergence, on direction, is the crux of this appraisal.
- Bias-correction (Stockwell). Once Stockwell corrects abstainer definitions and controls study quality, the low-volume estimate moves to RR 0.97, 95% CI [0.88, 1.07] — no significant protection. Using occasional drinkers rather than lifetime abstainers as the referent, «there would be no evidence of health protective effects for low-volume drinkers or any other category of drinker». (Stockwell et al., 2016)
- Mendelian randomization (Millwood) — genetic variants (ALDH2, ADH1B) that lower alcohol tolerance serve as lifetime-exposure instruments, immune to reverse causation and lifestyle confounding. The genetically-predicted relationship is monotone with no protective lower arm — total stroke RR 1.38 (1.26-1.51) per 280 g/week. Its verdict: «the apparently protective effects of moderate alcohol intake against stroke are not mainly caused by alcohol itself, and are largely artifacts of reverse causation and confounding». (Millwood et al., 2019)
Two independent adjudications — a bias-corrected meta-analysis and a genetic natural experiment — reach
the same qualitative conclusion (the low-dose benefit is not causal). The [E-independent] lift is on
that direction, not on a pooled magnitude; the two address different outcomes.
What manufactures the arm
The arm appears with no true benefit behind it. The driving mechanism is former-drinker misclassification — people who quit because they became ill (sick-quitters) are counted among “abstainers,” making the referent group look unhealthy and everyone who still drinks look protected; former drinkers carry elevated risk (RR 1.38 [1.24-1.54]). Confounding by frailty compounds it: 27 of 30 candidate confounders for coronary disease were more prevalent among abstainers than among moderate drinkers. (Stockwell et al., 2016)
The decision rule
A protective lower arm stays unadjudicated until it survives a referent-correction or a genetic/quasi-experimental check — mechanistic plausibility and covariate adjustment are not enough. Excluding early follow-up (to remove the reverse-causation window) is the weak check: the alcohol U-shape survived it, yet dissolved under the strong checks. An arm that has passed only the weak check has not been adjudicated. The referent-correction and MR are the strong routes, and where a genetic instrument exists it is decisive.
The same pattern on a second outcome — dementia
The artifact replicates on brain outcomes. Observational studies find a J-shaped dementia curve in which non-drinking looks worse than light drinking — again because many non-drinkers previously drank heavily or stopped for illness. Alzheimer’s-disease Mendelian randomization removes it: «any relationship between not drinking and AD is due to survivor bias». (Livingston et al., 2024) The same referent-contamination-plus-MR adjudication that dissolved the mortality arm dissolves the dementia one — a different literature reaching the same verdict.
A gold dose-response meta-analysis puts numbers on that J, and it is the textbook instance the discipline above dissolves. Xu pooled 11 prospective cohorts (73,330 participants, 4586 all-cause-dementia cases) into a nonlinear curve (p_nonlinearity < 0.05): the apparent protection is confined to at most 12.5 g/day (roughly one drink a day), risk bottoms at RR ~0.9 around 6 g/day, then climbs about 10% once intake passes ~38 g/day (~23 drinks/week). (Xu et al., 2017)
That ~6 g/day nadir is exactly the artifact-suspect lower arm: the same referent contamination (former heavy drinkers sitting among the abstainers) and the Alzheimer’s Mendelian randomization above apply to it unchanged, so the nadir is neither a target nor a safe dose. Xu itself does not read the protective arm as causal — it cautions that its findings need cautious interpretation, given varying methods and no standard definition of intake. (Xu et al., 2017) The upper arm — the ~10% rise above ~38 g/day — marks where the dementia signal becomes unmissable, not a ceiling to drink up to, and it lines up with the heavy-midlife harm below. Xu is not an independent line of evidence here: it shares its Qingdao/Fudan authorship with other cognition sources the wiki holds, so it quantifies the shape rather than corroborating the verdict.
Effect and dose-response shape, outcome by outcome
With the low-dose arm adjudicated as artifact, each outcome traces its own curve, and the shapes differ. Each outcome below is its own card: read its direction, magnitude, and dose-response shape separately, keep the four evidence-states (harm / no meaningful effect / benefit / insufficient) apart, and let no single card’s threshold govern another. One recurring trap: the dose at which a harm becomes statistically detectable marks an edge of the evidence base, not a level below which the exposure is safe. This page gives doses in grams of ethanol (1 standard drink = 10 g) with the drink equivalent.
All-cause mortality — harm above a sex-specific dose, no net benefit below
Low-dose drinking buys no net benefit — the apparent J-curve is an abstainer-bias / sick-quitter artifact (established above; the adjudicated estimate is null, not the raw 0.86). The updated pooled analysis adds where harm becomes detectable, and that dose is sex-specific:
- Pooled: significant elevation begins at 45-64 g/day (~4.5-6.5 drinks), RR 1.19 (95% CI 1.07-1.32), rising to 1.35 at >=65 g/day; the 25-44 g/day band is null (1.05, ns).
- Women: medium-volume RR 1.21 (1.08-1.36) from ~25 g/day (~2.5 drinks), and any drinking is already elevated (1.22, 1.02-1.46).
- Men: null below 45 g/day.
- Across all levels, women carried higher mortality RR than men — sex interaction P=.001.
Read the dose correctly. 45 g/day (25 for women) is where cohort data can resolve an elevation above noise — an evidence-detectability edge, not a safe threshold. It does not license “below 45 g/day is fine”: other outcomes (cancer, below) are already accruing risk far below it. The dose is specific to this endpoint and this studied range (categories up to >=65 g/day); it is not a cross-outcome licence.
Stroke and blood pressure — monotone causal harm
For stroke the genetic (Mendelian-randomization) estimate — using variants that lower alcohol tolerance as instruments for lifetime exposure, so it is immune to reverse causation — is monotone from zero with no protective lower arm:
- Total stroke RR 1.38 (95% CI 1.26-1.51) per 280 g/week (~28 drinks/wk, ~4/day); intracerebral haemorrhage 1.58 (1.36-1.84); ischaemic stroke 1.27 (1.13-1.43).
- Blood pressure rises +4.3 mmHg (3.7-4.9) genetic and +4.8 mmHg (4.5-5.1) conventional per 280 g/week — the two routes converge, so here the confounding is specific to the protective claim, not to the BP effect itself.
The stroke harm is causal (state HARM) and rises across the whole studied range. No low-dose level has been shown exempt: the curve carries no threshold. Blood pressure here is a surrogate — a marker on the causal pathway, not the outcome itself; it counts because its transmission to hard vascular outcomes is well evidenced, never as a harm in its own right (a marker moved is not a life lost). -> Surrogate Outcomes
Interventional confirmation, and where the BP lever concentrates (Roerecke 2017). The Millwood
slope is genetic and observational; a meta-analysis of 36 randomised alcohol-reduction trials (2,865
participants) supplies the interventional leg and grades the dose-response GRADE-high. Cutting alcohol
lowered systolic pressure by about −3.1 mmHg overall, but the effect is concentrated above ~2 drinks
per day — roughly −0.9 mmHg per drink/day of baseline intake, with no significant BP reduction in
people already drinking two or fewer per day
(Roerecke et al., 2017). This third,
independent method confirms that alcohol raises blood pressure
causally ([E-independent] with the genetic route above), while refining the shape of the BP
surrogate: a single linear MR coefficient cannot resolve a low-dose knee, whereas the reduction trials
locate one, so for a light drinker cutting alcohol buys little or no BP change. It refines the surrogate
only — it neither contradicts the monotone stroke harm (a hard outcome, driven across the whole range)
nor softens the no-safe-level verdict, which rests on cancer’s harm from zero, not on blood pressure.
(Millwood et al., 2019; inferred from Roerecke et al., 2017)
Coronary heart disease / MI — the honest boundary (insufficient)
Coronary disease is a separate card from stroke — “alcohol harms the heart” must not collapse the two. For myocardial infarction the genetic estimate is null: RR 0.96 (95% CI 0.78-1.18) per 280 g/week, p=0.69 — «little net effect». The case count was limited, and the authors caution that «some real benefit or hazard cannot be excluded», with any BP harm possibly «offset by cardio-protective changes in other factors».
State this as INSUFFICIENT, not benefit and not harm: coronary heart disease is the one outcome where a small real benefit is not ruled out. The stroke harm above does not transport to it, and the cancer/stroke harms do not erase this open question.
Cancer — monotone from zero, no lower threshold
Alcoholic drinks are an IARC Group 1 (established) human carcinogen, and WCRF grades alcohol a cause of cancers of the mouth/pharynx/larynx, oesophagus, breast, colorectum, liver and stomach. Cancer’s shape differs from mortality’s: risk is present at every dose, with no floor below which it disappears.
«Even small amounts of alcoholic drinks can increase the risk of some cancers – there is no level of consumption below which there is no increase in the risk of at least some cancers.» (World Cancer Research Fund & American Institute for Cancer Research, 2018)
That “at least some cancers” hides a site-by-site split, and a gold dose-response meta-analysis resolves it. Bagnardi pooled 572 studies (486,538 cancer cases) across 23 sites and found the no-threshold shape holds for a specific set, not for every cancer -> Alcohol and Cancer Risk. At three sites risk is already significantly raised at light drinking (~1 drink/day, <=12.5 g ethanol): female breast RR 1.04 (95% CI 1.01-1.07), oral cavity and pharynx 1.13 (1.00-1.26), and oesophageal squamous-cell carcinoma 1.26 (1.06-1.50) — the curve climbs from the first drink. Bagnardi states it directly: «alcohol drinking was associated with cancer of the oral cavity and pharynx, oesophagus (SCC) and female breast even at low doses.» (Bagnardi et al., 2014)
At other sites the light-drinking estimate sits at unity — colorectum RR 0.99 (0.95-1.04), stomach 0.99 (0.92-1.06) — and risk becomes detectable only at moderate or heavy intake, so those curves are shallower or start higher, a threshold-like shape. Heavy drinking then multiplies the aerodigestive sites severalfold (oral cavity and pharynx RR 5.13 (4.31-6.10), oesophageal SCC 4.95 (3.86-6.34)). (Bagnardi et al., 2014)
So no safe threshold is site-specific, not uniform — and that resolution keeps the overall verdict intact rather than softening it: because the breast and aerodigestive sites carry no floor, no level of drinking is safe for cancer as a whole, even though other sites need moderate-or-heavy intake before their risk resolves. WCRF’s own gram-figures line up with the same split: colorectal from ~30 g/day (~2 drinks), liver and stomach from ~45 g/day (~3 drinks) mark the edge of the graded evidence base, not safe levels — where the data thin out, below which risk still accrues. (World Cancer Research Fund & American Institute for Cancer Research, 2018)
Kidney cancer shows a genuine protective association (up to ~30 g/day), but it is «far outweighed by the increased risk for other cancers» — a single-site benefit inside a net-harmful whole-organism ledger (net-effect rule, not a reason to drink). (World Cancer Research Fund & American Institute for Cancer Research, 2018)
The cancer card governs the cross-outcome reading: because cancer risk starts at the first drink, the mortality “45 g/day” figure can never be read as a whole-body safe level.
Cognition and dementia — harm at heavy midlife intake
Heavy midlife drinking (>21 units/week) raises dementia risk: IPD HR 1.22 (95% CI 1.01-1.48). The apparent protection of light-versus-none is again the low-dose artifact adjudicated above (former heavy drinkers among the “non-drinkers”; AD-MR attributes the not-drinking association to survivor bias), not a benefit. (Livingston et al., 2024)
State HARM at heavy intake; the light arm is artifact, not benefit.
Injury and non-cancer liver disease — named gaps, no direction inferred
The decision menu lists two patient-important outcomes that no dedicated dose-response arm in the fabric covers; this page names them as gaps rather than answering them:
- Injury — held only inside GBD’s 23-outcome population aggregate, with no standalone dose-response curve extracted.
- Non-cancer liver disease (alcoholic liver disease / cirrhosis) — the fabric holds liver cancer and a MASLD definitional ceiling only; there is no dedicated ALD/cirrhosis dose-response arm.
This page infers no direction for either (insufficient evidence held, distinct from no-effect), and consolidates both with the other named gaps later.
The active axis: total ethanol, drinking pattern, or beverage type?
The previous section indexed every effect to average volume — grams of ethanol per week. Average volume is the dominant axis: total ethanol drives the stroke, cancer, and mortality gradients, and it does so regardless of what the drink is. But volume is not the whole story. Two further axes are live — how the same weekly total is spread across sessions, and what beverage delivers it. The rest of this section asks whether either does independent work, holding a favourable-beverage story to exactly the bar the harm claims had to meet.
Drinking pattern — a real second axis at fixed volume
Roerecke’s meta-analysis (14 studies, 4718 ischaemic-heart-disease events) isolates pattern by comparing drinkers matched on average volume. In a random-effects model the «pooled relative risk of irregular heavy drinking occasions compared with regular moderate drinking was 1.45 (95% confidence interval: 1.24, 1.70)» (Roerecke & Rehm, 2010), with I²=53.9%. Concentrating a week’s drinks into heavy sessions raises heart-disease risk at the same weekly total.
Two features keep this finding robust rather than another J-curve artifact. The contrast excludes abstainers and former drinkers, so «it is unlikely that a sick-quitter effect … influenced our findings» (Roerecke & Rehm, 2010) — the misclassification that dissolved the protective lower arm does not apply here. And non-differential measurement error biases a pooled estimate toward the null, so the effect is if anything understated.
Read the finding precisely, because Roerecke’s referent — regular moderate drinkers — is presumed cardioprotective, and that presumption inherits the earlier verdict: it does not survive adjudication. So this is not evidence that moderate drinking protects. What it robustly shows is a within-drinker effect at fixed volume: among people who already drink, spreading the same total out beats bingeing it. Roerecke’s own summary — «the cardioprotective effect of moderate alcohol consumption disappears when, on average, light to moderate drinking is mixed with irregular heavy drinking occasions» (Roerecke & Rehm, 2010) — is a statement about pattern, not a licence to drink. Decision form: if you drink, do not concentrate a week’s drinks into one or two heavy sessions — a lever available even to someone unwilling to cut total intake.
Beverage type: the red-wine / resveratrol story, tested and failed
The remaining candidate axis is beverage type — the French paradox claim that red wine’s benefit lies not in ethanol but in a polyphenol, resveratrol. The resveratrol story is a favourable-beverage claim, so under symmetric standards it earns the same scrutiny as the harm claims: a testable component claim, and Semba tested it. Semba’s InCHIANTI cohort (783 adults 65+, 9-year follow-up) measured resveratrol directly — 24-hour urinary metabolites rather than a wine-drinking questionnaire — and found «total urinary resveratrol metabolite concentration was not associated with inflammatory markers, cardiovascular disease, or cancer or predictive of all-cause mortality. Resveratrol levels achieved with a Western diet did not have a substantial influence on health status and mortality risk» (Semba et al., 2014).
The same data expose the collinearity trap. Urinary resveratrol tracks wine intake: the «Spearman correlation between alcohol consumption in grams per day and total urinary resveratrol metabolite concentrations was 0.67 (P < .001)» (Semba et al., 2014), and the biomarker is «a valid biomarker of wine consumption» (Semba et al., 2014). So a study crediting resveratrol is largely measuring ethanol under a polyphenol label — the component collects the credit that belongs to the drink carrying it.
Symmetric standards cut both ways here, so separate what Semba establishes from what it does not. Semba is a single moderate-tier cohort, a labelled-weak null counterweight, and no gold-tier resveratrol-outcome meta-analysis exists, so it does not settle wine’s merits by authority. It does refute a folklore inference: at the doses a drinker actually reaches, the presumed active component does no measurable work, so drink-red-wine-for-the-resveratrol is removed. It is not evidence that wine is worse than other drinks, and it does not rule out a supraphysiologic supplement dose (100-1000x dietary — a different exposure, untested here). The wine-specific benefit is unmeasured-or-absent — a claim that was tested and failed, not a proven wine-harm and not a proven wine-benefit.
Beverage type at matched ethanol — a named gap
That leaves the sharper question the fabric cannot answer: at matched grams of ethanol, does wine differ from beer or spirits on a hard outcome? No meta-analysis compares beverage types at equal ethanol on mortality, stroke, or cancer. It stays a named gap, not a null, and the wiki does not rank drinks on it. The evidence supports something narrow but firm: whatever ethanol does, it does regardless of the glass; the pattern lever is real; and the one beverage-specific benefit anyone tested came back empty.
These per-person and within-drinker axes scale up to a population question — and there the appraisal meets published guidance.
Population scale and the guidance clash
Move from the individual drinker to the population and a second kind of evidence appears: not a per-person risk ratio but a count of cases and deaths across whole nations. Two population-attributable-fraction (PAF) models supply it, and both point the same way as the per-outcome appraisal above.
GBD finds the harm-minimizing level is zero. Weighting all 23 alcohol-linked outcomes by global disease burden, «the level of alcohol consumption that minimised harm across health outcomes was zero (95% UI 0.0-0.8) standard drinks per week» (1 standard drink = 10 g ethanol) (Griswold et al., 2018). Alcohol led the risk-factor ranking at ages 15-49 and was tied to 2.8 million deaths in 2016 (Griswold et al., 2018).
The minimum and its offset, read together. GBD does recover a small protective minimum for ischaemic heart disease and diabetes — a relative risk near 0.86 (0.80-0.96) for men and 0.82 (0.72-0.95) for women at roughly 0.9 drinks/day. But «these protective effects were offset by the risks associated with cancers, which increased monotonically with consumption» (Griswold et al., 2018). Netted across outcomes, the safest level lands at zero: the single-outcome benefit is real and is simply outweighed — the same benefit-inside-net-harm pattern the coronary and kidney-cancer arms show, now at population scale.
Rumgay adds the cancer count, and it reinforces no-threshold with scale. «An estimated 741 300 (95% UI 558 500-951 200), or 4.1% (3.1-5.3), of all new cases of cancer in 2020 were attributable to alcohol consumption» (Rumgay et al., 2021). Light-moderate drinking is not exempt: «moderate drinking still contributed one in seven alcohol-attributable cases and more than 100 000 cancer cases worldwide» (Rumgay et al., 2021) — the <20 g/day band alone, and even intake up to ~10 g/day (about one drink) contributed 41 300 cases. The population form of no lower threshold: cancer cases accrue from the lightest drinking.
These are population counts, not per-person risks. A PAF answers how many cases across a population would not have occurred under lifetime abstention — prevalence combined with relative risk — not the risk faced by one drinker. It is therefore not commensurable with the cohort risk ratios in the per-outcome section: those size an individual’s change in risk; these size a national case load. Rumgay borrows its underlying RRs from WCRF’s cancer review and GBD from its own re-meta-analysis, so the four figures here add magnitude and reach, not effect sizes to line up against the earlier RRs.
Name the axis, and stop. GBD states the clash with published advice directly: «Our results show that the safest level of drinking is none. This level is in conflict with most health guidelines, which espouse health benefits associated with consuming up to two drinks per day» (Griswold et al., 2018). Neither side is in error; they stand in different places. Population guidance optimizes an average outcome under communicability and safety-at-scale constraints, while an individual reads the same evidence against personal baseline risk and preferences. This pressure runs toward guidance permitting more than a population-minimizing rule would; the appraisal names that it exists and which direction it runs, and stops short of pricing it.
Confidence, and where this could be wrong
How confident should any of this make someone? Moderate — GRADE-certainty MEDIUM, not high. Magnitude and certainty are two separate axes: the alcohol association is large and consistent across outcomes, but a large consistent association is not by itself a causal effect — the causal read here leans on Mendelian randomization (MR: genetic variants standing in for lifetime exposure, immune to reverse causation), and MR carries assumptions that can fail. So the strength of the finding does not buy it certainty; the two are priced independently.
Four things could still change the read here:
-
The genetic evidence is one study in one population. Millwood’s MR rests on the ALDH2 rs671 variant, common in East Asians and rare in Europeans, so the genetic instrument transports imperfectly to Western populations. What carries the causal read is not the MR alone but its convergence with the Western-heavy, bias-corrected observational meta-analysis — two independent routes, not one. (inferred from Millwood et al., 2019)
-
Measurement error flattens, it cannot manufacture. Drinkers underreport their intake, which makes a real dose-response gradient look shallower, never steeper. This cuts one way only: it works against a protective low-dose arm, so it cannot rescue the J-curve — the artifact verdict is if anything conservative. (inferred from Millwood et al., 2019)
-
The loop is open. This is a coherence-and-source-fidelity appraisal, not a test against realized outcomes: no step here grades a claim against what actually happened to people who did or did not drink. A clean appraisal is a well-warranted read of the evidence, not a validated prescription.
-
Four gaps are named, not filled. The fabric holds no dedicated dose-response arm for injury (it appears only inside GBD’s 23-outcome aggregate) and none for non-cancer liver disease (alcoholic liver disease / cirrhosis — only liver cancer and a MASLD-definitional ceiling are held); no meta-analysis compares beverage types at matched grams of ethanol, so the drinks are not ranked; and no acetaldehyde / IARC-monograph mechanism source is held, so any statement about how alcohol causes cancer stays marked. None of these is inferred in either direction — they are absences flagged so a reader knows where the evidence stops.
Mandatory caveats
- Open loop. This wiki grades internal coherence and fidelity to its sources — never whether a recommendation actually improves outcomes in the world. A clean appraisal is not a validated result.
- Appraise, do not prescribe. This page is a general, population-level appraisal, not medical advice. Selecting a target, screening for individual contraindications (pregnancy, liver disease, medication interactions, dependence), and managing them are prescriber acts, and they need information this document does not hold.
- A general appraisal, applied per person. Your sex, baseline cardiovascular and cancer risk, drinking pattern, dependence history, and realistic alternative decide the individual weighting. Alcohol’s rank is itself stratum-dependent: for a heavy or binge-pattern drinker it is a big rock — one of the few large, high-certainty levers — that dominates the ranking until it is addressed; for someone who barely drinks it is a small remaining lever, and reporting that ceiling — that there is little left to gain here — is itself the result, not a failure to find one. -> Shared Modifiable Levers Across Age-Related Diseases
- Coherence, not validity. The causal reading rests on the Mendelian-randomization assumptions (instrument validity, no pleiotropy) and on one MR study in one (East Asian) population; the convergence with the Western-heavy bias-corrected meta-analysis is what carries it.
- Health axis only. This cut weighs alcohol on longevity, cardiovascular, cancer, and cognitive outcomes; social ritual, pleasure, and cost are real and yours to weigh, not appraised here.
Evidence box
Question ’What is the effect of alcohol on each patient-important outcome (all-cause and CV mortality, cardiovascular disease, cancer, liver, cognition/brain, injury) — in which direction, how large, for whom, how certain — and what is the dose-response shape (monotonic, U/J-shaped, or threshold, and does the shape differ by outcome)? Does drinking pattern, beverage type, or matrix modify the effect, or is total ethanol the dominant axis?‘ Evidence included 12 sources — 6 gold, 5 high, 1 moderate Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 1 source(s) below the gold evidence bar feed this page: Semba (cohort, moderate). Each labelled by tier; none load-bearing for the core claims. Last updated 2026-09-04 · Independently reviewed: No · Full edit history