Opens the alcohol cluster. The famous finding that moderate drinkers outlive abstainers — the
J-curve — is the textbook case of a protective lower arm that dissolves under scrutiny. Two
independent methods say the protection is largely not real.
The observational J-curve — the appearance
- All-cause mortality (Stockwell, 87 studies). Unadjusted, the «classic J-shaped curve» appears: low-volume drinkers (1.3-24.9 g/day) «RR = 0.86, 95% CI [0.83, 0.90]» vs abstainers; occasional drinkers «0.84 [0.79, 0.89]»; former drinkers elevated «1.22 [1.14, 1.31]». (Stockwell et al., 2016)
- Vascular disease (Millwood, China Kadoorie, 500k). Self-reported intake had «U-shaped associations with the incidence of ischaemic stroke… intracerebral haemorrhage… and acute myocardial infarction; men who reported drinking about 100 g of alcohol per week… had lower risks of all three diseases than non-drinkers or heavier drinkers.» (Millwood et al., 2019)
Why the lower arm is an artifact — two independent adjudications
1. Bias-corrected observational (Stockwell). Correcting for how abstainers are defined erases the protection. The driving bias is former-drinker misclassification (sick-quitters who quit because they are ill, counted as abstainers, making the referent look unhealthy):
- fully adjusted for abstainer biases + study quality, low-volume «RR = 0.97, 95% CI [0.88, 1.07]» — no significant protection; former drinkers rise to «1.38 [1.24, 1.54]».
- the 13 bias-free studies show «no significant protection for low-volume drinkers… RR = 0.90, 95% CI [0.76, 1.06]»; using occasional drinkers (not abstainers) as the referent, «there would be no evidence of health protective effects for low-volume drinkers or any other category.»
- the corrected pattern is «more consistent with a linear dose response than a J-shaped curve.»
2. Mendelian randomization (Millwood). Genetic variants (ALDH2 rs671, ADH1B) that strongly lower alcohol tolerance act as instruments free of reverse causation and lifestyle confounding. The genotype-predicted (causal) relationship is monotonic, with no protective lower arm:
- ischaemic stroke RR per 280 g per week 1.27 (1.13-1.43); intracerebral haemorrhage «1.58 (1.36-1.84)»; total stroke «1.38 (1.26-1.51)» — «no suggestion of increased stroke risk at very low levels» that would indicate an abstinence penalty, and «no material deviation from log-linear relationships.»
- the causal mechanism is shared by both methods where the observational result is real: blood pressure rises 4.8 mm Hg (4.5-5.1) per 280 g per week (conventional) and «4.3 mm Hg (3.7-4.9)» (genetic) — convergent, so the confounding is specific to the protective claim, not to alcohol’s BP effect.
Millwood’s 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.» Stockwell’s: «low-volume alcohol consumption has no net mortality benefit compared with lifetime abstention or occasional drinking.» (Millwood et al., 2019)
Two independent methods, one conclusion [E-independent]
| Parameter | Stockwell 2016 | Millwood 2019 | Same claim? |
|---|---|---|---|
| Method | meta-analysis, bias-corrected | Mendelian randomization (genetic instruments) | independent — different data, methods, populations |
| Outcome | all-cause mortality | stroke, myocardial infarction | NO — different outcomes |
| Moderate-drinking effect | RR 0.97 (0.88-1.07), ns | monotonic harm (stroke); no protective arm | — |
| The claim reached | no net protection from low-volume drinking | protective J-arm is non-causal | YES — moderate drinking is not cardioprotective |
The convergence is on the qualitative claim (the protective lower arm is not causal), reached by a bias-correction route and a genetic route that share no data or lineage and neither references the other as antecedent — a genuine independent-backing lift, not a shared-study echo. The effect sizes are not matched (different outcomes), so this is an E on the direction, not a pooled magnitude. Corroborated across The U-Shaped Association Artifact.
The updated meta-analysis, and where harm starts [Zhao 2023]
Zhao et al. 2023 updates the observational analysis to 107 studies, 4.8 million people and confirms the artifact finding: fully adjusted, low-volume drinking is «RR, 0.93; 95% CI, 0.85-1.01; P = .07» vs lifetime abstainers and only «a RR of 0.97» against the cleaner occasional-drinker referent — no significant protection. Only «21» of 107 studies were free of abstainer bias, and stripping the covariates drives the low-volume estimate to «0.86 (0.83-0.88)» — the protection is manufactured by the biases.
What it adds that the corpus lacked — an actionable harm threshold, and it is sex-specific. Significant all-cause mortality elevation begins at:
- pooled: 45 g/day — «45 to 64 g per day (RR, 1.19; 95% CI, 1.07-1.32)», rising to «1.35» at >=65 g/day; 25-44 g/day pooled is null (1.05, ns);
- women: from 25 g/day — medium-volume «RR, 1.21; 95% CI, 1.08-1.36», and any drinking is already elevated («1.22; 95% CI, 1.02-1.46»);
- men: from 45 g/day — 25-44 g/day is null for men.
- «Across all levels of alcohol consumption, female drinkers had a higher RR of all-cause mortality than males (P for interaction = .001).»
This is the first how-much-is-too-much number the corpus holds for alcohol, and the sex asymmetry is the decision-relevant part — a woman’s threshold sits at roughly half a man’s.
F, not a third E witness (laundering guard). Zhao 2023 is Stockwell’s own group — it «updates our
earlier systematic review» (its ref 8 is Stockwell), same
team and method. It refines the observational arm (more studies, occasional referent, the
threshold); it is not an independent method, so the [E-independent] convergence above stays
Stockwell ∥ Millwood, unchanged.
(Zhao et al., 2023)
The honest boundary — stroke is not myocardial infarction
MR shows clear causal harm for stroke and blood pressure, but for myocardial infarction the genetic estimate is null: RR per 280 g per week 0.96 (0.78-1.18), p=0.69 — «little net effect». Millwood suggests alcohol’s BP harm «could be offset by cardio-protective changes in other factors», and cautions the MI case count was limited so «some real benefit or hazard cannot be excluded». So “alcohol is uniformly harmful” holds for stroke and BP; for coronary heart disease the causal picture is genuinely unresolved, and that is the one place a small real benefit is not excluded. (Millwood et al., 2019)
Interventional confirmation — reducing alcohol lowers BP, confined to >2 drinks/day [Roerecke 2017]
The BP slope above (Millwood) is observational + genetic. Roerecke’s 2017 SR+MA of 36 randomised trials (2865 participants) supplies the missing interventional leg — the RCT dose-response of reducing alcohol on blood pressure, with the dose-response graded GRADE-high: «we rate the quality of evidence for a dose-response association as high». (Roerecke et al., 2017)
- Overall, cutting alcohol lowered blood pressure by «–3·13 mm Hg (95% CI –3·93 to –2·32) for systolic blood pressure and –2·00 (–2·65 to –1·35) for diastolic blood pressure with substantial between-study heterogeneity (I² 82·0% and 79·5%, respectively)» — heterogeneity that baseline intake almost entirely explained (75.4% SBP, 93.4% DBP variance). (Roerecke et al., 2017)
- Dose-dependent, with a threshold at ~2 drinks/day. Stratum SBP means by baseline band (Table 2, all trials): <=2 drinks/day -0.18 (-1.02 to 0.66), NS; 3 drinks -1.18 (-2.32 to -0.04); 4-5 drinks -3.00 (-3.98 to -2.03); >=6 drinks -5.50 (-6.70 to -4.30) [DBP >=6: -3.97 (-4.70 to -3.25)]. (Roerecke et al., 2017) Meta-regression: «β = –0·91 mm Hg» SBP (and −0·75 DBP) «per one drink per day» of baseline intake, p<0·0001. (Roerecke et al., 2017) Studied range: baseline weighted means 15 / 30 / 49 / 76 g/day across the four strata (single trials to ~380 g/day) — so the curve is estimated over the drinking range, not extrapolated.
- Below the threshold the lever is empty — the decision-relevant part. «People drinking two drinks or fewer per day did not have a signifi cant reduction in blood pressure when they reduced their alcohol consumption to near abstinence, suggesting that this amount of alcohol intake does not increase blood pressure.» (Roerecke et al., 2017) So the alcohol->BP lever is confined to the >2 drinks/day stratum; for a light drinker, cutting alcohol buys no BP — a route-(a)/(b) stratification, not a universal recommendation. The dose-response held «in healthy participants and people with hypertension or other CVD risk factors». (Roerecke et al., 2017)
Third independent leg for the alcohol->BP causal claim [E-independent]
The BP slope on this page was already reached two ways (Millwood conventional-observational and genetic MR). Roerecke adds a third, independent method — pooled RCTs of alcohol reduction — converging on the same causal claim. No author overlap with Millwood (Roerecke/Kaczorowski/Tobe/Gmel/Hasan/Rehm vs Millwood/Walters/Chen), different design, data, and direction of manipulation, and neither references the other as antecedent — a genuine independent-backing lift, not a shared-study echo.
| Parameter | Millwood 2019 (conventional + genetic MR) | Roerecke 2017 (RCT interventional) | Same quantity? |
|---|---|---|---|
| Route | observational cohort + Mendelian randomization | random-effects MA of 36 alcohol-reduction RCTs | no — independent designs (the point of E) |
| Manipulation | genetically/self-reported higher vs lower intake | reduction of usual intake toward abstinence | mirror image (increase vs decrease) |
| SBP effect | +4.8 (conv, 4.5-5.1) / +4.3 (genetic, 3.7-4.9) mm Hg per 280 g/week | -0.91 mm Hg per one drink/day baseline (β) | roughly — 280 g/wk ~= 3.3 drinks/day, so Millwood ~1.3-1.4 vs Roerecke ~0.91 mm Hg per drink/day: same sign + order of magnitude |
| Shape | log-linear (MR assumes linearity) | dose-dependent with a threshold at 2 drinks/day | no — Roerecke adds a threshold the MR cannot see (F) |
| Establishes | alcohol causally raises BP (reverse-causation-immune) | reducing alcohol lowers BP (interventional) | same causal claim, reached independently -> E |
The alcohol->BP causal effect is thus triangulated across observation, genetics, and randomisation, and Roerecke refines the Millwood log-linear picture with a threshold (an F riding on the E): the BP effect is a >2-drinks/day phenomenon, flat below it. (inferred from Millwood et al., 2019; Roerecke et al., 2017)
But BP is the SURROGATE — the CV events are modelled, not measured
Roerecke measured blood pressure, a surrogate. Its downstream cardiovascular payoff is a model, not a trial outcome: for the UK it projected «more than 7000 inpatient hospitalisations and 678 cardiovascular deaths prevented» per year — but that figure rests on «standard methods for comparative risk assessment analyses (eg, Global Burden of Disease studies)», run «assuming persistence of lower alcohol consumption within 1 year, and no lag time of eff ects on CVD outcomes». (Roerecke et al., 2017) The BP->event transmission is assumed via the risk-assessment machinery, not observed; no hard CV endpoint was measured, and «the physiological mechanisms for alcohol’s eff ect on blood pressure and hypertension are still unknown». (Roerecke et al., 2017) The BP drop is «similar to that of other health behaviour changes, such as physical activity» or weight-loss diets (Roerecke et al., 2017) — the BP->event step itself is adjudicated at Blood Pressure Lowering and Cardiovascular Events.
No protective arm for the BP outcome — the interventional check the artifact concept demands. Where alcohol’s mortality/IHD J-curve carries a spurious protective lower arm (above), the BP dose-response has none: reducing alcohol never raises BP, and below the 2-drink threshold it is simply flat. A randomised-reduction design is the strongest form of the interventional check -> The U-Shaped Association Artifact asks for — but note the outcome is distinct (BP, not mortality), so this confirms the concept on a different endpoint, it does not itself adjudicate the mortality J.
What this probes [PRIOR handle — not scored here]
Alcohol is the canonical instance of the telos’s open [PRIOR] #2 — U/J-shaped associations are
frequently artifacts of reverse causation or confounding. This ingest builds the evidence that in the
alcohol case the prior holds (the protective arm is artifact, shown two ways). Per the ingest
contract, the fabric records this; the [PRIOR] is scored in a separate pass, against this and the
still-open sodium J-hypothesis -> Sodium Intake and Blood Pressure, The U-Shaped Association Artifact.
Second outcome — dementia, and the same J-curve artifact replicates
Excessive alcohol is one of the 14 modifiable dementia risk factors -> Dementia Prevention and Modifiable Risk Factors, and the dementia literature reproduces this page’s verdict on a new outcome. Heavy midlife drinking (>21 units/week) raises dementia risk (IPD «HR 1.22, 1.01-1.48»), while the apparent protection of light-vs-none is again an artifact — the observational J-curve «is probably because many non-drinkers have previously had high alcohol consumption» or other illness that stopped them drinking. (Livingston et al., 2024) Mendelian randomization agrees: it «finds a causal relationship between alcohol consumption and earlier age of onset of AD», and «suggests that any relationship between not drinking and AD is due to survivor bias». The same sick-quitter / referent-contamination + MR adjudication that dissolved the mortality J-curve dissolves the dementia one -> The U-Shaped Association Artifact. This is a second outcome reaching the same conclusion by a different literature (dementia cohorts + AD-MR), not a restatement.
The dementia dose-response, first-hand — the J-curve quantified, and its protective arm UNADJUDICATED at source [2026-09-04, Xu]
Livingston’s dementia verdict above is qualitative and secondhand. Xu 2017 supplies the first-hand dose-response — a gold MA of 11 prospective cohorts, 73,330 participants, 4586 all-cause-dementia (ACD) cases (NOS>=8), a dedicated dose-response study of the alcohol -> dementia curve (the Commission’s own HR 1.22 harm cell traces to a separate Whitehall IPD, not to Xu). It draws the exact J-curve Livingston describes:
- The shape (ACD, p_nonlinearity < 0.05): a protective lower range of «0 and 7.5 drinks/week (Fig. 3a) or 12.5 g/day», nadir RR ~= 0.9 at roughly 6 g/day / 4 drinks/week, and a harm knee above 23 drinks/week or 38 g/day. (Xu et al., 2017) Studied-range honesty: the nadir and harm knee are interior spline features read off Fig. 3b’s confidence band — the nadir RR ~0.9 is NOT tabulated with a numeric CI, so it is a curve feature, not a precise point estimate, and the g/day arm leans on a single unit-native cohort plus drinks->grams conversion.
- Protection is wine-only — «current drinker (RR 0.67; 95% CI 0.48-0.94) or light-to-moderate drinker (RR 0.58; 95% CI 0.39-0.87)» for wine, while beer highest-vs-lowest is elevated (RR 1.84, 1.01-3.34) and liquor is null. (Xu et al., 2017) A benefit confined to the health-conscious drinker’s beverage is the healthy-user signature, not an ethanol dose-response — and the resveratrol/polyphenol rescue Xu offers is the one the held Semba biomarker null already refutes (above).
The protective arm is UNADJUDICATED — Xu ran only the WEAK checks and concedes every strong-check gap [inferred from @xu2017alcohol]. This is what the first-hand source adds over the secondhand cell: the dose-response referent is the sick-quitter-contaminated «lowest category» (mostly abstainers), and Xu concedes «we cannot exclude the potential influences of including former drinkers, who may quit drinking due to underlying diseases and have a high risk of dementia, in the reference group»; no referent-correction and no Mendelian randomization were run — Xu names the missing check outright («warrants further validation using more advanced approach, such as Mendelian randomization») — plus conceded residual confounding and self-report misclassification. [EXTRACTED (Xu - Alcohol Consumption Dementia 2017) chunk 01] So at the primary-source level the dementia protective arm has cleared only the checks the mortality J-curve also survived before referent- correction + MR removed it; it is the held AD-MR (Livingston, above) that supplies the strong check adjudicating it as artifact, not Xu. The harm arm (>38 g/day) is the causally coherent half — vascular/hypertensive-mediated, consistent with this page’s stroke/BP verdict.
Type-F, not independent-E. Xu is a single observational dose-response MA on the same cohort
literature and the same sick-quitter/reverse-causation machinery as the mortality instances — a
cross-outcome worked case, not a second independent witness; it does not lift the page’s
[E-independent] convergence (Stockwell || Millwood). (inferred from Xu et al., 2017)
Third outcome — cancer, and here there is no safe threshold
WCRF’s Third Expert Report grades alcohol a cause of many cancers — mouth/pharynx/larynx, oesophagus (squamous cell), breast, colorectum, liver and stomach (convincing/probable) — and IARC classes alcoholic drinks a Group 1 (established) human carcinogen. (World Cancer Research Fund & American Institute for Cancer Research, 2018) The recommendation is blunt: «For cancer prevention, it’s best not to drink alcohol.»
The cancer axis behaves differently from the mortality axis on threshold. Where Zhao’s all-cause harm begins at a dose (45 g/day pooled, 25 g/day women), the cancer risk has no lower threshold: «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) The site-specific numbers in the matrix are the edge of the evidence base, not safe thresholds (the CLAUDE.md rule: a guideline number is first a marker of where the data thin out): breast «No threshold … was identified» (FN38); the colorectal judgement is based on intakes above ~30 g/day (~2 drinks, FN37) and the liver/stomach judgements on intakes above ~45 g/day (~3 drinks, FN36) — i.e. that is where the graded evidence sits, not a level below which alcohol is safe. (World Cancer Research Fund & American Institute for Cancer Research, 2018)
A genuine J-arm — instantly outweighed. Alcohol «helps protect against kidney cancer (at least up to 30 grams or two drinks per day), but this is far outweighed by the increased risk for other cancers.» (World Cancer Research Fund & American Institute for Cancer Research, 2018) So a real protective association can coexist with net harm — the net-effect-not-the-intended-effect rule: a single-site benefit is not a reason to drink when the whole-organism cancer ledger is negative.
Not an independent-E witness (same-body / outcome-extension guard). WCRF is a guideline synthesis
of the same observational cohort literature, on a different outcome (cancer incidence) — it neither
shares nor is independent of Stockwell/Millwood/Zhao’s methods. It extends the outcome menu (cancer,
with a stronger no-threshold claim) rather than corroborating the mortality/stroke finding by an
independent route, so the page’s [E-independent] convergence (Stockwell ∥ Millwood) is unchanged.
The cross-outcome pattern — no protective arm survives on mortality, stroke, dementia, or cancer — is
the wiki’s synthesis, not WCRF’s claim. (inferred from World Cancer Research Fund & American Institute for Cancer Research, 2018)
The per-drinker site-specific dose-response (which cancers, at what dose, with light-drinking risk for aerodigestive sites and breast) lives on Alcohol and Cancer Risk — Bagnardi 2014’s gold dose-response MA quantifies WCRF’s qualitative grades into RRs by site, and pairs with Rumgay’s burden leg (below) as the effect x burden composite.
A distinct axis — drinking PATTERN, holding volume fixed [Roerecke 2010]
Every arm above is indexed to average volume (grams/week, drinks/day). Roerecke’s SR+MA (14 studies, 31 estimates, 4718 IHD events) opens a second axis the volume curves cannot see: at the same average intake, concentrating the alcohol into irregular heavy occasions carries higher ischaemic-heart-disease risk than spreading it out.
- Irregular heavy drinking occasions («60 g of pure alcohol or 5 drinks per occasion at least monthly») vs regular moderate drinking at comparable average volume: pooled random-effects «relative risk of irregular heavy drinking occasions compared with regular moderate drinking was 1.45 (95% confidence interval: 1.24, 1.70)» (fixed-effects 1.36), I²=53.9%; a detrimental effect «even for drinkers whose average consumption is moderate.» (Roerecke & Rehm, 2010)
- The comparison excludes abstainers and former drinkers, so this is not a sick-quitter artifact: “Because we did not include an abstainer group in our analysis and used risk estimates that separated former drinkers from their analysis, it is unlikely that a sick-quitter effect … influenced our findings.” And it is conservative — non-differential misclassification biases the pooled estimate toward the null. (Roerecke & Rehm, 2010)
- Roerecke’s verdict: «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)
Why this is a refinement, not a tension with the artifact verdict above — the contrasts are different quantities:
| Parameter | Roerecke 2010 | Millwood 2019 (this page) | Same quantity? |
|---|---|---|---|
| Contrast | irregular-heavy vs regular-moderate, at equal average volume | genetically higher vs lower intake (MR) | NO — pattern contrast vs volume contrast |
| Referent | regular moderate drinkers | genetically lower-alcohol | NO |
| Outcome | ischaemic heart disease (MI + coronary death) | myocardial infarction (and stroke) | partial |
| Effect | RR 1.45 (1.24, 1.70), heavy-episodic vs regular | MI RR 0.96 (0.78-1.18) per 280 g/wk, null | NO — different comparison |
The fourth column is NO on three rows: Roerecke measures a within-drinker pattern effect at fixed volume; Millwood measures the drinking-vs-not causal effect. They are consistent — a given volume can still be worse when binged even if drinking carries no net causal MI benefit. What Roerecke presumes (that regular-moderate is cardioprotective vs not drinking) inherits this page’s artifact critique; what it robustly shows — pattern modifies risk at fixed volume — stands regardless. Honest composite: volume is the dominant axis, pattern is a real second axis, and the decision-relevant form is don’t concentrate a week’s drinks into one or two heavy sessions — a lever available even to someone unwilling to cut total intake.
This fills a gap GBD names explicitly. GBD flagged the hole itself: «drinking patterns within a year are assumed to be consistent; however, past work shows that drinking patterns, rather than average levels of consumption such as standard daily drinks, might be related to different levels of risk». (Griswold et al., 2018) Roerecke (2010) supplies exactly the pattern-risk function GBD’s population model could not — the composite covers a blind spot GBD’s own model explicitly names (an F-refinement across the two), not a contest. (Griswold et al., 2018; inferred from Roerecke & Rehm, 2010)
The beverage-matrix story — it’s the red wine / resveratrol is null at dietary doses [Semba 2014]
A common rescue of the J-curve is that the benefit is not ethanol but the polyphenols in red wine — the French paradox, attributed to resveratrol. Semba’s InCHIANTI cohort (783 adults 65+, 9-year follow-up) tests it with a biomarker instead of a food-frequency proxy: 24-hour urinary resveratrol metabolites.
- Resveratrol was null on every outcome: «total urinary resveratrol metabolite concentration was not associated with inflammatory markers, cardiovascular disease, or cancer or predictive of all-cause mortality.» Lowest-vs-highest quartile mortality HR «0.80 (95% CI, 0.54-1.17)», ns; incident CVD and cancer also flat across quartiles. (Semba et al., 2014)
- The verdict is scoped to dietary doses: «Resveratrol levels achieved with a Western diet did not have a substantial influence on health status and mortality risk of the population in this study.» (Semba et al., 2014)
- The biomarker is really a wine-intake marker: urinary resveratrol correlated with alcohol intake at Spearman «0.67 (P < .001)», and «urinary resveratrol levels are a valid biomarker of wine consumption.» (Semba et al., 2014)
So the presumed active component of red wine does no measurable work at the doses a drinker actually gets — the same failure mode as coffee’s caffeine (present but inactive for the outcomes) or the refined-vs-whole-grain boundary (null on hard outcomes) -> Is the Food Category Doing Any Work: a category benefit credited to a named sub-component that, measured directly, carries nothing. It does not rule out a supraphysiologic-dose supplement effect (the trials use doses orders of magnitude above dietary, and one was halted early for renal toxicity) — but it removes the drink-red-wine-for-the-resveratrol inference. Whatever ethanol does (stroke, BP, cancer, above) it does regardless of beverage; the wine-specific benefit is unmeasured-or-absent, not established — do not let a well-lit polyphenol biomarker stand in for the outcome that matters (the streetlight caveat). Semba is a single moderate-tier cohort (a labelled null counterweight — no gold resveratrol-outcome MA exists), so it refutes a folklore claim; it is not an independent-E witness for the mortality/stroke verdict. (inferred from Semba et al., 2014)
Population scale — no safe level, and PAF is not a per-person effect [GBD 2018; Rumgay 2021]
The arms above are per-person risks. Two population-attributable-fraction (PAF) models add the population magnitude — a different unit, not a larger version of the RRs, and read as such.
GBD 2016 — the external guidance-null anchor. Across all 23 outcomes weighted 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 drink = 10 g ethanol); alcohol «led to 2·8 million deaths» in 2016 and was the leading risk factor at ages 15-49. (Griswold et al., 2018) GBD reaches this by re-doing the meta-analysis with a controlled reference category — the same sick-quitter correction Stockwell/Zhao applied — so it is population-scale reinforcement of the same artifact finding, not an independent method. It does find a residual protective minimum for IHD and diabetes («0·86 (0·80-0·96) for men and 0·82 (0·72-0·95) for women» at ~0·9 drinks/day) but «these protective effects were offset by the risks associated with cancers, which increased monotonically». (Griswold et al., 2018) It states the guidance clash plainly: «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)
Rumgay 2021 (IARC) — the cancer-burden leg, and it reinforces no-threshold with scale. «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», top sites oesophagus (189 700), liver (154 700), breast (98 300). Crucially for the no-safe-threshold claim (WCRF, above), light-moderate drinking is not exempt: «moderate drinking (<20 g per day) contributed 103 100 (13·9%; 95% UI 82 600-207 200) cases», i.e. «moderate drinking still contributed one in seven alcohol-attributable cases and more than 100 000 cancer cases worldwide», and even «drinking up to 10 g per day contributed 41 300 (35 400-145 800) cases» (~1 drink). (Rumgay et al., 2021)
Read PAF correctly (the unit guard). A PAF is how many population cases would not have occurred under lifetime abstention — prevalence × RR — NOT the risk to an individual drinker, and not commensurable with the cohort RRs on this page. Rumgay’s RRs are borrowed from WCRF’s CUP; GBD’s are its own re-meta-analysis. So these two enter the page as co-membership (the same cross-outcome verdict at population scale), not as effect sizes to line up against Stockwell/Millwood/Zhao. What they add is the magnitude and the guidance-null exercise: even the fraction attributable to moderate drinking is a six-figure case count — the population-scale form of no protective arm survives, and cancer has no lower threshold. (inferred from Griswold et al., 2018; Rumgay et al., 2021)
A different outcome axis — alcohol drives EATING UP, and its own calories are not compensated [Kwok 2019]
Every arm above indexes a disease outcome (mortality, stroke, BP, dementia, cancer). Kwok’s SR+MA (22 crossover/RCT studies, 701 participants, younger adults 18-37 y) opens a distinct energy-balance / weight channel the disease curves cannot see: what alcohol does to subsequent food intake in the same occasion — the aperitif effect / passive over-consumption. The finding is that alcohol’s energy is additive, not compensated:
- No compensation. «All twenty-two studies consistently demonstrated that participants did not reduce their food energy intake to compensate for the energy consumed from alcoholic beverages.» (Kwok et al., 2019)
- Food intake rises, total intake rises more. Alcohol vs a non-alcoholic comparator increased «food energy intake and total energy intake … by weighted mean differences of 343 (95 % CI 161, 525) and 1072 (95 % CI 820, 1323) kJ, respectively». (Kwok et al., 2019) Food EI +343 kJ (~82 kcal, 12 studies, I2=82.5%); total EI +1072 kJ (~256 kcal, 8 studies, I2=73.7%). The total exceeds the food increase because the beverage’s own energy is added on top and not offset — that gap is the non-compensation made quantitative. Decision framing: «a relatively modest alcohol dose may lead to an increase in food consumption». (Kwok et al., 2019)
- Not confined to heavy drinking. Subgroup by dose: «both low-dose and high-dose alcohol increased food energy intake. Low-dose alcohol increased food energy intake to a greater extent compared with high-dose alcohol, although overlapping 95 % CI were observed.» (Kwok et al., 2019) No subgroup modifier (dose, sex, comparator type) was established — every between-group difference had overlapping CI — so this is not a low-dose-exempt effect. (Low dose here = <30 g or <0.6 g/kg; a route-(a) prognostic split, not a demonstrated effect-modification.)
This is an amplifying-direction instance of Net Effect vs Intended Effect. The compensation literature usually runs the attenuating way — the body offsets an intervention, shrinking its naive effect (Exercise Energy Compensation: added exercise partly compensated by eating more / moving less). Alcohol is the mirror: the intervention (a drink) provokes the organism to consume more, so its naive energy cost UNDER-states the net energy surplus. Same principle (net != intended via whole-organism response), opposite sign.
Symmetric-standards caveats — this cascade inflates alcohol’s harm, so it is held to the protective claims’ bar:
- Acute, not chronic. These are within-session crossover feeding-study outcomes, not long-term intake or a hard endpoint. The intake -> weight step is inferred, not measured: body-weight change was reported in only 3 short trials (13 d-10 wk), 2 null and one «increased significantly by a mean of 0·9 (SE 0·4) kg» that «could not be directly attributable to the consumption of a specific beverage». (Kwok et al., 2019)
- The food-EI arm carries upward-bias risk; «the asymmetrical funnel plots and Egger’s regression test suggests small-study effects may exist in the meta-analysis for food energy intake (Egger’s test: P = 0·002)», and a HKSJ sensitivity analysis widened the CI though it «did not change our conclusion (weighted mean difference 343 kJ, 95 % CI 109, 577 kJ)». (Kwok et al., 2019) The total-EI arm had no small-study effects (Egger P=0.8), so the larger, more decision-relevant estimate is the cleaner one.
- Mechanism is unsettled (expectancy, disinhibited restraint, satiety hormones, CNS neurotransmitter pharmacology — all flagged «unclear» by the authors), so the direction (intake up) is what transports, not any one pathway. (inferred from Kwok et al., 2019)
A likely second route to the same weight channel — flagged, not asserted. Beyond this direct intake effect, a common belief holds that a nightcap disrupts sleep architecture and worse sleep raises next-day intake (alcohol -> sleep -> eating). The fabric takes no position on that route — it is an open gap awaiting an alcohol -> sleep-architecture SR (no such source is held yet), and must not be read as established here. Kwok evidences only the direct within-occasion cascade, not the sleep-mediated one.
Not an independent-E witness. Kwok extends the outcome menu (energy balance / weight) with a
different endpoint; it neither corroborates nor is independent of the mortality/stroke methods, so the
page’s [E-independent] convergence (Stockwell || Millwood) is unchanged. What it adds is a fourth
non-disease consequence consistent with the whole page: alcohol offers no protective arm on mortality,
stroke, dementia, or cancer, and on energy balance it actively pushes intake up.
Limits
- All-cause mortality (Stockwell) and vascular disease (Millwood) are different endpoints — matched only at the level of is moderate drinking protective. Neither covers the other’s outcome set.
- Measurement error flattens, it does not manufacture. Both note self-reported intake is underreported; Millwood states this would make the real dose-response shallower, not create the monotonicity — so it cannot rescue the protective arm.
- Millwood is one MR study in one (East Asian) population; ALDH2 is common there and rare in Europeans, so the instrument transports imperfectly. The convergence with a Western-heavy observational meta-analysis is what carries it.
- Coherence, not validity (R1): the causal read rests on the MR assumptions (instrument validity, no pleiotropy — Millwood checks the latter via women as a negative control).
A guidance family has abandoned the moderate-drinking position — NNR2023 [2026-08-27, NNR revisit]
The Nordic Nutrition Recommendations 2023 give the guidance-family read that matches this page’s verdict: no protective floor is granted, and abstention is the reference. NNR’s recommendation is that «NNR2023 recommends avoiding alcohol intake. If alcohol is consumed, the intake should be very low.» (Nordic Council of Ministers, 2023) The warrant is the threshold claim this page’s artifact-verdict rests on — its science advice states «No safe lower limit for alcohol consumption has been established.» (Nordic Council of Ministers, 2023), reached «since no threshold for safe level of alcohol consumption has currently been established for human health» (Nordic Council of Ministers, 2023).
- Classification: guidance-family agreement (Layer-1 attribution), NOT independent backing. NNR
is a guideline synthesis resting on the same evidence base already on this page — GBD’s safest
level is none, WCRF/IARC’s no-safe-threshold-for-cancer, and the standard cohort/MR literature —
so it does not raise confidence the way an independent method would. No
[E-independent]tag. Its value is that a national/regional guidance body has, in its current cycle, dropped the moderate-drinking-is-fine position — the guidance null on this question now agrees with the fabric rather than opposing it. - Counter-passage check. NNR states no protective exception (no cardioprotective carve-out for moderate intake), so there is no divergence to file; the position is fully aligned. It is a population-standpoint recommendation (guidance divergence class 1) framed as harm-minimization, consistent with the artifact read of the lower J-curve arm.
No safe level refined to an age- and region-conditional threshold — the same-body revision [Bryazka 2022, GBD 2020]
The «no safe level» anchor on this page is GBD 2016 (Griswold, above). The GBD 2020 Alcohol Collaborators (Bryazka, lead) re-ran the analysis one cycle later and conditioned the threshold on background disease rates — the beyond-summary move. Instead of one global optimum, it estimates a theoretical-minimum-risk exposure level (TMREL) and a non-drinker-equivalence level (NDE) per region x age x sex x year from burden-weighted dose-response curves across 22 outcomes. One standard drink = «10 g of pure ethanol». (Bryazka et al., 2022)
- The young: still ~zero. For ages 15-39 (2020) the TMREL ranged 0 (95% UI 0-0) to 0·603 (0·400-1·00) standard drinks/day and the NDE 0·002 (0-0) to 1·75 (0·698-4·30) across 21 regions — near-abstinence minimises harm because the young carry little CVD burden and alcohol’s damage is injury/violence-dominated. Bryazka is «leading risk factor for mortality among males aged 15-49», 1·78 million (1·39-2·27) alcohol deaths in 2020. (Bryazka et al., 2022)
- 40+: a non-zero, J-shaped threshold. For ages 40+ (2020) the burden-weighted RR curve was «J-shaped for all regions», TMREL 0·114 (0-0·403) to 1·87 (0·500-3·30) standard drinks/day and NDE 0·193 (0-0·900) to 6·94 (3·40-8·30) — a non-zero optimum arises only where high background CVD/diabetes rates let alcohol’s small modelled benefit on those outcomes offset its harms. (Bryazka et al., 2022) (Sampled range 0-100 g/day pure ethanol; the upper regional TMREL bounds sit near the studied edge and carry wide UIs — read them as region-conditional, never a personal target.)
- The global headline moved off zero — but by re-weighting, not new causal evidence. Global age-standardised both-sexes TMREL: 0 (0-0·800) with GBD-2016 RRs+DALYs -> 0·534 (0-1·00) once only the DALY weights update to 2020 -> 0·511 (0·400-0·700) with the updated RRs too. (Bryazka et al., 2022) Most of the shift is the changing disease-composition weight (0 -> 0·534), not the RR update (0·534 -> 0·511) — the optimum moved because the population’s disease mix was re-weighted, not because new evidence made alcohol more protective.
- The decision shift Bryazka draws: stratify guidance by AGE, not SEX. TMREL/NDE did not vary significantly by sex or year but varied strongly by age; the estimates «do not support low consumption guidelines that differ by sex». (Bryazka et al., 2022) (Males were still 76·9% [73·0-81·3] of harmful-amount consumers — a prevalence fact, not a threshold difference.)
Classification — same-body F-refinement that partly contradicts, NOT independent-E. Bryazka is the
GBD 2020 Alcohol Collaborators; Griswold (above) is the GBD 2016 Alcohol Collaborators — the
same collaboration, next cycle, and 17 of the 22 outcome RR curves are carried over directly from GBD
2016 (only five — ischaemic heart disease, ischaemic stroke, intracerebral haemorrhage, type 2
diabetes, tuberculosis — were re-estimated on +71 studies; lower respiratory infection was dropped for
insufficient evidence). (Bryazka et al., 2022) So this
is a within-collaboration revision on largely shared inputs, not an independent method — no
[E-independent] tag. It is a type-F refinement (the universal «no safe level» is bounded to an
age/region-conditional threshold) that partly contradicts (the strict global «zero» is overturned for
older adults in high-CVD regions). What survives: ~zero for the young everywhere, and the qualitative
artifact verdict below.
The 40+ protective arm is a MODEL OUTPUT, not established as causal — the U-shape gate. The J-shape at 40+ is a burden-weighting of the observational IHD/T2D dose-response RR curves, so it inherits whatever abstainer / sick-quitter / reverse-causation bias sits in those inputs -> The U-Shaped Association Artifact. Bryazka’s inputs are not referent-corrected — it uses a «reference group of non-drinkers» (Bryazka et al., 2022), the sick-quitter-contaminated referent that Stockwell/Zhao showed erases the protection, not the never-drinker/occasional referent. MR-BRT carries a bias covariate for whether an input study adjusted for sick-quitter bias, but Bryazka concedes «it is possible that relative risk estimates did not account and adjust for all sources of bias, including measurement bias and selection bias, as well as the potential impacts of reverse causality». (Bryazka et al., 2022)
Decisively, Bryazka itself reports the strong check largely nulls the protection: a recent Mendelian
randomization meta-analysis had «67% of studies on cardiovascular disease and 75% of studies on diabetes
reporting a null association with alcohol».
(Bryazka et al., 2022) So the non-zero 40+ TMREL does
not validate a real protective effect — a J surfacing in a burden-weighted model is neither a
referent-correction nor a genetic check, and by this page’s own adjudication (Millwood MR: monotonic
stroke harm, null MI) the cardioprotective arm feeding it is largely artifact. The age-conditioning is
the decision-relevant refinement; the protective arm is not. A same-collaboration update on borrowed
RRs refines scope without adding an independent method, so page confidence: stays medium — the
artifact verdict is reinforced, not the protective claim.
(inferred from Bryazka et al., 2022; Griswold et al., 2018; Millwood et al., 2019)