The effect/dose leg of the alcohol-cancer decision: how much a drinker’s risk rises, at which sites, and at what dose. Alcohol is a big-rock Layer-1 exposure and an IARC Group 1 (established) human carcinogen; this page holds the per-drinker site-specific dose-response, the companion to the population-burden leg on Alcohol and Mortality and Vascular Disease (WCRF grades · Rumgay case counts). The anchor is Bagnardi 2014, a gold dose-response meta-analysis of «572 studies, including 486 538 cancer cases» across 23 cancer types. (Bagnardi et al., 2014)
The cancer axis is not the mortality axis. The all-cause/vascular curve carries a J whose protective lower arm is largely artifact -> The U-Shaped Association Artifact. The cancer curve for the aerodigestive sites and breast is monotonic from zero, with no protective arm to adjudicate — so the U-artifact machinery does not transfer here. Do not import the mortality J-curve onto cancer.
Reading the doses
Bagnardi bins by the midpoint of each study’s exposure category, all RRs versus nondrinkers and occasional drinkers:
- Light — midpoint <=12.5 g/day (~<=1 standard drink; the meta-analysis equates 12.5 g = 1 drink).
- Moderate — midpoint <=50 g/day (~2-4 drinks/day).
- Heavy — midpoint >50 g/day (>~4 drinks/day).
The continuous dose-response (Figure 3) models risk to ~100 g/day; any RR below is stated over that studied range, not extrapolated. A category midpoint is not a threshold — it marks where the graded data sit.
Site-specific dose-response — RR by site and dose
Pooled RRs vs nondrinkers/occasional drinkers (Bagnardi 2014, Figure 2). «for those neoplasms there was a clear dose–risk relationship.» (Bagnardi et al., 2014)
| Site | Light | Moderate | Heavy | Studies |
|---|---|---|---|---|
| Oral cavity & pharynx | 1.13 (1.00-1.26) | 1.83 (1.62-2.07) | 5.13 (4.31-6.10) | 52 |
| Oesophageal SCC | 1.26 (1.06-1.50) | 2.23 (1.87-2.65) | 4.95 (3.86-6.34) | 54 |
| Larynx | 0.87 (0.68-1.11) ns | 1.44 (1.25-1.66) | 2.65 (2.19-3.19) | 41 |
| Liver | 1.00 (0.85-1.18) ns | 1.08 (0.97-1.20) ns | 2.07 (1.66-2.58) | 36 |
| Gallbladder | 1.23 (0.84-1.83) ns | 0.88 (0.68-1.13) ns | 2.64 (1.62-4.30) | 8 |
| Female breast | 1.04 (1.01-1.07) | 1.23 (1.19-1.28) | 1.61 (1.33-1.94) | 118 |
| Colorectum | 0.99 (0.95-1.04) ns | 1.17 (1.11-1.24) | 1.44 (1.25-1.65) | 66 |
| Stomach | 0.99 (0.92-1.06) ns | 0.97 (0.90-1.04) ns | 1.21 (1.07-1.36) | 39 |
| Pancreas | 0.95 (0.89-1.01) ns | 1.03 (0.97-1.09) ns | 1.19 (1.11-1.28) | 39 |
| Lung | 0.84 (0.79-0.88) | 0.98 (0.92-1.05) ns | 1.15 (1.02-1.30) | 34 |
The established causal set matches IARC/WCRF: «Alcohol increases risk of cancer of oral cavity and pharynx, oesophagus, colorectum, liver, larynx and female breast.» Pancreas, prostate and melanoma are the emerging set — «accumulating evidence» only (prostate 1.09 (0.98-1.21) heavy, driven by North American studies; melanoma heavy not evaluable). (Bagnardi et al., 2014)
The decision-relevant finding — risk is NOT confined to heavy use
For three sites the elevation is significant already at light drinking (~1 drink/day), so the curve rises from the first drink with no safe lower threshold: «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)
- Breast — light RR 1.04 (1.01-1.07): only a ~4% relative rise, but breast is the highest-incidence female cancer, so a small relative effect is a large absolute burden (Rumgay’s 98,300 breast cases below). A high-baseline-risk woman (family history) gains more absolute benefit from the same reduction -> Baseline Risk and the Relative-Absolute Split.
- Oral cavity/pharynx light 1.13 (1.00-1.26) and oesophageal SCC light 1.26 (1.06-1.50) — larger relative effects on lower-baseline cancers.
No safe threshold is SITE-SPECIFIC, not blanket. Colorectum, larynx, stomach, pancreas and lung show no significant light-drinking effect — their risk becomes significant only at moderate or heavy intake (a threshold-like feature). So no safe level is established for the aerodigestive-plus-breast sites, weaker or absent elsewhere. Stating it as a universal all-cancer claim overreads the data.
Mechanism — directional, genetic corroboration
(inferred from Bagnardi et al., 2014)
«Acetaldehyde, the first metabolite of ethanol, is accountable for part of the carcinogenicity of alcohol drinking on the liver and the upper aerodigestive tract.» (Bagnardi et al., 2014) The strongest corroboration is a genetic gradient (high on the mechanistic gradient — a natural-experiment flavour, though Bagnardi reports it as subgroup heterogeneity, not a formal MR): light drinking raised aerodigestive risk «in Asian countries only, where 28–45% of the population has a variation of the gene ALDH2» — the variant that slows acetaldehyde clearance. (Bagnardi et al., 2014) That the effect-modifier tracks the enzyme predicted by the mechanism is directional support for causation, discounted for the observational base and marked as mechanism, not asserted as an MR result. A dedicated Mendelian-randomization study of alcohol -> site-specific cancer is not held.
What discounts the magnitudes — observational caveats
The direction and no-threshold read are well-warranted (gold MA + IARC Group 1 + mechanism converge), but every estimate is observational, so the magnitudes carry uncertainty and the light-drinking figures most of all:
- Measurement error — self-reported intake; «an underreporting of alcohol consumption in drinkers may partly or largely explain the association with light alcohol drinking.» Ask what instrument produced the dose before trusting a light-dose RR -> Measurement Error in Dietary Assessment. (Bagnardi et al., 2014)
- Referent contamination biases toward the null (conservative) — former drinkers who quit while ill sit in the nondrinker referent, «thus diluting the risk of cancer among drinkers», so the true light-drinking harm may be understated, not overstated. (Bagnardi et al., 2014)
- Residual confounding by smoking — strongest where drinking and smoking co-occur (aerodigestive, lung, stomach). Lung is the tell: «alcohol consumption was not associated with lung cancer risk in never smokers», so the lung signal is plausibly confounded and the aerodigestive magnitudes (though mechanism-supported) inherit some of the same concern. (Bagnardi et al., 2014)
- Volume is not independence — 572 studies share dietary-assessment instruments and confounding structures; the pooled precision overstates the independent evidence.
The inverse-association sites are not a reason to drink
Kidney (light 0.92, moderate 0.79), thyroid (0.81) and the lymphomas (Hodgkin 0.63 heavy, non-Hodgkin 0.75) show inverse associations. (Bagnardi et al., 2014) These are single-site protective arms, and two facts neutralize them for the decision: (i) the lymphoma inverse is plausibly artifact — «early symptoms of lymphomas may cause subjects to either quit or reduce their drinking» (reverse causation) (Bagnardi et al., 2014); (ii) the kidney benefit is real but «far outweighed by the increased risk for other cancers» on the whole-organism ledger -> the net-effect-not-the-intended-effect rule (Alcohol and Mortality and Vascular Disease). A protective association at one site is not a threshold below which alcohol is safe.
The burden composite — Bagnardi x Rumgay (type F, not a tension)
Bagnardi (per-drinker RR) and Rumgay 2021 (population case burden) measure different quantities and are complementary legs, not rival estimates. The parameter table before any cross-source claim:
| Parameter | Bagnardi 2014 | Rumgay 2021 (IARC) | Same quantity? |
|---|---|---|---|
| What it measures | per-drinker RR vs nondrinkers, by site x dose | population cases attributable in 2020 (Levin PAF x incidence) | No — relative effect vs absolute burden |
| Unit | RR (dimensionless) | cases; % of all cancers | No |
| Light/moderate arm | breast light 1.04; aerodigestive light 1.13-1.26 | moderate <20 g/day 103,100 cases (13.9%); up to 10 g/day 41,300 cases | No — same DIRECTION (light not exempt), different quantity |
| RR provenance | its own pooled MA | borrowed from WCRF CUP | — not independent |
Rumgay: «741 300 … or 4·1% … of all new cases of cancer in 2020 were attributable to alcohol consumption», top sites oesophagus (189,700), liver (154,700), breast (98,300). (Rumgay et al., 2021) The composite beats either alone (type F): Bagnardi says how much per drinker per site, Rumgay says how many cases that is at population scale — and the two agree on the decision-relevant point that light-to-moderate drinking is not exempt. This is not type-E: Rumgay’s RRs are borrowed from WCRF’s CUP and Bagnardi is part of the same observational base, so they are not independent routes to one claim.
The decision
- Alcohol is a Layer-1 big rock; cancer is one leg of its harm ledger (mortality · vascular · cancer · dementia — see the nucleus). Unlike the contested mortality J, the cancer axis offers no protective arm to offset — it argues monotonically for less, from the first drink for the aerodigestive and breast sites.
- Frame as reduction/abstention. Absolute benefit scales with the person’s baseline cancer risk (route (a) — no subgroup claim needed): the same RR reduction returns more to a high-baseline individual. Even light drinkers hold a small, reducible aerodigestive/breast risk.
- Precision follows the decision. For a person choosing whether to drink at all, the site-resolved magnitudes matter less than the robust qualitative facts — established carcinogen, dose-dependent, no safe threshold for the aerodigestive/breast sites.
Synthesis
Value type: F (claim-refinement / quantification) — Bagnardi converts the qualitative IARC/WCRF alcohol causes these cancers grades into per-site dose-response RRs, and the Bagnardi x Rumgay composite pairs effect-size with population burden; the composite beats either source alone. A G-gap is named: no Mendelian-randomization source on alcohol -> site-specific cancer is held, so the ALDH2 genetic gradient stays directional mechanism, not a formal natural experiment. Not type-E — Bagnardi, WCRF and Rumgay share the observational evidence base (Rumgay’s RRs are WCRF’s). The cross-outcome pattern (no protective arm survives on mortality, stroke, dementia or cancer) is stated on the nucleus; this page adds the site-and-dose resolution and the explicit cancer-axis-vs-mortality-axis shape distinction -> The U-Shaped Association Artifact.