Alcohol and longevity: what the cohorts say and what the genetics say

Self-reported drinking has a J-shaped curve against mortality in 40,386 Chinese men, and the same study's Mendelian randomization finds a linear harm signal: 27 per cent more cardiovascular disease and 20 per cent more death per standard drink once heavy drinkers are excluded. The burden-of-proof review of alcohol and ischaemic heart disease reaches the same verdict — the cohort protection is not there once bias is priced in.

mixedStudies disagree, or the mechanism is contested.Revised 25 Oct 2026

A single empty plain unlabelled wine glass and a plain water glass side by side on a dark wooden table beside a folded grey linen napkin

Every J-shaped curve in this field has the same kink at zero: the reference group is people who chose not to drink, not people who were assigned not to.

There is a curve in almost every alcohol paper and it always bends the same way: down, then up. Drink a little, the line says, and you live longer than the people who drink nothing. Drink a lot and the line climbs back past where it started. That J has a name in this literature and it has survived about thirty years of criticism.

In 2024 a group published something unusual for the field: instead of picking a study design and reporting it, they took the whole evidence base for alcohol and ischaemic heart disease, ranked it by how well each design can answer that particular question, and asked what survives the ranking.

What survived the ranking

The paper, in Nature Communications, is framed as a burden-of-proof study. Cohort and case-control data show low to moderate consumption associated with decreased ischaemic heart disease risk — specifically, intake inversely related to ischaemic heart disease and myocardial infarction morbidity in both sexes, and to ischaemic heart disease mortality in males.

Pooled Mendelian randomization data show no association.

Two designs, same question, opposite signs, and the framework’s verdict is that the randomized-by-genetics data are the stronger instrument for a causal claim. The authors’ own summary of why is the sentence to carry out of this article: self-reported versus genetically predicted alcohol use data yield conflicting findings about the alcohol–ischaemic heart disease relationship.

The J-curve in one cohort, and the straight line beside it

The clearest single demonstration comes from a 2022 cohort in the American Journal of Clinical Nutrition: 40,386 Chinese men, of whom 17,676 were also genotyped for the rs671 variant in the aldehyde dehydrogenase 2 gene. Over 303,353 person-years they recorded 2,406 incident cardiovascular diseases and 3,195 all-cause deaths.

Self-reported intake: J-shaped, with decreased risk for light drinkers (25 grams a day or less) and for moderate drinkers (more than 25 up to 60 grams a day), against current non-drinkers.

Genetically predicted intake: a straight line upwards. Incident cardiovascular disease hazard ratios across increasing tertiles of genetically predicted consumption were 1 (reference), 1.18 (95 % CI 1.01 to 1.38) and 1.22 (95 % CI 1.03 to 1.46), with p-trend 0.02 overall, 0.03 for coronary artery disease and 0.02 for stroke.

Then the sentence that makes the paper: after excluding heavy drinkers, the risk of incident cardiovascular disease and all-cause mortality was increased by 27 per cent and 20 per cent per standard drink of genetically predicted alcohol consumption.

The authors draw the conclusion the design permits — that the evidence extends to a harmful effect on total cardiovascular disease including coronary artery disease and on all-cause mortality, and that even among light-to-moderate male drinkers lowering consumption has health value.

Why the kink sits at zero

The 2021 BMC Medicine analysis is the best-documented J-curve in the field and it makes the mechanism visible. It combined de novo analyses of 14,386 patients with previous myocardial infarction, angina or stroke in UK Biobank (median follow-up 8.7 years, 1,640 deaths, 2,950 subsequent events) and 2,802 patients from the Health Survey for England and the Scottish Health Survey (median follow-up 9.5 years, 1,257 deaths), with 12 published studies covering 31,235 patients, 5,095 deaths and 1,414 subsequent events: 48,423 patients in total.

Alcohol was associated with all assessed outcomes in a J-shaped manner relative to current non-drinkers. Risk reduction peaked at 7 grams a day for all-cause mortality (relative risk 0.79, 95 % CI 0.73 to 0.85), 8 grams a day for cardiovascular mortality (0.73, 95 % CI 0.64 to 0.83) and 6 grams a day for cardiovascular events (0.50, 95 % CI 0.26 to 0.96).

The final line is the one that explains the shape: in the few studies that excluded former drinkers from the non-drinking reference group, the reductions among light-to-moderate drinkers were attenuated.

That is the entire mechanism in one clause. The reference group in most alcohol cohorts is “people who currently do not drink”, and that group contains people who used to drink heavily and stopped — people whose cardiovascular health is already changed. Move them out and the benefit shrinks.

The cancer literature has no J

For colorectal cancer the two designs do not merely disagree, one of them is unambiguous. A 2022 analysis in International Journal of Cancer pooled 32 cohort studies and ran two-sample Mendelian randomization alongside an epigenetic analysis. Genetic predisposition to alcohol drinking was causally associated with increased colorectal cancer risk, odds ratio 1.79 (95 % CI 1.23 to 2.61). Two alcohol-related methylation sites mapped to the COLCA1/COLCA2 gene were identified as causally associated with increased risk, and in UK Biobank the carriage of risk alleles at interacting polymorphisms increased risk further with alcohol consumption (p-interaction 0.027 and 0.016).

No threshold appears in that design, because a genetic instrument does not have an arm that people volunteer for.

The four questions this literature turns on

Is the reference group current non-drinkers or never drinkers — because that single choice moves the J. Is the outcome mortality, cardiovascular events or biomarkers, because the shapes differ between them. Was the alcohol measured by self-report, by record, or by instrument — and note that the burden-of-proof review’s whole point is that self-report and genetics disagree about the same people. And how long was follow-up, because a J-shaped curve needs decades to resolve and a five-year trial cannot see one.

The genetic studies have a known limit of their own: they estimate the effect of lifetime predisposition, which is not the same as the effect of a drink someone chooses to have tonight.

What this evidence does not show

It does not show that the J-curve is fabricated; the cohort signal is large, replicated across four countries in the analysis above, and remains the reason guidelines have never told anyone to start drinking. It does not show that a person who drinks moderately should stop — these papers are about population risk, not individual decisions, and the burden-of-proof authors call for better methods and randomised emulation rather than for a verdict. Nothing here addresses the difference between alcohol types, drinking patterns, or socioeconomic confounding beyond what the excluded-former-drinker analysis touches. And the Mendelian randomization estimates are about genotype, not about anyone’s actual intake.

This article summarises published cohort analyses and Mendelian randomization studies. It is not medical advice and it is not a recommendation for any individual.