The Most Comforting Statistic in Medicine
For decades, the science seemed settled. Study after study found that people who drank a glass or two of wine daily lived longer than people who drank nothing at all. Plotted on a graph, the relationship between alcohol consumption and death formed a distinctive J-shape: teetotalers on the left side dying sooner, moderate drinkers sitting in the comfortable dip at the bottom, and heavy drinkers climbing the right arm toward early graves, a pattern so consistent across cohorts in North America, Europe, and Asia that it became one of the most replicated findings in modern epidemiology. Your evening Pinot wasn't just harmless, according to the data; it was medicinal.
Physicians cited the J-curve when patients asked about wine, the American Heart Association acknowledged a possible cardiovascular benefit, and dietary guidelines worldwide encoded the idea that moderate consumption could be protective. It was wrong.
The Ghost in the Reference Group
The problem was hiding in the control group. When epidemiologists compare drinkers to "abstainers," who exactly are the abstainers? In most cohort studies, the nondrinker category lumped together people who had never touched alcohol with former drinkers who had quit, often because they were already sick: liver damage, cancer diagnoses, cardiovascular events, medication interactions, or a doctor's firm instruction to stop. The sickest people in any population are disproportionately likely to have stopped drinking, and when researchers dumped them into the "abstainer" bin, they made nondrinkers look far unhealthier than they actually were.
Alcohol epidemiologists have a name for this: the "sick quitter" bias, a distortion debated since the 1980s but never comprehensively measured until a team led by Jinhui Zhao and Tim Stockwell at the University of Victoria assembled every cohort study they could find on alcohol and all-cause mortality published between 1980 and 2021, totaling 107 studies, 724 risk estimates, 4,838,825 participants, and 425,564 recorded deaths.
Two Analyses, Two Realities
The unadjusted pooled analysis produced exactly what three decades of research predicted: a textbook J-curve in which low-volume drinkers (roughly one to two standard drinks per day) had a relative risk of 0.86 compared to abstainers, a 14% mortality advantage, while occasional drinkers consuming less than a single drink per day looked similarly protected at RR 0.84, and former drinkers showed elevated mortality at RR 1.22. The familiar pattern held perfectly.
Then the team adjusted. They corrected for former-drinker bias by flagging studies that had mixed ex-drinkers into their abstainer reference groups, controlled for median cohort age (older cohorts accumulate more sick quitters), and adjusted for sex, country, publication year, follow-up duration, drinking pattern, and whether each study had accounted for preexisting heart problems, socioeconomic status, race, diet, exercise, BMI, and smoking.
The J-curve collapsed. Low-volume drinkers went from a 14% mortality advantage to a non-significant 7% advantage (RR 0.93, 95% CI 0.85β1.01, P = .07), occasional drinkers landed at RR 0.96 (95% CI 0.86β1.06, P = .41), and the protective dip that had shaped public health messaging for a generation was no longer statistically distinguishable from zero. The benefit of a daily drink had been an artifact of comparing healthy drinkers to a reference group contaminated with sick people who had recently stopped.
The Genetic Confirmation
Observational epidemiology, no matter how carefully adjusted, can never fully eliminate confounding. But Mendelian randomization can sidestep the problem entirely by using genetic variants as a natural experiment: if the random genetic differences that predispose people to drink more or less alcohol produce a J-shaped mortality pattern, confounding becomes an unlikely explanation for the curve.
In 2024, Kassaw and colleagues at the University of South Australia tested this in 278,093 UK Biobank participants using 94 genetic variants as instruments, and over a median 12.6-year follow-up during which 20,834 participants died, the analysis found a strictly linear dose-response with no trace of the J-curve: every standard-unit increase in genetically predicted alcohol intake raised all-cause mortality risk by 27% (OR 1.27, 95% CI 1.16β1.39), the test for nonlinearity returned P β₯ 0.21 for every mortality outcome, and cancer, cardiovascular, and digestive-disease mortality all rose linearly. The genetic evidence didn't just fail to support the J-curve. It found a straight line running the wrong direction.
The Strongest Counterargument
The International Scientific Research Forum on Alcohol Research published a detailed rebuttal. Their central objection is that Stockwell's team has, across three sequential papers (2007, 2016, 2023), selectively excluded studies to engineer a preferred conclusion, and they note that the supplementary eFigure 4 in the paper itself appears to show a J-shaped curve even after adjustments. Dozens of individual cohort studies published since 2006 continue to find protective associations for moderate drinkers even after excluding former drinkers from the reference group. And the biological plausibility of a protective mechanism is real: randomized human trials demonstrate that moderate alcohol intake raises HDL cholesterol, improves endothelial function, and reduces inflammatory markers. The Global Burden of Disease study (Bryazka et al., 2022) found that the theoretical minimum risk exposure level was not zero but approximately 5 grams per day. These are substantive objections from researchers with decades of expertise, and ignoring them would mean dismissing both observational and mechanistic evidence that runs into the thousands of papers.
What We Didn't Prove
No study has randomly assigned people to drink or abstain over decades and measured mortality, and the former-drinker bias, though it explains a large portion of the J-curve's protective signal, may not explain all of it. Cardiovascular mortality was not analyzed separately from all-cause mortality in this paper, and several heart-disease-specific meta-analyses still find a modest protective effect. The Mendelian randomization evidence is limited to a predominantly white British cohort, and 94 genetic variants explain only a fraction of variance in alcohol consumption, limiting power for detecting nonlinearity. Cultural drinking patterns, beverage type, and meal context went unaccounted for in both studies.
The Bottom Line
The belief that moderate drinking extends life rested on a measurement error: decades of studies made abstainers look unhealthy by mixing in former drinkers who had quit because they were already dying. Fix the comparison group and the benefit vanishes. The genetic evidence agrees. This doesn't mean a glass of wine is particularly dangerous for a healthy person, but "it's good for you" was never real.
What You Can Do
If you enjoy wine with dinner and have no contraindicating health conditions, the absolute risk at one drink per day remains small, but stop framing it as a health practice. If a physician has recommended moderate drinking for cardiovascular benefit, share the Zhao 2023 meta-analysis and the Kassaw 2024 Mendelian randomization study and ask whether the recommendation still holds. For anyone who doesn't currently drink, these findings eliminate the last evidence-based reason to start. Researchers designing future cohort studies should use lifetime abstainers as the reference group and record reasons for abstention at enrollment.