The cardioprotection that wasn't: how Mendelian randomization dismantled the moderate-drinking hypothesis

The J-curve story was one of epidemiology's most durable findings for forty years. Study after study, through the 1970s, 1980s, and 1990s, arrived at the same shape: abstainers, people drinking one to two drinks a day, and heavy drinkers all arranged along a J or U, with the middle group apparently doing better than either extreme. Cardiovascular disease—the leading cause of death across high-income countries—was the most prominent outcome linked to the curve. Light drinking, the story went, reduced coronary heart disease risk, perhaps by raising HDL cholesterol, perhaps by lowering platelet aggregation, perhaps through a handful of other plausible mechanisms that could be sketched on a whiteboard.

The story was intuitive, pleasurable for drinkers to hear, and backed by a formidable body of observational data. It also turned out to be built on a methodological problem that ran directly through the center of how the studies were designed.

The abstainer category was contaminated

The most consequential methodological flaw in J-curve research is now called "sick quitter bias" or "abstainer bias." In most early studies, the comparison group labeled "abstainers" or "non-drinkers" was not a clean population of lifetime teetotalers. It was a mixture: people who had never drunk, people who had drunk lightly in youth, and—critically—people who had stopped drinking because they were already sick.

Former heavy drinkers who quit after a cardiac event, a cancer diagnosis, or a worsening chronic illness often spent the rest of their lives counted as non-drinkers. Because their health had already deteriorated before quitting, they elevated the apparent mortality of the abstainer group. By comparison, stable light drinkers in midlife often had better jobs, better diets, more physical activity, and fewer comorbidities driving the comparison. That cluster of advantages is sometimes called the "healthy drinker effect."

A 2017 analysis in Addiction by Naimi and colleagues formally quantified the selection problem. Examining 87 published studies of alcohol and mortality, the authors found that most did not exclude former drinkers from the abstainer comparison group, and many did not distinguish lifetime teetotalers from people who had recently quit for health reasons.[3] The apparent protective association with light drinking was systematically stronger in studies with the most contaminated comparison groups and weaker—or absent—in studies that took greatest care to isolate lifetime abstainers. The J-curve was not a stable empirical finding; it was a statistical echo of study design.

What Mendelian randomization showed

Observational epidemiology cannot randomize people into lifetime drinking habits. No trial can. But genetics offers a partial solution.

Some people carry variants in ALDH2 (acetaldehyde dehydrogenase 2) or ADH1B (alcohol dehydrogenase) that make alcohol metabolism slower and more aversive. Individuals with certain ALDH2 alleles flush, feel nauseated, and experience elevated acetaldehyde levels when they drink. As a result, they tend to drink less—not by choice, not because of prior illness, but because of genetic endowment set before any disease developed. These variants are distributed essentially randomly in the population and are not themselves associated with the cardiovascular risk factors that confound observational studies. They can therefore function as a natural instrument: a proxy for "tends to drink less" that sidesteps the selection problems plaguing conventional cohort analyses. This approach is called Mendelian randomization.

In 2014, Holmes and colleagues published the first large-scale Mendelian randomization analysis of alcohol and cardiovascular disease, drawing on individual participant data from 56 epidemiological studies covering 261,991 individuals.[2] The observational analysis within the same dataset reproduced the familiar J-curve: light drinkers appeared to have lower coronary heart disease risk than abstainers, just as the literature had always shown. The Mendelian randomization analysis told a different story. Genetically predicted lower alcohol consumption was associated with lower systolic blood pressure and lower coronary heart disease risk—not higher. When confounders were removed through genetic instruments, the apparent benefit of light drinking did not survive. The direction reversed.

In 2019, Millwood and colleagues extended the Mendelian randomization approach using the China Kadoorie Biobank, a prospective cohort of 512,715 adults recruited across ten Chinese regions.[4] The setting provided unusual analytical leverage: drinking rates among Chinese women were very low for cultural reasons—not health reasons—while rates among Chinese men were substantially higher. Variants in ALDH2 predicted drinking behavior differently by sex in this cohort, creating a natural experiment within the population. The observational data again showed what looked like a protective pattern for moderate drinking. The Mendelian randomization analysis again reversed it: genetically predicted lower alcohol intake was associated with lower risk of stroke and ischemic heart disease. The J-curve's protective signal disappeared when confounding was removed through genetic instruments.

Across both studies, the genetic evidence placed cardiovascular risk in a roughly linear or monotonic relationship with alcohol exposure—the more alcohol, somewhat more risk—rather than a J-shape. That is a fundamentally different dose-response model than the one that dominated the cardioprotection literature for four decades.

The 2018 Lancet assessment: totaling the burden

In the same year as the Millwood study's preliminary publication, the GBD 2016 Alcohol Collaborators published a major burden analysis in The Lancet covering 195 countries and territories across 1990 to 2016.[1] The paper's headline conclusion—"the safest level of drinking is none"—was sometimes treated as an overstatement by critics who focused narrowly on coronary heart disease. But the paper's full accounting included cancer, liver disease, injury, and total mortality rather than heart disease alone.

For a 40-year-old woman drinking one standard drink per day, the authors estimated that 918 out of 100,000 drinkers would develop an alcohol-attributable condition over one year, compared with 914 at zero intake.[1] The excess risk at that dose was small—four additional attributable cases per 100,000—but it was net positive across all health outcomes combined. Alcohol's cancer and injury burden at any dose offset the narrow band of possible cardiovascular protection. The authors wrote explicitly that accounting for all health outcomes eliminated the net benefit that cardiovascular-focused analyses had appeared to show in certain age and sex subgroups.

The industry research problem

In parallel with the scientific reassessment, an institutional problem surfaced.

In 2018, reporting by the New York Times and other outlets revealed that five major alcohol companies—including Anheuser-Busch InBev, Heineken, and Diageo—had quietly contributed the majority of funding for a $100 million NIH-registered randomized trial called MACH15, which was designed to test whether moderate drinking reduced cardiovascular events. Internal documents obtained by investigators showed that the trial's principal researchers had been extensively involved in pitching the study to industry funders with language suggesting the design was likely to show a benefit. The NIH Office of the Inspector General and the NIH Director's Advisory Committee reviewed the matter and found that the principal investigators had failed to maintain required independence from their industry funders. The NIH cancelled the study in June 2018. No randomized trial designed under clean conditions to test a cardioprotective effect of moderate alcohol has been launched since.

The episode matters not only as an ethics case. It illustrates the institutional pressure that shaped the research environment from which the J-curve literature emerged. Meta-research analyses of funding bias in alcohol research have consistently found that industry-funded studies are significantly more likely to reach conclusions favorable to moderate drinking than independently funded studies examining the same outcomes.

Where the evidence boundary sits today

The honest summary is narrower than either "alcohol is definitely protective" or "any drink is immediately dangerous." But the evidence has decisively moved away from the cardioprotective framing.

The observational J-curve was statistically real but appears to have been largely an artifact: contaminated comparison groups, the healthy-drinker effect, and lifestyle confounders drove most of the association.[3] Mendelian randomization evidence in two large, independent settings with different population genetics reversed the apparent protective direction.[2][4] The full health burden of alcohol at any dose, accounting for cancer, liver disease, and injury, appears net positive rather than net negative at essentially any intake level.[1] No major health authority—WHO, the CDC, Health Canada, the UK Chief Medical Officers—currently recommends drinking for cardiovascular benefit. Canada's 2023 updated guidance adopted a "no safe amount" framework for cancer risk at the population level.

What the evidence cannot yet resolve with certainty is the magnitude of risk at very low doses for individuals with otherwise excellent cardiovascular profiles. A person drinking one glass of wine three or four times per week alongside a diet high in vegetables, regular exercise, and no smoking is probably not at acute risk. But the epidemiological case that this pattern actively reduces cardiovascular risk—relative to not drinking—has not survived the methodological scrutiny of the past decade.

The cardioprotection that generations of light drinkers were told they were receiving was a number that moved when the measuring instruments improved. What remains once the sick quitters are removed, the genetic instruments are applied, and the cancer and injury burden is added back is a flat-to-slightly-positive risk slope, not a J.

cronfeed.work