And Then You're Dead

The Autopsy / Case File · Chronic Disease Progression

The Attributable Fraction

Almost nobody in America is recorded as having died of obesity. The condition has an ICD-10 code, E66, and it is almost never the underlying cause on a death certificate, because the certifier is asked to name what started the chain of events — and obesity does not kill anyone directly. It kills through heart disease, through diabetes, through kidney failure, through a dozen cancers. So the number of obesity deaths in the United States is not counted. It is calculated. And the calculations disagree with each other enormously.

Four peer-reviewed attempts to estimate obesity-related deaths in the United States around the year 2000 produced numbers of roughly 112,000, 186,000, 280,000, and 365,000 — the highest and lowest separated by more than a factor of three, with one of them requiring a published correction that lowered it by 35,000 after it had already circulated as a headline. None is fraudulent. All are competent. The distance between them is a matter of which relative risks were used, which ages were included, whether smokers were separated out, and whether the model looked at a person's weight today or the highest weight they ever reached.

And Then You're Dead · September 2026 · 10 min read

Join the discussion →
Two federal health researchers sit at a conference table reviewing a bar chart of national dietary survey data displayed on a wall screen and a laptop, with measuring spoons and portion-size models on the table in front of them
The director of CDC's National Health and Nutrition Examination Survey and the head of USDA's dietary survey group review national intake data in 2009, with portion models on the table. NHANES — measured heights and weights on a national sample, linked to death records — is the input every competing estimate of obesity-attributable mortality is built from. The disagreement is not about the survey. It is about what is done with it. Photo: USDA, public domain, via Wikimedia Commons.

Start with the thing that makes this hard. When a person with a body mass index of 41 dies of a heart attack at 58, the certificate records a myocardial infarction, probably with coronary atherosclerosis on the line below it, and possibly type 2 diabetes and hypertension in Part II. The National Center for Health Statistics defines the underlying cause as the disease or injury that initiated the train of morbid events leading directly to death, and it is that single condition — the one at the bottom of Part I — that becomes the cause in every national mortality statistic. Obesity, in that framework, is upstream of the chain rather than in it. It is a reason the chain existed at all.

Why the Certificate Can't Answer It

This is not an accident of paperwork; it is what the instrument is designed to do. Part I asks the certifier to write the sequence: the immediate cause on line (a), then each antecedent condition below it, with the underlying cause lowest. Part II is reserved for other significant conditions contributing to death but not resulting in the underlying cause. A physician who believes obesity is the reason a patient is dead has one place to put that belief, Part II, and no obligation to use it. Most do not. Obesity has consequently never appeared on any list of the country's leading causes of death — not because it doesn't kill people, but because the counting system was never built to see risk factors, only diseases.

So the field does something else. It uses a population attributable fraction: take the relative risk of death associated with a given body mass index category, estimated from a cohort study; take the share of the population in that category, estimated from a survey with measured heights and weights; combine them to get the proportion of deaths that would not have occurred if everyone in the high-risk category had the mortality risk of the reference category; multiply by total deaths. The arithmetic is simple. Every input is contestable.

0
excess deaths associated with obesity in the U.S. in 2000, in the CDC/NCHS analysis published in JAMA in 2005 (95% CI: 53,754 to 170,064)
0
deaths attributed to poor diet and physical inactivity in the U.S. in 2000 by a different CDC team the year before — after a published correction cut the figure from 400,000

Four Numbers for the Same Year

In 1999, Allison and colleagues published in JAMA the first rigorous national estimate, drawing hazard ratios from five prospective cohorts — Alameda, Framingham, Tecumseh, the American Cancer Society's Cancer Prevention Study I, and the NHANES I follow-up study — plus published results from the Nurses' Health Study, and applying them to 1991 body mass index distributions. Their mean estimate was 280,184 obesity-attributable deaths annually, range 236,111 to 341,153. Restricting the hazard ratios to nonsmokers and never-smokers and applying those to the whole population raised it to 324,940. That is the origin of the "about 300,000 deaths a year" figure that became standard in public health argument for a decade.

In 2004, Mokdad and colleagues at CDC published "Actual Causes of Death in the United States, 2000," which reframed mortality around modifiable behaviors rather than diseases. Tobacco led at 435,000 deaths, 18.1 percent of all US deaths. Poor diet and physical inactivity came second, originally published as 400,000 and projected to overtake tobacco; JAMA issued a correction in January 2005 revising it to 365,000, or 15.2 percent. Note what the category is: diet and inactivity, the behaviors, not obesity, the physiological state. The two get quoted interchangeably and are not the same quantity — one is a behavioral exposure, the other an outcome of it.

Then in April 2005, Flegal, Graubard, Williamson and Gail — working at CDC's own National Center for Health Statistics — published a directly competing estimate in the same journal. They pooled relative risks from NHANES I, II and III with mortality follow-up, applied them to the BMI distribution measured in NHANES 1999–2002, and adjusted for age, sex, race, smoking and alcohol. Their answer for obesity, meaning a BMI of 30 or above: 111,909 excess deaths, with a confidence interval running from 53,754 to 170,064 — roughly a third of Allison's number, produced inside the same agency that had just published the 400,000 figure. Two further findings were more disruptive still: overweight, a BMI between 25 and 30, was associated with 86,094 fewer deaths than normal weight rather than more, and underweight with 33,746 excess deaths of its own.

The disagreement is not about whether obesity kills people. It is about which epidemiological convention you accept, and each convention moves the answer by six figures.

The Weight-Loss Problem

The most consequential methodological objection is not about statistics at all. It is about who ends up in the "normal weight" box. Almost every one of these estimates classifies people by their weight at a single moment — the day they were surveyed. That group therefore contains two very different populations mixed together: people who have always been at a normal weight, and people who were formerly obese and have since lost weight. Some of that weight loss is deliberate and healthy. A great deal of it is the weight loss that accompanies cancer, heart failure, COPD, and the last years of life. Stokes and Preston, writing in PNAS in 2016, used NHANES data from 1988 through 2010 linked to death records through 2011 to compare models built on weight at survey against models built on a person's maximum attained weight. The maximum-weight models performed substantially better, and the reason was exactly this: mortality among people who had lost weight was exceptionally high, and folding them into the normal-weight reference category makes normal weight look more dangerous and obesity look less dangerous than either is. Their conclusion was that studies based on body mass index at survey — which is most of them — have consistently underestimated the effect of excess weight on US mortality.

Two years later, Preston, Vierboom and Stokes applied the maximum-BMI approach to the question of why American mortality had stopped improving. Introducing maximum lifetime BMI into a Cox model controlling for age and sex raised the annual rate of mortality decline by 0.54 percent, with a confidence interval of 0.45 to 0.64 percent. In plain terms: without the rise in body weight between 1988 and 2011, American mortality would have been falling meaningfully faster than it did. They estimated that the increase in maximum BMI over that period reduced life expectancy at age 40 by 0.9 years as of 2011, and accounted for 186,000 excess deaths in that single year.

What Age Does to the Answer

The other major fault line is age. Most deaths occur in old age, so an attributable-fraction estimate is extremely sensitive to whatever relative risk it assumes for older adults — and the cohorts that supply relative risks tend to underrepresent them, or exclude the sickest of them to control for reverse causation. Masters and colleagues, in the American Journal of Public Health in 2013, used nineteen waves of the National Health Interview Survey linked to individual National Death Index records from 1986 through 2006, modeling both the age variation in obesity's mortality effect and the cohort variation in how long people had been obese. They found a substantially stronger association at older ages than earlier work, and an attributable share rising across successive birth cohorts: 15.6 percent of adult deaths for white men and 5.0 percent for Black men, 21.7 percent for white women and 26.8 percent for Black women. Those are proportions, not counts, and they sit far above what a 112,000-death estimate implies against roughly 2.4 million annual American deaths.

A parallel approach skips attributable fractions entirely and reads the certificates in bulk, using every cause listed rather than only the underlying one. Adair and Lopez, in BMC Medicine in 2020, defined obesity-related cardiovascular mortality as any cardiovascular disease appearing anywhere on the certificate alongside diabetes, chronic kidney disease, obesity, lipidemias or hypertensive heart disease. On that definition, half of all US cardiovascular multiple-cause mortality qualifies, against 40 percent in Australia — and the rate was rising about 3 percent a year, fastest at the youngest ages.

What Follows From Picking One

None of this is academic housekeeping. The choice of estimate determines whether obesity is described as the second-leading preventable cause of death in America or as a problem an order of magnitude smaller than tobacco. It determines how a cost-effectiveness model values a bariatric procedure or a GLP-1 prescription, which prevention programs get funded, and how large a benefit a regulator may claim. And it sets the background against which current prevalence is read: NCHS put adult obesity at 40.3 percent during August 2021 through August 2023 and severe obesity at 9.4 percent, the latter up from 7.7 percent a decade earlier even as overall obesity held roughly flat. Whatever the per-death multiplier turns out to be, it is being applied to two in five American adults.

And Then You're Dead

The certificate will say heart disease. It will be accurate. Somewhere behind it is a number nobody can produce without first choosing a method, and the honest range of that number spans a quarter of a million deaths a year. This is what an uncounted cause of death looks like from the inside: not a cover-up, not a mystery, just a question the recording system was never designed to be asked, answered instead by models that agree on the direction and not the size.

That's it. That's the whole thing.

Comments

Loading…

Comments are anonymous and moderated.

Sources