H Heuristics Digital Reports

H Heuristics · Digital Report № 2026-11 · September 2026

The Bars Do Not Add Up

What the global risk-factor ranking measures, what it cannot tell you, and where to intervene instead

Sixty-seven million attributable deaths, in a world where fifty-six million people died. The arithmetic is not broken; the question being asked of it is.

AuthorHunter Hughes
InstitutionH Heuristics
Published13 September 2026
Report №2026-11
Reading time15 min

Abstract

The Global Burden of Disease ranking of deaths by risk factor is an accurate and valuable chart, and it is routinely read as something it cannot be: a priority list in which the longest bar names the first thing to fix. Three features of the data defeat that reading. The bars are attributable burdens, each computed against its own counterfactual with everything else held constant, so the same death is counted in several of them; adding the thirty-four charted values gives about 66.9 million deaths in a year when roughly 56 million people died worldwide. The chart also double-counts internally, listing air pollution as an aggregate of 4.9 million alongside its two components of 3.41 and 1.64 million. And ten of the thirty-four bars are components of a single behaviour — what people eat — split into separate nutritional channels totalling 15.3 million attributable deaths that are not 15.3 million distinct people.

The ranking also mixes categories that sit at different points on the same causal chains. The top twelve entries contain physiological states, behaviours, environmental exposures and dietary components. High blood pressure, the longest bar, is not a policy competing with smoking; it is a measurement of a state produced by salt, body weight, alcohol and inactivity, four of which appear separately and lower down the same chart. A further consequence is that the ordering is partly an artefact of where category boundaries were drawn: splitting diet into ten components shortens each bar and pushes diet down, while aggregating air pollution into one row lengthens its bar and pushes it up.

Redrawn as a causal map rather than a league table, the chart yields a clearer instruction. The three longest bars all sit in the middle of the chain, where no instrument reaches at population scale except by acting on their own causes or by treating individuals who already have them. The operable drivers sit upstream: tobacco, air pollution, diet composition, alcohol and physical inactivity. Tobacco and air pollution are distinctive in being simultaneously large, upstream and directly operable, reaching the endpoints both through the physiological states and around them, which is why they dominate the effective-intervention literature.

The intervention answer is unusually concrete and already written down. The World Health Organization's twenty-nine 'best buys' for noncommunicable disease return roughly fourfold in social and economic benefit for every dollar invested over five years, and sevenfold if sustained to 2035; full implementation would save 12 million lives and prevent 28 million heart attacks and strokes by 2030, with economic gains above a trillion dollars, at a cost of around 0.6 per cent of low- and middle-income countries' gross national income. The pattern within that list contradicts the framework most discussions of this chart produce: the interventions carrying the measured population-level effect are overwhelmingly fiscal and regulatory — tobacco and alcohol taxation, sugary-drink taxes, mandatory sodium reformulation, advertising restrictions, emissions standards — rather than educational. Sodium reformulation is the clearest instance, because it lowers consumption without requiring any individual to decide anything. A final caution against reading budgets off the ranking: the same air-pollution burden is estimated at 4.9 million deaths in the 2017 data charted here and at 8.1 million in the Health Effects Institute's 2024 assessment using 2021 data and revised exposure–response functions, a difference larger than most bars on the chart.


Executive Summary

One of the most widely circulated charts in global health ranks the world's causes of death by risk factor. It is an excellent chart. It is also routinely read as a priority list, which is the one thing it cannot be.

FINDING 01

The bars overlap, so they cannot be summed

Add the thirty-four bars and the total comes to about 66.9 million deaths, in a year when roughly 56 million people died. Each bar is computed against its own counterfactual, holding everything else constant, so the same death is counted in several of them. That is not an error in the chart; it is what an attributable burden is.

FINDING 02

The bars are not the same kind of thing

The list mixes physiological states, behaviours, environmental exposures and dietary components — measurements taken at different points on the same causal chains. "High blood pressure" is not a policy that competes with "smoking"; it is largely an outcome of several other entries on the list.

FINDING 03

The longest bars are the least operable

The top three are physiological states. Nothing acts on blood pressure directly except the things further down the ranking: salt, alcohol, weight, inactivity. Read as a map rather than a league table, the chart points to a short list of upstream levers — and those levers are costed, evidenced, and largely unbought.

The chart in question is Our World in Data's rendering of the Institute for Health Metrics and Evaluation's Global Burden of Disease estimates for 2017. It is accurate, carefully sourced, and one of the most useful summaries in the field. The problem is not the chart but the inference people draw from it: that the ranking is a to-do list, and that the longest bar names the first thing to fix.

66.9m
Deaths obtained by adding all thirty-four bars together
Author's sum of the charted values
~56m
People who actually died worldwide that year, from all causes combined
IHME, Global Burden of Disease 2017
10 of 34
Bars that are components of a single thing — the diet — split into separate rows
The chart as published
$4 → $7
Returned per dollar invested in WHO's "best buys" over five years, rising to sevenfold if sustained to 2035
World Health Organization

This report does two things. It explains what the chart actually measures and why the ordinary reading of it fails, which takes the first half. Then it converts the ranking into a causal map and reads the intervention question off that instead — where the answer turns out to be unusually concrete, well costed, and mostly fiscal rather than educational.

Sixty-seven million attributable deaths, in a world where fifty-six million people died. The arithmetic is not broken; the question being asked of it is. On summing attributable burdens

1. What the Chart Measures

Comparative risk assessment answers a precise question, and it answers it separately for every risk on the list.

The Global Burden of Disease study estimates, for each risk factor, the deaths that would not have occurred in a counterfactual world where exposure to that factor sat at a theoretical minimum-risk level — and where everything else was held constant. That last clause is the whole of the matter. Each bar is the answer to its own question, computed independently of the others.

Consider a sixty-year-old who dies of a heart attack, and who was overweight, hypertensive, diabetic, sedentary, a smoker, ate a high-salt diet and lived in a polluted city. Ask "would this death have occurred if he had not smoked?" and the answer may be no. Ask "would it have occurred if his blood pressure had been normal?" and the answer may also be no. Both answers are correct. The death appears in both bars, and in several others.

This is not a flaw in the method. It is the only sensible way to answer the question that each bar poses, and the alternative — partitioning each death among its causes — would require causal weights nobody can estimate. The GBD approach is deliberate, standard and right. It simply produces quantities that behave differently from the ones most readers assume.

What the chart supports, and what it does not

It supports comparing the attributable burden of one risk against another; tracking a single risk over time; and establishing that a given exposure is a large problem worth acting on. It does not support adding the bars; computing what share of deaths a risk represents by dividing into the total; allocating a budget in proportion to bar length; or concluding that eliminating the top risk would avert its full quoted number of deaths once the others are also addressed.


2. The Bars Do Not Add Up

The overlap is not a subtlety at the margin. It is larger than any single bar on the chart.

Figure 1 — The sum of the parts against the whole

Adding the thirty-four charted values gives 66.9 million. Global deaths from all causes in 2017 were approximately 56 million on the same study's estimates (IHME; see also the GBD 2017 mortality analysis in The Lancet). The excess is the overlap between risks, and it exceeds the largest single bar.

Two features of the chart make the point visible without any external data at all.

The first is that it double-counts within itself. "Air pollution (outdoor & indoor)" appears at 4.9 million, and "Outdoor air pollution" (3.41 million) and "Indoor air pollution" (1.64 million) appear as separate bars further down. An aggregate and its two components are plotted on the same axis. A reader summing the column adds air pollution twice.

The second is the treatment of diet. Ten of the thirty-four bars are dietary components — sodium, whole grains, fruits, nuts and seeds, vegetables, seafood omega-3, fibre, legumes, calcium, red meat — together totalling 15.3 million attributable deaths. These are not 15.3 million distinct people. They are largely the same cardiometabolic deaths, attributed ten times through ten nutritional channels of a single underlying thing: what people eat.

Notice what that does to the ranking. Splitting diet into ten components makes each component's bar short and pushes diet down the list; aggregating air pollution into one row makes its bar long and pushes it up. The ordering is partly an artefact of where the analyst drew the category boundaries — a decision made for epidemiological reasons that has nothing to do with which lever a health minister should pull.


3. Four Different Kinds of Thing

The ranking places on one axis quantities that sit at entirely different points in the causal chain, and that are acted on by entirely different instruments.

Figure 2 — The top twelve, coloured by what kind of thing each is

The same published values, recoloured. The aggregate "air pollution" row is omitted here because its two components are shown separately. Values as charted, from Our World in Data on IHME GBD 2017.

Table 1 — What each kind is, and who can act on it

Kind Examples on the chart Position in the chain Instrument that acts on it
Physiological state High blood pressure, high blood sugar, obesity Downstream — produced by other entries on the list Clinical treatment, and upstream action on its own causes
Behaviour Smoking, alcohol use, low physical activity Midstream — individually enacted, structurally shaped Price, availability, advertising rules, built environment
Environmental exposure Outdoor and indoor air pollution, unsafe water Upstream — not chosen by the person exposed Emissions regulation, fuel switching, infrastructure
Dietary composition Sodium, whole grains, fruits, nuts, vegetables Midstream — ten measured facets of one behaviour Reformulation mandates, price, agricultural policy
Deprivation Child wasting, low birth weight, vitamin-A deficiency Upstream — a consequence of poverty, not of choice Income support, food systems, maternal health services

Categories are the author's, applied to the chart's own entries. The point is not the exact taxonomy but that a single ranking mixes levels of a causal chain.

Once the categories are visible, the ordinary reading of the chart becomes hard to sustain. "High blood pressure" is the longest bar, and it is not a thing anybody can do anything about directly at population scale. It is a measurement — a physiological state produced by salt intake, body weight, alcohol, inactivity and age, four of which appear separately on the same chart, lower down, with shorter bars.


4. From a Ranking to a Causal Map

Redraw the same entries as a chain rather than a league table, and the intervention question answers itself.

Figure 3 — The same risks, arranged by causal position

Risk factors arranged by causal position Three columns. On the left, operable upstream drivers: diet composition, alcohol, physical inactivity, tobacco and air pollution. In the middle, the physiological states that carry the chart's three longest bars: high blood pressure, high blood sugar and obesity. On the right, the deaths. Arrows run from the left column into the middle and from the middle to deaths, while tobacco and air pollution also act directly on deaths, bypassing the middle column. OPERABLE DRIVERS MEASURED STATES OUTCOME Diet composition salt · grains · fruit · nuts Alcohol use 2.84m attributable Physical inactivity 1.26m attributable Tobacco 7.10m attributable Air pollution 4.90m attributable High blood pressure 10.44m — longest bar High blood sugar 6.53m Obesity 4.72m DEATHS cardiovascular cancer respiratory diabetes tobacco and air pollution also act directly, bypassing the middle column no instrument acts on these directly

Schematic, using the chart's own entries and values. Arrows show the dominant direction of causation, not its full complexity; age, genetics and other determinants are omitted.

Drawn this way, the ranking's strangeness becomes obvious. The three longest bars on the published chart are all in the middle column. Every arrow that reaches them comes from the left. And no health ministry has an instrument that acts on the middle column at population scale, except by acting on the left column or by treating individuals once they are already there.

The chart's three longest bars are in the middle column. None of them is a place to intervene.

Tobacco and air pollution are worth separating out because they behave differently: both contribute to the middle column and both also act directly on the endpoints, through carcinogenesis and respiratory damage. They are, in this picture, the two entries that are simultaneously large, upstream and directly operable — which is exactly why they dominate the effective-intervention literature.


5. Where to Intervene

Reading the map rather than the ranking produces a short list — and it is not the list that usually gets produced.

Three criteria follow from the map. An intervention should be upstream, acting on the left column rather than the middle; operable, meaning an instrument exists that a government can actually apply, rather than an outcome it can merely wish for; and evidenced, with measured effects at population scale rather than in principle.

The World Health Organization maintains a costed list that satisfies all three: the twenty-nine "best buys" for preventing noncommunicable disease, covering tobacco, alcohol, unhealthy diets and physical inactivity. Their returns are unusually well quantified.

Figure 4 — What the best buys return

Every US$1 invested returns roughly fourfold in social and economic benefit over five years, and sevenfold if sustained to 2035. Full implementation would save 12 million lives and prevent 28 million heart attacks and strokes by 2030, with economic gains above US$1 trillion; spending about 0.6 per cent of low- and middle-income countries' gross national income would put 90 per cent of them within reach of the SDG target on NCD mortality. Source: WHO, as reported by Health Policy Watch; see the WHO's own summary of the best buys.

Table 2 — Interventions mapped to the causal map, not to the ranking

Lever Acts on Reaches which of the long bars
Tobacco taxation, advertising bans, plain packaging Tobacco, directly Smoking; and, through the cardiovascular pathway, blood pressure
Mandatory sodium reformulation and front-of-pack labelling Diet composition, without requiring any behaviour change Sodium; high blood pressure
Taxes on alcohol and sugar-sweetened beverages Alcohol and diet, through price Alcohol; obesity; high blood sugar
Clean household fuels and emissions standards Air pollution, directly Indoor and outdoor air pollution
Primary-care hypertension and diabetes management The middle column, individually Blood pressure and blood sugar, once already present

5.1 Price and reformulation, not exhortation

There is a pattern in that table worth stating plainly, because it contradicts the framework that almost every discussion of this chart produces. The interventions that carry the measured population-level effect are overwhelmingly fiscal and regulatory: taxes, mandates, standards, bans. Public education and awareness campaigns appear in every list of recommendations and carry comparatively little of the effect.

The WHO's own framing is blunt about it: taxes on tobacco, alcohol and sugary drinks are among the most effective steps a government can take, and they raise revenue rather than costing it. Brazil roughly halved its smoking rate through sustained increases in tobacco taxation. Mexico's tax on sugary drinks reduced consumption while generating significant public revenue. Thailand channels tobacco and alcohol tax receipts into a dedicated health-promotion foundation.

Sodium reformulation is the clearest case of all, because it requires no behaviour change whatsoever. When manufacturers are required to lower the salt content of bread and processed food, consumption falls without any individual deciding anything. The instrument acts on the left column and the middle column moves, and nobody has to be persuaded of anything — which is precisely why it works where advice does not.


6. When the Method Changes

A final reason not to treat the ranking as a budget allocation: the numbers move when the methodology is revised, and sometimes by a lot.

The chart puts air pollution at 4.9 million deaths for 2017. The Health Effects Institute's State of Global Air 2024, using 2021 data and updated exposure–response functions, attributes 8.1 million deaths a year to air pollution and ranks it the second-leading risk factor for death worldwide.

Figure 5 — One quantity, two published estimates

Not a contradiction: different years, different data and revised concentration–response relationships, particularly for cardiovascular endpoints. But the difference is larger than most of the bars on the chart, and it moves air pollution from fourth place to second.

The lesson is not that either estimate is wrong. It is that a ranking sensitive to methodological revision is a poor foundation for allocating budgets, because the allocation would have to be redone every time a coefficient is updated — and the underlying physical reality would not have changed at all.

This is a further argument for intervening on the left column. Whether air pollution killed 4.9 million people or 8.1 million, the instruments are the same: cleaner fuels, emissions standards, and getting combustion out of the places where people live and breathe. Those do not need re-ranking when the epidemiology is revised. An earlier report in this series takes up what that implies when the same combustion is also a climate problem, and another examines why cleaning the air and cooling the planet are not the same intervention.


7. An Agenda

How to use the chart, and what to do about what it shows.

7.1 Read it correctly

7.2 Intervene upstream

7.3 Prefer instruments that do not require persuasion

7.4 Buy the best buys

7.5 Conclusion

The risk-factor chart is a genuine achievement and deserves its circulation. What it does not deserve is the reading it usually gets, in which the bars are a queue and the longest one goes first. The bars overlap so heavily that they sum to twenty per cent more deaths than occurred; they mix physiological states with behaviours, exposures and nutrients; and the ordering shifts when an analyst redraws a category boundary or an epidemiologist revises a coefficient.

Redrawn as a causal map, it says something clearer and more useful than the ranking did. The largest measured burdens sit in the middle of the chain, where no policy instrument reaches. The instruments sit upstream, on a short list of drivers: tobacco, air pollution, what is in the food, alcohol, and whether people move. That list is neither surprising nor new, which is rather the point — the difficulty has never been identifying it.

What is genuinely useful is the second observation: that the effective instruments are fiscal and regulatory rather than educational, and that they are already specified, costed and evaluated by the WHO at a fourfold return. A well-drawn chart of where people die can tell you a great deal about the world. It cannot tell you what to do, and the answer to that question turns out to be written down somewhere else, and mostly unbought.


References

Every quantitative claim above is attributed inline. The principal sources are collected here.


Metadata

Keywords
global burden of diseaseattributable burdencomparative risk assessmentrisk factorsmortalitydouble countingnoncommunicable diseaseWHO best buystobacco taxationsodium reformulationair pollutionhypertensiondata interpretationpublic health policy
Topics
Public Health Institutions & Governance Emerging Markets Technology
JEL classification
I18, I12, Q53, H51, I15 — government policy and regulation of health; health behaviour; air pollution; government expenditure on health; health and economic development
Data and method
This report analyses the Our World in Data rendering of IHME Global Burden of Disease 2017 attributable-death estimates, together with the WHO's costed 'best buys' for noncommunicable disease prevention and the Health Effects Institute's State of Global Air 2024. The 66.9 million total is the author's own sum of the thirty-four values as charted, and is presented as such; the comparison figure of approximately 56 million all-cause deaths is from the same GBD 2017 study. Global death totals and the WHO return-on-investment figures were verified by search in September 2026. Figures 1, 2, 3 and 5 replot published values without modelling. Table 1's taxonomy of risk-factor kinds and the causal arrangement in Figure 3 are the author's, applied to the chart's own entries; Figure 3 shows dominant directions of causation rather than full causal complexity, and omits age, genetics and other determinants. The report does not dispute the GBD methodology, which it describes as deliberate and correct, and argues only against a particular inference commonly drawn from its output. The report is analytical rather than predictive.
Report
H Heuristics Digital Report № 2026-11 · Published 13 September 2026
Licence
CC BY-NC-ND 4.0
Cite as
Hunter Hughes (2026). The Bars Do Not Add Up: What the global risk-factor ranking measures, what it cannot tell you, and where to intervene instead. H Heuristics Digital Report 2026-11. https://digitalreports.hheuristics.com/reports/risk-factor-chart-misread/
↑ Back to top