What Does the 745,000 Deaths Estimate Measure?

Pega et al. (2021) modeled how many people in each country worked long hours, multiplied that exposure through relative risks taken from two earlier systematic reviews of prospective cohorts, and applied the resulting population-attributable fractions to WHO Global Health Estimates of total deaths. The exposure side ran on 2,324 cross-sectional surveys and 1,742 quarterly datasets through a four-stage pipeline: a multilevel model of working-hours categories by country, sex, age and year, a multinomial logit model of transitions between those categories, a microsimulation over a ten-year exposure window, and then the risk assessment itself. Exposure was estimated for 194 countries and attributable burden for 183.

The headline totals are 745,194 deaths (UR 705,786-784,601) and 23.3 million DALYs (UR 22.2-24.4 million), split into 346,753 ischemic heart disease deaths (UR 319,658-373,848) and 398,441 stroke deaths (UR 369,826-427,056). Expressed as attributable fractions, long hours were assigned 3.7% (UR 3.4-4.0) of ischemic heart disease deaths and 6.9% (UR 6.4-7.5) of stroke deaths, and 5.3% (UR 4.9-5.6) and 9.3% (UR 8.7-9.9) of the respective DALYs. On the exposure side, the model put 488 million people at 55 or more hours per week in 2016.

Two framing points matter before the number gets repeated. The reference year is 2016 and the paper appeared in 2021, so this is not a current annual toll and it does not describe last year. And while the paper calls 488 million people 8.9% (UR 8.6-9.1) of the global population, that percentage behaves like a share of the working-age population rather than of everyone alive, so the safer phrasings are the raw 488 million or about 9% of working-age people.

Stroke, not heart disease, is the larger half of the estimate at 398,441 of the 745,194 deaths. Those stroke deaths were modeled using a relative risk of 1.35 (95% CI 1.13-1.61) for working 55 or more hours per week compared with 35 to 40 hours. That figure comes from Descatha et al. (2020) and it is the estimate for non-fatal stroke incidence, pooled from 7 studies and 162,644 participants, rated moderate quality with sufficient evidence of harmfulness.

The same review produced a separate estimate for stroke mortality at 55 or more hours: RR 1.08 (95% CI 0.89-1.31), pooled from 10 studies and 664,647 participants. That confidence interval crosses 1, the finding was rated low quality, and the evidence was classified as inadequate for harmfulness. In other words, the direct evidence that long hours are associated with having a stroke is considerably stronger than the direct evidence that they are associated with dying of one, and the model used the stronger of the two to generate deaths.

The authors state the bridge openly. Because there is no evidence that the relative risks for mortality and morbidity differ, they assume the population-attributable fractions are the same for fatal and non-fatal events, and apply the same fractions to produce both deaths and DALYs. That is a documented modeling assumption rather than a concealed error, and it may well be right. It is also load-bearing for more than half of the headline, which is why it belongs next to the number every time the number is used.

The ischemic heart disease half rests on firmer footing. Its input, RR 1.17 (95% CI 1.05-1.31) from Li et al. (2020), is itself a fatal-event estimate, pooled from 16 studies and 726,803 participants and rated moderate quality with sufficient evidence of harmfulness. The WHO and ILO press phrasing about a 35% higher risk of stroke and a 17% higher risk of dying from ischemic heart disease maps onto exactly these two risk ratios.

The paper is not internally consistent about causation. Its abstract conclusion uses the word causes, while its discussion concedes that certainty about a causal association is not available and rests the case on the fact that most of the high-quality evidence came from prospective cohort studies. All the underlying evidence is observational and the authors accept that residual confounding cannot be ruled out. Associational language is therefore the defensible choice, even though the abstract of the source does not use it.

The grading itself was contested. The authors record that some experts disagreed with the sufficient evidence rating for ischemic heart disease specifically. Underlying evidence quality was rated moderate rather than high, and the risk increases are modest in size. Potential mediators such as smoking and physical inactivity were difficult to separate out, so some part of the pattern may run through behavior that accompanies long hours rather than through the hours themselves.

Generalizability is the other open question. The cohort evidence came predominantly from high-income countries but was applied to 194 countries, while the largest modeled burdens fall in South-East Asia and the Western Pacific. South-East Asia carried the highest exposure prevalence at 11.7% (UR 10.8-12.5) against 3.5% (UR 3.5-3.6) in Europe. Effect modification by socioeconomic status remains unclear, all working-hours data are self-reported with quality varying by country, and the model assumes a ten-year latency between exposure and clinical disease while assigning each individual to their highest exposure category in the window. Shift work and seasonal variation were not accounted for.

Independence deserves a plain statement. This is not an outside audit of the WHO and ILO numbers, it is the Joint Estimates output itself: the lead author was a WHO technical officer, the burden was assessed against WHO Global Health Estimates, and funders included the US NIOSH and CDC and the German Federal Ministry of Health. Two of the model inputs are papers co-authored by members of the same author group, which is normal for a joint estimates program and still worth knowing. Alexis Descatha, first author of the stroke review whose relative risk drives the stroke half, is a co-author here and declared personal fees from the University of Angers, Angers University Hospital Center and Elsevier, the journal publisher, plus a regional grant, outside the submitted work.

The threshold is the operational content. Everything in this estimate concerns 55 or more hours in a week measured against a 35 to 40 hour reference, and the bands in between were largely null in the source reviews. Ischemic heart disease incidence came out at RR 0.98 (95% CI 0.91-1.07) for 41 to 48 hours and RR 1.05 (0.94-1.17) for 49 to 54 hours; ischemic heart disease mortality at RR 0.99 (0.88-1.12) and RR 1.01 (0.82-1.25); stroke incidence at RR 1.04 (0.94-1.14) and RR 1.13 (1.00-1.28), the last of which touches 1 at its lower bound. The evidence draws its line at 55 hours, not at overtime in general.

The burden is also unevenly distributed, which tells you where to look first. Exposure prevalence was 13.2% (UR 12.7-13.7) among males and 4.5% (UR 4.2-4.8) among females, and males accounted for 538,852 of the 745,194 deaths, the 72% male share quoted in the WHO and ILO announcement of 17 May 2021. Numbers and rates rose with age up to about 70 years, with the 60 to 74 age groups carrying the larger stroke burden. Regionally, South-East Asia was assigned the largest attributable ischemic heart disease burden at 159,832 modeled deaths (UR 135,442-184,242) against 16,942 (UR 15,878-18,005) in Africa, and the largest attributable stroke burden at 158,987 modeled deaths (UR 141,968-176,006) against 18,285 (UR 17,162-19,409) in the Americas.

The trend figures are the easiest to conflate, because the disease-specific rises and the combined figure come from different sources. Between 2000 and 2016 the global prevalence of working 55 or more hours per week rose 9.3% (UR 4.3-14.6), attributable ischemic heart disease deaths rose 41.5% (UR 27.9-56.5) and attributable stroke deaths rose 19.0% (UR 7.8-31.1). The widely quoted 29% increase in the combined total since 2000 comes from the WHO and ILO announcement rather than from the paper abstract; it is consistent with those disease-specific figures, and it should be attributed to the announcement when it is cited.