Is There a Weekly Hours Threshold Where Risk Jumps?

Kivimäki et al. (2015) pooled 25 studies from 24 cohorts in Europe, the USA and Australia and reported a graded series for stroke against a reference of 35 to 40 hours per week. The 41 to 48 hour band came in at RR 1.10 (95% CI 0.94-1.28, p=0.24), an estimate not statistically distinguishable from no difference because the interval crosses 1. The 49 to 54 hour band reached RR 1.27 (95% CI 1.03-1.56, p=0.03), and 55 hours or more reached RR 1.33 (95% CI 1.11-1.61, p=0.002), with a test for trend at p<0.0001.

That pattern is what makes a 55 hour line defensible as shorthand and misleading as biology. The bands were drawn by the analysts, and the data behind them describe a graded trend rather than a discontinuity at one particular hour. Nothing switches on at hour 55. What changes is that the estimate moves far enough from 1, with a tight enough interval, for the association to hold up as a statistical finding.

The stroke estimate also survived stress testing. The authors report that the excess risk persisted in analyses addressing reverse causation, in models adjusted for other risk factors, and across different methods of ascertaining stroke, with estimates ranging from RR 1.30 to 1.42. That range brackets the main figure rather than collapsing toward 1, which is the reassuring direction for a sensitivity analysis to move.

The WHO and ILO re-examined the same question using the Navigation Guide and GRADE. Descatha et al. (2020) pooled 22 studies covering 839,680 participants across 8 countries and 3 WHO regions. For 55 hours or more, stroke incidence came in at RR 1.35 (95% CI 1.13-1.61) from 7 studies and 162,644 participants with follow-up of 1 to 20 years and heterogeneity near zero (I-squared 3%), graded as sufficient evidence for harmfulness. The 49 to 54 hour band (RR 1.13, 95% CI 1.00-1.28, p=0.04) was graded limited evidence, and the 41 to 48 hour band (RR 1.04, 95% CI 0.94-1.14) was graded inadequate, with the review team describing itself as uncertain about the effect at that exposure.

Stroke mortality is a different story. No exposure band reached significance, including 55 hours or more at RR 1.08 (95% CI 0.89-1.31), and every mortality estimate was graded low quality (Descatha et al., 2020). The durable claim is about the incidence of stroke, not death from it. Any summary that blurs the two outcomes is stating something firmer than either review found.

The two reviews are not separate lines of evidence. Rugulies and Siegrist appear as authors on both, and the WHO and ILO review pools primary studies drawn from the same IPD-Work Consortium cohorts that Kivimäki et al. (2015) used. Read Descatha et al. (2020) as a partially overlapping re-appraisal under a formal grading system rather than as confirmation from a fresh evidence base, which is a meaningful difference when you are weighing how much two agreeing papers add.

RR 1.33 means the pooled rate of stroke among people reporting 55 or more hours was 33% higher in relative terms than among people reporting 35 to 40 hours. It does not say how many additional strokes that represents, and the paper does not report absolute risk. Stroke is uncommon in working-age populations, so a 33% relative increase on a low baseline is a small absolute increase. Converting the ratio into an expected number of cases in a workforce is not something this evidence supports.

The stroke arm rests on 1,722 events across 3.8 million person-years, with a mean follow-up of 7.2 years (Kivimäki et al., 2015). That is a modest event count behind a number this widely quoted. It does not make the finding wrong, and the confidence interval already carries that uncertainty, but it does mean the estimate should be read as a range rather than as a precise value.

The coronary heart disease half is weaker and should not be quoted with equal confidence. It pooled 603,838 people free of coronary heart disease at baseline, 5.1 million person-years, a mean follow-up of 8.5 years and 4,768 events, returning RR 1.13 (95% CI 1.02-1.26, p=0.02), an interval whose lower bound barely clears 1. The authors' own summary is that "the association with coronary heart disease is weaker" (Kivimäki et al., 2015). The dose-response series that paper published runs for stroke, and its coronary result travels as one pooled estimate against the same 35 to 40 hour reference rather than band by band. Per-band estimates for the cardiac outcome do exist, in the later WHO and ILO review of ischemic heart disease (Li et al., 2020), where the bands below 55 hours are largely null.

The cohorts were working-age employees in Europe, the USA and Australia, extended by the WHO and ILO review to 8 countries across three WHO regions. These are predominantly high-income, formally employed populations. The exposure under study was weekly hours, not shift timing, so the findings do not transfer automatically to shift workers as a group, to informal economies, or to the self-employed.

Hours were self-reported in every included study and in most cases measured once at baseline, and the primary adjustment covered age, sex and socioeconomic status only. Authors and later reviewers flag residual confounding from health behaviors, workload and sleep duration, along with possible selection effects. Subgroup analyses in the WHO and ILO review found no evidence of differences by region, age, sex, socioeconomic status or stroke type (Descatha et al., 2020). Read that as no detected difference rather than as a demonstration that risk is uniform, because a subgroup comparison carries less statistical power than the pooled estimate and a real difference inside a smaller stratum can go unseen.

The literature search behind Kivimäki et al. (2015) closed on August 20, 2014, which makes the primary anchor more than a decade old. Descatha et al. (2020) is the current WHO and ILO position and should travel with it. Citing the 2015 paper on its own presents a snapshot of the field as though it were the field's current state, and it also drops the mortality null that the later review added.