Does Night Shift Work Shorten Your Life?

The most repeated claim about this study is the one it never makes. Gu et al. (2015) followed 74,862 US female registered nurses from June 1988 to June 2010, roughly 1.5 million person-years, and recorded 14,181 deaths, of which 3,062 were cardiovascular and 5,413 were cancer. Every result is a hazard ratio comparing nurses at a given level of lifetime rotating night exposure against nurses who never worked rotating nights. There is no absolute risk in it, no years-of-life-lost calculation, and no life-expectancy estimate.

The honest reading of the headline figure, a hazard ratio of 1.11 (95% CI 1.06, 1.17) at 6 to 14 years, is roughly an 11% higher relative rate of death during the follow-up window. It is not a statement that anyone lost a measurable amount of life. The paper reports no years-of-life-lost figure, so any number of years off your life did not come from it, and the distinction matters the moment the finding gets quoted back to a night team.

Exposure was asked once, in 1988, as total years worked rotating night shifts of at least three nights per month in addition to days or evenings in that month, across eight response categories. Person-years split roughly 41% never, 41% at 1 to 5 years, 11% at 6 to 14 years, and 7% at 15 or more years, so most of the cohort sat at the low end of the exposure range.

At 1 to 5 years, all-cause mortality came in at a hazard ratio of 1.01 (95% CI 0.97, 1.05) and cardiovascular mortality at 1.02 (95% CI 0.94, 1.11). Neither reaches statistical significance. The association appears in the two longest categories, 1.11 (95% CI 1.06, 1.17) at 6 to 14 years and 1.11 (95% CI 1.05, 1.18) at 15 or more years, with a p for trend below 0.001.

Notice what the two long-duration estimates do and do not do. For all-cause mortality they are identical at 1.11, so this cohort does not show risk climbing steadily with every additional decade. Cardiovascular mortality does step up across the categories, 1.19 (95% CI 1.07, 1.33) at 6 to 14 years and 1.23 (95% CI 1.09, 1.38) at 15 or more. A threshold somewhere around five to six years fits these data better than a smooth per-year gradient does.

Cardiovascular death is where the association concentrates, and within it the specific driver is ischemic heart disease: a hazard ratio of 1.22 (95% CI 1.02, 1.46) at 6 to 14 years and 1.31 (95% CI 1.09, 1.59) at 15 or more years, with a p for trend below 0.001. Cerebrovascular disease and other cardiovascular causes showed no significant association (Gu et al., 2015).

The cancer results are the ones most likely to be read past their adjustment. Age-adjusted, all-cancer mortality looked elevated at 1.10 (95% CI 1.01, 1.20) for 6 to 14 years and 1.20 (95% CI 1.09, 1.32) for 15 or more. After multivariable adjustment those fell to 1.04 (95% CI 0.95, 1.13) and 1.08 (95% CI 0.98, 1.19), neither statistically significant. The multivariable figures are the ones to quote, because the age-adjusted set runs higher across every endpoint and overstates the finding.

Lung cancer was the only site that stayed significant at 15 or more years, at 1.25 (95% CI 1.04, 1.51) after adjustment, down from an age-adjusted 1.44 (95% CI 1.20, 1.73). The authors identify smoking as the confounder driving that estimate. Breast cancer at 15 or more years was flat at 0.99 (95% CI 0.74, 1.33), and even at 30 or more years the estimate of 1.47 (95% CI 0.94, 2.32) did not reach significance.

Treat the isolated site findings as exploratory rather than as results. Kidney cancer came in at 1.72 (95% CI 1.03, 2.86) for 6 to 14 years, significant in that middle category alone, then fell to 1.39 (95% CI 0.75, 2.57) for 15 or more, with a p for trend of 0.048, which is the shape of noise rather than of a dose response. Myeloma sat at 1.56 (95% CI 0.93, 2.64) and 1.61 (95% CI 0.90, 2.88), with a p for trend of 0.08. Many endpoints were tested, most of those intervals cross 1, and the authors report that results were largely unchanged when they required a p value below 0.01.

This is a prospective observational cohort, and the authors write in associations throughout, describing a "potentially detrimental effect" on health and longevity and asking for future work on residual confounding. Nurses with longer night durations were older at baseline, heavier, more likely to be current smokers, drank less alcohol, ate less cereal fiber, and had more diabetes, hypertension, and hypercholesterolemia. Adjustment was extensive, covering smoking and pack-years, BMI, diet, activity, alcohol, menopausal status, hormone use, and an education-based socioeconomic proxy, but residual confounding at hazard ratios of 1.1 to 1.2 cannot be excluded.

The exposure measure is narrow and fragile. It was self-reported once in 1988, never updated across 22 years of follow-up, and the authors concede it cannot be validated, arguing that any resulting misclassification would tend to push estimates toward the null. It also did not distinguish rotating from permanent night shifts, so this is not evidence about permanent night teams. In the sister nurses cohort, fewer than 10% of nurses worked permanent nights.

The population is narrow too. Everyone in it was a US female registered nurse, aged 30 to 55 at enrollment in 1976, predominantly white, and relatively homogeneous in socioeconomic status. There are no men and no other occupations. Exposure dates from 1988 and follow-up closed in 2010, before much of the current thinking on fatigue management and rotation design took hold, which is a reason to be careful about transplanting these numbers onto a modern roster.

Two further limits should travel with the numbers. In the smoking stratification, the all-cause association at 15 or more years was stronger among current smokers, at 1.21 against 1.11 in the full sample, but it remained statistically significant among never smokers, with a p for trend of 0.05, and the cardiovascular and ischemic heart disease estimates were comparable among never smokers. Separately, the authors acknowledge that night shift work is difficult to separate from lack of sleep, although restricting the analysis to nurses sleeping 6 to 8 hours did not materially change the results.