Is Shift Work Linked to Depression?

Torquati et al. (2019) searched the literature to August 2018 and pooled 7 prospective cohort studies covering 28,431 unique participants and 14 separate adjusted risk estimates, comparing shift workers with non-shift workers. The review screened 1,902 retrieved records down to 639 by title and abstract and 51 at full text, excluded 2 otherwise eligible studies because their data were reported in continuous format, and included 7. For depressive symptoms the pooled estimate was ES = 1.33; 95% CI = 1.02, 1.74; I2 = 31.5%, which the paper describes as a 33% higher risk of depressive symptoms in shift workers than in non-shift workers.

Two details govern how far that number travels. The measure is a ratio of adjusted estimates pooled under a random-effects model, which the authors label an effect size rather than a clean risk ratio, so it describes the relative standing of two groups and not a count of extra cases in your building. And the lower confidence bound of 1.02 leaves the pooled result almost no margin above 1.0, which makes the finding positive but imprecise.

The combined adverse mental health outcome pooled to ES = 1.28; 95% CI = 1.02, 1.62; I2 = 70.6%, with heterogeneity high enough that the single pooled figure hides real disagreement between the studies feeding it. Anxiety symptoms (ES = 1.20; 95% CI = 0.85, 1.69) and general poor mental health symptoms (ES = 1.18; 95% CI = 0.72, 1.91) both crossed no difference and were not statistically significant. This review is therefore not evidence that shift work moves anxiety.

Female shift workers were more likely than female non-shift workers to experience depressive symptoms (OR = 1.73; 95% CI = 1.39, 2.14). The male estimate was not statistically significant, with a confidence interval running from 0.49 to 2.65, wide enough to be compatible with a sizable association in either direction and with no association at all.

The metaregression is the most useful part of the paper for anyone reading across studies rather than within one. Gender explained 90.00% of the heterogeneity in results and overall risk of bias explained 0.61%. Every other moderator tested returned a negative I2, which the authors interpret as zero. The abstract states the same point as gender differences explaining more than 90% of heterogeneity.

One caution on reading that split. The male confidence interval is wide because the male subgroup is thinner, so a non-significant male estimate is an absence of evidence rather than evidence of absence. The paper also uses the term gender throughout, including in the metaregression, and nothing in it separates gender from sex, or either from the occupations, caring loads, and household circumstances that differ by gender inside these cohorts.

The female subgroup estimate is printed two different ways. The abstract and the metaregression paragraph give 1.73 (95% CI = 1.39, 2.14), while the paragraph describing Figure 3 gives ES = 1.78 with the identical interval of 1.39, 2.14. Two different point estimates cannot share one confidence interval, so at least one of them is a typographical error. The male estimate has the same problem, printed as 1.14 in one paragraph and 1.25 in another, both carrying the interval 0.49, 2.65.

The labeling is muddled as well. The abstract attaches the female figure to depressive symptoms, the Figure 3 paragraph attaches the female subgroup to combined adverse mental health outcomes, and the discussion compares it against the 1.28 combined estimate. The defensible reading, and the one the wider literature quotes, is the abstract version: female shift workers were more likely than female non-shift workers to experience depressive symptoms (OR = 1.73; 95% CI = 1.39, 2.14).

None of this makes the paper unusable, and it is worth saying that the direction and the confidence interval are stable across every printing. It does mean that anyone repeating a number from it should quote the abstract version and say so, because a citation to 1.78 and a citation to 1.73 both trace back to the same subgroup analysis in the same article.

All 7 included studies were prospective cohorts, not trials, and no participant was assigned to a schedule. Phrasing such as shift workers were more likely to report depressive symptoms is accurate; shift work causes depression is not. The authors keep to associational language throughout and conclude only that shift workers are at increased risk for poor mental health. Nothing here tells you what happens to symptoms when a schedule changes.

The outcomes are symptoms, not diagnoses. The included studies used GHQ-12, GHQ-28, HADS-14, and a self-report HPQ checklist, and the authors state that this approach would be insufficient for diagnostic purposes, which is why they write about symptoms rather than diagnoses. Reporting the finding as a rise in clinical depression or major depressive disorder would misstate what was measured.

Quality is mixed. Risk of bias was high in more than half the studies and moderate in the remainder, with exposure assessment and attrition above 20% the two weak domains in most of them. Shift work exposure was self-reported in 6 of the 7 studies and drawn from company records in one. The methods mention funnel plots and the Egger test, but no numeric result appears in the article body, so no publication-bias claim can be made in either direction.

Scope matters too. Mean participant age was 39.7 years across a range of 18 to 60, sample sizes ran from 298 to 9,765, follow-up was one year or less in 3 studies and between 2 and 10 years in 4, and occupations were nurses in 2 studies, factory employees in 1, and unspecified in 4. The tables carry the label United States, 2018, which is journal house style for where the analysis was performed; the cohorts themselves are not US populations. The search also stopped in August 2018, so cohorts published in the years since are not represented.