The finding is a well-powered NULL. Across 10 prospective studies (1.4 million women; 4660 breast cancer cases among night-shift workers), pooled breast cancer incidence for ever vs never night shift work was RR 0.99 (95% CI 0.95 to 1.03) — no association, with narrow intervals and no significant heterogeneity. (Travis et al., 2016)
«Altogether, these studies included 4660 women with breast cancer who had worked night shifts (Figure 2). When results from the 10 studies were combined, the weighted average RR was 0.99 (95% CI¼ 0.95 to 1.03) for any night shift work compared with none. There was no statistically signifi- cant heterogeneity across studies (P ¼ .052).» (Travis et al., 2016)
This is the occupation cluster’s circadian channel — a fourth distinct occupational exposure alongside physical demand, psychosocial strain, and cognitive stimulation -> The Physical Activity Paradox. Unlike those three, its patient-important verdict is a credible no-meaningful-effect, not a harm or a benefit.
Magnitude, with the long-duration arm and its bounds
The dose axis is duration of night shift work, and the null holds — indeed tightens — at the high end:
| Exposure contrast | Studies | Pooled RR (95% CI) | Note |
|---|---|---|---|
| Ever vs never | 10 | 0.99 (0.95-1.03) | 4660 exposed cases |
| >=20 years vs never | 8 | 1.01 (0.93-1.10) | narrow — excludes a moderate association |
| >=20 years (updated NHS) | 8 | 0.97 (0.90-1.06) | re-run with newer Nurses’ Health results |
| >=30 years vs never | 4 | 1.00 (0.87-1.14) | sparse — <1000 long-duration cases total |
The confidence intervals carry the decision, so they are stated with the estimate, not stripped to a point. The long-duration interval is what licenses the strong reading:
«Confidence intervals for the incidence rate ratios are narrow, even for 20 or more years of night shift work (RR ¼ 1.01, 95% CI ¼ 0.93 to 1.10), so these findings exclude a moderate association of breast can- cer incidence with long duration night shift work. The main limitation of the present findings is that an in- crease in the relative risk of breast cancer incidence of only a few percent cannot be ruled out.» (Travis et al., 2016)
Evidence state: no meaningful (moderate) effect — NOT insufficient evidence. The expectancy test passes: 1.4M women, 4660 exposed cases, intervals tight enough to rule out a moderate RR even at >=20 years. What survives is the honest residual — a few-percent relative increase cannot be excluded, because fewer than 1000 long-duration cases have yet accrued (already >4x the case count the 2007 IARC review had). That residual is a bounded uncertainty, not an open question.
Why the prospective restriction is the load-bearing move
The value here is not only the null but why it overturns an earlier positive signal. The prior evidence base for a night-shift/breast-cancer link was case-control-heavy; Travis’s meta-analysis restricts to prospective designs, five of which (0.8M women) postdate the previous meta-analyses.
«Restriction to pro- spective studies is important when trying to detect or refute moderate hazards as it avoids the moderate biases that can re- sult from retrospective methodology.» (Travis et al., 2016)
This is a live occupational-epidemiology instance of the prospective-overturns-case-control discriminator -> Upgrading Observational Evidence, the exact structure of Willett’s dietary-fat -> breast-cancer reversal (strong case-control associations that prospective studies did not reproduce, diagnosed as artifacts of noncomparable case-vs-control data collection). Same outcome, same design-class flip, a different exposure — recall bias in retrospective exposure ascertainment manufactures a moderate association that the prospective ascertainment dissolves. (inferred from Travis et al., 2016)
The alternative explanations were tested, not assumed away
The null is robust to the three ways it could be spurious:
- Residual confounding. Night-shift workers differed (more likely obese, smokers, on sleep medication, evening chronotype), so residual confounding «cannot be completely excluded» — but the null held «either in minimally adjusted or in multivariable-adjusted analyses». (Travis et al., 2016)
- Screening bias (a masked association). The hypothesis that shift workers’ different screening uptake hid a real effect was tested directly: in the Million Women Study there was no difference by shift category in the proportion of cancers detected by screening — the masking mechanism did not operate. (Travis et al., 2016)
- Exposure misclassification. Self-reported night work had good repeatability, and long-duration workers «are likely to have had substantial exposure… so misclassification is unlikely to have been so great as to have masked any material risk». (Travis et al., 2016)
- Effect modification. No modification by sleep characteristics, diurnal preference, family history, or recency; and a nurse-specific sensitivity analysis (MWS, >=10 years as a nurse) showed no significant increase — so the earlier nurse-cohort signals are not an occupation-specific exception. (Travis et al., 2016)
That each artifact route was probed and closed is what moves this from insufficient to credible no-effect.
The IARC pole — now held, with its own evidence basis
Travis draws an explicit conclusion against the IARC 2007 classification of shift work as a probable (Group 2A) human carcinogen:
«The totality of the current prospective evidence suggests that night shift work, including long-term night shift work, has little or no effect on breast can- cer incidence. The IARC 2007 shift work review was necessarily based on limited epidemiological evidence, and, although fur- ther follow-up is desirable, the prospective evidence now avail- able shows that classification of night shift work as a probable human (breast) carcinogen is no longer justified.» (Travis et al., 2016)
The IARC source is now held, so its reasoning can be read in its own words. The 2A rests on a two-legged basis — the standard IARC combination rule of limited human evidence AND sufficient animal evidence:
«On the basis of “limited evidence in humans for the carcinogenicity of shift-work that involves nightwork”, and “suffi cient evidence in experimental animals for the carcinogenicity of light during the daily dark period (biological night)”, the Working Group concluded that “shift-work that involves circadian disruption is probably carcinogenic to humans” (Group 2A).» (Straif et al., 2007)
The classified exposure is precisely “shift-work that involves circadian disruption”, not shift-work per se. IARC’s own human leg was explicitly limited — and its limits are the exact evidence class Travis’s prospective restriction was built to supersede:
«Six of eight epidemiological studies from various geographical regions, most notably two independent cohort studies of nurses engaged in shift- work at night» [refs 2,3] «have noted a modestly increased risk of breast cancer in long-term employees compared with those who are not engaged in shiftwork at night.» (Straif et al., 2007)
IARC itself flagged these as «limited by potential confounding» and by inconsistent shift-work definitions, «with several focused on a single profession» (Straif et al., 2007) — case-control/retrospective-heavy, single-profession (the two Schernhammer Nurses’ Health cohorts among them), precisely the evidence Travis’s prospective MA re-runs and overturns.
Hazard vs risk — a not-joined DISTINCTION, not a two-sided tension
With IARC’s own basis now readable, the counter-passage check resolves the relation cleanly, and it is not a type-D tension. The two sources answer different questions (not-joined check (ii), different scope/unit -> Non-Sugar Sweeteners for the same hazard-vs-risk split on aspartame):
- IARC does hazard identification — can circadian disruption cause cancer under some circumstances, in any tissue? — and its affirmative rests primarily on the sufficient-ANIMAL leg (light-at-night tumour studies), the human leg being only limited.
- Travis quantifies population risk — by how much does breast cancer incidence change in real shift-working women, at realistic exposure? — and finds a well-powered null.
A true hazard driven by animal + mechanistic evidence is logically compatible with a null population risk estimate, so the apparent contradiction largely dissolves without either side being wrong. The sharp, decision-relevant consequence of holding IARC’s own basis: Travis’s human-epidemiology null bounds (and, she argues, removes) IARC’s human leg, but a breast-cancer-incidence null in women does not by itself address the animal + mechanism leg the 2A actually rests on. That asymmetry is exactly why this is hazard-vs-risk rather than a clean refutation. (Straif et al., 2007; inferred from Travis et al., 2016)
Where the two DO join issue — the human-evidence question — the difference is a legitimate class, not a process defect. IARC 2007’s limited human leg and Travis 2016’s prospective null are different evidence bases nine years apart: Travis’s five newest prospective studies (0.8M women) postdate the meta-analyses IARC relied on. This is the different-evidence-base + lag guidance-difference class (reasons 2 and 4), NOT a process defect (reason 5) — IARC applied a structured hazard-classification procedure faithfully on the evidence it had; the human evidence simply improved after it reported. On that human leg specifically, the current prospective evidence favours Travis. (Straif et al., 2007; inferred from Travis et al., 2016)
Bearing on the circadian-oncogenesis mechanism
Sleep Regularity and Mortality holds a circadian-oncogenesis pathway (irregular sleep-wake timing -> clock desynchrony -> cancer), admitted directionally as mechanism. Night shift work is the most-studied occupational circadian disruptor, so this null is a well-powered human test of that pathway’s oncogenic potency in breast tissue — and it is negative for a moderate effect. This attenuates the magnitude one should attach to the circadian-cancer mechanism, without refuting it: night-shift schedule and general population sleep-timing irregularity are different exposures (not-joined check (ii): different unit and disruptor), and the pathway may still operate for other cancers or via other disruptors. Held as a bound on mechanism weight, not a mechanism refutation. (inferred from Travis et al., 2016)
The IARC statement is the mechanistic-evidence pole of that same pathway, and holding it sharpens the bound. IARC’s sufficient-animal leg is exactly this mechanism: «More than 20 studies investigated the eff ect of constant light, dim light at night, simulated chronic jet lag, or circadian timing of carcinogens, and most showed a major increase in tumour incidence» (Straif et al., 2007), with the proposed route running via melatonin suppression, clock-gene (Per2/PERIOD) deregulation, and impaired immunosurveillance (NK-cell/Th1-Th2). So the pathway has sufficient animal + molecular backing yet a null realized effect on breast-cancer incidence in exposed women — the F-refinement is that mechanism potency in a model system does not transport to a measurable population risk on this outcome (net-effect-not-intended; whole-organism compensation). This is directional mechanism, not a human-outcome finding. (inferred from Straif et al., 2007)
Decision relevance
- For a night-shift worker weighing breast cancer risk: on the current prospective evidence, night shift work — including long-term — has little or no effect on breast cancer incidence; a moderate risk is excluded, only a few-percent increase remains unexcludable. This is a reassurance decision-change against the widely-communicated 2A probable carcinogen message.
- The lever, if any, is elsewhere. Shift work’s plausible health costs sit on other outcomes and pathways (metabolic, cardiovascular, sleep-loss) -> Job Strain and Coronary Heart Disease, Sleep Regularity and Mortality — not on this cancer outcome. Do not let the breast cancer null read as a clean bill for shift work overall.
- Symmetric standards. This null was extracted with the same rigor a positive finding gets — the narrow CIs, the tested artifact routes, and the prospective-design upgrade are what earn it medium confidence, exactly as they would earn a harm finding.
Confidence: MEDIUM — why
- Gold meta-analysis, large and prospective — 1.4M women, 10 prospective cohorts, narrow pooled CIs, low heterogeneity despite design/population/exposure-definition differences; the artifact routes (confounding, screening, misclassification, effect modification) were each probed and closed.
- Held below high because: observational throughout (no randomization is possible for this exposure); a few-percent relative increase is not excluded at long durations (<1000 cases); and the IARC-2A relation, while now readable from both poles, is a hazard-vs-risk distinction whose animal + mechanism leg the breast-cancer null does not adjudicate.
- Coherence, not validity (R1): the pooled data show no moderate association in these cohorts; that night shift work causes no breast cancer is the inference the design supports directionally, and the open loop (no realized-outcome check) stands.
Self-critique [run 2026-08-20, before commit — IARC pole landed]
- The NULL is not softened or inflated. Magnitudes are stated with CIs and studied range (ever /
=20y / >=30y); the residual few-percent uncertainty is kept visible; the evidence state is named no-meaningful-effect, distinguished from insufficient.
- The IARC relation is a not-joined DISTINCTION, not a two-sided tension — verified against IARC’s own quoted basis. The counter-passage check now runs on the held IARC source (not a Travis paraphrase): the two answer different questions (hazard identification vs population-risk magnitude, not-joined check (ii)), and IARC’s own two-legged basis (limited human + sufficient animal) is quoted. Where they join issue — the human-evidence leg — the difference is classed different-evidence-base + lag (2/4), explicitly NOT a process defect (5): IARC applied its procedure faithfully on 2007 evidence.
- The added value beyond either source alone (F/B). Holding IARC’s own basis lets the page make a move neither source makes: the 2A rests primarily on the animal leg, so Travis’s human null bounds the human leg but does not adjudicate the animal + mechanism basis. That is quoted-and-reasoned, not asserted; both new INFERRED tags name IARC’s quoted leg and mark the reading as the wiki’s.
- No over-claim on the mechanism. IARC’s sufficient-animal evidence is admitted as directional mechanism only (animal/molecular,), never dressed as a human-outcome finding; the transport-gap reading against Travis’s null is explicitly the wiki’s synthesis.
sources:dual test satisfied. Straif carries distinct extracted claims on this page (the 2A two-legged basis, the limited-human leg, the animal-evidence quote), so it earns asources:slot — not a corroboration-only pad.