Are night shifts during pregnancy linked to a higher risk of miscarriage?

Begtrup et al. (2019) built the cohort by linking the Danish Working Hour Database, which holds payroll data on all Danish public hospital employees, with Danish national registers on births and on hospital admissions for miscarriage. That produced 22,744 pregnant women and 377,896 pregnancy weeks of follow-up, which the authors report as an average of 19.7 weeks per woman, a figure that does not reconcile with the cohort itself, since 377,896 divided by 22,744 is 16.6 and a window running from week 4 to week 22 caps the average at 19. Miscarriage risk during pregnancy weeks 4 to 22 was modeled with Cox proportional hazards regression using time-varying exposure and a fixed set of adjustments for potential confounders.

The headline result is that women who had two or more night shifts the previous week had an increased risk of miscarriage after pregnancy week 8 compared with women who did not work night shifts, at HR 1.32 with a 95% confidence interval of 1.07 to 1.62. Read the scale before you read the size. A hazard ratio describes the rate at which events occur among women still pregnant at each point in follow-up, so 1.32 means a 32% higher hazard, and the interval runs from roughly 7% higher to roughly 62% higher. The association is clearly positive and its magnitude is not pinned down precisely.

A hazard ratio is a comparison, not a tally, and on its own it tells you nothing about how many losses sit behind the estimate. The published abstract reports the ratio and its interval rather than the count of miscarriages in each exposure group, so this page cannot give you the absolute numbers. If someone hands you event counts attributed to this cohort, ask which table of the paper they came from before those figures travel into a policy document.

The reason this paper carries weight is measurement, not design. Earlier observational work indicated an association between night work and miscarriage, but exposure was usually reconstructed from recall or job titles, which the authors argue precludes causal inference. Payroll data gives exact, prospectively recorded shift counts that no one had to remember or estimate. Better exposure measurement strengthens the inference; it does not convert a cohort study into an experiment.

The threshold is the authors' own framing. Begtrup et al. (2019) conclude that the study corroborates earlier findings and indicates "a lowest observed threshold level of two night shifts per week". Lowest observed is doing a lot of work in that sentence. It means two shifts was the smallest weekly exposure at which this dataset detected elevated risk, given its sample size, its follow-up, and how it counted miscarriages. A level at which no signal was detected is not a level shown to be harmless.

The paper also reports that the cumulative number of night shifts during pregnancy weeks 3 to 21 was associated with miscarriage risk in a dose-dependent pattern, which is what you would expect if the association tracks exposure rather than something incidental to who works nights. Category-level hazard ratios for that cumulative dose response are not stated in the abstract and the full text sits behind a paywall, so no per-band figure can be quoted here. Anyone citing specific dose-response numbers for this study should be asked where they read them.

One practical gap matters for policy writing. The paper's operational definition of a night shift could not be confirmed from the primary source, and secondary descriptions disagree about the clock window used. If you plan to write a weekly night-shift cap into a pregnancy accommodation policy, define your own window explicitly rather than implying it came from this study. Two night shifts per week is where the published evidence starts showing a signal, not a validated safety limit and not a regulatory line.

Payroll-recorded exposure was a genuine advance, and it does nothing for the other half of the measurement problem. Miscarriages in Begtrup et al. (2019) were identified from hospital admissions, so losses managed in primary care, or never presenting clinically at all, are absent from the count, and the losses least likely to reach a hospital are the earliest ones. Restricting the headline result to losses after pregnancy week 8 is the authors' response to that gap rather than a claim that early pregnancy is unaffected.

Residual confounding is the second live objection. A payroll database records when a shift happened, not what happened during it, so physical workload such as lifting or time spent on your feet sits outside the exposure measure, and in a hospital that workload plausibly travels with night rostering. Women who worked nights also differed from those who did not in socioeconomic status, occupation, number of previous pregnancies, and number of previous miscarriages. Any of those could shape the result, and adjustment can only handle what was measured.

Selection into and out of night work is the third. Exposure was time-varying, so a woman counts as exposed only in the weeks her payroll record shows night shifts, and anyone moved off nights during pregnancy stops contributing exposed weeks from that point onward. Whoever is still rostered on nights late in follow-up is therefore not a random subset of those who started, and healthy-worker effects and reverse causation are ordinary hazards in this literature rather than exotic ones. It is consistent with all of this that the authors write only that night work may confer an increased risk.

This is not a sample of shift workers in general. It is Danish women employed in public hospitals across the five Danish administrative regions, whose hours are captured in a national payroll database. Danish labor protections, staffing norms, parental-leave rules, and healthcare access differ from most other markets, and hospital night work differs from night work in retail, hospitality, manufacturing, or logistics in workload, autonomy, and how quickly a schedule change can be arranged. Applying the 32% figure to those settings is an extrapolation, not a finding.

The vintage matters too. The paper was published in 2019 on payroll data from the 2007 to 2015 era, and it was funded by the Danish Working Environment Research Fund and NordForsk (grant 74809). Its payroll-recorded shift counts remain unusual in this literature, where exposure is more often reconstructed from recall or from job titles, which is worth saying plainly, but it predates the changes in healthcare staffing patterns that followed 2020. Treat it as the best available measurement of a real association in one well-documented workforce, and expect the next good study to refine rather than confirm it.