Study design

Register-based historical cohort study.

Participants

All women born between 1957–1997 and living in Denmark between 1977–2017 were eligible for inclusion in the study cohort and identified in the Danish Civil Registration System [20]. The unique personal identification number provided to each resident enabled linkage to other national registers. Immigrants were eligible for inclusion if immigration occurred before age 20. Persons with an outcome of interest before age 12 were excluded. Follow-up commenced at age 12, immigration, or start of follow-up (January 1, 1977), whichever came last. Censoring occurred due to emigration, death, end of follow-up (December 31, 2017), or an event of interest, whichever came first. The study was approved by the Danish Health Data Authority. No informed consent is need for register-based studies in Denmark. A complete list of definitions and registers used can be found in the S1 Table.

Reproductive history

In the Danish Medical Birth Register [21], the number of live births and stillbirths were identified. The lower gestational threshold for stillbirths in this register was 28 gestational weeks before 2004 and 22 gestational weeks after. Therefore, pregnancies ending at earlier gestational ages were identified in the Danish National Patient Register [22]. A pregnancy loss was defined as the registered loss of a pregnancy before viability and included miscarriage, blighted ovum (ie. a pregnancy with a visible gestational sac, but no visible embryo), and missed abortion (ie. a nonviable pregnancy remaining in the uterus). Recurrent pregnancy loss was defined as three consecutive pregnancy losses without a delivery or induced abortion in between or a specific diagnosis of recurrent pregnancy loss. Recurrent pregnancy loss was subdivided by whether the sequence was preceded by a delivery (secondary) or not (primary). As complications of early pregnancy could lead to multiple hospital contacts during clinical care, an algorithm using restriction periods was used (S1 Appendix) to ascertain each pregnancy was only counted once. Therefore, additional pregnancy outcomes were also identified: ectopic pregnancy, molar pregnancy, and induced abortion, to separate individual pregnancies. The latter were identified in the Danish Register of Legally Induced Abortions [23].

Exposure

Pregnancy loss was the exposure of interest and defined hierarchically with no pregnancy losses at the lowest level and recurrent pregnancy loss at the highest level, as these women were a priori hypothesized to have a higher probability of underlying immune disorder. Women could increase exposure level during follow-up but not revert, and could only contribute risk time in one category at a given time during follow-up (categories: no pregnancy losses, 1 pregnancy loss, 2 pregnancy losses, ≥3 non-consecutive pregnancy losses [not fulfilling criteria for recurrent pregnancy loss], primary recurrent pregnancy loss, and secondary recurrent pregnancy loss).

Outcome

The primary outcome of interest was an incident diagnosis of multiple sclerosis in the Danish Multiple Sclerosis Register [24]. In case the date of diagnosis or date of symptom onset was given as the calendar year, June 1 was chosen as the date of diagnosis or symptom onset. Kinship information was available since 1968 in the Civil Registration System. As the Danish Multiple Sclerosis Register was only available for women, multiple sclerosis among first degree relatives (mother, listed father, or full sibling) was identifiable since 1977 in the Danish National Patient Register using the International Classification of Diseases and Related Health Problems, 8th revision (ICD-8), and 10^th revision (ICD-10) codes for multiple sclerosis: ICD-8: 340, ICD-10: G35. Additional outcomes examined were amyotrophic lateral sclerosis, Guillain-Barré syndrome, and myasthenia gravis, all identified in the Danish National Patient Register.

Covariates

The study included the following potential confounders in the main analysis: (i) the number of live births (categories: 0, 1, 2, and ≥3), (ii) obtainment of a bachelor’s degree or higher education (categories: yes, no, or unknown), (iii) family history of multiple sclerosis, defined as mother, listed father, or full-sibling diagnosed with multiple sclerosis (categories: yes, no, or unknown), (iv) calendar period (categories: 1977–1989, 1990–1999, 2000–2009, and 2010–2017), and (v) maternal age (categories: 12–19, 20–24, 25–29, 30–34, 35–39, 40–44, 45–49, and 50–61 years). Persons with missing data were grouped in a separate ‘unknown’ category in analyses.

Statistical analyses

Follow-up was conducted in a time-dependent manner and risk time in years with three decimal places, was split at changes in age, other covariates, or exposure status. The number of events and person-years was summarized, and crude incidence rates per 10,000 person-years were calculated. The primary model estimated crude and adjusted incidence rate ratios (IRR) using Bayesian Poisson regression. A sensitivity analysis used a negative binomial model to better account for potential overdispersion. The Bayesian framework was chosen as it provided an intuitive interpretation of uncertainty in estimates and modeled the complete posterior distribution based on the statistical model, data, and prior beliefs. The chosen priors in the main analysis were modestly informative. A sensitivity analysis used minimally informative uniform priors, approximating estimates which would be obtained by a corresponding frequentist model. The IRR and 95% credible interval (CI) were extracted from the posterior distribution. Convergence of chains was confirmed for all analyses [25]. For detailed specifications of the statistical analyses including chosen priors, see the S2 Appendix.

Sensitivity analyses

To assess the robustness of the findings, we performed a series of sensitivity analyses. Due to the risk of type I error due to multiple comparisons, these should be interpreted as exploratory. First, as patients diagnosed with multiple sclerosis may have had a prolonged period of prodromal symptoms, a secondary analysis used the date of symptom onset as the event date. The date of symptom of onset was assessed and registered by the clinician upon diagnosing multiple sclerosis. Second, as women who never achieve pregnancy may include a higher proportion of women with underlying medical conditions, these women were excluded. Third, as smoking status (categories: yes/no) is associated with multiple sclerosis, we conducted an analysis including this covariate [26]. Smoking status has in Denmark routinely been assessed for live births and stillbirths since 1997. This analysis only included women with non-missing data in these variables (i.e., complete case analysis) and women were included at the delivery where their smoking status was first registered. The last known values for these variables were used until a newer was available. Fourth, as treatment of pregnancy loss was mainly provided in a hospital setting in Denmark before the year 2000, we conducted an analysis concluding the study this year to minimize exposure misclassification. Fifth, exposure to stillbirth (categories: 0, ≥1) was examined, as stillbirth has been associated with other diseases such as cardiovascular disease [27]. Sixth, an analysis used minimal-informative priors. Finally, we conducted an analysis using a negative binomial model where the covariate calendar year was modeled to have a linear effect, and the covariate age was modeled using a cubic spline with five knots. The knots were chosen as the quartiles of the variable in the full dataset. This analysis improved modelling in case of overdispersion and non-linear effects of age on the outcome.

Analyses were conducted using R version 4.0.3 [28], and Bayesian models were fit using Stan version 2.21.2 [29] using the R interface package rstanarm. A two-sided 95% CI that did not overlap 1.00 was considered statistically significant. The study was reported using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [30].

Limitations

This study has several limitations. First, as multiple sclerosis is often not diagnosed until years after initial symptoms, the study could have been susceptible to reverse causality. To further explore this, a sensitivity analysis was conducted using the date of symptom onset as the event date. This did not change the findings materially.

Second, women who never achieve pregnancy may comprise a heterogenous group not necessarily comparable with other women. To increase the homogeneity of the comparator group, women who were never registered pregnant were excluded in a sensitivity analysis. Again, this did not change the significance of the results. Third, although the primary analysis adjusted for many important confounders, residual confounding could not be ruled out. Therefore, an analysis further adjusted for smoking status as a proxy for unmeasured lifestyle factors, and the resulting estimates were materially unchanged. Smoking status was only known from year 1997 an onwards for women with a delivery (ie. live- or stillbirth), and consequently cannot necessarily be extrapolated to women with no prior deliveries. Therefore, a post-hoc sensitivity analysis (S3 Table), explored the effect of identifying smokers by hospital diagnosis code (ICD-10: F17) or fulfilling a prescription for a smoking cessation drug (Anatomical Therapeutic Chemical Classification code: N07BA), this did not change the significance of the results. Fourth, as the exposure only included clinical pregnancy losses, that is, those treated in a hospital setting and not including those only treated by private practitioners, misclassification of the exposure may have biased results. Assuming non-differential misclassification, this would bias our results towards the null. Until the year 2000, pregnancy losses in Denmark were routinely evaluated in a hospital setting [9]; therefore, a secondary analysis evaluated the effect of ending the study on December 31, 1999. Although the statistical power was reduced, this did not change the significance of the results. Further, a study found that a diagnosis of miscarriage in the National Patient Register had a high validity as the diagnosis was confirmed in 114 out of 117 hospital records [36].

Fifth, conducting Bayesian models facilitates and necessitates incorporating prior evidence and knowledge into the probability of an outcome. However, in the current study, only sparse evidence existed to guide the model, and the prior was used to guide posterior estimates into a plausible range. To further investigate the effect of this, a secondary model assessed the effect of minimally informative uniform priors on estimates. This analysis only changed the results minimally due to the abundant data guiding the model. Fifth, a sensitivity analysis was designed to better handle potential overdispersion and non-linear effects of the covariate age. This analysis did not change results substantially. Misclassification of the outcome was assumed to be minimal due to the high completeness (91%) and validity (94%) of the Danish Multiple Sclerosis Registry. Additional outcomes were based on diagnosis codes in the Danish Patient Register which have shown positive predictive values ranging from 83.8 to 92.5% [24, 37–39]. Family linkage was based on the Civil Registration System and has to our knowledge not been validated. Correct assignment of the mothers is likely very accurate as is occurs right after birth. Family history of multiple sclerosis was based on diagnosis codes in the Danish National Patient Register which have been validated against the Danish Multiple Sclerosis Registry and shown a completeness of 92.8% and validity of 95.1% [40]. In the cohort missing values were rare: 9.7% had missing information on one or both parents leading to unknown family history of an outcome, and 3.7% had missing information on educational status. We acknowledge this could be a potential source of bias.

Although the population was large, some exposure groups and outcomes were rare. Therefore, the possibility of insufficient power to accurately detect a small increase in risk cannot be excluded, and future studies may well aim to reproduce estimates for the outcomes of multiple sclerosis and Guillain-Barré syndrome after exposure to recurrent pregnancy loss.