Data sources
We linked data from four population-based nationwide registries in Denmark: Danish Civil Registration System [17], Danish Medical Birth Registry [18], Danish National Patient Registry [19], and Danish National Prescription Registry [20]. Additional file 1: Table S1 provides a detailed description of all data sources, including specific types of data originating from each.
Study design, population, and period
We included all pregnancies in Denmark that started and ended between 01 January 2002 and 31 December 2013. Pregnancies ending in a live birth or a stillbirth (≥22 gestational weeks) were identified in the Danish Medical Birth Registry. Pregnancies ending earlier than 22 gestational weeks in abortive outcomes were identified from hospital diagnoses recorded in the Danish National Patient Registry. Starting in 2007, the Danish National Patient Registry had information on congenital malformations identified during second-trimester therapeutic pregnancy terminations.
Exposure
The Danish National Prescription Registry provided information on dispensings for oseltamivir at outpatient (community) pharmacies. The following mutually exclusive categories of oseltamivir exposure during pregnancy were defined: exposure during the first trimester regardless of exposure in the second or third trimester; exposure during the second or the third trimester but not in the first trimester; and no exposure at any time during pregnancy (the reference category). Because organogenesis is complete in the first trimester, we examined association of first-trimester oseltamivir exposure with congenital malformations. For the remaining birth outcomes, oseltamivir exposure at any trimester was considered.
Outcomes
Congenital malformations, identified from diagnoses recorded at therapeutic second trimester abortions (2007–2013), at stillbirth, and up to 1 year postnatally in liveborn infants, were classified according to the major EUROCAT categories [3]. The Danish National Patient Registry is nearly 99% complete for diagnoses of congenital malformations [21]. For congenital heart defects, we used an algorithm developed specifically for the Danish National Patient Registry, based on the EUROCAT-specified diagnostic codes combined with therapeutic cardiac procedures [22]. The positive predictive value of this algorithm, estimated on a random sample of cases observed in this study, was 94.6% (95% confidence interval 89.2% to 97.7%). The other pregnancy outcomes were stillbirth at ≥22 weeks of gestation; foetal death (spontaneous or induced abortion before 22 weeks of gestation); preterm birth (gestational age 22- < 37 weeks) among live and stillbirths; small for gestational age (SGA) (birth weight below 10th percentile of the sex- and gestational-week-specific weight distribution) among live and stillbirths; and low 5-min Apgar score (< 7) among live births. For non-singleton pregnancies, a given outcome was considered present if recorded in at least one foetus/newborn.
Covariates
We assessed the following covariates based on their known associations with the birth outcomes: maternal age at conception, calendar year of conception, smoking as reported at the first prenatal visit (for live and stillbirths); pre-pregnancy body mass index (BMI) for live and stillbirths; mode of delivery; parity; marital status; birth of a previous child with a malformation (since 1994); indicators of maternal health care utilization (hospitalizations, visits to hospital outpatient specialist clinics, emergency room visits, dispensings for specific drug classes); maternal inpatient or outpatient morbidity (respiratory disease, cardiovascular disease, haematological disease, diabetes, neurological disease, liver or kidney disease, rheumatic disease, inflammatory bowel disease, obesity, immunodeficiency, disorders of female pelvic organs/genital tract, hospital contact for injury or poisoning); maternal outpatient dispensings for antidepressants, antiepileptics, antidiabetics, antihypertensives, drugs for ulcer/gastroesophageal reflux, oral contraceptives, drugs for in-vitro fertilization, thyroid hormones, systemic corticosteroids, non-steroidal anti-inflammatory drugs, opiates, and systemic anti-infective agents other than oseltamivir. Data on all diagnoses originated from inpatient or outpatient hospital diagnoses (secondary care), while data on medication dispensings originated from primary care and outpatient prescribing. The covariates were ascertained during 12 months preconception. Information on influenza status was not available from any data source. Definitions of the study variables appear in Additional file 1: Table S2 and Table S3.
Statistical analyses
We described the distributions of the pregnancy characteristics according to exposure to oseltamivir using appropriate descriptive statistics. For all outcomes except spontaneous or induced abortions, prevalence was used as the measure of occurrence. Crude and adjusted odds ratios (ORs) were computed using logistic regression. Pregnancies that ended before the second trimester were excluded from the analyses of second/third trimester exposure. For abortions, incidence rate was used, with hazard ratios estimated via Cox’s proportional-hazards regression, with oseltamivir exposure treated as a time-varying variable [23]. All estimates were reported with 95% confidence intervals (CIs).
Confounding was addressed using two approaches: propensity score matching and conventional adjustment using multivariate regression with generalised estimating equations to account for within-woman correlation. Propensity score matching was considered superior in control of measured confounding, while conventional adjustment allowed use of all available observations and provided the context to evaluate the direction of estimates’ change in response to tighter confounding control [24].
A propensity score for each pregnancy was computed, using logistic regression, as the probability of an oseltamivir dispensing given the covariates. Separate propensity scores were computed for the first-trimester and for the second/third-trimester exposure. Unexposed pregnancies were matched to exposed pregnancies on propensity score using nearest-neighbour matching with a caliper width of 0.2 standard deviations of the logit of the propensity score [25]. The balance of baseline characteristics was assessed post-matching, using standardised mean differences, whereby a value of ≤0.1 was considered indicative of balance. Per protocol, up to 100 oseltamivir-unexposed pregnancies were planned to be matched to each oseltamivir-exposed pregnancy. Post-matching assessment of the resulting balance indicated that only 1:1 matching achieved the target covariate balance. This 1:1 matched sample was used in propensity-score analysis, as it was deemed to remove most of the measured confounding. The covariates included in the propensity scores and the balancing statistics before and after matching are described in Additional file 1: Tables S4–S5, and Figure S1.
In conventionally adjusted analyses, we included all covariates with prevalence ≥5% or those inducing a >10% change in the crude OR. The final model included binary variables for parity (0 vs. > 0); marital status; smoking; obesity (BMI ≥30 kg/m2 or a hospital-based diagnosis of obesity); any chronic illness (cardiovascular disease, haematological disease, diabetes, neurological disease, liver or kidney disease, rheumatic disease or inflammatory bowel disease); and respiratory disease. In addition, all models included variables for mother’s age at conception (as a cubic spline) and for prior delivery of a child with a malformation. Smoking is not recorded for pregnancies ending in abortive outcomes; therefore the sensitivity analyses that contained such pregnancies were not adjusted for smoking.
The main analyses were conducted based on pregnancies ending in a live or stillbirth using propensity score to control for confounding. Since confounding by indication was expected to persist in this setting, several prespecified and post hoc sensitivity analyses were conducted for the malformation outcomes to obtain indirect evidence on confounding extent. First, we repeated the main analyses while including malformation diagnoses from terminated pregnancies (for pregnancies in 2007–2013). Second, we excluded mothers with a prior delivery of a child with a malformation. Third, we assessed risks of malformations associated with dispensing for oseltamivir during the main organogenesis period (gestational weeks 4–10 [26]). Fourth, we conducted several ‘negative control’ analyses [27]: examining effects of oseltamivir dispensing during periods etiologically implausible with respect to inducing major malformations (12 to 3 months preconception; second/third trimester of pregnancy). Fifth, we repeated the analysis replacing first-trimester exposure to oseltamivir with first-trimester exposure to penicillin, which is an anti-infective agent without evidence of teratogenicity [28] but presumed to correlate with presence of an infectious process, including fever. Finally, we examined the distribution of specific types of congenital heart defects for potential clustering, as clustering would support a causal association.
The analyses were conducted using SAS®, version 9.4 (Cary, NC, USA). Results were presented only when the individual cell counts in tables exceeded 5 observations, as specified by the Danish Data Protection Law (www.datatilsynet.dk) and/or regulations of Statistics Denmark (www.dst.dk).