Data sources
CPRD GOLD is a UK-based repository of anonymised general practice data and makes up part of one of the largest resources of primary care data in the world.32 It covers approximately 7% of the UK’s population, representative by age, sex, and ethnicty.32 The primary care data in CPRD GOLD are linked to practice-level Index of Multiple Deprivation (IMD) scores and Hospital Episode Statistics Admitted Patient Care for most English practices.32
The CPRD GOLD Pregnancy Register is described in detail elsewhere;33 in short, it contains pregnancy episodes with estimated pregnancy dates and outcomes linked to patient identifiers in CPRD GOLD (see Supplementary Information S1). In the current study, the Pregnancy Register was prepared in accordance with recommendations from the authors of the register algorithm.34
Pregnancy episodes with an uncertain outcome were recoded where Hospital Episode Statistics data were available, and dates were updated using imputed values as per the Pregnancy Register algorithm, whereby a set gestational length is used depending on the updated outcome (see Supplementary Information S2). Pregnancy episodes ending in ‘unknown outcome’ that were not recoverable using Hospital Episode Statistics were excluded.
Antidepressants
All antidepressants approved for treating depression in the UK were identified and stratified by class (see Supplementary Table S1). Briefly, prescription end date was used in conjunction with the pregnancy start date and the end date of trimester one to identify whether an antidepressant prescription occurred within, or overlapped with, the first trimester to identify those ‘exposed’ to antidepressants (see Supplementary Information S3).
Dose was standardised for each medication using the distribution of dose in milligrams; low (≤25th percentile), medium, and high (>75th percentile) doses (see Supplementary Information S3). In instances where multiple doses were prescribed in trimester one, individuals were classified with the highest dose they received in the first trimester.
Confounders
Confounders were chosen a priori based on the literature. The primary adjustment set contained: age; year of pregnancy; IMD quintile (as a proxy for practice-level socioeconomic deprivation); history of miscarriage and severe mental illness; smoking; parity; use of high-dose folic acid, antipsychotics, and antiseizure medications; number of primary care consultations in the 12 months before pregnancy; and whether an individual has ever been diagnosed with depression and anxiety before the start of pregnancy (see Supplementary Information S4, Supplementary Table S2, and Supplementary Figure S1).
Depression and anxiety were identified using pre-defined, expert-verified codelists in primary care (Read codes) and Hospital Episode Statistics Admitted Patient Care (International Classification of Diseases, 10th Revision codes).
Ethnicity35 and body mass index (BMI) around the start of pregnancy contained >10% missing data, thus were dropped from the primary adjustment set and included in sensitivity analysis.
Analysis
Baseline characteristics of the eligible sample are described by first trimester antidepressant use. All analyses were performed using complete records for covariates.
First, those prescribed antidepressants in trimester one were compared with those who were not, using Cox proportional-hazards models. ‘Incident’ users (those who were not prescribed antidepressants in the 3 months before pregnancy but were in trimester one) contributed non-prescribed time to the analysis until the start of their antidepressant prescription; ‘prevalent’ users (those prescribed in the 3 months before pregnancy and into trimester one) only contributed exposed time to the models. Censoring occurred at the earliest of other loss (see Supplementary Information S5), reaching 24 weeks’ gestation, or study end (31 December 2018). Cluster-robust standard errors (clustered by pregnant individual) were employed to account for those who contributed multiple pregnancies to the analysis.
To enhance clinical interpretability, the absolute confounder-adjusted risks (1 minus survival) were estimated using Breslow’s baseline estimator and these were integrated with the hazard ratio (HR) through the G-formula (that estimates the average outcome that would be seen if everyone took antidepressants during trimester one) and bootstrapping for standard errors (1000 repetitions).
The model was run restricted to those with evidence of depression or anxiety in the 12 months before pregnancy and to those with ‘severe’ depression or anxiety, as defined by administered scale standardised scores (like the nine-item Patient Health Questionnaire, Supplementary Information S6) in the 12 months before pregnancy. In addition, receipt of an antidepressant prescription was compared with none in trimester one among those who were prescribed antidepressants in the 3 months before pregnancy.
In an exposure-discordant pregnancy analysis, pregnancies in the same individual were compared. This approach accounts for time-stable confounders, like genetic liability to miscarriage, by design because in a single individual time-stable confounders have the same effect on all pregnancies and therefore are ‘adjusted’ for in the analysis.36,37 A stratified Cox model adjusted for the primary adjustment set (except history of miscarriage) was used, where each stratum in the model represented an individual with ≥2 exposure-discordant pregnancies (see Supplementary Information S6).
Propensity-score matching was also performed, following the stepwise process laid out by Desai and Franklin.38 The propensity score included both confounders and predictors of the outcome (see Supplementary Table S2)39 and was restricted to first pregnancies. The propensity score was estimated using logistic regression, then the final iteration of balancing criteria were applied: exposed and unexposed pregnancies were matched 1:1 without replacement using a caliper of 0.2, and exact matching on number of primary care visits before pregnancy. The Love plot representing the balance achieved by the above criteria can be found in Supplementary Information S6.3 and Supplementary Figures S2–S7.
For the secondary analyses, ‘prevalent’ (≥1 prescription for antidepressants in the 3 months before and during trimester one) and ‘incident’ (≥3 months clear of antidepressant prescriptions before pregnancy; ≥1 prescription during trimester one) antidepressant users were compared with the unexposed group. Analysis was also restricted to those with any depression or anxiety, as well as ‘severe’ illness before pregnancy (see Supplementary Table S3). Individual antidepressant class was compared with the unexposed group. In addition, low, medium, and high doses of antidepressant in trimester one were compared with the unexposed group (see Supplementary Information S3).
In sensitivity analysis, all above analyses were restricted to those with Hospital Episode Statistics data, owing to pregnancy outcome modifications (see Supplementary Information S2). The primary Cox model where exposure was redefined as ≥2 antidepressant prescriptions in trimester one was performed to limit exposure misclassification. The potential for differential pregnancy exclusion, potential bias in the complete records analysis,40 and potential bias introduced by competing events (that is, other early pregnancy losses) were investigated.
All analyses were performed in Stata (version 17.0) and R (version 4.3.1).
