Antidepressants in pregnancy: applying causal epidemiological methods to understand service-use outcomes in women and long-term neurodevelopmental outcomes in exposed children

Identification of pregnancies and linkage of women to offspring

A validated set of algorithms that identify pregnancies within the CPRD is now integrated within the CPRD as a pregnancy register. 58 This register enables the identification of the dates, stages and outcomes of pregnancies within the CPRD. 58 The CPRD has also developed a probabilistic mother–baby link, which allows the patient identifier numbers of mothers and the patient identifier numbers of their live-born offspring to be linked, enabling the construction of an intergenerational cohort. 59

Linkage of Clinical Practice Research Datalink data with other resources

For consenting CPRD practices in England, which represent approximately 60% of patients in our data, it was possible to link the anonymised primary care records with other data sources. These sources include the Hospital Episode Statistics (HES), which has separate registers for inpatient admissions, outpatient care and accident and emergency (A&E) attendance in England. Linkage with the Office for National Statistics (ONS) mortality data and the Census small-area socioeconomic data was also available for this same subset.

The use of CPRD data for this project was approved by the CPRD’s Independent Scientific Advisory Committee (reference 17_225).

Pregnant women’s cohort

Figure 1 shows the derivation of the pregnant women’s cohort used for the main analysis.

FIGURE 1.

Derivation of the pregnant women’s cohort.

From the eligible sample of 344,720 pregnancies, we excluded:

• records where the general practice was not yet up to standard (UTS) (n = 61,704). CPRD defines practices as being ‘up to standard’ if they have provided data on a consistent basis

• records where the patient had not yet registered with her current general practice 1 year prior to the date of conception, as recorded in the pregnancy register (n = 93,638)

• records suggesting that the woman had transferred out of the general practice while pregnant (n = 15,627)

• records with less than 2 years’ follow-up beyond the pregnancy end date (n = 18,569)

• records that showed overlap with a preceding or successive pregnancy episode (n = 23,691)

• records of any successive pregnancy episode that started less than 4 years after a prior episode had ended to minimise biased results arising owing to the possibility of women being pregnant again or trying to conceive during follow-up (n = 32,930).

We set aside pregnancies for which antidepressants had been prescribed for indications other than depression (n = 18,458; these were used in additional analyses described in Variation by indication: depression compared with other indication for antidepressant prescribing); therefore, 80,103 pregnancies were included to study women’s primary care service use outcomes. Of these pregnancies, 45,358 were eligible for record linkage to study secondary care service outcomes. Among these, data on inpatient admission were available for pregnancies that had started on or after 1 April 1997 (n = 43,662); outpatient treatment data were available for pregnancies starting on or after 1 April 2003 (n = 35,674); and A&E attendance data were available for pregnancies starting on or after 1 April 2007 (n = 25,697).

Mother and child cohort

Figure 2 shows the derivation of the mother and child cohort.

FIGURE 2.

Derivation of the mother and child cohort.

From the eligible sample of 344,720 pregnancies, we excluded:

• records where the general practice was not yet UTS, as defined by CPRD (n = 61,704)

• records where the patient had not yet registered with her current general practice 1 year prior to conception (n = 93,638)

• records where the patient had transferred out of the general practice while still pregnant (n = 15,627)

• records that showed overlap with a preceding or successive pregnancy (n = 26,357)

• pregnancies not recorded to have resulted in a live birth (n = 72,565)

• live deliveries that could not be linked with offspring patient records (n = 15,298)

• pregnancies that were recorded to have lasted less than 22 gestational weeks (n = 542)

• any offspring who transferred out of their general practice (n = 10,404) or died (n = 5) before the age of 4 years.

Given that the CPRD pregnancy register includes only the first child in cases of multiple deliveries, we identified an additional 546 children with at least 4 years’ follow-up in the mother–baby link data set, by matching the mother’s patient identification number and exact date of delivery. Setting aside mothers who were likely to have been prescribed antidepressants for indications other than depression (n = 8485) and children followed up for less than 4 years owing to being born after 2013 (n = 6367), we were able to include 34,274 children in the offspring cohort (mean age at end of follow-up 10.04 years, range 4–22 years).

General practitioner consultations

During pregnancy and in each of the four consecutive 6-month follow-up periods, we counted the number of days on which women had consulted with their general practice. General practitioner (GP) consultations were required to have been face to face or by telephone and the staff member was required to be a doctor, nurse (including community psychiatric nurse) or psychologist. Report Supplementary Material 4 contains the operational definitions used in deriving this outcome. Following the same definitions, we counted the number of days on which women consulted with their general practice specifically for further episodes of depression or self-harm (including suicide attempts) during pregnancy and follow-up periods. Symptoms of depression and self-harm were identified in the consultation records using validated Read code lists (see Report Supplementary Materials 1 and 5) where the consultation type equalled ‘symptom’, ‘examination’, ‘diagnosis’, ‘administration’ or ‘presenting complaint’.

Referrals made by the general practitioner

We constructed a set of binary variables to indicate whether women had been referred by their GP to secondary services for depression or self-harm/suicide attempts during pregnancy or in each of the four consecutive 6-month follow-up periods. Referrals were defined as the presence of a referral record with a medical code for depression or self-harm/suicide attempt, where the NHS referral specialty classification equalled ‘mental handicap’, ‘mental illness’, ‘child and adolescent psychiatry’, ‘forensic psychiatry’, ‘psychotherapy’, ‘old age psychiatry’, ‘clinical psychology’, ‘learning disabilities’, ‘adult psychiatry’ or ‘community psychiatric nurse’, or where the Family Health Services Authority referral classification equalled ‘psychiatry’.

Inpatient admissions

Using linked HES data, we constructed a set of binary variables to indicate whether women had been admitted as an inpatient for a mental health problem [International Classification of Diseases, 10th Edition (ICD-10): F00–F99] or for intentional self-harm/suicide attempt or self-harm of undetermined intent (ICD-10: X60–X84 and Y10–Y34) during pregnancy or in each of four consecutive 6-month follow-up periods. These inpatient admissions were defined by a record where the main specialisation of the consultant equalled ‘accidents and emergency’, ‘learning disability’, ‘adult mental illness’, ‘forensic psychiatry’ or ‘psychotherapy’; where their treatment specialisation equalled ‘accidents and emergency’, ‘clinical psychology’, ‘learning disability’, ‘forensic psychiatry’, ‘psychotherapy’, ‘eating disorders’, ‘liaison psychiatry’, ‘perinatal psychiatry’, ‘mental health recovery and rehabilitation service’ or ‘mental health dual diagnosis service’; and where the method of admission equalled ‘A&E or dental casualty department’, ‘request for immediate admission by GP’, ‘consultant clinic’, ‘admission via mental health crisis resolution team’ or ‘other means’.

Outpatient treatment

Using linked HES data, we constructed a set of binary variables to indicate whether women had used outpatient services for a mental health problem during pregnancy or in each of four consecutive 6-month follow-up periods. Outpatient contacts were defined by outpatient records where the treatment specialty of the consultant equalled ‘adult mental illness’, ‘child and adolescent psychiatry’, ‘forensic psychiatry’ or ‘psychotherapy’.

Accident and emergency department attendance

Using linked HES data, we counted the number of instances women had presented to A&E services during pregnancy or in each of four consecutive 6-month follow-up periods. We considered only first A&E attendances (excluding any planned or unplanned follow-up for a prior attendance) and excluded attendances for assault, sports injuries or firework injuries, or where the patient had died on arrival to the A&E unit.

Autism spectrum disorder

We examined the primary care clinical and referral records of linked offspring for the presence of autism spectrum disorder (referred to hereafter as autism) using a validated Read code list (see Report Supplementary Material 6). Offspring were considered positive on outcome if they had a primary care record that indicated autism, autism spectrum disorder, autistic disorder, Asperger syndrome, atypical autism, childhood autism, infantile autism, autistic psychopathy or pervasive developmental disorder, and if they had a record of autism when they were at least 4 years of age. A recent study60 validated the CPRD diagnosis, as recorded in the CPRD, against the clinical records for a subsample and reported a positive predictive value of 91.4%. Given that the HES data were available for only a subsample, and registered diagnoses are recorded in less than 5% of all outpatient attendances (during which most autism-related consultations would happen), we did not use linked data to supplement the autism diagnoses. 61

Intellectual disability

Following the same approach, we examined the primary care records of linked offspring for presence of intellectual disability that had been diagnosed when the child was at least 4 years of age. Children were considered positive on outcome if their primary care records indicated Read codes related to intellectual disability (see Report Supplementary Material 7 for the list of Read codes). These codes were similar to those used by previous studies of intellectual disability within CPRD,62,63 although we did not include codes for autism (i.e. someone with a code of autism would be counted as having an intellectual disability only if there were additional codes related to intellectual disability in their medical record). Similar to autism, linked data were not used to supplement these diagnoses because these data were available for only a subset and the HES outpatient registers had less than 5% recording of diagnostic data. 61

Attention deficit hyperactivity disorder

Primary care diagnoses of ADHD were identified by the presence of medical codes pertaining to ADHD or therapy records that indicated that the child had been prescribed ADHD medication when they were at least 4 years of age. Children were considered positive on outcome if they had a recorded Read code related to ADHD (see Report Supplementary Material 8) or if they had been prescribed any of the following ADHD medications: methylphenidate, dexamphetamine, atomoxetine, dextroamphetamine, amphetamine with dexamphetamine, or lisdexamphetamine (see Report Supplementary Material 9). Similar methods have been used to identify ADHD in previous CPRD studies. 64 As above, linked data were not used to supplement these diagnoses because these data were available for only a subset and the HES outpatient registers had incomplete recording of diagnostic data. 61

Multivariable regression

We used multivariable regression to estimate the maternal and child outcomes associated with initiating (described further in Chapter 3) or continuing (described further in Chapter 4) an antidepressant into pregnancy. For each outcome, we first estimated crude associations and then controlled statistically for the range of potential confounders described in Covariates. Further detail on the selection and specification of multivariable regression models is provided in Chapters 3 and 4.

Propensity score-matched regression

Alongside conventional multivariable regression, we carried out all analyses in subsets of the data for which we matched treatment groups on propensity scores for initiation and continuation of antidepressants during pregnancy. Propensity score matching (PSM) is a commonly used method in pharmaco-epidemiology that allows the identification of pairs of observations that are similar in all measured characteristics, except for treatment status. 68 It, therefore, aims to achieve balanced treatment groups, allowing for a like-with-like comparison that would be achieved by randomisation in RCTs. It has been argued that PSM may provide a more effective approach to minimising confounding bias than traditional multivariable regression methods because it can incorporate large numbers of potential covariates that may overwhelm traditional regression models. 69 However, the main constraint of PSM is that the groups can be matched only on characteristics that are measured, so confounding by unmeasured characteristics is still possible. 32 Furthermore, in analyses using PSM, individuals who cannot be matched for being too dissimilar are excluded from the analysis, which can affect statistical power because of reduced numbers.

In this study, we estimated propensity scores using classification and regression tree (CART) models69 separately for our two comparisons to match mothers who initiated antidepressants with mothers who received no treatment (see Chapter 3) and to match mothers who continued antidepressants into pregnancy with mothers who discontinued antidepressants before pregnancy (see Chapter 4). Further details are provided in the respective chapters.

Instrumental variable analysis

Instrumental variable regression is a statistical technique that can allow the estimation of causal effects in the presence of unmeasured confounding. 70 This is where unobserved characteristics of patients influence their likelihood of being prescribed antidepressants and at the same time influence risk of outcome, resulting in a confounded treatment effect. The rationale for IV analysis, in this particular context, is that the clinical decision to prescribe an antidepressant in pregnancy can be viewed as being influenced by three factors: first, whether or not the GP deems it safe to prescribe antidepressants to a pregnant patient given potential concerns about teratogenicity; second, the characteristics of the patients themselves, for instance their clinical characteristics, including severity of depression during pregnancy; and, third, the propensity of the physician to prescribe antidepressants. Using a well-specified IV, we can, therefore, distinguish between variability in treatment decisions owing patient characteristics (which may confound the treatment effect) and variability in treatment decisions as a result of whether or not GPs are willing to prescribe antidepressants in pregnancy (which is not determined by the characteristics of their current patient). For this reason, IV analysis can overcome unmeasured treatment-outcome confounding and, therefore, identify the causal effect of treatment on outcome. Broadly following methods proposed in earlier work,71 we aimed to capture as an IV the GPs previous prescribing practice of antidepressants in a pregnant patient given potential concerns around risks. Given that a GP’s views on medication safety cannot be directly observed, we proxied this by the number of times that they had issued an antidepressant in prior consultations with other pregnant patients. The validity of the result then depends on the extent to which the following assumptions are tenable: first, the instrument associates with the treatment (relevance assumption); second, the IV should influence only the outcome through the treatment variable (the exclusion restriction); and third, the IV does not share a common cause with the outcome (i.e. there are no confounders of the instrument outcome relationship) (the independence assumption). 70 If these IV assumptions are met, IV analysis can estimate the causal effect of treatment on an outcome.

We used IV analysis separately for women who initiated an antidepressant or continued with an existing prescription into pregnancy. Further details are provided in Chapter 5.

Matched treatment-discordance designs

Another approach to account for unmeasured confounding is the matched treatment-discordance design. This design is also commonly referred to as a sibling design when the matching is based on siblings to study outcomes in offspring of treatment or exposure discordant pregnancies. 72,73 We use the term treatment-discordance design because we have used this approach to study women’s outcomes across pregnancies, as well as outcomes for the offspring across pregnancies.

In this design, we consider consecutive pregnancies to the same woman that differed in terms of treatment status. For example, a woman may have not taken antidepressants in the first pregnancy but initiated an antidepressant in a second pregnancy, or she may have discontinued antidepressants in the first pregnancy but then continued antidepressants in a second pregnancy. These being pregnancies to the same women, any observed or unobserved characteristics that remain stable between pregnancies cannot confound the treatment–outcome association when they are analysed as matched pairs. For this reason, matched treatment-discordance designs are robust against both observed and unobserved confounders that are constant between pregnancies. Further detail on the selection and specification of statistical models used for these analyses is provided in Chapter 6.

Negative control analysis

We examined the risk of offspring neurodevelopmental outcomes where antidepressants were prescribed before but not during the pregnancy. 74 If prescription of antidepressants before the gestational period is associated with increased risk of an adverse outcome, it is unlikely that these associations are because of the effect of in utero exposure to the medication and would, therefore, suggest confounding by other characteristics. Further details of the method and results of these analyses are presented in Chapter 7.

Variation by indication: depression compared with other indication for antidepressant prescribing

To explore potential confounding by the indication, where the severity of depressive symptoms during pregnancy may influence both the likelihood of treatment and the risk of adverse outcome, we compared associations where antidepressants had been issued for depression with associations where antidepressants were likely to have been issued for other indications. A stronger association of antidepressants prescribed for depression is suggestive of confounding by the indication. Methods and results pertaining to these analyses are described in Chapter 8.

Association by timing of initiation in pregnancy

To identify potentially sensitive periods in fetal development, we compared risk of offspring neurodevelopmental problems where antidepressants were initiated in the first trimester with where they were issued in the second or third trimester. Further detail of the methods and the results of these analyses are presented in Chapter 9.

Dose response of associations

To assess dose–response relationships of antidepressant use with offspring neurodevelopmental disorders, we categorised the dose of antidepressants prescribed to each woman into low, moderate and high. It should be noted that, although such associations may highlight any dose–response relationships, they remain vulnerable to the possibility of confounding by the severity of the indication. Further detail of the methods and results for these analyses are presented in Chapter 9.

Associations for type of antidepressants

We examined associations with offspring neurodevelopmental outcomes where women had been prescribed SSRIs, tricyclic antidepressants (TCAs) or other types of antidepressants during pregnancy. Where women were issued different types of antidepressants during the same pregnancy, pregnancies counted independently to each risk estimate (e.g. women who were prescribed a SSRI and TCA were considered in the analysis of either drug type). Further detail and results for these analyses are presented in Chapter 9.

Associations by serotonin transporter receptor affinity

We examined the risk of offspring neurodevelopmental outcomes in relation to the serotonin transporter (SERT) affinity of antidepressant medications. 46,50 For these analyses, we compared women who were prescribed antidepressants in pregnancy with women who were not prescribed antidepressants in pregnancy. Further detail and results for these analyses are presented in Chapter 9.

Associations for specific antidepressant medications

Where we had sufficient numbers to enable statistical analyses, we report the associations of specific medications with neurodevelopmental outcomes. Where women were prescribed different medications within the same pregnancy, we counted them independently towards the risk estimates for all medications prescribed and then limited our analyses to women prescribed only a single medication within the same pregnancy as a sensitivity analysis. Further detail and results for these analyses are presented in Chapter 9.

Study cohorts

Depending on the outcome under investigation, we used the pregnant women’s cohort or the mother and child cohort for analysis, as described Chapter 2, Study cohort selection, to use the largest available sample size relevant to each outcome.

Statistical analysis

First, we compared the characteristics of women in each arm of our target trial to assess differences in covariate distributions.

Logistic regression models with cluster-robust variances were used to estimate the relative odds associated with initiating an antidepressant in pregnancy for each of the following binary outcomes:

• whether or not women consulted with their GP for depression or self-harm during pregnancy and in each of the four consecutive 6-month follow-up periods

• whether or not women had been referred by their GPs to specialist services during pregnancy and each of the four consecutive 6-month follow-up periods

• whether or not they had been admitted as an inpatient or outpatient to specialist mental health services during pregnancy and each of the four consecutive 6-month follow-up periods

• whether or not they were still or again on antidepressants 2 years after the pregnancy end date

• whether or not children resulting from the study pregnancies had been diagnosed with autism, ADHD or intellectual disability.

We used negative binomial regression models with cluster-robust variances to estimate incidence rate ratios for the following count outcomes:

• the number of days on which the mother had consulted with her GP during pregnancy and further follow-up periods

• the number of days on which the mother had consulted with her GP specifically for depression during pregnancy and further follow-up periods

• the number of times the mother had attended A&E services during pregnancy and further follow-up periods.

To account for differential length of follow-up between treatment groups, for instance because of differences in time of initiation or length of pregnancy, we included the natural logarithm of a time-at-risk variable in our models, constraining its regression coefficient to one.

Propensity score-matched regression

Using a CART model69 with 15,000 iterations in RStudio version 1.0.153 (The R Foundation for Statistical Computing, Vienna, Austria), we estimated a continuous score capturing women’s propensity to initiate an antidepressant in pregnancy based on their other measured characteristics. By contrast with propensity score estimation by parametric methods (where a single model is chosen to predict the data), a CART model uses a multitude of potential models, including interaction terms, and optimises its prediction across these. For this reason, CART models are well suited to predictive data modelling problems, such as propensity score estimation, because they do not depend on subjective decisions regarding the specification of the predictive model. Using the estimated propensity scores, we matched pregnancies during which women initiated antidepressants with pregnancies during which they received no treatment during the study period. Matches were carried out in a 1 : 1 ratio, without replacement and not allowing the propensity scores of matched pairs to differ by more than 0.2 standard deviations (SDs). We evaluated the quality of the matching algorithm by comparing standardised mean differences in covariate distributions before and after matching (Figures 5 and 6; see Report Supplementary Materials 21–23 for PSM analyses carried out in cohort subsets for which linked data were also included) and then exported the matched data sets to Stata 15.1/MP for statistical analysis. In these analyses, no further statistical adjustments for covariates were made because the groups were sufficiently balanced on the propensity score.

FIGURE 5.

Standardised absolute mean differences in covariate distributions between initiators and non-initiators of antidepressants in the women’s cohort before (blue) and after (orange) matching on propensity scores. Note: (1) ASAM = the Average Standardised Absolute Mean difference for all covariates.

FIGURE 6.

Standardised absolute mean differences in covariate distributions between initiators and non-initiators of antidepressants in the offspring cohort, before (blue) and after (orange) matching on propensity scores. Note: (1) ASAM = the Average Standardised Absolute Mean difference for all covariates.

Descriptive statistics of initiators versus non-initiators of antidepressants

Table 1 describes the characteristics of the study population of the women’s cohort by treatment status. There were 18,978 pregnancies in which women had evidence of depression during the pregnancy or within the preceding 12 months and, of these, antidepressants were initiated in 6177 pregnancies. Women who initiated an antidepressant during pregnancy were, on average, 0.3 years older and had seen the GP more frequently in the year prior to pregnancy than women who were not initiated on an antidepressant. At the start of pregnancy, women who were thereafter initiated on an antidepressant were more likely to have a history of physical comorbidities and, except for psychosis and bipolar affective disorder, were more likely to have a history of other psychiatric illness. They were also more likely to have been prescribed medications other than antidepressants in the year prior to or during pregnancy, to be current or ex-smokers, and to have a record of severe depression in their medical histories when they became pregnant. The treatment groups also differed in terms of area of residence within the UK and calendar year of the pregnancy.

Table 2 describes the characteristics of the study population of the mother and child cohort by treatment status in terms of women with depression who were initiated on an antidepressant during pregnancy compared with women who were not initiated on an antidepressant and had the child’s record linked with a minimum follow-up period of 4 years. There were 8478 pregnancies in which women had evidence of depression and, of these pregnancies, antidepressants were initiated in 2649. The characteristics of women initiating in this cohort were largely similar to those described in Table 1 for the women’s cohort, barring that there was no age difference between women in the treatment groups observed.

Table 3 provides descriptive statistics for the outcomes evaluated (number and percentages for categorical outcomes and the average number of events with SD for count outcomes) in the regression and propensity score analysis. We used the maximum data available for each outcome under investigation; given that linked data were available for only a subset of women, these analyses included a smaller number of women.

All neurodevelopmental conditions were relatively rare and were observed in less than 2% of the sample in either group. The prevalence of autism (1.85%) and ADHD (1.7%) was slightly greater in children of women who initiated antidepressants for depression in the main sample than in children of women who had depression but were not initiated on antidepressants (1.58% for autism and 1.01% for ADHD).

Results of multivariable regression and propensity score-matched analysis

To control for differences in measured characteristics between treatment groups, we examined associations between treatment status and the various outcomes while adjusting statistically for covariates and matching on the propensity to initiate antidepressants (Tables 4–6).

Table 4 presents results relating to women’s use of primary care services during pregnancy and within each of the four additional 6-month follow-up periods. Crude regression estimates suggested that women who had initiated an antidepressant consulted more frequently with their GPs, for any reason or specifically for depressive symptoms, during or up to 2 years after pregnancy than women who received no antidepressants. These women were also more likely to be prescribed an antidepressant medication at the end of the 2-year follow-up period. These associations remained after statistical adjustment for measured differences between treatment groups and in propensity score-matched analysis.

Differences between treatment groups in odds of consulting with the GP for episodes of self-harm were imprecise owing to a small number of observations (see Table 4). Where estimates were sufficiently powered, we observed greater odds of GP consultations for self-harm between 6 and 12 months after the pregnancy end date associated with initiating an antidepressant in multivariable regression analyses (OR 2.81, 95% CI 1.15 to 6.85). The apparently protective effect of initiating an antidepressant in terms of consulting for self-harm between 18 and 24 months after pregnancy (OR 0.11, 95% CI 0.01 to 0.92) was based on less than three treated individuals who experienced the outcome and is, therefore, likely to be unreliable.

With regard to GP referrals to secondary care services for depression during pregnancy, there was weak evidence that these were less likely among women who had initiated antidepressants compared with women who did not initiate antidepressants in pregnancy when the data were examined using multivariable regression analyses (OR 0.72, 95% CI 0.49 to 1.06) and comparably stronger evidence when using propensity score-matched regression analyses (OR 0.59, 95% CI 0.38 to 0.90). Conversely, we observed evidence for a twofold increased odds of referral to secondary services between 12 and 18 months after the pregnancy end date among women who had initiated antidepressants using both multivariable and propensity score-matched regression models.

Examining differences in continued need for antidepressants at the end of follow-up, we observed that women who had been initiated on an antidepressant in pregnancy were twofold more likely than women who had not been prescribed in the year before or during the study pregnancy to be prescribed an antidepressant 2 years after the pregnancy end date in regression (OR 2.16, 96% CI 1.95 to 2.39) and propensity score analyses.

The results of the analyses presented in Table 4 were very similar when repeated in the subset of the data with record linkages available to enable additional control for deciles of IMD as a covariate (see Report Supplementary Material 24).

The results pertaining to women’s use of secondary care services are presented in Table 5. These were broadly consistent with results for primary care outcomes in suggesting that odds of in-patient admission or out-patient treatment for a mental health problem were greater among women who had initiated antidepressants. However, low statistical power resulted in wide confidence intervals (CIs) around some estimates. While crude regression analyses suggested that A&E attendances were more common among women who had initiated antidepressants, these associations did not persist on statistical adjustment for potential confounders and/or matching on propensity scores.

Associations between initiation of an antidepressant and diagnoses related to neurodevelopmental problems in offspring are reported in Table 6. While we observed little evidence of associations in crude and multivariable regression analyses, there was some evidence for increased odds of offspring autism with initiation of an antidepressant in propensity score-matched analyses (OR 1.64, 95% CI 1.01, 2.66). There was evidence of increased odds of offspring ADHD with initiation of antidepressants in crude regression analyses, although this association attenuated on statistical adjustment for potential confounders and in propensity score-matched regression analyses, albeit with wide CIs. We observed little evidence for a difference in odds of offspring intellectual disability with initiation of an antidepressant during pregnancy.

Repeating the analysis presented in Table 6 on a smaller subset with availability of linked data to enable further adjustment for deciles of IMD led to similar point estimates in multivariable regression for all neurodevelopmental outcomes. The point estimates were attenuated for autism, and inflated for ADHD in the propensity score analysis, albeit with wide CIs due to smaller numbers (see Report Supplementary Material 25).

Study cohorts

As described in the previous chapter, depending on the outcome under investigation we used the pregnant women’s cohort and the mother and child cohort for analysis. Each of these made optimal use of the available data, as described in Chapter 2, Study cohort selection.

Statistical analysis

We first compared the characteristics of women in each arm of our target trial, that is women who continued an antidepressant into pregnancy with women who discontinued prior to becoming pregnant, to assess differences in covariate distributions.

Logistic regression models with cluster-robust variance estimators were used to estimate the relative odds of the following outcomes: GP consultations for self-harm, GP referrals to specialist services, admission as an inpatient or outpatient to specialist mental health services, prescription status 2 years after the pregnancy end date, and diagnoses relating to autism, ADHD or intellectual disability in offspring from the age of 4 years.

We used negative binomial regression models with cluster-robust variance estimators to estimate incidence rate ratios for the number of days on which the mother had consulted with her GP, consulted with her GP specifically for depression and attended A&E services during pregnancy and further follow-up periods. To account differential length of follow-up between treatment groups, we included the natural logarithm of a time-at-risk variable in our models, constraining its regression coefficient to one.

Propensity score-matched regression

Using a CART model with 15,000 iterations,69 we estimated a continuous score capturing women’s propensity to continue an antidepressant in pregnancy based on their other measured characteristics. Using the estimated propensity scores, we then matched pregnancies where women continued antidepressants with pregnancies where women discontinued prior to conception. Matches were carried out in a 1 : 1 ratio, without replacement, with a calliper of 0.2 SDs. We evaluated covariate imbalance between treatment groups before and after matching (Figures 8 and 9; see Report Supplementary Materials 26–28 for PSM analyses carried out in cohort subsets in which linked data were also included) and then exported the matched data sets to Stata 15.1/MP for further analysis. No further statistical adjustment for covariates were made as the groups were sufficiently balanced on the propensity score.

FIGURE 8.

Standardised absolute mean differences in covariate distributions between continuers and discontinuers in the women’s cohort, before (blue) and after (orange) matching on propensity scores. Note: (1) ASAM = the Average Standardised Absolute Mean difference for all covariates.

FIGURE 9.

Standardised absolute mean differences in covariate distributions between continuers and discontinuers in the mother and child cohort, before (blue) and after (orange) matching on propensity scores. Note: (1) ASAM = the Average Standardised Absolute Mean difference for all covariates.

Descriptive statistics of continuers versus discontinuers of antidepressants

Table 7 describes the characteristics of the study population of the women’s cohort by treatment status. There were 61,125 pregnancies in which women had a prior prescription of antidepressants for depression; of these 37,278 women continued the antidepressant into their pregnancy and 23,847 discontinued the antidepressant at least 2 months before the start of pregnancy. Women who continued antidepressants into pregnancy were, on average, 1.2 years older when they became pregnant and had consulted with their GPs on 1.2 more occasions in the year prior to pregnancy than women who had discontinued antidepressants prior to becoming pregnant. On becoming pregnant, women who had continued antidepressants were more likely to have a history of physical health comorbidities and a range of psychiatric comorbidities, to have been prescribed other medications during the study period and to be current or ex-smokers than women who discontinued antidepressants. They were also more likely to have records of both mild and severe depression in their medical histories, to have been prescribed multiple antidepressants at the same time or to have switched from one antidepressant to another during the study period than women who discontinued. There was evidence of a difference between treatment groups in the area of residence within the UK and calendar year of the pregnancy.

Table 8 describes the characteristics of the study population of the mother and child cohort by treatment status used for these analyses (i.e. these refer to women whose pregnancies could be linked to the child’s records), with a minimum follow-up of 4 years. There were 25,796 pregnancies in which women had a prior prescription of antidepressants for depression; of these pregnancies, 15,295 women continued the antidepressant into their pregnancy and 10,501 discontinued the antidepressant by the start of pregnancy. The characteristics of women who continued antidepressants into pregnancy compared with those who discontinued in this cohort were similar to those described above for the women’s cohort.

Table 9 provides descriptive statistics for the outcomes evaluated (number and percentages for categorical outcomes and the average number of events with SD for count outcomes) in the regression and propensity score analysis. We used the maximum data available for each outcome under investigation and given that linked data were available for only a subset of women those analyses included fewer women. All neurodevelopmental conditions were relatively rare and observed in less than 2% of either group. The prevalence of autism, ADHD and intellectual disability appeared similar in all groups irrespective of whether the women continued or discontinued antidepressants during pregnancy.

Results of multivariable and propensity score-matched regression analysis

To control for differences in measured characteristics between treatment groups, we examined associations between treatment status and the various outcomes while adjusting statistically for covariates, and matching on the propensity to initiate antidepressants, as shown in Tables 10–12.

For results relating to women’s use of primary care services during pregnancy and each of the four additional 6-month follow-up periods, see Table 10. Although crude associations suggested that women who had continued antidepressants had consulted with their GPs more frequently for any reason, statistical adjustment for covariates revealed a slight protective effect. However, this finding was not replicated in propensity score-matched regression analyses in which control for potential confounders would have been more efficient. In terms of GP consultations specifically for depression, women who had continued antidepressants consulted more frequently with their GPs than women who had discontinued antidepressants, with consistent associations observed for crude, multivariable and propensity score-matched regression analyses. In general, women who had continued an antidepressant were more likely to have consulted for self-harm or to have been referred to specialist mental health services for depression during pregnancy and in further follow-up periods. Women who had continued antidepressant in pregnancy also had over two-fold odds of still being prescribed antidepressants at 2-year follow-up than women who had discontinued. Repeating these analyses on a subsample with linked data for further adjustment with decile of IMD led to similar results (see Report Supplementary Material 29).

In term of outcomes related to secondary care (see Table 11), crude associations suggested that women who had continued an antidepressant were more likely than women who had discontinued an antidepressant to have received inpatient or outpatient treatment for a mental health problem during pregnancy or in further follow-up periods. Statistical adjustment for potential confounders and/or matching on propensity scores resulted in weaker associations, but odds generally remained elevated. Crude associations suggested that women who continued antidepressants had attended A&E services more frequently than those who discontinued antidepressants, although these associations attenuated to the null on statistical adjustment for potential confounders or after matching on the propensity to continue antidepressant treatment.

See Table 12 for a description of the associations between continuation of an antidepressants and diagnoses of neurodevelopmental disorders in offspring. We observed no differences in odds for autism or for intellectual disability for mothers who continued or discontinued antidepressants during pregnancy. Although we observed lower odds of offspring intellectual disability among women who had continued taking antidepressants, this association attenuated to the null on statistical adjustment for potential confounders and in propensity score-matched regression analyses. Repeating these analyses on the subset of the sample with linked data to further adjust for deciles of IMD led to broadly similar results with wider CIs (see Report Supplementary Material 30).

Identification of the general practitioner

Many women in our sample will have been seen by different GPs during the course of their pregnancies. Therefore, we aimed to identify the GP who, by issuing or not issuing an antidepressant, had determined the woman’s the treatment status, and defined the IV on that GP’s prior prescribing behaviour. That is, for patients who initiated or continued an antidepressant into pregnancy, we identified the first GP to have prescribed an antidepressant in pregnancy, whereas for patients who received no treatment or discontinued an existing prescription, we identified the GP who had attended the first consultation after the estimated date of conception. Where antidepressants were issued as part of a repeat prescription, we did not consider prescriptions for which there had been no face-to-face or telephone consultation. We made these exclusions because it may be likely that, in these instances, the GPs would have been unaware of their patient being pregnant.

Treatment variables

We considered two treatment variables. The first treatment variable captured the initiation of an antidepressant during pregnancy compared with receiving no antidepressant treatment, and the second captured the continuation of an antidepressant into pregnancy compared with discontinuing at least 2 months prior to conception, as described in Chapter 2, Definition of treatment groups.

Outcome variables

Our outcome variables for these analyses included a diagnosis of (1) autism, (2) ADHD or (3) intellectual disability when the child was at least 4 years of age.

Statistical analysis

For all observations with a non-missing value on the IV, we estimated the associations using ordinary least squares (OLS) regression (using a linear probability model for the binary outcomes) to allow for comparison of these estimates with associations estimated using the IV method applying two-stage least squares regression. All associations were estimated using robust standard errors to account for clustering of patients treated by the same GP.

Selection and evaluation of the instrumental variable

We found an IV based on eight prior consultations with other pregnant patients to be optimally associated with the treatment variables, while including an optimal number of observations in IV analyses. For the IV based on eight prior consultations, we checked associations of the instrument with potential confounders of the treatment–outcome association using bias component plots. 75 Results from these analyses suggested weak associations of the instrument with the following covariates: (1) number of GP visits in the year prior to pregnancy; (2) region of residence in the UK; (3) calendar year of the pregnancy; (4) concurrent use of multiple antidepressants; (5) maternal age; and (6) Charlson Comorbidity Index score. We, therefore, adjusted for these characteristics in the analytical models.

Results of standard and instrumental variable analysis

The results of the standard and IV analyses are presented in Table 13. Among observations with a non-missing value on the IV, there was no observed association between initiation or continuation of antidepressants and any of the neurodevelopmental outcomes in either the standard or the IV approach, although CIs were wide in all analyses. Using Wu–Hausman F-tests, there was little evidence for differences in effect size between OLS regression and IV estimates for these outcomes. This does not provide any evidence that residual confounding can explain our OLS results.

Definition of treatment discordance

Treatment discordance was defined as (1) having initiated an antidepressant in one pregnancy and having received no treatment in another pregnancy, or (2) having continued an antidepressant in one pregnancy and having discontinued antidepressants in another. All women contributing at least two pregnancies that were discordant in terms of treatment status were considered in the analysis.

Definition of outcome discordance

Statistical analysis

For binary outcomes, we first estimated associations using standard logistic regression models to assess whether or not the associations observed in Chapters 3 and 4 of this report were present in this subset of the data. Associations were estimated using robust standard errors, clustering on women’s patient identification numbers, as our estimates were based on comparisons of more than one pregnancy in the same women. We then used fixed-effects logistic regression models to estimate the matched treatment–outcome association. We followed a similar approach for count outcomes, first estimating associations using standard negative binomial regression and second using a fixed-effects negative binomial regression model to estimate the matched treatment–outcome association. We statistically adjusted all associations for calendar year and maternal age at delivery.

Definition of the treatment variable and its negative control

As a negative control, we identified all women who had discontinued an antidepressant at least 3 months prior to becoming pregnant, comparing them with a reference group that had not been prescribed antidepressants at all. To contrast the effect of the negative control with that of the actual treatment, we combined women who had initiated or continued an antidepressant in pregnancy and compared them with the same reference group. Any woman who had been prescribed antidepressants solely during the grace period, as described in Chapter 2, was not considered for these analyses.

Outcome variables

These analyses were relevant only for outcomes for which the timing of prescription in relation to the pregnancy period could have potentially influenced risk. We, therefore, carried them out only for neurodevelopmental outcomes in offspring where exposure to an antidepressant in utero may be a potentially causal mechanism of any observed association (i.e. autism, ADHD and intellectual disability).

Statistical analysis

We used logistic regression to compare associations with outcomes for the actual treatment variable (prescription during pregnancy) with its negative control (prescription before but not during pregnancy). Associations were estimated using cluster robust variance estimators to recognise the presence of consecutive pregnancies to the same women. We adjusted all associations for the potential confounders described in Chapter 2, Covariates.

Pregnancies where antidepressants had not been issued for depression

To identify pregnancies during which antidepressants had been issued for other indications than depression, we selected those where prescription had occurred during the study period, but no current or past depressive symptoms were present in women’s medical records. Readers should note that the pregnancies identified here are the same as those set aside as described in Chapter 2, Study cohort selection.

Treatment variables

As described in Chapter 2, Design: observational cohorts emulating target randomised controlled trials, and elsewhere in this report, we considered associations with the following treatment variables: (1) initiating an antidepressant during pregnancy compared with receiving no antidepressant treatment during the study period, and (2) continuing an antidepressant into pregnancy compared with discontinuing the antidepressant.

Outcome variables

Methods

We limited these analyses to the timing of initiation of antidepressants in pregnancy because we could clearly identify the point at which women were first exposed to antidepressants during pregnancy. We coded a time-specific exposure variable to indicate exposure in the first trimester compared with prescriptions issued in the second or third trimesters (these latter time points had to be combined because of small cell counts). We then used logistic regression models with cluster-robust variances to estimate the relative odds of offspring autism, ADHD and intellectual disability associated with initiating an antidepressant in the first trimester, or in the second or third trimesters, compared with offspring born to women who did not initiate an antidepressant. In addition, to providing crude estimates, we adjusted the ORs for potential confounding variables (see Chapter 2, Covariates).

Results

Table 19 shows the results of the analysis estimating relative odds of offspring neurodevelopmental disorders by timing of initiation of antidepressants in pregnancy. Overall, associations between antidepressants during pregnancy and offspring odds of autism, ADHD or intellectual disability did not appear to vary with timing of initiation. We observed weak evidence for an association between first trimester initiation of an antidepressant and offspring ADHD, although a similar association was observed for second or third trimester initiation of an antidepressant with wide CIs.

Methods

We identified the generic drug category and daily dose in milligrams for each individual prescription of an antidepressant. Using this information, we calculated tertiles of distributions of daily doses in milligrams separately for each of 34 generic drug categories, and then combined this information in a single dose level variable (first tertile defined as low doses, second tertile as moderate doses and third tertile as high doses). In case of pregnancies during which prescriptions had been issued at different dose levels, we used the highest daily dose prescribed. Using logistic regression models with cluster-robust variances, we estimated relative odds of offspring neurodevelopmental disorders associated with being prescribed a low, moderate or high dose of antidepressants during pregnancy, compared with not having been prescribed antidepressants while pregnant. We provide both crude and statistically adjusted estimates.

Results

There was some evidence for a dose–response association between antidepressants prescribed to the mother during pregnancy and offspring odds of autism, although the CIs around estimates for low, moderate and high doses overlapped (Table 20). There was no clear evidence for dose–response association with offspring ADHD or intellectual disability.

Method

We categorised individual antidepressants prescribed during pregnancy into the following groups: (1) SSRIs included prescriptions for citalopram, escitalopram, fluoxetine, fluvoxamine, paroxetine and sertraline; (2) TCAs included prescriptions for amitriptyline, clomipramine, dosulepin, doxepin, imipramine, lofepramine, mianserin, nortriptyline and trimipramine; and (3) ‘other’ antidepressants included agomelatine, duloxetine, isocarboxazid, mirtazapine, moclobemide, nefazodone, trazodone, phenelzine, reboxetine and venlafaxine. In cases where women had been prescribed different types of antidepressants during pregnancy, they counted independently towards analyses for each type (e.g. pregnancies where a SSRI and TCA had been prescribed were considered in analyses of SSRIs as well as in analyses of TCAs). We used logistic regression models with cluster-robust variances to estimate relative odds of offspring neurodevelopmental disorders associated with being prescribed a SSRI, a TCA, or an ‘other’ antidepressant during pregnancy, compared with not having been prescribed antidepressants while pregnant. We provide both crude and statistically adjusted estimates.

Results

The results of the analyses comparing no antidepressant prescription for depression in pregnancy with prescription of antidepressants grouped into SSRIs, TCAs and other antidepressants are provided in Table 21. We observed greater adjusted odds of autism among children whose mothers had been prescribed SSRIs (OR 1.26, 95% CI 1.04 to 1.53) or TCAs (OR 1.58, 95% CI 1.12 to 2.24) during pregnancy, although CIs for other antidepressants were wider, probably reflecting smaller numbers. There was little evidence for association between the type of antidepressant issued during pregnancy and later risk of ADHD or intellectual disability in resulting offspring.

Methods

We identified prescriptions where antidepressants with low, moderate or high SERT affinity had been issued during pregnancy: (1) low-affinity medications included desipramine, nortriptyline, amoxapine, doxepin, trimipramine, trazodone, nefazodone and mirtazapine; (2) moderate-affinity medications included citalopram, imipramine, fluvoxamine, amitriptyline and venlafaxine; and (3) high-affinity medications included escitalopram, fluoxetine, paroxetine, sertraline, duloxetine and clomipramine. These groupings were based on previous work on this topic46,50 but it is important to note that the empirical evidence behind these remains limited and, therefore, results should be viewed with caution. Where women were prescribed antidepressants with different affinities, they were counted independently in each analysis. We used logistic regression models with cluster-robust variances to estimate relative odds associated with being prescribed a low, moderate or high SERT affinity medication compared with not having been prescribed antidepressants during pregnancy, providing both crude and statistically adjusted estimates.

Results

The point estimates of offspring odds of all neurodevelopmental outcomes were lower for higher-affinity antidepressants than those for lower-affinity antidepressants, although the CIs for all estimates overlapped (Table 22). There were increased odds of autism among children whose mothers had been prescribed moderate-affinity antidepressants (OR 1.50, 95% CI 1.16 to 1.94), and increased odds of ADHD among children whose mothers had been prescribed low-affinity antidepressants (OR 1.96, 95% CI 1.06 to 3.64), compared with children whose mothers had not been prescribed antidepressants during pregnancy. It should be noted that low- or moderate-affinity antidepressants are generally used for the treatment of more severe depression; therefore, these associations may be consistent with residual confounding by the severity of depression.

Methods

Analyses of individual medications were limited by small cell counts. We will, therefore, report results only for which we had at least five observations within each cell of bivariate tables of exposure and outcome variables. For example, we required at least five instances in which women had been prescribed citalopram while pregnant with linked offspring who were later diagnosed with autism to report this result. Where women had been prescribed multiple medications during pregnancy, they counted independently towards analyses of each medication. We used logistic regression models with cluster robust variances to estimate associations with specific antidepressant medications, providing both crude and adjusted relative risk and odds estimates.

Results

The results of the associations of individual medications with neurodevelopmental outcomes are presented in Table 23. There were some variations in the outcomes in relation to individual medications. The adjusted odds of offspring autism were greater when mothers had been prescribed amitriptyline, citalopram, lofepramine or paroxetine during pregnancy. There was weak evidence that the odds of offspring ADHD were increased among mothers who had been prescribed amitriptyline or mirtazapine while pregnant, but little evidence for increased odds of intellectual disability with prescription of any individual antidepressant medications during pregnancy. The CIs of all these results are wider owing to smaller numbers contributing to the analyses.