Antidepressant treatment and risk of subsequent bipolar disorder in adolescents with unipolar depression

Abstract

Background

Selective serotonin reuptake inhibitors (SSRIs) are a common treatment for depression and anxiety in adolescents but are associated with an increased incidence of bipolar disorder (BD). Whether this relationship is causal remains unclear.

Objective

We applied a quasi-experimental design to national registry data, using an instrumental variable (IV) approach (regional variation in prescribing practice) to investigate for a causal relationship between adolescent SSRI treatment and subsequent risk of BD.

Methods

We used national electronic health register data on individuals born 1991–1998 followed to maximum age 32 years, looking at individuals diagnosed with unipolar depression in adolescence. Using regional variation in prescribing practice as an IV, we compared risk of BD in adolescents prescribed vs not prescribed SSRIs (fluoxetine, sertraline or citalopram).

Findings

In non-IV analyses, adolescents who were prescribed SSRIs had an increased risk of BD, in keeping with previous research. Subsequent IV analyses, however, did not support a causal relationship between SSRI treatment and BD risk, either in the short or long term.

Clinical implications

These findings do not support a causal relationship between SSRI treatment and risk of BD. Rather, they suggest that the apparent relationship between SSRI treatment and later BD may be a result of unmeasured confounding.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Many observational studies have shown an association between selective serotonin reuptake inhibitor (SSRI) treatment in adolescence and an increased risk of subsequent mania or hypomania. These findings, however, are potentially biased by unmeasured confounding factors, particularly confounding by indication, where the underlying severity of depression is itself a risk factor for mania/hypomania as well as being an indication for SSRI treatment.

WHAT THIS STUDY ADDS

• Using an instrumental variable design to better account for unmeasured confounding, this quasi-experimental study found no evidence of a causal relationship between SSRI treatment and subsequent short- or long-term risk of mania/hypomania. This suggests that previously observed associations are likely due to unmeasured confounding factors, notably depression severity, rather than by SSRI treatment itself.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• These findings provide important reassurance regarding the safety of SSRI treatment in terms of mania/hypomania risk in adolescents and highlight the value of quasi-experimental methods for evaluating causal effects of psychiatric treatments.

Introduction

Bipolar disorder (BD) is a serious mental disorder, which typically emerges in late adolescence or early adulthood¹ and is a major contributor to disability in young people.² Up to 50–80% of individuals with BD first present to mental health services with a depressive (rather than a manic or hypomanic) episode,³,⁸ frequently resulting in treatment with antidepressant medication, usually selective serotonin reuptake inhibitors (SSRIs). Many studies, however, have shown that SSRI treatment of depression is associated with an increased incidence of subsequent manic episodes,⁹,¹⁶ especially in adolescents.¹⁰,^{1214 16} Whether or not this relationship is causal remains debated.^{11 19 20} Given that as many as half of all individuals diagnosed with BD in the population had attended child and adolescent psychiatry services, most commonly with depression,^{21 22} there is a clear need for robust studies on the relationship between SSRI treatment and later risk of BD in this population.

In short-term, controlled studies, SSRIs have not generally been associated with an increased incidence of BD, though, notably, most such studies have been small and patients were usually also prescribed mood stabilising medications.²³,²⁶ Longer-term studies have raised greater concern. One recent large observational study,¹⁸ for example, found that, while SSRI treatment in children and adolescents was not predictive of shorter-term manic switches, there was an increased risk of mania when patients were followed to 1 year.

Randomised controlled trials (RCTs) remain the gold standard for causal inference. However, they tend to have short-term follow-up periods, small sample sizes and unrepresentative samples. For example, recent research has demonstrated that the majority of trials of SSRIs in adolescents with depression explicitly excluded young people with suicidality, even though suicidal thoughts and behaviour are common features of depression. These factors make it difficult to use RCTs to determine the real-world risk of mania associated with SSRI treatment.

Routinely collected observational data, on the other hand, can provide large representative samples and a long-term follow-up with minimal attrition, which enables investigations of long-term treatment effects on rare outcomes like mania. Observational studies based on routinely-collected data, however, may be subject to bias from confounding-by-indication, which occurs when the indication for a treatment (eg, severity of depression as an indication for prescribing SSRIs) is related to the outcome of interest (eg, BD), leading to a spurious association between the treatment and the outcome.²⁷

One analytical approach that can be adopted when there are unmeasured confounding factors—such as illness severity—is an instrumental variable (IV) approach.^{28 29} This approach relies on the presence of a variable (termed ‘instrument’) that is: (1) associated with the exposure, (2) unrelated to the confounders, and (3) unrelated to the outcome other than through its association with the exposure.³⁰,³³

One candidate IV is physician prescribing propensity.^{34 35} SSRI prescribing propensity has been shown to vary significantly between different physicians and across clinical centres.³⁶ A number of studies have leveraged this variability in prescribing practices for psychiatric medications,³⁷,⁴⁰ including SSRIs,⁴¹ in order to estimate causal effects on a range of health outcomes. In the current study, involving electronic health records of individuals born 1991–1998 and followed to maximum age 32 years, we used regional-level SSRI prescribing practices for adolescent depression as an instrument to investigate for a causal relationship with BD incidence. From a causal perspective, if the IV analyses indicate an increased risk of BD, this would support a causal effect of SSRIs. If, on the other hand, the IV analyses show no association, this suggests that the observed relationship in conventional analyses is likely driven by unmeasured confounding factors, such as depression severity, rather than a direct causal effect of SSRIs.

Methods

This study is reported according to the Strengthening the Reporting of Observational Studies in Epidemiology Using Mendelian Randomization: the STROBE-MR statement.⁴² Although originally designed for MR studies, the STROBE-MR checklist is applicable to any IV analysis, as MR is a subclass of IV methods.⁴²

Data sources

We used Welsh electronic health records within the Welsh Secure Anonymised Information Linkage (SAIL) data bank,⁴³,⁴⁷ which contains administrative data for the population of Wales—approximately 3.1 million in mid-2022.⁴⁸ SAIL’s Information Governance Review Panel granted approval to conduct this research (IGRP Number 1635). Information on the study variables until the end of the study period (ie, November 2023) was identified from a range of datasets in the SAIL Databank: (1) Welsh Demographic Service Dataset (WDSD) (1990–study end), (2) Patient Episode Database for Wales (PEDW) (1995–study end), (3) Outpatient Database for Wales (2004–study end) and (4) Welsh Longitudinal General Practice Dataset (WLGP) (2000–study end).

Study subjects and variables

Cohort

We included individuals born 1991–1998, alive until study end, registered with a Welsh GP before age 13 years, had attended child and adolescent psychiatry services and had an adolescent (age 13–18 years) unipolar depression record (online supplemental eFigure 1). Time zero was defined as the date of the earliest recorded depression in adolescence. This date marked the start of the observation period for all individuals, irrespective of subsequent exposure status.

We included individuals with an adolescent depression record, regardless of whether it is an incident or a prevalent event of depression (ie, preceded by childhood depression events). Depression records were identified using Read (V.2) codes for depression symptoms and diagnoses from GP data (WLGP) and International Classification of Diseases, 10th Revision (ICD-10) codes for diagnoses from inpatient data (PEDW), adapted from John et al⁴⁹ and reviewed by a consultant child and adolescent psychiatrist on the study team (IKe) (online supplemental eTable 1).

The UK clinical guidelines (National Institute for Health and Care Excellence) are that antidepressant medication should only be prescribed following a diagnosis of depression by a child and adolescent psychiatrist.⁵⁰ Therefore, eligibility for prescribing should only arise once seen in the child and adolescent psychiatry services, and young people diagnosed with depression in primary care would therefore not be eligible to enter our sampling frame. Contact with child and adolescent psychiatry services was defined as any outpatient appointment or inpatient admission associated with a psychiatry specialty before age 18 years. Inpatient contacts were admissions beginning before age 18 years, associated with either a primary diagnosis of any mental disorder (any ICD-10 F code), or where a specialty code for the admission was for a relevant psychiatry specialty. Outpatient contacts were appointments before age 18 years, associated with a psychiatry specialty code or a GP record of a Read (V.2) code denoting mental health service contact (online supplemental eTable 2). Read codes were adapted from Joseph et al⁵¹ and reviewed by IKe.

Exposure

Information on SSRI prescriptions was extracted from WLGP (see Read (V.2) codes adapted from John et al⁴⁹ in online supplemental eTable 3). Exposure was defined as receiving at least one prescription of the three most commonly prescribed SSRIs in adolescence—fluoxetine, sertraline or citalopram. We used a continuous exposure variable by calculating cumulative SSRI doses prescribed over several exposure windows (figure 1). The windows began on the date of the earliest adolescent depression record, to make exposed and unexposed individuals comparable for time-at-risk for outcome. To assess outcomes following a short-term exposure, we defined six short-term windows corresponding to 1–6 months of cumulative exposure. We also defined four long-term windows for 1–4 years of exposure.

Figure 1. Design of the investigation. GP, general practice; SSRI, selective serotonin reuptake inhibitor.

We standardised each SSRI prescription by the defined daily dose of the corresponding SSRI (fluoxetine: 20 mg, sertraline: 50 mg and citalopram: 20 mg)—based on the WHO anatomical therapeutic chemical code recommendation^{52 53}—to standardise dosage across SSRI types. The assignment and accumulation of possible SSRI exposure began on the same date as cohort entry (time zero). Exposure windows of varying duration (eg, 1 month, 6 months, 1 year) were defined from this date, and individuals’ cumulative exposure was calculated within each window. Cumulative exposure was further adjusted for the duration of each window, by dividing the standardised dose by the number of months in the window, assuming each prescription represented approximately 1 month of medication. This yielded a cumulative exposure score ≥0; for example, 0=no SSRI prescription in the window, 1=daily SSRI prescription at its defined daily dose throughout the window.

Outcome

Our outcome was an incident diagnosis of BD, mania or hypomania, following an exposure window. Individuals with any diagnosis recorded prior to the end of an exposure window were excluded (figure 1, online supplemental eFigure 1). The outcome variable was coded binary (incidence vs no incidence). Outcome diagnoses were identified via Read (V.2) codes from GP records (WLGP) or ICD-10 codes from inpatient records (PEDW), adapted from Carr et al⁵⁴ and Kuan et al⁵⁵ and reviewed by IKe. (online supplemental eTable 4).

For each exposure window, outcome follow-up began immediately after the end of that window. For example, for a 1-month exposure window, follow-up for bipolar disorder began 1 month after the earliest adolescent depression record for all individuals. Equivalent follow-up windows were therefore applied to both exposed and unexposed individuals to ensure comparability of risk periods. To prevent immortal-time bias, individuals who developed the outcome (bipolar disorder, mania or hypomania) prior to or during an exposure window were excluded from the corresponding window’s analysis, regardless of their exposure status. This ensured that both exposed and unexposed individuals were equally at risk from time zero onwards.

Instrumental variable

Earlier research has demonstrated that physicians’ prescribing preferences can vary across healthcare systems and regions, and this variation can be exploited for quasi-experiments.^{29 56 57} We used regional-level SSRI prescribing propensity as the IV. Region/cluster was defined as the catchment area of a GP. We omitted clusters with <15 individuals with adolescent depression (online supplemental eFigure 1), to ensure sufficient observations per cluster for valid IV estimation.^{58 59} We constructed 10 IVs—one per exposure window—by taking the cluster-level average of the standardised cumulative exposure score in that window. A ‘leave-self-out approach’ was used to ensure that an individual’s own prescription did not contribute to their cluster’s prescribing propensity.

For each of the 10 exposure windows (1 month through 4 years), we performed a separate IV analysis, to assess the incident outcome following that window (figure 1, online supplemental eFigure 1). While the IVs based on the long-term exposure helped us estimate the risk of bipolar disorder following long-term treatment, short-term windows helped us capture immediate risks following treatment on individuals who would be excluded in longer windows simply because of the treatment duration.

Covariates

The clinical confounders we considered were: severity of depression symptoms,^{18 60 61} age at depression diagnosis,^{18 61} psychiatric comorbidities, hospitalisation, psychiatric treatment prior to depression diagnosis,¹⁷,^{1960 62} history of substance use¹⁸⁶⁰,⁶³ and parental psychiatric disorders (including depressive and bipolar disorder)¹⁷,¹⁹⁶¹(online supplemental eFigure 2, Directed Acyclic Graph).

Of these, from GP data (WLGP), we were able to measure year and age at earliest adolescent depression record and SSRI prescriptions prior to that record. From inpatient data (PEDW), we measured history of inpatient child and adolescent psychiatry service contact prior to the earliest adolescent depression record (as proxy for the unobserved psychiatric illness severity). Additionally, at cluster level, we derived the percentage of the population who attended child and adolescent psychiatry services.

Socio-demographic confounders measured were: year of birth (WDSD), sex at birth (WDSD; male/female), childhood socioeconomic deprivation (WDSD; 2014 Welsh Index of Multiple Deprivation quintiles, with quintile 1 as most deprived,⁶⁴ urban/rural residence in childhood (WDSD; 2011 rural–urban classification by the Office for National Statistics,⁶⁵ low birth weight (National Community Child Health Database; as birth weight <2500 g) and winter birth (WDSD; births in December–February). Additionally, at cluster level, we derived the cluster size (ie, number of individuals registered to each GP), percentage of the population with urban residence, and the percentage of the population in each deprivation quintile (online supplemental eFigure 2, directed Acyclic Graph).

Statistical analysis

All analyses were conducted using Stata/SE V.19; no artificial-intelligence or machine-learning techniques were used. We described clinical and socio-demographic characteristics of the final sample, overall, and stratified by SSRI exposure (vs no exposure) within 4 years of the earliest adolescent depression record (table 1).

Table 1. Demographic, socioeconomic and clinical characteristics of subjects with depression, with and without SSRI prescription.

	Subjects with adolescent depression (among 1991–1998 birth cohort attending CAP services)
	Overall	Prescribed SSRI*	Not prescribed SSRI*
Total sample†	6615	4168 (63.01)	2447 (36.99)
Female	4788 (72.38)	3152 (75.62)	1636 (66.86)
Missing	1 (0.02)	1 (0.02)	0
Year of birth
1991	549 (8.30)	320 (7.68)	229 (9.36)
1992	573 (8.66)	359 (8.61)	214 (8.75)
1993	602 (9.10)	391 (9.38)	211 (8.62)
1994	719 (10.87)	450 (10.80)	269 (10.99)
1995	777 (11.75)	487 (11.68)	290 (11.85)
1996	984 (14.88)	644 (15.45)	340 (13.89)
1997	1139 (17.22)	713 (17.11)	426 (17.41)
1998	1272 (19.23)	804 (19.29)	468 (19.13)
Socioeconomic deprivation
Quintile 1 (most deprived)	2337 (35.33)	1513 (36.30)	824 (33.67)
2	1589 (24.02)	1009 (24.21)	580 (23.70)
3	1050 (15.87)	652 (15.64)	398 (16.26)
4	798 (12.06)	484 (11.61)	314 (12.83)
5 (least deprived)	841 (12.71)	510 (12.24)	331 (13.53)
Urban	5142 (77.73)	3266 (78.36)	1876 (76.67)
Low birth weight	716 (10.82)	472 (11.32)	244 (9.97)
Missing	28 (0.42)	14 (0.34)	14 (0.57)
Winter birth	1569 (23.72)	1009 (24.21)	560 (22.89)
Inpatient CAP service contact prior to adolescent depression	275 (4.16)	184 (4.41)	91 (3.72)
Age at earliest adolescent depression record, median (IQR), years	16.45 (15.29–17.27)	16.73 (15.69–17.37)	15.87 (14.78–16.91)
Age at first SSRI prescription following adolescent depression, median (IQR), years	NA	17.61 (16.75–18.39)	NA
SSRI prescription prior to adolescent depression	366 (5.53)	336 (8.06)	30 (1.23)
Diagnosis of bipolar disorder, mania or hypomania‡	343 (5.19)	266 (6.38)	77 (3.15)
Age at diagnosis of bipolar disorder, mania or hypomania, median (IQR), years‡	21.70 (19.01–24.67)	21.80 (19.23–24.72)	21.04 (17.30–24.23)

No. (%) presented unless specified otherwise; column % presented unless specified otherwise. For binary variables, only one category is presented for non-missing observations since the other category is complementary.

^*At least one prescription of fluoxetine, sertraline or citalopram within 4 years of earliest adolescent depression record.

^†Row % presented.

^‡Lifetime diagnosis until end of follow-up.

CAP, Child and Adolescent Psychiatry; NA, not applicable; SSRI, selective serotonin reuptake inhibitor.

Online supplemental eTable 5 presents a comparative outline of our quasi-experimental study with that of a hypothetical RCT designed to address the same research question. We first analysed the data in the conventional fashion—that is, by only accounting for measured confounders. We then used the IV analysis approach as an attempt to account for both measured and unmeasured confounders. All analyses accounted for clustering of individuals at the GP level.

Conventional (non-IV) analysis

We accounted for baseline-measured confounders via an outcome regression approach, an inverse probability weight (IPW) approach and a doubly-robust approach. In the outcome regression approach, we regressed the outcome directly on the exposure and the measured covariates.

In the IPW approach, we first predicted the propensity of exposure—both marginal and conditional on measured confounders—using a generalised linear model with log link and gamma family, to correctly model the right-skewed exposure distribution. We then adjusted for the stabilised exposure weights (ratio of marginal and conditional propensity score) to estimate the outcome.

The doubly-robust approach combined IPW with outcome regression, yielding consistent estimates if either the exposure model or the outcome model was correctly specified.⁶⁶

In all three approaches, we estimated average marginal effects from probit regressions, representing risk differences per one-unit increase in cumulative exposure (table 2).

Table 2. Conventional analyses showing risk differences for bipolar disorder/mania/hypomania incidence for cumulative SSRI prescription.

Exposure windows	Outcome regression		Inverse probability weight	Doubly robust
	Unadjusted*† (95% CI)	Adjusted*‡ (95% CI)	Adjusted*‡ (95% CI)	Adjusted*‡ (95% CI)
1 month	0.000 (−0.006 to 0.007)	0.001 (−0.007 to 0.008)	0.004 (−0.006 to 0.015)	0.006 (−0.005 to 0.017)
2 months	0.003 (−0.005 to 0.011)	0.004 (−0.006 to 0.013)	0.009 (−0.003 to 0.021)	0.012 (−0.001 to 0.025)
3 months	0.004 (−0.005 to 0.013)	0.004 (−0.005 to 0.014)	0.010 (−0.003 to 0.023)	0.014 (−0.000 to 0.028)
4 months	0.004 (−0.006 to 0.013)	0.005 (−0.006 to 0.015)	0.009 (−0.004 to 0.023)	0.014 (−0.001 to 0.028)
5 months	0.004 (−0.006 to 0.014)	0.004 (−0.006 to 0.015)	0.009 (−0.005 to 0.023)	0.013 (−0.002 to 0.028)
6 months	0.003 (−0.007 to 0.014)	0.004 (−0.006 to 0.015)	0.009 (−0.004 to 0.023)	0.012 (−0.002 to 0.027)
1 year	0.006 (−0.005 to 0.016)	0.008 (−0.002 to 0.017)	0.012 (−0.001 to 0.024)	0.014 (0.001 to 0.027)
2 years	0.009 (−0.002 to 0.020)	0.011 (0.002 to 0.021)	0.013 (0.002 to 0.024)	0.016 (0.005 to 0.027)
3 years	0.009 (−0.001 to 0.019)	0.013 (0.004 to 0.022)	0.015 (0.004 to 0.025)	0.019 (0.008 to 0.030)
4 years	0.011 (0.002 to 0.020)	0.014 (0.006 to 0.022)	0.016 (0.006 to 0.025)	0.019 (0.009 to 0.029)

The risk difference represents the absolute percentage-point change in the probability of the outcome per one-unit increase in cumulative SSRI exposure.

^*Analyses accounted for clustering of observations at GP level. Those who had outcome diagnosis prior to the intervention window were removed from the analysis.

^†Outcome-only probit model.

^‡Probit model accounting for sex, urban residence, deprivation, year of birth, low birth weight, winter birth, age at earliest adolescent depression record, year of earliest adolescent depression record, inpatient CAP service attendance prior to earliest adolescent depression record, SSRI prescription prior to earliest adolescent depression record, population size in GP cluster, percentage of GP population with a CAP service attendance, percentage of GP population from urban areas, percentage of GP population in each deprivation quintile.

CAP, Child and Adolescent Psychiatry; GP, general practice; SSRI, selective serotonin reuptake inhibitor.

IV analyses

For the IV analysis, our estimand was a local average treatment effect, which corresponds to the average causal effect among individuals for whom the SSRI prescribing propensity would vary across the clusters. We used the two-stage least squares regression method to estimate β coefficients that represent risk differences—ie, percentage-point changes in the probability of outcome per one-unit increase in cumulative exposure. We also used an IV probit model to account for the binomial outcome distribution. In both IV analyses, we additionally adjusted for all baseline-measured confounders (table 3).

Table 3. Instrumental variable analysis showing risk differences for bipolar disorder/mania/hypomania incidence for cumulative SSRI prescription.

Instruments (regional-level SSRI prescribing propensity) for different exposure windows	Outcome (incident diagnosis of bipolar disorder/mania/hypomania)
1 month
No. of outcomes/No. of total individuals	322/6594
F-statistic from first stage of IV 2SLS	20.69
Risk difference (95% CI) from IV 2SLS	−0.049 (−0.158 to 0.061)
Risk difference (95% CI) from IV probit	−0.049 (−0.165 to 0.067)
2 months
No. of outcomes/No. of total individuals	–
F-statistic from first-stage of IV 2SLS	23.13
Risk difference (95% CI) from IV 2SLS	−0.028 (−0.161 to 0.105)
Risk difference (95% CI) from IV probit	−0.028 (−0.167 to 0.111)
3 months
No. of outcomes/No. of total individuals	–
F-statistic from first-stage of IV 2SLS	25.21
Risk difference (95% CI) from IV 2SLS	−0.032 (−0.181 to 0.117)
Risk difference (95% CI) from IV probit	−0.031 (−0.187 to 0.125)
4 months
No. of outcomes/No. of total individuals	–
F-statistic from first-stage of IV 2SLS	23.85
Risk difference (95% CI) from IV 2SLS	−0.023 (−0.187 to 0.140)
Risk difference (95% CI) from IV probit	−0.022 (−0.192 to 0.149)
5 months
No. of outcomes/No. of total individuals	–
F-statistic from first-stage of IV 2SLS	25.27
Risk difference (95% CI) from IV 2SLS	−0.013 (−0.168 to 0.143)
Risk difference (95% CI) from IV probit	−0.011 (−0.173 to 0.151)
6 months
No. of outcomes/No. of total individuals	310/6582
F-statistic from first-stage of IV 2SLS	23.43
Risk difference (95% CI) from IV 2SLS	−0.012 (−0.182 to 0.159)
Risk difference (95% CI) from IV probit	−0.010 (−0.189 to 0.168)
1 year
No. of outcomes/No. of total individuals	302/6574
F-statistic from first-stage of IV 2SLS	23.02
Risk difference (95% CI) from IV 2SLS	−0.041 (−0.202 to 0.120)
Risk difference (95% CI) from IV probit	−0.042 (−0.210 to 0.126)
2 years
No. of outcomes/No. of total individuals	282/6554
F-statistic from first-stage of IV 2SLS	19.02
Risk difference (95% CI) from IV 2SLS	−0.032 (−0.186 to 0.123)
Risk difference (95% CI) from IV probit	−0.030 (−0.188 to 0.127)
3 years
No. of outcomes/No. of total individuals	249/6521
F-statistic from first-stage of IV 2SLS	14.68
Risk difference (95% CI) from IV 2SLS	−0.031 (−0.181 to 0.118)
Risk difference (95% CI) from IV probit	−0.031 (−0.183 to 0.121)
4 years
No. of outcomes/No. of total individuals	221/6493
F-statistic from first stage of IV 2SLS	10.62
Risk difference (95% CI) from IV 2SLS	−0.011 (−0.162 to 0.140)
Risk difference (95% CI) from IV probit	−0.012 (−0.164 to 0.139)

For 2, 3, 4 and 5 months windows, no. of outcomes and total observations not presented, as differences in no. of outcomes or observations across those windows are <5, leading to a secondary disclosure risk.

Analyses adjusted for sex, urban residence, deprivation, year of birth, low birth weight, winter birth, age at earliest adolescent depression record, year of earliest adolescent depression record, inpatient CAP service attendance prior to earliest adolescent depression record, SSRI prescription prior to earliest adolescent depression record, population size in GP cluster, percentage of GP population with a CAP service attendance, percentage of GP population from urban areas, percentage of GP population in each deprivation quintile. Analyses also accounted for clustering of observations at GP level. Those who had outcome diagnosis prior to intervention window were removed from the analysis.

The risk difference represents the absolute percentage-point change in the probability of the outcome per one-unit increase in cumulative SSRI exposure.

CAP, Child and Adolescent Psychiatry; GP, general practice; IV, instrumental variable; 2SLS, two-stage least squares; SSRI, selective serotonin reuptake inhibitor.

A valid IV requires four assumptions: relevance, independence, exclusion restriction and monotonicity.³⁰,³³ In online supplemental eMethods, we provide a detailed description and empirical examination of these assumptions through either direct or falsification tests.⁶⁷ As a measure of relevance (ie, strength of association between IV and exposure), we reported the Anderson-Rubin first-stage F statistic (table 3), computed with robust SEs to avoid size and power distortions associated with conventional F statistic estimation.⁶⁸

Analyses stratified by type of SSRI

We also performed the analyses stratified by type of SSRI. This was, however, only feasible for fluoxetine (online supplemental eTables 10–15, eFigures 13–18), using regional-level fluoxetine prescribing propensity as the IV. For sertraline and citalopram, we did not have sufficient statistical power for a stratified analysis.

Findings

Descriptive statistics

Our sample comprised 6615 (72.38% female) individuals registered in 224 GP clusters (table 1). Of them, 63.01% were exposed to SSRI prescriptions within 4 years of their earliest depression record. Median age at first adolescent depression record was 16.45 (IQR: 15.29–17.27) years. Sample characteristics were similar between exposed and unexposed individuals. Mania/hypomania was more prevalent in the exposed group (n=266, 6.38%) than in the unexposed group (n=77, 3.15%). We have also presented the cumulative absolute risk curves following each exposure window, stratified by the group exposed and the group unexposed within that window (online supplemental eFigures 19, 20).

Conventional (non-IV) analysis

Conventional analyses indicated that long-term exposure was associated with a greater risk of the outcome (table 2). The doubly-robust analysis demonstrated statistical evidence for this for all long-term exposure windows, but in the outcome regression and IPW analyses, there was no evidence for the 1 year window.

Instrumental variable analysis

Our IV—regional-level SSRI prescribing propensity—was associated with the exposure, with first-stage robust F statistic ranging from 10.62 to 25.27 across exposure windows (table 3). We found no relationship between the instrument and the outcome, either in the short- or long-term (table 3). Regional variation in SSRI prescribing propensity is presented in online supplemental eTable 6. Findings from the investigations of IV assumptions are described in online supplemental eMethods and presented in online supplemental eTables 7–9 and 12–15, eFigures 3–12.

Stratified analysis for fluoxetine prescription

Over one-third of participants (37.37%) received at least one fluoxetine prescription within 4 years of adolescent depression (online supplemental eTable 10). Conventional analyses showed that greater cumulative fluoxetine exposure was linked to higher risk of the outcome (online supplemental eTable 11). However, IV analyses—excluding the 1–4 months windows due to weak instruments (first-stage partial-F <10)—found no evidence of an association between fluoxetine exposure and the outcome (online supplemental eTable 13).

Discussion

Recent research has shown that a substantial proportion (up to half) of individuals who developed BD had, at some stage in childhood, attended child and adolescent psychiatry services, most frequently for treatment of depression and anxiety,^{21 22} typically involving SSRI medication. Given the high proportion of BD cases that emerge in patients of child and adolescent psychiatry services, it is crucial to carefully analyse for a potential causal relationship between SSRI treatment in this cohort and risk of subsequent BD.

Consistent with previous research, we observed an association between SSRI treatment and subsequent risk of mania/hypomania. Using more rigorous causal methods, however—specifically utilising an IV design leveraging regional variability in SSRI prescribing practices—we found no evidence to support a causal relationship between SSRI treatment in adolescence (either short- or long-term) and subsequent mania/hypomania.

Previous electronic health record research on antidepressant treatment of depression has reported an increased risk of subsequent mania.¹⁴,¹⁶¹⁸ These findings, however, were based on analyses that may be confounded. For example, adolescents prescribed SSRIs showed higher rates of new-onset mania in US claims data¹⁴ and UK health registry data¹⁶ through conventional analyses (assuming no unmeasured confounding). A recent Swedish registry study, which investigated mania risk in children treated with antidepressants, observed an increased risk when followed to 1 year.¹⁸ That study used inverse probability of treatment weighting, which applies weights to patient characteristics in an attempt to achieve comparability between treatment groups with respect to observed baseline characteristics. Registry studies, however, typically lack information on patients’ symptom severity, meaning this crucial difference cannot be accounted for, introducing the critical issue of confounding-by-indication. That is, the indication for treatment (depression severity), rather than the treatment itself (SSRIs), may be responsible for the outcome (mania). Specifically, it may be the case that individuals with more severe depression are more likely to develop mania, and it may also be the case that more severe depression is treated with SSRIs. This would create a spurious association between SSRI treatment and risk of mania, even in studies applying sophisticated observational analytic methods such as IPW.^{18 60}

An IV approach, on the other hand, has the potential to circumvent the issue of confounding-by-indication by inducing an as-if random variation in the treatment, provided certain assumptions hold.²⁸,³³³⁵ We evaluated these assumptions through a series of direct and falsification tests (discussed more in online supplemental eMethods). Although the assumption of relevance was met, the validity of the IV analysis also largely depends on the assumptions of exclusion restriction and conditional unconfoundedness, neither of which can be fully verified empirically.⁶⁷ While our testing provided falsification evidence supporting these assumptions—specifically, the lack of association between the IV and most measured confounders, and the null finding in the control sample reduced-form analysis—the possibility of unmeasured confounding or alternative indirect pathways remains. Regional prescribing variation may partly represent underlying differences in service capacity and provisions. This is, however, central to the IV’s construction: regional prescribing propensity operationalises a composite of clinician preference and service context that shapes antidepressant prescribing propensity. Adjusting the IV analyses for key regional-level factors (deprivation, urbanicity, psychiatric service attendance and service-level population burden) may have accounted for any major service-level variation embedded in the instrument. Furthermore, we performed a sensitivity IV analysis of the regional-level outcome and the regional-level instrument, clustering by region and adjusting for service indicators; this yielded consistent results (see online supplemental eTable 16). Findings from these assessments suggest that the regional-level SSRI prescribing propensity was likely a valid instrument for our investigation.

Our conventional (non-IV) analyses demonstrated that long-term SSRI treatment for adolescent depression was associated with an increased probability of subsequent mania/hypomania in the current cohort, consistent with previous observational studies.^{12 16 18 26} In subsequent IV analyses, however, we found no evidence to support a causal association. This suggests that the apparent relationship between SSRI treatment in adolescents and subsequent risk of BD demonstrated in the current (and previous) studies using non-IV methods is likely to be confounded, rather than reflecting a true causal relationship.

Strengths and limitations

We used electronic health records that cover approximately 86% of the Welsh population,⁶⁹ supporting the generalisability of our findings. The use of prospectively captured data reduced the risk of information bias. It furthermore allowed investigations of long-term and uncommon outcomes such as incident mania in youth. A quasi-experimental design through an IV analysis approach enabled us to estimate a causal association, overcoming bias from confounding-by-indication.

Limitations include the fact that a medication was prescribed does not necessarily mean it was consumed, which is an inherent limitation of all studies where medicine consumption is not directly supervised; it is likely, however, that individuals who received more prescriptions consumed more medication. A small subgroup (~5%) had SSRI prescriptions prior to time zero (ie, their earliest adolescent depression record). To account for potential differential risk carried into follow-up from past exposure, we adjusted for a binary indicator of prior SSRI use at baseline. While this mitigates—but does not entirely remove—prevalent-user bias, a sensitivity analysis restricted to new users (excluding individuals with pre-baseline exposure) produced results consistent with the main analysis (online supplemental eTable 17), suggesting limited influence on the overall estimates. Welsh registry data do not reliably differentiate between bipolar I and II diagnoses, meaning it was not possible for us to look at these diagnoses separately. Our findings apply specifically to adolescents; we could not investigate this question in younger children due to the low prevalence of SSRI prescription in this age group. As is typical of IV methods, our estimates represent a local average treatment effect, applying to the subgroup of adolescents whose antidepressant treatment status would be influenced by their regional prescribing propensity. Consequently, our results may not strictly generalise to the wider population of all adolescents with depression.

Conclusions

Our findings suggest that the apparent link between SSRI treatment of depression in youth and later risk of bipolar disorder observed in prior studies may be confounded by unmeasured factors, such as depression severity. Using an IV approach, we found no evidence to support a causal relationship between SSRI treatment in adolescents and subsequent risk of bipolar disorder, in both the short- and long-term. Given the high proportion of incident bipolar disorder cases at a population level that emerge in people who attend child and adolescent psychiatry services—most commonly for treatment of depression and anxiety—these findings provide important reassurance around SSRI safety in terms of bipolar disorder risk.

Data availability statement

Data may be obtained from a third party and are not publicly available.