Multivariable Prediction Modeling of Antidepressant Initiation in Unipolar Depressed Adolescents: A Secondary Analysis of the IMPACT Trial

Abstract

Introduction:

This secondary analysis of data collected in a randomized controlled trial (RCT) for the treatment of depression in adolescents aimed to test prediction models relating antidepressant (AD) initiation to clinical variables.

Methods:

The primary study was an RCT where adolescents (ages 11–17) with depression were assigned one of three outpatient psychotherapies over 86 weeks. The current study tested five registered prediction models using data on adolescents not taking ADs at baseline (N = 337). Outcomes of interest included: AD initiation, change in depression severity, and self-injurious thoughts and behaviors (SITBs).

Results:

Findings from registered analytic strategies were not consistent with our a priori hypotheses; rather we unexpectedly observed a relationship between initiation of AD and increased risk of suicide attempts and suicidal ideation during the same time interval (p > 0.01). Sensitivity analyses found that: (1) higher depressive symptom severity and self-harm each predicted future AD initiation (p < 0.05), and (2) new-onset SITB was associated with AD initiation (p < 0.01).

Conclusions:

Taken together, our results suggest that depression symptoms severity and SITBs may prompt AD initiation. Researchers may wish to further explore causal pathways relevant to the association ADs between SITBs. Clinicians need to be cognizant of high-quality guideline recommendations when prescribing ADs to adolescents.

Background

Worldwide, Major Depressive Disorder occurs in ∼4.5% and 3.0% of adolescent girls and boys, respectively (World Health Organization, 2017). It is the leading cause of global burden of disease in youth (Gore et al., 2011) and is a major factor in risk for death by suicide (Renaud et al., 2008).

The UK NICE guideline recommendations for the management of depression in children and young people have been appraised as high quality (Bennett et al., 2018). This guideline indicates that antidepressants (ADs) should only be prescribed for moderate-to-severe depression and given even more consideration if there is no response to initial psychosocial interventions (NICE, 2019). A meta-analysis of randomized controlled trials (RCTs) for adolescent depression reported that fluoxetine, in particular, is significantly effective in reducing symptoms of depression, although not all studies agree that this medication is of specific added value over a psychological treatment (Cipriani et al., 2016; Davey et al., 2019; Hetrick et al., 2021).

Self-injurious thoughts and behaviors (SITBs; Nock et al., 2007) are common in the context of depression in adolescents (Hawton et al., 2012). There have been multiple historical efforts to build consensus with respect to the SITB nomenclature (O'Carroll et al., 1996; Silverman et al., 2007a; Silverman et al., 2007b). Despite these efforts there continue to be discrepancies among experts in the field on how best to conceptualize SITBs (De Leo et al., 2021). For the purposes of this article, SITB encapsulates both suicidal ideation and self-harm; and self-harm includes both suicide attempts and nonsuicidal self-injury (NSSI). This conceptualization of self-harm is consistent with that found in the relevant NICE guideline (NICE, 2022).

Evidence suggests a complex relationship with respect to how ADs are associated with increased (Sharma et al., 2016) and/or decreased (Gibbons et al., 2006) risk of SITB. The UK NICE guideline on thef long-term management of self-harm (NICE, 2012) has also been appraised as high quality (Courtney et al., 2019). This guideline recommends, “Do not offer drug treatment as a specific intervention to reduce self-harm” (NICE, 2012). This recommendation was repeated in the newly released NICE guideline for the assessment and management of self-harm (NICE, 2022). This recommendation was supported by a recent Cochrane systematic review on interventions for the management self-harm (Witt et al., 2021). Few studies have examined predictors of AD initiation.

The IMPACT trial was a major RCT of adolescents with depression (ages 11–17 years old, N = 465) rolled out across three regions (East Anglia, North London, and the Northwest England), each providing five routine NHS clinics (n = 15) in the United Kingdom from whom all patients were recruited (Goodyer et al., 2017a). Detailed data were collected throughout the trial, including the initiation of ADs. The primary purpose of the trial was to test changes in symptom and functioning outcomes for participants assigned to one of the following planned programs: (1) up to 20 sessions of Cognitive–Behavioral Therapy (CBT), (2) up to 28 sessions of Short-Term Psychoanalytic Psychotherapy (STPP), or (3) up to 12 sessions of a manualized Brief Psychosocial Intervention (BPI). This was a pragmatic superiority trial with participants recruited between 2010 and 2013, intended to represent real-world clinical practice.

The main finding of the trial was that there were no significant differences found between treatment arms on the primary outcome of changes in self-reported depressive symptoms on the Mood and Feelings Questionnaire (MFQ) at a nominal 36, 52, and 86 weeks postrandomization. The rate of improvement was also not significantly different between treatment arms (Goodyer et al., 2017b). ADs, including fluoxetine, were allowed to be prescribed in any arm if clinical progress was judged to be not occurring at any point during the manualized treatment. Prescribing was therefore not a controlled arm nor determined by preplanned protocol rules before randomization.

The primary hypothesis was that the two specialist treatments (CBT, STPP) would be superior in reducing self-reported depression symptoms sum scores when compared with the reference treatment of BPI by end of study, which was 1 year after the end of treatment. The published findings showed that the null hypothesis was supported with no differences in symptom reduction between the treatment arms. Around 37% of patients in each arm were taking an AD (fluoxetine >85%) during psychological treatment. A full description of the trial can be found elsewhere (Goodyer et al., 2017b).

Leveraging baseline data from the IMPACT trial, Cousins et al. (2016) have performed a cross-sectional case–control study exploring clinical characteristics associated with a participant being prescribed an AD before randomization in the parent study. Of the original 465, 457 participants had baseline prescribing data coded. Of these, 89 (19.5%) had been prescribed ADs before randomization. Using multinomial logistic regression modeling, they reported that depression symptom severity in boys (odds ratio [OR]: 1.07, 95% confidence interval [CI]: 0.002–0.13 for each point increased on the MFQ) and self-harm in girls (OR: 1.03, 95% CI: 0.0001–0.065 for each point increase on the Risk Taking and Self-harm Inventory for Adolescents—Self-harm subscale) was associated with having been prescribed an AD. The investigators suggested a possible gender bias in prescribing patterns. More can be learnt with regard to longitudinal prescribing patterns and their relationship to symptom profiles, including symptom severity, SITB, and gender.

In the context of the IMPACT trial, we aimed to test hypotheses with respect to temporal relationships between clinical variables and AD prescription, which had not yet been investigated. We wanted to test hypotheses that:

(H1) initiation of ADs would be associated with a smaller reduction or worsening of depressive symptoms between the two prior time points, self-harm (relative to no self-harm) in girls and higher depression symptom severity in boys;

(H2) there would be a greater degree of symptom improvement in time intervals where AD medication is initiated for at least 4 weeks relative to times when AD medication is not initiated; and,

(H3–H5) the reduction in SITB between time points before and after AD initiation will be greater in those starting ADs relative to those not starting ADs.

Methods

Preregistered details of the analyses described in this study are available at Courtney 2022a. The hypotheses and analytic strategies were generated and registered after data collection was completed, but before any analyses examining relationships to these variables. For this trial, data were collected and analyzed as intention to treat on adolescents with depression in all three regions of the United Kingdom in the parent trial (East Anglia, North London, and the North-West). AD prescribing was collected by the Child and Adolescent Service Use Schedule (CASUS, Byford et al., 1999) as free text data by the research assistant.

The free text most often consisted of medication name, dose, and date of initiation. The free text was newly coded by the lead investigator as “initiated” or “not initiated” during the interval of interest, and whether there was exposure to the AD for 4 weeks before the relevant time point. Ambiguous free text data were counted as “missing.” Only participants who were coded as “not on antidepressants at baseline” are included in this analysis. Data collection time points included baseline (at randomization) and then at weeks 6, 12, 36, 52, and 86. Differences between those who were prescribed ADs at baseline versus those who were not prescribed ADs are detailed elsewhere (Cousins et al., 2016).

The 33-item MFQ (Angold and Costello, 1987) was used to measure depressive symptoms on a continuous scale. The range of the MFQ is 0–66; a score of >25 was used to represent clinically important depressive symptoms in the primary study (Goodyer et al., 2017b). The Columbia Suicide Severity Rating Scale (C-SSRS) (Posner et al., 2011) was used to assess the presence or absence of SITB at each time interval. The presence of a suicide attempt (i.e., “actual attempt” as opposed to “aborted attempted,” “interrupted attempt,” or “preparatory behavior”) or NSSI was captured by relevant single items on the C-SSRS. The presence of suicidal ideation was captured if the score was ≥2 on the ideation severity subscale (“nonspecific active suicidal thoughts”). Outcome assessors but not patients were blinded to randomized psychotherapy treatment arm.

Model 1a

To test hypothesis 1 (H1), we conducted logistic regression with mixed-effects modeling where the dependent variable was the initiation relative to noninitiation/continuation of an AD between the two most recent time points (k-1 and k). ADs included fluoxetine, sertraline, citalopram, escitalopram, paroxetine, fluvoxamine, venlafaxine, desvenlafaxine, or duloxetine. Approximate dates of initiation of AD medication were reported by patients and caregivers to research assistants at each follow-up assessment. Independent fixed variables in the model included age, gender (not differentiated from sex in the dataset), and region. Independent time-varying variables included time since randomization, change in MFQ score in the prior interval (from k-2 to k-1; representing extent of change in depressive symptoms), absolute MFQ score at the prior time point (k-1; representing depression severity), and its interaction with gender; as well as self-harm at the prior interval (k-2 to k-1; representing NSSI or suicide attempt) and its interaction with gender.

Model 1b

After viewing the results of this analysis, we also conducted a registered sensitivity analysis Cox regression to test H1, with discrete time to initiation of AD as the outcome of interest (Courtney, 2022b). Time of origin is the time of enrollment in the study, which is highly correlated with time of initiating psychotherapy and assumed to be associated with time of initiation of help seeking. Predictor variables include gender, time-varying depression symptom severity (MFQ), and time-varying self-harm (C-SSRS). This analysis was adjusted for presence or absence of youth-informant of any anxiety-related diagnoses (including generalized anxiety disorder, panic disorder, social phobia, obsessive-compulsive disorder, posttraumatic stress disorder, separation anxiety disorder, and/or agoraphobia; excluding specific phobia) at baseline as these may be other reasons for prescribing ADs during the trial.

Model 1c

To further clarify results, we conducted a nonregistered sensitivity analysis with gender interaction terms dropped from Model 1b.

Model 2

To test hypothesis 2 (H2), we conducted linear regression mixed-effects modeling with change in MFQ score in the previous interval (k-1 to k) as a continuous dependent variable. Fixed effects included age, gender, and region. The time-varying variables included time since randomization, and a composite variable where both an AD was initiated in the previous interval (k-1 to k) and reported exposure to an AD for at least 4 weeks before the time point of interest.

Models 3–5a

To test hypotheses 3–5 (H3–H5), we conducted three logistic regression analyses with mixed-effects modeling. For each of the models, the dependent variables were presence/absence of a suicide attempt (Model 3), NSSI (Model 4), and suicidal ideation (Model 5a) in the previous interval (k-1 to k), respectively. Fixed effects were identical to those analyzed testing H2. The time-varying variables were time since randomization and exposure to an AD in the recent interval (k-1 to k).

Model 5b

After seeing the results of the analyses for H3 to H5, it was decided to conduct a nonregistered sensitivity analysis, where the dependent variable was new onset of any SITB (i.e., suicide attempt, NSSI, or suicidal ideation).

For all mixed-effects models, random effects were included in the model to account for correlated outcomes at the level of the individual participant over time as well as correlated outcomes at the level of the site. Each mixed-effects model was tested for significance at the level of p < 0.01 significance (0.05 divided by 5 hypotheses) to account for multiple testing (Staffa and Zurakowski, 2020). The Cox regression sensitivity analysis was also tested for significance at p < 0.01. Two-tailed tests were used. The linearity assumptions of the proposed models were checked using graphical output and existence of high influential cases was examined by visualizing the Cook's distance values.

There were some deviations from the registered analyses. Time since starting therapy was originally planned to be a covariate in the mixed-effects models; however, this was very highly correlated to time since randomization, and was therefore used interchangeably in the models. Region was initially planned as a random effect; however, the models did not converge, so was changed to a fixed effect. AD initiation in the recent interval and exposure to the AD in the 4 weeks before the time point were initially planned as separate independent variables for H2–H5; however, their combined occurrence is more relevant to the hypotheses and was therefore merged into one variable, “Initiation of AD in recent interval (k-1 to k) AND AD exposure over 4 weeks just before time point k” for these tests; if both conditions were met, this variable was coded as “1,” if not the variable was coded as “0.” Nesting by therapist was also planned; however, a significant minority of youth participants had unique therapists, rendering this analysis component infeasible. Sex and gender were not differentiated in this trial; this report will use the term “gender” and “girl/boy” to describe relevant differences.

Multiple imputation was conducted on the raw data (M = 40) for each of the mixed-effects models using Mplus version 8.3. Auxiliary variables included sex, score on the Risk Taking and Self-harm Inventory for Adolescents (Vrouva et al., 2010), the Revised Children's Manifest Anxiety Scale (Reynolds and Paget, 1983), and the Rosenberg Self-Esteem Scale (Rosenberg, 1965) at all time points. Censoring was used to manage missing data in the Cox regression analysis.

The original trial was registered with Current Controlled Trials, number ISRCTN83033550. The secondary analyses were registered with As Predicted, numbers 15291 and 74196.

The study was approved by the Cambridgeshire 2 Research Ethics Committee (reference 09/H0308/137) and all the local NHS provider trusts in each region and all participants and caregivers (e.g., parents) signed written informed consent.

Results

The original sample includes 465 participants. Forty-one had missing or conflicting information about being on ADs at baseline. Of the remaining 424, 87 were classified as having been on ADs at baseline. The resulting sample analyzed here included 337 participants. Table 1 outlines clinical variables with respect to available data, before imputation.

Table 1.

Clinical Characteristics of the Sample (N = 337)

Variable	Mean (SD)	n/N with available data (%)
Age (years)	15.6 (1.4)
Girl gender		261/337 (77)
Anxiety disorder		128/337 (38)
At randomization, lifetime history of SITB
Suicide attempt		112/327 (34)
NSSI		169/292 (58)
Suicidal ideation		237/331 (71)
MFQ score
At randomization (n = 337)	45.5 (10.3)
Week 6 (n = 227)	35.8 (12.7)
Week 12 (n = 242)	33.6 (13.8)
Week 36 (n = 233)	27.1 (16.1)
Week 52 (n = 239)	24.8 (16.7)
Week 86 (n = 259)	23.7 (16.1)

MFQ, Mood and Feelings Questionnaire; NSSI, nonsuicidal self-injury; SITB, self-injurious thoughts and behaviors.

Over the course of the trial, 66 of 337 of participants newly started an AD (18%; 50 [76%] of these were girls). Of these 66, 41 (62%) remained on the AD for the 4 weeks before the associated time point (30/41, or 73%, were girls). Of these 41, initiation of the AD occurred by 6, 12, 36, 52, and 86 weeks for 20, 4, 10, 2, and 5 participants, respectively. Twelve of the 41 were randomized to BPI, 13 to CBT, and 16 to STPP. Thirty of the 41 were prescribed fluoxetine, 5 sertraline, 4 citalopram, and 1 amitriptyline. Overall, in this subsample, 77 had a new-onset SITB (64 were girls), with 29 having had a first ever suicide attempt (23 were girls), 32 with new-onset NSSI (29 were girls), and 28 with new-onset suicidal ideation (20 were girls) throughout the 86-week data collection period.

Table 2 describes the findings from our original a priori hypotheses. None of the hypothesized independent variables predicted the outcome of interest at a threshold of p < 0.01; in fact, in testing H3 and H5, we observed that more suicide attempts and suicidal ideation in the interval k-1 to k was associated with increased odds of AD initiation in the same interval at p < 0.01; this is the opposite direction of what was hypothesized.

Table 2.

Linear Mixed Effects Models Testing Hypotheses Regarding the Relationships Between Antidepressant Initiation and Clinical Variables (Depression Symptom Severity and Self-Injurious Thoughts and Behaviors) As Well As Relevant Sensitivity Analyses

Model 1a (registered analysis for H1): logarithm of the odds of AD initiation during recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	−3.51	1.72	−6.88 to −0.14	0.04
Age	0.01	0.10	−0.19 to 0.22	0.89
Gender (girl coded as “1”)	0.19	1.20	−2.16 to 2.54	0.87
Time since randomization	0.0007	0.0008	−0.0009 to 0.0023	0.39
MFQ at prior time point (k-1)	0.011	0.023	−0.033 to 0.056	0.62
Self-harm in prior interval (k-2 to k-1)	0.073	0.86	−1.62 to 1.77	0.93
Region
2	−0.11	0.37	−0.83 to 0.61	0.77
3	−0.12	0.33	−0.77 to 0.53	0.72
Change in MFQ in prior interval (k-2 to k-1)	−0.002	0.013	−0.029 to 0.025	0.91
Gender × MFQ at prior time point (k-1) interaction	0.004	0.025	−0.045 to 0.053	0.86
Gender × self-harm in prior interval (k-2 to k-1) interaction	−0.40	0.90	−2.16 to 1.36	0.66

Model 2 (registered analysis for H2): change in MFQ in recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	6.25	3.60	−0.79 to 13.3	0.08
Age	0.11	0.23	−0.34 to 0.55	0.63
Gender (girl coded as “1”)	−0.41	0.77	−1.92 to 1.10	0.59
Time since randomization	−0.009	0.002	−0.013 to −0.006	<0.001
Region
2	0.02	0.82	−1.59 to 1.63	0.98
3	−0.96	0.74	−2.42 to 0.50	0.20
Initiation of AD in recent interval (k-1 to k) and AD exposure over 4 weeks just before time point k	−0.04	1.55	−3.08 to 3.00	0.98

Model 3 (registered analysis for H3): logarithm of the odds of at least 1 suicide attempt in recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	−3.87	1.90	−7.61 to −0.13	0.04
Age	0.06	0.12	−0.17 to 0.28	0.64
Gender (girl coded as “1”)	0.50	0.40	−0.28 to 1.27	0.21
Time since randomization	−0.001	0.001	−0.002 to 0.000	0.17
Region
2	−0.19	0.43	−1.04 to 0.66	0.66
3	−0.44	0.41	−1.25 to 0.35	0.27
Initiation of AD in recent interval (k-1 to k) and AD exposure over 4 weeks just before time point k	1.16	0.44	0.29 to 2.03	0.009^**

Model 4 (registered analysis for H4): logarithm of the odds of NSSI in recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	−0.018	2.18	−4.31 to 4.27	0.99
Age	−0.079	0.14	−0.35 to 0.19	0.57
Gender (girl coded as “1”)	1.09	0.49	0.13 to 2.06	0.03
Time since randomization	−0.002	0.001	−0.003 to −0.001	0.002
Region
2	−1.09	0.54	−2.16 to −0.03	0.045
3	−2.05	0.53	−3.11 to −0.96	0.00
Initiation of AD in recent interval (k-1 to k) and AD exposure over 4 weeks just before time point k	0.69	0.42	−0.13 to 1.5	0.10

Model 5a (registered analysis for H5): logarithm of the odds of suicidal ideation in recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	−1.94	1.55	−5.00 to 1.11	0.21
Age	0.06	0.10	−0.13 to 0.26	0.52
Gender (girl coded as “1”)	0.55	0.33	−0.10 to 1.19	0.10
Time since randomization	−0.001	0.0004	−0.002 to −0.001	0.00
Region
2	−0.35	0.34	−1.03 to 0.33	0.31
3	−0.72	0.32	−1.36 to −0.08	0.03
Initiation of AD in recent interval (k-1 to k) and AD exposure over 4 weeks just before time point k	1.34	0.37	−0.002 to −0.001	0.000^**

Model 5b (nonregistered sensitivity analysis related to H3–H5): logarithm of the odds new-onset self-injurious thought and behaviors in recent interval (k-1 to k)	ß	SE	95% CI	p
Constant	−1.56	1.35	−4.22 to 1.10	0.25
Age	−0.75	0.85	−0.24 to 0.09	0.39
Gender (girl coded as “1”)	0.28	0.30	−0.31 to 0.88	0.36
Time since randomization	−0.003	0.00	−0.003 to −0.001	0.00
Region
2	0.75	0.32	0.13 to 1.38	0.02
3	0.59	0.32	−0.04 to 1.21	0.07
Initiation of AD in recent interval (k-1 to k) and AD exposure over 4 weeks just before time point k	1.02	0.37	0.30 to 1.75	0.005^**

Values in bold indicate independent variable of interest.

p < 0.01.

AD, antidepressant; CI, confidence interval; SE, standard error.

Table 3 describes findings from H1 sensitivity analyses using Cox regression analysis to predict AD initiation. Our analysis did not confirm our registered prediction model for H1. The nonregistered Cox regression analysis with the gender interaction terms dropped did reveal that higher depression symptom severity at the prior time point and self-harm in the prior interval predicted AD initiation at a two-tailed threshold of p < 0.05.

Table 3.

Cox Regression Analysis as a Sensitivity Analysis Examining the Relationship Between Clinical Variables and Time to Antidepressant Initiation

Model 1b (registered sensitivity analysis for H1): hazard of time to AD initiation	ß	SE	HR	95% CI of HR	p
Constant	−3.73	1.14	0.024	0.001–0.39	0.01
Age	0.04	0.08	1.04	0.88–1.22	0.63
Gender (girl coded as “1”)	−0.03	0.66	0.97	0.27–3.55	0.97
Anxiety disorder	0.31	0.21	1.37	0.90–2.08	0.14
Time in therapy	0.000	0.00	1.00	1.00–1.00	0.99
Region	−0.198	0.13	0.82	0.63–1.07	0.14
MFQ at prior time point (k-1)	0.03	0.02	1.03	0.99–1.07	0.18
Self-harm in prior interval (k-2 to k-1)	−0.004	0.66	0.98	0.27–3.63	1.00
Gender × MFQ total at prior time point (k-1)	−0.011	0.02	0.99	0.95–1.03	0.61
Gender × self-harm in prior interval (k-2 to k-1)	0.806	0.71	2.24	0.56–8.93	0.26

Model 1c (nonregistered sensitivity analysis for H1): hazard of time to AD initiation with interaction terms removed	ß	SE	HR	95% CI of HR	p
Constant	−3.50	1.37	0.03	0.002–0.445	0.01
Age	0.03	0.08	1.03	0.88–1.21	0.72
Gender (girl coded as “1”)	−0.06	2.81	0.94	0.54–1.63	0.83
Anxiety disorder	0.31	0.21	1.36	0.89–2.06	0.16
Time in therapy	0.00	0.00	1.00	0.99–1.00	0.99
Region	−0.20	0.13	0.82	0.63–1.06	0.13
MFQ at prior time point (k-1)	0.02	0.01	1.02	1.00–1.04	0.04^*
Self-harm in prior interval (k-2 to k-1)	0.66	0.28	1.93	1.11–3.35	0.02^*

Values in bold indicate independent variable of interest.

p < 0.05.

Discussion

This secondary analysis examined models of the relationship between AD initiation and clinical factors, namely sex/gender, depression symptom severity, and SITBs. None of our results were consistent with our a priori modeled hypotheses; although when the gender terms were dropped in sensitivity analyses, we observed some patterns of interest.

The relationship between depression symptom severity and AD initiation was expected to be bidirectional with the temporal relationship being important; higher severity leading to greater odds of subsequent AD initiation (included in H1), whereas AD initiation with ≥4 weeks exposure was expected to be associated with improvement of depressive symptoms (included in H2). Models 1a and 1b did not support the first part of the expected relationship, nor any interaction with gender. These results are in contrast with Cousins et al. (2016) finding of an interaction between depression severity and gender and taking ADs. One possible explanation is that physicians may use different decisional processes around medication prescription knowing that a youth is receiving psychotherapy (as was the case in the current sample) versus before the initiation of therapy (as was the case in Cousins et al. sample).

Only when the gender interaction terms were dropped in the unregistered sensitivity Cox regression analysis (Model 1c) did we find that greater depression severity scores predicted subsequent AD prescription, that is, we did not observe evidence of prescribing biases with respect to gender. This latter finding is in keeping with the NICE guideline recommendation to use depression symptom severity to guide decision making around AD initiation (NICE, 2019).

The NICE guideline recommendations state, “Following multidisciplinary review, offer fluoxetine if moderate to severe depression in a young person (12–18 years) is unresponsive to a specific psychological therapy after 4 to 6 sessions.” (NICE, 2019). It is notable that at the 36-week time point, the mean MFQ score was 27.1 for the 233 participants with available data, that is, more than half of the remaining sample was above the clinical cutoff. By contrast, only two trial participants were newly started on ADs between weeks 36 and 52, suggesting a missed opportunity to further optimize treatment consistent with guidelines recommendations. Despite this apparent omission, we also observed continued reductions in MFQ scores at weeks 52 and 86 in the sample, challenging the importance of the above guidelines recommendation.

Treatment arm was not included in the models as previous results showed no difference in number of participants on ADs throughout the trial by treatment arm (Goodyer et al., 2017b). Differential speed of improvement between treatment arms may have also affected prescribing patterns; however, a recent trajectory analysis of the IMPACT data suggests no difference in the speed of improvement in each arm (Davies et al., 2020), supporting the omission of the “therapy type” variable from the model. Moreover, Model 2 did not support the second part of the bidirectional relationship, where initiation and ≥4 weeks of exposure to an AD was not significantly associated with improvement in MFQ scores in the same time interval. Although this relationship is interesting to examine in a naturalistic context, establishment of efficacy is better suited for RCTs specifically designed to test this hypothesis.

We also examined temporal relationships between AD initiation and SITB (both as a composite outcome, and more specific outcomes); expecting to find that the presence of self-harm predicts subsequent AD initiation and AD initiation with ≥4 weeks of exposure would be associated with absence of SITB. When gender interaction terms were included in the model, we did not observe that the presence of self-harm in the prior interval predicted AD initiation (Models 1a and 1b). Of interest, when the gender interaction terms were dropped (as in Model 1c), the presence of self-harm did predict subsequent AD initiation, even when controlling for depression symptom severity. This relationship was also unexpectedly supported in Model 3 where AD initiation and ≥4 weeks exposure was associated with the occurrence of at least one suicide attempt in the same interval.

These findings are partially consistent with the findings of Cousins et al. (2016) that self-harm was associated with taking ADs; however, their analyses only found this relationship in girls, whereas our analysis did not observe effect modification by gender. That is, we again did not observe evidence of prescribing bias by gender. Taken together, these results suggest that clinicians may be starting ADs in response to a self-harm event, in contrast with the NICE guideline recommendation to “not offer drug treatment as a specific intervention to reduce self-harm” owing to lack of evidence for this approach (NICE, 2022; NICE, 2012). Further work on whether clinicians are using self-harm to make decisions around ADs is needed to clarify this relationship.

There are other potential drivers to consider here: for example, it is possible that new-onset SITB prompts adolescents or their caregivers to be more willing to start ADs when it was already indicated for their depressive illness; or may prompt caregivers to request that the adolescent start an AD in the belief that SITBs indicate a worsening of their clinical state overall. Integrated care pathways and measurement-based care are potential facilitators of ensuring that care is following guidelines recommendations (Courtney et al., 2020).

There are several strengths to this study. First, we developed our registered analytic strategies before examining the longitudinal relationships from these variables (though after data collection). These strategies were based on findings from prior work regarding baseline data in the entire IMPACT sample. Second, we used a Bonferroni correction for our registered analytic strategies to limit the possibility of a spurious finding owing to multiple testing. The strategies to reduce vulnerability to bias were either not applied or not reported in most other secondary analyses of RCTs for the treatment of depression in adolescents (Courtney et al., 2022). Some limitations are also important to consider. First, coding of AD prescribing at baseline was discrepant from Cousins et al. (2016); where they identified 8 participants with missing data on this variable, and the current study identified 41 as missing/ambiguous.

Differences between the studies may reflect a more conservative approach with more participants with ambiguous data classified as “missing” in the present article. The CASUS tool may benefit from further refinement to increase consistency of data collection and interpretation. Next, data on failed AD trials before the trial were not consistently collected and omitted from this analysis as a result. This could have helped contextualize our findings. Furthermore, our nonsignificant findings may be owing to inadequate power with respect to each variable (e.g., limited number of AD initiations and SITB variables), or time intervals that are too long to assess these relationships (e.g., the interval between the last two time points is 34 weeks).

Moreover, the outcomes used in the prespecified models for H3 to H5 reflect presence or absence of SITB, which is a proxy for “reduction of SITB” described in the original hypotheses. The lack of precision in the data, including variable time intervals in which the SITB data were collected, did not allow for an accurate analysis of change in SITB. Findings here would benefit from validation in individual patient data meta-analyses. Regardless, the analyses that were performed do shed light on the questions we were aiming to answer.

Conclusion

When examining the analyses all together, we found mixed evidence that clinicians are using depression symptom severity (consistent with NICE guideline recommendations) and SITBs (in contrast to NICE guideline recommendations) to make decisions around AD prescriptions in adolescents with depression. Future studies can be designed to examine these relationships in more detail, including the direction of causal pathways and drivers of this relationship.

Clinical Significance

In our secondary data analysis of an RCT for the treatment of adolescents with depression receiving psychotherapy, we found evidence that depression symptom severity (but not worsening, or stagnation of response) predicts AD initiation. We did not find evidence that AD initiation was associated with later improvement in symptoms. We observed an unexpected relationship where SITBs were associated with later AD initiation. This result is contrasted to the NICE guideline recommendation to avoid prescribing medications to target self-harm. Clinicians need to be cognizant of guidelines recommendations around the use of medications and further management of self-harm.

Footnotes

Acknowledgments

The authors thank Shirley Reynolds for providing input on how to code and analyze the data. The authors also thank Sarah Hetrick and Allan Young for their reflections and feedback on the article. Funding for the original IMPACT trial was provided by National Institute Health Research, United Kingdom. Funding for this secondary analysis was provided by the Cundill Centre for Child and Youth Depression.

Authors' Contributions

D.C.: Conceptualization (equal), methodology (equal), writing original draft (lead), analysis (equal), review and editing (equal). M.A.: Methodology (equal), analysis (equal), review and editing (equal). W.W.: Methodology (equal), analysis (equal), review and editing (equal). S.C.: Methodology (equal), analysis (equal), review and editing (equal). P.W.: Methodology (equal), review and editing (equal). P.S.: Conceptualization (equal), methodology (equal), review and editing (equal). I.G.: Conceptualization (equal), methodology (equal), review and editing (equal).

Data Availability Statement

Data available on request due to privacy/ethical restrictions.

Disclosures

No competing financial interests exist.

References

Angold

, Costello

EJ.

Mood and Feelings Questionnaire (MFQ). Duke University: Developmental Epidemiology Program; 1987. Available from: https://devepi.duhs.duke.edu/measures/the-mood-and-feelings-questionnaire-mfq/ [Last accessed: October 20, 2022].

Bennett

, Courtney

, Duda

, et al. An appraisal of the trustworthiness of practice guidelines for depression and anxiety in children and youth. Depress Anxiety, 2018; 35(6):530–540; doi: https://doi.org/10.1002/da.22752

Byford

, Harrington

, Torgerson

, et al. Cost-effectiveness analysis of a home-based social work intervention for children and adolescents who have deliberately poisoned themselves. Results of a randomised controlled trial. Br J Psychiatry, 1999; 174:56–62; doi:10.1192/bjp.174.1.56

Cipriani

, Zhou

, Del Giovane

, et al. Comparative efficacy and tolerability of antidepressants for major depressive disorder in children and adolescents: A network meta-analysis. Lancet, 2016; 388(10047):881–890; doi: https://doi.org/10.1016/S0140-6736(16)30385-3

Courtney

Antidepressants in the IMPACT RCT: A Secondary Analysis. (#15291) University of Pennsylvania, Philadelphia, Pennsylvania; 2022a. Available from: https://aspredicted.org/mu382.pdf [Last accessed: April 5, 2023].

Courtney

Prescription patterns in IMPACT Study; Survival Curve Analysis. (#74196). University of Pennsylvania, Philadelphia, Pennsylvania; 2022b. Available from: https://aspredicted.org/ps53w.pdf [Last accessed: April 5, 2023].

Courtney

, Bennett

, Henderson

, et al. A way through the woods: Development of an integrated care pathway for adolescents with depression. Early Interv Psychiatry, 2020; 14(4):486–494; doi: https://doi.org/10.1111/eip.12918

Courtney

, Duda

, Szatmari

, et al. Systematic review and quality appraisal of practice guidelines for self-harm in children and adolescents. Suicide Life Threat Behav, 2019; 49(3):707–723; doi: https://doi.org/10.1111/sltb.12466

Courtney

, Watson

, Krause

, et al. Predictors, moderators, and mediators associated with treatment outcome in randomized clinical trials among adolescents with depression. JAMA Netw Open, 2022; 5(2):e2146331; doi: https://doi.org/10.1001/jamanetworkopen.2021.46331

10.

Cousins

, Whitaker

, Widmer

, et al. Clinical characteristics associated with the prescribing of SSRI medication in adolescents with major unipolar depression. Eur Child Adolesc Psychiatry, 2016; 25(12):1287–1295; doi: https://doi.org/10.1007/s00787-016-0849-y

11.

Davey

, Chanen

, Hetrick

, et al. The addition of fluoxetine to cognitive behavioural therapy for youth depression (YoDA-C): A randomised, double-blind, placebo-controlled, multicentre clinical trial. Lancet Psychiatry, 2019; 6(9):735–744; doi: https://doi.org/10.1016/s2215-0366(19)30215-9

12.

Davies

, Neufeld

, van Sprang

, et al. Trajectories of depression symptom change during and following treatment in adolescents with unipolar major depression. J Child Psychol Psychiatry, 2020; 61(5):565–574; doi:10.1111/jcpp.13145.

13.

De Leo

, Goodfellow

, Silverman

, et al. International study of definitions of English-language terms for suicidal behaviours: A survey exploring preferred terminology. BMJ Open, 2021; 11(2):e043409; doi: https://doi.org/10.1136/bmjopen-2020-043409

14.

Gibbons

, Hur

, Bhaumik

, et al. The relationship between antidepressant prescription rates and rate of early adolescent suicide. Am J Psychiatry, 2006; 163(11):1898–1904; doi: https://doi.org/10.1176/ajp.2006.163.11.1898

15.

Goodyer

, Reynolds

, Barrett

, et al. Cognitive behavioural therapy and Short-Term Psychoanalytical Psychotherapy versus a brief psychosocial intervention in adolescents with unipolar major depressive disorder (IMPACT): A multicentre, pragmatic, observer-blind, randomised controlled trial. Lancet Psychiatry, 2017a;4(2):109–119; doi: https://doi.org/10.1016/S2215-0366(16)30378-9

16.

Goodyer

, Reynolds

, Barrett

, et al. Cognitive-behavioural therapy and Short-Term Psychoanalytic Psychotherapy versus brief psychosocial intervention in adolescents with unipolar major depression (IMPACT): A multicentre, pragmatic, observer-blind, randomised controlled trial. Health Technol Assess, 2017b;21(12):1–94; doi: https://doi.org/10.3310/hta21120

17.

Gore

, Bloem

PJN

, Patton

, et al. Global burden of disease in young people aged 10–24 years: A systematic analysis. Lancet, 2011; 377(9783):2093–2102; doi: https://doi.org/10.1016/S0140-6736(11)60512-6

18.

Hawton

, Saunders

, O'Connor

RC.

Self-harm and suicide in adolescents. Lancet, 2012; 379(9834):2373–2382; doi: https://doi.org/10.1016/S0140-6736(12)60322-5

19.

Hetrick

, McKenzie

, Bailey

, et al. New generation antidepressants for depression in children and adolescents: A network meta-analysis. Cochrane Database Syst Rev, 2021; 5:CD013674; doi: https://doi.org/10.1002/14651858.cd013674.pub2

20.

NICE: National Institute for Health and Care Excellence. Depression in Children and Young People: Identification and Management. NICE Guideline [NG134]; 2019. Available from: https://www.nice.org.uk/guidance/ng134 [Last accessed: October 20, 2022].

21.

NICE: National Institute for Health and Care Excellence. Self-Harm: Longer-term Management. National Clinical Guideline Number 133. British Psychological Society (UK): Leicester, UK; 2012.

22.

NICE: National Institute for Health and Care Excellence. Self-harm: Assessment, Management and Preventing Recurrence [NG 225]; 2022. Available from: https://www.nice.org.uk/guidance/ng225/chapter/Recommendations [Last accessed: October 20, 2022].

23.

Nock

, Holmberg

, Photos

, et al. Self-injurious thoughts and behaviors interview: Development, reliability, and validity in an adolescent sample. Psychol Assess, 2007; 19(3):309–317; doi: https://doi.org/10.1037/1040-3590.19.3.309

24.

O'Carroll

, Berman

, Maris

, et al. Beyond the Tower of Babel: A nomenclature for suicidology. Suicide Life Threat Behav, 1996; 26(3):237–252; doi: https://doi.org/10.1111/j.1943-278X.1996.tb00609.x

25.

Posner

, Brown

, Stanley

, et al. The Columbia–Suicide Severity Rating Scale: Initial validity and internal consistency findings from three multisite studies with adolescents and adults. Am J Psychiatry, 2011; 168(12):1266–1277; doi: https://doi.org/10.1176/appi.ajp.2011.10111704

26.

Renaud

, Berlim

, McGirr

, et al. Current psychiatric morbidity, aggression/impulsivity, and personality dimensions in child and adolescent suicide: A case-control study. J Affect Disord, 2008; 105(1–3):221–228; doi: https://doi.org/10.1016/j.jad.2007.05.013

27.

Reynolds

, Paget

KD.

National normative and reliability data for the Revised Children's Manifest Anxiety Scale. School Psych Rev, 1983; 12(3):324–336; doi: https://doi.org/10.1080/02796015.1983.12085047

28.

Rosenberg

Society and the Adolescent Self-Image. Princeton University Press: Princeton, NJ; 1965. Available from: https://doi.org/10.1515/9781400876136 [Last accessed: October 20, 2022].

29.

Sharma

, Guski

, Freund

, et al. Suicidality and aggression during antidepressant treatment: Systematic review and meta-analyses based on clinical study reports. BMJ, 2016; 352:i65; doi: https://doi.org/10.1136/bmj.i65

30.

Silverman

, Berman

, Sanddal

, et al. Rebuilding the Tower of Babel: A revised nomenclature for the study of suicide and suicidal behaviors. Part 1: Background, rationale, and methodology. Suicide Life Threat Behav, 2007a;37(3):248–263; doi: https://doi.org/10.1521/suli.2007.37.3.248

31.

Silverman

, Berman

, Sanddal

, et al. Rebuilding the Tower of Babel: A revised nomenclature for the study of suicide and suicidal behaviors. Part 2: Suicide-related ideations, communications, and behaviors. Suicide Life Threat Behav, 2007b;37(3):264–277; doi: https://doi.org/10.1521/suli.2007.37.3.264

32.

Staffa

, Zurakowski

Strategies in adjusting for multiple comparisons: A primer for pediatric surgeons. J Pediatr Surg, 2020; 55(9):1699–1705; doi: https://doi.org/10.1016/j.jpedsurg.2020.01.003

33.

Vrouva

, Fonagy

, Fearon

PRM

, et al. The risk-taking and self-harm inventory for adolescents: Development and psychometric evaluation. Psychol Assess, 2010; 22(4):852–865; doi: https://doi.org/10.1037/a0020583

34.

Witt

, Hetrick

, Rajaram

, et al. Interventions for self-harm in children and adolescents. Cochrane Database Syst Rev, 2021; 3(3):CD013667; doi: https://doi.org/10.1002/14651858.cd013667.pub2

35.

World Health Organization. Depression and Other Common Mental Disorders: Global Health Estimates. World Health Organization: Geneva; 2017; pp. 1–24. Available from: https://apps.who.int/iris/handle/10665/254610 [Last accessed: October 20, 2022].