Predictors, Moderators, and Mediators Associated With Treatment Outcome in Randomized Clinical Trials Among Adolescents With Depression

Key Points Question What do randomized clinical trials for treatment of adolescents with depressive disorders report about the predictors, moderators, and mediators associated with outcomes? Findings In this scoping review of 33 RCTs with results described across 81 publications, variable domains reported as significant in at least 3 RCTs with respect to depression outcomes included age, sex/gender, baseline depression severity, early response to treatment, sleep changes, parent-child conflict, overall psychopathology, suicidal ideation, hopelessness, functional impairment, attendance at psychotherapy sessions, and history of trauma. A small minority of publications reported efforts to minimize bias through a priori hypothesis testing and adjustment for multiple comparisons. Meaning Variables identified in this review can be incorporated into rigorous research designs to further test the optimization of care for adolescents with depression.


Introduction
Depressive disorders in adolescents (DD-A) are prevalent, 1 impairing, 2,3 and associated with suicide. 4,5In the United States, rates of depressive symptoms and suicide in adolescents have increased over the past 10 to 15 years. 6,7Current treatment approaches have limited benefit. 8,9The application of precision medicine hopes to improve outcomes by offering "treatment strategies that take individual variability into account." 10Clinicians treating DD-A are expected to practice precision medicine. 11Researchers of DD-A treatment also need to understand how clinical factors are associated with outcomes to guide the development and testing of new treatment approaches. 12formation from randomized clinical trials (RCTs) can elucidate these factors, as rigorous data collection is conducted at set time points under controlled treatment conditions.A good understanding of the variables associated with depression severity outcome in RCTs for the treatment of DD-A can indicate what works for whom and how. 12,13Two previous evidence syntheses that have examined such variables 14,15 have included a very limited set of relevant studies.
An up-to-date, systematic, and comprehensive examination of what is currently known about predictors, moderators, and mediators derived from RCTs for the treatment of DD-A can inform the extent to which clinicians can practice precision medicine and guide trialists on interventions targeting specific mechanisms of action.The aims of this scoping review were to (1) identify the predictors, moderators, and mediators that have been studied to date in published RCTs of treatment for DD-A, (2) map out the reported findings from their analyses to guide further hypothesis testing, and (3) describe the extent to which a priori hypotheses and adjustments for multiple comparisons are reported in published analyses.

Methods
Methods are detailed in our registration, 16 preprint, 17 and published protocol. 18,19A scoping review design was chosen for this study's aims, as it is most appropriate for mapping out broad concepts, describing the extent of the available literature, and identifying gaps. 20,21This is in contrast to a systematic review design, intended to summarize the literature to answer specific questions, or meta-analyses, in which data are reanalyzed. 21We applied scoping review methods outlined by the Joanna Briggs Institute 21 and followed the Preferred Reporting Items for Systematic Reviews and Meta-analyses Extension for Scoping Reviews (PRISMA-ScR) checklist for scoping reviews. 20clusion criteria were English-language RCTs that assessed treatments of depressive disorders in adolescents (ages 13-17 years) in which (1) depression was defined as diagnoses of major depressive disorder, dysthymia/persistent depressive disorder, or depressive symptoms more severe than an established cutoff on a validated measure of depression symptom severity; (2) treatment interventions included biological interventions (eg, antidepressants), psychosocial interventions (eg, psychotherapy), or service delivery models (eg, collaborative care arrangements between mental health specialists and primary care); (3) a test of any predictor, moderator, or mediator associated with depression outcomes was conducted.Exclusion criteria were RCTs evaluating bipolar depression, peripartum depression, premenstrual dysphoria, minor depression, or seasonal affective disorder; RCTs targeting the prevention of depression or recurrence of depression; and economic analyses.Additional exclusion criteria, which were protocol deviations, were conference abstracts, dissertations, and studies with sample sizes less than 50.
We used a planned search strategy (eAppendix in Supplement 1) with the following databases: MEDLINE, Embase, APA PsycInfo, and CINAHL.The search date limits were from inception of the respective database to February 6, 2020.Source selection was performed by 3 investigators (D.B.C., P.W., and B.W.C.C.) with established interrater reliability (Fleiss κ = 0.93).Results were extracted and coded in duplicate by 3 of us (D.B.C., P.W., and K.R.K.) with respect to reported analysis type (univariable or multivariable), statistical significance, direction of effect size, reporting of a priori hypotheses, and adjustment for multiple comparisons.Definitions of predictors, moderators, and mediators from the literature were applied to categorize variables and analyses.Baseline predictors were defined as baseline variables that were associated with depression outcomes, independent of treatment group. 13Moderators (ie, effect modifiers) were defined as baseline variables that were associated with differential outcomes between treatment groups. 13,22,23Postbaseline predictor variables (including time-varying covariates) were defined as variables measured during or after treatment that were associated with depression outcomes, independent of treatment groups. 13Mediators were defined by an analysis of (1) the relationship between an independent variable correlated with treatment group and a postbaseline mediating variable, (2) the relationship between the mediating variable and a dependent outcome, and (3) the extent to which these 2 relationships account for the direct relationship between the independent variable and dependent variable. 13,23r each paper, 2 of 3 investigators (D.B.C., P.W., and K.R.K.) independently extracted the findings in duplicate with respect to end point depression outcomes, ie, outcomes relating to the measurement of depressive symptoms using an evaluator-rated scale (eg, the Childhood Depression Rating Scale-Revised [CDRS-R] 24 ) or self-rated scale (eg, Mood and Feelings Questionnaire 25 ).The most common examples of depression outcomes were continuous outcomes of symptom reduction on a scale score over time or dichotomous outcomes of response or remission. 19Regardless of treatment exposure, response was most often defined as a specified percentage decrease in depression scale scores (eg, 50% reduction on the CDRS-R) or a rating of much improved or very much improved on the Clinician's Global Impression-Improvement subscale. 26Remission was most often defined by an end point scale score below a specific cutoff or no longer meeting criteria for major depressive disorder. 19Greater reduction in symptom scale scores over time or greater proportions of responders and/or remitters represented favorable outcomes.Predictors, moderators, and mediators associated with other outcomes (eg, suicidal ideation, function) were not extracted in this review.

JAMA Network Open | Psychiatry
We created detailed tables describing the findings by publication, RCT sample size, and independent variable of interest.Given that P values were universally described across analyses and articles to characterize results, we categorized reported findings as significant or not significant, depending on the threshold set by the articles' authors; most often this threshold was P Յ .05,unless adjustments for multiple comparisons were made.Aggregated summary tables were created for the reporting of high-level findings from (1) secondary analyses relating to the same RCT sample and (2) predictor, moderator, and mediator variable thematic domains. 27We did not carry out our initial plan to extract effect sizes, which is more appropriate for focused individual patient data metaanalyses.In duplicate, we performed a preliminary risk of bias assessment relevant to secondary analyses; namely, with respect to (1) a priori model development and (2) correction for multiple testing.

Results
Of 98 RCTs identified in total in our search, 33 RCTs (Table 1; eTable 1 in Supplement 2)  reported results of at least 1 predictor, moderator, or mediator tested for association with depression outcome.
Most interventions studied were antidepressants, psychotherapy, and their combination.Analysis of predictors, moderators, and mediators associated with outcomes were reported across the 81 individual publications associated with these RCTs (eTable 2 in Supplement 2),  identified through the citation selection process shown in the Figure.

Baseline Predictors
Among 23 RCTs, 166 different variables were tested as baseline predictors, which we grouped into 53 domains.Baseline variables were analyzed as predictors associated with outcomes in 358 instances and, of these, 269 (75%) were reported as not significant.Baseline variable domains reported as significant with respect to depression outcome in at least 3 RCTs included age, depression severity, parent-child conflict, overall psychopathology, suicidal ideation, hopelessness, and functional impairment (Table 2).

Moderators
Across 21 RCTs, 117 unique variables were tested with respect to moderators that we grouped into 41 domains.In these RCTs, baseline variables were tested as moderators in 197 instances; of these, 159 (81%) were reported as not significant.Baseline variable domains reported as significant when tested as moderators in at least 3 RCTs included sex/gender, depression severity, and history of trauma (Table 3).

Postbaseline Predictors
Across 16 RCTs, we identified 107 unique variables tested as postbaseline predictors, grouped into 19 variable domains.In these RCTs, variables were tested as postbaseline predictors in 114 instances, and 68 results (60%) were reported as not significant.Postbaseline variable domains reported as significant with respect to depression outcomes in at least 3 RCTs included early response to treatment, sleep changes, and attendance at psychotherapy sessions (Table 4).

Mediators
Only 5 publications 28,61-64 across 4 RCTs conducted formal mediation analyses.A total of 16 variables were tested for mediation.In the Treatment for Adolescents with Depression Study (TADS), which compared fluoxetine, CBT, their combination, and placebo among 439 participants, reduction in perfectionism 61 and increase in active motivation 64 (in contrast to precontemplative, contemplative, and maintenance stages of change) mediated depression outcomes; however, these relationships were not specific to any treatment group.In a subgroup of Latinx youth studied by Reyes-Portillo and colleagues, 63 improvements in measures of relationship functioning with peers and family partially mediated the relationship between interpersonal psychotherapy and improvements in depression outcome.Kaufman

Significant in multivariable analysis
Suicidal ideation   mediators was developed a priori; 20 (25%) reported that their models were developed as post hoc tests; and 49 (60%) did not report when the model to be tested was developed.Of the 81 publications, only 10 (12%) reported any adjustment for multiple comparisons; 15 (19%) reported that their models were not adjusted for multiple comparisons; and 56 (69%) did not report on whether adjustments were made for multiple comparisons.Only 2 publications 65,66 reported both a priori model development and adjustment for multiple comparisons.Each of these articles found that higher baseline symptom severity and higher baseline parent-child conflict were associated with unfavorable depression outcomes in multivariable analyses.

Discussion
To optimize the application of principles of precision medicine for the management of DD-A, this scoping review is the first to broadly map out the literature with respect to predictors, moderators, and mediators associated with treatment response in RCTs.Most variables reviewed were classified as not significant.Variable domains reported as significant with respect to their association with outcomes in at least 3 RCTs included age, sex/gender, baseline depression severity, early response to treatment, sleep changes, parent-child conflict, overall psychopathology, suicidal ideation, hopelessness, functional impairment, attendance at psychotherapy sessions, and history of trauma.
In the 2 studies that indicated a priori model development and adjusted significance levels for multiple comparisons, both baseline symptom severity and baseline parent-child conflict were associated with unfavorable depression outcome.Only 5 publications reported results of mediation analyses, and no mediation findings have been replicated to date between RCTs.
Next steps can include the examination of variables identified in this review in individualized patient data meta-analyses (eg, Zhou et al 118 ); this examination may include the use of machine learning strategies in large data sets to clarify their relative importance.If these variables continue to show an important association with outcomes, investing in RCTs designed to specifically examine differential effects of treatment based on the variables is warranted.For example, results from an RCT comparing trauma-focused CBT to depression-focused CBT in adolescents who meet criteria for both DD-A and have a history of trauma could be very helpful in making tailored treatment decisions.
To further assess the role of early response or nonresponse to treatment, the use of adaptive trial designs (eg, Gunlicks-Stoessel et al 115 ) or the study of measurement-based care (eg, Courtney et al 119 ) can also be pursued.
Gaps to be addressed in the literature reviewed here include vulnerability to bias, heterogeneity of findings, and predictors, moderators, and mediators omitted from analyses.Multiple factors can contribute to potential bias.First, analyses of predictors, moderators, and mediators are unlikely to be adequately powered, as sample size calculations are typically made on primary analyses, increasing the probability of type II errors (ie, the risk of not detecting of an important association between independent and dependent variables, when there truly is one). 120Moreover, secondary analyses are often at risk of both publication bias and retrospective bias. 121The set of predictor, moderators, and mediator analyses in our review are susceptible to these biases in that only 12 of the 81 publications identified reported that their predictor, moderator, and mediator analyses were developed a priori.Next, multiple testing without adjustment of P values also renders secondary analyses vulnerable to type I errors (the risk of detecting an association between variables when there is none). 122We found that only 10 of the 81 publications made P value adjustments for multiple testing.The Bonferroni correction was the only method of adjustment we observed.There is ongoing debate regarding whether the Bonferroni correction is too conservative. 123,124Lastly, each variable tested was studied in the context of specific interventions and comparators; results from one trial may or may not be generalizable to adolescents exposed to treatments that were not studied in that specific RCT.
The reported results of the reviewed studies were also quite heterogeneous.Heterogeneity in reported findings was found both between and within RCTs.For example, comorbid anxiety disorder was associated with worse depression outcomes in TADS on multivariable analysis, 67 but it was not found to be associated with outcomes in the Treatment of Resistant Depression in Adolescents (TORDIA) study. 68Within the TADS trial, there was a discrepancy regarding whether a diagnosis of attention-deficit/hyperactivity disorder (ADHD) was associated with differential outcomes between groups across different types of subanalyses 69,70 ; 1 article 69 reported that compared with those without ADHD, participants with ADHD were more likely to have favorable depression outcomes from combined medication and therapy treatment relative to placebo while another article 70 did not find this association.Within-trial heterogeneity may also be attributed to different measurement instruments being used.For example, Birmaher and colleagues found that poor family functioning at baseline, as measured by the Conflict Behavior Questionnaire, 125 was associated with unfavorable depression outcome in a large psychotherapy trial, 29 but another baseline measure of family functioning (the McMaster Family Assessment Device 117 ) was not associated with depression outcomes in the same trial. 65ere are multiple variable domains of interest that have been omitted by the predictor, moderator, and mediator analyses in RCT samples.For example, potential biomarkers of treatment response, like electroencephalogram patterns, functional magnetic resonance imaging findings, inflammatory markers, heart-rate variability, polysomnography findings, genetic markers, and cortisol levels have not been reported in the included RCTs.The role of comorbid borderline personality disorder has also not been reported.The association of treatment implementation factors, like quality and extent of training in psychotherapy, with outcomes has not been examined.
The effect of clinician fidelity to the psychotherapy model was only reported in 1 publication 71 of 11 trials that examined in-person psychotherapy and was reported to be not significant.In psychotherapy trials, the extent to which skill acquisition is associated with depression severity outcomes has also only been evaluated in 1 publication 62 examining mediators of the effects of group CBT in adolescents with comorbid depression and conduct disorder.These variables all require further exploration.
To improve the quality and clinical usefulness of secondary analyses moving forward, investigators examining predictors, moderators, and mediators should incorporate methods to minimize retrospective and publication bias, increase harmonization of variable domain measurement, and standardize reporting of predictor, moderator, and mediator analyses.To mitigate the effects of retrospective and publication bias, researchers can preregister access to data, research questions, and analysis decision-trees. 126A priori documentation of these processes (eg, through Open Science Framework 127 ) means that investigators can still have flexibility in their analytic approach, but are less prone to erroneously highlighting interesting results that may lead to research waste.Increasing harmonization of variable domain measurement can be done through the development of a core outcome set, where all prospective trials are collecting a minimum set of common data in a standardized fashion. 128This process can facilitate the pooling of data across studies which, in turn, facilitates adequately powered analyses.Such a process is under way. 129vestigators would also benefit from the use of widely recognized and standardized guides on how to conduct and report predictor, moderator, and mediator analyses.Recently published guides include The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis Or Diagnosis (TRIPOD), 130 the Instrument for Credibility of Effect Modification Analyses (ICEMAN), 22 and the Guideline for Reporting Mediation Analyses of Randomized Trials and Observational Studies (AGReMA) 131 checklists.
Authors of recent predictor, moderators, and mediator analysis publications have advised caution with the clinical implications of using results for applying precision medicine to the treatment of DD-A. 70,72The high proportion of not significant findings in our review is consistent with this recommendation.The practice of evidence-based medicine integrates the use of the best available evidence, clinician expertise, and patient values. 132Clinicians need to acknowledge the very limited evidence currently available to support the practice of precision medicine when treating DD-A and rely more on their own expertise and patient values until the field is further advanced.For most

Limitations
There are a number of limitations to consider in this scoping review.In the absence of an established method for syntheses of predictor, moderator, and mediator analyses, we had to use consensusbased methods to aggregate independent variables into domains and coding strategies.In the absence of a universally recognized authority on what constitutes an adequate statistical model, the models were not assessed for quality and may also be a source of variation in findings.Our coding system also favored highlighting significant findings over not significant ones, as was required to simplify complex findings.Moreover, we relied on the quality of reporting in the publication to code the various analyses.If reporting in a given publication was unclear, it is possible that codes were misclassified with respect to the actual analysis undertaken.Additionally, we categorized results with respect to a P value threshold as significant and not significant.There are multiple critiques of using and interpreting P values to describe the importance of findings as well as active debate on the use of P values altogether. 136-138

Conclusions
Our scoping review highlights the limited extent to which the literature on predictors, moderators, and mediators can be used to inform further research into precision medicine principles as they apply to the treatment of DD-A.The field would benefit from the use of recognized and established processes and reporting guidelines for predictor, moderator, and mediator analysis publications and related evidence syntheses.In practice, it is important for clinicians to acknowledge uncertainty with respect to matching treatment to patient profiles.
likely to benefit from combination of fluoxetine and CBT or CBT alone relative to fluoxetine alone or placebo; low SES more likely to benefit from combination of fluoxetine and CBT or fluoxetine alone relative to CBT alone or placebo (Curry et al, High severity more likely to benefit from combination fluoxetine and CBT relative to monotherapies or placebo (Curry et al, 67 2006; Foster et al, 70 2019); lower BDI score more likely to benefit from SNRI relative to SSRI (Asarnow et al, 68 2009); if higher severity, more likely to benefit from C-CBT relative to TAU (Merry et al, If anxiety present, more likely to benefit from IPT-A relative to TAU (Mufson et al, conditions, more likely to benefit from CBT with medications relative to medications alone (Asarnow et al, discord, more likely to benefit from combination fluoxetine and CBT or fluoxetine alone relative to CBT alone or placebo; if low marital discord, more likely to benefit from combination fluoxetine and CBT relative to fluoxetine alone, CBT alone, or placebo (Amaya et al, combination of medication and psychotherapy, fluoxetine alone, and CBT alone had similar results, all more likely to benefit relative to placebo; without ADHD, more likely to benefit from combination medication and psychotherapy relative to fluoxetine alone, followed by CBT and placebo (Kratochvil et al, If prior episodes, more likely to benefit from CBT relative to life skills group (Rohde et al, Abbreviations: ADHD, attention-deficit/hyperactivity disorder; BDI, Beck Depression Inventory; BMI, body mass

Table 1 .
Study Characteristics of Included Original Randomized Clinical Trials for the Treatment of Depressive Disorders in Adolescents 117), psychological factors (eg, perfectionism, rumination, self-esteem, cognitive distortions, motivation), history of traumatic events (eg, physical or sexual abuse), recent stressful events (eg, death of loved one or moving), baseline coping skills (eg, avoidance strategies, problemsolving skills), and problem-solving orientation (eg, optimistic vs pessimistic perceptions about problem solving).Figure.Study Flow Diagram16 530 Total articles 12 347 Abstracts or titles screened 2234 Full texts screened 81 Articles included 4183 Duplicates removed 10 113 Rejected at abstract or title 2375 Not an adolescent sample 3050 Not depression 1778 Not treatment of depression 2728 Not original randomized clinical trial 157 Protocol 6 Ongoing study 19 Not in English 2136 Rejected at full text 1153 Not an adolescent sample 326 Not depression 107 Not treatment of depression 401 Not original randomized clinical trial 13 Protocol 58 Not in English 83 Not addressing predictors, moderators, or mediators 12 Conference abstract JAMA Network Open | Psychiatry Variables Associated With Treatment Outcome in RCTs Among Adolescents With Depression JAMA Network Open.2022;5(2):e2146331.doi:10.1001/jamanetworkopen.2021.46331(Reprinted) February 1, 2022 5/22 Downloaded From: https://jamanetwork.com/ on 09/28/2023 Treatment-specific factors tested as postbaseline predictors associated with depression outcome included attendance at psychotherapy sessions, cognitive behavioral therapy (CBT) homework completion, exposure to specific therapy components (eg, motivational interviewing, problem-solving), medication factors (eg, adherence, dosage), posttreatment symptoms, and early response (ie, a dichotomous outcome of response at a relatively early point in the study, such as 2-4

Table 2 .
28d colleagues62tested multiple potential mediators of group CBT for adolescents with depression and conduct disorder, including working alliance with the therapist, group cohesion, skill use, dysfunctional attitudes, and automatic thoughts.Only changes in automatic thoughts mediated the depression outcome.Siilarly, Smith and colleagues28found that changes in ruminative thinking mediated the benefits of computerized CBT. Baeline Variables Tested as Predictors of Depression Symptom Severity Outcomes in RCTs of Treatment for DD-A Preliminary Risk of Bias AssessmenteTable 10 in Supplement 2 outlines the results of our preliminary risk of bias assessment.Of the 81 publications, only 12 (15%) reported that at least 1 model evaluating predictors, moderators, and/or JAMA Network Open | Psychiatry Variables Associated With Treatment Outcome in RCTs Among Adolescents With Depression JAMA Network Open.2022;5(2):e2146331.doi:10.1001/jamanetworkopen.2021.46331(Reprinted) February 1, 2022 6/22 Downloaded From: https://jamanetwork.com/ on 09/28/2023

Table 2 .
Baseline Variables Tested as Predictors of Depression Symptom Severity Outcomes in RCTs of Treatment for DD-A (continued)

Table 3 .
Baseline Variables Tested as Moderators of Depression Symptom Severity Outcomes in RCTs of Treatment of Depressive Disorder in Adolescents

JAMA Network Open | Psychiatry
11,lescents presenting with DD-A, we advise broadly applying recommendations from high-quality clinical practice guidelines (eg, NICE guidelines) as a starting point.11,133Tofurtherpersonalizecare, a shared decision-making model can be used, 134 in which treatment decisions are made through the active elicitation of patient values, the collaborative discussion of benefits and risks of treatment options in light of these values, and the extent of the evidence to support these options.Measuring response to treatment and changing treatment if there is no response (ie, measurement-based care135) can also optimize the personalization of care.