This study was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement for reporting systematic reviews and meta-analyses [47]. The systematic review protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO, registration number: CRD42019119725).

We conducted a comprehensive literature search in the following six electronic databases from database inception to September 13, 2019: PubMed, EMBASE, PsycINFO, CINAHL, Web of Science, and Cochrane Central Register of Controlled Trials (CENTRAL). An information specialist was consulted for the search. The search strategy included combinations of relevant medical subject headings and text-based search terms covering computerized (or internet, digital, eHealth, online, smartphone, or web-based), CBT (or cognitive, behavior, therapy, treatment, or intervention), and adolescent (or child, young person, teenager, or youth). The complete search strings are documented in Multimedia Appendix 1. In addition, we manually searched reference lists of included studies and relevant previous reviews, and searched international trial registers for eligible studies, which resulted in one additional record.

We included RCTs in which computer-based, internet-based, or smartphone-based cognitive behavioral therapy targeting anxiety, depression, or both was compared to an active treatment control condition or passive control condition. The study population involved adolescents or young adults with a mean age between 12 and 25 years and elevated symptoms of anxiety or depressive disorder (ie, either a formal diagnosis or an elevated score on a standardized self-report measure representing at least a mild-to-moderate symptom level). We only included studies with an English abstract available, those that were published in peer-reviewed journals or were PhD theses, and those that contained outcome data on a continuous anxiety or depressive symptom measure that allowed for calculation of effect sizes. If effect sizes could not be calculated, authors were contacted to retrieve the necessary information.

The intervention needed to be primarily delivered via technology (ie, computer, internet, or smartphone). Interventions were categorized as CBT if (1) they were explicitly described as such by the authors of the article and we found no reason to disagree or (2) all authors of this review agreed that the description of the main intervention components could be regarded as CBT. The control condition was defined as active treatment control (ie, face-to-face CBT or treatment as usual [TAU]) or passive control (ie, waiting list/no treatment or information control). Studies in which the control condition involved an active self-help website (ie, including both psychoeducation and exercises) focused specifically on anxiety or depression were excluded. Studies in which the control condition involved a monitoring control website that did not include active self-help content were included. Comorbid psychiatric or medical disorders were not used as an exclusion criterion.

Two authors (CC and MS) conducted the study selection in a stepwise manner. First, titles and abstracts of all studies were independently screened for potential eligibility. Any disagreements were discussed until consensus was reached. Second, the full papers of all included abstracts were independently screened according to the inclusion and exclusion criteria. In case of discrepancy or uncertainty regarding inclusion, a third author (MB) was consulted until consensus was reached.

Information on study characteristics, participant characteristics, and mental health outcomes was extracted from each study and included in an Excel spreadsheet. Data extraction was conducted by one reviewer (CC) and checked by a second reviewer (MS). Study characteristics included authors, country, year of publication, study design, recruitment setting (ie, clinical, general population, or schools), inclusion and exclusion criteria, primary outcome measures, and descriptions of the experimental intervention and comparator, including focus of the intervention, information on guidance, number of treatment modules, and adherence rates. Participant characteristics included sample size, mean age, gender, primary diagnostic type (ie, anxiety, depression, or both; either based on diagnosis or an elevated symptom level), and baseline symptom levels. Means and SDs of the outcome measures of anxiety and depressive symptoms at posttreatment assessment were extracted. If available, means and SDs at short-term follow-up (ie, 1-5 months) and long-term follow-up (ie, 6-12 months) were extracted as well.

If possible, we utilized effect sizes of the intention-to-treat sample; if these were not available, we used effect sizes of the completer sample. In case of multiple outcome measures of anxiety and depressive symptoms, we selected the primary outcome measure as stated by the authors. If the authors did not specify any primary outcome measure of anxiety or depressive symptoms, we selected a well-validated and widely-used outcome measure of these symptoms that was used at every time point of the study (ie, both at posttest and follow-up, if applicable). If both an active treatment control and passive control were utilized in a single RCT, outcomes from both conditions were extracted. Our main meta-analyses were conducted separately for active treatment control (ie, face-to-face CBT or face-to-face TAU) and passive control (ie, waiting list, information control, or no treatment). In our subgroup analyses, data from all control groups per study were included. As the inclusion of multiple comparisons of one study in a meta-analysis violates the assumption of independence, we divided the n of the intervention group evenly across comparators, which is a procedure recommended by the Cochrane guidelines [48].

The quality of each included study was assessed following the guidelines provided by the Cochrane risk-of-bias tool for randomized trials, version 2.0 (RoB 2) [49]. Risk of bias was examined in the following five domains: (1) bias arising from the randomization process; (2) bias due to deviations from intended interventions; (3) missing outcome data; (4) bias in measurement of the outcome; and (5) bias in selection of the reported result. Each domain was rated as either low risk of bias, some concerns, or high risk of bias. A total score was calculated for each study by adding up the following values for each domain: “0” for low risk of bias, “1” for some concerns, and “2” for high risk of bias.

The overall quality of the evidence for each outcome was assessed using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach [50]. The quality of each outcome was assessed for the following domains: (1) risk of bias; (2) inconsistency of results (ie, heterogeneity); (3) indirectness of evidence; (4) imprecision of results; and (5) suspected publication bias. In case of limitations in one domain, the evidence for each outcome was downgraded by one or two levels. Subsequently, the overall evidence for each outcome across domains was categorized as high, moderate, low, or very low, representing the level of certainty of the effect estimates. Both RoB 2 and GRADE assessments were conducted independently by two reviewers (CC and MS), and any disagreements were resolved by discussion until consensus was reached. Cohen κ coefficients were calculated to determine the interrater reliability.

A random-effects meta-analysis was conducted with the Comprehensive Meta-Analysis software (CMA version 3), using the SMD to calculate pooled mean effect sizes (Hedges g). Effect sizes were calculated by subtracting the mean posttest score of the treatment group from the mean score of the comparison group, and dividing the result by the pooled standard deviation of the two groups. Posttreatment means and SDs were compared between the intervention and control groups. Effect sizes of 0.2, 0.5, and 0.8 are considered to be small, medium, and large, respectively [51]. In addition, we calculated the number needed to treat (NNT), using the Kraemer & Kupfer [52] formula. The NNT indicates the total number of patients who need to be treated in order to achieve one additional positive outcome [53].

Heterogeneity was assessed by calculating the I² statistic, which indicates how much overall variance should be attributed to between-study variance, with a value of 25% representing low heterogeneity, 50% representing moderate heterogeneity, and 75% representing high heterogeneity. In addition, we calculated the 95% CIs around I² by using the noncentral chi-square approach in the “heterogi” module of STATA [54,55].

Subgroup analyses were conducted to examine the influence on the difference between intervention and control conditions for (1) the diagnostic focus of the intervention (ie, anxiety, depression, or both); (2) age group (ie, adolescents with mean age ≤18 years or young adults with mean age >18 years); (3) the level of guidance (ie, guided or self-guided); (4) the adherence rate, defined as the percentage of participants in the intervention group who completed all treatment modules at posttreatment (ie, low, defined as ≤50%, or high, defined as >50%); (5) recruitment type (ie, clinical, community, or university/school); and (6) the number of treatment modules (ie, <5, 5-9, or 10-14 modules). Subgroup analyses were conducted across studies with interventions focused on anxiety, depression, or both, with multiple control groups per study included. Subgroup analyses were conducted using the mixed-effects model, in which the effect sizes within the subgroups are pooled with the random-effects model, whereas the fixed-effects model is used to test for significant differences between the subgroups. Subgroup analyses involving age, guidance level, and adherence rates were planned a priori based on the literature. In addition, a subgroup analysis on the diagnostic focus of the intervention was planned a priori to test whether it was justified to conduct all subgroup analyses on the total set of studies, including interventions aimed at anxiety, depression, or both. Recruitment type and number of treatment modules were examined post-hoc as these reflect potential sources of heterogeneity. For all six subgroup analyses, a Bonferroni-corrected α level of P<.008 was used to account for multiple testing. In addition, bivariate meta-regression analyses were conducted to explore the associations of the mean age of study participants, adherence rate, and risk of bias with effect sizes. Analyses with age and adherence rate were planned a priori, whereas risk of bias level was included post-hoc.

Publication bias was examined as follows. First, the funnel plot of effect sizes was visually inspected. Second, the Duval and Tweedie trim and fill procedure was used to calculate an adjusted pooled effect size that accounts for missing studies due to publication bias [56]. Third, the Egger test was used to quantify the bias captured by the funnel plot [57]. In accordance with the Cochrane guidelines [48], publication bias was only examined in meta-analyses with at least 10 studies.

This study provides an up-to-date meta-analysis examining the effects of cCBT on anxiety and depressive symptoms in adolescents and young adults compared with active treatment and passive controls, differentiating between posttreatment and follow-up. Our results indicate that cCBT is beneficial for reducing anxiety and depressive symptoms at posttreatment in adolescents and young adults compared with passive controls, with small to medium effect sizes. For cCBT aimed at depressive disorders or depressive and anxiety disorders, we found a pooled effect size of g=0.51 regarding depressive symptoms, which corresponds to an NNT of 3.55. For cCBT aimed at anxiety disorders or anxiety and depressive disorders, we found an effect size of g=0.50 regarding anxiety symptoms. After adjustment for missing studies owing to a minor indication of publication bias, the effect size lowered slightly to g=0.44, corresponding to an NNT of 4.10. Compared with active treatment controls, the pooled effect size regarding depressive symptoms was in favor of controls (g=−0.70). However, the effect size was not significant and heterogeneity was very high. For anxiety symptoms, cCBT and active treatment controls showed similar effects (g=0.04). Subgroup analyses did not reveal any differences between groups; however, owing to the small number of studies, the statistical power to detect small differences was limited. Meta-regression analyses showed no associations between age, adherence rate, or risk of bias and effect sizes.

Overall, this study shows robust evidence of the effectiveness of cCBT in reducing anxiety and depressive symptoms in adolescents and young adults compared with passive controls. Our results are largely in line with those of previous studies on cCBT in children and young people aged up to 25 years [36], adolescents and young adults aged 12 to 25 years [37] and 10 to 25 years [43], and children and adolescents aged up to 18 years [44]. However, these studies generally reported somewhat larger effect sizes (range 0.62-0.77) regarding both depressive symptoms and anxiety symptoms [36], anxiety symptoms [37], or depressive and/or anxiety symptoms [44] compared with passive controls [36,37,44]. Similarly, research in adults found larger effect sizes for cCBT regarding depression and anxiety compared with passive controls (g=0.90) [34]. Compared with the effects of traditional face-to-face CBT in children and adolescents with anxiety disorders against waiting list controls (NNT=3.0) [19], we found a somewhat lower NNT for cCBT against passive controls (NNT=4.10) after adjusting for potential publication bias. Effect sizes in our study were similar to those found in a meta-analysis comparing face-to-face CBT for depression (0.60) and anxiety disorders (0.48) to passive controls in college and university students [21].

Our results suggest that the effects of cCBT do not differ from those of active treatment controls (ie, face-to-face CBT or face-to-face TAU) regarding anxiety symptoms, but may be inferior to active treatment controls regarding depressive symptoms, although the effect size was not significant. These findings are in line with those of previous meta-analyses in youth across three early studies that were also included in the current meta-analysis [37,41]. Although our meta-analysis included three additional studies, the number of RCTs comparing cCBT with face-to-face treatment in adolescents and young adults remains small. With regard to depressive symptoms, heterogeneity was very high (I²=90.63), and the pooled effect size should be interpreted with caution. Hence, more research directly comparing both treatments is needed to determine whether cCBT is effective compared with face-to-face treatment controls in adolescents and young adults. Research in adults has shown largely equivalent effects of cCBT on both anxiety and depressive symptoms compared with face-to-face CBT [34,35]. However, the number of studies directly comparing cCBT with face-to-face CBT in adults remains limited as well.

This study also aimed to investigate the effectiveness of cCBT at short-term and long-term follow-ups. However, the number of studies reporting follow-up effects was limited. Regarding long-term effects, cCBT was effective in reducing depressive symptoms compared with passive controls, with a small effect size (g=0.27). Our results indicated no long-term follow-up effect for cCBT on anxiety symptoms compared with active treatment controls. Meta-analyses of short-term follow-up data on depressive symptoms indicated similar effects for cCBT compared with active treatment and passive controls. Only Välimäki et al [43] investigated the follow-up effects of cCBT and other internet-based interventions in adolescents and young adults, reporting mid-term and long-term effects on depressive symptoms and mid-term effects on anxiety symptoms. However, they only reported mean differences, and no standardized effect size or NNT. In addition, they did not separately examine the effects of cCBT. Moreover, their selection of studies was based on interventions aimed at depression and not anxiety. Therefore, their results are not easily comparable to those of the current study.

Importantly, owing to the small number of studies reporting follow-up effects, the power to detect small effect sizes was limited in both this study and the study of Välimäki et al [43]. In contrast, the number of studies reporting follow-up effects of cCBT in adults is substantially larger. A recent meta-analysis found 29 trials that reported short-term follow-up effects and 15 trials that reported long-term follow-up effects [34]. cCBT for depressive disorder or anxiety disorder was found to be effective at short-term follow-up (ie, 3-6 months) and long-term follow-up (ie, 9-18 months) compared with posttreatment effect sizes, with small effect sizes across disorders (ie, g=0.15 and g=0.22, respectively). In contrast with the small effect sizes identified in our study, Andersson et al [82] found very large effect sizes for cCBT regarding depressive or anxiety symptoms in adults (g=1.31 across 10 studies) at long-term follow-up of 2 to 5 years compared with mainly passive controls. However, the authors noted that it was unclear whether randomization remained intact over the follow-up period. In summary, in order to determine the long-term effects of cCBT in adolescents and young adults, it is of great importance that future studies include follow-up assessments. Studies comparing cCBT with active treatment control conditions should aim to maintain randomization during the entire follow-up period.

Furthermore, this study aimed to explore whether respondents’ age, guidance level, and treatment adherence were associated with effect sizes. No moderators of treatment effects could be identified. We found no differences in effect sizes for adolescents and young adults regarding anxiety or depressive symptoms, and no association between respondents’ mean age and effect sizes. Previous studies that examined the moderating role of age in meta-analyses among youth reported mixed results. Pennant et al [37] found a higher effect size in young adults compared with adolescents regarding anxiety symptoms, but not depressive symptoms. However, the authors noted that these groups also differed in terms of symptom level, which may have caused the difference in effect sizes. Ebert et al [36] and Podina et al [42] found a higher effect size in adolescents compared with children [36], whereas others [38,44] did not find evidence for such a moderating role of age. Regarding the absence of an association between guidance level and effect sizes, our results correspond with those of Pennant et al [37] in the same age groups. Studies in children and adolescents found mixed results, with Podina et al [42] reporting higher effect sizes for lower levels of guidance and Grist et al [44] reporting higher effect sizes for higher levels of guidance. The lack of an association between adherence rates and effect sizes in our study contrasts findings in adults with depression and anxiety [33,83]. However, most studies included in this meta-analysis did not report the most common operationalization of treatment adherence, and several did not report any information on treatment completion. Post-hoc subgroup analyses and meta-regression analyses found no association of the diagnostic focus of the intervention, risk of bias, recruitment type, or number of sessions with effect sizes.

This study included a thorough evaluation of the risk of bias and quality of evidence, which indicated an overall high risk of bias in 22 out of 24 studies, and, accordingly, low to moderate overall quality of evidence. The high risk of bias was mainly due to an increased risk of bias in measurement of the outcome caused by self-report or unblinded use of observer-rated measures. However, as self-report measures allow both treatment and outcome measures of studies on computerized interventions to be completed entirely from the participant’s home, they are commonly used in studies on cCBT. As such, using the Cochrane RoB tool 2.0, a high risk of bias in measurement of the outcome is inevitable in many studies on computerized interventions for depression and anxiety. Nevertheless, studies ideally should aim to complement self-report measures with blinded observer-rated outcomes, although measures of anxiety and depressive symptoms will always remain subjective to some extent, even when observer-rated outcomes are used.

Several limitations should be considered when interpreting the results. First, the number of studies in the meta-analyses was limited, especially with regard to short-term and long-term follow-up effects. Therefore, the power to detect small effect sizes was limited. Likewise, subgroup analyses consisted of a small number of comparisons, and the lack of relevant differences in most subgroup analyses might be caused by low statistical power. Second, the included studies showed large variations in intervention content, treatment intensity, and outcome measures. Heterogeneity was considerable in the majority of analyses, and pooled effect sizes should be interpreted with caution. Third, most studies had a high risk of bias owing to the use of self-report measures and/or inappropriate handling of missing data. Overall quality was low for most comparisons because of the high risk of bias and, in most cases, considerable heterogeneity. Lastly, almost all studies were conducted in high-income countries, and most studies in young adults were conducted among university students. Hence, generalizability of these results to other populations may be limited.

In the rapidly growing field of computerized mental health treatment in adolescents and young adults, new interventions are developed at a fast pace. Since the publication of the most recent previous meta-analysis in adolescents and young adults [43], six new studies were published, which have been included in our meta-analysis. However, the evidence base in young people remains limited compared with the large body of research in adults, and the quality of RCTs is often low. In addition, most RCTs have compared cCBT to passive control conditions, which appears to lead to an overestimation of effects. It is of utmost importance to compare the effects of cCBT with gold standard face-to-face treatment in order to determine whether cCBT can provide an equally effective alternative. Furthermore, more rigorous high-quality research in accordance with the CONSORT and CONSORT eHealth guidelines for conducting and reporting RCTs [84,85] is needed. In particular, future studies should minimize risk of bias by appropriately handling missing data and, ideally, complementing the use of self-report questionnaires with blinded observer-rated measures. Future research should also include larger sample sizes and longer follow-up periods in which randomization is maintained in case of active control groups [82] and should report adherence rates. Finally, future studies should investigate the effect of cCBT in lower-educated samples, as well as young people from low-income countries, for whom face-to-face mental health treatment is often unavailable [86-88]. When high-quality evidence in adolescents and young adults accumulates, future researchers will be able to draw stronger conclusions on the effectiveness of cCBT compared with both active treatment and passive controls and to determine differences in effect sizes for various subgroups and populations.

This meta-analysis provides robust evidence for the effectiveness of cCBT in the treatment of anxiety and depressive disorders in adolescents and young adults compared with passive controls, with small to medium posttreatment effect sizes. Furthermore, our results indicate that effects of cCBT are similar to those of active treatment controls in reducing anxiety symptoms. Regarding depressive symptoms, however, the results remain unclear, since heterogeneity was high and the number of studies comparing cCBT with active treatment controls was small. No moderators of treatment effects could be identified. cCBT appears to be a promising treatment option for young people, of whom most do not receive face-to-face treatment [23-25]. Importantly, this study also demonstrates the need for more methodologically high-quality research in this population, including active treatment control groups and long-term follow-up assessments.