Effectiveness of current psychological interventions to improve emotion regulation in youth: a meta-analysis

One of the most comprehensive systematic reviews by Aldao et al. [6] looked at the relationship between six different ER strategies and four different psychopathologies, including depression, anxiety, eating disorders and substance abuse. The authors found that the six strategies, avoidance, problem-solving, reappraisal, suppression, rumination and acceptance, were all associated with the different types of psychopathology. More specifically, they found that avoidance and suppression were positively associated with anxiety, depression and eating disorders, while rumination was positively associated with anxiety, depression, eating- and substance-abuse disorders. Problem-solving and reappraisal correlated negatively with psychopathological symptoms, while acceptance showed no significant association with depressive or anxiety symptoms. Further moderator analyses demonstrated that age (child vs. adult) significantly moderated the association between suppression, problem-solving, and depression, with adults showing significantly larger effect sizes than children. Age group was however not a significant moderator for the links between rumination and depression.

Aldao’s systematic review results were primarily based on data derived from adult studies, with only six of the 114 studies including data on children or adolescent samples. However, similar findings demonstrating the close association between emotion dysregulation and psychopathology have also been reported for studies focusing on young populations. Schäfer et al. [7] summarized the evidence for different ER strategies in youth exhibiting sub-clinical symptoms of anxiety and depression. Similarly, they found that depression and anxiety had the strongest positive association with avoidance and rumination; but the strongest negative association with acceptance. Their review focused on adolescents in the ages of 13–18 years with sub-clinical symptoms, therefore no conclusions could be made regarding younger groups or those displaying severe clinical symptoms. Evidence from studies looking at other youth mental disorders such as attention-deficit/hyperactivity disorder, conduct disorder, eating disorders and borderline-personality disorder have reported similar patterns [8‐11].

Most research looking at ER in clinical populations has focused on emotion dysregulation and strategies to regulate negative emotions, such as sadness or anger; while ER as an ability or positive ER strategies (e.g. savouring or gratitude) have been widely neglected [12]. Hence in the present review the term emotion dysregulation will refer to having difficulties, while the term emotion regulation will refer to abilities or skills. Furthermore, it will include both, strategies to regulate positive as well as negative emotions. Additionally, Aldao et al. [13, 14] highlighted that effective ER does not come down to mere down regulation of negative emotions and upregulation of positive emotions, but whether the individual is able to flexibly apply strategies that match the respective situation. Hence the present systematic review will also include ER measures that assess flexible ER.

Despite the growing evidence highlighting the importance of ER in the development of youth psychopathology, it remains unclear whether ER difficulties are a risk factor or a consequence of psychopathology. Until now, only a few studies have shed light on the nature of this relationship. McLaughlin et al. [15] investigated emotion dysregulation patterns in adolescents exhibiting different psychopathological symptoms (i.e. depression, anxiety, aggression, eating pathology) before and after a seven month period. They found that emotion dysregulation (in their study a latent factor based on low emotional understanding, emotion expression and ruminative response to distress), predicted increased symptoms seven months later for all psychopathologies, but depression. Emotion dysregulation on the other hand was not predicted by earlier psychopathological symptoms. Due to the limited availability of longitudinal studies so far, another way to explore this relationship would be through intervention studies that include mediation analyses, based on which one could conclude whether changes in ER lead to changes in psychopathology.

Due to findings relating ER to a wide range of mental disorders, ER has been argued to represent a transdiagnostic core feature underlying these disorders [16]. Transdiagnostic frameworks propose that multiple mental disorders are caused and maintained by a similar subset of underlying processes, which finds further support in high comorbidity rates among the disorders and observations showing that different disorders respond to similar treatments. Following this, it has been suggested that psychological interventions can be improved by having an increased focus on ER [17], as it is the case in most third wave interventions including mindfulness-based cognitive therapy (MBCT), dialectical-behavioural therapy (DBT) or acceptance-based behavioural therapy (ACT). Evidence from the adult literature indicates that promising psychological interventions did improve ER, and that these improvements mediated decreases in psychopathological symptoms [18]. Sloan et al. [19] recently examined whether changes in ER related to symptom reduction in anxiety, depression, substance abuse, eating and borderline personality disorder. They found that the use of maladaptive ER strategies improved following treatment, regardless the type of intervention or disorder. However, their systematic review only included participants older than 13 years.

The present meta-analysis aims to summarize the effectiveness of psychological interventions to improve ER in youth. To our knowledge there is no meta-analysis that has looked at research involving youth samples. Moreover, it focuses on emotion dysregulation, related strategies as well as ER abilities and related strategies. Finally, mediation analyses of changes in ER and psychopathology in response to interventions will be summarized.

We aim to answers the following research questions:

We followed the PRISMA guidelines for the present systematic review [20]. The literature search of the electronic databases was conducted on the 4th of December, 2017 and updated on the 9th of April, 2018 using the following electronic databases: Ovid/Medline, Ovid PsychINFO and Web of Science (a detailed overview of the search strategy can be found in the supplementary materials). Identified publications were downloaded from the databases and saved to a reference manager on the dates specified above. If relevant literature reviews were identified during the abstract screening process (see below), we manually screened their reference lists for further important publications. Our literature search was restricted to peer-reviewed journal articles written in English. Peer-reviewed publications have been assumed to increase the inclusion of studies with higher research quality.

All identified articles were added to a systematic review software (Eppi-Reviewer). Duplicates were removed and abstracts and titles were screened based on the inclusion and exclusion criteria. The method section of each paper was screened for valid ER measures. All studies with a valid ER measure entered the full-text screening stage. A second researcher (D.M) randomly reviewed and rated 25% of the selected title and abstract papers. Where there was a disagreement (4%) regarding the inclusion of a study, the two researchers reviewed the article and discussed its eligibility until an agreement was achieved.

Information relating to study characteristics including: authors, year of publication, study design, intervention type, definition and measurement of ER, comparison group, study results (including sample size, age group, participation rate, attrition, relevant clinical and ER outcomes) and information to determine any study bias was extracted from each study. Correlations between changes in ER and clinical outcomes were collected if reported. Coding options for categorical variables are provided in the supplementary materials.

Two researchers (BM and DM) independently assessed the methodological quality of the included studies (interrater agreement = 98%) using the Effective Public Health Practice Project Quality Assessment tool (EPHPP). The EPHPP evaluates the quality of each study based on their rating, ranging from strong, moderate to weak, across the following six categories: selection bias, study design, the presence of confounding variables, blinding, data collection methods, and participant withdrawals and drop-outs. The EPHPP has been reported suitable for systematic reviews and evidence has shown good content and construct validity [23, 24].

A primary analysis was conducted to detect any influential studies in the data-set. This was done through the “metaninf” command in Stata, which indicates each study’s impact on the overall effect size if that study is omitted from the analysis. Furthermore, we assessed each studies level of heterogeneity through a Galbraith plot (“galbr” command in Stata) [25]. Studies with a great impact on the overall effect size and larger than expected level of heterogeneity, were regarded as influential studies. Subsequent meta-analyses were conducted with and without these studies, in order to identify their respective impact on the results. In line with current recommendations for meta-analysis models in psychology, we conducted two random effects models: one with emotion dysregulation as a primary outcome and one with ER abilities as the primary outcome [26]. To explore sources of heterogeneity we conducted a series of sub-group analyses. Subgroup analyses help identify whether there are differences in effect size or heterogeneity due to study-level factors (see “Meta-regression and subgroup-analyses” for more detail below). Furthermore, we conducted a meta-regression with effect size as the dependent variable and intervention type, age group, control group and quality rating as the predictor variables. A combination of these two approaches has been recommended [27]. In order to answer the second research question we conducted a second meta-regression, with effect sizes of psychopathological symptoms as the dependent variable, and effects sizes of improved ER as the predictor variable.

Treatment effect was estimated using the weighted mean effect size Hedges’ g. Hedges’ g is interpreted like Cohen’s d, with effect sizes ranging from small (0.2), medium (0.5) to large (0.8) [28]. Hedges’g (see Formula 1) and the standard error were calculated based on standardized mean-differences, standard deviations and sample sizes. This data was entered into Stata and the “meta” command was used to conduct the random effects models.

Formula 1—Hedges’ g.

For studies with multiple treatment groups, the decision on how to include them, was made on a case-by-case basis with regards to the research question. In accordance with the Cochrane handbook the following options were considered [29]:

Heterogeneity between the studies was assessed with the Q statistic, I² and T². The Q test follows the chi-square distribution and estimates the probability of sampling error being the only cause for variance. A significant Q test indicates that heterogeneity is present. However, it does not provide sufficient information about the source of heterogeneity. Therefore, I² and T² were also taken into account. T² describes the between-study variance, while I² describes what proportion of the observed variance in the effect estimates is due to systematic differences between the studies rather than sampling error. Smaller values of I² suggest that the observed heterogeneity is mostly random, while larger values suggest study-level differences. The following levels of heterogeneity have been identified for I²: low: I² = 25%, medium: I² = 50%, and high: I² = 75% [30]. We also calculated 95% prediction intervals (PI; see Formula 2) [31], which aim to predict the range of possible population parameters in future empirical studies (e.g., we expect that in future studies 95% of the true effects lie within this interval). Hence, PI’s are different from confidence intervals, which estimate the precision of the mean effect size in the general population.

Formula 2—prediction interval

A meta-regression was performed to identify possible moderating effects of certain between study-level characteristics. The meta-regression was conducted with the “meta regress” command in Stata 16. Categorical variables are automatically dummy-coded by the software and the resulting estimates indicate how the effect size of each subgroup differs with respect to the chosen reference group. Furthermore, separate subgroup analyses with each relevant moderator were conducted to explore potential sources of heterogeneity and their impact on the overall effect size. With respect to the present research question the following subgroup analyses were conducted:

Publication bias was visually assessed with the help of a funnel plot. No publication bias was assumed if the points in the scatter plot form the shape of a funnel, while an asymmetrical shape suggests a publication bias. Furthermore, the Egger’s test was applied to test for small-study effects whereby precision seems to be related to the effect size estimate. Fail-safe N statistics were not performed due to unreliability [32]

To assess whether improvements in psychopathological symptoms were associated with changes in ER, a meta-regression was conducted, with effect sizes of psychopathological symptoms as the dependent variable, and effects sizes of improved ER as the predictor variable.

The search identified 1418 articles. After duplicates (n = 171) were removed 1250 papers were included for the abstract and title screening. 1049 articles were excluded based on the abstract and title screening. Of the remaining 201 papers, 122 papers had to be excluded due to missing ER measures. In total, 79 studies entered the full-text screening, of which 34 studies matched the selection criteria and provided sufficient data. Another 17 studies, matched the criteria, but the authors had to be contacted to provide additional information that could not be derived from the published article. During the data extraction phase 30 studies were excluded. Four of those were excluded because the authors were not accessible [33‐36]. Finally, 21 independent studies were included in the meta-analysis, from which 33 treatment effects were extracted (19 emotion dysregulation, 14 emotion regulation; see Fig. 1 for study selection process).

The characteristics of the included studies are summarized in Table 1. For studies with multiple treatment conditions, where both treatment conditions were assumed to have an effect on ER, both groups were included in the analysis, by splitting the control group in half and pairing it with each treatment group. For the remaining studies (k = 3), the second treatment group was excluded. All of the included studies showed a large variety regarding the type of ER measure and intervention employed a detailed description of these is provided in the supplementary materials, see Table 5. CBT was the most commonly employed intervention (k = 16) and almost all interventions included some kind of CBT components. Eight studies stated to specifically address emotion dysregulation (i.e. emotion regulation training). Four interventions targeted specific ER strategies (i.e., rumination or mindfulness; see Table 1).

Quality ratings for each study are shown in Table 1. All studies were randomized control studies, however nine studies reported baseline differences between the groups, while two studies did not provide any information on potential baseline differences. One study did not provide any information about the control condition, six studies compared the intervention with a treatment as usual condition.

The first random effects model was based on the original 19 effect sizes from 17 independent studies, which indicated a medium treatment effect (g = 0.52), 95% CI [− 0.86, − 0.18], p < 0.001). Due to large heterogeneity I² = 90.87% (Q = 129.64, df = 18, p < 0.001), we decided to run the “metainf” command and a Galbraith plot to identify highly influential studies [25, 37]. The results (see Plots 1 and 2 in supplementary materials) indicated that two studies, one by Slee et al. [38] and one of Livheim and colleague’s studies (based in Sweden [39]) had a significant impact on the overall effect size, while also contributing to a large amount of heterogeneity. We regarded these studies as highly influential studies and removed them from the main model, which effectively decreased the level of heterogeneity by I² = 18.05% [37]. (Results of the full and the reduced meta-analysis model are presented in Table 2 and supplementary materials). Results of the reduced model are discussed in more detail below.

The forest plot and confidence intervals (CI) show that eight studies significantly reduced emotion dysregulation (CIs are entirely on the negative side), while the remaining studies (k = 9) showed no significant treatment effect (Fig. 2). Overall, the results indicate a medium treatment effect (g = − 0.46), 95% CI [− 0.67, − 0.26], p < 0.001). The confidence interval (no value of 0 is present), and the z statistic (z = − 4.44, p < 0.001) suggest that the null hypothesis (H₀: intervention had no impact on emotion dysregulation) can be rejected. The Q statistic (Q = 54.06, df = 16, p < 0.001) indicated that the effect sizes differed significantly across the studies. I² of 72% suggests that most of the observed variance was due to differences on a study-level. T² of 0.12 suggests a small amount of absolute dispersion. Calculation of the 95% PI [− 0.67, − 0.25] suggests that the true effect size of a similar future study would fall within this range in 95% of the time. Most of the PI lies in the negative range, thereby indicating that interventions would be effective in most settings [30, 31].

The original random effects model was based on 14 effect sizes from 13 independent studies with ER abilities as an outcome. The full model indicates a treatment effect of (g = 0.43, 95% CI [0.18, 0.69], p < 0.001).The metaninf and the Galbraith plot suggested two influential studies, Livheim et al. [40] and Essau et al. [41] (see supplementary material Plots 3 and 4). In comparison to Essau et al. (N = 638), the study by Livheim (N = 25) was significantly underpowered, hence we decided to remove this study from the following analysis. (Results of the full and the reduced model are both presented in Table 3). The forest plot of the reduced model indicated that three studies [41‐43] showed a significant positive effect, while the remaining 10 studies had no significant effects (see Fig. 3). Overall, the results suggest a small treatment effect (g = 0.36, 95% CI [0.14, 0.58], p < 0.001). Based on the CI and the z statistic (z = 3.22, p < 0.001), the null hypothesis that the intervention has no impact on ER was rejected. The Q statistic (Q = 66.56, df = 12, p < 0.001) suggests that effect sizes differed significantly across the studies. I² of 70.8% suggests that most of the observed variance was due to differences on a study level (e.g., sampling error). T² of 0.10 suggests a small amount of between-study variance. The 95% PI = [0.14, 0.58] is in the positive range, suggesting that future studies will most likely find a positive effect size within this range [30, 31].

To explore possible causes of heterogeneity and investigate whether effect sizes varied for certain subgroups, a meta-regression and subgroup analyses were conducted (see Tables 2, 3, 4).

The contour-enhanced funnel plot (see Fig. 4) shows an asymmetric pattern. Visual inspection of the funnel plot indicates more studies on the left side. Furthermore, we see missing data points at the top and bottom of the funnel, for both the significant (light grey) and non-significant (dark grey) areas. In the case of a publication bias, we would expect to see missing studies in the non-significant areas. The present funnel plot seems to rather suggest a gap for studies including larger sample sizes. Most of the studies included in this review involved similar, small to medium-size samples (great density in the middle), which can result in spuriously increased effect sizes. Therefore, we conducted the Egger’s test, which was significant, thereby suggesting a bias, due to small-study effects (z = − 2.22, p < 0.05).It has been reported however that funnel-plot asymmetry can be caused by publication bias, as well as other factors such as poor methodological quality or between study heterogeneity [32]. Due to the large amount of heterogeneity in our analysis we performed the Egger’s test again, this time taking into account between-study heterogeneity, as a result of different types of interventions, ER measures and control groups. We found that heterogenity due to different intervention types, significantly influenced the results of the the Egger’s test, which was nonsignificant when intervention type was added to the model (z = − 1.31, p = 0.19).

Only two studies reported whether changes in ER were associated with changes in psychopathology. Slee et al. [43] investigated adolescents engaging in deliberate self-harm, and found that changes in ER difficulties partially mediated decreases in deliberate self-harm. The second study [45] found that changes in acceptance mediated decreases in anxiety and depression. Our meta-regression indicated a significant positive relationship between larger effect sizes of reduced ED and larger effect sizes of reduced psychopathology (see Fig. 5; β = 0.76, t = 2.93, p = 0.01). In other words, studies showing greater effectiveness in reducing ER difficulties were also more effective in reducing psychopathological symptoms (Figs. 6, 7, 8, 9).

Despite the limited evidence for the causal role of ER in the development and treatment of psychopathology in youth, the present findings encourage further development and evaluation of interventions that target ER specifically.

The average effect sizes suggest that interventions effectively change ER, irrespective of the type of the intervention program. However, the validation of existing interventions represents an important area for future work. As the present systematic review demonstrates there is a significant variety across intervention protocols in the way they target emotion regulation, and it is not clear yet which of the included components effectively enhance emotion regulation. Furthermore, there is still limited evidence with respect to different age groups and psychopathologies, which is of particular importance. First, research has shown that emotion regulation does not develop in a linear pattern, but that different developmental stages are characterised by certain advancements and deficits [46, 47]. For instance Cracco et al. [46] and Zimmermann et al. [47] have demonstrated that there is a significant shift in adolescents’ emotion regulation patterns (e.g., access to strategies, use of adaptive vs maladaptive strategies), which current interventions do not seem to take into consideration. Therefore, we argue that more efforts need to be made to increase our understanding of what works for who and when, so that relevant changes can be implemented in current clinical treatment plans. Secondly, young people frequently display a wide range of psychopathological symptoms and comorbidities [48, 49]. This can make interventions that have been designed for single-disorder symptoms, less suitable for this group. Thus the present review supports existing recommendations that ER interventions are effective in reducing a wider range of psychopathological symptoms by targeting underlying processes, which makes them highly suitable for young populations with high rates of comorbidities [17]. Furthermore, our results suggests the potential of transdiagnostic treatments being added as adjunctive modules in existing treatment protocols. This approach has already found support in adult studies where ER interventions in combination with CBT have resulted in better mental health and wellbeing outcomes than CBT alone [50].

The results were based on a relatively small number of studies, which primarily involved small to medium sized samples. It can be assumed that the variety in populations, intervention settings (e.g., digital, inpatient and outpatient, schools) and use of ER measures lead to large between-study variation, which may have biased our findings. With respect to the latter it has been highlighted recently that meta-analyses with an increased psychometric focus could provide more insights regarding the impact of measurement error on outcome biases [51]. In the present meta-analysis, only 11 of the 19 studies reported information on reliability, which did not allow us to correct for measurement error. Hence, we highly encourage future meta-analysts to also consider bias due to measurement error. Moreover, there was a great variety between interventions, even though CBT formed the basis of most interventions. However, due to the limited amount of data available, it was impossible to provide further insights regarding the impact of these study artifacts on the overall effect size.

Furthermore, due to missing evidence from longitudinal mediation analyses, the present study could only partly address the second research question whether changes in ER precede changes in psychopathology. Only two studies [43, 45] reported whether changes in ER were associated with changes in psychopathology. Both studies found that changes in ER mediated decreases in psychopathology. Similarly, our meta-regression showed a significant positive relationship between effect sizes of improved ER difficulties and effect sizes of improved psychopathology. Moreover, most studies only assessed changes in anxiety or depression even though a wider range of symptoms was reported at baseline. Due to the current lack of research reporting on ER outcomes in relation to different psychopathology outcomes, we were not able to conduct more specific mediation analyses. Similar issues have been raised in previous systematic reviews [52]. We recommend that future research includes measures of ER so that underlying mechanisms of change can be identified.

The quality of the included studies ranged from weak to strong. Even though we focused primarily on RCTs, there was a significant lack of high quality studies. The limited evidence may have made it difficult to detect differences in effect sizes relating to study quality. Moreover, it has frequently been pointed out that the level of quality found in primary research has a significant impact on the quality of any systematic review, due to the fact that systematic reviews rely on data from existing studies. Following this we can only emphasize that future research needs to focus on the delivery of more high quality studies that provide high-quality research outcomes. In line with this, we acknowledge that while we had hoped to identify more high-quality studies by excluding non-peer-reviewed articles, the exclusion of such unpublished data may have resulted in biased outcomes. Although our publication bias assessment did not clearly indicate the presence of a publication bias, this may have been due to the high level of heterogeneity. However, we would like to highlight that there was significant lack of large-sample size studies that included a comprehensive psychopathology assessment and targeted youth populations.

Further RCTs including larger sample sizes, different age groups and mental disorders are needed. While evidence suggests that research has widely neglected populations under the age of 25, future research should specifically address youth populations between the ages of 10 and 12 years. They form an interesting age group as research has emphasized a significant drop in ER skills at this age [47]. Furthermore, studies involving youth mostly investigate anxiety or depressive symptoms, while only a few have looked at ER in relation to other mental disorders. Similarly, interventions with a specific focus on ER often target specific disorders. Considering the suggested transdiagnostic nature of ER, future studies should involve participants from a broader psychopathological spectrum.

To increase our understanding of ER interventions and associated change mechanisms, future research needs to assess and actually report ER processes. A large number of studies was excluded, due to missing ER assessment. This can not only improve future interventions, but would also reduce the exploratory nature of current interventions. In line with this we suggest that future research should focus on the impact of measurement error in their studies. As mentioned above, studies included a wide range of ER measures, which have been based on different theories and models around ER. Thus, a psychometric meta-analysis of current ER measures, would be highly beneficial to the field.

Finally, we found that the investigation of positive ER strategies and ER abilities is still widely neglected. Although, past research has highlighted that adaptive ER strategies, as opposed to maladaptive strategies, were more strongly related to psychopathology in youth [53]. The opposite has been reported in adult studies [54]. We identified only one study that assessed a positive ER strategy [55]. This could be related to the fact that positive psychology is still a rather young field in comparison to the traditional CBT approaches or that the use of ER strategies has been less frequently studied in youth populations. Nevertheless, in line with previous research [12, 53] and our findings, we argue for a greater focus on the positive dimension of ER especially in researching and working with young populations.