Self‐Criticism as a Mechanism of Change in Compassion‐Based Interventions: A Meta‐Analytic Structural Equation Modelling Approach and Systematic Review

ABSTRACT

Compassion‐based interventions (CBIs) are effective transdiagnostic treatments for a wide range of psychopathological symptoms. While their efficacy is well established, their mechanisms remain insufficiently understood. Self‐criticism (SC) has been proposed as a promising mechanism of change (MoC), given its central role in psychopathology and its theoretical link with CBIs. This study synthesises evidence from randomised controlled trials to examine whether reductions in SC mediate the effects of CBIs on psychopathological symptoms. Systematic searches in PubMed, Web of Science, Scopus and ProQuest identified 57 randomised controlled trials (n = 4877) of CBIs in adults reporting SC and psychopathological symptom outcomes, analysed using One‐Stage Meta‐Analytic Structural Equation Modelling (OMASEM). SC partially mediated the effects of CBIs on symptom reduction (e.g., depression, anxiety, eating disorders; indirect effect = −0.054, 95% CI [−0.085, −0.030]). Stronger mediation effects were observed in studies prescribing more daily home‐practice minutes and shorter intervention duration. At posttreatment, CBIs achieved moderate reductions in SC (r = −0.20, 95% CI [−0.27, −0.13]) and psychopathological symptoms (r = −0.21, 95% CI [−0.27, −0.15]). Moreover, SC was positively correlated with psychopathological symptoms (r = 0.48, 95% CI [0.42, 0.54]), supporting partial mediation by SC. Considerable heterogeneity was observed across interventions and designs; 76.4% showed high risk of bias. These findings support SC as a potential MoC in CBIs and provide quantitative evidence for process‐based formulations of CBIs. The results also highlight the clinical relevance of directly targeting SC in therapeutic practice and contribute to understanding the psychological processes underlying CBIs.

Summary

• Self‐criticism mediates the relationship between compassion‐based interventions and psychopathology.

• Compassion‐based interventions reduce both self‐criticism and psychopathology.

• Higher levels of self‐criticism are associated with greater psychopathology and may therefore be relevant to clinical assessment.

• Self‐criticism may represent a relevant mechanism of change in compassion‐based interventions and may be useful to monitor during treatment.

• Self‐criticism may constitute a therapeutic target for optimising the effectiveness of compassion‐based interventions.

1. Introduction

Although compassion has been cultivated for centuries as a spiritual practice (Buddhaghosa 1991), compassion‐based interventions (CBIs) have grown substantially in the West in recent years and are increasingly recognised as transdiagnostic treatments (Kirby 2025). CBIs have been shown to be effective in reducing psychopathological symptoms, including anxiety, depression, posttraumatic stress disorder (PTSD) and eating disorders (e.g., Kirby et al. 2017; Millard et al. 2023; Winders et al. 2020), with small‐to‐moderate effects typically observed for anxiety and depression (Kirby et al. 2017; Ferrari et al. 2019; Han and Kim 2023; Luo et al. 2023; Fan et al. 2022). A meta‐analysis of CBIs (Kirby et al. 2017) identified six validated protocols, including compassion focused therapy (Gilbert 2014), compassion cultivation training (Jazaieri et al. 2013), mindful self‐compassion (Neff and Germer 2013), cultivating emotional balance (Kemeny et al. 2012), cognition‐based compassion training (Pace et al. 2009) and loving‐kindness meditation (Galante et al. 2014). Since then, additional approaches have also been investigated, such as attachment‐based compassion therapy (García‐Campayo et al. 2016) and compassion mind training (CMT; Matos et al. 2017), reflecting continued development in the field. Despite differences in their origin and methodology, these interventions share a foundation: the systematic cultivation of compassion (Kirby et al. 2017). This cultivation can be framed within Buddhist contemplative traditions or secular/evolutionary approaches. In the former, meditation emphasises direct experience and mental settling to support self‐transcendent, nondual compassion (i.e., reduced self–other boundaries; Gilbert and Van Gordon 2023). In contrast, secular approaches conceptualise compassion as a trainable prosocial motive/skill regulated by biopsychosocial systems (Gilbert and Van Gordon 2023). From this relational framework, compassion is expressed in three flows: toward others, received from others and toward oneself (self‐compassion; Gilbert et al. 2017). Both perspectives aim to alleviate and prevent suffering and converge on the need for sustained practice (Gilbert and Van Gordon 2023). At a pragmatic secular midpoint, Strauss et al. (2016) offer an operational definition of compassion integrating cognitive, affective and motivational components: recognising suffering, connecting with it through empathy, tolerating the distress it evokes and acting to alleviate it.

While CBIs are empirically supported, the mechanisms underlying their efficacy remain largely unexplored (e.g., Kirby and Gilbert 2019). Mechanisms of change (MoCs) refer to the dynamic, progressive and multilevel processes through which therapeutic change unfolds (Hayes et al. 2019). Identifying these MoCs is crucial for advancing theoretical models (Holmes et al. 2018; Tortella‐Feliu et al. 2016), and for optimising CBI protocols, enhancing effectiveness and improving real‐world applicability (Holmes et al. 2018; Kazdin and Nock 2003; Kraemer et al. 2002). Although research on MoCs in CBIs remains limited, several candidates have been proposed (e.g., compassion flows, fear of compassion, emotion regulation, common humanity and empathic concern, affect, self‐reassurance and self‐criticism [SC]; Matos et al. 2024; Paetzold et al. 2023; Roca et al. 2021; Sommers‐Spijkerman et al. 2018). Among these possible MoCs, the present study focuses on SC, given its strong theoretical relevance and well‐established association with psychopathology (e.g., Cox, Fleet, and Stein 2004; Gilbert 2009a, 2014).

SC refers to the tendency to judge oneself harshly across thoughts, behaviours and core aspects of identity, often resulting in feelings of inadequacy, worthlessness and failure (Blatt and Luyten 2009; Gilbert et al. 2004; Shahar 2015). Under certain conditions, SC may serve a constructive, self‐regulatory function, particularly when expressed in a nonhostile emotional tone and directed toward goal attainment (Gilbert 2022; Gilbert and Irons 2005; Kannan and Levitt 2013). More commonly, however, SC is associated with maladaptive psychological processes characterised by negative self‐evaluations, excessive self‐scrutiny, harsh self‐judgement and a critical inner dialogue (Blatt et al. 1974; Gilbert et al. 2004; Shahar 2015). These processes often elicit distressing emotional responses, including shame, guilt, anger, contempt, and self‐disgust (Whelton and Greenberg 2005), as well as persistent experiences of disapproval and internal judgement (Blatt et al. 1974; Kannan and Levitt 2013; Shahar 2015). Moreover, some authors argue that SC is not a unitary process but rather varies in terms of its emotional basis, form and function (Gilbert et al. 2004). Specifically, two main forms can be distinguished: the ‘inadequate self,’ which is linked to competitive motives and fear of inferiority and characterised by feelings of failure and inadequacy; and the ‘hated self,’ which involves a persecutory way of relating to oneself and is characterised by self‐contempt, disgust and hatred (Gilbert 2015; Gilbert et al. 2004). These differences are relevant, as SC appears to be more pathological and resistant to change when accompanied by feelings of hatred, disgust and self‐contempt (Castilho et al. 2017; Gilbert et al. 2004; Werner et al. 2019). From this perspective, shifting toward a more affiliative and compassionate way of relating to oneself may be particularly relevant, given its potential implications for psychopathology and treatment resistance.

From a theoretical perspective, the conceptualisation of SC varies across the frameworks underlying CBIs. Within Gilbert's model of emotional regulation systems (Gilbert 2009a, 2014; Gilbert and Van Gordon 2023), SC is viewed as a powerful activator of the threat system (Longe et al. 2010). The model distinguishes three interacting systems (threat, drive and soothing) that regulate affect and motivation. Accordingly, CBIs aim not only to reduce threat but to cultivate the soothing system, associated with safeness, care and affiliation, thereby restoring balance with threat (Gilbert 2009a, 2014; Gilbert and Van Gordon 2023). In this model, the threat system is considered dominant, and its overactivation has been linked to an increased risk of psychopathology (Gilbert 2009b). Given its relevance, certain CBIs have been specifically developed to target SC, such as CMT (Gilbert and Procter 2006), by replacing punitive self‐attack with compassionate self‐correction (Gilbert and Woodyatt 2017).

Moreover, SC and self‐reassurance are associated with different psychological, physiological and neurophysiological responses. SC is related to behavioural inhibition, processing and resolution of errors, whereas self‐compassion involves processes of self‐soothing and comforting oneself (Longe et al. 2010). Although they share common mechanisms related to self‐referential processing and mentalising, which are associated with the default mode network (with no significant differences found between SC and self‐reassurance in core default mode network regions), significant differences were found in other areas (Kim et al. 2020). Specifically, SC was associated with greater activation of the anterior insula, anterior cingulate cortex and amygdala, whereas self‐reassurance was associated with reduced activity in these regions. Consistent with this reduced threat reactivity, compassion training was associated with increased HRV, suggesting greater parasympathetic regulation and promoting emotion regulation (Kim et al. 2020). Therefore, although these dimensions of safety and threat can coexist, cultivating self‐compassion may help buffer or regulate the emotional and physiological impact of SC (Slavich et al. 2023).

Taking a different theoretical perspective, Neff (2003) describes SC as a form of self‐judgement, a harsh self‐evaluative process that contrasts with self‐kindness. In her model, self‐compassion comprises three bipolar dimensions: self‐kindness versus self‐judgement, common humanity versus isolation and mindfulness versus over‐identification (Neff 2003; Neff 2022, 2023). Accordingly, self‐compassion is framed as a bipolar construct, with SC at one pole and compassionate responding at the other (Neff 2016, 2020, 2022). Taken together, these perspectives highlight that SC remains closely intertwined with self‐compassion, making it a central therapeutic target for reducing psychopathology.

From an empirical perspective, accumulating evidence suggests that SC is implicated in numerous forms of psychopathology. Associations have been documented with depressive disorders (Luyten et al. 2007), eating disorders (Dunkley and Grilo 2007), social anxiety (Cox, Fleet, and Stein 2004), posttraumatic stress disorder (Cox, MacPherson, et al. 2004) and borderline personality disorder (Kopala‐Sibley et al. 2012). SC may operate not only as a transdiagnostic vulnerability factor (Zaccari et al. 2024), but also as a core process in the onset and maintenance of various psychological disorders (Gilbert and Irons 2005; Warren et al. 2016). Importantly, longitudinal evidence further supports the proposed role of SC in psychopathology, showing that SC predicts psychosocial impairment and increases in depressive symptoms in clinical samples up to 4 years later (Dunkley et al. 2009). Furthermore, individuals with remitted depression continue to report elevated levels of SC compared with individuals with no history of depression (Ehret et al. 2015), suggesting that elevated SC may persist beyond the remission of depressive symptoms and may represent a vulnerability factor for recurrent or chronic depression. Moreover, higher pretreatment SC is generally associated with poorer treatment outcomes across various treatment modalities (such as cognitive–behavioural, interpersonal and psychodynamic psychotherapies, as well as pharmacotherapy among others; Löw et al. 2020; Marshall et al. 2008). Although Marshall et al. (2008) suggested that the association between pretreatment SC and treatment outcomes may vary across treatment modalities, Löw et al. (2020) found no significant moderating effect of treatment type, suggesting that this negative association may be relatively consistent across treatment modalities. Given its role in psychopathology and its strong theoretical alignment with the goals of CBIs, reductions in SC may represent a key mechanism through which CBIs exert beneficial effects.

Recent meta‐analyses indicate that CBIs significantly reduce SC, with effect sizes ranging from small to large (g/d = 0.30–0.92; Ferrari et al. 2019; Petrocchi et al. 2024; Vidal and Soldevilla 2023; Wakelin et al. 2022). Effects are typically stronger versus waitlist (WL; g = 0.67) than active controls (g = 0.37; Ferrari et al. 2019). Despite this evidence, studies examining SC as a mediator remain scarce. Only Sommers‐Spijkerman et al. (2018) directly tested SC as a mediator: reductions in SC significantly mediated CBI‐related improvements in single‐mediator models (lower depression/anxiety and higher well‐being), but SC was not a unique mediator in multiple‐mediator models, suggesting shared variance with other MoCs. Although CBIs likely involve multiple mechanisms, SC's mediating role remains unclear, and no meta‐analytic estimate has yet been provided; this study aims to fill that gap.

This paper has two aims: (1) to systematically review RCT evidence on the effects of CBIs on SC and psychopathology, including study characteristics, outcomes, and methodological quality; and (2) to test whether SC mediates the effect of CBIs on psychopathological symptoms. To inform the mediation model, we estimated pooled associations for CBI–SC, CBI–psychopathology and SC–psychopathology.

More specifically, we conducted a systematic review and a mediation meta‐analysis using the OMASEM methodology (Jak et al. 2021; Jak and Cheung 2020). This methodology integrates structural equation modelling with meta‐analytic techniques (Jak et al. 2021). This approach offers several advantages (Ng et al. 2023): it moves beyond bivariate associations by enabling multivariable mediation analyses, supports the examination of moderators to clarify when mediation occurs (Jak and Cheung 2020), and is more effective for synthesising indirect effects (van Zundert and Miočević 2020). It also maximises the use of partial data from RCTs, which is important given the limited number of trials testing SC as a mediator (Sommers‐Spijkerman et al. 2018).

2. Method

2.1. Search Procedure

A comprehensive literature search was conducted following PRISMA guidelines (Page et al. 2021) without date restrictions; it was initially performed in September 2024 and updated in August 2025, covering studies published in peer‐reviewed scientific journals. Eligible studies were retrieved from PubMed, Web of Science, Scopus and ProQuest databases (see Supporting Information, Section S1 for the full list). Additionally, when papers met all criteria except outcomes but cited another study reporting the relevant variables, the companion article was actively sought and screened; one such article met the criteria and was included. Search terms targeted four key constructs: compassion, SC, intervention and psychopathology. Search strategies for each database are detailed in Tables S1, S2, S3 and S4, in Section S1 of the Supporting Information.

2.2. Inclusion and Exclusion Criteria

The inclusion criteria were defined using the PICO framework (Speckman and Friedly 2019). Eligible studies included adult populations (P; ≥18 years) from nonclinical, clinical or physical health condition samples. The intervention (I) was required to primarily target compassion (≥50% of content; ≥2 weeks). For the comparison (C), eligible studies were randomised controlled trials in which participants were randomly allocated to any control condition (e.g., WL, treatment‐as‐usual, active control or another psychological intervention). Finally, the outcomes (O) of interest were studies providing pre‐ and postintervention assessments of SC and psychopathological symptoms. Psychopathological symptoms had to be assessed using a validated questionnaire designed to measure symptoms across different psychopathological domains, such as depression, anxiety, trauma, among others. Given the conceptual heterogeneity surrounding SC, the instrument selection was based on Rose and Rimes (2018), who reviewed the measurement properties of SC questionnaires (the full list is reported in Supporting Information, Table S5). Although their review included only measures explicitly assessing SC, we also considered the Five Facet Mindfulness Questionnaire (FFMQ; Baer et al. 2006) as an SC measure, because its ‘Non‐judging of Inner Experience’ subscale captures SC‐related content, consistent with our framework linking SC to nonjudgemental awareness of internal experiences (Baer et al. 2006).

Studies were excluded if they were not published in English or their full text was not available. We also excluded non–peer‐reviewed, editorial materials, conference abstracts, proceedings, reviews, meta‐analyses, case reports, letters and patents as well as studies in which required data were unavailable or not provided by authors.

2.3. Study Selection

Study selection proceeded in three stages: removal of duplicates, title/abstract screening and full‐text eligibility assessment. The process was managed in Hubmeta (Steel et al. 2023), and the final data extraction was organised in a spreadsheet. Two evaluators (D.C. and A.P.) independently selected studies, achieving 86% agreement (Cohen's κ = 0.86); disagreements were resolved by a third reviewer (A.C.).

2.4. Data Summary

In this study, extracted information is presented in structured tables summarising key study features, including design, sample characteristics, assessment instruments, intervention type and control condition. We extracted descriptive statistics (e.g., frequencies/percentages) on study context (e.g., country, population type, clinical status) and measures (SC and psychopathology instruments). Variables extracted included participant demographics, design features (sample size, control condition), intervention characteristics (format, duration, frequency), outcome data (means, standard deviations and SC–psychopathology correlations) and methodological indicators (response, dropout, follow‐up).

2.5. Data Quality

To assess the methodological quality, two independent reviewers (A.P. and D.C.) evaluated the risk of bias using the Cochrane Collaboration's RoB 2.0 tool (Higgins et al. 2024). Discrepancies were resolved through discussion until consensus. RoB 2.0 evaluates bias across randomisation, deviations from intended interventions, outcome measurement, missing outcome data and selective reporting.

2.6. Effect Size Calculation and Data Analysis

In this meta‐analysis, we focused on SC as a mediator of the effects of CBIs on psychopathology. Meta‐analytic mediation models were tested using OSMASEM (described in the next subsection). However, OSMASEM could not accommodate certain analyses, including assessing and correcting publication bias, identifying outlying correlations, or incorporating moderator variables that varied within studies. To address these limitations, we conducted a second set of analyses in which each individual bivariate association (CBI–SC, CBI–psychopathology and SC–psychopathology) was meta‐analysed separately.

2.6.1. Meta‐Analytic Mediation Model

To explore whether SC mediated the relationship between CBIs and psychopathology, we applied the OSMASEM method (Jak et al. 2021; Jak and Cheung 2020). The approach requires correlation matrices across studies, including correlations among the three variables of interest. Since the intervention indicator is dichotomous, point‐biserial correlations between CBIs and psychopathology or SC were generally not available. Therefore, we first calculated Cohen's d using the means, standard deviations and sample sizes reported for each group (CBIs vs. control) on both SC and psychopathology. These Cohen's d values were then converted into point‐biserial correlation coefficients using the following formula (De Jonge et al. 2025):

$$r_{\mathrm{pb}}=\frac{d}{\sqrt{d^2+4-\frac{8}{N}}},$$

where N refers to the total sample size. Total scores for psychopathology and SC were used, and higher scores in SC and psychopathology indicated greater levels of these variables.. In some studies, scores on SC were reversed or appeared to be reversed. Detailed criteria for identifying and reversing these scores are provided in Supporting Information, Section S3. In such cases, the scores were inverted (i.e., the sign of the correlation was changed). A negative correlation between the intervention and SC or psychopathology indicated that CBIs were associated with lower levels of these outcomes. For the association between SC and psychopathology, Pearson correlation coefficients were typically reported, so no transformation was necessary.

Once the correlation matrices for all studies were prepared, the OSMASEM method estimated the desired SEM using the studies' correlation matrix data (Jak and Cheung 2020). Full‐information maximum likelihood was used to address any missing correlations. The SEM model fitted in this study was the one presented in Figure 1. Additionally, we estimated the indirect effect of CBIs on psychopathology. Significant direct and indirect effects were interpreted as partial mediation, whereas a significant indirect effect with a nonsignificant direct effect indicated full mediation; indirect effects of 0.01, 0.04 and 0.09 were considered small, medium and large, respectively (Jak et al. 2021).

FIGURE 1.

Meta‐analytic mediation model fitted into this meta‐analysis.

One advantage of the OSMASEM approach is its ability to examine study‐level moderators on SEM parameter estimates. To evaluate whether any study characteristic influenced the indirect effect of CBIs on psychopathology, we tested their moderation effects on parameters a and b, given that the indirect effect is calculated as the product of these two parameters. Moderator analyses were conducted only when there were at least 10 studies available for quantitative moderators and at least five studies for qualitative moderators. Studies with missing data on moderator variables were excluded from the analyses.

In this meta‐analysis, multiple correlation matrices could be extracted from a single study, typically due to several psychopathology outcomes. However, the OSMASEM method is not currently designed to account for the statistical dependency among correlation matrices derived from the same study. As suggested by Bilici et al. (2025), we averaged the correlation matrices within each study to produce a single, study‐level average matrix. Accordingly, we applied this within‐study averaging procedure. A limitation of this approach is that categorical moderator variables that varied within studies (e.g., type of meditation used) could not be examined using the OSMASEM method. However, this limitation was partially addressed by analysing these moderators in a second step, through separate meta‐analyses for each bivariate relationship.

2.6.2. Meta‐Analysis of Overall Effects

Correlation coefficients were converted into Fisher's Z scores prior to the analyses. Each bivariate relationship was synthesised separately. In some studies, multiple correlations were available for a given bivariate relationship (e.g., multiple psychopathology outcomes in the same study). Therefore, we applied three‐level models to account for the dependency that emerged from effect sizes extracted from the same study (Van den Noortgate et al. 2013, 2015). To further protect the parameter estimates from model misspecification, we applied robust variance correction to the results obtained from three‐level models (Fernández‐Castilla et al. 2021; Pustejovsky and Tipton 2022). Predictive intervals, which indicate the range within which future effect sizes are expected to fall, were also calculated (IntHout et al. 2016), with a slight modification to the formula to include the standard error obtained after applying the robust variance correction. Likelihood ratio tests were used to assess whether the between‐study and within‐study variances were substantial.

Moderator analyses were conducted for each bivariate relationship using mixed‐effects, three‐level models with study characteristics as independent variables. For more details on these moderator analyses, see Supporting Information, Section S4. Finally, a set of sensitivity analyses were performed. Firstly, we explored the potential presence of publication bias through the visual inspection of the funnel plots. Secondly, selection models (Vevea and Woods 2005) were implemented to obtain the overall effect corrected by the potential presence of severe or moderate publication bias. In this selection model, the researcher must select the probability that a given study is included based on the reported p value. To model publication bias, selection‐model weights were set to 1 (p < 0.05), 0.75 (0.05 < p < 0.50) and 0.50 (p > 0.50; moderate bias) and 1 (p < 0.05), 0.60 (0.05 < p < 0.50) and 0.30 (p > 0.50; severe bias). Because selection models assume independent effects, we randomly selected one effect size per study, ran both models 200 times and reported the mean corrected estimate. Thirdly, to examine the impact of exceptionally large or small effect sizes, studentised deleted residuals were computed for each effect size (Viechtbauer and Cheung 2010), and those falling outside the ±1.96 range were identified as potential outliers. Additionally, Cook's distance was calculated for each study, using a threshold of 0.45 to flag potentially influential studies (Viechtbauer and Cheung 2010). The primary analyses were then re‐run after removing both outlier effect sizes and influential studies. This manuscript reports pooled correlation coefficients both with and without these data points.

3. Results

3.1. Characteristics of the Included Studies

3.1.1. Study and Sample Characteristics

From 5920 records initially identified, 57 studies (n = 4877) met the inclusion criteria and were retained for the systematic review and the meta‐analysis (Figure 2). Sample sizes ranged from 18 (Asano et al. 2022) to 250 (Kuchar et al. 2023). Samples were nonclinical (49.12%, k = 28), clinical (40.35%, k = 23) or physically ill (10.53%, k = 6). The publications span a 10‐year period, with the earliest study published in 2015 (Shahar et al. 2015) and cover 24 countries (see Table S6 of the Supporting Information for country distribution). Notably, one study was conducted across two countries, and 24.56% of the studies (k = 14) were carried out in non‐Western contexts.

FIGURE 2.

PRISMA flow diagram of the study identification and screening process.

As regards comparison groups, 59.4% (k = 38) of studies used a WL, 37.5% (k = 24) employed active controls unrelated to contemplative interventions and 3.1% (k = 2) included an active control condition that trained meditation (Brito‐Pons et al. 2018; Collado‐Navarro et al. 2021). Some studies appear in multiple categories due to having more than one control group (see the Supporting Information, Table S10, Column 9). Concerning follow‐up, 41 studies (71.93%) included at least one follow‐up assessment (follow‐up timing in Supporting Information, Table S10, Column 12).

3.1.2. Measures of SC

Regarding the assessment of SC, the most commonly used instruments were the Self‐Judgement subscale of the Self‐Compassion Scale (SCS; Neff 2003; 89.47%, k = 51); the Inadequate‐Self subscale of the Forms of Self‐Criticising/Attacking and Self‐Reassuring Scale (FSCRS‐IS; Gilbert et al. 2004; 35.09%, k = 20); and the Hated Self subscale of the same scale (FSCRS‐HS; Gilbert et al. 2004; 21.58%, k = 18). Additional instruments are listed in Table S5 (Supporting Information), and study‐level measures are reported in Supporting Information, Table S10, Column 10.

3.1.3. Measures of Psychopathology

A range of psychopathological symptoms was assessed across the studies, with depressive symptoms being the most frequently measured (98.35%, k = 56), followed by anxiety symptoms (75.44%, k = 43). Other studies assessed eating disorder symptoms (5.26%, k = 3), PTSD (4.92%, k = 3), personality disorders (3.28%, k = 2) and addiction (3.28%, k = 2). Table S5 (Supporting Information) presents all instruments used to assess psychopathological symptoms, while Supporting Information, Table S10 (columns 10 and 11) outlines their distribution across studies.

3.1.4. Risk of Bias Summary

As shown in Figure 3, most studies were rated as having a high risk of bias (76.4%), while 23.6% were rated as having some concerns. The high‐risk classification was primarily due to deviations from the intended intervention (70.9%), since treatment adherence and other related factors were generally not evaluated, and missing outcome data (38.2%), which is associated with significant attrition or inadequate handling of missing data. Studies rated as having some concerns were mostly affected by the selection of reported results, with 87.3% showing concerns due to the absence of preregistration or insufficient specification of variables and planned analyses. Despite the overall high risk of bias, some domains consistently showed low risk. Specifically, 100% of the studies were rated low risk for outcome measurement and 87.3% were rated low risk for the randomisation process. For further details, see Section S6 of the Supporting Information, Tables S7 and S8.

FIGURE 3.

Summary of risk of bias by domain (% of studies).

3.2. Systematic Review Results: Descriptive Findings Across Outcomes

Of the 57 studies included in the review, seven employed more than one control group, yielding a total of 64 independent comparisons (defined as one intervention group versus one control condition). Accordingly, all reported percentages refer to comparisons rather than studies.

Regarding SC, 50% of comparisons (m = 32) did not report data on changes in SC. This omission was typically because the SCS self‐judgement subscale was not reported. CBIs produced statistically significant reductions in SC in 37.5% of comparisons (m = 24), while 12.5% (m = 8) found no significant changes. Among the 24 comparisons that reported significant reductions in SC, 20.83% (m = 5) did not provide follow‐up data, 41.67% (m = 10) maintained the effects and 37.5% (m = 9) did not. When examined by control group type, among WL comparisons (m = 38), 60.5% (m = 23) reported SC outcomes: 52.6% (m = 20) found significant reductions and 7.9% (m = 3) no effect. Among the 20 studies with improvements in SC, 20% (m = 4) did not include follow‐up data, 45% (m = 9) maintained the effects over time and 35% (m = 7) did not. Among active nonmeditation controls (m = 24), 37.5% (m = 9) reported SC outcomes: 16.7% (m = 4) found significant reductions, and 20.8% (m = 5) found no differences. Of the four studies reporting SC improvements, two did not report follow‐up data and one maintained effect whereas one did not. Finally, two studies comparing CBIs with an active meditation control unrelated to compassion (Brito‐Pons et al. 2018; Collado‐Navarro et al. 2021) did not report any SC outcomes.

Regarding psychopathological symptoms, 3.1% of comparisons (m = 2) did not report data, 65.6% (m = 42) showed statistically significant reductions in at least one psychopathological outcome and 31.3% (m = 20) found no significant change. Among the 42 comparisons showing significant improvements, 35.71% (m = 15) did not include a follow‐up assessment; 40.48% (m = 17) maintained the effects at follow‐up, whereas 23.81% (m = 10) did not. By control type, in WL (m = 38), 97.44% (m = 36) reported psychopathology outcomes: 72.7% (m = 28) showed significant improvements and 21.1% (m = 8) found no change. Of WL improvements (m = 28), 71.43% (m = 20) included follow‐up, with 42.86% (m = 12) maintaining effects and 28.57% (m = 8) not doing so. Among active nonmeditation control comparisons (m = 24), 58.3% (m = 14) showed significant improvements, while 41.7% (m = 10) found no change. Among the 14 comparisons reporting improvements, 50% (m = 7) did not report follow‐up data, 35.7% (m = 5) reported maintained effects and 14.3% (m = 2) did not. Finally, across the two comparisons that compared explicit CBIs with an active meditation control unrelated to compassion, neither found significant differences in psychopathological outcomes.

3.3. Meta‐Analytic Mediation Results: The Role of SC

Table 1 presents the meta‐analytic mediation analyses. The indirect effect of CBIs on psychopathology via SC was statistically significant and of large magnitude, and the direct effect also remained significant, indicating partial mediation. Table 2 shows that only meditation duration (minutes) and intervention length (weeks) moderated the model. Longer meditation duration was associated with greater reductions in SC (b = −0.011, p = 0.021), without moderating the SC–psychopathology path; the indirect effect remained significant and slightly increased. Also, longer interventions were associated with a weaker SC–psychopathology association (b = −0.039, p = 0.001) and a slightly smaller, yet still significant, indirect effect.

TABLE 1.

Results of the meta‐analytic mediation model run with OSMASEM.

CBIs → SC (a)	SC → Psyc. (b)	CBIs → Psyc. (c)	Indir. CBIs on Psyc. (a*b)
−0.173 [−0.229, −0.121] k = 39 n = 3730	0.313 [0.188, 0.439] k = 26 n = 2373	−0.130 [−0.188, −0.072] k = 55 n = 4555	−0.054 [−0.085, −0.030]

Abbreviations: CBIs = compassion‐based interventions; k = number of studies; n = total sample size; Psyc. = psychopathology; SC = self‐criticism.

TABLE 2.

Moderator analyses in OSMASEM.

Participant characteristics
	Omnibus test	CBIs → SC (a)	SC → Psyc. (b)	Indir. CBIs on Psyc. (a*b)
Mean age (k = 55)	${\chi }^2$ = 1.73 (p = 0.422)	0 (p = 0.874)	−0.007 (p = 0.174)	−0.046 [−0.075, −0.024]
Mean age IG (k = 52)	${\chi }^2$ = 1.35 (p = 0.508)	−0.001 (p = 0.809)	−0.006 (p = 0.231)	−0.05 [−0.082, −0.025]
Mean age CG (k = 52)	${\chi }^2$ = 1.40 (p = 0.496)	0 (p = 0.863)	−0.007 (p = 0.219)	−0.050 [−0.082, −0.025]
Percentage of men (k = 56)	${\chi }^2$ = 2.12 (p = 0.346)	0.001 (p = 0.604)	−0.003 (p = 0.182)	−0.052 [−0.081, −0.029]
Percentage of men IG (k = 54)	${\chi }^2$ = 1.73 (p = 0.422)	0 (p = 0.758)	−0.004 (p = 0.186)	−0.054 [−0.085, −0.029]
Percentage of men CG (k = 54)	${\chi }^2$ = 12.19 (p = 1)	0.001 (p = 1)	−0.004 (p = 1)	−0.050 [−0.076, −0.029]
Percentage of graduated (k = 35)	${\chi }^2$ = 1.10 (p = 0.576)	−0.001 (p = 0.493)	0.003 (p = 0.133)	−0.052 [−0.089, −0.019]
Percentage of graduated IG (k = 35)	${\chi }^2$ = 1.26 (p = 0.533)	−0.001 (p = 1)	0.003 (p = 1)	−0.052 [−0.089, −0.019]
Percentage of graduated CG (k = 35)	${\chi }^2$ = 0.38 (p = 0.826)	0 (p = 1)	0.003 (p = 1)	−0.052 [−0.09, −0.019]
Percentage of experience with meditation (k = 9)	${\chi }^2$ = 0.92 (p = 0.632)	0.001 (p = 1)	−0.004 (p = 1)	−0.024 [−0.097, 0.048]
Percentage of experience with meditation IG (k = 8)	${\chi }^2$ = 2.42 (p = 0.299)	0.001 (p = 1)	−0.007 (p = 1)	−0.007 [−0.074, 0.064]
Percentage of experience with meditation CG (k = 8)	${\chi }^2$ = 0.94 (p = 0.625)	0.001 (p = 1)	−0.004 (p = 1)	−0.007 [−0.091, 0.084]
Participant categories
General (1): k = 28	1 vs. 3: ${\chi }^2$ = 0.508 (p = 0.776)	1 vs. 3: −0.035 (p = 0.539)	1 vs. 3: −0.048 (p = 0.690)	1 vs. 3: −0.053 [−0.090, −0.025]
Physically ill (2): k = 7	1 vs. 2: ${\chi }^2$ = 0.639 (p = 0.726)	1 vs. 2: −0.075 (p = 0.468)	1 vs. 2: −0.089 (p = 0.640)	1 vs. 2: −0.058 [−0.093, −0.032]
Mentally ill (3): k = 22	2 vs. 3: ${\chi }^2$ = 0.542 (p = 0.723)	2 vs. 3: 0.052 (p = 0.638)	2 vs. 3: 0.194 (p = 0.467)	2 vs. 3: −0.014 [−0.139, 0.120]
Percentage of White participants (k = 21)	${\chi }^2$ = 1.06 (p = 0.590)	0 (p = 0.996)	0.002 (p = 0.289)	−0.060 [−0.112, −0.021]
Ethnic categories: White, k = 7; African American, k = 5; Asian, k = 1; Diverse, k = 10	White vs. African American: ${\chi }^2$ = 0.497 (p = 0.780)	0.054 (p = 0.593)	0.077 (p = 0.564)	−0.074 [−0.151, −0.023]
	White vs. diverse ${\chi }^2$ = 3.343 (p = 0.188)	0.199 (p = 0.057)	0.011 (p = 0.949)	−0.094 [−0.219, −0.027]
	African American vs. diverse ${\chi }^2$ = 1.405 (p = 0.495)	0.140 (p = 0.241)	−0.033 (p = 0.833)	−0.084 [−0.202, −0.007]

Intervention characteristics
	Omnibus test	CBIs → SC (a)	SC → Psyc. (b)	Indir. CBIs on Psyc. (a*b)
Follow‐up categories (k = 57): yes, k = 17; no, k = 40	${\chi }^2$ = 0.883 (p = 0.643)	−0.051 (p = 0.388)	−0.057 (p = 0.649)	−0.047 [−0.101, −0.011]
Delivery (k = 57): online, k = 24; face‐to‐face, k = 33	${\chi }^2$ = 3.239 (p = 0.198)	−0.065 (p = 0.213)	−0.175 (p = 0.136)	−0.057 [−0.104, −0.021]
Modality (k = 52): group, k = 33; individual, k = 24	${\chi }^2$ = 0.718 (p = 0.698)	−0.014 (p = 0.787)	0.090 (p = 0.439)	−0.029 [−0.108, 0.0005]
Percentage of attrition (k = 54)	${\chi }^2$ = 2.46 (p = 0.293)	0.003 (p = 0.12)	0.001 (p = 0.941)	−0.055 [−0.087, −0.03]
Session length (minutes; k = 48)	${\chi }^2$ = 4.74 (p = 0.093)	−0.001 (p = 0.099)	−0.003 (p = 0.062)	−0.059 [−0.093, −0.033]
Number of sessions (k = 57)	${\chi }^2$ = 1.23 (p = 0.540)	−0.003 (p = 0.629)	−0.010 (p = 0.273)	−0.054 [−0.084, −0.03]
Session interval (days; k = 51)	${\chi }^2$ = 0.64 (p = 0.725)	0.006 (p = 0.471)	0.010 (p = 0.579)	−0.055 [−0.092, −0.026]
Meditation duration (minutes; k = 21)	${\chi }^2$ = 5.76 (p = 0.056)	−0.012 (p = 0.021)	−0.008 (p = 0.489)	−0.065 [−0.139, −0.023]
Intervention length (weeks; k = 57)	${\chi }^2$ = 9.70 (p = 0.008)	−0.011 (p = 0.139)	−0.039 (p = 0.001)	−0.051 [−0.078, −0.031]

Abbreviations: CBIs, compassion‐based interventions; CG, control group; IG, intervention group; Psyc., psychological symptoms; SC, self‐criticism.

3.4. Meta‐Analysis for Each Bivariate Relationship

3.4.1. Effects of CBIs on SC

A total of 39 studies reporting 74 effect sizes were synthesised (forest plot in Supporting Information, Figure S1). The overall correlation between CBIs and SC was negative and moderate (r = −0.20, 95% CI [−0.27, −0.13], 95% PI [−0.56, 0.22]). Significant heterogeneity was found both between‐studies (${\widehat{\sigma }}_{\text{studies}}^2=$ 0.04, LRT = 16.24, p < 0.001) and within‐study (${\widehat{\sigma }}_{\text{within}}^2=0.01$, LRT = 10.09, p < 0.001). After correcting for publication bias, the overall effect remained similar (r = −0.24 for severe bias and r = −0.23 for moderate bias). Figure S4 shows the funnel plot. Four effect sizes were detected as potential outliers (Z = −0.92, Z = −0.83 and Z = −0.71 from Carter et al. 2023, and Z = −1.02 from Gharraee et al. 2018). After removing them, the overall effect slightly decreased to −0.16 but remained statistically significant (95% CI [−0.22, −0.11]). No influential studies were detected.

As regards the moderator analyses, the type of SC measure moderated the effectiveness of CBIs on SC (see Supporting Information, Table S9). CBIs were significantly less effective in reducing SC when measured with FSCRS‐HS (r = −0.104, m = 17) than with FSCRS‐IS (r = −0.263, m = 19; Z = −3.27, p = 0.003).

3.4.2. Effects of CBIs on Psychopathological Symptoms

A total of 125 effect sizes from 55 studies were synthesised to calculate the correlation between CBIs and psychopathology (forest plot in Supporting Information, Figure S2). The correlation was negative and moderate (r = −0.21, 95% CI [−0.27, −0.15], 95% PI [−0.56, 0.20]). Significant heterogeneity was detected between‐studies (${\widehat{\sigma }}_{\text{studies}}^2=$ 0.04, LRT = 30, p < 0.001) and within‐study (${\widehat{\sigma }}_{\text{within}}^2=$ 0.01, LRT = 11.69, p < 0.001; funnel plot in Supporting Information, Figure S5). After applying a publication bias correction, the overall effect slightly increased (r = −0.23 for moderate bias and r = −0.25 for severe bias). Nine effect sizes were detected as potential outliers (Z = −0.79 from Savari et al. 2021; Z = −0.80 from Carter et al. 2023; Z = −0.77 and Z = −0.90 from Gu et al. 2022; Z = 0.51 from Weibel et al. 2017; Z = −0.85 from Sajjadi et al. 2023; Z = 0.22 and Z = 0.20 from Kuchar et al. 2023; Z = 0.23 from Stevenson et al. 2019). After removing them, the overall effect size remained stable (r = −0.21). No influential studies were detected.

The type of control group moderated the relationship between CBIs and psychopathology (F = 16.26, p < 0.001), with CBIs being more effective when compared to a WL (r = −0.262, m = 80) than to active control (r = −0.066, m = 29; difference = 0.196, Z = 3.03, p = 0.007). Meditation length (in minutes) also influenced outcomes. Longer meditation duration was associated with greater psychopathology reductions (b = −0.008, p = 0.034). Finally, the type of psychopathology also moderated this relationship: the effect of CBIs was stronger on depression (r = −0.235, m = 62) than on anxiety (r = −0.175; m = 51; difference = 0.060, t = 2.56, p = 0.015). All moderator results are reported in Supporting Information, Table S9.

3.4.3. Association Between SC and Psychopathology

The synthesis of the 99 effect sizes (from 26 studies) shows a positive and strong correlation between SC and psychopathology (r = 0.48, 95% CI [0.42, 0.54], 95% PI [0.10, 0.74]; forest plot in Supporting Information, Figure S3). Significant heterogeneity was found both between (${\widehat{\sigma }}_{\text{studies}}^2=0.02$, LRT = 19.58, p < 0.001) and within‐studies (${\widehat{\sigma }}_{\text{within}}^2=0.02$, LRT = 57.06, p < 0.001). The funnel plots can be found in Figure S6. After correcting for publication bias, the overall effect remained stable (0.48 for severe bias and 0.49 for moderate bias). Seven effect sizes were potential outliers (Z = 1.21, Z = 1.03 and Z = 0.99 from Leboeuf et al. 2022; Z = 1.16 from Feliu‐Soler et al. 2017; Z = 1.10 from Kelly and Carter 2015; Z = 0.064 and Z = 0.08 from Paucsik et al. 2024). After removing them, the overall effect fell slightly to 0.47. No influential studies were detected and no study characteristic significantly moderated the overall effect (Supporting Information, Table S9).

4. Discussion

Understanding the MoCs through which an intervention produces therapeutic effects is crucial not only for advancing theoretical knowledge but also for enhancing clinical effectiveness (Holmes et al. 2018). In CBIs, these processes remain insufficiently explored despite growing evidence of efficacy in reducing psychopathological symptoms. To address this gap, this study examined whether SC mediates the beneficial effects of CBIs on psychopathology.

This review and mediation meta‐analysis indicate that research on the concurrent effects of CBIs on SC and psychopathology is relatively recent (with the first study in 2015) and growing, but remains methodologically heterogeneous in design, delivery and sample size, limiting comparability. The evidence is culturally narrow (24.56% non‐Western samples), constraining generalisability beyond WEIRD contexts. Measurement is uneven: psychopathology symptoms are mostly anxiety/depression, with PTSD, EDs, personality disorders and addictions underrepresented; SC was assessed with the SCS in ≈89% of studies, but the self‐judgement subscale was often not reported. Risk‐of‐bias results are concerning (≈76% high risk), with inadequate missing‐data handling and limited preregistration, highlighting the need for stronger adherence monitoring, appropriate missing‐data methods and preregistration. Despite these limitations, studies reported significant reductions in SC and psychopathology, consistent with prior meta‐analyses and promising for clinical application (e.g., Kirby et al. 2017; Ferrari et al. 2019; Han and Kim 2023; Luo et al. 2023; Wakelin et al. 2022). Overall, the field is expanding and the findings are encouraging, but greater conceptual and methodological standardisation is required to consolidate evidence on how CBIs influence SC and mental health outcomes.

The study found that SC partially mediated the relationship between CBIs and psychopathological symptoms, suggesting that CBIs may exert therapeutic effects in part by reducing SC. This aligns with Gilbert's (2009b, 2014) three‐systems model (threat, drive, soothing), under which compassion is theorised to activate the soothing system (linked to safeness and care) and thereby down‐regulate threat reactivity. Within this framework, SC is viewed as one relevant trigger of the threat system (Longe et al. 2010), whereas self‐reassurance, a compassionate way of responding to oneself, acts as a protective mechanism that helps to calm the mind and reduce self‐critical thoughts (Gilbert 2014). Consistent with this view, self‐reassurance can buffer SC's impact on mental health, particularly on depressive symptoms (Petrocchi et al. 2019). Therefore, reducing SC through CBIs not only alleviates emotional suffering but also promotes healthier emotion regulation processes involved in psychopathology. Notably, the only study testing SC as a mediator offers mixed support. Sommers‐Spijkerman et al. (2018) reported significant SC mediation in single‐mediator models but not when multiple mechanisms were modelled concurrently, possibly due to measurement and sample characteristics. Firstly, SC was assessed with FSCRS‐Inadequate‐Self, which typically shows smaller change estimates than more severe facets (e.g., Hated‐Self), potentially underestimating SC‐mediated effects. Secondly, the nonclinical sample may have reduced observable effects due to lower baseline psychopathology (Biermann et al. 2021). Overall, these findings suggest that, while SC is a plausible MoC, future research should test its interactions with other MoCs and their temporal sequencing to clarify how CBIs work.

Moreover, although mediation remained significant, two moderators emerged: daily meditation duration and overall intervention length. Longer daily meditation practice (ranging from 5 to 40 min) was associated with a slightly stronger indirect effect through SC. Evidence from other contemplative interventions, such as mindfulness‐based programs, on practice duration is mixed. Some findings support a positive association between practice duration and symptom improvement (Parsons et al. 2017), or nonjudgemental internal experience (Cebolla et al. 2017), whereas others report null effects when comparing shorter versus longer practices (Fincham et al. 2023; Palmer et al. 2023) or even the opposite pattern (Strohmaier et al. 2021). This heterogeneity suggests that outcomes may depend not only on practice quantity but also on practice quality. Recent work highlights compassion practice quality as a predictor of better outcomes (Andreu et al. 2022; Navarrete et al. 2021; Zeng et al. 2023). Moreover, preliminary data from an ecological momentary assessment study suggests that quality may outperform cumulative adherence in predicting change (Palacios et al. 2024). Future research should therefore assess both the amount and quality of practice to clarify mechanisms.

Moreover, overall intervention moderated SC mediation. Longer interventions showed smaller effects on psychopathology, reducing SC's mediating role. In our mixed clinical/subclinical dataset, SC may decrease early, with other mechanisms (e.g., self‐reassurance) becoming more influential later, which could explain weaker SC mediation in longer interventions. This aligns with Sommers‐Spijkerman et al. (2018), who hypothesise for clinical samples that sequence, mitigating SC first so as to ‘make room’ for self‐reassurance, whereas their empirical results come from subclinical samples, in which they propose consolidating self‐reassurance before reducing SC. Since we did not analyse additional mediators, this remains speculative. Future studies should model time‐ and sample‐dependent mediation, include multiple concurrent mediators, and account for ‘dose’ and content proportion (modules/topics) to elucidate how pathways of change are redistributed. Although the moderation terms were statistically significant, their incremental effects were modest. Even so, the findings underscore the need for further investigation into the MoCs and suggest the importance of identifying additional explanatory variables that may serve as alternative pathways of therapeutic benefit.

As a secondary aim, CBIs showed a significant and moderate negative effect on SC, supporting their ability to target this process and extending prior meta‐analytic evidence (Millard et al. 2023; Vidal and Soldevilla 2023; Wakelin et al. 2022). Additionally, the systematic review also supported this pattern, with most studies reporting decreased SC. Moreover, effects varied by SC measure: the FSCRS distinguishes inadequate‐self (failure/not good enough) from hated‐self (contempt/disgust and self‐rejection; Gilbert et al. 2004). Our results suggest CBIs reduce less severe SC (inadequate‐self) more than the more hostile hated‐self, which appeared less responsive to change. Although one meta‐analysis found larger hated‐self than inadequate‐self reductions, it included only seven RCTs (Vidal and Soldevilla 2023). By contrast, our findings align with evidence that hated‐self is more rigid and resistant to change (Castilho et al. 2017; Werner et al. 2019). This suggests that CBIs may be more effective when SC is rooted in inadequacy and less so when it stems from deeper, more hostile self‐relating patterns. One possible explanation may lie in differences in the emotional tone of these two forms of SC. Whereas inadequate‐self is characterised by feelings of failure, inadequacy and not being good enough, hated‐self involves a more persecutory way of relating to oneself, characterised by self‐contempt, disgust and hatred (Gilbert 2015; Gilbert et al. 2004). This more hostile emotional basis may contribute to the greater rigidity and resistance to change of hated‐self, making a shift toward a more affiliative and compassionate way of relating to oneself particularly relevant, but potentially more difficult to achieve.

Moreover, CBIs showed a significant and moderate negative effect on psychopathological symptoms. This aligns with our systematic review, where significant improvements were more common than null effects. This result is consistent with previous meta‐analyses supporting the efficacy of CBIs in reducing psychological symptoms (Kirby et al. 2017; Millard et al. 2023; Winders et al. 2020), indicating direct clinical benefits beyond intermediate mechanisms. In addition, three moderators were identified: the type of control group, the duration of the meditation practice and symptom type. CBIs showed stronger correlations with reduced psychopathological symptoms when compared to passive controls than to active controls; also the review similarly showed larger effects versus WL. This finding aligns with previous meta‐analyses, which have reported larger effects in reducing symptoms (particularly anxiety and depression) when CBIs are compared to passive controls, as opposed to active ones (Wilson et al. 2018). Meditation duration showed a small moderation effect, but evidence remains limited and contradictory, as already discussed above when examining its moderating role in the mediation model, where a greater number of practice minutes translated into slightly smaller reductions in psychopathology, although the effect was subtle. Finally, these interventions were found to have a greater effect on depressive symptoms than on anxiety symptoms. Previous literature reports mixed findings on this matter: while one meta‐analysis contradicts our results (Holzer et al. 2024), others support them, reporting larger effect sizes for the reduction of depressive symptoms compared to anxiety symptoms (Ferrari et al. 2019; Han and Kim 2023; Kirby et al. 2017). These discrepancies highlight the need to clarify disorder‐specific effects and the MoCs underlying these differential outcomes.

Finally, a significant positive association was observed between SC and psychopathology, with a large effect size, indicating that higher levels of SC are consistently linked to greater psychological symptoms. This finding is consistent with previous research identifying SC as a transdiagnostic factor involved in both the development and maintenance of various mental disorders (Gilbert and Irons 2005; Warren et al. 2016; Wilson et al. 2019; Zaccari et al. 2024). Specifically, SC has been associated with multiple disorders (e.g., depression, anxiety, obsessive‐compulsive disorder, PTSD, eating disorders, schizophrenia and personality disorders; Gilbert and Irons 2005; Zaccari et al. 2024). Moreover, early onset of SC has been shown to predict later development of different forms of psychopathology (Gilbert and Irons 2005), and its persistence contributes to symptom maintenance, since individuals with high levels of SC tend to respond more poorly to both psychological and pharmacological treatments compared to those with lower levels (Warren et al. 2016).

Furthermore, evidence suggests that the relationship between SC and depressive symptoms may be mediated by fear of compassion (Joeng and Turner 2015). Fear of compassion is defined as avoidant or fearful reaction to affiliative and compassionate emotions, which may be perceived as threatening or unpleasant (Gilbert et al. 2011). In particular, individuals with higher levels of comparative SC may experience greater fear of self‐compassion, as they may perceive self‐compassion as interfering with their own or others' high expectations, which may, in turn, contribute to greater depressive symptoms (Joeng and Turner 2015). In other words, fear of compassion may attenuate the buffering effects of self‐compassion on psychopathology while amplifying the impact of perceived threats (Matos et al. 2022). This highlights the need for future research to develop models that include multiple mediators, such as fear of compassion, to better understand how to reduce the impact of SC on psychopathology and, ultimately, increase the effectiveness of CBIs.

In summary, the findings deepen our understanding of how CBIs work by shedding light on the underlying mechanisms that support their effectiveness. Although CBIs have demonstrated beneficial effects across a range of mental health outcomes (e.g., Kirby et al. 2017; Millard et al. 2023; Winders et al. 2020), the mechanisms underlying these effects are still insufficiently understood. Identifying SC as a potential MoC contributes to a process‐based research agenda by shifting the focus from whether CBIs work toward understanding how and why they may work. It also directs attention toward a more dynamic, process‐oriented understanding of these interventions.

This contribution has several important implications. From a theoretical perspective, it represents a shift from a predominantly efficacy‐focused approach toward understanding the processes through which CBIs exert their effects, a perspective well established in research on other psychotherapies (Dragioti et al. 2017; McNally 2007), but comparatively less developed in CBIs (e.g., Kirby and Gilbert 2019). This shift is important for moving beyond established evidence of efficacy toward identifying pathways to optimise effectiveness. From a methodological perspective, OMASEM demonstrates the value of innovative meta‐analytic approaches for investigating mediations in a field where individual studies have rarely directly tested it.

By integrating evidence across studies, this approach helps narrow the set of candidate processes warranting further investigation in idiographic and longitudinal designs, thereby opening new opportunities for process‐oriented research in CBIs. From a research perspective, these findings provide a basis for future studies to examine the temporal dynamics of SC, which typically require greater resources, and determine whether changes in SC precede improvements in psychopathological symptoms, a key step toward establishing SC as MoC. Finally, from a clinical perspective, SC may be an important process to assess and monitor throughout treatment, potentially providing information about therapeutic progress. Developing or strengthening treatment modules specifically targeting SC may also offer a promising avenue for optimising CBIs, although future research is needed to determine whether explicitly targeting SC improves treatment outcomes.

4.1. Strengths and Limitations

This study has several strengths. First, using OMASEM enabled a more precise test of mediation than traditional bivariate meta‐analyses (Ng et al. 2023) and allowed moderator detection (Cheung 2022), improved power and maximised partial data use across trials (Ng et al. 2023). Second, the review was comprehensive, combining quantitative and qualitative syntheses to identify broad patterns and methodological heterogeneity. Study‐level descriptions also contextualised the state of CBI research and sources of heterogeneity. Finally, this work fills a gap by providing the first mediation‐based synthesis of SC in CBIs, supporting more mechanism‐informed intervention development.

This study also has several limitations. Firstly, although the OMASEM methodology improves the estimation of mediation effects, the analysis was constrained by the lack of primary data, as 34 out of 61 studies did not report the necessary correlations or failed to respond to data requests. Furthermore, OMASEM does not experimentally manipulate mediators; thus, indirect effects reflect meta‐analytic statistical mediation, not causal mechanisms (Kazdin 2007).

Second, substantial methodological heterogeneity across protocols, duration, samples and measures limits comparability. In addition, although eligibility was restricted to interventions in which compassion constituted at least 50% of the programme content, some interventions also incorporated components from other therapeutic approaches (e.g., ACT or DBT). However, overlap with mindfulness‐based interventions should be interpreted differently, as many CBIs incorporate mindfulness practices (e.g., Gilbert and Procter 2006; Jazaieri et al. 2013; Neff and Germer 2013), and mindfulness practice may also be associated with compassion‐related processes (Brito‐Pons et al. 2018). Thus, differences between approaches may partly reflect a differential emphasis on specific techniques rather than entirely distinct therapeutic mechanisms. Nevertheless, this heterogeneity with other components warrants some caution in interpreting the findings, as the observed effects may partly reflect the contribution of these additional therapeutic components.

Third, risk of bias was often high, especially due to deviations from intended interventions and missing outcome data, potentially weakening internal validity and the robustness of pooled conclusions.

Fourth, it is difficult to determine whether changes in SC precede clinical improvement, since most studies included only pre‐ and postintervention assessments. Finally, the majority of studies focused on symptoms of depression and anxiety, with limited representation of other psychopathologies such as trauma, addictions or eating disorders, which restricts the generalisability of findings to other clinical populations and their transdiagnostic character.

4.2. Future Directions

Several challenges remain for future research on CBIs. One priority is to design studies that analyse the temporal sequencing of mechanisms to confirm SC as a MoC (Kazdin and Nock 2003). Current evidence increasingly highlights that SC fluctuates dynamically within individuals, even over short time spans (Fraser et al. 2025; Mey et al. 2023). Therefore, methodologies featuring repeated measurements, such as ecological momentary assessment, offer a promising avenue to examine the dynamics between SC and symptoms. Furthermore, due to the number of studies excluded for lack of data and the high variability observed in protocols, advancing toward more open science practices is essential. Sharing datasets and protocols would enhance transparency, facilitate cross‐study comparisons and help clarify the underlying mechanisms. Reducing the risk of bias is also another priority; it involves ensuring proper assessment of treatment adherence, applying appropriate analyses in the presence of missing data, preregistering studies and reporting any deviations from planned protocols. Moreover, more studies conducted in non‐WEIRD contexts are needed. Future research could use frameworks such as ADDRESSING (Hays 2001) to systematically identify which groups remain underrepresented and to guide culturally‐sensitive adaptations of CBIs. In parallel, greater attention should also be given to underrepresented psychopathologies, such as eating disorders, personality disorders, PTSD and addictions, so as to improve the generalisability of findings across different clinical conditions. Additionally, because the present meta‐analysis treated psychopathology as a composite construct (i.e., averaging across symptom domains), future work should estimate domain‐specific mediation models to test whether the mediating role of SC differs across specific symptom clusters. Finally, exploring other proposed MoCs, such as compassion flows, fear of compassion, common humanity, empathic concern and adaptive emotion regulation, and integrating them into unified models would be valuable. This approach could advance process‐based psychotherapy toward greater precision and efficacy.

5. Conclusions

This work shows that SC partially mediates the effect of CBIs on psychopathology, with a slightly varying indirect effect depending on two moderators: the daily duration of meditation practice and the overall length of the program. This mediation suggests that part of the clinical benefit of CBIs can be explained by the reduction of patterns of self‐disapproval and self‐judgement, thus making SC a target process within CBIs. In line with Gilbert's (2009a,b) three‐system model of emotion regulation, cultivating compassion helps to balance the threat system, which may be activated by SC, thereby promoting a more adaptive mode of emotional regulation that is less associated with psychopathology. In terms of application, these findings reinforce the importance of assessing and directly addressing SC, incorporating strategies to recognise and reframe the internal critical dialogue. However, the conclusions should be interpreted with caution due to a risk of methodological bias (a lack of systematic evaluation of adherence to practice and inadequate handling of missing data), as well as heterogeneity in psychopathology measures and inconsistency in SC scales. Future studies should examine how SC interacts dynamically with other MoCs within multivariate models and promote greater methodological standardisation.

Disclosure

This study was preregistered in PROSPERO (CRD42023480668). In line with transparency standards, the full search strategies are reported in the Supporting Information (Section S1), and the PRISMA 2020 guidelines (Page et al. 2021) were followed throughout the review process.

Ethics Statement

Ethical approval was not required for this study because it is a meta‐analysis of previously published studies and did not involve direct interaction with human participants, the collection of personal data, or the use of human biological samples.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Search strategy applied in ProQuest database.

Table S2: Search strategy applied in Scopus database.

Table S3: Search strategy applied in WOS database.

Table S4: Search strategy applied in PubMed database.

Table S5: Instruments used to measure key variables.

Table S6: Frequency of countries in which studies were conducted.

Table S7: Risk of bias per study and domain.

Table S8: Overall distribution of risk of bias across domains.

Figure S1: Relationship between CBI and self‐criticism.

Figure S2: Relationship between CBI and psychopathology.

Figure S3: Relationship between self‐criticism and psychopathology.

Figure S4: Funnel plot of the relationship between CBI and self‐criticism.

Figure S5: Funnel plot of the relationship between CBI and psychopathology.

Figure S6: Funnel plot of the relationship between self‐criticism and psychopathology.

Table S9: Moderator analysis for each bivariate Pearson correlation.

Table S10: Characteristics of included studies and intervention details.

Data Availability Statement

The dataset and relevant code used to reproduce the results are openly available in the OSF repository: https://osf.io/7wsyf/.