The Impact of Mindfulness-Based Programmes on Self-Compassion in Nonclinical Populations: a Systematic Review and Meta-Analysis

Titles and abstracts from three online databases (PsycINFO, PubMed and Web of Science) up to 26th January 2020 were systematically searched using the terms [(Mindful* OR MBCT OR MBSR OR medit* OR “compassionate mind” OR “compassion focused” OR CFT OR ACT OR “acceptance and commitment” OR DBT OR “dialectical behaviour therapy”) combined with (depres* OR anx* OR stress OR wellbeing OR “well-being” OR “well being” OR “mental health” OR “psychological health”) combined with (“self-compassion” OR “self compassion” OR compassion) combined with random*]. All abstracts identified were screened by the first author. Full texts from the screen were reviewed and assessed for eligibility. A subset (12 studies) of screened full texts were reviewed by a second reviewer, and any disagreements were resolved by discussion. Cohen’s kappa could not be calculated for inter-rater reliability; however, the two raters agreed on 11 studies with the exception of one inclusion where discussion was required.

The manual search was conducted by the first author who hand-searched the reference lists of all eligible studies at the full-text screening stage. If inclusion could not be determined via details in the published study, authors were contacted directly by the researcher for further information. Where studies were part of wider trials, or linked to other published articles, these were consulted when further information was required.

Studies were included for review according to the following criteria: (1) RCTs; (2) participants were aged 18 or over; (3) programme was mindfulness-based, which for this review was specified as a programme encouraging home practice for at least 2 weeks duration (as to encompass brief interventions but exclude “one-off” practices), and at least half the practices consisting of mindfulness meditation; (4) data included validated psychometric subjective measures of self-compassion pre- and post-programme; and (5) data included validated psychometric subjective measures of one or more of state depression, anxiety or stress. Exclusion criteria were as follows: (1) participants recruited on the basis of a diagnosable mental or physical health problem via either a clinical setting or meeting clinical threshold on standardized measures and (2) not published in the English language.

Extraction of the following data was conducted: (1) participants, including number, method of recruitment, inclusion/exclusion criteria and other demographic characteristics; (2) intervention, including format (e.g. MBSR, MBCT, compassion-based), duration, amount of home practice encouraged and delivery method (e.g. group, self-help); (3) control group details; (4) programme adherence; (5) outcomes, including primary outcomes of interest and measures used; and (6) statistical findings regarding self-compassion and measures of depression, anxiety or stress were extracted. For the SCS, effect sizes for total scores were calculated for inclusion in the meta-analysis. Correlations between change in self-compassion and distress post-MBP were also extracted. Subscale scores were extracted if available. Authors were contacted for required data if not reported directly in the text (this was not possible with unpublished studies).

A meta-analysis was conducted for self-compassion outcomes, using the statistical software program Review Manager version 5.3 (The Cochrane Collaboration 2014). Effect sizes for total self-compassion scores were calculated using Hedges g (pre-post change for the MBP group minus the pre-post change for the control group, divided by the pre-programme standard deviation multiplied by Hedges correction factor (Carlson and Schmidt 1999; Hedges 1981; Morris 2008)). This effect size has been found to provide a better estimate in terms of bias, precision, and heterogeneity of variance, and takes into account baseline differences between groups (Morris 2008). Effect sizes were interpreted as 0.2 = small, 0.5 = medium and 0.8 = large (Cohen 1992). Effect sizes for studies with a design that randomized participants to three groups were calculated by combining mean, standard deviation and sample size for either MBP (where two MBP conditions were delivered) or control groups (where two different control conditions were used) to provide a weighted mean and estimated standard deviation. This is preferable to entering multiple comparisons for a single study, to avoid unit-of-analysis error resulting from unaddressed correlations between the estimated programme effects from multiple comparisons (The Cochrane Collaboration 2011).

A random-effects model was selected to examine the difference in pre-post self-compassion scores between the MBP group and control group in each study, resulting in an overall effect size based on weighted averages of the effect sizes. This model was chosen over a fixed-effects model due to the different populations sampled in the review and the degree of heterogeneity of MBPs included, which fits with the model’s assumptions that observed estimates of effect size may vary systematically across studies (Borenstein et al. 2009; Field and Gillett 2010). A random-effects model also allows for findings to be generalized beyond the subset of included studies. A forest plot was created for a visual representation of effect sizes for each study and allowed for inspection of outliers.

Heterogeneity of effect sizes was assessed using Cochran’s Q statistic (χ²) to investigate whether there was significant variation between studies, and the I² index to measure the magnitude of heterogeneity. A significant Q statistic (p < 0.05) indicates that the variance of the results may be due to varying effect sizes, resulting from possible sampling or methodological variance across the included studies. I² is interpreted as the degree of variability among effect sizes across studies due to heterogeneity rather than chance/sampling error, reported as a percentage. This can be interpreted as 0% indicates homogeneity; 25% indicates small levels of heterogeneity; 50% is medium; and 75% is large (Huedo-Medina et al. 2006).

Post hoc subgroup analyses were performed, to investigate potential sources of heterogeneity. As there were no pre-specified hypotheses regarding potential moderating variables that would impact on the effect on self-compassion, these analyses were considered exploratory. The moderating effects of explicit compassion components (MBP with no added compassion/kindness meditations or focus vs. MBP with additional compassion/kindness components), psychological distress as a participation inclusion criterion (minimum distress threshold measure vs. no threshold) and type of control group (wait-list vs. active control) were investigated. This was assessed by categorizing the studies according to the proposed moderator variable (or if applicable, categorizing experimental groups within a study) and comparing their separately pooled effect sizes, using a between-group heterogeneity statistic Q_between (Hedges and Olkin 1985). A significant moderator effect is indicated by a statistically significant Q_between.

A meta-regression was not indicated due to the number of studies that were available for analysis (The Cochrane Collaboration 2011); however, effect sizes for the association between self-compassion and distress were calculated using Fisher’s Z transformation for correlation coefficient r to avoid underestimation of effect (Borenstein et al. 2009) and weighted according to sample size with standard error 1/ √ (N − 3) and entered into a random-effects model. This was then back-transformed to r values for ease of reporting. Effect sizes were interpreted as small = 0.1, medium = 0.3 and large = 0.5 (Cohen 1992).

Publication bias was assessed by creating a funnel plot of effect sizes against standard error and visually inspecting the plot for signs of asymmetry. Asymmetry could be due to potential bias due to the under-reporting of negative or null results. Orwin’s (1983) failsafe N was calculated to determine the number of studies with null results (set at d = 0) needed to reduce the effect size to a negligible level (set at d = 0.1). This statistic was chosen due to the ability to select a nonzero effect size for the missing studies. The trim and fill method (Duval and Tweedie 2000) was considered to impute missing studies and compute an adjusted mean effect size taking missing studies into account; however, this method was not used due to the bias introduced by the heterogeneity found in the analysed studies, which has been found to affect the validity of this method (Peters et al. 2007; Terrin et al. 2003). This analysis strategy was used for total SCS scores both immediately post-programme and at follow-up in separate models to avoid unit-of-analysis error (The Cochrane Collaboration 2011). It was also hoped to synthesise effect sizes for SCS subscales; however, calculation and reporting of scores in the negative subscales (self-judgement, isolation and overidentification) varied between studies (using reversed items vs. nonreversed) and precluded any estimate of overall effect across the studies.

Studies meeting the criteria for review were assessed for methodological quality using an adapted version of the Quality Assessment Tool for Quantitative Studies (National Collaborating Centre for Methods and Tools 2008). The use of this checklist was modified for the purposes of the current study by (1) omitting an overall score, given the questions around the suitability of their application (Centre for Reviews and Dissemination 2009); (2) altering the weighting of scores in the Selection Bias component, where self-referral rating was changed from “weak” to “moderate”, as the nature of voluntary psychological programmes such as MBPs is more likely to produce studies that recruit via self-referral; (3) altering the Study Design component to include power calculation and removing use of a RCT design as a criterion for “strong”, as only RCTS were being assessed in the current review; (4) focusing on the outcomes of interest for the current review when assessing for quality of data collection methods; and (5) developing ratings for the Intervention Integrity and Analysis components, with a particular focus on the analysis of self-compassion outcomes rather than the study’s primary outcomes. This resulted in a set of 16 questions comprising eight components: (Selection Bias, Design, Confounders, Blinding, Data Collection Methods, Withdrawals and Dropouts, Intervention Integrity, Analysis) with each of these components receiving a “strong”, “moderate” or “weak” rating. No studies were removed based on quality assessment, as this can lead to overexclusion, therefore potentially limiting the validity of the results (Meline 2006).

Figure 1 illustrates the process of study selection according to PRISMA guidelines (Moher et al. 2009). The literature search yielded 849 articles, of which 307 were duplicates. After removal of duplicates, 598 titles and abstracts were screened and 141 articles were reviewed based on the full text. This resulted in the final selection of 26 studies, published between the years 2005 and 2020, conducted in the USA (n = 12), UK (n = 4), Spain (n = 2), Australia (n = 1), Belgium (n = 1), Canada (n = 1), Hong Kong (n = 1), the Netherlands (n = 1), Norway (n = 1), Romania (n = 1) and Sweden (n = 1). A summary of study characteristics can be found in Table 1.

All but one (which used random sampling) of the studies recruited using opportunity sampling, with 13 recruiting from universities and 13 recruiting from the community, reflecting an overrepresentation of the student population. A number of the studies sampling from the community targeted specific groups from a variety of sources, such as mothers, carers and specific workplaces. While most of the studies had no specific inclusion criteria in terms of psychological distress, seven of the studies only recruited participants with a minimum score on a measure of interest (e.g. neuroticism, perfectionism, stress). The total number of participants included across the studies was 2025, with sample sizes ranging from 13 to 83 for an MBP group and between 13 and 100 for a control group, reflecting a large range of sample sizes with many that were pilot studies and/or underpowered. A large range of ages with predominantly female participants comprised the studies’ participant sample; weighted mean age was 35.2 years across the 22 studies that reported age and mean percentage female was 78% (range 46–100%).

Of the included RCTs, 11 were labelled as pilot studies; these studies represented the smallest sample sizes; however, some of the RCTs were also underpowered. In 23 of the studies, participants were randomized into one of two groups (MBP vs control) and in three studies, a three-group randomized design was used (either two different MBPs vs control, or MBP vs active control and wait-list). The majority (21) of the included studies used wait-list control groups; however, five studies used other evidence-based psychological programmes for the target problem for their active control condition (CBT-based self-help, psychoeducational self-help, relaxation response programme, progressive muscle relaxation) and another study used an activity matched for duration and relaxing auditory content (classical music listening). Self-compassion was not a sole primary outcome in any of the studies, but was listed alongside other measures of mindfulness and psychological symptoms as primary outcomes in seven of the studies. One study only analysed self-compassion as a mediating factor but also reported pre-post-programme scores. As well as measuring self-compassion (and other variables, including depression, anxiety or stress) at baseline and post-programme, nine studies took follow-up measures (between 8 weeks and 1 year post-programme). The current sample of studies therefore favoured wait-list control in their RCT design and were primarily interested in indices of distress alongside self-compassion following MBPs, rather than assessing the association between the two.

All of the included studies used the SCS in either the long (n = 15) or short (n = 11) form to measure self-compassion. The majority of studies (24) used total score (either average or summed) as an index of self-compassion, and one study used only the positive subscales (Self-Kindness, Common Humanity and Mindfulness) of the SCS as a composite score. Another study presented the subscales of the SCS (which were also detailed in seven other studies) without reporting a total score. In terms of general measures of psychological distress, stress was measured in the majority of the studies (21), while depression was investigated in 14 and anxiety in 13; nine measured all three.

The majority (23) of the included studies described the MBP as MBSR- or MBCT-based, the remaining three studies used a 2-week (Gu et al. 2018) or 8-week (Huberty et al. 2019) self-help MBP or “brief integrated practice” group-delivered mindfulness programme (Arredondo et al. 2017). Of these studies, 11 described the MBP used as MBSR/MBCT adapted for their target population and eight incorporated additional kindness and compassion components. Group format was the dominant delivery method of MBP across the sample of studies (19 delivered weekly sessions, one study delivered sessions twice weekly and one study had twice weekly sessions in the first 2 weeks followed by weekly). For the group-based programmes, session duration ranged from 75 to 120 min, and total number of sessions ranged from three to nine (median number of sessions = 8). Although the majority of studies described their MBP as adapted MBCT/MBSR, only a small number incorporated a retreat into their programme (one half day, two full days and one with two full days). Total contact time ranged from 4.5 to 16 h (median = 16 h). Of the 21 group-delivered programmes, 15 reported facilitators with specific training in delivering MBPs (one who was a peer facilitator); the remaining studies described facilitators as being trained in CBT and had experience of mindfulness (n = 1), staff from a university counselling service (n = 2), licenced psychologists (n = 1) and one of the authors of the paper (n = 1), or did not report the expertise or training of facilitators (n = 2). A smaller number of studies investigated MBPs delivered through self-help (three with access to a website, one via a mobile app and one through a book), all of which provided extra information (either via psychoeducation webpages or book chapters) and encouraged daily home practice. Duration of self-help programmes ranged from 2 to 8 weeks. Home practice was reported as part of the MBP in 23 of the included studies, with 16 detailing frequency and duration of recommended formal practice. All but one of these studies stipulated daily practice with minimum durations ranging from 10 to 45 min (median = 20 min). MBPs based around the MBSR or MBCT group-delivered format with required home practice dominate the current sample of studies, with many adopting content for their selected sample. However, use of the retreat day was limited and not all studies reported using trained facilitators to deliver the MBP.

Twenty-two studies reported adherence to the MBPs (and active control condition if applicable) delivered, and definitions of adherence were wide ranging: from attending at least half of the sessions (or completing half of the self-help content) to attending seven out of eight sessions. These studies reported adherence rates of 48–100%, median = 80%. Although home practice was incorporated into the majority of programmes used in the studies, measures of duration or frequency of home practice were not always reported. When detailed, self-report of home practice showed a wide range across the studies; 14 studies reported average duration (mean or median) of home practice ranging from 15 to 120 min weekly and frequency between 2 and 7 days per week. Only two studies (Armstrong and Rimes 2016; Potharst et al. 2019) statistically investigated the link between adherence and self-compassion, both finding a significant association between either amount of home practice or number of sessions attended with SCS scores.

We were also interested in the consistency of significant change in self-compassion and psychological distress (depression, anxiety, stress). Of the 19 studies that found a statistically significant post-programme increase in self-compassion, two studies found no significant change in depression or anxiety (stress was not measured in either of these studies) and one study found no significant change in stress (however, the same study found significant improvements in depression and anxiety). Of the six studies that found no significant difference in self-compassion pre- and post-programme, two found significant change in one of depression, anxiety or stress. Four studies (Benn et al. 2012; Gu et al. 2018; Shapiro et al. 2005; Ștefan et al. 2018) formally assessed self-compassion as a mediating factor in psychological distress (indexed as stress in all of the studies). Three of these studies found a significant mediating effect of self-compassion on stress; however, one study did not report such an effect. Only two of these studies examined change in self-compassion that temporally preceded change in stress by using either mid-programme self-compassion (Gu et al. 2018) or post-programme self-compassion with follow-up stress (Benn et al. 2012). Overall, the majority of studies found increases in both self-compassion and distress; however, this was not a consistent picture. When investigated, there was inconsistent support for self-compassion as a mediating factor in change for MBPs, with variation in how mediation was assessed.

Four studies (Arredondo et al. 2017; Hwang et al. 2019; O’Donnell 2017; Shapiro et al. 2005) did not report sufficient data for inclusion in the meta-analysis (pre- and post-programme mean, pre-programme standard deviation and sample size were all required); therefore, effect sizes from 22 studies were calculated. Three studies required calculation to combine MBP groups (Bayot et al. 2020; Mistretta et al. 2018) and control groups (Gu et al. 2018). The forest plot (Fig. 2) shows effect sizes (Hedges g for MBP vs control pre-post difference in self-compassion scores) for each study. The analysis revealed a medium effect size on self-compassion (g = 0.60, 95% CI = 0.41 to 0.80, p < 0.001). Tests for heterogeneity revealed a significant amount of medium to large heterogeneity (Q(21) = 65.97, p < 0.001, I² = 68%).

Given the significant amount of medium to large heterogeneity found, post-hoc assessment of potential moderators for self-compassion was conducted. Comparing “standard” MBPs (n = 17, g = 0.53, 95% CI 0.30 to 0.77) with MBPs incorporating additional compassion/kindness components (n = 6, g = 0.78, 95% CI 0.58 to 0.98) revealed no significant difference in self-compassion between the two groups (Q_between(1) = 2.46, p = 0.12). For this subgroup analysis only, the data from the standard mindfulness and compassion-added groups in Bayot et al.’s (2020) study were entered separately. Comparing studies that used a minimum threshold for psychological distress as a participation inclusion criterion (n = 6, g = 0.76, 95% CI 0.21 to 1.31) with studies that used no distress threshold (n = 16, g = 0.55, 95% CI 0.37 to 0.74), no significant difference was found (Q_between(1) = 0.50, p = 0.48). Comparing studies that used an active control condition (n = 5, g = 0.61, 95% CI 0.07 to 1.14) to those that used a wait-list control condition (n = 17, g = 0.61, 95% CI 0.41 to 0.79), no significant difference was found (Q_between(1) < 0.01, p = 0.97). For this subgroup comparison only, the data from Gu et al.’s (2018) study were separated into the active and the wait-list control groups.

Visual inspection of the funnel plot (Fig. 3) indicated possible asymmetry, with larger studies showing a tendency towards a lower effect size and a lack of smaller studies showing a lower effect. However, Orwin’s failsafe N revealed that a total of 110 studies with a null effect size would need to be located to reduce the mean effect size to below 0.1. It is unlikely that such a number of studies with a null result exist, suggesting that the mean effect size found would be robust to publication bias. Due to heterogeneity, trim and fill analysis was not performed.

Seven studies provided sufficient follow-up data for inclusion in the meta-analysis. Effect sizes are shown in the forest plot in Fig. 4. The analysis revealed a medium effect size (g = 0.53, 95% CI = 0.21 to 0.85, p = 0.001). Tests for heterogeneity revealed a significant amount of medium heterogeneity (Q(5) = 17.83, p = 0.007, I² = 66%). The small number of studies made interpretation of the funnel plot (Fig. 5) difficult; however, it appeared to have reasonable symmetry. Orwin’s failsafe N found a total of 30 studies with a null effect size would need to be located to reduce the mean effect size to below 0.1.

Six studies provided correlations between change in self-compassion and stress, two studies reported correlations for change in self-compassion and depression and one study for anxiety. Only the correlations with stress were entered into a meta-analysis. Correlation coefficient r ranged from − 0.61 to − 0.41. The effect size found was not significant (r = − 0.52, CI − 0.99 to 0.88, p = 0.57).

Ratings for each of the categories for all studies can be found in Table 2. Of the 26 studies included, four were rated as weak on none of the categories, five in one category, 12 in two categories, four in three categories and one in four categories. In terms of Selection Bias, the studies that were rated as weak were due to a low percentage of recruited participants taking up the study (n = 8), insufficient information or evidence of unrepresentative sampling (n = 3) or both (n = 1). For Study Design, only one study was rated as weak, due to inappropriate randomisation introducing possible bias (allocation to wait-list or programme was based on enrolment in the programme in either the autumn or the summer). Five studies did not measure potential confounders, and three did not control for the confounding factors identified, resulting in a weak rating for the Confounders category. All studies scored either moderate or strong for Blinding and Data Collection Methods, reflecting the use of sufficiently valid and reliable self-report measures. Two studies were rated as weak for Withdrawal and Dropouts, due to reporting 40% or more dropout. For Intervention Integrity, studies were rated weak due to lack of description of programme integrity (n = 4) or less than 60% adherence to the programme (n = 5). The rating of the category was complicated by the variation of criteria for programme adherence between the studies, where described the authors’ own criteria were used. Thirteen studies lacked sufficient power to detect possible effects and were therefore rated weak for the Analysis category. Of these, six studies were described as pilot studies.

In reviewing the methodological quality of the studies, a number of strengths and weaknesses were identified. Most studies relied on opportunity sampling, introducing potential bias such as recruiting participants who had a positive attitude towards MBPs. However, any MBP in a nonresearch setting would also be open to invitation, therefore likely reflecting uptake in the real world. Females were overrepresented in most of the samples, therefore making it difficult to generalise the findings to men. As the studies were RCTs, there were low numbers of confounding factors, reducing the risk of selection bias. Most of the studies used wait-list control groups. While these are particularly helpful when testing a programme in a new population, as was the case in a number of the included studies, it remains unclear whether increased self-compassion was a result of MBPs specifically or other therapeutic components (e.g. group membership, psychoeducation). Although exploratory, one counter to this could be the lack of significant difference in effect size when comparing studies with a wait-list or active control. In terms of blinding, researchers were often not blinded to the condition which potentially introduced investigator effects in procedures such as communication with participants; however, the use of self-report measures minimises the effect when compared with researcher-conducted interviews or observational ratings. Commonly, participants were blinded to the hypotheses of the study; however, the nature of the studies precluded being blinded to condition, leading to possible social desirability bias in self-report responses. Most of the studies reported reasonable adherence to the programmes, suggesting that the MBPs used were feasible and acceptable to the participants; however, few investigated any link between adherence and self-compassion. One particular weakness identified in the studies was a lack of power due to small sample sizes that may have missed potential effects and analyses that did not take into account dropouts, which may have exaggerated the effects of MBPs. However, a number of underpowered studies were pilot studies. Both of these issues may have contributed to the heterogeneity found in the study sample.

All the studies used the SCS, which is one of the only validated and most commonly used measure of self-compassion (Neff 2003b, however see Gilbert et al. 2017). There have been a number of issues identified with this measure; for example, Williams et al. (2014) re-assessed the factor structure with meditators, nonmeditators and those in remission from repeat episodes of depression and could not find evidence for an overarching single factor of self-compassion, suggesting that reporting the six factors is preferable. Another study has questioned the validity of combining the positive (Self-Kindness, Mindfulness, Common Humanity) and negative (Overidentification, Isolation, Self-Judgement) aspects of the SCS (Costa et al. 2016). Neff (2016) suggested that either the single (particularly in the case of SCS-SF) or the six-factor items should be reported; however, this does not account for the issue that a person could score high on both negative and positive elements and receive the same score (Gilbert et al. 2017). Further, the overlap of mindfulness and self-compassion in the SCS may complicate the findings, as noted by MacBeth and Gumley (2012). This leaves unclear whether an increase in self-compassion following MBPs is driven by an increase in the mindfulness subscale much more than the other factors. Considering these issues, it may be beneficial for all future studies to report the six subscales of the SCS, or for a more psychometrically robust measure of self-compassion to be developed (Williams et al. 2014; Gilbert et al. 2017).

This review contributes further to evidence regarding the influence of MBPs on self-compassion and was able to produce a quantitative synthesis that summarised effect sizes across studies. However, a relatively small sample size for the meta-analysis precluded further investigation of the heterogeneity of effects found through methods such as meta-regression. Further, all of the post hoc subgroup analyses consisted of unbalanced small groups and were consequently underpowered; therefore, clear conclusions regarding the lack of significant difference between potential moderating factors are not possible. Meta-analysis of self-compassion and distress associations was similarly affected by very low sample sizes. Considering the inclusion criteria, there was no exclusion based on sample size or pilot study status. This may have resulted in inclusion of studies with limited power to detect effects; however, a number of the pilot studies included were sufficiently powered. Broader inclusion criteria regarding format, content and length of programme may have enabled further investigation of more diverse programmes. For example, programmes that draw on mindfulness but may have another focus (e.g. Acceptance and Commitment Therapy, Compassionate Mind Training), or single-dose meditations may provide further information about how mindfulness influences self-compassion. Confining results to studies that also measured state depression, anxiety or stress may also have restricted the number of studies that measured self-compassion in different contexts (particularly those that targeted community samples and measured psychological processes rather than symptoms). The focus on RCTs facilitated statistical analysis of overall effect and reduced the risk of bias from noncontrolled study designs; however, this excluded a number of studies that investigated self-compassion and its change over time in MBPs. The findings from this review revealed a number of potentially beneficial avenues for future research. We propose the following specific recommendations:

Studies assessing self-compassion (and other potential mechanistic outcomes) would benefit from further investigation into the link between self-compassion and psychological wellbeing following MBPs. This could take the form of reporting correlations or partial correlations between self-compassion and measures of psychological distress/wellbeing, or preferably taking follow-up or mid-intervention measures and conducting mediation analyses. This would enable future meta-regressions of such results.

While the current review found that MBPs with additional compassion components did not reach statistical significance in improving self-compassion, the low number of studies preclude firm conclusions. Further studies that directly compare MBPs with or without explicit compassion material are needed.

More studies should determine which “ingredients” of MBPs influence difference outcomes. Given theoretical suggestions about how self-compassion comes about during MBP participation (Feldman and Kuyken 2011; Kabat-Zinn 1994, 1996), particular attention should be paid to the attitude of the facilitator, perhaps utilising independent raters to determine compassionate responses during group sessions. Also of interest would be to compare programmes that do or do not contain inquiry, for example, programmes that are facilitated as opposed to other formats (e.g. self-help book, Internet, mobile app). Content of meditations (e.g. attention versus attitude) could also be examined and compared with self-compassion and other outcomes in mind.

Other measures of self-compassion should be investigated. All of the studies in the current review used versions of the SCS; however, given the debate around the conceptualisation of self-compassion, it would be useful to explore more recent tools, such as the Compassion Engagement and Action Scale (Gilbert et al. 2017). Reporting subscales of the SCS would also enable researchers to potentially tease out results based on varying conceptualisations of self-compassion (e.g. positive and negative subscales).

Given the wide range of programme adherence and home practice reported across the studies in this review, future studies may benefit from assessing how this relates to self-compassion (and other outcomes). For example, is there a minimum amount of practice that influences outcomes and is any relationship between these factors linear? Further, a more consistent method of reporting programme adherence in MBPs would be advantageous: for example, a score that incorporates a proportion of participants attending a sufficient number of group sessions and amount of home practice. Alternatively, an adherence score could be calculated for individual participants and either covaried for in analyses or testing for associations between adherence and outcome.

Considering the minority of studies using self-help MBPs that measured self-compassion, further work assessing this method of delivery is recommended, particularly given the emerging evidence of its efficacy (Blanck et al. 2018; Cavanagh et al. 2014; Spijkerman et al. 2016) and accessibility of mindfulness via books, websites and apps (e.g. Shore et al. 2018). This would enable future moderator analyses to compare self-help MBPs with more traditional group formats.

Use of active control conditions is recommended. Given the inconclusive moderator analysis in the current review, it is of interest to determine which other interventions do or do not increase self-compassion, for example CBT or relaxation programmes. This is of particular interest given the theoretical links between mindfulness and compassion.

Studies with more rigorous analytic strategies are needed. This would involve conducting power analyses and adequately accounting for dropout when determining sample size, taking baseline measures into account when analysing post-programme change and use of ITT analysis.

Self-compassion has been measured in a number of studies that were predominantly investigating effects of MBPs on psychological distress in nonclinical populations. A meta-analysis suggested that MBPs significantly improve self-compassion. However, this was not a consistent effect across all studies and reasons for this heterogeneity require further investigation. Further work assessing self-compassion as a mediating factor of change in psychological distress following MBPs would benefit theory and practice of MBPs.