An Investigation of the Factor Structure of the Self-Compassion Scale

The importance of self-compassion has long been recognised in Buddhist and Eastern philosophical traditions but, only recently, has its importance as a research construct distinct from other psychological constructs such as mindfulness (Kuyken et al. 2010) or self-esteem (Neff and Vonk 2009) been acknowledged. This has led to considerable growth in research examining the role of self-compassion particularly in the aetiology of both physical and mental well-being (Barnard and Curry 2011). Although researchers such as Gilbert (2009) have suggested definitions of self-compassion, one of the most widely used definitions is that put forward by Neff (2003b) who conceptualised self-compassion as follows:

Being touched by and open to one’s own suffering, not avoiding or disconnecting from it, generating the desire to alleviate one’s suffering and to heal oneself with kindness. Self-compassion also involves offering non-judgmental understanding to one’s pain, inadequacies and failures, so that one’s experience is seen as part of the larger human experience. (Neff 2003b, p. 87)

Within this definition, Neff conceptualised self-compassion as being composed of the following three components:

(a) self-kindness – extending kindness and understanding to oneself in instances of perceived inadequacy or suffering rather than harsh judgment and self-criticism, (b) common humanity – seeing one’s experiences as part of the larger human experience rather than seeing them as separating and isolating, and (c) mindfulness – holding one’s painful thoughts and feelings in balanced awareness rather than over-identifying with them in an exaggerated manner (Neff and Lamb 2009, p. 864)

Evidence for a link between self-compassion and mental well-being is increasing (for review see Barnard and Curry 2011). What is more, enhancing self-compassion may also have physical health benefits (Hall et al. 2013). Self-compassion has been shown to be a more accurate predictor of overall well-being than self-esteem (Neff and Vonk 2009) and it accounted for additional variance in anxiety and depression beyond that explained by self-esteem (Gilbert 2009). Self-compassion may protect against emotional distress. In a recent meta-analysis, MacBeth and Gumley (2012) found an association between self-compassion and lower levels of depression, anxiety and stress. Although the majority of studies were cross-sectional, the findings suggested that greater self-compassion was associated with mental well-being and that self-compassion may be associated with a reduction in some forms of emotional distress.

The main assessment tool used was the Self-Compassion Scale (SCS; Neff 2003a). Concerns have been raised that by measuring ‘negative’ components of compassion; specifically that the SCS is measuring self-criticism, rumination and social isolation, rather than self-compassion (MacBeth and Gumley 2012; Muris 2015). In a more recent meta-analysis, Muris and Petrocchi (2017) found that as the total score includes the negative components, then it might lead to an overestimation of the relationship with symptoms of psychopathology as the negative components are more strongly associated with psychopathology (r = .47 to –.50) than the positive components (r = − .27 to − .34). Neff (2016), however, described self-compassion as requiring an interaction between the positive and negative components of compassion and, as a consequence, she developed the SCS to assess compassion as per her definition (Neff 2003b).

According to Neff (2003a, b, 2016), the SCS has a six-factor structure with three positive and three opposing negative components that are interconnected. Specifically, the SCS assesses the following: (1) self-kindness: a person’s acceptance of personal flaws and ability to self-soothe in times of distress versus (2) self-judgement: expressions of self-critical or judgemental beliefs; (3) common humanity: the recognition of personal shortcomings as something that everyone experiences versus (4) feelings of isolation: feeling alone in their faults and (5) mindfulness: maintaining a non-judgemental awareness of thoughts and emotions versus (6.) over-identification with thoughts: becoming overwhelmed and wrapped up in emotions or thoughts. A series of confirmatory factor analyses were then used to evaluate the model fit. These showed that a six-factor correlated model was an ‘adequate fit’ to the data (Non-normed fit index (NNFI) = 0.90; Comparative Fit Index (CFI) = 0.91; Neff 2003a, b) and a higher-order model (NNFI = 0.88, CFI = 0.90) was also proposed as a reasonable fit (Neff 2003a, b) and was initially used to support the use of the SCS to give a total self-compassion score.

Since its original publication, the factor structure of the SCS has received considerable attention: studies have yielded mixed findings with some authors reporting support for the six-factor correlated model (Azizi et al. 2013; Castilho et al. 2015; Garcia-Campayo et al. 2014; Lee and Lee 2010; Mantzios et al. 2013) whereas other studies have been unable to replicate this factor solution (López et al. 2015; Petrocchi et al. 2013; Williams et al. 2014). Support for the higher-order model has been more sparse with only a few studies reporting it a fit to their data (Castilho et al. 2015; Cunha et al. 2016; Dundas et al. 2016).

As a result, many authors have proposed alternative factor structures which have included a single-factor model (i.e. an overarching single self-compassion construct; Deniz et al. 2008) and a four factor model where the positive factors are correlated and there is a distinct general negative factor (Zeng et al. 2016). The most widely proposed model is a two-factor solution comprised of self-compassion (total of the positive items) and self-coldness (total of the negative items; Gilbert et al. 2011). This solution has also been found when the SCS has been administered in Dutch (López et al. 2015) and Portuguese (Costa et al. 2015). Indeed, the majority of independent studies into the SCS have been carried out cross-culturally with researchers translating the scale into Greek (Mantzios et al. 2013), Iranian (Azizi et al. 2013), and Spanish (Garcia-Campayo et al. 2014) and evaluating the model fit of the adapted scales. This has led to some problems with translating the scale. López et al. (2015), for example, had to omit two of the items (self-kindness subscale item 5, ‘I try to be loving towards myself when I’m feeling emotional pain’; self-judgement subscale item 21, ‘I can be a bit cold-hearted towards myself when I'm experiencing suffering’) as the items did not translate into Dutch. This is not an uncommon occurrence as items are worded to suit the culture they are developed in and translation can change the context and meaning of items (Auer et al. 2000; Behling and Law 2000). It is not surprising, therefore, that adapting the scale for use in other cultures may slightly alter what is being measured which could affect item/factor loadings.

The incongruity in the factor structures found by previous researchers may suggest that the factor structure of the SCS is not stable and would benefit from further robust analyses. Indeed, Neff (2016) suggested that the higher-order structure may not be the most appropriate conceptualisation of compassion. Furthermore, recent studies (e.g. Neff et al. 2017; Tóth-Király et al. 2016) have investigated the factor structure further via alternatives to higher-order models and instead added a bi-factorial component alongside the six factors in the SCS model. Bi-factorial modelling assesses covariance between factors that arises from the presence of an overarching factor (in this case, self-compassion), whilst allowing the individual factors to retain and account for variance in their own subset of items (Reise et al. 2010).

Neff et al. (2017) found evidence supporting the six-factor correlated model in both non-clinical and clinical populations. In the non-clinical populations, the bi-factorial model was a comparable fit to the six-factor solution; however, it did not improve the model fit. Consequently, the authors suggested that further research using bi-factorial modelling was warranted. Since Neff (2016) suggested a bi-factorial model might best fit the measurement of self-compassion, several studies have employed this analysis using translated versions of the SCS in French, Brazilian Portuguese and Hungarian (Kotsou and Leys 2016; de Souza and Hutz 2016; Tóth-Király et al. 2016). For example, Tóth-Király et al. (2016) investigated the six-factor correlated and bi-factorial models using the Hungarian version of the SCS. The researchers compared model fit of the six-factor correlated model and the bi-factorial model using confirmatory factor analysis (CFA) and exploratory structural equation modelling (ESEM; combination of exploratory and confirmatory factor analysis techniques) and found that when using the CFA, neither model was an adequate fit to their data, but when using ESEM, both models fitted the data with the bi-factorial model being the best fit to the data. Although the focus on translated versions of the scale is welcome, there have been no independent replications of the bi-factorial model using the English language version of the scale.

With this in mind, the present study aimed to independently investigate the factor structure of the English language version of the SCS using both exploratory and confirmatory factor analytic techniques. Confirmatory factor analysis was used to compare the fit of the emergent exploratory factor structure to the alternative models described in the extant literature including the six-factor correlated model and the higher-order model (Neff 2003a, b) and the bi-factorial model proposed by Neff et al. (2017). An exploratory factor analysis was employed to explore if there was an alternative model that was a better fit to our data.

Six hundred and ninety-eight people started the online survey. Those who did not complete the self-compassion measure (n = 162) were excluded from the main analyses. This yielded a sample of 526 adults who completed the SCS at time 1 (t1). Chi-square tests showed that there were no significant differences on demographic variables between those who completed the SCS at both time points and those who only completed the SCS at baseline. The t tests revealed no differences between the EFA (t1) and the CFA (t2) samples in age or in any of subscales of the SCS. The majority of participants reported no experience of meditation or mindfulness (t1 N = 391 (74%); t2 N = 262 (79%)); of those who reported engaging in meditative practices, only 20–23% of people reported practising at least every couple of months.

As shown in Table 1, all of the SCS subscales and total score were all significantly inter-correlated, as anticipated. The subscales were moderately to highly correlated. Common humanity showed the lowest associations with the three negative subscales (self-judgement r = − .33, perceived isolation r = − .39 and over-identification with thoughts r = − .38). The SCS total score was most strongly correlated with the self-kindness (r = .81) and self-judgement (r = − .82) subscales.

The SCS is a widely used measure of self-compassion and its factor structure has received a great deal of research interest. This study provided an independent evaluation of the SCS’s factor structure and replicated the models evaluated by Neff et al. (2017). Specifically, the outcomes of this study echo those found in Neff et al.’s (2017) study, in particular the results from her student sample. We found the SCS to be reasonably reliable with both the overall scale and subscales having relatively high internal reliability and good test-retest reliability. Of the models we investigated, we found that the bi-factorial model consisting of the six-factor correlated model and an overarching self-compassion factor was the best fit to our data. This supports Neff’s (2016) conceptualisation of self-compassion as having six distinct factors that are influenced by a concurrent (self-compassion) factor and the use of the SCS to give both an overall self-compassion score, or to use the scores from individual subscales. The inclusion of a general self-compassion factor accounted for some of the shared variance between factors and improved the model fit across all of our fit criteria (TLI = 0.94, CFI = 0.95) and the AIC suggested that this model was the best fitting of all the models. When we ran the same analyses on the t1 data, the bi-factorial model did not fulfil all of our fit criteria (TLI < 0.9), but remained the closest fit to our data. In Neff et al.’s (2017) recent paper, the bi-factorial model was not as good a fit as the six-factor solution in any of the populations, but it still demonstrated an acceptable fit in all of the populations with the exception of the clinical sample. Van Prooijen and Van Der Kloot (2001), however, emphasised that there is never ‘one single true model’ as data are subject to individual differences.

The omega indices showed further support for the bi-factorial construction of self-compassion as the subscales ranged from ω = .80 to .93 and the total score had an ω of .96 suggesting that the subscales and the overarching scale are representative of both self-compassion and the six factors. With the inclusion of the overarching self-compassion factor, the ωH indices reduced to between .05 (self-kindness) and .46 (isolation). Lower ωH scores indicate that a greater proportion of that factor’s variance has been explained by the overarching self-compassion factor rather than the individual factor(s). Self-kindness, for instance, appeared to be comprised largely of self-compassion as the variance reduced by 88% (t2) when the overarching factor was included. We also calculated the percentage variance in total scores (89% of t1 and 94% of t2) explained by the overarching self-compassion factor (Table 5). Our findings echo the percentages reported by Neff et al. (2017) who found that the general self-compassion factor accounted for 90–95% of variance across their samples. These omegas indicate that both the scores from the specific factors and from the total score can be interpreted with confidence.

The six-factor correlated model was also a good fit to our t2 data and the fit was comparable with previous research (Neff 2003a, b: TLI = 0.9, CFI = 0.91. Neff et al., 2017: student sample: TLI = 0.92, CFI = 93. The present study: TLI = 0.92 and CFI = 0.93). In our model, the items loaded well onto the proposed factors with loadings ranging from .55 to .85. These were comparable to those from the student sample in Neff et al.’s recent paper (Neff et al. 2017). Very similar factor loadings were found when the CFA was run on the t1 data. Internal consistency was mostly good within the subscales. However, we found that perceived isolation was highly correlated with the over-identification and self-judgement factors. Correlations of this level (0.94 and 0.90 respectively) can indicate poor discriminant validity between subscales, but in some, as is probable in this case, they can be indicative of a shared latent variable that impacts upon the scale over and above the impact of the factors (Gaskin 2016). The inconsistencies in models found by previous research might suggest the latter may be the case and findings from this study support this conjecture.

The five-factor model from our EFA was not supported during the confirmatory procedures from either of our time points; however, this is not an unusual outcome in cross-validation studies as no parameters are set in EFA and the data are allowed to inform the model formation whereas the CFA procedure is run with more restrictions in place (van Prooijen and van der Kloot 2001). Neither the single-factor nor two-factor models fitted our data. This might suggest that the operationalisation of self-compassion is more complicated than it being a single construct or a sum of the positive and negative items. To address concerns regarding the inclusion of the negative components of self-compassion, some research has adopted the two-factor model to measure self-coldness and self-compassion scores (e.g. Gilbert et al. 2011). Although the fit indices were approaching a fit to the data, this model did not reach acceptable levels of fit; therefore, we found no support for using the SCS in this way. We found similar outcomes for the higher-order model to Neff et al. (2017). In higher-order models, the overarching factor accounts for all variance between the factors that load on to it. It does not allow the factors to retain any individual influence on the model. This does not fit with Neff’s (2016) conceptualisation of self-compassion as consisting of both self-compassion and interrelated components.

The lack of fit for the two-factor model should also alleviate some concerns that the SCS may be affected by the item scoring method (e.g. López et al. 2015, Muris and Petrocchi 2017). The fact that the single compassion factor was not a fit to our data supports Neff et al.’s (2017) ascertainment that although the SCS can be used to yield a total score compassion, it is not constructed of a single dimension.

Self-compassion is an important psychological construct and it is imperative that we advance our understanding of how it is optimally operationalised. Our findings support the view that compassion is a multi-faceted construct that is more complicated than being comprised exclusively of the positive components of self-compassion. Muris and Petrocchi (2017), however, reported greater associations between the negative components and psychopathology than the positive components of the SCS. These authors highlight the importance that the scoring method can have on a scale in that the reverse scored items might serve to inflate the self-compassion score, thus increasing the association between the self-compassion total score and psychopathology. The bi-factorial construct of the SCS affords researchers the opportunity to explore the impact of the individual factors as well as the overall total score and address this concern. In the present study, our inter-factor correlations were stronger between negative components of self-compassion than those between the positive ones. More research needs to be conducted into the mechanisms underlying the components of self-compassion and to explore how much impact each factor has under various circumstances or populations.

The study employed a student sample which was three quarters female. Model fit should be tested across other populations to establish what models of self-compassion are most appropriate in different populations and studies should investigate how gender impacts upon model fit. In future studies, modelling techniques ought to be reflective of the complexity of self-compassion and, as such, assess the presence of a shared compassion factor by using bi-factorial and other in-depth structural equation modelling techniques. Factorial and structural modelling techniques are continuing to develop, and in a recent investigation, Tóth-Király et al. (2016) applied exploratory structural equation modelling techniques to a Hungarian version of the SCS. Applying these techniques to the original language version of the SCS would allow for more rigorous testing of this important construct.

Our study found that the SCS can be used to give subscale totals and to give an overall total compassion score. Despite the scale’s extensive use, however, we have little understanding of how the six components of the SCS interact with each other and more work needs to be done to understand if all the factors contribute equally to a person’s compassion or if one area is potentially more important than another. In this vein, Muris and Petrocchi (2017) also emphasised the need for studies to investigate the predictive value of the different components of the SCS in the aetiology of psychopathology. To facilitate this, research into the SCS needs to move away from cross-sectional studies and employ more prospective designs which would also allow the investigation of the stability of self-compassion over time. Studies could also be designed to determine how self-compassion is affected by the presence of stressful life events, particularly events that increase feelings of self-criticism and failure (e.g. Tóth-Király et al. 2016), and allow exploration of the relationship between these and the latent variables of the SCS.

More research emphasis needs to be placed on the positive components of mental health rather than the negative aspects, and as such, self-compassion is an important area that deserves much more research attention. Thus far, research into self-compassion has primarily focussed on its association with mental health problems such as depression, anxiety and stress; however, how self-compassion is related to more complex mental health problems including experiences such as paranoia and distressing voices, self-harm and suicide is worthwhile. Moreover, the role of self-compassion in recovery merits more attention (Anthony 1993; Leamy et al. 2011). Our research reiterates Neff et al.’s (2017) findings that the SCS can be used as the six-factor and bi-factorial model, thereby further emphasising the complexity of self-compassion. Our findings also support the use of the SCS to give a total score as suggested by Neff and colleagues (Neff 2003a, b, Neff et al. 2017). However, in light of Muris and Petrocchi (2017) recent meta-analysis, further examination of the contributions of the individual factors, particularly the negative factors, to the overall self-compassion score is vital. In sum, further research into this complex construct is needed to establish the impact of the individual components on the models of the SCS and how these components interact within mental health and illness.