Journal of Personality Assessment

ISSN: 0022-3891 (Print) 1532-7752 (Online) Journal homepage: http://www.tandfonline.com/loi/hjpa20

Examining the Factor Structure of the Self-

Compassion Scale in Four Distinct Populations: Is
the Use of a Total Scale Score Justified?

Kristin D. Neff, Tiffany A. Whittaker & Anke Karl

To cite this article: Kristin D. Neff, Tiffany A. Whittaker & Anke Karl (2017) Examining the
Factor Structure of the Self-Compassion Scale in Four Distinct Populations: Is the Use of
a Total Scale Score Justified?, Journal of Personality Assessment, 99:6, 596-607, DOI:
10.1080/00223891.2016.1269334

To link to this article: https://doi.org/10.1080/00223891.2016.1269334

Published online: 31 Jan 2017.

Submit your article to this journal

Article views: 691

View Crossmark data

Citing articles: 22 View citing articles

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=hjpa20
JOURNAL OF PERSONALITY ASSESSMENT
2017, VOL. 99, NO. 6, 596–607
https://doi.org/10.1080/00223891.2016.1269334

Examining the Factor Structure of the Self-Compassion Scale in Four Distinct

Populations: Is the Use of a Total Scale Score Justified?
Kristin D. Neff,1 Tiffany A. Whittaker,1 and Anke Karl2
1
Department of Educational Psychology, University of Texas at Austin; 2Department of Psychology, University of Exeter, UK

ABSTRACT ARTICLE HISTORY

This study examined the factor structure of the Self-Compassion Scale (SCS) using a bifactor model, a Received 1 March 2016
higher order model, a 6-factor correlated model, a 2-factor correlated model, and a 1-factor model in 4 Revised 10 November 2016
distinct populations: college undergraduates (N D 222), community adults (N D 1,394), individuals
practicing Buddhist meditation (N D 215), and a clinical sample of individuals with a history of recurrent
depression (N D 390). The 6-factor correlated model demonstrated the best fit across samples, whereas
the 1- and 2-factor models had poor fit. The higher order model also showed relatively poor fit across
samples, suggesting it is not representative of the relationship between subscale factors and a general
self-compassion factor. The bifactor model, however, had acceptable fit in the student, community, and
meditator samples. Although fit was suboptimal in the clinical sample, results suggested an overall self-
compassion factor could still be interpreted with some confidence. Moreover, estimates suggested a
general self-compassion factor accounted for at least 90% of the reliable variance in SCS scores across
samples, and item factor loadings and intercepts were equivalent across samples. Results suggest that a
total SCS score can be used as an overall mesure of self-compassion.

The construct of self-compassion is drawn from Buddhist psy- Self-compassion is equally relevant, however, when suffering
chology, and was first operationally defined and introduced stems from one’s own imprudent actions or personal failures.
into the psychological literature over a decade ago. Neff (2003b, As Neff (2016) wrote, the various elements of self-compassion
2016) proposes that self-compassion is a type of self-to-self are conceptually distinct and tap into different ways that individ-
relating that represents a compassionate rather than uncompas- uals emotionally respond to pain and failure (with kindness or
sionate stance toward the self when faced with personal suffer- judgment), cognitively understand their predicament (as part of
ing: self-kindness versus self-judgment, a sense of common the human experience or as isolating), and pay attention to suf-
humanity versus isolation, and mindfulness versus overidentifi- fering (in a mindful or overidentified manner). Although these
cation. These components combine and mutually interact to components are separable and do not covary in a lockstep man-
create a self-compassionate frame of mind. Self-kindness entails ner, they do mutually affect one another. Put another way,
being gentle, supportive, and understanding toward oneself. self-compassion can be seen as a dynamic system that represents
Rather than harshly judging oneself for personal shortcomings, a synergistic state of interaction between the various elements of
the self is offered warmth and unconditional acceptance. Com- self-compassion.
mon humanity involves recognizing the shared human experi- Over the past decade research on self-compassion has
ence, understanding that all humans fail and make mistakes, expanded exponentially. Self-compassion has been consistently
and that all people lead imperfect lives. Rather than feeling iso- related to psychological health in the research literature, includ-
lated by one’s imperfection—egocentrically feeling as if “I” am ing increased positive outcomes such as happiness and life sat-
the only one who has failed or am suffering—one takes a isfaction and decreased negative outcomes such as anxiety and
broader and more connected perspective with regard to per- depression (Barnard & Curry, 2011; MacBeth & Gumley, 2012;
sonal shortcomings and individual difficulties. Mindfulness Zessin, Dickhauser, & Garbade, 2015). Most of the research on
involves being aware of one’s present moment experience of self-compassion has been conducted using the Self-Compassion
suffering with clarity and balance, without running away with a Scale (SCS; Neff, 2003a), which assesses overall trait levels of
dramatic storyline about negative aspects of oneself or one’s life self-compassion. Items are written in a face-valid manner and
experience—a process that is termed overidentification. Self- measure the cognitive and emotional behaviors associated with
compassion can be directed toward the self when suffering more compassionate and fewer uncompassionate responses to
occurs through no fault of one’s own—when the external cir- feelings of personal inadequacy and general life difficulties.
cumstances of life are simply painful or difficult to bear. Sample items (Neff, 2003a) are “I try to be loving toward myself

CONTACT Kristin D. Neff [email protected] Department of Educational Psychology, University of Texas at Austin, 1912 Speedway, Stop D5800,
Austin, TX 78712–1289.

© 2017 Taylor & Francis

 EXAMINING THE FACTOR STRUCTURE OF THE SCS 597

when I’m feeling emotional pain” (self-kindness), “I’m disap- providing convergent validity for the SCS and thus support for
proving and judgmental about my own flaws and inadequacies” the construct validity of self-compassion (Neff, 2016). For
(self-judgment), “When things are going badly for me, I see the instance, self-compassion interventions have been found to
difficulties as part of life that everyone goes through” (common increase optimism, happiness, life satisfaction, self-efficacy, and
humanity), “When I think about my inadequacies it tends to body appreciation; to decrease rumination, depression, anxiety,
make me feel more separate and cut off from the rest of the stress, and body shame (Albertson, Neff, & Dill-Shackleford,
world” (isolation), “When I’m feeling down I try to approach 2014; Neff & Germer, 2013; Shapira & Mongrain, 2010; Smeets,
my feelings with curiosity and openness” (mindfulness), and Neff, Alberts, & Peters, 2014); and to positively affect physio-
“When something upsets me I get carried away with my logical responses to stress (Arch et al., 2014.) Similarly, experi-
feelings” (overidentification). Neff (2016) argued that the mental studies designed to induce a self-compassionate mood
mind-state of self-compassion represents the relative balance of (i.e., responding to writing prompts that foster self-kindness,
compassionate and uncompassionate responses to suffering, common humanity, and mindfulness) have been shown to
which is why the SCS measures the lack of uncompassionate increase positive affect and motivation and also decrease nega-
responses as well as the presence of compassionate ones. tive emotions such as anxiety, shame, and depression (Breines
The SCS was developed in a sample of college undergradu- & Chen, 2012; Diedrich, Grant, Hofmann, Hiller, & Berking,
ates (Neff, 2003a). After identifying items that made sense to 2014; Johnson & O’Brien, 2013; Leary, Tate, Adams, Allen, &
students, Neff used exploratory factor analyses (EFAs) to iden- Hancock, 2007; Odou & Brinker, 2014).
tify items that loaded best on separate subscales representing SCS total scores also demonstrate good discriminant valid-
the six components of self-compassion. Confirmatory factor ity. Although a key feature of self-compassion is a lack of self-
analyses (CFAs) were used to provide support that scale items judgment, and SCS scores are moderately correlated with
fit as intended with the proposed a priori theoretical model self-criticism (Blatt, D’Afflitti, & Quinlan, 1976), total SCS
(Furr & Bacharach, 2008). An initial CFA found an adequate fit scores still negatively predict anxiety and depression when con-
to a six-factor intercorrelated model (nonnormed fit index trolling for self-criticism and negative affect (Neff, 2003a;
[NNFI] D .90, comparative fit index [CFI] D .91). A second Neff, Kirkpatrick, & Rude, 2007). In addition, SCS scores dem-
CFA found a marginal fit to a higher order factor model onstrate known groups validity: Individuals who practice Bud-
(NNFI D .88, CFI D .90) that accounted for the intercorrela- dhist meditation have higher total SCS scores, as would be
tions between subscales. The factor structure of the scale was expected given the Buddhist origins of the construct
cross-validated in a second student sample, and adequate fit (Neff, 2003a; Neff & Pommier, 2013). Scale scores demonstrate
was found for the six-factor intercorrelated model (NNFI D excellent convergent validity in terms of consistency with rat-
.92, CFI D .93) as well as the higher order factor model ings by observers. For instance, therapists were able to signifi-
(NNFI D .90, CFI D .92). Findings were interpreted as evidence cantly predict individuals’ SCS scores after a brief interaction
that the subscales could be examined separately or else that a (Neff, Kirkpatrick, & Rude, 2007), and there was a strong asso-
total score could be used to represent overall self-compassion ciation between self-reported and partner-reported scores on
levels, according the interests of the researcher. the SCS (multilevel modeling estimated the association to be
In the original publication (Neff, 2003a), total SCS scores .70) among couples in long-term romantic relationships
evidenced good internal reliability (Cronbach’s a D .92), as did (Neff & Beretvas, 2013). Similarly, high levels of agreement
the six subscales (with Cronbach’s a ranging from .75–.81). (intercorrelation coefficient [ICC] D .77) were found between
Test–retest reliability over a 3-week interval was also good for independent coders using SCS items to rate the level of self-
the total score (Cronbach’s a D .93) and six subscale scores compassion displayed in brief verbal dialogues (Sbarra, Smith,
(with Cronbach’s a ranging from .80–.88). Moreover, the inter- & Mehl, 2012). These findings suggest that the SCS measures
nal reliability of SCS scores has been found to be high across a behaviors that are clearly observable by others.
wide variety of populations (e.g., Allen, Goldwasser, & Leary, Because the SCS is the main measure available to assess self-
2012; Neff & Pommier, 2013; Werner et al., 2012). compassion, it is important to determine whether its factor
There is ample evidence for the idea that interpretation of structure is valid and replicable across various populations.
scores on the SCS has construct validity. First, the large body of Given that the SCS is designed to assess self-compassion as an
research indicating that higher total scores on the SCS are asso- overall construct, moreover, it is especially important to deter-
ciated with well-being is a source of predictive validity. For mine whether use of a total scale score is psychometrically
instance, higher total scores on the SCS have been associated justified. As Neff (2016) pointed out, most studies examining
with greater levels of happiness, optimism, life satisfaction, the factor structure of the SCS have been conducted in the
body appreciation, perceived competence, and motivation course of validating translations of the scale. The large majority
(Hollis-Walker & Colosimo, 2011; Neff, Hsieh, & Dejitthirat, of translations have replicated the six-factor structure of the
2005; Neff, Pisitsungkagarn, & Hsieh, 2008; Neff, Rude, & Kirk- SCS (e.g., Castilho, Pinto-Gouveia, & Duarte, 2015; Chen, Yan,
patrick, 2007), as well as lower levels of depression, anxiety, & Zhou, 2011; Garcia-Campayo et al., 2014; Hupfield &
stress, rumination, body shame, and fear of failure (Daye, Ruffieux, 2011; Petrocchi, Ottaviani, & Couyoumdjian, 2013).
Webb, & Jafari, 2014; Finlay-Jones, Rees, & Kane, 2015; Neff Although not all examined the higher order model, those that
et al., 2005; Raes, 2010). Higher scores on the SCS are also pre- did yielded inconsistent findings. For example, a higher order
dictive of healthier physiological responses to stress (Breines factor was found with a Chinese student and Portuguese clini-
et al., 2015; Breines et al., 2014). Moreover, these findings have cal and community samples (Castilho et al., 2015; Chen et al.,
been duplicated in research using non-self-report methods, 2011), but not with German and Italian student and
598 NEFF, WHITTAKER, KARL

community samples (Hupfield & Ruffieux, 2011; Petrocchi & Van der Veld, 2009), both of which are counterintuitive find-
et al., 2013), a Dutch community sample (L
opez et al., 2015), or ings. This is echoed in applied research settings in which tradi-
a second Portuguese clinical sample (Costa, Mar^oco, tional cutoff values do not tend to support complex, multifactor
Pinto-Gouveia, Ferreira, & Castilho, 2015). (e.g., five or more factors with five or more indicators per fac-
In the course of examining these translations, L
opez et al., tor) structures associated with good scales (see, e.g., Marsh,
(2015) conducted EFA and found that all the positive items Hau, & Grayson, 2005). This suggests that the cutoff values
loaded on one factor and all the negative items loaded on a sec- associated with model fit indexes in structural equation model-
ond factor, and argued that the SCS was bidimensional and ing are dependent on the model under investigation and its
that a total scale score should not be used (no CFA was con- properties. As such, the traditional model fit cutoff values are
ducted to support this factor structure). Of course, findings recommended as broad guidelines and not stringent bench-
could have been due to a method effect, given that EFA is marks (Marsh, Hau, & Wen, 2004; West, Taylor, & Wu, 2012).
highly sensitive to positive versus negative affect (DiStefano & Other markers of adequate model fit, particularly the effect
Motl, 2006). Costa et al. (2015) used CFA to compare a six- sizes associated with the relevant parameter estimates, should
factor uncorrelated model, a higher order model, a two-factor also be examined. Given the complex factors involved in deter-
uncorrelated model that separated positive and negative items, mining the quality of a psychometric measure—including the
and a two-factor model that included correlated errors soundness of its theoretical underpinnings and its validity and
designed to improve model fit, and found that the two-factor utility for facilitating research—it is useful to have a more sub-
model with correlated errors had the best fit. Note that the stantive estimate of the ability of a self-report measure to assess
researchers’ choice to examine an uncorrelated six-factor model its intended target than model fit alone.
was inconsistent with Neff’s (2003a) original approach, in One model that might support the interpretation of an over-
which the six factors are theoretically presumed to correlate. As arching self-compassion factor in addition to six separate sub-
Neff (2016) cautioned, moreover, care must be used before scale factors is a bifactor model (Reise, Bonifay, & Haviland,
assuming that findings obtained with translations generalize to 2013). In a bifactor model, individual scale items load on a gen-
the original language version of a scale given potential issues eral or “target” factor as well as a subscale or “group” factor.
with the quality of translations or cultural factors that could Considering how this model applies to the SCS, the target factor
affect findings (Behling & Law, 2000). is the general self-compassion factor and the group factors con-
Surprisingly few studies have attempted to replicate the fac- sist of the six subscale factors. As Reise, Moore, and Haviland
tor structure of the original English SCS. An important excep- (2010) discussed, in a bifactor model the general factor is mod-
tion is a study by Williams, Dalgleish, Karl, and Kuyken (2014) eled as directly influencing individual item responses, and the
that examined the factor structure of the SCS in a community ways in which individual items form group factors are also
(N D 821), meditator (N D 211), and clinical sample of individ- modeled. Moreover, none of the factors (target or group) are
uals with recurrent depression (N D 390) living in the United allowed to correlate in a bifactor model. This is because the sub-
Kingdom. Williams et al. (2014) reported that CFAs were used scale factors are posited to account for the shared variance in
in each sample to examine SCS item fit to a one-factor model, a their respective set of item responses over and above the vari-
six-factor correlated model, and a higher order model. The ance accounted for by the target (e.g., self-compassion) factor.
authors concluded that the one-factor and higher order models In contrast, a higher order model posits that although the target
did not fit the data acceptably. The six-factor correlated model factor explains the correlation among group factors, there is no
fit the data more favorably than the remaining models in all direct effect of the target factor on individual items—the link
populations examined, and demonstrated an adequate fit for between the target and items is modeled as being only indirectly
the community sample. The authors concluded that “the SCS related through the group factors. See Figure 1 for examples of a
may be better suited to measuring six hypothesized facets of bifactor model and higher order model as they apply to the SCS.
self-compassion … than for measuring an overarching con- It should be noted that the theoretical assumption that the tar-
struct (i.e., self-compassion)” and that “further research is get factor has no direct influence on individual item responses
needed to develop a more psychometrically robust measure of is a strong one. In contrast, “the bifactor model specifies that
self-compassion” (p. 10). there is a single (general) trait explaining some proportion of
Because of the good reliability and strong support for the common item variance for all items, but that there also are
predictive, convergent, and discriminant validity of the SCS, group traits explaining additional common variance for item
and because the measure has been used in such a large number subsets. The general and group factors are on equal conceptual
of empirical studies—the vast majority of which have used a footing and compete for explaining item variance—neither is
total scale score (Neff, 2016)—more investigation is warranted ‘higher’ or ‘lower’ than the other” (Reise et al., 2010, p. 547).
before concluding that the SCS should be redesigned. Some of Although the bifactor model was first developed in the early
the limitations of relying on model fit as a means of evaluating 20th century (Holzinger & Swineford, 1937), it was not well
the adequacy of a factor structure representing the items on a known or commonly used in the psychometric literature (at least
scale are that it depends on sample size and other model prop- in the United States) when the SCS was first developed (Reise
erties. For instance, simulation studies in this area have found et al., 2013). Neff (2016) argued that the bifactor model is a more
that model fit indexes suggest poor model fit of the correct accurate way to represent her original conceptualization of self-
model with an increase in the number of indicators per factor compassion (Neff, 2003a), given that self-compassion is theo-
(Marsh, Hau, Balla, & Grayson, 1998) and that model fit rized to directly manifest in the particular ways that individuals
decreases as the size of factor loadings increases (Saris, Satorra, respond to suffering (as represented by SCS scale items).
 EXAMINING THE FACTOR STRUCTURE OF THE SCS 599

Figure 1. Comparison of a bifactor model (top) and higher order model (bottom). Note: SC D self-compassion; SK D self-kindness; SJ D self-judgment; CH D common
humanity; IS D isolation; MI D mindfulness; OI D overidentification.

One benefit of using a bifactor model is that it allows for the SCS is bidimensional. We examined the properties of the SCS
straightforward calculation of the percentage of total score vari- in a sample of undergraduates (students), community adults
ance that is accounted for by the general target factor, each (community), and individuals practicing Buddhist meditation
group factor, and error. Although Williams et al. (2014) did (meditators). Because we did not have direct access to a clinical
not find support for the use of a total SCS score based on tradi- sample, we obtained permission to reanalyze data from the
tional model fit criteria, it might still be defensibly used. For sample of individuals with a history of recurrent depression
instance, if the large majority of observed variance in SCS previously examined by Williams et al. (2014; clinical).
scores is explained by the general target factor, this would pro-
vide some sense of confidence that the scale could be used to
measure the general factor even in the presence of multidimen- Methods
sionality. In contrast, if the majority of variance in scores is not Participants
attributable to the general target factor, this would argue
against the use of a total score. Reise et al. (2013) suggested .75 Students
or higher as an ideal amount of variance explained by a general This sample included a total of 222 undergraduates (84 male
factor to confidently use a total scale score. and 138 female, M age D 20.94, SD D 2.03) who were randomly
This study examined the psychometric properties of the selected from an educational-psychology subject pool at a large
original English version of the SCS using a bifactor, higher Southwestern university. In terms of ethnicity, 57% of the sam-
order, and six-factor correlated model in various populations, ple self-identified as White, 22% as Asian, 14% as Hispanic, 3%
with the main goal of determining whether or not the use of an as Black, and 4% as other.
overall SCS score (in addition to the six subscale scores) is justi-
fied. We also examined the fit of a one-factor model (to be con- Community
sistent with Williams et al., 2014), and a two-factor correlated The sample of community adults was recruited from Mechani-
model that separated positive and negative items to test the cal Turk, an online survey research recruitment method that
claims of L
opez et al. (2015) and Costa et al. (2015) that the samples from the general public. Mechanical Turk has been
600 NEFF, WHITTAKER, KARL

found to be much more nationally representative of the general Specifically, CFAs were conducted to examine a one-factor,
population than college samples (Buhrmester, Kwang, & two-factor correlated, six-factor correlated, higher order, and a
Gosling, 2011). Participants were directed to Survey Monkey to bifactor model. Mplus software (version 7.4; Muth
en &
take the study, and were paid 75 cents for completing it (see Muth
en, 1998–2016) was used when conducting all of the
Buhrmester et al., 2011, for supporting evidence of validity at CFAs. Because item responses on the SCS are ordinal in nature,
low payment levels). The sample had 1,394 participants (35% ranging from 1 (almost never) to 5 (almost always), maximum
male and 65% female; M age D 36.01, SD D 12.88). The ethnic likelihood robust (MLR) estimation in Mplus was used to esti-
breakdown was 77% White, 7% Black, 6% Asian, 6% Hispanic, mate model parameters.1
and 6% other. Model fit was evaluated globally using the comparative fit
index (CFI), the Tucker–Lewis Index (TLI; also known as the
Meditators NNFI), the root mean square error of approximation (RMSEA)
This sample included a total of 215 meditators (30% male and with accompanying 90% confidence interval (CI), and the stan-
70% female; M age D 47.40, SD D 11.59). The ethnic break- dardized root mean square residual (SRMR). Because the goal
down was 87% White, 2% Asian, 2% Hispanic, 2% Black, and of the study was to determine if use of a total SCS score is justi-
7% other. Participants were recruited via an e-mail that invited fied, and whether the SCS is a “good enough” measure of self-
them to complete an online questionnaire via Survey Monkey. compassion or if it needs to be redesigned, we evaluated fit
E-mails were sent to individuals affiliated with Seattle Insight indexes using the liberal criteria used in the study conducted by
Meditation Society, Spirit Rock, the Insight Meditation Society, Williams et al. (2014). In other words, we felt the conclusion
and other similar groups. Participants reported a wide range in that a total score should not be used and that the SCS should
meditation experience from beginner to advanced (1–20 years be redesigned should only be drawn if model fit failed to meet
of meditation practice). The average length of meditation prac- these more liberal criteria. Thus, models associated with CFI
tice for the sample was 6.67 years (SD D 3.86). Fifty-three per- and TLI values greater than or equal to .90 were deemed as
cent of the participants identified as Buddhist, 26% identified acceptably fitting models (Bentler & Bonnett, 1980). Models
as having no religious affiliation, 12% identified as Christian, associated with RMSEA values equal to or less than .10 would
6% as other, and 3% as Jewish. indicate acceptable model fit using the liberal criteria (Browne
& Cudeck, 1993; MacCallum, Browne, & Sugawara, 1996). The
Clinical RMSEA value is associated with a 90% CI that provides further
This is the same sample that was analyzed in the Williams et al. evidence of acceptable model fit if the upper limit does not
(2014) study. It initially included 405 participants (23% male exceed a value of .10 (West et al., 2012). SRMR values equal to
and 77% female; M age D 50.16, SD D 11.08). Participants or less than .10 would also indicate adequate model fit using
were recruited through primary care settings in the United the liberal criteria (Hu & Bentler, 1999). Note that chi-square
Kingdom. Criteria for this group included having a diagnosis of test statistics were not used because they tend to be highly sen-
recurrent major depressive disorder in full or partial remission sitive to sample size (Marsh, Balla, & McDonald, 1988), distri-
according to the Diagnostic and Statistical Manual of Mental butional assumption violations of the data (Bentler & Bonnett,
Disorders (4th ed. [DSM–IV]; American Psychiatric Associa- 1990), and the type of model misspecification present in the
tion, 1994), having three or more previous major depressive data (Gallini & Mandeville, 1984). Akaike’s (1973) information
episodes, and being 18 or older. For a full description of the criterion (AIC) for each model was also documented. This cri-
sample, see Williams et al. (2014). Note that as in the original terion is comparative, meaning that models associated with the
study, participants with any missing data on the SCS were smallest AIC values are selected as models that would most
excluded, leaving a remaining sample of N D 390. likely cross-validate in subsequent samples. Because a unique
feature of this study involved examining the bifactor model in
multiple samples, multiple-group techniques were also used to
Measures
examine the equivalence of the bifactor model’s parameter esti-
Self-compassion mates across the four populations.
Participants in all four samples completed the SCS (Neff, 2003a) One of the advantages of a bifactor model is that it allows for
described earlier. This 26-item self-report measure includes 5 the calculation of different indexes that represent the percent-
self-kindness items, 5 self-judgment items, 4 common human- age of variance in scores attributable to all of the factors as well
ity items, 4 isolation items, 4 mindfulness items, and 4 over- as the percentage of variance in scores attributable to the gen-
identification items. Responses are given on a 5-point Likert eral factor only. The omega index, which is a ratio of true score
scale ranging from 1 (almost never) to 5 (almost always). Nega- variance to total variance and corresponds to internal consis-
tive items are reverse-coded so that higher scores represent a tency reliability (Hancock & Mueller, 2001), represents the per-
lower frequency of these responses. For scoring purposes, centage of variance in the total scores accounted for by all of
means are calculated for each subscale, and a grand mean is cal-
culated that represents an overall self-compassion score. 1
Normal theory maximum likelihood (ML) estimation has been shown to produce
accurate parameter estimates for CFAs with ordered variables having five or
more categories (Beauducel & Herzberg, 2006; Rhemtulla, Brosseau-Liard, &
Psychometric analyses Savalei, 2012). Nonetheless, MLR estimation was used to adequately correct for
underestimated standard errors and inaccurate test statistics that tend to occur
In this study, five different models were used to examine the with ordered categorical variables when using ML estimation (Rhemtulla et al.,
factor structure of the SCS in each of the four samples. 2012).
 EXAMINING THE FACTOR STRUCTURE OF THE SCS 601

Table 1. Descriptive statistics for the Self-Compassion Scale (SCS) total and subscales across samples.

Student (N D 222) Community (N D 1,394) Meditator (N D 215) Clinical (N D 390)

M SD A M SD a M SD a M SD a

SCS 3.11 .67 .94 3.00 .76 .94 3.66 .61 .95 2.56 .62 .91
SK 3.07 .77 .86 2.92 .88 .86 3.61 .56 .82 2.50 .82 .80
SJ 3.00 .81 .85 3.11 .96 .85 2.64 .78 .89 3.64 .78 .78
CH 3.20 .80 .81 3.09 .92 .81 3.83 .79 .82 2.90 .95 .80
IS 2.87 .84 .77 3.16 1.01 .77 2.37 .82 .83 3.67 .83 .75
MI 3.29 .78 .80 3.23 .86 .80 3.95 .68 .83 2.94 .81 .74
OI 3.05 .90 .80 3.01 1.01 .83 2.49 .75 .82 3.69 .79 .70

Note. SCS D Total SCS score; SK D Self-Kindness subscale; SJ D Self-Judgment subscale; CH D Common Humanity subscale; IS D Isolation subscale; MI D Mindfulness
subscale; OI D Overidentification subscale. Note that the SJ, IS, and OI subscales were reverse-coded before calculating the total SCS score.

the factors. Omega hierarchical (omegaH; McDonald, 1999) is 2006). The internal consistency estimates associated with the
an index used to estimate the percentage of variance in the total subscales and the overall self-compassion scale ranged from .70
scores that is attributed to the general or target factor (Hancock to .95 with the majority falling above .80. Thus, the SCS total
& Mueller, 2001; Reise et al., 2013). and subscale scores demonstrated acceptable internal reliability
For each of the four samples examined, standardized esti- estimates of Cronbach alpha.
mates from the bifactor model were used to calculate both Table 2 shows the model fit indexes associated with the five
omega and omegaH indexes to determine the reliability associ- models analyzed for each sample. (It should be noted that the
ated with the SCS as well as the amount of variance on the SCS fit indexes obtained for the clinical sample in this study differed
total scores that could be attributed to an overall self-compas- slightly from those reported in the original Williams et al.
sion factor, respectively. To determine the amount of reliable (2014) study because slightly different model estimation proce-
variance (i.e., not due to error) in the SCS scores attributed to dures were used. However, results did not differ in any substan-
the general
“self-compassion” factor, omegaH is divided by tive way.) As seen in Table 2, the bifactor model demonstrated
omega vvH . adequate fit using the liberal criteria based on the CFI, RMSEA,
and SRMR indexes in the student, community, and meditator
samples. The TLI supported adequate model fit in the medita-
Results
tor sample, marginal fit in the student and community samples,
Descriptive statistics, including means, standard deviations, and inadequate fit in the clinical sample. For the higher order
and internal consistency reliability alphas, are presented in model, the RMSEA and SRMR indexes demonstrated adequate
Table 1. Internal consistency coefficients are generally deemed fit across all samples, whereas the TLI suggested inadequate
acceptable when they are .80 or above (Lance, Butts, & Michels, model fit across all samples. The CFI indexes were marginal for
the student, community, and meditator samples and inade-
quate for the clinical sample. The six-factor correlated model
Table 2. Model fit across samples. demonstrated adequate model fit based on the CFI, TLI,
Model CFI TLI RMSEA [90% CI] SRMR AIC RMSEA, and SRMR indexes in the student, community, and
meditator samples. In the clinical sample, the RMSEA and
Student sample (N D 222) SRMR indexes supported adequate model fit, whereas the CFI
One-factor .79 .77 .09 [.08, .10] .08 14438.04
Two-factor correlated .88 .87 .07 [.06, .08] .06 14191.63 and TLI indexes suggested inadequate fit of the six-factor corre-
Six-factor correlated .93 .92 .05 [.05, .06] .05 14047.81 lated model. The two-factor correlated and single-factor models
Higher order .89 .88 .07 [.06, .07] .07 14153.43 generally had poor fit across the four samples according to all
Bifactor .91 .89 .06 [.05, .07] .06 14098.65
Community sample (N D 1,394) fit criteria. The AIC suggested that, in all four samples, the
One-factor .74 .72 .10 [.09, .10] .08 100063.94 six-factor correlated model would likely cross-validate the best,
Two-factor correlated .88 .87 .07 [.06, .07] .05 97446.58 followed by the bifactor model, followed by the higher order
Six-factor correlated .94 .93 .05 [.04, .05] .04 96229.10
Higher order .89 .88 .06 [.06, .07] .08 97214.10 factor model, followed by the two-factor intercorrelated model,
Bifactor .91 .89 .06 [.06, .06] .07 96823.54 and finally the one-factor model.
Meditator sample (N D 215) Standardized factor loadings obtained with the six-factor
One-factor .74 .72 .11 [.10, .11] .09 12651.00
Two-factor correlated .87 .86 .08 [.07, .09] .06 12258.42 correlated model of individual SCS items on their intended sub-
Six-factor correlated .93 .92 .06 [.05, .07] .06 12071.63 scales are presented in Table 3, along with standardized factor
Higher order .89 .88 .07 [.06, .08] .08 12190.95 loadings obtained with the bifactor model of individual SCS
Bifactor .91 .90 .07 [.06, .07] .07 12115.61
Clinical sample (N D 390) items on the general self-compassion factor. The loadings on
One-factor .64 .61 .11 [.11, .12] .10 28263.28 the subscales in the six-factor correlated model were all statisti-
Two-factor correlated .83 .82 .08 [.07, .08] .07 27495.59 cally significantly different from zero2 (at p < .001) and ranged
Six-factor correlated .88 .87 .07 [.06, .07] .06 27307.50
Higher order .80 .78 .09 [.08, .09] .10 27639.30 in magnitude from .38 to .89. The loadings on the general
Bifactor .84 .81 .08 [.07, .08] .09 27474.55

Note. CFI D comparative fit index; TLI D Tucker–Lewis Index; RMSEA D root mean 2
Statistical significance was assessed using z scores associated with each of the
square error of approximation; CI D confidence interval; SRMR D standardized loadings. The z scores are calculated by dividing the loading value by its respec-
root mean square residual; AIC D Akaike’s information criterion. tive standard error.
602 NEFF, WHITTAKER, KARL

Table 3. Factor loadings of Self-Compassion Scale (SCS) items on their intended subscale as found in the six-factor correlated model (SF) and on the general self-compas-
sion factor as found in the bifactor model (GF) across samples.
Student N D 222 Community N D 1,394 Meditators N D 215 Clinical N D 390

SF GF SF GF SF GF SF GF

Self-kindness items
5. I try to be loving towards myself … .71 .62 .75 .56 .76 .51 .75 .45
12. When I’m going through a very hard … .76 .66 .80 .58 .77 .59 .81 .52
19. I’m kind to myself when I’m … .75 .67 .80 .60 .84 .61 .75 .53
23. I’m tolerant of my own flaws … .74 .75 .38 .33 .40 .43 .50 .56
26. I try to be understanding and patient … .75 .70 .76 .63 .79 .60 .56 .46
Self-Judgment items
1. I’m disapproving and judgmental about … .73 .64 .76 .71 .82 .75 .68 .61
8. When times are really difficult, I tend to … .70 .62 .76 .69 .78 .69 .63 .51
11. I’m intolerant and impatient towards … .65 .61 .69 .62 .75 .68 .62 .49
16. When I see aspects of myself that I … .80 .75 .81 .76 .87 .79 .75 .64
21. I can be a bit cold-hearted towards … .74 .68 .72 .66 .71 .60 .54 .48
Common Humanity items
3. When things are going badly for me, I … .57 .38 .60 .38 .66 .45 .59 .42
7. When I’m down and out, I remind … .75 .46 .75 .43 .73 .40 .66 .34
10. When I feel inadequate in some way, I … .80 .51 .79 .42 .80 .51 .75 .47
15. I try to see my failings as part of the … .75 .55 .69 .49 .73 .50 .80 .51
Isolation items
4. When I think about my inadequacies, it … .66 .60 .76 .72 .65 .70 .70 .63
13. When I’m feeling down, I tend to feel … .65 .56 .75 .65 .79 .59 .62 .51
18. When I’m really struggling, I tend to … .63 .54 .72 .60 .81 .63 .56 .44
25. When I fail at something that’s … .75 .72 .77 .72 .75 .68 .70 .66
Mindfulness items
9. When something upsets me I try to keep … .68 .57 .66 .49 .59 .32 .52 .25
14. When something painful happens I try … .75 .60 .78 .58 .84 .52 .76 .47
17. When I fail at something important … .79 .73 .78 .61 .85 .51 .68 .54
22. When I’m feeling down I try to … .61 .52 .65 .51 .70 .55 .62 .48
Overidentified items
2. When I’m feeling down I tend to obsess … .79 .70 .80 .77 .80 .75 .66 .64
6. When I fail at something important to … .72 .71 .78 .78 .77 .71 .69 .66
20. When something upsets me I get … .65 .55 .71 .64 .65 .50 .50 .44
24. When something painful happens … .66 .59 .68 .62 .69 .53 .52 .50

Note. All reported factor loadings were significant at p < .001.

factor in the bifactor model were all statistically significant to the data from the two comparison samples without imposing
(at p < .001) and ranged from .25 to .79. Tabachnick and Fidell factor loading or item intercept equality constraints across the
(2007) suggested that factor loadings of .32 or above are consid- samples. Then, the bifactor model in which all of the factor
ered meaningful, and Comrey and Lee (1992) suggested load- loadings were constrained to be equal across both samples was
ings of .55 and above are good because it means that estimated and compared to the baseline model. The bifactor
approximately 30% of the variance in the item is explained by model with item intercepts constrained across both samples
the factor. Using these criteria, the vast majority of items had was subsequently estimated and compared to the bifactor
good factor loadings on their respective subscale factors in the
six-factor correlated model, and all loadings were meaningful. Table 4. Correlations between the Self-Compassion Scale (SCS) subscale factors in
Similarly, the large majority of items had good factor loadings the six-factor correlated model for the student sample (N D 222), community sam-
on the general factor for the student, community, and medita- ple (N D 1,394), meditator sample (N D 215), and clinical sample (N D 390).
tor samples in the bifactor model, although fewer had good Student Community Meditator Clinical
loadings for the clinical sample. In all samples, however, all
SK-SJ ¡.82 ¡.72 ¡.77 ¡.56
items except one (mindfulness Item 9 in the clinical sample) SK-CH .77 .71 .72 .81
had loadings on the general self-compassion factor in the bifac- SK-IS ¡.78 ¡.63 ¡.61 ¡.49
tor model that would be considered meaningful. SK-MI .88 .83 .85 .76
SK-OI ¡.75 ¡.63 ¡.70 ¡.50
Table 4 presents the factor intercorrelations for the six-fac- SJ-CH ¡.48 ¡.44 ¡.53 ¡.34
tor correlated model across the four samples, which ranged SJ-IS .86 .86 .83 .81
from .34 to .97 and were all statistically significant (at p < SJ-MI ¡.68 ¡.60 ¡.64 ¡.39
SJ-OI .92 .91 .91 .91
.001). Multiple-group analysis was then conducted to assess CH-IS ¡.50 ¡.48 ¡.54 ¡.46
strong or scalar measurement invariance. Specifically, we tested CH-MI .80 .79 .77 .87
the equivalence of the factor loadings and the item intercepts CH-OI ¡.48 ¡.48 ¡.46 ¡.46
IS-MI ¡.78 ¡.63 ¡.58 ¡.52
for the bifactor model across the different samples. Because the IS-OI .88 .90 .82 .97
clinical sample had slightly lower model fit, invariance was MI-OI ¡.78 ¡.72 ¡.71 ¡.57
tested using the clinical sample as the reference group com-
Note. SK D Self-Kindness subscale; SJ D Self-Judgment subscale; CH D Common
pared to the other three samples. A baseline model was first Humanity subscale; IS D Isolation subscale; MI D Mindfulness subscale; OI D
estimated in which the bifactor model was fitted simultaneously Overidentification subscale. All correlations were significant at p < .001.
 EXAMINING THE FACTOR STRUCTURE OF THE SCS 603

Table 5. Omega estimates of explained variance from the bifactor model across of self-compassion. Likewise, the clinical sample reported the
samples. lowest levels, which would be expected given their diagnosis of
Omega v OmegaH vH General SC factor recurrent major depressive disorder. Similarly, meditators
reported the highest levels of self-kindness, common humanity,
Student .95 .90 .95
Community .95 .89 .93 and mindfulness, and the lowest levels of self-judgment, isola-
Meditator .96 .90 .94 tion, and overidentification, whereas the clinical sample dis-
Clinical .94 .85 .90 played the opposite pattern.
Note. Omega D the proportion of variance in the total score accounted for by all Liberal fit criteria were used in interpreting the meaning of
factors; omegaH D the proportion of variance in the total score accounted for by results to determine if the SCS can measure a general factor of
the self-compassion (SC) factor; general SC factor D the amount of reliable vari- self-compassion as intended or if it needs to be redesigned.
ance (not due to error) that is accounted for by the SC factor.
CFAs were used to examine a bifactor model, a higher order
model, a six-factor correlated model, a two-factor correlated
model with only constrained factor loadings. As proposed by model, and a one-factor model. Because of arguments that fit
Cheung and Rensvold (2002), measurement invariance was measures alone should not be the deciding factor in determin-
supported if the DCFI between comparison models was equal ing the validity of a scale’s factor structure (e.g., Morgan,
to .01 or less. The DCFI between the baseline and the con- Hodge, Wells, & Watkins, 2015), we also used the bifactor
strained factor loading model comparing the clinical to the stu- model to calculate omega indexes to better inform judgments
dent, community, and meditator samples was less than .01, concerning the dimensionality of the SCS. Results generally
suggesting that the relationship between corresponding items found evidence for the idea that a total SCS score can be inter-
and the SCS subscale factors and the overall SC factor were preted using a bifactor model structure, but not a higher order
similar. The DCFI between the constrained factor loading model structure.
model and the constrained intercept model was .01 or less for The higher order model demonstrated relatively poor fit
all three sample comparisons, suggesting that item intercepts across samples, even in a student sample that was similar to the
are equal across the sample comparisons. This indicates that one in which the SCS was first developed. Whereas the RMSEA
there is no systematic variation across samples that would sug- and SRMR indexes generally suggested adequate fit of the
gest that some outside variable is systematically producing dif- higher order model in all four samples, the CFI and TLI indexes
ferences in the scores. suggested marginal or inadequate fit in all four samples. This
The estimates from the bifactor model were used to calculate suggests that the higher order model is not representing the
the omega indexes as previously described. The overall omega relationships among items satisfactorily, and is not the best
index .v/, omegaH, and the percentage of reliable variance way to understand the relationship between subscales or to jus-
accounted for by the self-compassion factor for each sample tify the use of a total SCS score. The fact that the higher order
are presented in Table 5. As seen in the first column, the overall model does not allow for modeling of the direct impact of the
omega index .v/ in all of the samples was .94 or greater, indi- target variable of self-compassion on individual item responses,
cating that the large majority of total variance in the scores but instead posits an indirect effect only, does not theoretically
could be attributed to both the self-compassion and the six sub- align with the conceptual underpinnings of the SCS and is
scale factors. Thus, the variance in the total score due to error most likely the source of the poor model fit.
was never greater than 6% in any of the samples. The second The bifactor model, in contrast, which simultaneously
column displays the omegaH indexes for each sample, which examines the contribution of a general factor and group factors
ranged from .85 to .90, indicating that a large majority of the to item variance in multidimensional measures, was found to
total variance in the scores can be attributed to the general, demonstrate better fit to the data than the higher order model.
overall self-compassion factor. Further, results presented in the Using liberal model fit criteria, the bifactor model generally
third column suggest that the overall self-compassion factor demonstrated acceptable fit according to most of the fit indexes
accounts for 90% to 95% of the reliable variance (i.e., not due in the student, community, and meditator samples. The bifac-
to error) in the total scores. tor model demonstrated suboptimal fit in the clinical sample,
however, indicating that the SCS might be operating slightly
differently for clinical populations. Although loadings on the
Discussion
general self-compassion factor in the bifactor model were gen-
The purpose of this study was to examine the factor structure of erally good for the three nonclinical samples (Comrey & Lee,
the SCS across a variety of populations, and, in particular, to 1992), there were fewer items with good loadings for the clini-
address the question of whether the use of an overall self-com- cal sample. It could be that individuals with a history of depres-
passion score is justified. Results suggested that the SCS had rel- sion have different patterns of emotional regulation, meaning
atively good psychometric properties and that a total SCS score that some items function a bit differently in relation to a gen-
could be reliably interpreted in four different populations—stu- eral self-compassion construct for these individuals, although
dent, community, meditator, and clinical—although findings the precise reasons underlying these differences will need to be
for the clinical sample were more mixed. First, it should be examined in future research. Still, when factor loadings and
highlighted that mean levels of self-compassion differed across item intercepts were compared across samples, those for the
samples in a theoretically consistent manner. As would be clinical sample were highly similar to those found for the other
expected given the Buddhist roots of the self-compassion con- three samples. Moreover, all but one item (mindfulness Item 9)
struct, the meditator sample reported the highest overall levels had meaningful loadings on the general factor according to
604 NEFF, WHITTAKER, KARL

Tabachnick and Fidell’s (2007) criteria. This suggests that the particular outcome, or to examine group differences in the six
SCS had the same basic factor structure in all four populations, aspects, it is probably best to examine the six subscales in addi-
although fit was not as good for the clinical population. tion to an overall score rather than examining one or more of
Fit indexes also supported the six-factor correlated structure the subscales completely on their own. These are interdepen-
of the SCS across samples. CFA results indicated that the six- dent parts of a whole, and should be understood and examined
factor correlated model demonstrated adequate model fit based as such.
on the CFI, TLI, RMSEA, and SRMR indexes in the student, Note that the one- and two-factor models examined for the
community, and meditator samples, although fit was deemed SCS had poor fit in all samples. Although no theorists we are
suboptimal based on the CFI and TLI indexes in the clinical aware of have argued that the SCS is unidimensional, there are
sample. In all of the samples, however, the vast majority of fac- some who have argued that the SCS is bidimensional, with a
tor loadings for the six-factor correlated model between an single self-compassion factor consisting of all the positive items
item and its respective subscale factor were good (Comrey & and a single self-criticism factor consisting of all the negative
Lee, 1992), and all loadings were meaningful (Tabachnick & items (Costa et al., 2015; Gilbert, McEwan, Matos & Rivis,
Fidell, 2007). As with the general self-compassion factor, when 2011; L
opez et al., 2015; Muris, 2015). Because social mentality
factor loadings and item intercepts for the six subscales were theory (Gilbert, 2005) posits that self-compassion taps into the
compared across samples, those for the clinical sample were mammalian caregiving system (associated with the parasympa-
highly similar to those found for the other three samples. thetic nervous system), whereas self-criticism taps into the
AIC comparisons indicated that the six-factor correlated threat defense system (associated with the sympathetic nervous
model would be most likely to replicate in all four samples, fol- system), Gilbert and colleagues have argued that positive and
lowed by the bifactor model, the higher order model, the two- negative self-affect should not be represented by an overall scale
factor correlated model, and finally the one-factor model, score (Gilbert et al., 2011). However, given that the sympathetic
respectively. One might argue that because the six-factor corre- and parasympathetic nervous systems continuously interact
lated model provided the best fit to the data, that the SCS sub- and covary (Porges, 2001), from our point of view there is no
scales should be used instead of a total score. According to reason why a single summary score cannot be used to assess
Reise et al. (2013), however, “even in the presence of multidi- the relative balance of system components.
mensionality, total scale scores justifiably can be interpreted” Muris (2015) also proposed that the positive and negative
(p. 132). First, note that the six factors were shown to have a items of self-compassion form separate constructs, and that
high degree of intercorrelation across samples, suggesting that self-compassion should only be assessed using the positive
the subscale factors are operating in concert with a “system” items because the negative items are conflated with psychopa-
view of self-compassion. A benefit of using a bifactor model in thology. However, a recent study by Krieger, Berger, and Holt-
psychometric analyses is that instead of relying on model fit cri- forth (2016) that used cross-lagged analyses to determine
teria alone (which is a debatable issue in itself), one can esti- whether changes in self-compassion led to changes in depres-
mate how much variance in the total score is explained by a sion or the reverse, found not only that self-compassion pre-
general factor as well as subscale factors, allowing for a more dicted depression (rather than the reverse) but that findings
nuanced and tangible measure of a scale’s ability to be used as were the same whether one examined a total scale score, a posi-
intended. Omega values indicated that at least 94% of the vari- tive factor only, or a negative factor only. They interpreted their
ance in the total scores was due to the general self-compassion findings as evidence that self-compassion should be considered
factor as well as the six subscale factors. Thus, the largest an overall construct rather than two separate constructs, and
amount of variance in the SCS total scores that could be attrib- current findings are congruent with this interpretation.
uted to error in any sample was 6%. Results also indicated that There are also theoretical problems with collapsing the three
90% or more of the reliable variance in total SCS scores was positive and three negative components into two separate fac-
attributable to an overall self-compassion factor in all popula- tors given that it would obscure important differences between
tions examined, including the clinical sample. This value components themselves. For instance, it would make it impos-
greatly exceeds the value of 75% suggested by Reise et al. (2013) sible to distinguish factors such as self-kindness and mindful-
as warranting confidence in the use of a total scale score. Over- ness, which are likely to tap into differing neurological and
all, these findings provide support for the use of a total SCS physiological systems (Engen & Singer, 2016). Our findings
score as a reliable measure of self-compassion, even in clinical that the six-factor intercorrelated and bifactor models had a
populations. better fit than the two-factor model suggest that the SCS can be
Of course, one could also argue that because the vast major- seen as having six subscale factors and a general factor of self-
ity of variance was explained by the general self-compassion compassion simultaneously, rather than being comprised of
factor, that the six subscales should not be examined indepen- two factors, one positive and one negative.
dently. Our position is that because the six-factor correlated The large majority of researchers to date have chosen to
model was found to have the best fit according to model com- examine self-compassion as an overall construct, most likely
parisons using the AIC across samples, it is valid to examine because it is simpler to conceptualize self-compassion as a sin-
the six subscales independently. However, researchers should gle state of mind that encompasses the compassionate versus
keep in mind that these are all aspects of self-compassion and uncompassionate ways that individuals emotionally respond,
not wholly separate entities. In other words, although it is pos- cognitively understand, and pay attention to their feelings of
sible to examine the six subscales to answer questions such as personal inadequacy and experiences of suffering. Also,
which aspect of self-compassion is the strongest predictor of a research interest in self-compassion is often motivated by its
 EXAMINING THE FACTOR STRUCTURE OF THE SCS 605

potential implications for psychological interventions, and pro- Albertson, E. R., Neff, K. D., & Dill-Shackleford, K. E. (2014). Self-compas-
grams designed to teach self-compassion tend to affect all its sion and body dissatisfaction in women: A randomized controlled trial
components simultaneously (Neff, 2016). It is simpler to ana- of a brief meditation intervention. Mindfulness, 6, 444–454.
Allen, A. B., Goldwasser, E. R., & Leary, M. R. (2012). Self-compassion and
lyze and report such findings with the use of a total SCS score, well-being among older adults. Self and Identity, 11, 428–453.
and current findings support the use of the SCS in this manner. American Psychiatric Association. (1994). Diagnostic and statistical man-
For particular research questions, however, it might also be ual of mental disorders (4th ed.). Washington, DC: Author.
appropriate to analyze the six subscales separately in addition Arch, J. J., Brown, K. W., Dean, D. J., Landy, L. N., Brown, K. D., & Lau-
to using a total score. For instance, one might want to deter- denslager, M. L. (2014). Self-compassion training modulates alpha-
amylase, heart rate variability, and subjective responses to social evalu-
mine if some people struggle with certain aspects of self- ative threat in women. Psychoneuroendocrinology, 42, 49–58.
compassion more than others or understand which compo- Barnard, L. K., & Curry, J. F. (2011). Self-compassion: Conceptualizations,
nents of self-compassion are most strongly related to particular correlates, & interventions. Review of General Psychology, 15, 289–303.
outcome variables. K€ orner et al. (2015), for example, used Beauducel, A., & Herzberg, P. Y. (2006). On the performance of maximum
regression analyses to examine which of the six self-compassion likelihood versus means and variance adjusted weighted least squares
estimation in CFA. Structural Equation Modeling, 14, 186–203.
components most strongly predicted depressive symptoms in a Behling, O., & Law, K. S. (2000). Translating questionnaires and other
large, nonclinical community sample, and found that isolation research instruments: Problems and solutions. Thousand Oaks, CA:
predicted 18% of the variance in symptoms, followed by over- Sage.
identification and self-kindness, which each predicted 2%, and Bentler, P. M., & Bonnett, D. G. (1980). Significance tests and goodness of
mindfulness and self-judgment, which each predicted 1%. fit in the analysis of covariance structures. Psychological Bulletin, 88,
588–606.
These types of studies are useful as they provide clues about Bentler, P. M., & Bonnett, D. G. (1990). Significance tests and goodness of fit in
which aspects of self-compassion might be important to target the analysis of covariance structures. Psychological Bulletin, 88, 588–606.
in interventions designed to address particular outcomes like Blatt, S. J., D’Afflitti, J., & Quinlan, D. (1976). Experiences of depression in
depressive symptomology. One of the benefits of the SCS is young adults. Journal of Abnormal Psychology, 65, 383–389.
that it can be used flexibly to meet particular research needs. Breines, J. G., & Chen, S. (2012). Self-compassion increases self-improve-
ment motivation. Personality and Social Psychology Bulletin, 38, 1133–
In summary, findings of this study using a bifactor model 1143.
support the conclusion that a total SCS scale score can be Breines, J. G., McInnis, C. M., Kuras, Y. I., Thoma, M. V., Gianferante, D.,
appropriately used to measure the general construct of self- Hanlin, L., & Rohleder, N. (2015). Self-compassionate young adults
compassion, whereas findings using a six-factor correlated show lower salivary alpha-amylase responses to repeated psychosocial
model suggest it can also be used to examine the six constituent stress. Self and Identity, 14, 390–402.
Breines, J. G., Thoma, M. V., Gianferante, D., Hanlin, L., Chen, X., &
components of self-compassion. Given that the higher order Rohleder, N. (2014). Self-compassion as a predictor of interleukin-6
model was not well supported in any of the samples examined, response to acute psychosocial stress. Brain, Behavior, and Immunity,
and that a bifactor model appears to be a superior way to model 37, 109–114.
Neff’s (2003b) conceptualization of self-compassion, results Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model
suggest that future attempts to validate translations of the SCS fit. In K. A. Bollen & J. S. Long (Eds.), Testing structural equation mod-
els (pp. 136–162). Newbury Park, CA: Sage.
or to examine the properties of the SCS in specific populations Buhrmester, M., Kwang, T., & Gosling, S. D. (2011). Amazon’s Mechanical
should not attempt to justify the use of a total SCS score using Turk: A new source of inexpensive, yet high quality, data? Perspectives
a higher order model. Instead, researchers should examine a on Psychological Science, 6(1), 3–5.
bifactor model (including estimating the amount of reliable Castilho, P., Pinto-Gouveia, J., & Duarte, J. (2015). Evaluating the multifac-
variance that can be attributed to an overall self-compassion tor structure of the long and short versions of the Self-Compassion
Scale in a clinical sample. Journal of Clinical Psychology, 71, 856–870.
score) in addition to a six-factor correlated model to investigate Chen, J., Yan, L., & Zhou, L. (2011). Reliability and validity of Chinese ver-
validity. sion of Self-Compassion Scale. Chinese Journal of Clinical Psychology,
19, 734–736.
Cheung, G. W., & Rensvold, R. B. (2002). Evaluating goodness of fit
Acknowledgments indexes for testing measurement invariance. Structural Equation
Modeling, 9, 233–255.
We are grateful to Willem Kuyken for his comments on a draft of this arti- Comrey, A., & Lee, H. (1992). A first course in factor analysis. Hillsdale, NJ:
cle and the trial team for allowing us to use the data. Erlbaum.
Costa, J., Mar^ oco, J., Pinto-Gouveia, J., Ferreira, C., & Castilho, P. (2015).
Validation of the psychometric properties of the Self-Compassion
Funding Scale: Testing the factorial validity and factorial invariance of the mea-
The clinical sample examined in this study was drawn from the PREVENT sure among borderline personality disorder, anxiety disorder, eating
Trial, a project funded by the National Institute for Health Research disorder and general populations. Clinical Psychology & Psychotherapy,
Health Technology Assessment Programme (Project Number 08/56/01). 23, 460–468.
This trial is reported in full in the Lancet (doi:https://doi.org/10.1016/ Daye, C. A., Webb, J. B., & Jafari, N. (2014). Exploring self-compassion as a
S0140-6736(14)62222-4). refuge against recalling the body-related shaming of caregiver eating
messages on dimensions of objectified body consciousness in college
women. Body Image, 11, 547–556.
References Diedrich, A., Grant, M., Hofmann, S. G., Hiller, W., & Berking, M. (2014).
Self-compassion as an emotion regulation strategy in major depressive
Akaike, H. (1973). Information theory and an extension of the maximum disorder. Behaviour Research and Therapy, 58, 43–51.
likelihood principle. In B. N. Petrov & F. Csaki (Eds.), Second interna- DiStefano, C., & Motl, R. W. (2006). Further investigating method effects
tional symposium on information theory (pp. 267–281). Budapest, Hun- associated with negatively worded items on self-report surveys. Struc-
gary: Akademiai Kiado. tural Equation Modeling, 13, 440–464.
606 NEFF, WHITTAKER, KARL

Engen, H. G., & Singer, T. (2016). Affect and motivation are critical in con- Marsh, H. W., Hau, K.-T., & Grayson, E. (2005). Goodness of fit in struc-
structive meditation. Trends in Cognitive Sciences, 20, 159–160. tural equation modeling. In A. Maydeu-Olivares & J. J. McArdle (Eds.),
Finlay-Jones, A. L., Rees, C. S., & Kane, R. T. (2015). Self-compassion, Contemporary psychometrics (pp. 275–340). Mahwah, NJ: Erlbaum.
emotion regulation and stress among Australian psychologists: Testing Marsh, H. W., Hau, K.-T., & Wen, Z. (2004). In search of the golden rules:
an emotion regulation model of self-compassion using structural equa- Comment on hypothesis-testing approaches to setting cutoff values for
tion modeling. PloS One, 10(7), e0133481. fit indexes and dangers in over-generalizing Hu and Bentler’s (1999)
Furr, R. M., & Bacharach, V. R. (2008). Psychometrics: An introduction. findings. Structural Equation Modeling, 11, 320–341.
Thousand Oaks, CA: Sage. McDonald, R. P. (1999). Test theory: A unified approach. Mahwah, NJ:
Gallini, J. K., & Mandeville, G. K. (1984). An investigation of the effects of Erlbaum.
sample size and specification error on the fit of structural equation Morgan, G., Hodge, K., Wells, K., & Watkins, M. (2015). Are fit indices
models. Journal of Experimental Education, 53, 9–19. biased in favor of bi-factor models in cognitive ability research? A com-
Garcia-Campayo, J., Navarro-Gil, M., Andr
es, E., Montero-Marin, J., parison of fit in correlated factors, higher-order, and bi-factor models
L
opez-Artal, L., & Demarzo, M. M. (2014). Validation of the Span- via Monte Carlo simulations. Journal of Intelligence, 3, 2–20.
ish versions of the long (26 items) and short (12 items) forms of Muris, P. (2015). A protective factor against mental health problems in
the Self-Compassion Scale (SCS). Health and Quality of Life Out- youths? A critical note on the assessment of self-compassion. Journal
comes, 12(1), 4. of Child and Family Studies, 25, 1461–1465.
Gilbert, P. (2005). Compassion: Conceptualizations, research and use in Muth
en, L. K., & Muth
en, B. O. (1998–2016). Mplus user’s guide (7th ed.).
psychotherapy. London, UK: Brunner-Routledge. Los Angeles, CA: Muth
en & Muth
en.
Gilbert, P., McEwan, K., Matos, M., & Rivis, A. (2011). Fears of compas- Neff, K. D. (2003a). Development and validation of a scale to measure self-
sion: Development of three self-report measures. Psychology and Psy- compassion. Self and Identity, 2, 223–250.
chotherapy: Theory, Research and Practice, 84, 239–255. Neff, K. D. (2003b). Self-compassion: An alternative conceptualization of a
Hancock, G. R., & Mueller, R. O. (2001). Rethinking construct reliability healthy attitude toward oneself. Self and Identity, 2, 85–102.
within latent variable systems. In R. Cudeck, S. du Toit, & D. S€ orbom Neff, K. D. (2016). The Self-Compassion Scale is a valid and theoretically
(Eds.), Structural equation modeling: Present and future (pp. 195–216). coherent measure of self-compassion. Mindfulness, 7, 264–274.
Lincolnwood, IL: Scientific Software International. Neff, K. D., & Beretvas, S. N. (2013). The role of self-compassion in roman-
Hollis-Walker, L., & Colosimo, K. (2011). Mindfulness, self-compassion, tic relationships. Self and Identity, 12, 78–98.
and happiness in non-meditators: A theoretical and empirical examina- Neff, K. D., & Germer, C. K. (2013). A pilot study and randomized con-
tion. Personality and Individual Differences, 50, 222–227. trolled trial of the mindful self-compassion program. Journal of Clinical
Holzinger, K. J., & Swineford, R. (1937). The bifactor method. Psychome- Psychology, 69, 28–44.
trika, 2, 41–54. Neff, K. D., Hsieh, Y., & Dejitthirat, K. (2005). Self-compassion, achieve-
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance ment goals, and coping with academic failure. Self and Identity, 4, 263–
structure analysis: Conventional criteria versus new alternatives. Struc- 287.
tural Equation Modeling, 6, 1–55. Neff, K. D., Kirkpatrick, K., & Rude, S. S. (2007). Self-compassion and its
Hupfield, J., & Ruffieux, N. (2011). Validierung einer deutschen Version link to adaptive psychological functioning. Journal of Research in Per-
der Self-Compassion Scale [Validation of a German Version of the sonality, 41, 139–154.
Self-Compassion Scale]. Zeitschrift f€ ur Klinische Psychologie und Psy- Neff, K. D., Pisitsungkagarn, K., & Hsieh, Y. (2008). Self-compassion and
chotherapie, 40, 115–123. self-construal in the United States, Thailand, and Taiwan. Journal of
Johnson, E. A., & O’Brien, K. A. (2013). Self-compassion soothes the savage ego- Cross-Cultural Psychology, 39, 267–285.
threat system: Effects on negative affect, shame, rumination, and depressive Neff, K. D., & Pommier, E. (2013). The relationship between self-com-
symptoms. Journal of Social and Clinical Psychology, 32, 939–963. passion and other-focused concern among college undergraduates,
K€orner, A., Coroiu, A., Copeland, L., Gomez-Garibello, C., Albani, C., Zen- community adults, and practicing meditators. Self and Identity, 12,
ger, M., & Br€ahler, E. (2015). The role of self-compassion in buffering 160–176.
symptoms of depression in the general population. PLoS ONE, 10(10), Neff, K. D., Rude, S. S., & Kirkpatrick, K. (2007). An examination of self-
e0136598. compassion in relation to positive psychological functioning and per-
Krieger, T., Berger, T., & Holtforth, M. (2016). The relationship of self-com- sonality traits. Journal of Research in Personality, 41, 908–916.
passion and depression: Cross-lagged panel analyses in depressed patients Odou, N., & Brinker, J. (2014). Exploring the relationship between rumina-
after outpatient therapy. Journal of Affective Disorders, 202, 39–45. tion, self-compassion, and mood. Self and Identity, 13, 449–459.
Lance, C. E., Butts, M. M., & Michels, L. C. (2006). The sources of four Petrocchi, N., Ottaviani, C., & Couyoumdjian, A. (2013). Dimensionality
commonly reported cutoff criteria. What did they really say? Organiza- of self-compassion: Translation and construct validation of the Self-
tional Research Methods, 9, 202–220. Compassion Scale in an Italian sample. Journal of Mental Health, 23,
Leary, M. R., Tate, E. B., Adams, C. E., Allen, A. B., & Hancock, J. (2007). 72–77.
Self-compassion and reactions to unpleasant self-relevant events: The Porges, S. W. (2001). The polyvagal theory: Phylogenetic substrates of a
implications of treating oneself kindly. Journal of Personality and Social social nervous system. International Journal of Psychophysiology, 42,
Psychology, 92, 887–904. 123–146.
L
opez, A., Sanderman, R., Smink, A., Zhang, Y., van Sonderen, E., Ran- Raes, F. (2010). Rumination and worry as mediators of the relationship
chor, A., & Schroevers, M. J. (2015). A reconsideration of the Self- between self-compassion and depression and anxiety. Personality and
Compassion Scale’s total score: Self-compassion versus self-criticism. Individual Differences, 48, 757–761.
PloS One, 10(7), e0132940. Reise, S. P., Bonifay, W. E., & Haviland, M. G. (2013). Scoring and model-
MacBeth, A., & Gumley, A. (2012). Exploring compassion: A meta-analysis ing psychological measures in the presence of multidimensionality.
of the association between self-compassion and psychopathology. Clin- Journal of Personality Assessment, 95, 129–140.
ical Psychology Review, 32, 545–552. Reise, S. P., Moore, T. M., & Haviland, M. G. (2010). Bifactor models and
MacCallum, R. C., Browne, M. W., & Sugawara, H. M. (1996). Power anal- rotations: Exploring the extent to which multidimensional data yield
ysis and determination of sample size for covariance structure model- univocal scale scores. Journal of Personality Assessment, 92, 544–559.
ing. Psychological Methods, 1, 130–149. Rhemtulla, M., Brosseau-Liard, P. E., & Savalei, V. (2012). When can cate-
Marsh, H. W., Balla, J. R., & McDonald, R. P. (1988). Goodness-of-fit gorical variables be treated as continuous? A comparison of robust con-
indexes in confirmatory factor analysis: The effect of sample size. Psy- tinuous and categorical SEM estimation methods under suboptimal
chological Bulletin, 103, 391–410. conditions. Psychological Methods, 17, 354–373.
Marsh, H. W., Hau, K.-T., Balla, J. R., & Grayson, E. (1998). Is more ever Saris, W. E., Satorra, A., & Van der Veld, W. M. (2009). Testing structural
too much? The number of indicators per factor in confirmatory factor equation models or detection of misspecifications? Structural Equation
analysis. Multivariate Behavioral Research, 33, 181–220. Modeling, 16, 561–582.
 EXAMINING THE FACTOR STRUCTURE OF THE SCS 607

Sbarra, D. A., Smith, H. L., & Mehl, M. R. (2012). When leaving your ex, Werner, K. H., Jazaieri, H., Goldin, P. R., Ziv, M., Heimberg, R. G., &
love yourself: Observational ratings of self-compassion predict the Gross, J. J. (2012). Self-compassion and social anxiety disorder. Anxi-
course of emotional recovery following marital separation. Psychologi- ety, Stress & Coping, 25, 543–558.
cal Science, 23, 261–269. West, S. G., Taylor, A. G., & Wu, W. (2012). Model fit and model selection
Shapira, L. B., & Mongrain, M. (2010). The benefits of self-compassion and in structural equation modeling. In R. H. Hoyle (Ed.), Handbook of
optimism exercises for individuals vulnerable to depression. The Jour- structural equation modeling (pp. 209–231). New York, NY: Guilford.
nal of Positive Psychology, 5, 377–389. Williams, M. J., Dalgleish, T., Karl, A., & Kuyken, W. (2014). Examining
Smeets, E., Neff, K., Alberts, H., & Peters, M. (2014). Meeting suffering the factor structures of the Five Facet Mindfulness Questionnaire and
with kindness: Effects of a brief self-compassion intervention for female the Self-Compassion Scale. Psychological Assessment, 26, 407–418.
college students. Journal of Clinical Psychology, 70, 794–807. Zessin, U., Dickhauser, O., & Garbade, S. (2015). The relationship between
Tabachnick, B. G., & Fidell, L. S. (2007). Using multivariate statistics (5th self-compassion and well-being: A meta-analysis. Applied Psychology:
ed.). Boston, MA: Pearson. Health and Well-Being, 7, 340–364.