Factor structure and psychometric properties of a Swedish version of the Sussex-Oxford Compassion Scales (SOCS)

Abstract

The Sussex-Oxford Compassion Scales (SOCS) are recently developed measures of compassion, which have showed support for a five-factor structure for both other-compassion (SOCS-O) and self-compassion (SOCS-S). The study aimed to validate the Swedish translations of both the SOCS-O and the SOCS-S. A sample of adult participants was randomly split into either an exploratory sample (N = 403) or a replication sample (N = 402). The exploratory sample was used for both exploratory factor analysis and confirmatory factor analysis. In the replication sample, we (1) used CFA to validate results from the exploratory sample, (2) assessed measurement invariance for different groups (gender, nationality, age), and (3) evaluated psychometric properties using local fit. Results from both sub-samples support the presence of five-factor models for both SOCS-O (using 19 items) and SOCS-S (using 20 items). For both scales, measurement invariance is supported for all grouping variables, and local psychometric properties indicate good internal consistency with fairly good discriminant and convergent validity. This study supports the five-factor model of both other-compassion and self-compassion, respectively, and shows that the Swedish versions xof both SOCS-O and SOCS-S are reliable and valid instruments that can be used to index compassion with general adult populations in Sweden and Finland.

Keywords:

• compassion
• self-compassion
• SOCS-O
• SOCS-S
• factor structure
• psychometric properties

Introduction

Compassionate emotions can be directed both towards oneself (self-compassion) and toward others (compassion). With emphasis on the latter, Goetz et al. (2010, p. 351) defined compassion as “the feeling that arises in witnessing another’s suffering and that motivates a subsequent desire to help”. It has been argued that empathy is involved in the process of compassion, but that the two emotions still are distinct from each other. That is, while compassion involves feeling for another person, the affective component of empathy concerns feeling what another person is feeling (i.e., feeling with another person) (Goetz et al., 2010). The resulting feeling of compassion involves the motivation to help, which in turn, is characterized by an upregulation of positive affect (Engen & Singer, 2015). Even though compassion has emotional connotations, Strauss et al. (2016, p. 26) have argued that compassion is more than an emotion, as the concept also includes “perceptiveness or sensitivity to suffering, understanding of its universality, acceptance, nonjudgement, and distress tolerance, and intentions to act in helpful ways.”

Compassion is considered an evolutionarily advantageous trait for group survival and flourishing as it is related to altruism and cooperative behavior (Goetz et al., 2010). Findings indicate that it is possible to increase compassion through several different interventions (Kirby, 2017). There is evidence that short-term training in compassion increased activation in brain regions associated with social cognition and emotion regulation when seeing other individuals suffer, which was also followed by an increment in prosocial behavior—a core feature of compassion (Weng et al., 2013). A meta-analysis has also shown that compassion training is related to increments in well-being and decrements in different aspects of mental illness (Kirby et al., 2017).

In a review, Neff (2003a) defined the concept self-compassion as containing the three components: self-kindness, common humanity, and mindfulness. Even though compassion and self-compassion might seem like two opposing traits (with the former being other-oriented and the latter self-oriented), Strauss et al. (2016) pointed out that they share a common foundation stemming from the Buddhist philosophy, where the concepts are seen as being closely related. Both traits have also been related to social functioning and altruism (Dzwonkowska & Żak-Łykus, 2015). The component of common humanity, which is a part of both concepts, seems to explain a larger part of their common prosocial connection (Fuochi et al., 2018). Considering the similarity between oneself and humanity as a whole, self-compassion has been proposed to differ from other more individualistic self-concepts (Neff, 2003a). For example, the concept of self-compassion has been proposed to avoid the pitfalls related to efforts trying to increase self-esteem, which can also lead to elevated narcissism and self-centeredness (Barnard & Curry, 2011; Neff, 2003a; Neff, 2003b). Instead, self-compassion has been suggested as a promising concept for both sustainable individual well-being and social connectedness (Barnard & Curry, 2011; Dzwonkowska & Żak-Łykus, 2015; Fuochi et al., 2018; Neff, 2003a). Even though there seems to be a correspondence between compassion and self-compassion, some previous studies have found only a weak, and at times non-significant, association between the two (Hwang et al., 2008; López et al., 2018; Neff, 2003b; Neff & Pommier, 2013; Tatum, 2012), which led Strauss et al. (2016) to speculate that existing instruments on compassion and self-compassion fail to capture the common core of the two traits.

Due to both the variety of definitions of the concept compassion and the limitation of psychometric properties of available measures, Strauss et al. (2016) proposed a new definition of compassion based on the common elements earlier used in the literature: (1) recognizing suffering, (2) understanding the universality of suffering, (3) emotional resonance (i.e., feeling for those who suffer and connecting with their distress), (4) tolerating uncomfortable feelings aroused in response to a suffering person, and (5) motivation to act or acting to alleviate suffering (Strauss et al., 2016). Subsequently, Gu et al. (2017) used this definition to develop and evaluate a scale using a combination of self-report items from previous compassion measures and some newly developed items. While they found a general support for the five-factors, the factor associated with tolerating uncomfortable feelings tended to be less connected to the core construct. Following this study, Gu et al. (2019) gathered a group of both experts and non-experts from all continents in order to generate new items for the five elements as proposed by Strauss et al. (2016). They generated new items for compassion towards others, but also for compassion towards oneself (i.e., self-compassion). These two scales contain 20 items each and are a part of the Sussex-Oxford Compassion Scales (SOCS), which include compassion for others (SOCS-O) and self-compassion (SOCS-S; Gu et al., 2019). When testing the psychometric properties using two separate samples (health care workers and students), they found the best fit for a five-factor hierarchical model showing both good internal consistency, and good convergent and discriminant validity in relation to other measures. In both samples, Gu et al. (2019) also found moderate correlations between compassion for others and self-compassion, indicating a common ground for the concepts, but with an important distinction between them, as the moderate correlation means that there are other, non-overlapping, explanatory factors for the two concepts (Gu et al., 2019).

Since the original work on SOCS by Gu et al. (2019), the psychometric properties of the scales have been tested in translations to Slovak, Korean, Dutch, and Italian (Halamová & Kanovsky, 2021; Kim & Seo, 2021; Krijger et al., 2022; Lucarini et al., 2022). Findings from both the Korean and the Italian population (Kim & Seo, 2021; Lucarini et al., 2022) found support for five-factor hierarchical models for the SOCS-O (which is in line with the original study, Gu et al., 2019). In addition, SOCS-O showed adequate internal consistency, and convergent and discriminant validity in relation to other scales, which was also the case for the SOCS-S in the Kim and Seo study (2021). Halamová and Kanovsky (2021) found similar support for the factor structure of the SOCS-O with a Slovak sample, but their results of the SOCS-S indicated some concerns regarding the unidimensionality of the scale; the general self-compassion factor explained little variance within the model, likely because of the grouping of subscales with Recognizing and Understanding on the one hand, and Feeling, Tolerating, and Acting on the other. However, in a Dutch sample, Krijger et al. (2022) found support for five-factor hierarchical models of the SOCS-S using three different samples (crisis line volunteers, military personnel, and nursing students). Similar to the Halamová and Kanovsky study (2021), Krijger et al. (2022) found that Feeling, Tolerating, and Acting correlated more strongly to other convergent scales in comparison to Recognizing and Understanding, indicating a sort of grouping of these subscales here too. As a summary of previous findings, support has been found for both the factor structure and the psychometric properties of both the SOCS-O and the SOCS-S, while some concerns have been shown with a grouping of the subscales with the SOCS-S.

When it comes to differences between genders on both compassion for others and self-compassion, with populations of a wide array of ages, there has been a clear tendency that women score higher than men on compassion for others (López et al., 2018; Tatum, 2012; Weisberg et al., 2011), while some studies have shown no difference between women and men on self-compassion (Kang et al., 2018; López et al., 2018; Neff & Pommier, 2013; Sun et al., 2016; Tatum, 2012). The same pattern has also been shown for both compassion for others and self-compassion for the SOCS (Gu et al., 2019; Kim & Seo, 2021). On the contrary, meta-analytic work (see Yarnell et al., 2015) suggests that men score higher on self-compassion than women, but the effect size tends to be small.

In the previous studies of the SOCS, only Kim and Seo (2021) and Krijger et al. (2022) have tested measurement invariance. Kim and Seo (2021) tested invariance for both SOCS-O and SOCS-S across genders and found partial support for measurement invariance of both scales. Krijger et al. (2022) only tested invariance for SOCS-S and found support for measurement invariance for gender in two of the three samples, and measurement invariance for age groups in all three samples.

From the best of our knowledge, the SOCS have not yet been tested in populations from any of the Nordic countries. The main aim of this study was to examine the factor structure and psychometric properties of a Swedish translation of the SOCS-O and the SOCS-S by using an adult sample from both Sweden and Finland, adding to the variety of populations that the SOCS have been validated in. We aimed to test SOCS thoroughly by both exploring factor structure in a data-driven approach and comparing these outcomes with an a priori factor structure based on previous research of the scales. For this purpose, the participants were randomly split into two subsamples (an exploratory sample and a replication sample). In the exploratory sample we conducted both exploratory factor analysis (EFA) and confirmatory factor analysis (CFA) with independent subsamples. To evaluate if the results could be reproduced, we used CFA again in the independent replication sample to validate the results from the exploratory sample. In the replication sample, we also (1) examined whether the scales showed similar structure between different groups (gender, nationality, and age), and (2) examined reliability and validity of the scales. We expected to find support for the previous found five-factor (Recognizing, Universality, Feeling, Tolerating, and Acting) hierarchical models with an overarching factor for respective scales. Another aim of the study was to examine the association between SOCS-O and SOCS-S and possible gender differences between the scales.

Materials and methods

Participants

In order to reach a broad range of ages within different geographical areas in Sweden and in areas with a high density of Swedish speakers in Finland the participants (Swedish speaking adults ≥ 18 years old) were recruited via two different universities’ webpage and marketed posts on social media (Facebook and Instagram). Of the 1180 participants who opened the survey, 363 terminated the survey prior to the completion of all stages and 12 completed the survey but did not provide answers to all questions. In line with a complete case approach design, cases with any missing data were excluded, leaving a total of 805 participants, between the ages 18-83 years (Mean age = 45.13, SD = 14.04). Nationalities of the participants were Swedish (70.4%), Finnish (27.5%), and Other (2.1%). Of all participants, 21.7% were students, whereas 78.3% were not, and 87.0% of the total samples were women.

Participants were informed about the anonymous and confidential handling of data, as well as the voluntary nature of participation with a possibility of leaving the survey to end their participation. Participants were informed about the aim of the study to develop measures that could help us better understand social interactions, and they were required to give an informed consent before beginning to complete the survey. Due to the anonymity of the subjects, the informed consent, and earlier ethical approval of similar surveys and subjects, the local ethical committee approved the study and did not find it necessary to seek ethical approval from the regional ethics committee.

Procedure

The survey was administered using Qualtrics (http://www.qualtrics.com) between 14^th of March to 5^th of May 2020, and was presented to participants as a study concerning humans’ social functioning and preferences, with the purpose of improving both the measurement and the understanding of social interactions. The survey started with four demographic questions (age, gender, nationality, and whether the participant was a student or not) followed by the two SOCS questionnaires and two other questionnaires (measuring attachment security and empathy) not reported here. The four questionnaires were presented in a random order for the participants. To prevent participants from taking the survey more than once, Qualtrics placed a cookie in participants’ web browsers, excluding them from multiple partaking.

Measures

Compassion

The SOCS-O, measuring compassion, contained 20 items divided into five subscales with four items for each. Items were rated on a scale from 1 to 5 (1 = not at all true, 5 = always true). See Table S1 in supplementary materials for the items in Swedish and the Swedish instruction text for answering the items in Appendix A in supplementary materials. Cronbach’s α for the respective subscales were .866 (Recognizing), .816 (Universality), .703 (Feeling), .700 (Tolerating), and .846 (Acting).

Table 1 CFA of the SOCS-O in the exploratory sample: Model fit indices for all assessed factor models.

Model	χ^2 (df)	χ^2/df	CFI	SRMR	RMSEA	AIC
One-factor	671.488 (170)	3.95	.653	.108	.121	7374.249
Three-factor	307.123 (167)	1.84	.903	.086	.065	6956.863
Five-factor	233.559 (160)	1.46	.949	.057	.048	6890.112
Five-factor mod.	242.784 (142)	1.71	.944	.058	.053	6431.330
Five-factor hier.mod.	236.793 (148)	1.60	.937	.064	.055	6437.147

Note. CFI = Comparative Fit Index, SRMR = Standardized Root Mean Square Residual, RMSEA = Root Mean Square Error of Approximation, hier.mod. = hierarchical modified.

Self-compassion

The SOCS-S, measuring self-compassion, had 20 items with four for each of the five subscales. Items were rated on a scale from 1 to 5 (1 = not at all true, 5 = always true). See Table S2 in supplementary materials for the items in Swedish and the Swedish instruction text for answering the items in Appendix B in supplementary materials. Cronbach’s α for the five subscales in this study was .813 (Recognizing), .848 (Universality), .808 (Feeling), .779 (Tolerating), and .834 (Acting).

Translations

Permission to translate the scales was given by the first author of the study reporting the development and testing of the SOCS (Gu et al., 2019). The first draft of a translation was then done by two of the authors (AS and ÖS) of this study, both having Swedish as their native language. The translations were sent for a back-translation to a translation agency where an experienced Swedish-to-English back-translator performed it.

The translation was then validated through a quantitative process, recommended by Sperber (2004), in which all items from the scales of the original English version were presented in pairs with the back-translated English version and were judged by an independent panel on both language comparability and similarity of interpretation. The panel consisted of nine individuals who all have a proficiency in English, and all but one had at least a bachelor’s degree in psychology.

From both the quantitative results and the written qualitative feedback of the back-translated items from the fourth author of this study (JG), some corrections were made before sending the original items side-by-side with the Swedish translations and the back-translations to the fifth author of this study (BJ), for further written qualitative feedback. Some final revisions were made from the comments, yielding the final versions of the translations.

Data analytic procedures

In order to both test and cross-validate the factor structures of the scales, the 805 participants were randomly split into either an exploratory sample (N = 403) or a replication sample (N = 402). The exploratory sample consisted of 53 men (Mean age = 42.87, SD = 15.44) and 345 women (Mean age = 45.66, SD = 13.66), whereas the replication sample consisted of 42 men (Mean age = 39.50, SD = 13.78) and 355 women (Mean age = 45.67, SD = 14.04). The exploratory sample was further split into two subsamples, one for the EFA (N = 202) and one for the CFA (N = 201). With more than 200 participants for each analysis, an adequate number of participants were reached for both EFA and CFA in the exploratory sample (Kyriazos, 2018). As we used the whole subsample for the CFA in the replication sample, a good number of participants were reached there with over 300 participants (Kyriazos, 2018; Tabachnik & Fidell, 2019).

The EFA in the exploratory sample allowed us to compare the best fitting exploratory model with our a priori models. Parallel analysis was used to determine the number of factors in the EFA. The estimation method used for factor extraction was Maximum Likelihood with oblique rotation (promax).

Using CFA in both the exploratory and confirmatory samples, we then tested the a priori five-factor hierarchical model (with an overarching compassion or self-compassion factor loading on the five hypothesized factors) and contrasted this model with a one-factor model (with a compassion or self-compassion factor loading on all items), a possible alternative factor model (found from the results of the EFA), and a five-factor model (with correlations between factors). As the data in this study both showed some deviations to normality and were of ordinal character, we used a robust maximum likelihood estimator with a Satorra-Bentler scaled test statistic for the CFAs (see e.g. Li, 2016).

Four model fit indices were analyzed according to a priori cutoffs assessing measurement model validity. For this purpose, we used the Root Mean Square Error of Approximation (RMSEA), the Standardized Root Mean Residual (SRMR), the Comparative Fit Index (CFI), and the relative chi square statistics (χ²/df). We mainly used cutoffs recommended by Schermelleh-Engel et al. (2003) for assessing model fit. With respect to the RMSEA, values below .05 are considered a good-fitting model, values between .05–.08 indicates an acceptable fit, and values between .08–.10 a mediocre fit. SRMR values under .05 indicates a good fit to the data, while under .10 is considered acceptable. CFI values larger than .97 suggest a good fit and values above .95 an acceptable fit; however, a more liberal threshold of .90 for acceptable fit have also been used previously (Gu et al., 2019; Williams et al., 2014), and for the purpose of this study we consider this a mediocre fit. For χ²/df, an acceptable fit is reached when χ²/df is below 3, and a good fit below 2. We also used the Akaike information criterion (AIC) to compare fit between the different models. The lowest AIC indicates the best fit, and this fit index usually rewards more parsimonious models (Schermelleh-Engel et al., 2003). After the assessment of fit of the factor structures, the modification indices were examined to detect improvement of the models. We chose to not cross-correlate between different factors as these correlations complicate interpretation of the factor model.

In the confirmatory sample, the final models of the SOCS-O and the SOCS-S were also tested for measurement invariance by using several multi-group models that compared invariance between gender (men and women, N = 42 and 355), nationality (Swedish and Finnish, N = 283 and 111), and age (younger than 45 years; older/equal to 45 years, N = 185 and 217). Invariance was assessed using a sequential strategy testing the invariance at different levels; configural, metric, scalar, and residual invariance (Putnick & Bornstein, 2016). Even though the sample size was 402, the more conservative cutoff values for the fit indices (as described by Chen, 2007) were used for gender and nationality invariance tests due to unequal distribution of participants. For these two multi-group models, decrements in CFI ≥ .005 together with increments in RMSEA ≥ .01 indicate non-invariance between groups. As the age groups had more equal group sizes, the less conservative cutoff values were used, i.e., decrements in CFI ≥ −.01 together with increments in RMSEA ≥ .015 indicates non-invariance between groups (Chen, 2007). Adequate steps of measurement invariance testing would mean equivalence in factor structure (configural invariance), an equal contribution from items to latent factors (metric invariance), an equally captured shared variance of items by the latent factors (scalar invariance), and a comparable item and error variance (residual variance; Putnick & Bornstein, 2016) across groups. Finally, the confirmatory sample was assessed for internal consistency, and convergent and discriminant validity. This assessment was done with guidelines for cutoffs from Hair et al. (2014). Good internal consistency, shown by composite reliability (CR), should be ≥ .7. For convergent validity, all standardized factor loadings within a factor, as well as average variance extracted (AVE), should be ≥ .5. An AVE of .5 or higher indicates that, on average, more variance is explained by the items’ factor loadings than the measurement error. Discriminant validity was tested by comparing the AVE of each factor with the maximum shared variance (MSV) with any other factor in the model. AVE should be larger than MSV as it indicates that more variance is explained within the factor than through intercorrelations with another factor.

The CFAs were computed using the R (R Core Team, 2021) package lavaan (Rosseel, 2012). All other statistical analyses were performed using JASP (JASP Team, 2022).

Results

SOCS-O – exploratory sample

When we first performed an EFA of all items in SOCS-O to evaluate the number of factors and factor loadings, the parallel analysis suggested that three factors should be extracted, which differs from the a priori theoretical assumption of five factors. Factor 1 contained items mostly from Feeling, Tolerating, and Acting, Factor 2 contained Recognizing items, and Factor 3 contained Universality items. Two cross-loadings were Feeling3 and Feeling4 which loaded over .32 on both Factors 1 and 2, whereas Tolerating4 loaded under .32 on all factors (see Table S3 supplementary materials for all loadings).

Table 3 Indicators of internal consistency and validity (and factor correlations) for the five-factor model of the SOCS-O (with 19 Items) in the replication sample.

Factor	CR	AVE	MSV		1	2	3	4	5
1. Recognizing	.883	.654	.561		–
2. Universality	.852	.593	.254		.248	–
3. Feeling	.734	.412	.939		.749	.448	–
4. Tolerating	.704	.443	.939		.574	.504	.969	–
5. Acting	.862	.611	.801		.515	.395	.895	.861	–

Note. CR = Composite Reliability, AVE = Average Variance Extracted, MSV = Maximum Shared Variance.

Estimates for the model fits for the CFA of the SOCS-O are presented in Table 1. The one-factor model had unacceptable fit on all of the four fit indices (χ²/df, CFI, SRMR, and RMSEA). The three-factor model (based on the items and corresponding factors from the EFA) reached good fit for χ²/df, acceptable fit for SRMR and RMSEA, and mediocre fit for the CFI. While the five-factor model yielded good fit according to χ²/df and RMSEA, the SRMR was acceptable and the CFI was mediocre (barely below the threshold for acceptable fit).

When examining correlations between items and their parent factor in the five-factor model, all loadings exceeded .5 except for Feeling2 (.491) and Tolerating4 (.415). As Feeling2 was close to .5, it was included in subsequent analyses. After the removal of Tolerating4, the modification indices suggested no substantial improvement in model fit, and no other modifications were done. In this modified five-factor model, the χ²/df, CFI, and SRMR were very similar to the original five-factor model, the AIC was slightly improved, but the RMSEA showed somewhat worse fit.

Finally, when the five-factor hierarchical model was examined, a Heywood case appeared (as the relationship between the overarching compassion factor and Feeling showed a standardized coefficient of 1.03 and a residual variance of −.07) and this residual variance was fixed to 0 (which is in line with recommendations by Dillon et al. 1987) in the subsequent analyses. After the removal of Tolerating4, the modification indices suggested improved fit by cross-correlating Recognizing with Feeling4 (modification index = 22.663), but this was not suitable. The modified five-factor hierarchical model showed similar χ²/df, CFI, SRMR, RMSEA, and AIC compared to the five-factor modified model. In the exploratory sample of the SOCS-O, standardized correlations between the overarching compassion and underlying factors in the five-factor hierarchical model where: Recognizing at .56, Universality at .44, Feeling at 1.00, Tolerating at .96, and Acting at .86. All items correlated over .5 with corresponding factors (see Table S4 in supplementary materials for specific loadings).

Table 4 CFA of the SOCS-S in the exploratory sample: model fit indices for all assessed factor models.

Model	χ^2 (df)	χ^2/df	CFI	SRMR	RMSEA	AIC
One-factor	863.544 (170)	5.08	.558	.149	.142	8562.040
Three-factor	337.681 (167)	2.02	.891	.085	.071	7919.947
Five-factor	281.280 (160)	1.76	.923	.079	.061	7874.283
Five-factor hier.	296.368 (165)	1.80	.916	.090	.063	7880.503

Note. CFI = Comparative Fit Index, SRMR = Standardized Root Mean Square Residual, RMSEA = Root Mean Square Error of Approximation, hier. = hierarchical.

SOCS-O – replication sample

In this sample, only the five-factor models were analyzed as those showed substantially better fit than the one- and the three-factor models in the exploratory sample. Estimates for the model fits for the CFA of the SOCS-O are presented in Table 2. The five-factor model yielded good fit according to SRMR, just outside the threshold for good fit for χ²/df and RMSEA, and a mediocre fit (but close to acceptable) for the CFI.

Table 2 CFA of the SOCS-O in the replication sample: Model fit indices for all assessed factor models.

Model	χ^2 (df)	χ^2/df	CFI	SRMR	RMSEA	AIC
Five-factor	323.344 (160)	2.02	.944	.049	.050	13599.087
Five-factor mod.	252.368 (142)	1.78	.961	.043	.044	12615.827
Five-factor hier.mod.	276.795 (148)	1.87	.954	.047	.047	12633.365

Note. CFI = Comparative Fit Index, SRMR = Standardized Root Mean Square Residual, RMSEA = Root Mean Square Error of Approximation, hier.mod. = hierarchical modified.

The only item that correlated below .5 with its parent factor within the five-factor model was Tolerating4 (.386), and therefore, this item was removed. After the removal of Tolerating4, modification indices suggested to correlate Tolerating with Feeling1 (modification index = 21.707) and Recognizing to correlate with Feeling1 (modification index = 21.119), but these cross-factor modifications were not suitable. The modified five-factor model showed good fits for χ²/df, SRMR, and RMSEA, acceptable fit for the CFI, and better fit according to the AIC.

Finally, a five-factor hierarchical model was examined. As in the case with the exploratory sample, a Heywood case appeared in this model as the relationship between the overarching compassion factor and Feeling showed a standardized coefficient of 1.04 and a residual variance of −.04. This residual variance was yet again fixed to 0. After the removal of Tolerating4, modification indices suggested improved model fit by correlating Recognizing with Feeling4 (modification index = 32.63), and correlating Recognizing with Feeling (modification index = 24.13). These cross-factor modifications were not suitable. However, when looking at correlations between items and the parent factors, all loadings were > .5 except for Tolerating4, which was only weakly correlated with its parent factor (.377), and we therefore decided to remove this item. The modified five-factor hierarchical model showed (as with the exploratory sample) that the χ²/df, CFI, SRMR, RMSEA, and AIC were very similar to the five-factor modified model. In the replication sample of the SOCS-O, standardized correlations between the overarching compassion factor and underlying factors in the five-factor hierarchical model where: Recognizing at .65, Universality at .46, Feeling at 1.00, Tolerating at .96, and Acting at .88. All items correlated over .5 with corresponding factors (see Table S5 in supplementary materials for specific loadings).

Table 5 CFA of the SOCS-S in the replication sample: Model fit indices for all assessed factor models.

Model	χ^2 (df)	χ^2/df	CFI	SRMR	RMSEA	AIC
Five-factor	438.386 (160)	2.74	.915	.062	.066	16092.321
Five-factor modified	372.561 (159)	2.34	.935	.060	.058	16019.514
Five-factor hier.mod.	407.346 (164)	2.48	.926	.076	.061	16050.036

Note. CFI = Comparative Fit Index, SRMR = Standardized Root Mean Square Residual, RMSEA = Root Mean Square Error of Approximation, hier.mod. = hierarchical modified.

We chose the hierarchical model for invariance testing as it has similar fit indices as the non-hierarchical five-factor model but has more practical meaning as it includes the overarching concept of compassion measured by its five factors. Configural invariance was supported for all grouping variables as fit for CFI was mediocre to acceptable and fit for RMSEA acceptable to good. Also, results showed that all decrements in CFI were less than −.005, and all increments in RMSEA were less than .01 between all the steps of invariance testing for both gender and nationality. Furthermore, decrements in fit for CFI were less than −.01 and increments in fit for RMSEA were less than .015 for age. Therefore, measurement invariance for all grouping variables of the SOCS-O was supported (see specific results of the invariance test in Table S6 in supplementary materials).

Table 6 Indicators of internal consistency and validity (and factor correlations) for the five-factor modified model of the SOCS-S in the replication sample.

Factor	CR	AVE	MSV		1	2	3	4	5
1. Recognizing	.797	.496	.225		–
2. Universality	.878	.644	.193		.439	–
3. Feeling	.842	.576	.766		.409	.235	–
4. Tolerating	.774	.465	.766		.474	.335	.875	–
5. Acting	.842	.581	.697		.470	.292	.835	.832	–

Note. CR = Composite Reliability, AVE = Average Variance Extracted, MSV = Maximum Shared Variance.

As shown in Table 3, the SOCS-O with the five-factor model (without Tolerating4) showed good CR for all five subscales as the values were ≥ .7. Convergent validity was good for Recognizing, Universality, and Acting, but did not reach the AVE threshold of ≥ .5 for Feeling and Tolerating. AVE was larger than MSV for Recognizing and Universality, indicating good discriminant validity, but not for Feeling, Tolerating, and Acting.

SOCS-S – exploratory sample

We first performed an EFA in the exploratory sample for an indication of the appropriate number of factors using this method. The parallel analysis from the EFA of the SOCS-S also yielded three factors, which again differs from the a priori five-factor model. Results were overall similar to that of SOCS-O, Factor 1 contained mainly Feeling, Tolerating, and Acting items, Factor 2 contained Recognizing items (but also Tolerating1 and Feeling3), and Factor 3 contained Universality items. No cross-loadings were found (see Table S7 in supplementary materials for all loadings).

Table 7 Descriptive statistics of the 20-item versions of both SOCS-O and SOCS-S.

			SOCS-O
	Recognizing	Universality	Feeling	Tolerating	Acting
M	3.8	4.7	3.9	3.8 (3.9)	3.9
SD	0.6	0.4	0.5	0.6 (0.6)	0.6
			SOCS-S
	Recognizing	Universality	Feeling	Tolerating	Acting
M	4.0	4.7	3.1	3.2	3.3
SD	0.6	0.5	0.8	0.7	0.7

Note. Numbers in parenthesis are with the removal of the Tolerating 4 item from the scale.

Estimates for the fit indices for the CFA of the SOCS-S are presented in Table 4, showing a one-factor, three-factor, five-factor, and five-factor hierarchical model. The one-factor model had unacceptable fit on all four fit indices. The three-factor model with the items and corresponding factors found from the EFA reached acceptable fit for χ²/df, SRMR and RMSEA, but an unacceptable fit for CFI. The five-factor model showed good fit according to χ²/df, acceptable fit for SRMR and RMSEA, but a mediocre fit for CFI.

With respect to the loadings between items and their parent factors in the five-factor model, all but three relationships were ≥ .5. Those less than .5 were Feeling3 (.486), Tolerating1 (.468), and Tolerating4 (.485), but we decided to keep the items in the models as they were fairly close to .5. The modification indices suggested that correlating Feeling3 with Recognizing (modification index = 33.058), Tolerating3 with Feeling (modification index = 32.213), Feeling3 with Tolerating (modification index = 29.576), Tolerating2 with Feeling (26.025), and Tolerating3 with Recognizing (modification index = 21.573) would improve the model. However, these cross-factor modifications were not suitable.

The five-factor hierarchical model had similar fit indices as the five-factor model, with good fit according to χ²/df, acceptable fit for SRMR and RMSEA, and mediocre fit for CFI. Similar to the five-factor model, the same three items were the only ones with correlations to corresponding parent factors < .5, and these were once again remained in this model. Also, similar to the five-factor model, the modification indices suggested improvement only by cross-correlating items, and therefore the model was left unmodified. In the exploratory sample of the SOCS-S, standardized correlations between the overarching self-compassion factor and underlying factors in the five-factor hierarchical model where: Recognizing at .41, Universality at .21, Feeling at .93, Tolerating at .97, and Acting at .86 (see Table S8 in supplementary materials for specific loadings between items and corresponding factors).

SOCS-S – replication sample

In this sample, only the five-factor models were analyzed as those were substantially better than the one- and the three-factor models in the exploratory sample. Results from the CFA of the SOCS-S are presented in Table 5. The five-factor model showed acceptable fit for χ²/df, SRMR and RMSEA, and mediocre CFI.

In the five-factor model, all correlations between items and parent factors were over .5 except for Acting1 (which correlated .499), and all items were retained in the model. The modification indices suggested that correlating the residuals of Tolerating2 and Tolerating4 would improve model fit (modification index = 77.094). As these items also were judged as very similar, we allowed the residuals to correlate. This modification resulted in the five-factor modified model with slightly better model fit, but still at same threshold levels.

With the same modification (Tolerating2–Tolerating4) used for the five-factor hierarchical modified model, all four fit indices were in large comparable with the previous model, reaching same thresholds but with slightly worse fit for all indices. The AIC indicated a slightly worse fit for the five-factor hierarchical model compared to the modified five-factor model. In the replication sample of the SOCS-S, standardized correlations between the overarching self-compassion factor and underlying factors in the five-factor hierarchical model where: Recognizing at .50, Universality at .33, Feeling at .92, Tolerating at .94, and Acting at .90 (see Table S9 in supplementary materials for specific loadings between items and corresponding factors).

We chose the hierarchical model for invariance testing as it has similar fit indices as the non-hierarchical five-factor model but has more practical meaning as it includes the overarching concept of self-compassion measured by its five factors. Configural invariance was supported as CFI was mediocre and RMSEA was acceptable to good for all grouping variables (gender, nationality, and age). All steps of testing for age (by using less conservative thresholds than for gender and nationality) showed adequate changes of fit. For gender, metric invariance was supported by both fit indices, but there was some concern of scalar invariance with respect to CFI, and residual invariance was supported for both fit indices. Nationality did not fully meet the threshold for metric or scalar invariance with respect to CFI, however, invariance was found with respect to RMSEA. There were no concerns with respect to residual invariance. Therefore, measurement invariance for all grouping variables of the SOCS-S was supported (see specific results of the invariance test in Table S10 in supplementary materials).

As shown in Table 6, the SOCS-S with the five-factor modified model showed good CR for all subscales. Convergent validity was good for Universality, Feeling, and Acting, but not for Recognizing and Tolerating. Finally, discriminant validity was good for Recognizing and Universality, but not for Feeling, Tolerating, and Acting.

Correlations and mean values of the SOCS-O and the SOCS-S in the full sample

In the combined exploratory and replication samples (N = 805), the correlation between the SOCS-O and SOCS-S was r = .35 (p < .001). Even when the Universality subscales—which are almost identical for both scales—were removed, there was still a significant correlation at r = .27 (p < .001). Means and standard deviations for the scales are presented in Table 7.

As we got nested sub-samples from two different nationalities, we ran two intraclass correlation analyses for these two groups with all variables in the study to examine possible differences between them. The intraclass correlation coefficient (ICC) for the Swedish population (N = 567) was .202 with a 95% confidence interval from .182 to .224, whereas the ICC for the Finnish population (N = 221) was .254 with a 95% confidence interval from .219 to .296. With overlapping confidence intervals of the two sub-samples, data suggest that we cannot assume that the two sub-samples differ in their ratings on the SOCS.

We compared mean scores for the scales between genders. On the SOCS-O using 20 items, women (M = 4.05, SD = 0.40) scored significantly higher than men (M = 3.87, SD = 0.50), t(793) = 3.78, p < .001, d = 0.41. However, on the SOCS-S, the difference in mean scores between women (M = 3.66, SD = 0.51) and men (M = 3.57, SD = 0.47) was not significant, t(793) = 1.66, p = .098.

Discussion

This study assessed the factor structure and psychometric properties of a Swedish version of the SOCS. Unlike the previous theoretical conception of a five-factor solution for both compassion and self-compassion (Gu et al., 2017; Gu et al., 2019; Strauss et al., 2016), EFA results for both scales in this study suggest three-factor solutions that mainly have one factor for the combination of Feeling, Tolerating, and Acting, another for Recognizing, and a third for Universality. The three factors grouped together in the EFA (Feeling, Tolerating, and Acting) also show very strong covariation between each other in both scales, as seen by the results of the psychometric properties in the replication samples. Despite the EFA results and the high covariations of the mentioned triad of factors, CFA results from the exploratory sample show stronger support for five-factor solutions for both scales compared with both one- and three-factor solutions.

As far as we know, our study is the first to test a three-factor solution for SOCS-S. For SOCS-O, Lucarini et al. (2022) tested an exploratory three-factor solution by combining Feeling, Tolerating, and Acting into one factor, and by having the original Recognizing and Universality as two separate factors. When comparing this model to the one-factor and the five-factor solutions, they also found the best model fit for five-factor solutions. By comparing exploratory three-factor models with the hypothesized five-factor models, our findings give further support for the five-factor solutions for both SOCS-O and SOCS-S in Swedish speaking populations in Sweden and Finland. We also showed that the five-factor solution could be reproduced with similar fit in an independent sample, indicating that the results were robust.

Results from CFA of the SOCS-O showed a weak loading for the item Tolerating4 towards its parent factor (Tolerating) in the five-factor models. This was found both in the exploratory and the replication sample, leading to the removal of that item and leaving 19 items in the scale. When comparing the fit indices of 19- and 20-item versions with five factors in both samples, most fit indices were very similar between the two versions, but AIC favored the one using 19 items in both samples. Tolerating4 also had a particularly weak loading both in the original version of the scale (Gu et al., 2019) as well as in the Italian version of the scale (Lucarini et al., 2022), which raises the question whether this item should be removed or rephrased in future use of the SOCS-O. However, in the Korean version (Kim & Seo, 2021), this item posed no problem and showed a loading of .69 towards its parent factor. Inspecting the items based on three independent translating services online (Google Translate, Bing Microsoft Translator, and Reverso), the Korean version show high correspondence with the original English version. From our study, we cannot draw any conclusions with respect to why Tolerating4 have posed a problem in versions stemming from European countries. Future studies could investigate whether this difference could stem from cultural disparities.

A so-called Heywood case (standard coefficient over 1) appeared for the loading between the overarching compassion factor and Feeling in the five-factor hierarchical model of the SOCS-O in both the exploratory and the replication sample. This was found even though Heywood cases usually are rare in CFA with samples over 300 (as in the replication sample) with factors having at least three indicators (Hair et al., 2014). The complete case approach for handling missing data that was used in this study is known for increasing the risk for nonconvergence and could be one explanation for the Heywood case. On the other hand, this approach is recommended for studies using structural equation modeling (Hair et al., 2014). Also, considering that, for both samples, (1) the residual variance was close to zero, the (2) standard error of the factor was roughly similar to the other factors in the model, and (3) the fit of the model being adequate, one likely explanation for the Heywood case is that it stems from a sampling fluctuation (Dillon et al., 1987). Therefore, the decision was to fix the residual variance at zero, a solution that has been supported as appropriate for this kind of problem in empirical and simulated data (Dillon et al., 1987). This modification introduces an artificial perfect correlation between the overarching compassion factor and the Feeling factor. Considering Goetz et al.’s (2010) definition of compassion in which feeling for other individuals is central, it is possible that the Feeling component actually correlates quite strongly with compassion, but that sampling variability inflated the correlation in this study, thus causing a Heywood case. It is, however, not the first time, that this happens with the SOCS-O-scale, as Lucarini et al. (2022) also noticed a Heywood case for the Feeling factor. Future studies validating the SOCS-O could again assess the strength of association between the second-order compassion factor and all five factors, and to once again look into some of the suggestions for the problematic correlations between different factors that was found (but left unmodified) in the CFA of this study. The modification with highest impact for the 19 item five-factor hierarchical model of the SOCS-O in both the exploratory and the replication sample was to correlate the Feeling4 item with the Recognizing factor. It is easy to see that this item could be interpreted as a Recognizing item (especially so in a Swedish translation); rephrasing this item more in line with the Feeling concept could be considered.

CFA results of the SOCS-S with the exploratory sample showed that three items (Feeling3, Tolerating1, and Tolerating4) have somewhat low loadings (below .5), but we decided to keep these in the model as they were close to .5, but also as they have shown strong loadings previously (Gu et al., 2019; Kim & Seo, 2021). Results from the replication sample of the SOCS-S showed satisfactory loadings for the three aforementioned items, but with a loading just below .5 for another item (Acting1), and therefore no need in removing this item. In addition, this item has previously showed strong loadings in different samples (Gu et al., 2019; Kim & Seo, 2021). Therefore, we argue for the use of all 20 items in the SOCS-S.

The fit indices were quite similar between the five-factor models and the five-factor hierarchical models for both scales. Nevertheless, the fit indices were slightly better for the five-factor models in all instances except for χ²/df in the exploratory sample for SOCS-O, which was lower (better) for the hierarchical model. Few of the differences in fit indices in favor for the non-hierarchical models where however—as previously has been argued (Williams et al., 2014)—meaningful, as changes where not bigger than one point in the second decimal. The only meaningful differences in favor for the non-hierarchical models were SRMR with the exploratory sample for SOCS-S and χ²/df with the replication sample for SOCS-S; on the other hand, χ²/df was meaningfully in favor for the hierarchical model with the exploratory sample for SOCS-O. Given the theoretical and empirical basis for using the overarching concepts (compassion and self-compassion) as hierarchical factors for the five factors (Strauss et al., 2016; Gu et al., 2017; Gu et al., 2019), as well as close to similar fit between hierarchical and non-hierarchical models in this study, it could be argued that the value of using hierarchical models for SOCS-O and SOCS-S overrules the minor differences in favor for non-hierarchical models shown by fit indices in this study.

In a previous study (Gu et al., 2017), the Tolerating factor (using other items than in this study) was not well related to the core of Compassion, but here it was strongly related to its respective overarching factors. This shows that the items generated by Gu et al. (2019), and translated and tested here, support the Tolerating factor as part of both compassion and self-compassion. Several other studies analyzing different versions of SOCS in different populations have supported this too (Gu et al., 2019; Halamová & Kanovsky, 2021; Kim & Seo, 2021; Krijger et al., 2022; Lucarini et al., 2022), giving strong accumulated support for the Tolerating factor.

The CFAs showed that the five factors differ in strength towards respective overarching factor in both scales, but the patterns of the relationships between factors and overarching factors were quite similar for the two scales. The strong correlation between Feeling, Tolerating, and Acting with overarching factors could be expected based on the results from the EFA, where these factors, at large, were combined into one factor. This is also similar to previous findings (Halamová & Kanovsky, 2021) in which Feeling, Tolerating, and Acting have been more strongly correlated with overarching factors, in particular for the SOCS-S. Intercorrelations between that triad seem to cause some concerns with respect to discriminant validity for Feeling, Tolerating, and Acting for both scales in our study. While the convergent validity was good for all factors except Feeling and Tolerating in the SOCS-O, the convergent validity was good for all factors except for Recognizing and Tolerating in the SOCS-S. Both scales showed good internal consistency for all factors, which replicates findings from the original English versions (Gu et al., 2019) as well as other translated versions (Halamová & Kanovsky, 2021; Kim & Seo, 2021; Lucarini et al., 2022). However, some exceptions to internal consistency in the SOCS-S have previously been found in Halamová & Kanovsky (2021), as there was a problem with unidimensionality of the scale, and in Krijger et al. (2022) where the Recognizing scale was just below the threshold for adequate internal consistency in a nursing student sample. To the best of our knowledge, the present study was the first study to evaluate the local fit by assessing the convergent and discriminant validity of the factors within the models. The test-retest reliability should be considered in future validation studies.

Another approach in this study was to test the SOCS for measurement invariance. The invariance tests from both the SOCS-O and the SOCS-S showed no evidence of measurement non-invariance, neither for gender (men and women), nationality (Swedish and Finnish), nor age (under 45 and equal to or over 45 years), which further support the utility of the SOCS across different populations. These findings differ somewhat from previous findings in the Korean (Kim & Seo, 2021) and Dutch (Krijger et al., 2022) versions of the SOCS, in which only partial support was found for gender. Our study replicated Krijger et al.’s (2022) findings of full support for measurement invariance between two age groups. However, in our study of the SOCS-S, there were some instances in which changes in CFI fell outside the threshold for adequate change, but since invariance was met with respect to the RMSEA in all cases, measurement invariance was supported for all grouping variables in accordance with recommendations from Chen (2007).

One limitation of the study is the use of convenient sampling to gather participants. For the results to be more strongly generalized to the target populations (adults in Sweden and Finland), stratified sampling methods could have been used. However, results from the ICCs of the two nationalities show no statistical difference between ratings of the two, which mitigates the limitation somewhat. While efforts were made to reach out to a wide group of people (by using several channels), there were some uneven distributions of participants. There were, for example, more women than men and more Swedish than Finnish participants.

Correlational data of the combined samples show a moderate correlation between SOCS-O and SOCS-S when all the factors are included, but a weaker correlation when the Universality factor was excluded. As these correlations are similar in magnitude to findings from the English version (Gu et al., 2019), this study provides further support for compassion and self-compassion as related but distinct constructs. Note, however, that some previous findings have found only weak correlations between the constructs when using other measures and conceptualizations (Neff, 2003b; Neff & Pommier, 2013; Hwang et al., 2008).

Our results show that, while women score higher than men on compassion for others, there is no significant difference between the groups on self-compassion. These findings are fully consistent with previous studies comparing means between genders both with SOCS in other languages (Gu et al., 2019; Kim & Seo, 2021), but also with some studies using other measures of compassion for others and self-compassion (Kang et al., 2018; López et al., 2018; Neff & Pommier, 2013; Sun et al., 2016; Tatum, 2012; Weisberg et al., 2011). Therefore, this study further supports previous findings of women being more understanding and caring for other individuals.

To conclude, this study gives further support of the utility of the SOCS as measures of both compassion and self-compassion containing the five factors, first brought to light by Strauss et al. (2016), now tested with Swedish versions in general adult populations in Sweden and Finland. Both scales showed fit indices from mediocre to good with respect to the hierarchical models. Mediocre fits could however be questioned, as some argue for more conservative indices to assure the strength of factor models (Schermelleh-Engel et al., 2003). The scales are however supported by local psychometric properties showing good internal consistency and fairly good discriminant and convergent validity, as well as support for measurement invariance for three different grouping variables. We do however recommend continuous assessment of the SOCS to assure that the scales are psychometrically robust both over time and with different populations.

Supplementary Materials

Supplementary materials for this study are found on: https://osf.io/9y7kg/?view_only=0af165a8f7c84e7c9aa2d5842a29279d

Disclosure statement

The authors report there are no competing interests to declare.