Psychometric evaluation of the Swedish version of Rosenberg’s self-esteem scale

Abstract

Background: The widely used Rosenberg’s self-esteem scale (RSES) has not been evaluated for psychometric properties in Sweden.

Aims: This study aimed at analyzing its factor structure, internal consistency, criterion, convergent and discriminant validity, sensitivity to change, and whether a four-graded Likert-type response scale increased its reliability and validity compared to a yes/no response scale.

Methods: People with mental illness participating in intervention studies to (1) promote everyday life balance (N = 223) or (2) remedy self-stigma (N = 103) were included. Both samples completed the RSES and questionnaires addressing quality of life and sociodemographic data. Sample 1 also completed instruments chosen to assess convergent and discriminant validity: self-mastery (convergent validity), level of functioning and occupational engagement (discriminant validity). Confirmatory factor analysis (CFA), structural equation modeling, and conventional inferential statistics were used.

Results: Based on both samples, the Swedish RSES formed one factor and exhibited high internal consistency (>0.90). The two response scales were equivalent. Criterion validity in relation to quality of life was demonstrated. RSES could distinguish between women and men (women scoring lower) and between diagnostic groups (people with depression scoring lower). Correlations >0.5 with variables chosen to reflect convergent validity and around 0.2 with variables used to address discriminant validity further highlighted the construct validity of RSES. The instrument also showed sensitivity to change.

Conclusions: The Swedish RSES exhibited a one-component factor structure and showed good psychometric properties in terms of good internal consistency, criterion, convergent and discriminant validity, and sensitivity to change. The yes/no and the four-graded Likert-type response scales worked equivalently.

Keywords:

• Confirmatory factor analysis
• depression
• gender
• mental health
• self-esteem
• structural equation modeling
• well-being

Introduction

Self-esteem is a valuable personal asset and has, for example, shown to predict quality of life [1] and empowerment [2] among people with severe mental illness (SMI). It has also shown to be related to well-being across cultures and nations [3]. Self-esteem is thus regarded an important outcome from mental health services [4].

In order to assess effects of interventions, but also to describe, compare, and predict self-esteem, it is important with feasible and psychometrically sound measures. Rosenberg’s self-esteem scale (RSES) was developed for this purpose [5]. This 10-item scale exists in several language versions. For example, Schmitt and Allik [6] translated the RSES into 28 languages, administered it to almost 17,000 participants in 53 nations and found the RSES to be largely invariant across the included nations.

The construct behind the RSES tends to be more of a stable personality trait than an indicator of functional status [4]. It has been debated whether the RSES measures a single construct or the scale reflects more than one factor. It is often acknowledged that it consists of two sub-components, self-competence, and self-liking [7], where the former may be seen as one’s instrumental value and the latter as one’s intrinsic value [8]. There is also support for two factors where one is based on five positively worded items and the other on five negatively worded items [9]. Nevertheless, research generally indicates that the often found two components should be seen as two facets of a common factor and that the items produce a global self-esteem score [6,9,10].

Factors that have shown to distinguish between groups with different levels of self-esteem as measured by RSES are age, where an upside-down U relation was found [11]. Both younger and older people rated their self-esteem lower than the middle-aged group, peaking at 60 years. The same study also showed that female gender and low education were related with low self-esteem. Furthermore, the RSES has shown to be sensitive in identifying persons with depression and anxiety [12,13], and also in groups that do not primarily have a mental health disorder, such as stroke survivors and their spouses [14].

Despite extensive use of the instrument in research in Sweden [15–19], no psychometric testing seems to have been performed on a Swedish version of RSES. A literature search indicates that more than one translation exists, and two different response formats have been used. One is the four-graded Likert-type response scale originally proposed by Rosenberg [5] and the other is a dichotomous yes/no scale suggested by Oliver and colleagues [20] as more feasible with people with severe mental disorders.

The Swedish version of RSES needs to be investigated for psychometric properties. The aim of this study was to evaluate the psychometric properties of the RSES among people with SMI, addressing: factor structure, internal consistency, criterion, known-group and construct validity, and sensitivity to change. An additional aim was to test whether the increased level of variability in the four-graded Likert-type scale also increased the reliability and validity of the RSES.

Methods

The study was based on two samples. Sample 1 consisted of participants in a randomized controlled study evaluating rehabilitation for people with SMI. Outpatients attending community mental health services and service users from municipal psychiatric settings in southern and western Sweden were invited and participants completed the RSES using the yes/no response scale. The project was approved by the Regional Ethical Vetting Board in Lund, Reg. No. 2012/70. Sample 2 consisted of participants in a randomized controlled study investigating the effectiveness of an intervention focusing self-stigma in a sample of SMI. The study was performed in a larger city in western Sweden. The project was approved by the Regional Ethical Vetting Board in Lund, Reg. No. 2015/414. Sample 2 had completed the RSES with the four-graded Likert-type response scale and was included in the current study to enable testing of whether the yes/no scale and the four-graded Likert-type response scale were equivalent. All procedures, for both samples, followed the ethical standards of the Helsinki Declaration of 1975, as revised in 1983 and 2004, and all participants signed a written consent prior to the start of the data collection.

Participants

Sample 1 consisted of participants who had been randomized to one of two branches with active rehabilitation support [21]. They were treated as one sample for the purpose of the current study, since all participants received an intervention. The intervention was either a new activity-based intervention termed Balancing Everyday Life (BEL), provided by an occupational therapist, or standard rehabilitation (care as usual, CAU), also provided by an occupational therapist. BEL is a group-based rehabilitation program with 5–8 participants in each group. The group meets at 12 sessions, distributed over 16 weeks. A gatekeeper occupational therapist (staff member in the setting) identified clients according to the following inclusion criteria: aged 18–65 years, substance abuse not the main diagnosis, no comorbidity of dementia or developmental disorder, and enough command of Swedish to respond to questionnaires. The data used for the current study were collected before the intervention started and at completion after 16 weeks.

Sample 2 consisted of participants in a study investigating the effectiveness of a group-based anti-self-stigma intervention, narrative enhancement and cognitive therapy (NECT), with regard to changes in self-stigma, self-esteem, and subjective quality of life [22]. NECT is a structured, group-based intervention comprising 20 weekly 1-h sessions [23]. After screening for eligibility (scoring above a defined cutoff on a self-stigma questionnaire), 106 participants were included in a randomized controlled trial. Assessments were made at baseline and at termination of the intervention after 20 weeks. The intervention was given as an add-on to treatment as usual. Further inclusion criterion was an ability to read and speak Swedish since participant manuals are only available in Swedish.

The characteristics of the participants are shown in Table 1. Three participants in each sample were excluded because they had not completed the RSES, thus generating a total of 223 participants in Sample 1 and 103 in Sample 2. The two samples differed on all characteristics; Sample 1 consisting of more females and being younger, more often living with a partner, less often having a higher education and more often having an anxiety/depressive disorder. Regarding ‘other’ diagnoses, a neuropsychiatric disorder was the most common alternative in both samples.

Table 1. Characteristics of the participants.

Characteristics	Sample 1, N = 223	Sample 2, N = 103	p value
Sex [n (%) women]	163 (73.1)	53 (51.5)	<.001
Age [mean (SD)]	40.5 (11.1)	45.2 (11.2)	.001
Living with a partner [n (%)]	62 (28.2)	18 (18.2)	<.001
Education [n (%)]^a			.001
Nine-year compulsory school or lower	40 (19.3)	14 (13.6)
High school	123 (59.1)	45 (43.7)
College/university education	45 (21.6)	44 (42.7)
Diagnosis [n (%)]^b			<.001
Psychosis	45 (21.0)	51 (68.0)
Depression/anxiety disorder	110 (51.4)	12 (16.0)
Other	59 (27.6)	12 (16.0)

^a Missing data for 15 in Sample 1.

^b Diagnosis not available for 9 in Sample 1 and for 26 in Sample 2, mainly because these participants had recently developed mental illness and had not yet received a diagnosis.

Instruments used with both samples

The RSES [5] has five positively and five negatively worded items (cf. Table 2). The original response scale has four alternatives; strongly agree, agree, disagree, and strongly disagree. When developing the Lancashire Quality of Life Profile [20], the authors suggested a dichotomous response scale where the ‘strongly agree’ and ‘agree’ responses were collapsed into ‘yes’, and ‘strongly disagree’ and ‘disagree’ were collapsed into ‘no’. The authors argued that the four-graded Likert-type response scale was perceived as confusing by a test panel of mental health service users [24] and estimated that a simple yes/no response format would produce less ambivalence and suit people with a mental illness better. The scoring is yes = 1 and no = 0 for the positively worded items, whereas yes = 0 and no = 1 for the negatively worded items. The mean of the negative items is subtracted from the mean of the positive items. This gives a possible score that ranges between −1 and 1, and 0 indicates a neutral level of self-esteem. This dichotomous response scale has been used with RSES in several studies [13,18,24,25] but does not seem to have been psychometrically investigated. This scale was used with Sample 1, whereas Sample 2 responded to the widely used original four-graded Likert-type response scale where a rating of one represents the worst and four the best self-esteem [5,6].

Table 2. Standardized factor loadings for the two samples.

	Dichotomous response scale		Four-graded likert-type response scale
Item	Sample 1 (N = 223)	Sample 2 (N = 103)	Sample 2 (N = 103)
A person of worth	0.895	0.816	0.809
Have a number of good qualities	0.751	0.750	0.728
Feel that I am a failure	0.711	0.805	0.765
Able to do things as well	0.910	0.689	0.539
Do not have much to be proud of	0.867	0.664	0.629
Positive attitude toward myself	0.940	0.966	0.909
Satisfied with myself	0.781	0.877	0.858
Wish more respect for myself	0.636	0.666	0.578
Feel useless at times	0.911	0.877	0.845
Think I am no good at all	0.895	0.857	0.811

Italics indicate the negatively worded items.

The Manchester Short Assessment of Quality of Life (MANSA) [26] includes two quality of life estimates. One is a one-item estimate of general quality of life. The other is an index of satisfaction with 11 specified life domains, such as housing, economy, interpersonal relationships, and health. The items are rated according to a seven-point scale ranging from 1 = ‘could not be worse’ to 7 = ‘could not be better’. A Swedish MANSA version was used shown to have adequate construct validity and good internal consistency [27].

Background questionnaires were used as well. Both samples responded to questions concerning sociodemographic data and self-reported psychiatric diagnosis. The trustworthiness of self-reported diagnoses has previously been validated by comparing the resulting diagnostic groups on psychopathology based on interviewer ratings and finding logical differences, such as those diagnosed with mood disorders showing more depressive symptoms [28].

Instruments used with Sample 1

The self-mastery concept was measured by the Pearlin Mastery scale, which has dependably shown good internal consistency [29–31]. Seven statements are rated from 1 to 4. Two of these need to be reversed; then the ratings are summed into a score where a higher rating denotes stronger self-mastery. The Swedish version (Mastery-S) used in the current study has also been found valid and reliable [32].

The Global Assessment of Functioning (GAF) scale [33] generates two ratings, one of severity of symptomatology and one reflecting level of functioning. The scales range from 0 to 100, where a higher score indicates better functioning. Satisfactory inter-rater reliability has been demonstrated and GAF has been recommended for clinical use [34,35]. All research assistants who collected the data received training in performing the GAF rating by using videos and were calibrated against an expert GAF rater.

The participants rated their worker role by means of the Worker Role Self-assessment, WRS [36,37]. The worker role is conceptualized as beliefs in having and managing a role in working life. WRS was developed in Sweden and has shown appropriate psychometric properties in terms of internal consistency, construct validity, and test–retest reliability.

Data analysis

Both samples were used for the analyses if nothing else is indicated. The item structure was tested with confirmatory factor analyses (CFA) and included a between-group invariance test of the CFA. One CFA was tested for the dichotomous response scale and two different CFAs were tested for the four-graded Likert-type response scale, the first based on the four-level raw Likert ratings and the second based on a dichotomized version where ratings 1 and 2 were set to form one category and ratings 3 and 4 the other. The dichotomized version was used for invariance testing. Estimations were conducted with MPlus and we applied the Weighted Least Square Mean and Variance corrected estimator (WLSMV) because all items were categorical with a presumed approximate latent normal distribution. The invariance testing was conducted on dichotomous data. Since we used WLSMV estimation, the DIFFTEST implemented in Mplus was used to test for significant differences between models. Reliability was calculated using the R psych package. Two indices are presented: ω hierarchical, an index that estimates the reliability of the general factor of a scale, and Cronbach’s α. Both were based on the polychoric correlations (the same also used for the factor analysis).

Furthermore, we performed structural modeling using the RSES as a predictor of a number of criteria (quality of life, gender, age, diagnosis, and education level) to investigate the validity of the Swedish scales. Against the backdrop of research, women were expected to score lower on self-esteem, as well as those with a diagnosis of depression/anxiety (vs. other diagnoses) and the younger participants (vs. a middle-aged group). No group of older people was available for this study since no participants over 65 were included in Sample 1 and very few in Sample 2. We also compared these validities based on the full Likert scale and the dichotomized version, based on Sample 2.

Model fit was estimated with the χ², the comparative fit index (CFI: using above 0.95 as indicating excellent fit) and root mean square error of approximation (RMSEA: using 0.05 or below as indicating excellent fit) [38].

Construct validity may be investigated by a number of approaches, one of which is to compare the targeted construct with various indicator variables. Constructs that are theoretically similar are expected to show inter-correlations [39], and if the relationship with an established instrument assessing the indicator variable is strong, often determined as a correlation coefficient >0.5 [40], then this indicates convergent validity of the target construct. Conversely, constructs that are theoretically different would be expected to show only low correlations, below 0.30 [40]. If such a low correlation can be demonstrated with an established measure of a construct regarded as theoretically dissimilar, discriminant validity has been shown. Both convergent and discriminant validity were selected to address construct validity. Quality of life and self-mastery were regarded as theoretically close to self-esteem and thus suitable to assess convergent validity. Just as self-esteem, they are subjective estimates of the quality of an inner, personal experience, and all three have been found to be related in previous research [41–43]. The constructs chosen to reflect constructs regarded as theoretically different from self-esteem, thus discriminant validity, concerned symptomatology and level of functioning as rated by a research assistant and a self-rating of one’s worker role. The analyses testing convergent and discriminant validity were based solely on Sample 1.

The paired-samples t-test was used to assess sensitivity to change. The software used was the SPSS version 23 [44]. The alpha value for significant results was set at 0.05.

Results

CFA and internal consistency of the RSES

The measurement model of the RSES was tested with CFA. Previous research has supported a one-factor model and thus this model was tested first. The dichotomous response scale was used with Sample 1, and the loadings of the one-factor model are displayed in Table 2. The model fit was excellent, χ²(35) = 71.49, p < .001, CFI = 0.988, RMSEA = 0.068. The largest modification index was an error correlation between item 9 and 10, adding this revealed an even better fit, χ²(35) = 44.28, p < .001, CFI = 0.997, RMSEA = 0.037. Using Sample 2 and the dichotomized version of the scale (loadings in Table 2) also revealed very good fit, χ²(35) = 61.06, p < .004, CFI = 0.975, RMSEA = 0.085. Loadings were generally very strong (see Table 2), suggesting the scale to be highly reliable. The misfit in the model was mainly caused by two correlations, between items 1 and 2 and between items 6 and 7. Adding these two error correlations increased fit to an even higher level, χ²(33) = 37.56, p < .005, CFI = 0.996, RMSEA = 0.036. Cronbach’s α were 0.95 and 0.93 for Sample 1 and Sample 2, respectively. Hierarchical ω was 0.90 and 0.72 for Sample 1 and Sample 2, respectively. The lower level of hierarchical ω reflects the error correlations found in the CFA. These results generally support the one-factor model of the RSES in both samples, but there were some items that created small sub-factors, especially in Sample 2.

The one-factor CFA using the full information from the four-graded Likert-type response scale also revealed good fit, at least as regards CFI, χ² = 124.29, p < .001, CFI = 0.946, RMSEA = 0.157. The inconsistency between the fit indices could be caused by many small correlations between the latent error variables in the model, which was thus tested. There were a few modification indexes related to error correlations, the highest between items 1 and 2, items 6 and 7, and items 9 and 10. Adding these correlations increased fit, even if this was still not excellent, χ² = 55.35, p < .005, CFI = 0.986, RMSEA = 0.084. CFI was very high, indicating that the bulk of co-variability of the items was included in the model. Loadings were comparable to the dichotomized version (see Table 2). No other models were tested due to strong support for the one-factor model (with one error correlation). The reliability was 0.91 based on Cronbach’s α and the hierarchical ω was 0.80.

The loading structure based on Samples 1 and 2 was tested for invariance. The χ² difference (Δ) was tested between a model assuming the same loadings for both samples and a model where the loadings were free to vary between the samples. The more restrictive model, with fixed loadings between groups, was significantly different from the less restricted model that estimated the loadings separately, Δχ²(10) = 42.58, p < .001. On the other hand, there were only very small differences between the fit indices of the two models ΔCFI = 0.985–0.978, ΔRMSEA = 0.084–0.078, suggesting this difference to be small although statistically significant. Table 2 displays the loadings from both Sample 1 and Sample 2. The largest differences between the samples were found for items 4 and 5, where both loadings were stronger in Sample 1. In addition, we found a higher mean for the latent variable in Sample 2. This difference in mean was tested for significance using the restricted (loadings being equal) model and was found to be significant, Δχ²(1) = 8.825, p < .01 (latent mean difference was 0.308). This suggests that Sample 1 rated lower levels of self-esteem. To summarize, the results suggest that the construct validity of the RSES was supported in both samples and with both response scales.

Criterion validity

In addition to the structural validation, we also included criterion validation of the scales. A measure of quality of life (satisfaction with life domains) was used as the dependent variable and the latent self-esteem factor as the independent variable. In Sample 1, the standardized coefficient was β = 0.58 and in Sample 2 β = 0.64, p < .001, suggesting that self-esteem had a strong relation to quality of life in these samples. We also tested for differences on sex, age, education, and diagnosis (depression compared to all other diagnoses). Sex predicted self-esteem in both samples, suggesting males to rate higher, standardized β = 0.15, p = .033, and β = 0.22, p = .038, respectively. Age predicted self-esteem in Sample 1, β = 0.21, p = .004, but not in the second sample, β = −0.01, p > .05. Level of education did not predict self-esteem in any of the samples. Lastly, patients with a diagnosis of depression had lower levels of self-esteem in both samples, β = −0.35, p < .001 and β = −0.30, p = .017 for the respective samples.

We further compared the validity of RSES using two response scale versions for Sample 2, the original Likert scale and a version based on dichotomizing the Likert scale. It was found that the Likert scale version had a somewhat stronger significant relation to having a diagnosis of depression (β = 0.25 and 0.34, respectively), but on the other hand, the dichotomized version had a somewhat stronger relation to quality of life (β = 0.62 and 0.50, respectively).

Based on Sample 1, all correlations with indicator variables chosen to test convergent validity were r > 0.5 (Table 3), whereas those with variables expected to indicate discriminant validity were just above 0.2. All correlations but one were statistically significant at p < .001; that concerning GAF symptoms was p = .002.

Table 3. Correlations between RSES and the indicator variables based on Sample 1.

	Correlation (r)	95% CI for r
Indicators of convergent validity
Quality of life (general)	0.52	0.40–0.62
Satisfaction with life domains	0.52	0.40–0.62
Self-mastery	0.52	0.40–0.62
Indicators of discriminant validity
GAF symptoms	0.21	0.05–0.35
GAF functioning	0.25	0.10–0.39
Worker role self-rating	0.24	0.09–0.38

Sensitivity to change

A comparison of RSES differences between the measurement points for Sample 1, thus based on the yes/no response scale, indicated statistically significant changes from baseline to 16 weeks of rehabilitation (from −0.15 to −0.05, p = .002), suggesting sensitivity to change. Similarly, based on Sample 2 and the four-graded Likert-type response scale, but only including those who had received the anti-self-stigma intervention, a statistically significant improvement was found from baseline to completed intervention (from 2.47 to 2.71, p = .006, n = 39).

Discussion

The CFA showed that the Swedish RSES functioned well in both samples and that the response scales were equivalent. The four-graded Likert-type response scale did not increase reliability or validity, which is a unique finding. No previous research seems to have addressed this issue, and renders support to the argument that a yes/no response scale could be used to reduce ambivalence [24] without jeopardizing the psychometric properties of the RSES.

A one-factor model yielded strong support, which is in agreement with several previous studies [6,9,10]. In line with research on a number of language versions [6], strong internal consistency reliability was indicated. This said, there were small sub-factors (error correlations) that can be attributed to the different themes in the questionnaire, for example, the first 2 items are about self-worth, items 6, 7, and 8 concern attitudes toward the self, and 9 and 10 about a negative view in relation to competence. These sub-factors were somewhat more pronounced in Sample 2, also indicated by the somewhat lower hierarchical omega in Sample 2. This tendency, which thus seems to be related to the phrasing of items, could also be attributed to differences between the samples, such as more people with diagnoses of depression/anxiety in Sample 1. Depression might entail a reduced focus on the wording details in the items, resulting in a more generalized response. That would be in line with research showing that depression is linked with a variety of subjective well-being criteria [45], in turn shown to reflect a general construct [46].

Criterion validity in relation to quality of life (satisfaction with life domains) was excellent in both samples, further underscoring that both response scales functioned well. The strongest indication of criterion validity concerned depression, which is in agreement with international research [12,13]. Swedish RSES showed further criterion validity by discriminating between the sexes and diagnostic groups (depression vs. other). The RSES also discriminated between younger- and middle-aged participants in Sample 1, but not, however, in Sample 2. Sex differences have been identified in some studies [7,11], but not in others [13,47]. This may be due to methodological issues, such as the sample size and the target group. The current study investigated fairly large samples from a vulnerable group. Intersectionality between mental illness and gender may have accentuated the identified sex differences. The fact that RSES did not discriminate between groups based on education may be due to the fact that the participants, just as people in Sweden in general, were fairly highly educated – >80% in both samples had completed high school or more. There was thus no truly educationally disadvantaged group to compare with, which may have been the case in research based in other societies [11].

The findings also indicated good construct validity in terms of convergent and discriminant validity. Furthermore, RSES showed sensitivity to change. At re-assessment after the interventions, the participants’ self-esteem scores were significantly higher than at baseline in both samples. The participants had improved on other measures as well, such as quality of life and activity engagement in Sample 1 [21] and reduced stigma in Sample 2 [22], which supports that the RSES detected a true change. The evidence of the ability of RSES to detect change is also acknowledged in other research [15,48].

Focusing specifically on the two response formats, this study indicates that they were equally suited for distinguishing diagnostic groups (depressive vs. other disorders) and between the sexes and for detecting changes over time in self-esteem. Small deviations among findings concerning factor structure might rather be due to the phrasing of items, and how they were received by the respondents, than the response format. This study thus renders support to both response scales.

The fact that this study was based on people with mental illness deserves some attention. The previous psychometric testing of the RSES, although extensive and performed in several countries and language versions, is primarily built on surveys in different samples from the general population. By showing that the RSES was psychometrically sound when used with people with mental illness, the current study thus contributed with new knowledge. On the other hand, no sample from the general population was included, which makes testing in various target groups an important agenda for future research on the Swedish RSES.

Methodological concerns

The two samples were not equivalent and differed on all variables displayed in Table 1. The difference in diagnosis would also explain the lower RSES rating in Sample 1, which consisted of larger proportions of women and of people with a diagnosis of depression or anxiety disorder. The sample differences would not have an impact on this study’s ability to meet the study aim, however, and the findings must be accredited both internal and external validity.

Sample size in relation to the statistical methods is always an important issue. In CFA the requested sample size is dependent on the number of factors, the number of indicators and the strength of the loadings. With a one-factor model, the number of indicators used in the present study and very high loadings (0.80), the requirements would be less than 100 participants [49]. The present study is thus clearly on the safe side in that respect.

Admittedly, further psychometric testing of the Swedish RSES is warranted. The current study can report solid findings regarding the factor structure, but other analyses regarding construct validity, in terms of criterion, convergent, and discriminant validity, must be regarded as tentative since no gold standards were available to compare these with. Similarly, the tests of sensitivity to change were crude. More sophisticated methods, such as first conducting a pilot study to establish likely change characteristics and/or calculating standard response mean [50], should be applied in future attempts to establish RSES’s sensitivity to change.

Conclusions

The Swedish RSES showed good psychometric properties in terms of a one-factor structure, good internal consistency reliability, and expected group differences and associations that indicated criterion, convergent, and discriminant validity. This study also showed initial evidence of sensitivity to change. It can therefore be recommended for clinical and research purposes. The yes/no and the four-graded Likert-type response scales worked equivalently, indicating that either version may be used and the choice can be based on the target group. More psychometric testing seems warranted, however, such as test–retest reliability, predictive validity, and more detailed testing of sensitivity to change.

Acknowledgements

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Primary data may be obtained on request from the author.

Disclosure statement

The authors declare they have no competing interests.