The Rosenberg Self-Esteem Scale (RSES) is a widely used instrument that has been tested for reliability and validity in many settings; however, some negative-worded items appear to have caused it to reveal low reliability in a number of studies. In this study, we revised one negative item that had previously (from the previous studies) produced the worst outcome in terms of the structure of the scale, then re-analyzed the new version for its reliability and construct validity, comparing it to the original version with respect to fit indices.

In total, 851 students from Chiang Mai University (mean age: 19.51±1.7, 57% of whom were female), participated in this study. Of these, 664 students completed the Thai version of the original RSES - containing five positively worded and five negatively worded items, while 187 students used the revised version containing six positively worded and four negatively worded items. Confirmatory factor analysis was applied, using a uni-dimensional model with method effects and a correlated uniqueness approach.

The revised version showed the same level of reliability (good) as the original, but yielded a better model fit. The revised RSES demonstrated excellent fit statistics, with χ^{2}=29.19 (df=19, n=187, p=0.063), GFI=0.970, TFI=0.969, NFI=0.964, CFI=0.987, SRMR=0.040 and RMSEA=0.054.

The revised version of the Thai RSES demonstrated an equivalent level of reliability but a better construct validity when compared to the original.

The Rosenberg Self-Esteem Scale (RSES) is one of the most widely used self-esteem measures in social science research.

The factor structure of the RSES has been extensively studied, the debate focusing on whether it is a uni-dimensional or a two-dimensional model.

Among the negatively worded items that are attributed to the indeterminable factor structures, the most common item is "I wish I could have more respect for myself". Pullmann and Allik

In this investigation, we compared the model fit results obtained in previous studies with the results from this present study, using an independent sample.

In total, 851 students attending a university in northern Thailand, with ages ranging from 18 to 34 (mean±SD, 19.51±1.7) participated in this study. There were two sub-studies carried out within this project. In the first, 1,664 participants completed the Thai version of the original RSES (five positively worded items and five negatively worded items). In this group, the mean age was 19.87 (SD 1.85) (min-max=18-34), with 57% of the group being female. In the second study, 2,187 students participated and completed the revised version of the RSES. The mean age of this group was 18.63 (SD 0.63) (min-max=18-23), with 56% of this group being female (

The RSES was translated into Thai-with cultural adaptations, using the following steps. First, the authors translated the original English version of the RSES into Thai, then it was back-translated by a bilingual person (an English-Thai school teacher), who had not seen to the original RSES before. Cultural adaptations and comparisons of reading difficulty were checked. Third, both versions were compared and reviewed by consensus (comprising a bilingual psychologist and the authors), with a small number of disagreements found and corrected in this way. Finally, grammatical and printing errors were corrected before experimenting with the final version in a field trial. The Thai-RSES was tested for psychometric properties and found to demonstrate good reliability, and showed concurrent validity with attachment anxiety

The students were informed about the study after a class taken by a research assistant who was not otherwise associated with the class. Interested students were provided with a take-home pack containing an information sheet, questionnaires and an informed consent form. Each student later returned the completed questionnaires and the completed informed consent forms to the research assistant, who then separated the informed consent form from the anonymous data.

Two samples were independently analyzed. Data screening for factor analysis was conducted and found to be acceptable in both samples (i.e., an acceptable reliability; Cronbach's α>0.6), and all items showed skewness and kurtosis of <±2)(2). Missing values were managed by replacing them with the series mean. The sampling adequacy was good, with Kaiser-Meyer-Olkin (KMO) values of 0.91 for Group 1 and of 0.83 for Group 2. Bartlett's test of sphericity was significant in both samples (p<0.001)(3), and the maximum likelihood method, with an oblique rotation, was performed on the items.

Confirmatory Factor Analysis (CFA) was used to determine the fit and the number of factors to retain from the previously identified two-factor model. We chose to analyze and compare the results of both studies in terms of a uni-dimensional model with method effect, using the correlated uniqueness (CU) approach.^{2}/df being <3 (2). Modifications were made to the model after the initial analysis using modification indices, and internal consistency/reliability was determined by calculating Cronbach α coefficient.

There was no difference between the two groups in terms of age and gender distribution, and both groups scored higher in attachment anxiety scores than in attachment avoidance. There was no difference in both scales, including the depression scale scores, between the two groups.

Internal consistency was good, with a Cronbach's alpha of 0.86 in the first sample and 0.84 in the second sample. The mean rating of the items ranged from 2.23 to 3.31 in the original version, and from 2.95 to 3.36 in the revised version. The original version yielded factor loadings ranging from 0.277 to 0.808 - with communalities of 0.077 to 0.661; whereas, the revised version yielded factor loadings ranging from 0.361 to 0.814 - with communalities of 0.149 to 0.672 (

When calculating the CFA, a uni-dimensional construct with method effect testing and using a correlated uniqueness was adopted (as shown in ^{2}=29.19, df=19, n=187, p=0.063, GFI=0.970, Non-Normed Fit Index (NNFI) or TFI=0.969, NFI=0.964, CFI=0.987, SRMR=0.040 and RMSEA=0.054).

After investigating the correlation between the revised version and external measurements, the results were as expected. The attachment anxiety sub-scale and avoidance sub-scale correlated negatively with the revised RSES, as they did with the original version (r=-0.23, p<0.01; and r=-0.17, p<0.01, respectively). The same results occurred with the depression scale, TDI (r=-0.30, p<0.01).

It is clearly shown from the results that the mean score for item 5 increased from 2.23±0.82 to 3.18±0.66, meaning the total score for this scale increased significantly (t=4.0, p<0.01). More importantly, it improved on the factor loading and communality of the item, confirming that the assumption of response bias had been corrected. This is supported by a previous study by Marsh,

When compared to the original version the revised version demonstrated a comparable internal consistency, though it may be expected to produce a higher level of reliability than the original version if used with a sample size of similar magnitude to the original. In addition, the revised version produced an excellent model fit, with all the required criteria being met (goodness of fit>0.95, SRMR<0.08 and RMSEA<0.06, χ^{2}=29.19, df=19; with a p-value >0.05 indicative of a rejected H) - confirming the validity of the factor structure.

Besides item '5', item '7' appeared to be low in terms of communality both in the original and the revised version (h^{2}=0.194 and 0.149 respectively). In fact, both models yielded low values for communality - item 5 being the lowest and item 7 the second lowest, indicating that it is a poor indicator of this factor. Exploratory factor analysis showed that item 7 had a low factor loading (0.36) and a cross-loading on the other factor which led to relatively low communality on the designated factor (<0.2). Taking the content of item 7 into account, even though its meaning seems to be positive, that may not be the case; it could be regarded as 'neutral'. This unclear message may lead to it being a grey zone - with poor item-total correlation and ultimately producing unsatisfactory factor loading. All in all, item 7 should be further investigated and revised along the same lines we did with item 5. In addition, low factor loading and communality can also be attributed to sample size if the communality is not strong enough (less than 0.4), and if the size of the sample has a greater impact upon factor analysis outcomes.

Further studies of the revised model should be conducted employing a larger sample size, plus an invariance test of any gender differences should be addressed. Finally, a test-retest study should be conducted, since the present cross-sectional study limited our ability to draw conclusions regarding the stability of the construct.

The revised version of the Thai RSES demonstrated similar (good) levels of reliability to the original version, but showed a better construct validity.

Measurement model using the method factor for the RSES - comparing the original (A) and the revised versions (B). A: One factor - Global Self-Esteem, and the correlated uniqueness among five negative items and five positive items, B: One factor - Global Self-Esteem, and the correlated uniqueness among four negative items and six positive items. Global: Global Self-Esteem Scale, p1: item no.1 positively worded, n2: item no.2 negatively worded, e1: error term of item no. 1, double arrow headed lines: covariance lines. RSES: Rosenberg Self-Esteem Scale.

Demographic and psychometric characteristics of participants

RSES: Rosenberg Self-Esteem Scale, N: sample size, SD: standard deviation, NS: non-significant

A comparison of Mean, SD, Factor loadings, and communalities (h^{2}) between the original and revised version

^{*}re-worded to positive direction. F.L.: factor loading, M: mean

A comparison of the Fit Indices of RSES between the original and revised version

RSES: Rosenberg Self-Esteem Scale, NFI: Normed Fit Index, CFI: Comparative Fit Index, TLI: Tucker-Lewis Index, GFI: Goodness of Fit Index, SRMR: standardized root-mean-square residual, RMSEA: root-mean-square error of approximation