Cross-loading items.

The first of these issues relates to how cross-loading items should be dealt with. While there is no consensus in the literature concerning a minimum value above which a factor pattern coefficient is considered salient, even cross-loadings as low as .10 may sometimes exert salient effects on factor modelling [10]. This is particularly important because cross-loadings were common and substantive in both the original development study [7], as well as in the test adaptation work in Sweden [8]. For instance, examination of the item-factor loadings in the EFA reported by Piran and colleagues [7] shows that 11 of the 34 EES items cross-loaded on more than one factor at ≥ .25 and a majority (21 items) cross-loaded at ≥ .20. Likewise, although Kling et al. [8] did not report a full factor loading matrix, they noted that nine items had cross-loadings greater than .25.

This, in turn, has implications for how scholars should cross-validate the EES in test adaptation studies. Specifically, CFA may have lower utility as a method of cross-validation because CFA only allows items to load on to their respective hypothesised latent factor, while forcing cross-loadings to be zero [11–13]; that is, CFA assumes “pure” factors where these factors only load onto their a priori latent factors. However, given that a majority of EES items show at least some overlap across conceptually related facets, failure to account for cross-loadings will likely lead to biased estimates of factor correlations and associations between these factors and other variables [12, 14, 15], as well as at least some model misspecification [16]. Even apparently well-fitting CFA models–such as Piran and colleagues’ [7] 6-factor model–could hide these misspecifications given their ability to absorb unmodelled cross-loadings through an inflation of factor correlations, without letting them impact model fit, e.g., [13, 17].

Instead of relying on CFA, or indeed an EFA-to-CFA approach, an alternative method for assessing factorial validity where cross-loadings are assumed to be substantive is exploratory structural equation modelling (ESEM) [12, 17–19]. ESEM was specifically designed to integrate the best elements of both EFA and CFA, including the relaxation of the zero cross-loadings requirement of CFA (a feature typically limited to EFA), while also allowing researchers to obtain goodness-of-fit statistics, residual correlations, standard error estimates, tests of measurement invariance, and tests of associations between latent constructs (i.e., features typically limited to CFA). Moreover, through its “target” rotation method (i.e., where cross-loadings are “targeted” to be close to zero while all main loadings are freely estimated) [12, 17], ESEM allows for a theory-driven approach that can be used in a confirmatory manner. As a result, ESEM provides a different approach to address the aforementioned limitations of CFA for the assessment of factorial validity [17].

Higher-order modelling.

There also remains scope to re-examine the utility of both higher-order and bifactor modelling of the EES. In their work, Piran and colleagues [7] in fact found that a 6-factor bifactor-CFA (henceforth B-CFA) model had superior fit to a 6-factor high-order CFA model (henceforth HO-CFA), although the former was de-valued on theoretical grounds (i.e., some items, including all RO items, did not load on the G-factor). While both bifactor and higher-order models postulate the co-existence of a general factor with specific factors, a unique feature of bifactor models is in their assumption that general factors have direct (rather than indirect) effects on the indicators (items). As such, all items are loaded on a specific factor and one general factor, with the variance of indicators partitioned into three sources, namely the specific factor, the general factor, and measurement error. In contrast, in higher-order models, associations between the indicators and the higher-order factor are indirect and associations between the indicators and the unique part of the higher-order factor are also mediated by the lower-order factors.

However, such an assumption in higher-order models is often viewed as empirically implausible and problematic [20, 21]. Indeed, emerging consensus suggests that bifactor models should be preferred over higher-order models unless there are strong conceptual and theoretical justifications for the latter [22]. With regards to the EES, Piran and colleagues’ [7] suggestion that the specific factors in their HO-CFA model can be conceptualised as subscale scores representing domains of the experience of embodiment is not in and of itself an adequate justification for favouring a higher-order model over a bifactor model. In fact, we suggest that it is in fact the bifactor model that provides greater theoretical and empirical plausibility [13]. Before any firm conclusions can be drawn, however, it should be noted that bifactor models can be constructed using both CFA and ESEM (henceforth B-ESEM). As such, and following on from the discussion above, there is a need to assess whether a B-ESEM model presents optimal fit for EES data.

Negatively worded items.

A third issue that requires further consideration is the inclusion of negatively worded items. Notably, all items on the BUA factor were negatively worded and four of the six EES factors included both positively and negatively worded items. Although negatively worded items can sometimes be useful for minimising acquiescence, affirmation, or agreement biases, they also require greater cognitive effort when responding [23]. This, in turn, often leads to methods effects that result in spurious covariances among items. Such method effects can be viewed as “noise” variance that should be controlled in analyses or modelled as correlated uniqueness (CU) among the indicators [24]. Previous studies, however, have not controlled for negatively worded items in the EES, which may have resulted in biased parameter estimates.

Differential item functioning.

Finally, much more can also be done to better understand whether respondent characteristics affect the probability of endorsing a given item on the EES (i.e., whether EES items are affected by differential item functioning or DIF). Two characteristics that are worthy of investigation in this regard are age and body mass index (BMI). In terms of the former, embodiment itself is a lived process that is likely to be affected by the physical state of one’s corporeal body, which in turn means that experiences of embodiment may be affected by age [25]. In terms of the latter, it is possible that navigating societies that objectify women based on BMI may affect experiences of embodiment. In Swedish women, for instance, Kling and colleagues [8] reported a negative association between EES scores and women’s BMI. In light of these issues, a fuller understanding of the role of age and BMI on item responses to the EES would be useful.

Experience of embodiment.

Participants were asked to complete a novel Greek translation of the 34-item EES [7], with items rated on a 5-point scale ranging from 1 (strongly disagree; Greek: διαφωνώ απόλυτα) to 5 (strongly agree; Greek: συμφωνώ απόλυτα). To prepare a Greek translation of the EES, we followed the 5-step procedure recommended by Beaton et al. [30]. Specifically, two translators–one informed, and one uninformed–first independently forward-translated the EES instructions, items, and response options from English to Greek. Next, the two translations were examined by a third, independent translator who resolved any discrepancies and produce a synthesised translation. Third, the synthesised translation was then back-translated by two translators naïve to the EES back into English. Fourth, the forward- and back- translations were compared by an expert committee comprising all the translators, as well as all authors of the present study, who resolved any minor inconsistencies between versions. In the fifth and final stage, the translated EES was pre-tested in a sample of 16 women who broadly matched the target sample. Participants in the pre-test study provided qualitative feedback regarding their level of understanding, as well as suggestions for improvements to enhance comprehension (based on open-ended questions). This feedback was returned to the committee, who agreed that no further revisions were necessary. The EES items in English and Greek are reported in the Appendix within the (S1 Appendix).

Body appreciation.

Participants completed the 10-item from the Body Appreciation Scale-2 (BAS-2) [31]; Greek translation: [32]. Items were rated on a 5-point scale (1 = never, 5 = always). In the present study, McDonald’s ω for scores on this scale was .96 (95% CI = .95, .96).

Self-esteem.

Self-esteem was assessed using the 10-item Rosenberg Self-Esteem Scale (RSES) [33]; Greek translation: [34]. Items were rated on a 4-point scale (1 = strongly disagree, 4 = strongly agree. In the present study, McDonald’s ω for RSES scores was .91 (95% CI = .90, .92).

Life satisfaction.

Participants completed the 5-item Satisfaction with Life Scale (SLS) [35]; Greek translation: [36], which assesses one’s assessment of the quality of life on the basis of their own unique criteria. All items were rated on a 5-point scale ranging from 1 (strongly disagree) to 5 (strongly agree). In the present work, McDonald’s ω for scores on this scale was .90 (95% CI = .89, .91).

Eating restriction.

To measure eating restriction, we used the 5-item Restraint subscale of the Eating Disorder Questionnaire (EDE-Q) [37]; Greek translation: [38]. All items were rated on a 7-point scale, ranging from 0 (no days) to 6 (every day). In the present work, McDonald’s ω for scores on this subscale was .85 (95% CI = .83, .86).

Perfectionism.

To measure perfectionism, we used the High Standards subscale of the Almost Perfect Scale–Revised (APS-R) [39]; Greek translation: [40], a 7-item measure of adaptive perfectionism. All items were rated on a 7-point scale ranging from 1 (strongly disagree) to 7 (strongly agree). In the present work, McDonald’s ω for scores on this subscale was .86 (95% CI = .84, .87).

Internalisation of appearance ideals.

Participants were asked to complete the Internalisation-General subscale of the Sociocultural Attitudes Toward Appearance Questionnaire-3 (SATAQ-3) [41]; Greek translation: [42], a 9-item measure of the degree to which individuals internalise general appearance ideals. All items were rated on a 5-point scale (1 = strongly disagree, 5 = strongly agree). In the present study, McDonald’s ω for scores on this subscale was .94 (95% CI = .93, .95).

Demographics.

Participants were asked to provide their demographic details consisting of age, highest educational qualification, ethnicity, height, and weight. Height and weight data were used to compute BMI as kg/m².

Data treatment.

Data used in the present study are provided in S1 File. To examine the dimensionality of EES scores, our aim was to initially compare the utility of CFA and ESEM models, before cross-validating these results. To ensure adequate subsample sizes for both sets of analyses, the total dataset was first split using a computer-generated random seed, resulting in two split-half subsamples (n = 467 and 466, respectively). There were no significant differences between the two split-halves in terms of age and BMI, nor were there significant differences in the distribution of ethnic groups and educational qualifications (results available from the corresponding author). There were no missing data in our dataset.

First split-half subsample.

CFA and ESEM representations of the EES. In the first split-half subsample, all analyses were conducted using Mplus 8.8’s [43] robust weighted least squares estimator with mean and variance adjusted statistics (WLSMV). In a first step, the a priori 6-factor model of the EES was estimated using CFA and ESEM. In the CFA model, EES ratings were respectively explained by six correlated latent factors without cross-loadings. In the ESEM solution, all cross-loadings were freely estimated using a confirmatory oblique target rotation procedure, allowing us to rely on an a priori specification of the main indicators of each factor and “targeting” all cross-loadings to be as close to zero as possible [14, 44]. Additionally, in order to control for the methodological artefact introduced by the negatively worded EES items [24, 45], CFA and ESEM models were also separately estimated with CU between the following items: #21 and 31 (from the Agency and Functionality factor); #28 and 30 (from the Experience and Expression of Sexual Desire factor); #15, 16, 18 and 23 (from the Attuned Self-Care factor); and #19 (from the Resisting Objectification factor). Composite reliability of EES latent factors was estimated using McDonald’s omega (ω) [46].

Model fit. Model fit was examined using the following fit indices [47–49]: the root mean square error of approximation (RMSEA) and its 90% CI (values ≤ .08 indicate acceptable fit; ≤ .06 indicates good fit), the Tucker-Lewis index (TLI; values ≥ .90 indicate acceptable fit and > .95 indicate good fit), and the comparative fit index (CFI; values ≥ .90 indicate acceptable fit and > .95 indicate good fit). However, as highlighted by Morin et al. [13, 17], goodness-of-fit assessment is insufficient to guide model selection when contrasting ESEM and CFA. Therefore, Morin et al. [13, 17] also recommended carefully examining parameter estimates (i.e., loadings, cross-loadings, latent correlations, composite reliability) from ESEM and CFA models. Observation of reduced factor correlations in the ESEM solution coupled with generally well-defined factors presents additional evidence for the superiority of the ESEM solution over a similarly fitting CFA solution [13, 17].

Ethnic invariance. The model providing the optimal representation of the data (CFA or ESEM, with and without CU) was then retained for tests of measurement invariance across ethnicity in the following sequence [50]: (i) configural invariance; (ii) weak invariance (loadings); (iii) strong invariance (thresholds); (iv) strict invariance (uniquenesses); (v) invariance of CU (if the model with CU is retained); (vi) invariance of the latent variances/covariances; and (vii) invariance of latent mean factors. Model comparisons (i.e., with each model contrasted to the previous one) relied on changes (Δ) in CFI, TLI, and RMSEA. Invariance was supported when ΔCFI and ΔTLI were ≤ .01, and ΔRMSEA was ≤.015 [51, 52].

Differential item functioning and latent mean differences as a function of age and BMI. In a third step, a hybrid multiple indicators multiple causes (MIMIC) multiple-group model [53–55] was used to examine: (a) differential item functioning (DIF), that is, direct associations between the predictors (age and BMI) and the EES item responses over-and-above the association between the predictors and the EES latent factors; (b) the associations between predictors (age and BMI) and EES latent factors; and (c) the equivalence of these associations across the two ethnic groups (Greek and Greek-Cypriot). These models were developed from the most invariant multiple-group model identified in the ethnic invariance test, to which the age and BMI were included. Specifically, hybrid MIMIC models were estimated in the following sequence [18, 19]: (a) null effects model (paths from the predictors to the EES latent factors and item responses were constrained to be zero); (b) saturated model (paths from the predictors to the EES item responses were freely estimated, while paths from the predictors to the EES latent factors were constrained to be zero); and (c) factors only model (paths from the predictors to the EES latent factors were freely estimated, while paths from the predictors to the EES item responses were constrained to be zero). To ease interpretations, age and BMI were standardised prior to analyses. A substantial improvement in model fit (ΔCFIs-ΔTLIs > .01 and ΔRMSEAs >. 015 [55], in the factors-only and saturated models relative to the null effects model provides support for an association between predictors and the EES item responses. Additionally, improvement in model fit for the saturated model relative to the factors-only model provides support for DIF. These models were studied with all associations freely estimated across the Greek and Greek-Cypriot subsamples. Then, the most appropriate model was retained and compared to an alternative model in which all associations were constrained to be equal across Greek and Greek-Cypriot respondents.

Second split-half subsample.

Higher-order and bifactor representations of the EES. In the second split-half subsample, all analyses were conducted using Mplus 8.8’s [43] robust WLSMV. In a first step, the a priori 6-factor model of the EES was estimated using CFA and ESEM as described in 2.4.2.1. Higher-order representation (HO-CFA and HO-ESEM) of the EES was estimated using six a priori first-order factors (PBCC, BUA, AF, EESD, ASC, and RO) and one second-order factor (HO of experience of embodiment). The HO-ESEM was estimated using ESEM-within-CFA as recommended by Morin and Asparouhov [56]. Bifactor representation of the EES (B-CFA and B-ESEM) comprised one more factor than their CFA and ESEM counterparts. In these models, all factors were specified as orthogonal [13, 17] and all items had a main loading on both a global factor (G-factor of experience of embodiment) and on their six specific factors (S-factors: PBCC, BUA, AF, EESD, ASC and RO). Additionally, based on results obtained in the first split-half subsample, models were estimated with or without CU. Finally, McDonald’s ω [46] was used to estimate composite reliability of EES latent factors.

Model fit. Fit indices used to identify the optimal model were identical to those reported above. As highlighted by Morin et al. [13, 17] goodness-of-fit assessment is insufficient to guide model selection when contrasting CFA, ESEM, HO-CFA, HO-ESEM, B-CFA, and B-ESEM solutions. Instead, Morin et al. [13, 17] recommended a careful examination of parameter estimates (i.e., loadings, cross-loadings, latent correlations, composite reliability) from the various models. This comparison begins with a comparison of the CFA and ESEM models, where the observation of reduced factor correlations in the ESEM solution coupled with generally well-defined factors provides evidence in favour of the ESEM solution over a similarly fitting CFA solution [13, 17]. Next, the retained model should be contrasted to its higher-order and bifactor counterparts. In this second comparison, the observation of a well-defined G-factor coupled with at least a subset of well-defined S-factors supports the superiority of the bifactor solution over a similarly fitting first-order or higher-order solution [13, 17].

Ethnic invariance, DIF, and latent mean differences as function of age and BMI. These analyses were performed with the model providing the most optimal representation of the data (CFA, B-CFA, HO-CFA, ESEM, HO-ESEM, and B-ESEM), using the same strategies as those reported in Section 2.4.2.3.

Construct validity. Construct validity was examined in the total sample using a structural equation model (SEM) in which the EES factor structure was estimated based on model providing optimal representation of the data. In this model, the EES latent factors and the observed scores of convergent or concurent measures were all correlated. Correlations values ≤ .10 were considered weak, ~ .30 moderate, and ~ .50 strong [57].

CFA and ESEM representations of the EES.

Goodness-of-fit indices of all measurement models are reported in Table 1. The 6-factor CFA solution had poor fit to the data (TLI < .90 and RMSEA > .08), whereas the 6-factor CFA-CU reached acceptability for CFI and TLI (> .90), but not for RMSEA (> .08). Both ESEM models (with and without CU) showed substantial improvement in fit relative to their CFA counterparts (ESEM: ΔCFI = +.051, ΔTLI = +.039, ΔRMSEA = -.020; ESEM-CU: ΔCFI = +.045, ΔTLI = +.037, ΔRMSEA = -.020). Additionally, the 6-factor ESEM-CU solution had an improved level of fit (ΔCFI = +.008, ΔTLI = +.007, ΔRMSEA = -.004) relative to its non-CU counterpart. Although these results lend preliminary support to the ESEM-CU solution relative to the CFA-CU solution, we followed Morin et al.’s [13, 17] suggestions, turning our attention to the parameter estimates from these solutions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Goodness-of-fit statistics for the Experience of Embodiment Scale (EES).

https://doi.org/10.1371/journal.pone.0303268.t001

Parameter estimates for the 6-factor CFA-CU and 6-factor ESEM-CU solutions are reported in S1 and S2 Tables¹ in S1 File. In the 6-factor CFA-CU solution, factor loadings were all reasonably high (PBCC: M_λ = .854; BUA: M_λ = .680; AF: M_λ = .798; EES: M_λ = .833; ASC: M_λ = .712; RO: M_λ = .637), except for Item #19 on the RO factor. These factors presented acceptable coefficients of composite reliability (ω = .705 to .943, M_ω = .869) and their latent correlations remained high (r = .354 to .867; M_r = .593).

In the 6-factor ESEM-CU solution, factor loadings were generally acceptable (PBCC: M_λ = .632; BUA: M_λ = .498; AF: M_λ = .718; EES: M_λ = .730; ASC: M_λ = .564; RO: M_λ = .608) and accompanied by reasonably small cross-loadings (M_λ = .107), with the exception of Items #3, 16, 19, 21, and 22. Item #3 from the BUA factor was more strongly associated with the PBCC factor; Item #16 from the ASC factor presented a similar pattern of associations with the BUA factor; Item #19 from the RO factor was more highly associated with the BUA factor; Item #21 from the AF scale presented a similar pattern of associations with the ASC factor; and Item #22 from the ASC factor presented a similar pattern of associations with the AF factor. The composite reliability coefficients of the six factors were adequate (ω = .718 to .923; M_ω = .847). In contrast to the 6-factor CFA-CU solution, the latent factor correlations between scales were substantially reduced (r = .221 to .558; M_r = .352), thus supporting their distinguishability (see S2 Table in S1 File).

Ethnic invariance.

Goodness-of-fit statistics associated with the ESEM-CU solutions estimated separately in the Greek and Greek-Cypriot subsamples are presented in Table 1 (Models 2–1 and 2–2). These results revealed acceptable fit indices for all models (CFI and TLI > .90 or > .95; RMSEA ≤ .08). The goodness-of-fit statistics associated with the tests of measurement invariance conducted for the ESEM-CU (Models 2–3 to 2–9) provided support for complete measurement invariance (i.e., loadings, thresholds, uniqueness, CU, variances/covariances, and means) of the ESEM-CU model.

DIF as function of age and BMI.

Table 1 presents the results from the MIMIC models. These models were estimated starting from the most invariant model of the ethnic invariance test (model 2–9: invariance of means). Results revealed a substantial improvement in model fit in the saturated (model 3–2) and factors-only models (model 3–3) relative to the null effects model (model 3–1). This suggests that the predictors (age and BMI) are associated with EES responses. Additionally, the factors-only model resulted in a similar level of model fit compared to the saturated model (ΔCFI = -.001, ΔTLI = +.001, ΔRMSEA = -.001), supporting a lack of DIF as a function of predictors. Finally, the last model (model 3–4), built from the retained factors-only model, showed that relations between the predictors and the latent factors could be considered as equivalent across Greek and Greek-Cypriot subsamples.

Table 2 presents results from the invariant factors-only model. First, these results showed that age significantly and positively predicted scores on the BUA, ASC, and RO factors. More specifically, older participants tended to present significantly higher scores on these subscales. Second, BMI significantly and negatively predicted scores on the PBCC, BUA, EESD, and ASC subscales. Thus, individuals with higher BMIs tended to present significantly lower values on these subscales.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Relations between the EES latent factors and the predictors in the first and second split-half subsamples.

https://doi.org/10.1371/journal.pone.0303268.t002

Higher-order and bifactor representations of the EES.

Goodness-of-fit indices of all measurement models are reported in Table 1. The 6-factor CFA-CU solution had poor fit to the data (TLI < .90 and RMSEA > .08). Although the B-CFA-CU model reached acceptability (CFI and TLI > .90; RMSEA = .08), the HO-CFA-CU reached acceptability for CFI and TLI (> .90), but not for RMSEA (> .08). In contrast, the ESEM-CU, HO-ESEM-CU, and B-ESEM-CU solutions resulted in an acceptable fit to the data and in a level of fit that was substantially improved relative to their CFA-CU (ΔCFI = +.051, ΔTLI = +.041, ΔRMSEA = -.021), HO-CFA-CU (ΔCFI = +.053, ΔTLI = +.046, ΔRMSEA = -.024), and B-CFA-CU counterparts (ΔCFI = +.051, ΔTLI = +.052, ΔRMSEA = -.031). The comparison of the first-order, higher-order, and bifactor ESEM solutions suggested the superiority of the B-ESEM-CU compared to the ESEM-CU (ΔCFI = +.019, ΔTLI = +.030, ΔRMSEA = -.019) and HO-ESEM-CU solutions (ΔCFI = +.016, ΔTLI = +.023, ΔRMSEA = -.015). However, despite the results seeming to favour the B-ESEM-CU solution, as recommended by Morin et al. [13, 17], the selection of the optimal solution should be based on a careful examination of the parameter estimates, composite reliability, and factor correlations. To this end, the CFA-CU and ESEM-CU solutions were first contrasted, and then the most optimal of these representations with its higher-order and bifactor counterparts.

The detailed parameter estimates from all solutions are reported in S3-S8 Tables² in S1 File. In the CFA-CU solution, the factor loadings of the EES were satisfactory (|λ| = .308-.939, M_λ = .701) and associated with acceptable composite reliability coefficients (ω = .716 to .936, M_ω = .868). However, the factor correlations remained high (r = .563). In contrast, this factor correlation was substantially reduced in the ESEM solution (r = .334), suggesting that the slight level of overlap could in fact be explained by the presence of cross-loadings (M_|λ| = .108). Nevertheless, although most cross-loadings remain negligible, the results also suggest that (a) Item #3 from the BUA subscale was more strongly associated with the PBCC subscale; (b) Item #17 from the PBCC subscale presented a nearly similar pattern of associations with the ASC subscale; (c) Item #19 from the RO subscale was more highly associated with the BUA subscale; and (d) Item #22 from the ASC subscale presented a nearly similar pattern of associations with the AF subscale. Finally, the ESEM solutions also resulted in generally well-defined (|λ| = .197-.894, M_λ = .618) and reliable latent factors (ω = .671 to .906, M_ω = .841).

Based on these above results, the ESEM solution was favoured, and contrasted with the HO-ESEM-CU and B-ESEM-CU solutions. In the HO-ESEM-CU solution, the first-order factor loadings of the EES were generally acceptable (|λ| = .252-.901, M_λ = .618) and accompanied by reasonably small cross-loadings (M_λ = .107) with the exception of Items #3, 8,17, 19, and 22. The composite reliability coefficients of the six first-order factors were adequate (ω = .691 to .905, M_ω = .844). Additionally, the second-order factor was well-defined by five of the subscales of the EES (|γ| = .532-.761, M_γ = .631), but weakly defined by the RO scale (γ = .293) suggesting that this scale might tap into a different construct. The composite reliability coefficient of the second-order factor was adequate (ω = .754).

The B-ESEM-CU solution resulted in a well-defined G-factor (|λ| = .053-.853, M_λ = .584) reflecting participants’ experience of positive embodiment, except for one item (#19) that remained substantial in two S-factors (BUA and RO). Additionally, the composite reliability of the G-factor was excellent (ω = .971) and higher than of the HO-ESEM-CU solution (ω = .754). Moreover, results revealed that: (a) the PBCC was less well defined (|λ| = .024-.674, M_λ = .282, ω = .692) and this could be attributed to four items (Items #8, 9, 11, and 17) that essentially serve to define the G-factor; (b) the BUA (|λ| = .221-.446, M_λ = .359, ω = .720) and ASC (|λ| = .218-.539, M_λ = .380, ω = .693) S-factors remained reasonably well-defined, but with weaker loadings values than those of the G-factor; (c) the AF (|λ| = .402-.738, M_λ = .535, ω = .846) and EESD (|λ| = .533-.704, M_λ = .598, ω = .834) S-factors remained well-defined and with similar loadings values to those of the G-factor; and (d) the RO (|λ| = .340-.729, M_λ = .592, ω = .710) S-factor remained well-defined and with higher loadings values than those of the G-factor. Finally, most cross-loadings were reduced (except for Items #3, 19, 22, and 34) in the B-ESEM-CU solution (M_|λ| = .083) relative to the ESEM-CU and HO-CFA-CU (M_|λ| = .107) solutions. Therefore, the present results support the B-ESEM-CU representation of the data (i.e., B-ESEM-CU resulted in an improved level of fit to the data compared to all other solutions; the G-factor of experience embodiment is a well-defined and the EES S-factors are reasonably well-defined).

Ethnic invariance.

Table 1 presents goodness-of-fit statistics associated with the B-ESEM-CU solutions estimated separately in the Greek and Greek-Cypriot subsamples (Models 5–1 and 5–2). These results revealed good fit indices for all models (CFI and TLI > .95; RMSEA ≤ .06). Additionally, the goodness-of-fit statistics associated with the tests of measurement invariance conducted for the B-ESEM-CU (Models 5–3 to 5–9) provided support for the complete measurement invariance (i.e., loadings, thresholds, uniqueness, CU, variances/covariances, and means) of the B-ESEM-CU solution.

DIF as function of age and BMI.

Results from the MIMIC models are reported in Table 1. These models were estimated starting from the most invariant model of the ethnic invariance test (model 5–9: invariance of means). Results revealed a substantial improvement in model fit in the saturated (model 6–2) and factors-only models (model 6–3) relative to the null effects model (model 6–1). This supports the idea that the predictors (age and BMI) are associated with EES responses. Additionally, the factors-only model resulted in a similar level of model fit compared to the saturated model (ΔCFI = -.001, ΔTLI = +.001, ΔRMSEA = .000), supporting a lack of DIF as a function of predictors. Finally, the last model (model 6–4), built from the retained factors-only model, showed that relations between the predictors and the latent factors could be considered equivalent across subsamples.

Table 2 presents result from the invariant factors-only model. First, these results showed that age significantly and positively predicted the G-factor and the BUA S-factor of the EES. Additionally, age significantly and negatively predicted the EESD S-factor of the EES. More specifically, older participants tended to present significantly higher values on the G-factor and the BUA S-factor, and significantly lower values on the EESD S-factor. Second, BMI significantly and negatively predicted the G-factor and the BUA S-factor of the EES. Additionally, BMI significantly and positively predicted the AF, ASC and RO S-factors of the EES. Thus, individuals with higher BMIs tended to present significantly lower values on the G-factor and the BUA S-factor, and higher values on the AF, ASC and RO S-factors.

Construct validity.

Goodness-of-fit statistics from the SEM including the EES latent factors (B-ESEM-CU model) and the other measures were acceptable: χ²(470) = 1570.347, CFI = .973, TLI = .955, RMSEA = .050 (90% CI = .047, .053). As illustrated in Table 3, the EES G-factor was significantly and (a) positively correlated with body appreciation, internalisation of appearance ideals, life satisfaction, perfectionism, and self-esteem; and (b) negatively correlated with eating restriction. Results for the PBCC and EESD S-factors showed significant and positive correlations with eating restriction. Additionally, results for the BUA S-factor also showed significant and (a) positive correlations with body appreciation, internalisation of appearance ideals, perfectionism and self-esteem; and (b) negative correlations with eating restriction and perfectionism. Moreover, results showed that the AF S-factor was significantly and positively correlated with all measures. In addition, the ASC S-factor was significantly and positively correlated with self-esteem, and negatively correlated with perfectionism. Finally, the RO S-factor was significantly and (a) positively correlated with body appreciation and internalisation of appearance ideals; and (b) negatively correlated with eating restriction, life satisfaction, and perfectionism.³

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Construct validity analyses from the bifactor exploratory structural equation modelling representation with correlated uniqueness of the EES in the overall sample.

https://doi.org/10.1371/journal.pone.0303268.t003