This is a cross-sectional study. Samples in which this study was conducted were part of a major research project that aims to explore eating behaviours and emotional-related difficulties in adult women. This major project was approved by the Scientific Committee of Psychology Faculty (official letter n° 014/2013) and also by the Research Committee of Pontifical Catholic University of Rio Grande do Sul (PUCRS), Brazil. All participants provided written informed consent after being briefed about the aims of the study and its voluntary and confidential nature.

Two independent samples were used in this study. The general population group was composed of women who were awaiting transportation or services in the Porto Alegre¿s Bus Station and in three citizens’ bureaus, respectively. The clinical group comprised women with overweight or obesity in medical treatment for weight loss in a public hospital in Porto Alegre. Inclusions criteria: a) female; b) age between 18 and 60 years; c) 5 or more years of education. Moreover, in order to be part of the clinical group, participants had to present a Body Mass Index (BMI) ≥ 25 (kg/m2).

Two independent researchers (both Brazilian and fluent in English) adapted the BI-AAQ to Brazilian-Portuguese and made a back-translation to English. Then, a committee of three experts in both the underlying construct of BI-AAQ and in transcultural validations was formed with the purpose of evaluating and deciding which adaptation of each item should integrate the preliminary version of the scale. Afterwards, an analysis of each item of the preliminary version was conducted, taking into consideration the clarity of language, the practical pertinence for the target culture and the theoretical relevance, according to a 5-point Likert scale (see Cassepp-Borges, Balbinotti, & Teodoro, 2010). Then the Content Validity Coefficient was calculated (see Hernández-Nieto, 2002) and all 12 items presented CVC ≥ 8. Finally, a pilot study was conducted in a sample of 22 Brazilian adults (being 45.5%, n= 10 females), with an average of 13.59 years of education (SD=3.32), and of 32.45 years of age (SD=10.22). Regarding marital status, the majority were single (62.8%; n=15) followed by married (31.8%, n=7). This pilot study showed that BI-AAQ was clearly understood, so the final version was consolidated and the data collection was initiated.

The statistical analyses were performed in the following samples.

Sample 1: composed of two groups (general population, n=682; and clinical group, n=330), which comprises 1,012 participants. Table 1 presents the socio-demographic characteristics of the total sample and of each group.

Sample 2: this was a subgroup from the clinical group of Sample 1 (n= 294) composed of participants who responded the following set of measures: BI-AAQ, SCS, BES and EDI-2. This subgroup presented a mean age of 41.87 years old (SD= 11.47), an average of 14.66 (SD= 3.54) years of education and of 32.38 (SD= 5.25) of BMI. Regarding marital status, 34% (n= 100) were single, 54.8% (n= 161) were married/cohabiting, 9.9% (n= 29) divorced and 1.4% (n= 4) were widowed. Additionally, the majority of participants were employed (70.1%, n= 206), followed by unemployed (18.7%, n= 55) and retired (11.2%, n= 33).The clinical and non-clinical groups from Sample 1 were used to perform the analyses aimed to confirm the BI-AAQ's factorial structure, its measurement invariance across groups as well as its ability to detect differences between them. Sample 2 was used to conduct the concurrent validity and mediational analyses.

Body Image Acceptance and Action Questionnaire (BI-AAQ; Sandoz et al., 2013). This is a 12-item 7-point Likert-like scale (from 1 never true to 7 always true) of psychological (in)flexibility specifically developed for assessing (in)flexible responding to body-related thoughts and feelings. BI-AAQ resulted from a pool of 46 items adapted from other general AAQ versions, which were submitted to an item-total correlation analysis, where items with item-total correlations <.30 were omitted (n=17). The remaining 29 items were submitted to a Principal Factor Analysis (PFA) and to a Parallel Analysis (PA). Those results revealed a single factor structure accounting for 34.4% of the variance, where 26 items had factor loading above .40. So, in order to produce a shorter length measure, 12 items with factor loadings above .60 were retained. Finally, the 12-items version was re-examined in an independent sample where the PFA and the PA were replicated. Results supported the stability of the one-factor structure, which accounted for 54% of the variance (Sandoz et al., 2013). Additionally, the BI-AAQ's Portuguese validation have also performed a PFA and its results indicated a one-factor solution accounting for 63.36% of the variance (Ferreira et al., 2011). This one-factor 12-items solution was confirmed through CFA analyses in the Portuguese and Persian studies (Ferreira et al., 2011; Izaadi et al., 2013). Internal consistency analysis revealed an excellent Cronbach's alpha (α=.92 and α=.93) in the original validation sub-studies.

Self-Compassion Scale (SCS; Neff, 2003; Brazilian version: Souza & Hutz, 2016) is a 26-item 5-point Likert self-report measure (from 1=almost never to 5=almost always) developed to assess trait levels of self-compassion. Although SCS was initially designed to grasp three components that interact with each other (self-kindness versus self-judgment, common humanity versus isolation and mindfulness versus over-identification), it can also be used as a two-factor structure measure: one factor that assess self-compassion attitude (sum of self-kindness, common humanity and mindfulness) and one factor of a self-criticism attitude (sum of self-judgement, isolation and over-identification). This two-factor structure has adequate model fit and good internal consistency (α=.91 for self-compassion and α=.89 for self-criticism) (Costa, Marôco, Pinto-Gouveia, Ferreira, & Castilho, 2015).

Eating Disorder Inventory (EDI-2; Garner, Olmsted, & Polivy, 1983). Composed by eight subscales represented by a total of 64-items. The respondents have to rate the frequency in which what is described in each item reflects their experience, according to a 6-point Likert scale (from Always to Never) through which they rate the extent to which the items reflect their experience. For the current study, it was only used the Drive for Thinness (DT) subscale. The Portuguese validation found an excellent internal consistency for DT subscale (α= .91; Machado, Gonçalves, Martins, & Soares, 2001), corroborating the findings from the original study which observed α=.85.

Binge Eating Scale (BES; Gormally, Black, Daston, & Rardin, 1982). This measure is composed of 62 statements (arranged in 16 items) used to evaluate the severity of binge eating symptoms. For each item participants have to decide which statement (between three or four alternatives) better applies to them. Then, a score from 0 to 3 is assigned for each marked statement and the total score is calculated. Higher scores indicate higher binge eating severity. BES has high internal consistency in the original study (Gormally et al., 1982) as well as the Brazilian version (α=.89; Freitas, Lopes, Appolinario, & Sichieri, 2002).

Descriptive statistics were assessed through SPSS statistics software (v.20; SPSS Inc., Chicago, IL), while the Confirmatory Factor Analysis (CFA) and Multigroup Analysis (MA) were performed using AMOS software (v.19, SPSS Inc., Chicago, IL). Outliers were assessed through the squared Mahalanobis Distance (MD2) considering values of p1 and p2<.05 (Marôco, 2010). Normality was assessed according to values of Skewness (Sk) and Kurtosis (Ku), where values of Sk>|3| and Ku>|10| indicate severe violations of normality (Kline, 2010). Cases with missing values were excluded since they were completely at random and less than 5% of cases (Tabachnick & Fidell, 2014), according to the results from the Missing Value Analysis (MVA) procedure, which is provided by SPSS. A Confirmatory Factor Analysis (CFA) was performed to test if the one-factorial structure proposed by the authors of BI-AAQ (Sandoz et al., 2013) had good adjustment to the data. A combination of goodness-of-fit indices was used as suggested by Kline (2010): Tucker-Lewis Index (TLI), Comparative Fit Index (CFI), Goodness of Fit Index (GFI), Incremental Fit Index (IFI, also known as BL89), Relative Fit Index (RFI) and Normed Fit Index (NFI). TLI, RFI and NFI can vary between 0 and 1 - and the closer to 1, the better the fit of the model. CFI, IFI and GFI values ≥ .90 indicate a good fit (Hu & Bentler, 1999). Finally, the local adjustment was assessed both by standardized factor weights and the individual reliability of the items, considering of λ ≥ .50 and R2 ≥ .25 (Hair, Black, Babin, & Anderson, 2014), respectively. Maximum Likelihood was used as the estimation method (since the BI-AAQ response rate have 7 points–see Rhemtulla, Brosseau-Liard, & Savalei, 2012). Moreover, a chi-square test of differences was calculated (as described in Marôco, 2010) to test whether the final model has better fit to the correlational structure of the items than the original one. Internal consistency was assessed by Cronbach's alpha (values of α>.70 indicates adequate reliability - Kline, 2000). A multi-group analysis was conducted to test the measurement invariance across clinical and non-clinical groups (through the comparison of the unconstrained model with a model in which factor loadings were constrained; Byrne, 2016). Furthermore, the BI-AAQ's ability to detect differences between groups was analysed through a Student's T Test. Cohen's d was calculated to verify effect size which can be small (]0.2; 0.49]), intermediate (]0.5; 0.79] or large (≥.80) (Cohen, 1988). Concurrent validity was assessed by Pearson's correlations between the BI-AAQ, BES, SCS and Drive for Thinness (EDI Subscale). Finally, the mediational study was conducted through a Path Analysis using Maximum Likelihood as the estimation method. A bootstrap method with 2,000 resamples was performed, as it is one of the most reliable and powerful method to evaluate the significance of the direct, indirect and total standardized effects (Marôco, 2010) providing greater statistical power and confidence interval precision regardless of sample size, effect size or level of statistical significance (Mallinckrodt, Abraham, Wei, & Russell, 2006). A 95% bias-corrected confidence interval was considered (Kline, 2010).

The initial model was specified with all items pertained to a single factor, as proposed by the authors of BI-AAQ (Sandoz et al., 2013), which demonstrated a poor fit (χ2(54) =793.08; p<.001; CFI=.91; GFI=.87; TLI=.89; IFI= .91; RFI= .885; NFI= .90). Regarding local adjustment, all items presented high values of factor weights and individual reliability (λ≥ .64 and R2≥ .41, respectively) with exception of item 6 which presented λ≥ .44 and R2≥ .20.

The global model fit would be improved through the correlation between several pairs of item's errors according to the modification indices. However, it was only theoretically justifiable the correlation of the following pairs of item's errors: 2&3 (disturbing internal experiences regarding body shape and/or weight), 2&5 (excessive worry about the body), 7&8 (necessity to feel better or in control about body-related difficulties and its consequences for one's life), and 9&10 (influence of dissatisfaction with weight on one's life). It is a common procedure to add paths to the model correlating errors of items from the same factor, as long as these correlations have theoretical reasons (i.e. similar content or formulation; Marôco, 2010). Thus, the model was modified through the addiction of the aforementioned paths.

Even though the modified model presented an increase in the global fit indices (χ2(50) =424.27; p<.001; CFI=.95; GFI=.93; TLI=.94; IFI= .95; RFI= .93; NFI= .95), the local adjustment of item 6 still presented factor weights and individual reliability bellow the expected (λ≥ .45 e R2≥ .20, respectively). Furthermore, according to the values of “alpha if item deleted”, all items are contributing for the internal consistency of this measure with the exception of item 6 (its exclusion would increase the Cronbach's alpha from α=.93 to α=.94). Finally, the communalities’ analysis indicate that all items presented values ≥.47 with the exception of item 6, which presented the lowest value (.23). Thus, taking into consideration all these data together, we decided to exclude item 6 from the model. Therefore, a final model with 11 items was specified maintaining the same error's correlations that were made in the last model (see Figure 1). The final model presented a good global fit and did not show problems in the local adjustment (λ≥ .62; R2≥ .39).

Figure 1.

Standardized factor weights of the 11 items in BI-AAQ on the final model (N=1,012). χ2(40)=390.439; p<.001; CFI=.957; GFI=.932; TLI=.941; IFI=.957; RFI= .935; NFI=.952.

(0.27MB).

Additionally, a chi-square difference test confirmed that the final model presented a significantly improvement when compared to the modified model (Δχ2(10)=33.835; p< .05). It is also worth mentioning that the final model presented lower levels of MECVI (.438) than the modified one (.476), showing that the final model presents better validity in the studied populations.

The final model presented good fit indices, high values of factor weights and individual items reliability (λ≥ .55 e R2≥ .31). Figure 2 presents the standardized factor weights and individual reliability of the items in each group.

Figure 2.

Standardized factor weights of the final model for the two groups in study (general population, n=682; and clinical group, n=330) χ2(90)=472.247; p<.001; CFI=.949; GFI= .920; TLI=.938; IFI= .949; RFI= .924; NFI= .938.

(0.44MB).

No differences were detected in regard to factor weights (Δχ2(1)=1.911; p= .167), showing measurement invariance.

BI-AAQ presented a Cronbach's alpha of α=.94, α=.95 and α=.93, in the total sample, clinical and non-clinical groups, respectively.

Significant differences between clinical and non-clinical groups were found (t(622)= -11.19; p<.001), being the scores of the clinical group higher than the ones in the general population group: means of 45.98 (SD= 17.48) and 33.08 (SD= 16.60), respectively. Cohen's d showed an observed value of .76.

Correlations between the BI-AAQ and binge eating severity, drive for thinness, self-criticism and self-compassion were presented in Table 2 together with the values of internal consistency (Cronbach's alpha) obtained for all variables in the current study.

A mediational study tested a theoretical model in which the role of body image psychological inflexibility mediated the effect of self-criticism and drive for thinness on binge eating severity (Figure 3). This was a fully saturated model (containing 10 parameters), so the model fit indices were neither tested nor reported since fully saturated models always produce a perfect adjustment to the data.

Figure 3.

The path model (Sample 2, n=294). Note: Standardized coefficients are presented and all paths are significant at the .001 level.

(0.17MB).

Results showed a significant standardized total effect of self-criticism and drive for thinness on binge eating behavior (Self-criticism: β=.36; 95% CI=[.27; .45]; p=.001; Drive for thinness: β=.33; 95% CI=[.25; .41]; p=.001), as well as a positive and significant standardized indirect effect (through body image psychological inflexibility) of both variables on binge eating (Self-criticism: β=.17; 95% CI=[.11; .23]; p=.001; Drive for thinness: β=.16; 95% CI=[.11; .22]; p=.001). The standardized direct effect of the mediator on the criterion variable was significant (β=.40; 95% CI=[.30; .51]; p=.001). Finally, the standardized direct effect of self-criticism and drive for thinness still significant in the presence of the mediator (Self-criticism: β=.19; 95% CI=[.09; .28]; p=.001; Drive for thinness: β=.16; 95% CI=[.07; .25]; p=.004), which indicates a partial and significant mediation. The model explained 42% of binge eating severity.