We conducted a validation study of the EAT-26 between September 2019 and February 2020, in secondary education students enrolled in 5 schools in the Valencian Community, Spain. Four schools were in the province of Valencia and one in the province of Alicante. Of the 4 schools in the province of Valencia, 2 were in the city of Valencia and the other 2 in rural towns near the city with fewer than 30 000 inhabitants; the school in the province of Alicante was in a rural setting inland. The socioeconomic level of the areas where the study was conducted varied from low to high. All participating schools were part of the public education system, four were state schools and one a publicly funded private school. When it came to the educational programmes offered, 2 offered compulsory secondary education (ESO, escuela secundaria obligatoria), vocational education (FP, formación profesional) and non-compulsory secondary education (bachillerato), 2 offered ESO and bachillerato studies, and 1 offered ESO and VE programmes.21

Participants were recruited by convenience sampling. The sample size was calculated according to the most recent statistical recommendations, which yielded a minimum sample size of 500 participants.19 We excluded incomplete cases. The inclusion criteria were: (a) student aged 12 to 18 years (b) current enrolment in the school. We excluded students with known mental health disorders diagnosed by a physician.

• -

  Sociodemographic and clinical variables: age (in years, continuous variable); sex (male, female); nationality (Spanish, other); weight status based on Body Mass Index (BMI) (underweight, <18.5; normal weight, 18.5–24.9; overweight, 25–29.9; obesity >30).22

• -

  Eating Attitudes Test-26 (EAT-26): originally developed by Garner et al.7 and later validated in Spain by Gandarillas et al.12 It is a self-report measure to assess DEABs. It consists of 26 items in 3 dimensions: (a) diet, (b) bulimia and (c) oral control. The items are rated on a 6-point Likert scale (1 = never to 6 = always) that is subsequently transformed into a 4-point Likert scale (0 = never to 3 = always). A score greater than 20 is indicative of DEABs.10 The scale has exhibited a moderate to adequate internal consistency internationally (>0.70).7,10

• -

  SCOFF questionnaire: originally developed by Morgan et al.23 and subsequently validated in Spain by Rueda et al.24 It is a unidimensional instrument comprised of 5 items that assess different DEABs. Participants answer on a dichotomous scale (yes/no). Answering “yes” to 2 or more questions constitutes a positive result. In studies conducted in different countries, the internal consistency of the SCOFF questionnaire has ranged between 0.4 and 0.7, values considered sufficient for screening instruments.23,24

All participants who agreed to participate in the study were given a form to take home so that their parents or legal guardians could provide informed consent. The form included detailed information about the name, objectives and relevance of the study and about the confidentiality of the data. Participants in the study received another document containing all the variables under study.

We divided total number of observations (N = 718) into 2 samples, sample 1 (n = 359) and sample 2 (n = 359). All the analyses were carried out with the free software R (version 6.3). We assessed the performance of the instrument by calculating the symmetry, kurtosis and floor and ceiling effects. Asymmetry and kurtosis coefficients of less than −1.5 or more than 1.5, violate the assumption of a normal distribution.19 We considered that there was a floor or ceiling effect if more than 15% of the responses of participants were in the lowest or highest answer categories, respectively, which suggests that the measure cannot discriminate between subjects at either end of the scale.25 Applying the criteria proposed by Rhemtulla et al.,26 we treated the data as ordinal. We analysed the structure of the instrument using exploratory factor analysis (EFA) in sample 1. To determine whether the data were suited for EFA, we used the Kaiser-Meyer-Olkin (KMO) (suitable if KMO ≥0.7027), Bartlett’s test of sphericity (suitable if P < .0528) and the coefficient of determination (suitable if near zero). To determine the number of factors in the instrument, we used Horn’s parallel analysis.19,29 The EFA was performed with the psych package using the unweighted least-squares (ULS) method, recommended for categorical data without a normal distribution,30 and Promax oblique rotation using dimensions highly correlated to the construct being measured.19 The criteria used for the selection and finetuning of the items were: (a) saturation ≥0.40 and (b) elimination of Heywood cases (saturation ≥1).30 Subsequently, we assessed the structure obtained in the EFA with confirmatory factor analysis (CFA) in sample 2 with the lavaan package using the robust weighted least squares mean and variance (WLSMV) estimate, recommended for ordinal data.25 We adjusted the fit of the model using the comparative fit index (CFI), the Tucker-Lewis index (TLI) and the root mean square error of approximation (RMSEA), considering the following variables adequate: CFI > 0.90, TLI > 0.90 and RMSEA < 0.06.20,31,32 We proposed 4 statistical adjustments: (1) tau-equivalence, (2) congeneric model, (3) correlated errors (modification index >35 000) and (4) second-order model. We rejected models with Heywood cases, correlated errors that were incoherent at the theoretical level and negative variances.31 We assessed internal consistency through the calculation of the ordinal α coefficient, which offers a more precise estimate for scales with categorical outcomes. An α coefficient of 0.70 or greater indicates adequate reliability.20,33 We assessed convergent validity by calculating the product-moment correlation between the obtained factors and items of the EAT and the SCOFF questionnaire.34 We expected significant correlations between 0.2 and 0.7.20 Measurement invariance was assessed at 4 levels: (a) configurational invariance, b) metric invariance, (c) strong invariance and (d) strict invariance. We considered a change in the CFI (ΔCFI) of 0.010 or less and a change in the RMSEA (ΔRMSEA) of 0.015 or less insignificant in the comparison of different levels of measurement invariance.35,36

All participants agreed to participate in the study on a voluntary basis. The study was approved by the Ethics Committee (file no. 2016/0794). We safeguarded privacy and data confidentiality in adherence with regulation (EU) 2016/679 of the European Parliament and Council of April 27, 2016.

The recruited sample included 725 participants, and, after excluding incomplete questionnaires (n = 7; 0.96%), the final sample included 718 participants (99.04%). We ought to mention that none of the students refused to participate. Of the 718 students included in the analysis, 39% were male and 61% female, with mean ages of 14.4 years (SD, 1.8) and 14.7 years (SD, 2.0), respectively. Most were Spanish, had a social support network and a normal weight (Table 1).

Table 2 presents the performance of the scores of the EAT-26 items. We found floor and ceiling effects, asymmetry and kurtosis, and therefore considered the data to be ordinal.

We conducted 3 consecutive Horn parallel analyses with the corresponding EFAs on the initial set of 26 items. The items to be eliminated in each round were selected according to the preestablished criteria. Each EFA yielded a KMO of 0.70 or greater, a P value of less than 0.05 in Bartlett’s test and a coefficient of determination near zero. Eventually, we eliminated items 2, 4, 5, 15 and 25 due to factor loadings of less than 0.4. The resulting KMO (0.79), Bartlett’s test of sphericity (degrees of freedom (df) = 210; P = .000) and the coefficient of determination of less than 0.001 indicated that the correlation matrix of the 21 selected items (EAT-21) and the 6 factors suggested by the parallel analysis were suitable for EFA. The variance explained by each rotated factor was: factor 1 (23.98%); factor 2 (11.53%); factor 3 (8.34%); factor 4 (7.52%); factor 5 (4.97%) and factor 6 (4.91%) (Table 3).

Continuing with the analysis of the content of the items, the factors obtained in the EFA and the studies conducted by Maïano et al.8 and McEnery et al.,9 we labelled the factors as: factor 1, “fear of getting fat”; factor 2, “eating-related control”; factor 3, “vomiting-purging behaviour”; factor 4, “social pressure to gain weight”; factor 5, “dieting behaviour” and factor 6, “food preoccupation”.

Table 4 presents the results of the CFA for the models that fulfilled the predefined criteria. The 6-factor congeneric model obtained in the EFA exhibited an excellent fit, with factor loadings ranging between 0.44 and 0.95 (Fig. 1). The results of the second-order model in the CFA de revealed an acceptable fit, albeit marginally inferior to that of the congeneric model. The factor loadings of the items of the second-order model ranged between 0.40 and 0.95 (Fig. 2).

Figure 1.

Congeneric model.

Factor 1 (fear of getting fat), Factor 2 (eating-related control), Factor 3 (vomiting-purging behaviour), Factor 4 (social pressure to gain weight), Factor 5 (dieting behaviour), Factor 6 (food preoccupation).

(0.4MB).

Figure 2.

Second-order model.

Factor 1 (fear of getting fat), Factor 2 (eating-related control), Factor 3 (vomiting-purging behaviour), Factor 4 (social pressure to gain weight), Factor 5 (dieting behaviour), Factor 6 (food preoccupation).

(0.23MB).

The internal consistence, estimated with the ordinal α, was 0.81 for factor 1 (fear of getting fat) 0.66 for factor 2 (eating-related control), 0.77 for factor 3 (vomiting-purging behaviour), 0.74 for factor 4 (social pressure to gain weight), 0.67 for factor 5 (dieting behaviour) and 0.51 for factor 6 (food preoccupation).

Table 5 presents the results of the analysis of the product-moment correlation between the scores of the resulting EAT-21 factors and the SCOFF questionnaire. The total score of the EAT-21 was positively and significantly correlated to the SCOFF questionnaire score (r = 0.483; P < .001).

Table 6 presents the measurement invariance results. They suggests that it is reasonable to assume measurement invariance across the sexes, as none of the comparisons revealed a change in fit greater than a ΔCFI ≤ 0.010 or an ΔRMSEA ≤ 0.015.