Confirmatory factor analysis (CFA) model-building

Candidate multifactor models for CFA of the ChEAT were identified from the literature. Anton et al. (2006) and Lynch & Eppers-Reynolds (2005) evaluated factor analysis studies that were published before 2005. Both Lynch and Eppers-Reynolds (2005) and Anton et al. (2006) concluded that neither the models based on studies of adults, including Garner et al. (1982), nor the models based on studies of children, including Smolak and Levine (1994), had a good fit to new data they collected from non-clinical samples of children; therefore, they proposed new models.

We searched PubMed, PsychInfo, and Scopus on August 30, 2018 to capture additional factor models from studies published in 2005 or later. We used the following search strategy: (TITLE-ABS-KEY ((“eating attitudes test” AND “children”)) AND (“factor analysis” OR “psychometric” OR “IRT”))). We found thirteen studies that included factor analysis results for children between the ages of 8 and 14 (Ambrosi-Randić & Pokrajac-Bulian, 2005; Chiba et al., 2016; Elizathe, Murawski, Arana, & Rutsztein, 2012; Escoto Ponce De León & Camacho Ruiz, 2008; Maïano, Morin, Lanfranchi, & Therme, 2013; McEnery, Fitzgerald, McNicholas, & Dooley, 2016; Pilecki, Kowal, Woronkowicz, Kryst, & Sobiecki, 2013; Ranzenhofer et al., 2008; Rojo-Moreno et al., 2011; Sancho, Asorey, Arija, & Canals, 2005; Senra, Seoane, Vilas, & Sánchez-Cao, 2007; Teixeira et al., 2012; Theuwis, Moens, & Braet, 2009). Study characteristics and principal components analysis (PCA) or factor analysis (FA) results from Anton et al. (2006), Lynch and Eppers-Reynolds (2005), and the thirteen new studies are summarized in eTable 1, alongside the seminal Garner et al. (1982) and Smolak and Levine (1994) studies of the (Ch)EAT.

Skugarevsky et al. (2014) conducted a PCA for the ChEAT on the same PROBIT sample; therefore, we did not consider those results in CFA model building. However, for the reasons explained below, we examined the CFA fit of the Skugarevsky model in sensitivity analyses.

Factor constructs from early studies were often supported by later studies, but the specific items loading on each factor varied considerably among studies (eTable 1). Occasionally, a study did not test one or more items. The vomiting items, 9 and 26, were the items most often excluded from analyses, either because authors believed they were not appropriate for the age group (Sancho et al., 2005; Senra et al., 2007) or participants uniformly scored zero (Elizathe et al., 2012; Pilecki et al., 2013). Therefore, the number of studies with an estimated factor loading for an item varied. We based selection of factor-item pairings on the following criteria: (1) in more than one third of studies that analyzed the item, the item loaded on the same (or a synonymous) factor with a factor loading >0.70, or (2) in more than two thirds of studies that analyzed the item, the item loaded on the same (or a synonymous) factor with a factor loading >0.30. According to those selection criteria, we found support for a 5-factor model, although support for the vomiting factor was tenuous. We labelled the factors as ‘weight preoccupation’ (WP), ‘dieting’ (D5), ‘food preoccupation’ (FP), ‘social pressure [to eat or gain weight]’ (SP), and ‘vomiting [and purging]’ (VP). Box 1 lists the factor label abbreviations and other labels that we considered to be synonyms.

Abbreviations

Three studies (Anton et al., 2006; Lynch & Eppers-Reynolds, 2005; Maïano et al., 2013) proposed models that grouped Garner’s dieting items into three distinct factors instead of two, resulting in six factors. Those studies used the full 6-point scale for each item, which may be more sensitive to variation in relatively healthy children, and an adequate sample size for the potential number of model parameters, which is large given the ChEAT’s 26 items. We tried to use their results to identify a 6-factor model, but only the factor-item pairings for WP were similar across the studies; the remaining dieting-related items (those that did not load onto WP) did not load onto synonymous factors across the three studies. Moreover, none of the other twelve studies replicated the extra dieting-related factor. Therefore, we compared only a 5-factor model to a single-factor model as a latent-trait analog to the univariate summed ChEAT score and to the original 3-factor model (Garner et al., 1982) (using the 24 available items). Table 1 shows the factor-item pairings and the item wordings.