The Salzburg Emotional Eating Scale was developed by Meule et al. (39) to measure emotional eating behaviors by examining different food responses in possible positive and negative emotional states. On the basis of various scales developed in research focusing on emotional eating, such as the Emotional Eating subscale developed by Van Strien et al. (40), the Emotional Eating Scale by Arnow et al. (3), and the Emotional Overeating Scale developed by Masheb and Grilo (17), it was observed that these scales primarily investigated negative emotions (40, 3, 17). In contrast, the Positive and Negative Emotional Eating Scale encompasses both positive and negative emotions but has been shown to be reliable and valid only for female participants (45). The Emotional Appetite Questionnaire, which covers positive and negative emotional states and measures increased and decreased eating tendencies in response to emotions, was not developed based on experimental data during the scale development phase (46). The significance of the SEES lies in its detailed delineation of specific emotions and its ability to differentiate between overeating and undereating based on emotions, thus representing an expanded version of previous scales (39).
In Turkey, scales examining emotional eating behaviors include the Emotional Appetite Questionnaire, which covers both positive and negative emotions. However, the Turkish validation analyses for this scale have not been completed (47, 34). It is believed that the adaptation of the Salzburg Emotional Eating Scale to the Turkish population could offer a broader perspective for studies focused on emotional eating.
In scale adaptation studies, sample sizes of 100 participants are considered weak for validity and reliability analyses, 200 are considered sufficient, and 300 are considered good (48–50). The present study was conducted using data obtained from 303 participants, including students, academics, and staff at Namık Kemal University in Tekirdağ. Women constituted 89.1% of the participants. In Meule (39) and Ghafouri et al.'s (51) studies, the majority of participants were also women (89.1%, 82.9%, 74.4%; 90%), indicating a similar sex distribution across all three studies. The disproportionate distribution of gender, with women being the majority in the present study, is presumed to potentially impact the study results, similar to the original study.
For the validity analysis of the SEES, structural validity analysis methods, including content and surface validity, were used. During the translation phase of the SEES, as Coster and Mancini (52) and Bayık (53) suggested, translation was conducted by two commissions consisting of 6 and 4 expert individuals, respectively (52, 53). Subsequently, the Davis method was used to analyse content validity. According to Davis (41), an assessment by at least 2 and up to 20 experts is needed; hence, the evaluation was carried out by 6 expert individuals, revealing KMO values indicating suitability for content validity (> 0.80).
For structural validity, both exploratory and confirmatory factor analyses were employed. As Suhr (54) emphasized, the objective of exploratory factor analysis is to uncover and explore the factor structure underlying the statements representing the variables of an adapted scale. CFA aims to verify the adequacy of the original factor structure. These methods are frequently preferred for testing and validating models (54, 55).
Considering the criteria outlined by Beavers et al. (56), the KMO value (0.89) in factor analysis indicated a very good sample size for conducting factor analysis. Additionally, Bartlett's sphericity test (p < 0.001) confirmed sufficient relationships among the variables for factor analysis. Principal component analysis and scree plot methods were used in exploratory factor analysis to evaluate the 'number of factors or components' and 'factor loading of variables' to determine the structural characteristics of the variables (57). The EFA revealed that 20 items with factor loading values ranging from 0.393 to 0.931 formed four subscales. The study confirmed that the four subscales aligned with the adapted scale, maintaining compatibility with each other (39).
In various studies, a criterion value of 0.30 has been considered adequate for factor loading (58, 59, 60). In the present study, the factor loading of each item was > 0.30. A high factor loading indicated the potential presence of the item in the respective subscale. Exploratory factor analysis revealed that the 20 items with factor loadings ranging from 0.393 to 0.931 formed four subscales. The current study demonstrated that the scale consisted of four subscales, and the items constituting each subscale were compatible with the adapted scale (SEES) as per the original study (39).
Another analysis of construct validity was CFA. A confirmatory factor analysis was also conducted to assess construct validity. It has been reported that when ML and GLS methods are used for estimating factor loads in data that do not exhibit a normal distribution, they tend to inflate model fit indices and decrease standard error rates (61). Therefore, the method chosen for nonnormally distributed data is critical (43). ADF is an analysis method used when the data do not follow a normal distribution. However, this approach makes few assumptions about the distribution of data. Moreover, it requires a minimum sample size of n: 200 for 'simple' models and an extremely large sample size (n > 5,000) for 'complex' models (62). Additionally, ADF is prone to weak estimations when the model is misspecified due to the influence of sample size (63). The Bollen–Stine Bootstrap method is a second potential solution provided for multivariate nonnormally distributed data (64). Therefore, in the present study, the Bollen–Stine Bootstrap method was used for DFA.
In evaluations of fit indices derived from Bollen–Stine Bootstrap analysis, it has been stated that Bollen–Stine Bootstrap analysis can provide strong evidence by avoiding biases in estimation values (62, 65). Kaya and Çolakoğlu resorted to this method in their validity analysis (66). Yaman (67) mentioned that the application of the bootstrap procedure enhances the parametric statistical fit of the CFA (Yaman, 2016). In the present study, ML fit indices were evaluated within the Bollen–Stine Bootstrap framework. When examining model fit indices, the χ2 value for the hypothesized model, especially with a large sample size, typically indicates the rejection of the model. A CFI value of 0.96 suggests that this model represents a marginally good fit. The other model fit indices, χ2/sd at 2.46, RMSEA at 0.074, NFI at 0.94, TLI at 0.95, and IFI at 0.96, are within acceptable ranges, while the GFI at 0.87 and AGFI at 0.83 are very close to acceptable ranges. These data indicate that the scale fits well into the model. Meule et al. (2018) stated that within the context of confirmatory factor analysis, CFI = 0.917–0.932 and RMSEA = 0.051–0.073 are considered acceptable (Meule et al., 2018). Ghafouri et al.'s (51) study reported the model fit indices as CMIND/DF = 7.58, GFI = 0.91, CFI = 0.90, TLI = 0.89, and RMSEA = 0.061 (51). Similar to the current study, it was observed that some fit indices are very close to an acceptable range.
Reliability analyses revealed that the Cronbach's α coefficients of the SEES-TR subscales of happiness, sadness, anger, and anxiety ranged from 0.913 to 0.942. The Cronbach's α coefficient of the total scale was determined to be 0.924. In scale adaptation studies, a Cronbach's α reliability coefficient > 0.80 indicates good reliability, while a value > 0.90 indicates excellent reliability (68). A total scale and subdimensions > 0.90 suggest that the SEES-TR has excellent reliability. According to McDonald's ω coefficient, which is a more general form of Cronbach's α, there is no optimal reliability measure, and it has been reported that the omega (ω) coefficient should be indicated for reliability estimation instead of α (69, 70). Therefore, both reliability analysis methods were used in the current study. The omega coefficients of the total SEES-TR and the happiness, sadness, anger, and anxiety subscales were found to be 0.917–0.943. Watkins et al. (2017) mentioned that the omega coefficient should be above 0.75 (71), indicating that the current study's omega reliability coefficient is reliable. In a study by Meule et al. (2018), the Cronbach's α coefficient for the subdimensions of the scale ranged from α = 0.732–0.871. In the present study, the reliability coefficients of the total scale and subdimensions indicate that the reliability is consistent with that of the original scale (39).
Another reliability analysis, the stability of the scale over time, involved readministering the scale to a subset of participants after 3 weeks. There was no statistically significant difference, and the reliability coefficient between the two test results was calculated. For the sadness and anger subscales of the study, the reliability coefficients were found to be 0.777 and 0.825, respectively, which are above the reported sufficient reliability coefficient (72). The reliability coefficient for the anxiety subscale was 0.698, which is very close to the 0.70 adequacy level, while the reliability coefficient for the happiness subscale was 0.490. Ghafouri et al. (2021) noted that the lack of a retest analysis was a limitation of their study and emphasized the need for this analysis to be conducted in the future for scale (51). In the current study, the term 'reliable scale' was reinforced through the stability method.
In this study, the parallel form method, another reliability analysis method, was also used by sampling different items that could represent the same behavioral patterns to create two equivalent forms. These forms were simultaneously administered, and the DEBQ emotional subscale was used for the parallel-form method in accordance with the original study of the scale (39). The DEBQ emotional subscale showed a positive correlation with the sadness, anger, and anxiety subscales, demonstrating consistency with the SEES. A negative correlation between the SEES happiness subscale and the DEBQ emotional subscale was suggested (39); similarly, in the current study, it was found that the happiness subscale correlated negatively with the DEBQ emotional subscale but not significantly.
Although the SEES-TR could allow for a more detailed analysis of emotional effects on eating behavior, it is still potentially biased based on self-reports. Specifically, the scale requires participants to have significant awareness of fluctuations in their daily emotions and their impact on food intake. Responses to the scale will be shaped by this awareness. A Turkish validation and reliability study of the SEES was conducted on individuals with higher education levels. Hence, to enhance the applicability of the study to all segments of society, future studies involving individuals with lower levels of education are recommended. For studies focusing on the impact of emotions on eating behavior, it is suggested that a scale be used to explore a broader range of emotions and changes in food intake.