Cluster analysis
Based on the total scores of ED-15, DERS and UPPS-P negative urgency, three clusters were estimated by the hierarchical cluster analysis, which coincided with the optimal solution chosen by the K-means cluster analysis. The Silhouettes coefficient was used as a measure of the goodness of the final cluster solution and its mean value was .301 (min. -.086; max. .508), suggesting a fair fitting. Cluster 1 comprised 39.7% of the sample (n = 29), Cluster 2 also included 39.7% of the sample (n = 29) and Cluster 3 represented 20.5% of the participants (n = 15). The standardized scores for the three clusters are presented in Figure 1. Cluster 1 represented the participants with high ED-15 total scores relative to the sample mean, as well as the participants with lower than average total scores of DERS and UPPS-P negative urgency. Cluster 2 was characterized by participants with the highest scores in the three variables relative to the sample means. Cluster 3 represented the participants with the lowest scores in the three variables relative to the sample means.
Comparison among the Clusters
A series of one-way ANOVAs were conducted to test differences among the clusters in the main variables (Table 1). The results revealed significant differences among the three clusters in the total scores of ED-15, F (2, 70) = 22.18, p < .001, DERS, F (2, 32) = 63.22, p < .001, and UPPS-P negative urgency, F (2, 32) = 48.30, p < .001. In pairwise comparisons, the ED-15 total score in Cluster 1 was significantly higher than in Cluster 3 (mean difference = 2.09, p < .001), and the ED-15 total score in Cluster 2 was also significantly higher than in Cluster 3 (mean difference = 2.59, p <.001). No differences were found between the Clusters 1 and 2 in ED-15 total score (mean difference = -.50, p = .34). For DERS total score, Cluster 1 scored significantly lower than Cluster 2 (mean difference = -40.43, p < .001) and scored significantly higher than Cluster 3 (mean difference = 19.91, p = .04). Cluster 2 scored significantly higher than Cluster 3 (mean difference = 60.34, p < .001). For UPPS-P negative urgency, Cluster 1 also scored significantly lower than Cluster 2 (mean difference = -6.38, p < .001) and scored significantly higher than Cluster 3 (mean difference = 9.07, p = .04). Cluster 2 scored significantly higher than Cluster 3 (mean difference = 15.45, p < .001).
Additionally, the differences among the clusters in attitudes and eating behaviors, as well as in the dimensions of emotion dysregulation were analyzed (Table 1). Regarding eating attitudes, the MANOVA revealed a significant overall effect of the clusters on the subscales of the ED-15, F (4, 138) = 9.83, p < .001; Wilk’s λ = .61, partial η2 = .22. Univariate ANOVAs indicated that both weight and shape concerns, F (2, 70) = 21.29, p < .001, and eating concerns, F (2, 70) = 14.41, p < .001, were significantly different among clusters. Post-hoc comparisons suggested that weight and shape concerns in Cluster 1 were significantly higher than those in Cluster 3 (mean difference = 2.19, p < .001). Weight and shape concerns in Cluster 2 were also significantly higher than those in Cluster 3 (mean difference = 2.73, p <.001). Additionally, eating concerns in Cluster 1 were significantly higher than those in Cluster 3 (mean difference = 1.96, p < .001), as well as eating concerns in Cluster 2 were significantly higher than those in Cluster 3 (mean difference = 2.41, p <.001). No differences were found between the Clusters 1 and 2 in weight and shape concerns (mean difference = -.54, p = .35) or eating concerns (mean difference = -.46, p = .54).
For eating behaviors, Kruskal-Wallis tests showed significant differences among the clusters in binge-eating episodes, χ2 (2) = 9.28, p = .010, vomiting episodes, χ2 (2) = 10.40, p = .006, eating restraint days, χ2 (2) = 8.74, p = .013, and excessive exercise days, χ2 (2) = 8.47, p = .015. No significant differences were found among the clusters on laxative misuse days, χ2 (2) = 4.00, p = .135. Mann-Whitney tests with Bonferroni correction suggested that participants from Cluster 1 scored significantly higher than those from Cluster 3 in binge-eating episodes, U = 144.50, p = .021, vomiting episodes, U = 120.00, p = .003, and restraint days, U = 135.50, p = .026. For excessive exercise days, no differences were found between the Clusters 1 and 3, U = 160.00, p = .077. Participants from Cluster 2 scored significantly higher than those from Cluster 3 in binge-eating episodes, U = 112.00, p = .002, vomiting episodes, U = 112.50, p = .002, eating restraint days, U = 108.00, p = .004, and excessive exercise days, U = 117.00, p = .005. No differences were found between the Clusters 1 and 2 in all eating behaviors.
With respect to the dimensions of emotion dysregulation, the MANOVA revealed a significant overall effect of the clusters on the six subscales of the DERS, F (12, 130) = 10.65, p < .001; Wilk’s λ = .25, partial η2 = .50. Except for awareness, F (2, 70) = 2.03, p = .139, univariate ANOVAs indicated that strategies, F (2, 70) = 55.44, p < .001, nonacceptance, F (2, 70) = 28.87, p < .001, impulse, F (2, 70) = 39.05, p < .001, goals, F (2, 70) = 21.86, p < .001, and clarity, F (2, 70) = 24.10, p < .001, were significantly different among clusters. Post-hoc comparisons suggested that participants assigned to Cluster 1 scored significantly lower than those assigned to Cluster 2 in strategies (mean difference = -10.55, p < .001), nonacceptance (mean difference = -9.24, p < .001), impulse (mean difference = -8.79, p < .001), goals (mean difference = -4.31, p < .001) and clarity (mean difference = -4.55, p < .001). On the other hand, the Cluster 1 scored significantly higher than the Cluster 3 in strategies (mean difference = 6.60, p < .001), goals (mean difference = 4.71, p = .003) and lack of emotional clarity (mean difference = 3.41, p = .016). Finally, the Cluster 2 scored significantly higher than the Cluster 3 in strategies (mean difference = 17.15, p < .001), nonacceptance (mean difference = 10.69, p < .001), impulse (mean difference = 12.48, p < .001), goals (mean difference = 9.02, p < .001) and clarity (mean difference = 9.02, p < .001).
No significant differences were found among the clusters regarding age, BMI and durations of the ED or treatment (Table 1). Regarding the distribution of the DSM-5 ED diagnostics into each cluster (Table 2), Cluster 1 included equal proportions of patients with AN restricting type, BN and OSFED. Most patients in Cluster 2 were diagnosed with BN, followed by patients with AN restricting type. Cluster 3 did not include participants with AN binge eating/purging-type and most of the patients in this cluster were diagnosed with AN restricting type.
Characteristics of NSSI in each Cluster
Chi-square tests were conducted to determine the distribution of current (in the preceding week or month of the study) and past (within the previous several months or more than a year ago) NSSI across clusters. A significant relationship was found between cluster membership and engagement in current or past NSSI, χ2 (2) = 14.37, p = .001. While in Cluster 2, most participants (n = 17, 58.6%) reported NSSI during the preceding week or month of the study, in Clusters 1 (n = 18, 62.1%) and 3 (n = 15, 100%), more participants reported NSSI within the previous several months or more than a year ago.
There were no significant differences among the clusters in number of methods of NSSI, F (2, 70) = .24, p = .79. As outlined in Table 3, the most common method of NSSI in the three clusters was cutting.
To sum up, Cluster 1, compared to the sample mean, included participants with higher levels of eating pathology. It was also characterized by lower emotion dysregulation and negative urgency than Cluster 2, but both higher than Cluster 3. The most frequent diagnoses in this cluster were AN restricting type, BN and OSFED. Finally, Cluster 1 included more participants with past NSSI. Cluster 1 was labeled in this study as the “moderate severity cluster”.
Cluster 2 included the participants with the highest scores in the main variables, namely greater levels of eating pathology, emotion dysregulation, and negative urgency than the other clusters. The most frequent diagnoses in this cluster were BN and AN restricting type. Cluster 2 also included more participants with current NSSI. It was labeled as the “high severity cluster”.
Cluster 3 included the participants with the lowest levels of eating pathology, emotion dysregulation and negative urgency. The most frequent diagnosis was AN restricting type, and this cluster included more participants with past NSSI. Cluster 3 was labeled in this study as the “low severity cluster”.