The publication of the fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5; American Psychiatric Association, 2013) recognized binge-eating disorder (BED) as a formal and specific eating-disorder diagnosis and introduced a severity specifier comprising four levels. The severity levels are based on the frequency of binge-eating episodes per week defined as mild (1–3), moderate (4–7), severe (8–13), and extreme (⩾14). Ideally, severity indicators would index ‘not only the intensity of defining psychopathological features of the illness, but also the level of functional impairment’ (Giannini et al., Reference Giannini, Roberto, Attia, Walsh, Thomas, Eddy and Sysko2017, p. 914) and the appropriate level of care (e.g. outpatient, inpatient).
The DSM-5 BED severity indicator based on binge-eating frequency was not empirically identified (e.g. <3 v. 4–7 binge-eating episodes) and evidence is mixed regarding the indicator's clinical utility. For instance, some studies report a linear trend of greater binge-eating severity being associated with greater intensity of other eating-disorder symptoms, comorbid symptoms, and impairment (Dakanalis, Colmegna, Riva, & Clerici, Reference Dakanalis, Colmegna, Riva and Clerici2017a; Reference Dakanalis, Riva, Serino, Colmegna and Clerici2017b; Giannini et al., Reference Giannini, Roberto, Attia, Walsh, Thomas, Eddy and Sysko2017; Grilo, Ivezaj, & White, Reference Grilo, Ivezaj and White2015a; Reference Grilo, Ivezaj and White2015b). In contrast, other studies report no significant group differences on eating-disorder psychopathology across BED severity groups (Nakai et al., Reference Nakai, Nin, Noma, Teramukai, Fujikawa and Wonderlich2017; Smith et al., Reference Smith, Ellison, Crosby, Engel, Mitchell, Crow and Wonderlich2017) and three of the five studies that reported statistically significant findings highlighted the small effect sizes (Giannini et al., Reference Giannini, Roberto, Attia, Walsh, Thomas, Eddy and Sysko2017; Grilo et al., Reference Grilo, Ivezaj and White2015a, Reference Grilo, Ivezaj and White2015b). To date, three studies investigated the predictive significance of the DSM-5 BED severity indicator on treatment outcomes. Two report limited utility (Lydecker, Ivezaj, & Grilo, Reference Lydecker, Ivezaj and Grilo2020; Smith et al., Reference Smith, Ellison, Crosby, Engel, Mitchell, Crow and Wonderlich2017) and one reports significant differences (Dakanalis et al., Reference Dakanalis, Colmegna, Riva and Clerici2017a). Collectively, evidence suggests that either the currently-specified frequencies of binge-eating episodes are inadequate indicators of BED severity or that binge-eating episodes overall are an inaccurate metric by which to specify BED severity.
Indeed, an alternative BED severity indicator based on the presence of a clinically significant level of shape/weight overvaluation has been proposed (Grilo, Reference Grilo2013). Shape/weight overvaluation occurs when one's body shape or weight is one of the most important components defining one's self-worth. Prevailing theoretical conceptualizations of eating disorders propose that shape/weight overvaluation is what motivates and maintains engagement in eating-disorder behaviors (Fairburn, Cooper, & Shafran, Reference Fairburn, Cooper and Shafran2003). Accordingly, shape/weight overvaluation is widely considered to be the transdiagnostic, core eating-disorder psychopathology of anorexia nervosa, bulimia nervosa, and BED (Fairburn et al., Reference Fairburn, Cooper and Shafran2003). In support of this position, epidemiological (Coffino, Udo, & Grilo, Reference Coffino, Udo and Grilo2019) and clinical studies (Giannini et al., Reference Giannini, Roberto, Attia, Walsh, Thomas, Eddy and Sysko2017; Grilo et al., Reference Grilo, Ivezaj and White2015a; Reference Grilo, Ivezaj and White2015b; Kenny & Carter, Reference Kenny and Carter2018) report that people with BED with shape/weight overvaluation are characterized by significantly greater eating-disorder psychopathology and impairment than people with BED but without shape/weight overvaluation, with effect sizes generally moderate-to-large. Taken together, shape/weight overvaluation appears to more reliably and strongly signal severity in BED compared to the DSM-5 binge-eating frequency severity specifier. However, since shape/weight overvaluation is a continuous construct, dichotomizing may result in over-simplification and loss of important information (Altman & Royston, Reference Altman and Royston2006).
Overall, gaps exist in not only directly comparing the impact of binge-eating frequencies v. shape/weight overvaluation on the differentiation of BED severity but also in identifying the precise levels of binge-eating frequencies and/or shape/weight overvaluation that most meaningfully differentiate BED severity. To fill this gap, the current project uses structural equation model (SEM) trees – a form of exploratory data mining – to determine (1) which of the two severity indicators (i.e. binge-eating frequency v. shape/weight overvaluation) most meaningfully differentiates levels of BED severity and (2) the precise levels of the severity indicator (e.g. 2, 5, and 7 binge-eating episodes per week) that most significantly differentiate BED severities. We hypothesized that shape/weight overvaluation would be more important in BED severity differentiation relative to binge-eating frequency. Hypotheses were not specified for the precise levels of the severity indicators that would emerge, given that this aspect of analyses was exploratory. We also compared whether the empirically determined BED severity groupings explained more variance in clinical characteristics relative to the two existing BED severity classification schemes.
Exploratory data mining belongs to a family of statistical learning techniques that rely more on data to suggest relations among variables rather than on a priori assumptions about data. This inductive approach has identified important predictors and interactions among predictors of eating-disorder onset (Mehl, Rohde, Gau, & Stice, Reference Mehl, Rohde, Gau and Stice2019; Stice & Desjardins, Reference Stice and Desjardins2018). For example, certain identified interactions had not been known (i.e. interactions among body dissatisfaction, overeating, dieting, and thin-ideal internalization in the prediction of BED onset; Stice and Desjardins, Reference Stice and Desjardins2018) and therefore would not have been specified using traditional analytic methods (King & Resick, Reference King and Resick2014). In addition to being used to identify predictors of eating-disorder onset, exploratory data mining can also be used to empirically determine distinct subgroups of people with eating disorders. Of course, traditional statistical methods [e.g. analysis of variance (ANOVA)] also identify group differences but require group membership to be imposed a priori and, as emphasized above, the existing BED severity indicators have mixed support. Allowing patterns in the data to determine subgroups, however, removes bias or arbitrariness that may exist in current definitions of severity and may yield more clinically reliable criteria to determine severity.
One limitation of some exploratory data mining approaches is that they rely solely on the data and lack confirmatory or theory-based aspects. However, SEM Trees ‘combine the benefits of confirmatory and exploratory approaches to data analysis’ (Brandmaier, Oertzen, McArdle, & Lindenberger, Reference Brandmaier, Oertzen, McArdle and Lindenberger2013, p. 71) by (1) specifying a template or ‘outcome’ SEM, which is based on theory and evidence (confirmatory), and (2) recursively partitioning the data into subgroups that explain the most variance in the outcome model (exploratory). This combination of confirmatory and exploratory approaches allows us to simultaneously capitalize on – yet extend – knowledge about constructs that may capture the essence BED severity.
Method
Participants
Participants were 788 adult treatment-seeking men and women with BED enrolled in treatment studies evaluating pharmacological and/or psychological BED treatments. Participants ranged in age from 18 to 65 years and met full DSM-IV (American Psychiatric Association, 2004) BED research diagnostic criteria (which exceeds requirements for DSM-5-defined BED). Exclusion criteria included receiving concurrent treatment for eating or weight problems, being pregnant or diagnosed with a medical condition that influenced eating or weight (e.g. diabetes), or had select severe psychiatric diagnoses (e.g. psychosis, florid bipolar) that could either interfere with clinical assessment or dictated other treatments. This research received full review and approval from the Yale Institutional Review Board. All participants provided written informed consent prior to performing study procedures.
Procedure
At baseline evaluations prior to treatment, trained and monitored doctoral-level clinicians administered the Structured Clinical Interview for DSM Diagnosis (SCID; First, Spitzer, Gibbon, and Williams, Reference First, Spitzer, Gibbon and Williams1997) and the Eating Disorder Examination interview (EDE; Fairburn and Cooper, Reference Fairburn, Cooper, Fairburn and Wilson1993). The SCID was used to determine the BED diagnosis and age of BED onset as well as psychiatric comorbidity (see below). Kappa coefficients for interrater reliability for BED was 1.0 and ranged from 0.68 to 1.0 for other psychiatric diagnoses. The EDE was used to confirm BED diagnosis and to assess the full range of eating-disorder psychopathology including binge-eating behaviors and frequencies. Participants also completed a battery of psychometrically-established self-report measures and had their body weight and height measured objectively.
Measures and variables
Outcome model: latent BED severity
BED severity was modeled as a latent variable indicated initially by eating-disorder symptoms, body mass index (BMI), depressive symptoms, and the number of psychiatric comorbidities.
Eating-disorder symptoms. Eating-disorder symptoms were quantified by the EDE global score. To calculate the global score, 22 of the EDE items are averaged into four subscales: restraint, eating concerns, shape concerns, and weight concerns. The subscales are then averaged into a global score with higher values indicating greater severity. The global score was used in analyses because it captures overall eating-disorder psychopathology and the individual subscales do not replicate in factor analyses with BED samples (e.g. Grilo et al., Reference Grilo, Crosby, Peterson, Masheb, White, Crow and Mitchell2010). Inter-rater reliability was calculated on a portion of the sample (n = 113) and the intraclass correlation for EDE global score was 0.92.
BMI. BMI, calculated from measured weight and height, was included in the outcome model as a physical indicator of severity.
Depressive levels. Depressive levels were assessed with the Beck Depression Inventory (BDI; Beck and Steer, Reference Beck and Steer1987). The BDI is a 21-item, self-report measure of depressive levels during the past two weeks. Items are scored 0–3 and summed, where higher scores indicate more severe depressive levels. The BDI taps, in addition to depression, a broad range of negative affect, and correlates strongly with broad psychopathology (Grilo, Masheb, & Wilson, Reference Grilo, Masheb and Wilson2001). In this study, BDI scores were included in the outcome model as one of two indicators of comorbidity (i.e. to complement the SCID-based diagnostic comorbidity) given findings provide incremental value for identifying concurrent severity about categorical diagnoses (Grilo et al., Reference Grilo, Masheb and Wilson2001) and predicting treatment outcomes (Grilo, Masheb, & Crosby, Reference Grilo, Masheb and Crosby2012). In the present study, the BDI showed excellent internal consistency (α = 0.88).
Lifetime diagnostic comorbidity. The lifetime number of comorbid disorders was quantified by totaling the lifetime number of depressive, bipolar, anxiety, posttraumatic stress, obsessive–compulsive, somatic, substance (alcohol and drug), and eating disorders (not including BED), determined with the SCID diagnostic interview. The lifetime frequency of other diagnoses was included as the second indicator of comorbidity, given that 79–94% of people with BED have at least one other co-occurring form of psychopathology (Hudson, Hiripi, Pope, & Kessler, Reference Hudson, Hiripi, Pope and Kessler2007; Udo & Grilo, Reference Udo and Grilo2019).
Model covariates: binge-eating frequency and shape/weight overvaluation
The average weekly binge-eating frequency over the past 28 days (intraclass correlation = 0.94) and the severity of shape/weight overvaluation were used as the two model covariates (described further below). Both variables were assessed with the EDE. Shape overvaluation and weight overvaluation were assessed with two separate EDE items. However, the items were highly correlated (r = 0.83, p < 0.001). We, therefore, averaged the items and used this averaged variable for analyses.
Data analytic plan
Analyses were performed using R software (R Core Team, 2013) and the following packages: mice (van Buuren & Groothuis-Oudshoorn, Reference van Buuren and Groothuis-Oudshoorn2011), OpenMX (Boker et al., Reference Boker, Neale, Maes, Wilde, Spiegel, Brick and Fox2011; Reference Boker, Neale, Maes, Wilde, Spiegel, Brick and Brandmaier2012), lavaan (Rosseel, Reference Rosseel2012), semtree (Brandmaier et al., Reference Brandmaier, Oertzen, McArdle and Lindenberger2013), and stats (R Core Team, 2013).
Missing data
Missing data ranged from 0% (EDE variables) to 0.01% (BDI). Data were consistent with a pattern of missing completely at random, χ 2(10) = 8.13, p = 0.62. Missing data were imputed with single imputation.
SEM Trees
SEM Tree analyses took place in two parts: specification of an outcome model (confirmatory) and decision trees (exploratory). A latent variable of BED severity (referred to as the outcome model) was specified through confirmatory factor analysis. Model fit indices and factor loadings were inspected. Model fit statistics included model χ2, Root Mean Square Error of Approximation (RMSEA), Comparative Fit Index (CFI), and Tucker Lewis Index (TLI). ‘Good’ fit was indicated by RMSEA<0.08, CFI⩾0.90, and TFI⩾0.90. ‘Excellent’ fit was indicated by RMSEA < 0.05, CFI > 0.95, and TLI > 0.95 (Hu & Bentler, Reference Hu and Bentler1999).
Upon ensuring adequate model fit, the decision tree recursively partitioned data into subgroups that explained the maximum amount of variance in the outcome model of BED severity; the subgroups were based on specific values (i.e. ‘splits’) of the covariates. In SEM Trees, covariates are the variables for which one wants to identify discrete levels or subgroups of cases (e.g. subgroups based on binge-eating frequency or shape/weight overvaluation). Splits were determined using the ‘fair’ splitting criterion, which randomly divided the sample into two halves. In the first half, all possible values upon which the covariates could be split were compared, and the split value that resulted in the largest improvement in model fit was selected. The split value was then evaluated in the second half of the sample. The split value that resulted in the most improvement in model fit was then selected as the ‘split.’ These split values were used to obtain the precise levels of binge-eating frequency and/or shape/weight overvaluation that differentiated BED severity (e.g. see Fig. 1).
Results of random forest structural equation model tree analysis.
Note: sw_overval, shape and weight overvaluation; obe_wk, objective binge-eating episodes per week. resid1, residual variance of eating-disorder symptoms; resid2, residual variance of comorbid psychiatric diagnoses; resid3, residual variance of Beck Depression Inventory (BDI) depressive levels; v1, latent binge-eating disorder severity factor variance; m1, manifest mean eating-disorder symptoms; m2, manifest mean comorbid psychiatric diagnoses; m3, manifest mean BDI depressive levels.
SEM Forests
Decision trees have high interpretability – i.e. because a single tree is fit, its parameters (e.g. where ‘splits’ occur) are easily identifiable. However, because of inherent variability within a single tree, parameters may be unstable (e.g. an initial ‘split’ may be influenced by anomalies in the data, which could have downstream effects in how the rest of the tree is split). To counter this limitation, we completed SEM Forest analyses (Brandmaier, Prindle, McArdle, & Lindenberger, Reference Brandmaier, Prindle, McArdle and Lindenberger2016). SEM Forests are an ensemble method, where multiple decision trees were fit (n = 100) and one variable per split was compared (specified via mtry = 1). Each of the 100 decision trees could include different split values, which would compromise interpretability. Therefore, SEM Forest analyses yield an aggregate metric of variable importance to determine, within the forest of individual trees, the relative strength (in the scale of −2 log-likelihood) of the covariates in differentiating subgroups on the outcome model. In our case, variable importance quantified whether binge-eating frequency v. shape/weight overvaluation more robustly differentiated BED severity.
Planned comparisons
ANOVAs with orthogonal, planned contrasts were used to compare SEM Tree-derived groups on demographic and clinical characteristics. Effect sizes were indicated by η 2. Contrasts sequentially compared the group with lower severity to the groups with higher severity (e.g. Group 1 v. the combined Groups 2–5; Group 2 v. the combined Groups 3–5; etc.). Contrast effect sizes (r con) indexed the magnitude of group differences and were computed with the following formula: $\surd ($
t 2/t 2 + df) (Field, Reference Field2013).
To compare whether the SEM Tree-derived severity groups outperformed existing BED severity indicators, we created two additional grouping variables: one based on the binge-eating frequencies specified in the DSM-5 severity levels and another based on clinically significant shape/weight overvaluation (either EDE item rated as ⩾ 4). ANOVAs with planned, orthogonal contrasts sequentially compared the DSM-5 BED severity groups on demographic and clinical characteristics. T tests compared shape/weight overvaluation groups on demographic and clinical characteristics. Effect sizes were calculated for all group comparisons. We descriptively compared effect sizes for each BED severity classification scheme.
Variables were inspected for normality. Binge-eating frequency showed significant skew and kurtosis. Analyses were completed using the raw data and a log-transformed variable, and the pattern of results was highly similar. For ease of interpretation, results using the non-transformed binge-eating frequency variable are reported here. Levene's test was performed to determine whether heterogeneity of variance was present. Multiple variables exhibited significant heterogeneity of variance (see Tables 1–4). Welch-corrected ANOVAs and contrasts are reported as indicated.
Demographic, physical, and clinical characteristics compared among structural equation model tree-derived groups
SEM, structural equation modeling; BED, binge-eating disorder; overvaluation, shape and weight overvaluation; binge eating, weekly binge-eating frequency; ED, eating disorder; BDI, Beck Depression Inventory.
Note: See Fig. 1 for SEM Tree diagram depicting group splits.
* Indicates significant heterogeneity of variance and thus Welch-corrected ANOVA reported.
Structural equation model tree-derived groups' contrast results for demographic and clinical characteristics
BED, binge-eating disorder; overvaluation, shape and weight overvaluation; binge eating, weekly binge-eating frequency; ED, eating disorder; BDI, Beck Depression Inventory.
Note: r con is the effect size for the contrast.
* Indicates significant heterogeneity of variance and thus Welch-corrected contrasts are reported.
Descriptive statistics and group comparisons of demographic and clinical characteristics for Diagnostic and Statistical Manual for Mental Disorders-5 specified severity indicators
Overvaluation, shape and weight overvaluation; binge eating, weekly binge-eating frequency; ED, eating disorder; BDI, Beck Depression Inventory; r con, effect size for the contrast.
* Indicates significant heterogeneity of variance and thus Welch-corrected contrasts are reported.
Demographic and clinical characteristics and contrast results for groups based on shape/weight overvaluation clinical threshold of 4
Overvaluation, shape and weight overvaluation; binge eating, weekly binge-eating frequency; ED, eating disorder; BDI, Beck Depression Inventory; r con, effect size for the contrast.
* Indicates significant heterogeneity of variance and thus Welch-corrected contrasts are reported.
Results
Sem Trees and Forests
Confirmatory factor analysis of the BED severity outcome model yielded a one-factor solution. Model fit was excellent (χ2(2) = 5.38, p = 0.07; CFI = 0.99; TLI = 0.96; RMSEA = 0.05, 90% CI (0.00, 0.10), p = 0.46). The loadings for eating-disorder symptoms, depression levels, and comorbid psychiatric disorders were significant (ps < 0.001). The loading for BMI was small and nonsignificant (0.05, p = 0.65). BMI was thus removed. The revised model included only three indicators, resulting in a just-identified model for which fit could not be evaluated. Because all factor loadings in the revised model remained significant (p < 0.001), this model was used for SEM Tree analyses.
SEM Tree model results are shown in Fig. 1. The first split occurred at shape/weight overvaluation ⩾2.75. Among those with shape/weight overvaluation of <2.75, two subgroups emerged: those with shape/weight overvaluation <1.25 (Group 1, n = 100) and those with shape/weight overvaluation = 1.25–2.74 (Group 2, n = 128). Among those with shape/weight overvaluation ⩾2.75, another split occurred at shape/weight overvaluation ⩾4.25. This created one subgroup with shape/weight overvaluation = 2.75–4.24 (Group 3, n = 245). Among those with shape/weight overvaluation ⩾4.25, another split occurred at weekly binge eating frequency ⩾4.875. This split created two subgroups: those with shape/weight overvaluation ⩾4.25 and weekly binge-eating frequency <4.875 (Group 4, n = 195) and those with shape/weight overvaluation ⩾4.25 and weekly binge-eating frequency ⩾4.875 (Group 5, n = 120). As shown in Fig. 1, means for both the latent BED severity factor and its indicators increased across groups.
Because data were aggregated from multiple BED medication and psychological treatment studies, we inspected whether SEM Tree groups were associated with study type (online Supplementary Tables 1 and 2). Results indicated no consistent or meaningful associations between SEM Tree groups and study type.
SEM Forest analyses, which quantified how much each covariate contributed to model differentiation, revealed that splits that occurred on shape/weight overvaluation v. binge-eating frequency resulted in 190.40 v. 10.10 units of improvement in model fit, respectively. This indicates that shape/weight overvaluation had much more importance in empirically identifying BED severities subgroups as compared to binge-eating frequency.
SEM Tree group comparisons
Demographic and clinical characteristics for the SEM Tree-derived BED severity groups are displayed in Table 1. Chi-square and ANOVA results indicated that the groups did not significantly differ on sex, race, education, age, BMI, or age of BED onset. However, significant differences were found for shape/weight overvaluation, binge-eating frequency, eating-disorder symptoms, depression levels, and comorbid psychiatric diagnoses.
To identify where group differences emerged, orthogonal, planned contrasts sequentially compared the group with lower severity to the combined groups with higher severity (Table 2). Similar to the overall ANOVA result contrasts comparing groups on age, age of onset, and BMI were nonsignificant with small effect sizes. However, many group differences emerged on clinical characteristics. Specifically, comparing Group 1 to the combined Groups 2–5, the combined group had significantly greater shape/weight overvaluation, eating-disorder symptoms, depression levels, and comorbid psychiatric diagnoses. Binge-eating frequency was the only clinical characteristic where group differences did not emerge. The results of all other contrasts (i.e. Group 2 v. combined Groups 3–5, Group 3 v. combined Groups 4–5, and Group 4 v. Group 5) revealed a consistent pattern: groups with higher severity had significantly greater shape/weight overvaluation, binge-eating frequency, eating-disorder symptoms, depression levels, and comorbid psychiatric diagnoses than groups with lower severity. Four effect sizes were large, six were moderate, and ten were small.
Comparing SEM Tree-derived groups with existing severity indicators
Results of the DSM-5 ANOVAs and planned orthogonal contrasts are displayed in Table 3. Results revealed significant DSM-5 group differences on all outcome variables. Contrasts for the DSM-5 groups revealed that when comparing the Mild group to the combined Moderate–Extreme groups, the combined group had significantly greater shape/weight overvaluation, binge-eating frequency, eating-disorder symptoms, depression levels, and number of comorbid psychiatric diagnoses. When comparing the Moderate group to the combined Severe and Extreme group, the combined group had significantly greater binge-eating frequency. No other group differences emerged. When comparing the Severe group to the Extreme group, the Extreme group had significantly higher BMI and binge-eating frequency; groups did not differ on any other characteristics. Effect sizes for clinical characteristics were exclusively in the small range, with one exception of large effect sizes for binge-eating frequency.
Results for the clinically significant shape/weight overvaluation t tests are displayed in Table 4. People with the clinically significant shape/weight overvaluation had significantly greater shape/weight overvaluation, binge-eating frequency, eating-disorder symptoms, depression levels, and comorbid psychiatric diagnoses than people without clinically significant shape/weight overvaluation. Effect sizes for clinical characteristics were mixed among small, moderate, and large.
Effect sizes for all severity indicators are shown in Tables 1–4. The SEM Tree-derived indicator explained 9–93% of the variance in the clinical characteristics (mean = 35%). The DSM-5 indicator explained between 2 and 88% of the variance in clinical characteristics (mean = 19%). The clinically significant shape/weight overvaluation indicator explained between 0.06 and 71% of the variance in clinical characteristics (mean = 23%). Overall, the SEM Tree-derived indicator explained 1.84 more variance than the DSM-5 indicator and 1.52 more variance than the clinically significant shape/weight overvaluation indicator.
Discussion
This study empirically identified BED severity indicators based on shape/weight overvaluation and binge-eating frequency. Three primary results emerged. First, four splits were identified, creating five severity subgroups. The first three splits occurred at increasing intensities of shape/weight overvaluation. The final split occurred on an average of approximately five binge-eating episodes per week. Second, shape/weight overvaluation contributed to much more improvement in model fit than binge-eating frequency. Third, the SEM Tree-derived severity indicator explained between approximately 1.5–2 times more variance in clinical characteristics compared to existing BED severity classification schemes. Together, results suggest that BED severity may need to be considered in a nuanced way, based on a combination of cognitive and behavioral eating-disorder symptoms.
Findings confirmed that BED severity varies and highlighted five levels of BED severity. This is inconsistent with the DSM-5′s severity scheme in that only one of the four splits occurred on binge-eating frequency. This split occurred at an average of roughly five binge-eating episodes per week, which is not a value included in the DSM-5 BED severity levels. Another inconsistency with DSM-5's severity scheme comes from the finding that the SEM Tree-derived groups consistently differed on eating-disorder symptoms, depression levels, and psychiatric comorbidities, whereas DSM-5 severity groups consistently differed on only binge-eating frequency. The DSM-5-defined groups may not have strongly differentiated participants because binge-eating frequency is an insufficient BED severity criterion and/or clinical judgment was not incorporated into severity ratings, as DSM-5 recommends (Giannini et al., Reference Giannini, Roberto, Attia, Walsh, Thomas, Eddy and Sysko2017).
Multiple intensities of shape/weight overvaluation that corresponded to increasing levels of severity were identified, occurring at values of 1.25 (no – some importance), 2.75 (some – moderate importance), and 4.25 (moderate importance). This is a departure from other studies investigating shape/weight overvaluation as a BED severity indicator, which all used a binary cutoff of 4–5 on a 6-point scale. Our data-driven findings confirm that moderate shape/weight overvaluation differentiates levels of BED severity (Goldschmidt et al., Reference Goldschmidt, Hilbert, Manwaring, Wilfley, Pike, Fairburn and Striegel-Moore2010), yet also extends the literature by indicating that shape/weight overvaluation values of approximately 1 and 3 may also correspond to lower levels of severity. Additional studies are needed to confirm and validate these values as indicative of BED severity.
Shape/weight overvaluation contributed more strongly to group differentiation and explained more variance in clinical characteristics than the DSM-5 defined binge-eating frequency levels. The greater relative importance of shape/weight overvaluation v. binge-eating frequency in determining BED severity is consistent with network studies of eating-disorder psychopathology, which find that shape/weight overvaluation – rather than behavioral eating-disorder symptoms (binge eating, restriction, purging) – appears to be at the core of eating-disorder maintenance (e.g. Forrest, Jones, Ortiz, & Smith, Reference Forrest, Jones, Ortiz and Smith2018; DuBois, Rodgers, Franko, Eddy, & Thomas, Reference DuBois, Rodgers, Franko, Eddy and Thomas2017; Wang, Jones, Dreier, Elliott, & Grilo, Reference Wang, Jones, Dreier, Elliott and Grilo2019). For instance, in a BED symptom network, shape/weight overvaluation was among the most central symptoms, which in this case meant that these symptoms had the strongest and most frequent connections with all other symptoms in the network (Wang et al., Reference Wang, Jones, Dreier, Elliott and Grilo2019). Binge-eating frequency had relatively low centrality and was not highly connected with other eating-disorder symptoms. According to the network theory of psychopathology, highly central symptoms are those that are likely implicated in symptom and disorder maintenance (Borsboom & Cramer, Reference Borsboom and Cramer2013). Overall, this study, along with several others (see Grilo, Reference Grilo2013), suggests that incorporating shape/weight overvaluation into DSM-5 BED severity specification could increase diagnostic accuracy and inform treatment planning beyond just level of care (e.g. specific treatment targets). For example, one study found that overvaluation moderated treatment outcomes: patients with overvaluation appeared to benefit more from cognitive-behavioral methods than pharmacotherapy (Grilo et al., Reference Grilo, Masheb and Crosby2012) although further studies are needed (Grilo, White, Gueroguieva, Wilson, & Masheb, Reference Grilo, White, Gueroguieva, Wilson and Masheb2013).
A novel contribution of this study is that the SEM Tree identified an interaction in severity indication between shape/weight overvaluation and binge-eating frequency. Specifically, among people with the highest levels of shape/weight overvaluation (rated ⩾4.25), two subgroups of severity emerged: those with less than roughly five binge-eating episodes per week and those with five or more binge-eating episodes per week. This suggests that to fully operationalize BED severity, our definitions may need to be based on a combination of key cognitive and behavioral eating-disorder symptoms. According to the current results, distinct levels of shape/weight overvaluation may be the primary indicator of BED severity. A secondary indicator, which would apply to people with the greatest degrees of shape/weight overvaluation, would be engaging in five or more binge-eating episodes per week. Historically, the eating-disorders field has not often included in conceptual models possible interactions among variables at specific levels. One reason for this is that models must be testable and without using advanced statistical techniques, such as machine-learning approaches (e.g. decision trees), the field lacked the ability to hypothesize or identify the thresholds where variable interactions may be relevant. However, decision trees and other advanced statistical techniques are allowing for the identification of complex interactions among variables and may aid in the prevention, diagnosis, assessment, and treatment of eating disorders (Haynos et al., Reference Haynos, Wang, Lipson, Peterson, Mitchell, Halmi and Crowin press; Mehl et al. Reference Mehl, Rohde, Gau and Stice2019; Stice & Desjardins, Reference Stice and Desjardins2018).
We note several methodological strengths as context. The study group comprised a large, racially and ethnically diverse sample of patients with BED, with men accounting for approximately one-third of the sample. Participants were assessed and diagnosed using well-established interviews by doctoral-level research clinicians who had been trained and carefully monitored to assure reliability and prevent drift in diagnostic tasks. Our outcome model of latent BED severity was well rounded, as it included indicators related to eating-disorder psychopathology specifically and psychopathology more broadly (e.g. depression/negative affect).
Several limitations deserve note. First, the identified severity splits need to be replicated and validated across additional samples to account for within-sample anomalies. Second, analyses used cross-sectional data and future research should test whether the identified severity splits predict treatment outcomes. Third, the outcome model of latent BED severity and the covariates on which splits were determined were theoretically relevant yet not exhaustive. Different outcome models and/or covariates with other constructs implicated in BED maintenance (e.g. cognitive control, disrupted food reward processing; Balodis, Grilo, & Potenza, Reference Balodis, Grilo and Potenza2015; Kober & Boswell, Reference Kober and Boswell2018) might yield different severity splits and classification schemes. Fourth, analyses included participants who met full DSM-IV research diagnostic BED criteria. Results may not generalize to subthreshold BED and may not replicate when using different diagnostic criteria for BED, such as the International Classification of Diseases (11th Revision), which does not require an unusually large amount of food as a criterion for binge-eating (Stein et al., Reference Stein, Szatmari, Gaebel, Berk, Vieta, Maj and Reed2020). Fifth, findings may not generalize to groups with BED who differ on sociodemographic characteristics, are non-treatment-seeking, or who chose not to participate in research.
In conclusion, the current study used SEM Trees to empirically determine BED severity indicators and specifically compared the relative importance of shape/weight overvaluation v. binge-eating frequencies in BED severity differentiation. Results revealed four severity splits: three were defined by increasing intensity of shape/weight overvaluation and one was defined by greater binge-eating frequency. This final split was relevant only to people with the most extreme levels of shape/weight overvaluation. Shape/weight overvaluation contributed to much more improvement in model fit than binge-eating frequency. In addition, the empirically determined BED severity indicators consistently differentiated groups on eating-disorder symptoms, depression levels, and psychiatric comorbidities, and accounted for 1.5–2 times the variance in group differentiation relative to the existing BED severity indicators. Overall, these results indicate that shape/weight overvaluation more strongly accounts for BED severity differentiation than binge-eating frequency. Results also provide four specific and empirically determined thresholds of shape/weight overvaluation and binge-eating frequency by which to determine BED severity, including an interaction between these two symptoms to differentiate the highest levels of BED severity. These results set the stage for additional investigations to replicate and examine the prognostic and predictive validity of these empirically determined BED severity indicators.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291720002287
Acknowledgements
This research was supported, in part, by National Institutes of Health grant R01 DK49587 (CMG). CMG was also supported, in part, by National Institutes of Health grants R01 DK114075 and R01 DK112771. Funders played no role in the content of this paper.
Conflict of interest
None, though Dr Grilo reports several broader interests which did not influence this research or paper. Dr Grilo's broader interests include Consultant to Sunovion and Weight Watchers; Honoraria for lectures, CME activities, and presentations at scientific conferences; and Royalties from Guilford Press and Taylor & Francis Publishers for academic books.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
