DSM-5 specifies bulimia nervosa (BN) severity based on specific thresholds of compensatory behavior frequency. There is limited empirical support for such severity groupings. Limited support could be because the DSM-5’s compensatory behavior frequency cutpoints are inaccurate or because compensatory behavior frequency does not capture true underlying differences in severity. In support of the latter possibility, some work has suggested shape/weight overvaluation or use of single versus multiple purging methods may be better severity indicators. We used structural equation modeling (SEM) Trees to empirically determine the ideal variables and cutpoints for differentiating BN severity, and compared the SEM Tree groupings to alternate severity classifiers: the DSM-5 indicators, single versus multiple purging methods, and a binary indicator of shape/weight overvaluation.

Treatment-seeking adolescents and adults with BN (N = 1017) completed self-report measures assessing BN and comorbid symptoms. SEM Trees specified an outcome model of BN severity and recursively partitioned this model into subgroups based on shape/weight overvaluation and compensatory behaviors. We then compared groups on clinical characteristics (eating disorder symptoms, depression, anxiety, and binge eating frequency).

SEM Tree analyses resulted in five severity subgroups, all based on shape/weight overvaluation: overvaluation <1.25; overvaluation 1.25–3.74; overvaluation 3.75–4.74; overvaluation 4.75–5.74; and overvaluation ≥5.75. SEM Tree groups explained 1.63–6.41 times the variance explained by other severity schemes.

Shape/weight overvaluation outperformed the DSM-5 severity scheme and single versus multiple purging methods, suggesting the DSM-5 severity scheme should be reevaluated. Future research should examine the predictive utility of this severity scheme.

Eating-disorder severity indicators should theoretically index symptom intensity, impairment, and level of needed treatment. Two severity indicators for binge-eating disorder (BED) have been proposed (categories of binge-eating frequency and shape/weight overvaluation) but have mixed empirical support including modest clinical utility. This project uses structural equation model (SEM) trees – a form of exploratory data mining – to empirically determine the precise levels of binge-eating frequency and/or shape/weight overvaluation that most significantly differentiate BED severities.

Participants were 788 adults with BED enrolled in BED treatment studies. Participants completed interviews and self-report measures assessing eating-disorder and comorbid symptoms. SEM Tree analyses were performed by specifying an outcome model of BED severity and then recursively partitioning the outcome model into subgroups. Subgroups were split based on empirically determined values of binge-eating frequency and/or shape/weight overvaluation. SEM Forests also quantified which variable contributed more improvement in model fit.

SEM Tree analyses yielded five subgroups, presented in ascending order of severity: overvaluation <1.25, overvaluation = 1.25–2.74, overvaluation = 2.75–4.24, overvaluation ⩾4.25 with weekly binge-eating frequency <4.875, and overvaluation ⩾4.25 with weekly binge-eating frequency ⩾4.875. SEM Forest analyses revealed that splits that occurred on shape/weight overvaluation resulted in much more improvement in model fit than splits that occurred on binge-eating frequency.

Shape/weight overvaluation differentiated BED severity more strongly than binge-eating frequency. Findings indicate a nuanced potential BED severity indicator scheme, based on a combination of cognitive and behavioral eating-disorder symptoms. These results inform BED classification and may allow for the provision of more specific and need-matched treatment formulations.