Evaluating Bifactor Models of Psychopathology Using Model-Based Reliability Indices

There is has been a rapid increase in quantitative researchers applying the bifactor model to psychopathology data. The bifactor model, which typically includes a general p factor and internalizing and externalizing residual factors, consistently demonstrates superior model fit to competing models, including the correlated factors model, which typically includes internalizing and externalizing factors. However, the bifactor model’s superior fit might stem from its tendency to overfit noise and flexibly fit most datasets. An alternative approach to evaluating bifactor models that does not rely on fit statistics is model-based reliability assessment. Reliability indices, including omega/omega hierarchical, explained common variance, and percent uncontaminated correlations can be used to determine the viability of the general and specific psychopathology factors and the extent that the underlying data structure and its measurement is multidimensional. In this methodological review, we identified 49 studies published between 2009 and 2019 that applied the bifactor model to at least two separate symptom domains and calculated reliability indices from the standardized factor loading matrices. We also predicted variation in the p factor’s strength, indexed by the explained common variance, from study characteristics. We found that psychopathology measures tend to be multidimensional, with 57% of the variance explained by the p factor and the remaining variance explained by specific factors. By contrast, most of the variance in observed total scores (74%) was explained by the p factor, while relatively little of the variance in in observed subscale scores (37%) was explained by specific factors beyond the p factor. Finally, 62% of the variability in the p factor’s strength could be predicted by study characteristics, most notably the informant (in a simultaneous regression model), but also age, percent uncontaminated correlations, and the number of items (in separate regression models). We conclude that the latent structure of psychopathology is multidimensional, but its measurement is governed by a single dimension, the strength of which is predicted by study characteristics, particularly the informant.