Mechanistic arbitration between candidate dimensions of psychopathology

Transdiagnostic and dimensional alternatives to the Diagnostic and Statistical Manual of Mental Disorders have gained prominence in recent years. A key critique of these approaches, however, is that they are typically based on symptom correlations and as such are purely descriptive and therefore unlikely to have mechanistic grounding. We tested this idea empirically, conducting a large-scale comparison of competing factor solutions that allowed us to determine if the latent structure underlying the covariation of psychiatric symptoms has robust and specific cognitive correlates. In nine independent datasets, comprising N=7565 individuals including patients, healthy individuals, paid and unpaid participants, a broad set of age ranges and cognitive task variants that measured model-based planning and metacognition, we found that this was the case. Across datasets, the factors with the best fit to cognition were those derived from a first-order factor analysis on the maximal number of theoretically informed self-report symptoms available. These factors (‘Compulsivity and Intrusive Thought’ and ‘Anxious-depression’) performed better than thousands of engineered alternatives and performed twice as well as traditional questionnaire total scores. Crucially, this unsupervised approached based on symptom correlation only performed on-par with a partial least squares analysis, a supervised approach to deriving factors based on cognition. These results provide evidence that unsupervised factor analysis of psychiatric symptoms is a viable method for rethinking how we define mental health and illness, affording clear opportunities for enhancing our understanding of specific underlying mechanistic processes.