The recoverability, reliability, and generalizability of reward processing parameters and relation to mental health...

Theory-driven computational psychiatry attempts to use cognitive models of computation to understand how mental health problems might relate to cognitive processes such as learning and decision-making. However, the potential applications and relevance of this approach are contingent on several (often implicit) assumptions, including that computational parameters 1) are recoverable, 2) are reliable over time, 3) reflect conceptually similar processes across different tasks, and 4) relate to symptoms.

To illustrate and test these assumptions for a selection of commonly used cognitive tasks, we recruited a large online sample of participants (n = 548), who completed seven mental health questionnaires and five tasks. A subset of n = 115 re-completed the tasks 14 days later. For each task, five models were fit, and the winning model was selected through Bayesian model comparison.

The parameters of the winning models showed excellent recovery and good-to-excellent test-retest reliability. Parameters that were theoretically similar to each other were, however, generally not related, with relationships only found within the class of 'inverse temperature'-like parameters. Finally, only parameters from two of the five tasks (four-armed bandit and cognitive effort) related to symptoms, and these relationships were weak.

It seems that at least some of the implicit assumptions made when advocating for the use of computational models in psychiatry are not met for these popular tasks and computational models. Computational psychiatry researchers should carefully assess the assumptions and psychometric properties of their tasks and models as part of their early-phase development to ensure robustness of the field going forward.