Comparing the test-retest reliability of behavioral, computational and self-reported individual measures of reward and punishment sensitivity in relation to mental health symptoms

The field of computational psychiatry advocates for the use of behavioral task-derived computational measures to improve our understanding, diagnosis and treatment of neuropsychiatric disorders. However, recent meta-analyses in cognitive psychology suggest that behavioral and computational measures are less stable than self-reported surveys as assessed by test-retest correlations. If extended to mental health measures, this poses a serious challenge to the computational psychiatry agenda. To start addressing this question, we leveraged a well-validated reinforcement learning framework based on a neuro-computational model. We collected data from participants who performed a multi-context reinforcement-learning (RL) task twice (~5 months apart), so as to compare the reliability of behavioral measures, computational parameters, and psychological and mental health questionnaires. Despite the remarkable replicability of population-averaged measures, the test-retest reliability of individual metrics was low-to-very-low for behavioral and computational measures, while moderate-to-strong for questionnaires. Behavioral measures were essentially correlated only among themselves, while measures from questionnaires were also correlated with mental health symptoms. Overall, these findings challenge the translational potential of computational approaches for precision psychiatry.