Psychological inoculation against climate doom.

One of the key goals of computational psychiatry is to understand how mental health problems relate to (or are caused by) changes in cognitive processes such as learning and decision-making (Adams, Huys, & Roiser, ; Friston, Stephan, Montague, & Dolan, ; Huys, Maia, & Frank, ; Kishida, King-Casas, & Montague, ; Maia & Frank, ; Montague, Dolan, Friston, & Dayan, ; Paulus, Huys, & Maia, ). Computational psychiatry researchers often propose that identifying such cognitive changes might improve personalized approaches to treatment, mechanistic understanding of the development and maintenance of illness, and identification of biomarkers for clinical trials (Browning et al., ). However, this promise rests on a number of assumptions – which are increasingly coming under scrutiny (Brown, Chen, Gillan, & Price, ; Eckstein, Wilbrecht, & Collins, ; Eckstein et al., ; Haines et al., ; Karvelis, Paulus, & Diaconescu, ; Katahira, Oba, & Toyama, ; Mkrtchian, Valton, & Roiser, ; Palminteri, Wyart, & Koechlin, ; Pike et al., ; Schaaf, Weidinger, Molleman, & van den Bos, ; Schurr et al., ; Toyama, Katahira, & Kunisato, ; Vrizzi et al., ; Wilson & Collins, ). In this paper, we enumerate these assumptions (for a useful schematic, see Karvelis, Paulus, & Diaconescu, ), and present data from a test case that illustrates them further. The first assumption of the computational psychiatry approach is that model parameters can be recovered reliably – i.e. that the estimated parameters are similar when model-fitting is repeated; or, relatedly, model-fitting of synthetic data generated by known parameters outputs the original generating parameter values (Karvelis, Paulus, & Diaconescu, ; Wilson & Collins, ). If not, then this puts an upper bound on the strength of associations between parameters and other relevant variables that is likely to be observed, and limits statistical power (Wilson & Collins, ). The second assumption is that parameters are stable over time (test–retest reliability). Importantly, parameter values that vary substantially over time within an individual limit the power to detect the effect of an intervention in a repeated-measures design (Pike et al., ). Many computational psychiatry applications, particularly in ‘experimental medicine’ (such as the designation of intermediate endpoints for clinical trials, mechanistic studies to investigate the changes to parameters following interventions, or studies to screen candidate interventions) thus require high test–retest reliability.