Reliability of computational versus noncomputational metrics of working memory and episodic memory in serious mental illness.

Abstract

Computational models can be used to decompose overall behavior into individual parameters that are designed to isolate specific cognitive and neural processes and provide new information about psychiatric disorders. However, it is not yet clear whether these computational parameters have sufficient reliability to replace traditional noncomputational metrics, especially in severe mental illness. The present study asked whether computational parameters can be as reliable as, or even more reliable than, noncomputational performance metrics in healthy control participants (n = 105) and people with schizophrenia spectrum disorder (n = 72), bipolar disorder (n = 68), and major depressive disorder (n = 76). These participants were tested in working memory and episodic memory paradigms. We applied a computational model that yielded parameters corresponding to memory storage capacity, memory precision, and the rate of attention lapses. We also examined a noncomputational metric of performance, the mean error. Each participant was tested twice, separated by at least 28 days. In most cases, the computational parameters exhibited good-to-excellent split-half and test-retest reliability (>0.80) that was comparable with the reliability of the noncomputational performance metric. In one case, reliability was substantially better for the memory precision parameter than for the noncomputational metric. In another case, reliability for the memory precision parameter was substantially lower but was improved by the use of hierarchical Bayesian estimation. Thus, when applied thoughtfully, parameters derived from computational models can have excellent psychometric properties, providing a reliable means of isolating the factors that underlie individual and group differences in studies of mental illness. (PsycInfo Database Record (c) 2026 APA, all rights reserved).