Computational Cognitive Model Parameters as Dependent Variables in Regression

Computational cognitive models (CCMs) are powerful frameworks that posit latent cognitive parameters as drivers of human behavior, often employed on experimental data. Researchers use these latent variables to build theories about cognition and expand these theories by connecting cognitive variables to other, external ones such as personality traits or psychopathological symptoms. The go-to statistical tool for doing so is regression analysis, answering questions such as, “Does trait x predict cognitive variable y?”. Problematically, however, estimates of the cognitive variables (e.g., from hierarchical Bayesian estimation) are often treated as perfectly measured, inducing bias in regression coefficients and reducing power. While statistical methods of attenuating bias and increasing power have been adopted in traditional psychometrics, methods for doing so in the CCM literature are underdeveloped. In this work, I propose and evaluate a method for increasing the reliability of the regression coefficient for models in which a CCM parameter (i.e., a cognitive variable) is predicted from some external variable.