On the efficiency of neurally-informed cognitive models to identify latent cognitive states

Abstract

Psychological theory is advanced through empirical tests of predictions derived from quantitative cognitive models. As cognitive models are developed and extended, they tend to increase in complexity–leading to more precise predictions–which places concomitant demands on the behavioral data used to discriminate between candidate theories. To aid discrimination between cognitive models and, more recently, to constrain parameter estimation, neural data have been used as an adjunct to behavioral data, or as a central stream of information, in the evaluation of cognitive models. Such a model-based neuroscience approach entails many advantages, including precise tests of hypotheses about brain–behavior relationships. There have, however, been few systematic investigations of the capacity for neural data to constrain the recovery of cognitive models. Through the lens of cognitive models of speeded decision-making, we investigated the efficiency of neural data to aid identification of latent cognitive states in models fit to behavioral data. We studied two theoretical frameworks that differed in their assumptions about the composition of the latent generating state. The first assumed that observed performance was generated from a mixture of discrete latent states. The second conceived of the latent state as dynamically varying along a continuous dimension. We used a simulation-based approach to compare recovery of latent data-generating states in neurally-informed versus neurally-uninformed cognitive models. We found that neurally-informed cognitive models were more reliably recovered under a discrete state representation than a continuous dimension representation for medium effect sizes, although recovery was difficult for small sample sizes and moderate noise in neural data. Recovery improved for both representations when a larger effect size differentiated the latent states. We conclude that neural data aids the identification of latent states in cognitive models, but different frameworks for quantitatively informing cognitive models with neural information have different model recovery efficiencies. We provide full worked examples and freely-available code to implement the two theoretical frameworks.