Structural parameter interdependencies in computational models of cognition
Computational modeling of cognition allows measurement of latent psychological variables by means of free model parameters. The estimation and interpretation of these variables is impaired, however, if parameters strongly correlate with each other. We suggest that strong parameter intercorrelations are especially likely to emerge in models that combine a subjective value function with a probabilistic choice rule—a common structure in the literature. We demonstrate high intercorrelation between parameters in the value function and the probabilistic choice rule across several prominent computational models, including models on risky choice (cumulative prospect theory), categorization (the generalized context model), and memory (the SIMPLE model of free recall). Based on simulation studies, we show that the presence of parameter intercorrelations hampers estimation accuracy, in particular the ability to detect group differences on the parameters and to detect associations of the parameters with external variables. We show that these problems can be alleviated by changing the models’ error component, such as assuming parameter stochasticity or a constant error term. Our analyses highlight a common but often neglected problem of computational modeling of cognition and point to ways in which the design and application of such models can be improved.
Just wanted to leave a note of thanks for this talk. I (semi-)recently started working heavily with RL models, which are notorious for parameter correlations. I am very excited to see if your suggested remedies, especially parameter stochasticity, can help improve recovery, and thus support the reliability of my model-based inference. Do you...