Absolute performance of reinforcement-learning models for the Iowa Gambling Task.

Abstract

Decision-making deficits in clinical populations are often studied using the Iowa Gambling Task (IGT). Performance on the IGT can be decomposed in its constituent psychological processes by means of cognitive modeling analyses. However, conclusions about the hypothesized psychological processes are valid only if the model provides an adequate account of the data. In this article, we systematically assessed absolute model performance of the Expectancy Valence (EV) model, the Prospect Valence Learning (PVL) model, and a hybrid version of both models—the PVL-Delta model—using 2 different methods. These methods assess (a) whether a model provides an acceptable fit to an observed choice pattern, and (b) whether the parameters obtained from model fitting can be used to generate the observed choice pattern. Our results show that all models provided an acceptable fit to 2 stylized data sets; however, when the model parameters were used to generate choices, only the PVL-Delta model captured the qualitative patterns in the data. These findings were confirmed by fitting the models to 5 published IGT data sets. Our results highlight that a model’s ability to fit a particular choice pattern does not guarantee that the model can also generate that same choice pattern. Future applications of RL models should carefully assess absolute model performance to avoid premature conclusions about the psychological processes that drive performance on the IGT.