Poster #29 - Findings from an updated reinforcement learning model of the Iowa Gambling Task

• In Event: 3-106 - Poster Session 11
  In Poster Session: PS 11 Section - Cognitive Processes

Integrative Statement

The Iowa Gambling Task (IGT) is used to assess decision-making in developmental samples (Beitz, Salthouse, & Davis, 2014; Buelow & Suhr, 2009). Much of this IGT research operationalizes performance by subtracting the number of selections from good decks by the number of selections from bad decks (Toplak, West, & Stanovich, 2013). This measure has shown much utility in describing development both cross-sectionally and longitudinally (e.g., Hooper, Luciana, Conklin, & Yarger, 2004; Almy et al., 2017). However, this metric may be punitive regarding early exploration that is an integral part of learning the task, while also conflating various cognitive processes involved with decision-making (e.g., Steingroever, Wetzels, Horstmann, Neumann, & Wagenmakers, 2013). Several reinforcement-learning models of IGT performance have been developed, however, these models rarely consider the role of development. Unfortunately, with all of the different combinations of utility functions, learning functions, and how information is integrated, there are twenty different models of IGT performance (Hart et al., under review). We previously demonstrated that some of these models are better than others in terms of describing the actual choices of individuals. A new utility function was reported in a recent paper (Weiss-Cohen et al., 2018) and when combined with the other best-fitting elements from other models, has shown better overall fit relative to the models tested in previous work. We report the findings from this new model among a developmental sample. The five parameters from this model were estimated using Markov Chain Monte Carlo analyses and age-related changes were observed. The sample included 181 typically developing individuals (ages 9-23; 100 females). Age-related changes were observed for three of the five parameters. With increasing age, individuals paid less attention to gains, were less averse to losses, and made less random decisions in terms of using utilities to guide choices. This pattern may suggest that the IGT could become less of an emotional task with age, as there is less “value” attached to wins and losses which may allow for greater consistency and better learning during the task. Interestingly, there were no age-related changes for memory/learning, suggesting that working memory may not be sufficient for explaining age-related improvement on the IGT. This finding is similar to previous work on age-related changes in working memory (e.g., Satterhwaite et al., 2013). Neural correlates of the five parameters will be explored. In previous work with a different reinforcement model, the consistency parameter demonstrated a significant relation with thickness in the medial orbitofrontal cortex, consistent with the original findings presented by Bechara and colleagues (1994). This new model should demonstrate a similar relation, as the modification from the previous model reflects a different way of assessing utility. In previous work, there were minimal associations between reward and loss and putative brain structures that hypothetically encode these outcomes (e.g., insula). With an updated reinforcement model, the parameters may be better aligned to the cognitive processes and neural structures. Overall, the combination of computational modeling and neural data provide a rich and specific look at IGT performance across adolescence.

Authors

• Brandon Almy, University of Minnesota
  Presenting Author

• Brian Hart, University of Minnesota
  Non-Presenting Author

• Michael Kuskowski, University of Minnesota
  Non-Presenting Author

• Paul Collins, University of Minnesota
  Non-Presenting Author

• Monica Luciana, University of Minnesota
  Non-Presenting Author